File: | tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp |
Warning: | line 8169, column 15 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===----- CGOpenMPRuntime.cpp - Interface to OpenMP Runtimes -------------===// | |||
2 | // | |||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
4 | // See https://llvm.org/LICENSE.txt for license information. | |||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
6 | // | |||
7 | //===----------------------------------------------------------------------===// | |||
8 | // | |||
9 | // This provides a class for OpenMP runtime code generation. | |||
10 | // | |||
11 | //===----------------------------------------------------------------------===// | |||
12 | ||||
13 | #include "CGCXXABI.h" | |||
14 | #include "CGCleanup.h" | |||
15 | #include "CGOpenMPRuntime.h" | |||
16 | #include "CGRecordLayout.h" | |||
17 | #include "CodeGenFunction.h" | |||
18 | #include "clang/CodeGen/ConstantInitBuilder.h" | |||
19 | #include "clang/AST/Decl.h" | |||
20 | #include "clang/AST/StmtOpenMP.h" | |||
21 | #include "clang/Basic/BitmaskEnum.h" | |||
22 | #include "llvm/ADT/ArrayRef.h" | |||
23 | #include "llvm/Bitcode/BitcodeReader.h" | |||
24 | #include "llvm/IR/DerivedTypes.h" | |||
25 | #include "llvm/IR/GlobalValue.h" | |||
26 | #include "llvm/IR/Value.h" | |||
27 | #include "llvm/Support/Format.h" | |||
28 | #include "llvm/Support/raw_ostream.h" | |||
29 | #include <cassert> | |||
30 | ||||
31 | using namespace clang; | |||
32 | using namespace CodeGen; | |||
33 | ||||
34 | namespace { | |||
35 | /// Base class for handling code generation inside OpenMP regions. | |||
36 | class CGOpenMPRegionInfo : public CodeGenFunction::CGCapturedStmtInfo { | |||
37 | public: | |||
38 | /// Kinds of OpenMP regions used in codegen. | |||
39 | enum CGOpenMPRegionKind { | |||
40 | /// Region with outlined function for standalone 'parallel' | |||
41 | /// directive. | |||
42 | ParallelOutlinedRegion, | |||
43 | /// Region with outlined function for standalone 'task' directive. | |||
44 | TaskOutlinedRegion, | |||
45 | /// Region for constructs that do not require function outlining, | |||
46 | /// like 'for', 'sections', 'atomic' etc. directives. | |||
47 | InlinedRegion, | |||
48 | /// Region with outlined function for standalone 'target' directive. | |||
49 | TargetRegion, | |||
50 | }; | |||
51 | ||||
52 | CGOpenMPRegionInfo(const CapturedStmt &CS, | |||
53 | const CGOpenMPRegionKind RegionKind, | |||
54 | const RegionCodeGenTy &CodeGen, OpenMPDirectiveKind Kind, | |||
55 | bool HasCancel) | |||
56 | : CGCapturedStmtInfo(CS, CR_OpenMP), RegionKind(RegionKind), | |||
57 | CodeGen(CodeGen), Kind(Kind), HasCancel(HasCancel) {} | |||
58 | ||||
59 | CGOpenMPRegionInfo(const CGOpenMPRegionKind RegionKind, | |||
60 | const RegionCodeGenTy &CodeGen, OpenMPDirectiveKind Kind, | |||
61 | bool HasCancel) | |||
62 | : CGCapturedStmtInfo(CR_OpenMP), RegionKind(RegionKind), CodeGen(CodeGen), | |||
63 | Kind(Kind), HasCancel(HasCancel) {} | |||
64 | ||||
65 | /// Get a variable or parameter for storing global thread id | |||
66 | /// inside OpenMP construct. | |||
67 | virtual const VarDecl *getThreadIDVariable() const = 0; | |||
68 | ||||
69 | /// Emit the captured statement body. | |||
70 | void EmitBody(CodeGenFunction &CGF, const Stmt *S) override; | |||
71 | ||||
72 | /// Get an LValue for the current ThreadID variable. | |||
73 | /// \return LValue for thread id variable. This LValue always has type int32*. | |||
74 | virtual LValue getThreadIDVariableLValue(CodeGenFunction &CGF); | |||
75 | ||||
76 | virtual void emitUntiedSwitch(CodeGenFunction & /*CGF*/) {} | |||
77 | ||||
78 | CGOpenMPRegionKind getRegionKind() const { return RegionKind; } | |||
79 | ||||
80 | OpenMPDirectiveKind getDirectiveKind() const { return Kind; } | |||
81 | ||||
82 | bool hasCancel() const { return HasCancel; } | |||
83 | ||||
84 | static bool classof(const CGCapturedStmtInfo *Info) { | |||
85 | return Info->getKind() == CR_OpenMP; | |||
86 | } | |||
87 | ||||
88 | ~CGOpenMPRegionInfo() override = default; | |||
89 | ||||
90 | protected: | |||
91 | CGOpenMPRegionKind RegionKind; | |||
92 | RegionCodeGenTy CodeGen; | |||
93 | OpenMPDirectiveKind Kind; | |||
94 | bool HasCancel; | |||
95 | }; | |||
96 | ||||
97 | /// API for captured statement code generation in OpenMP constructs. | |||
98 | class CGOpenMPOutlinedRegionInfo final : public CGOpenMPRegionInfo { | |||
99 | public: | |||
100 | CGOpenMPOutlinedRegionInfo(const CapturedStmt &CS, const VarDecl *ThreadIDVar, | |||
101 | const RegionCodeGenTy &CodeGen, | |||
102 | OpenMPDirectiveKind Kind, bool HasCancel, | |||
103 | StringRef HelperName) | |||
104 | : CGOpenMPRegionInfo(CS, ParallelOutlinedRegion, CodeGen, Kind, | |||
105 | HasCancel), | |||
106 | ThreadIDVar(ThreadIDVar), HelperName(HelperName) { | |||
107 | assert(ThreadIDVar != nullptr && "No ThreadID in OpenMP region.")((ThreadIDVar != nullptr && "No ThreadID in OpenMP region." ) ? static_cast<void> (0) : __assert_fail ("ThreadIDVar != nullptr && \"No ThreadID in OpenMP region.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 107, __PRETTY_FUNCTION__)); | |||
108 | } | |||
109 | ||||
110 | /// Get a variable or parameter for storing global thread id | |||
111 | /// inside OpenMP construct. | |||
112 | const VarDecl *getThreadIDVariable() const override { return ThreadIDVar; } | |||
113 | ||||
114 | /// Get the name of the capture helper. | |||
115 | StringRef getHelperName() const override { return HelperName; } | |||
116 | ||||
117 | static bool classof(const CGCapturedStmtInfo *Info) { | |||
118 | return CGOpenMPRegionInfo::classof(Info) && | |||
119 | cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == | |||
120 | ParallelOutlinedRegion; | |||
121 | } | |||
122 | ||||
123 | private: | |||
124 | /// A variable or parameter storing global thread id for OpenMP | |||
125 | /// constructs. | |||
126 | const VarDecl *ThreadIDVar; | |||
127 | StringRef HelperName; | |||
128 | }; | |||
129 | ||||
130 | /// API for captured statement code generation in OpenMP constructs. | |||
131 | class CGOpenMPTaskOutlinedRegionInfo final : public CGOpenMPRegionInfo { | |||
132 | public: | |||
133 | class UntiedTaskActionTy final : public PrePostActionTy { | |||
134 | bool Untied; | |||
135 | const VarDecl *PartIDVar; | |||
136 | const RegionCodeGenTy UntiedCodeGen; | |||
137 | llvm::SwitchInst *UntiedSwitch = nullptr; | |||
138 | ||||
139 | public: | |||
140 | UntiedTaskActionTy(bool Tied, const VarDecl *PartIDVar, | |||
141 | const RegionCodeGenTy &UntiedCodeGen) | |||
142 | : Untied(!Tied), PartIDVar(PartIDVar), UntiedCodeGen(UntiedCodeGen) {} | |||
143 | void Enter(CodeGenFunction &CGF) override { | |||
144 | if (Untied) { | |||
145 | // Emit task switching point. | |||
146 | LValue PartIdLVal = CGF.EmitLoadOfPointerLValue( | |||
147 | CGF.GetAddrOfLocalVar(PartIDVar), | |||
148 | PartIDVar->getType()->castAs<PointerType>()); | |||
149 | llvm::Value *Res = | |||
150 | CGF.EmitLoadOfScalar(PartIdLVal, PartIDVar->getLocation()); | |||
151 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock(".untied.done."); | |||
152 | UntiedSwitch = CGF.Builder.CreateSwitch(Res, DoneBB); | |||
153 | CGF.EmitBlock(DoneBB); | |||
154 | CGF.EmitBranchThroughCleanup(CGF.ReturnBlock); | |||
155 | CGF.EmitBlock(CGF.createBasicBlock(".untied.jmp.")); | |||
156 | UntiedSwitch->addCase(CGF.Builder.getInt32(0), | |||
157 | CGF.Builder.GetInsertBlock()); | |||
158 | emitUntiedSwitch(CGF); | |||
159 | } | |||
160 | } | |||
161 | void emitUntiedSwitch(CodeGenFunction &CGF) const { | |||
162 | if (Untied) { | |||
163 | LValue PartIdLVal = CGF.EmitLoadOfPointerLValue( | |||
164 | CGF.GetAddrOfLocalVar(PartIDVar), | |||
165 | PartIDVar->getType()->castAs<PointerType>()); | |||
166 | CGF.EmitStoreOfScalar(CGF.Builder.getInt32(UntiedSwitch->getNumCases()), | |||
167 | PartIdLVal); | |||
168 | UntiedCodeGen(CGF); | |||
169 | CodeGenFunction::JumpDest CurPoint = | |||
170 | CGF.getJumpDestInCurrentScope(".untied.next."); | |||
171 | CGF.EmitBranchThroughCleanup(CGF.ReturnBlock); | |||
172 | CGF.EmitBlock(CGF.createBasicBlock(".untied.jmp.")); | |||
173 | UntiedSwitch->addCase(CGF.Builder.getInt32(UntiedSwitch->getNumCases()), | |||
174 | CGF.Builder.GetInsertBlock()); | |||
175 | CGF.EmitBranchThroughCleanup(CurPoint); | |||
176 | CGF.EmitBlock(CurPoint.getBlock()); | |||
177 | } | |||
178 | } | |||
179 | unsigned getNumberOfParts() const { return UntiedSwitch->getNumCases(); } | |||
180 | }; | |||
181 | CGOpenMPTaskOutlinedRegionInfo(const CapturedStmt &CS, | |||
182 | const VarDecl *ThreadIDVar, | |||
183 | const RegionCodeGenTy &CodeGen, | |||
184 | OpenMPDirectiveKind Kind, bool HasCancel, | |||
185 | const UntiedTaskActionTy &Action) | |||
186 | : CGOpenMPRegionInfo(CS, TaskOutlinedRegion, CodeGen, Kind, HasCancel), | |||
187 | ThreadIDVar(ThreadIDVar), Action(Action) { | |||
188 | assert(ThreadIDVar != nullptr && "No ThreadID in OpenMP region.")((ThreadIDVar != nullptr && "No ThreadID in OpenMP region." ) ? static_cast<void> (0) : __assert_fail ("ThreadIDVar != nullptr && \"No ThreadID in OpenMP region.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 188, __PRETTY_FUNCTION__)); | |||
189 | } | |||
190 | ||||
191 | /// Get a variable or parameter for storing global thread id | |||
192 | /// inside OpenMP construct. | |||
193 | const VarDecl *getThreadIDVariable() const override { return ThreadIDVar; } | |||
194 | ||||
195 | /// Get an LValue for the current ThreadID variable. | |||
196 | LValue getThreadIDVariableLValue(CodeGenFunction &CGF) override; | |||
197 | ||||
198 | /// Get the name of the capture helper. | |||
199 | StringRef getHelperName() const override { return ".omp_outlined."; } | |||
200 | ||||
201 | void emitUntiedSwitch(CodeGenFunction &CGF) override { | |||
202 | Action.emitUntiedSwitch(CGF); | |||
203 | } | |||
204 | ||||
205 | static bool classof(const CGCapturedStmtInfo *Info) { | |||
206 | return CGOpenMPRegionInfo::classof(Info) && | |||
207 | cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == | |||
208 | TaskOutlinedRegion; | |||
209 | } | |||
210 | ||||
211 | private: | |||
212 | /// A variable or parameter storing global thread id for OpenMP | |||
213 | /// constructs. | |||
214 | const VarDecl *ThreadIDVar; | |||
215 | /// Action for emitting code for untied tasks. | |||
216 | const UntiedTaskActionTy &Action; | |||
217 | }; | |||
218 | ||||
219 | /// API for inlined captured statement code generation in OpenMP | |||
220 | /// constructs. | |||
221 | class CGOpenMPInlinedRegionInfo : public CGOpenMPRegionInfo { | |||
222 | public: | |||
223 | CGOpenMPInlinedRegionInfo(CodeGenFunction::CGCapturedStmtInfo *OldCSI, | |||
224 | const RegionCodeGenTy &CodeGen, | |||
225 | OpenMPDirectiveKind Kind, bool HasCancel) | |||
226 | : CGOpenMPRegionInfo(InlinedRegion, CodeGen, Kind, HasCancel), | |||
227 | OldCSI(OldCSI), | |||
228 | OuterRegionInfo(dyn_cast_or_null<CGOpenMPRegionInfo>(OldCSI)) {} | |||
229 | ||||
230 | // Retrieve the value of the context parameter. | |||
231 | llvm::Value *getContextValue() const override { | |||
232 | if (OuterRegionInfo) | |||
233 | return OuterRegionInfo->getContextValue(); | |||
234 | llvm_unreachable("No context value for inlined OpenMP region")::llvm::llvm_unreachable_internal("No context value for inlined OpenMP region" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 234); | |||
235 | } | |||
236 | ||||
237 | void setContextValue(llvm::Value *V) override { | |||
238 | if (OuterRegionInfo) { | |||
239 | OuterRegionInfo->setContextValue(V); | |||
240 | return; | |||
241 | } | |||
242 | llvm_unreachable("No context value for inlined OpenMP region")::llvm::llvm_unreachable_internal("No context value for inlined OpenMP region" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 242); | |||
243 | } | |||
244 | ||||
245 | /// Lookup the captured field decl for a variable. | |||
246 | const FieldDecl *lookup(const VarDecl *VD) const override { | |||
247 | if (OuterRegionInfo) | |||
248 | return OuterRegionInfo->lookup(VD); | |||
249 | // If there is no outer outlined region,no need to lookup in a list of | |||
250 | // captured variables, we can use the original one. | |||
251 | return nullptr; | |||
252 | } | |||
253 | ||||
254 | FieldDecl *getThisFieldDecl() const override { | |||
255 | if (OuterRegionInfo) | |||
256 | return OuterRegionInfo->getThisFieldDecl(); | |||
257 | return nullptr; | |||
258 | } | |||
259 | ||||
260 | /// Get a variable or parameter for storing global thread id | |||
261 | /// inside OpenMP construct. | |||
262 | const VarDecl *getThreadIDVariable() const override { | |||
263 | if (OuterRegionInfo) | |||
264 | return OuterRegionInfo->getThreadIDVariable(); | |||
265 | return nullptr; | |||
266 | } | |||
267 | ||||
268 | /// Get an LValue for the current ThreadID variable. | |||
269 | LValue getThreadIDVariableLValue(CodeGenFunction &CGF) override { | |||
270 | if (OuterRegionInfo) | |||
271 | return OuterRegionInfo->getThreadIDVariableLValue(CGF); | |||
272 | llvm_unreachable("No LValue for inlined OpenMP construct")::llvm::llvm_unreachable_internal("No LValue for inlined OpenMP construct" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 272); | |||
273 | } | |||
274 | ||||
275 | /// Get the name of the capture helper. | |||
276 | StringRef getHelperName() const override { | |||
277 | if (auto *OuterRegionInfo = getOldCSI()) | |||
278 | return OuterRegionInfo->getHelperName(); | |||
279 | llvm_unreachable("No helper name for inlined OpenMP construct")::llvm::llvm_unreachable_internal("No helper name for inlined OpenMP construct" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 279); | |||
280 | } | |||
281 | ||||
282 | void emitUntiedSwitch(CodeGenFunction &CGF) override { | |||
283 | if (OuterRegionInfo) | |||
284 | OuterRegionInfo->emitUntiedSwitch(CGF); | |||
285 | } | |||
286 | ||||
287 | CodeGenFunction::CGCapturedStmtInfo *getOldCSI() const { return OldCSI; } | |||
288 | ||||
289 | static bool classof(const CGCapturedStmtInfo *Info) { | |||
290 | return CGOpenMPRegionInfo::classof(Info) && | |||
291 | cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == InlinedRegion; | |||
292 | } | |||
293 | ||||
294 | ~CGOpenMPInlinedRegionInfo() override = default; | |||
295 | ||||
296 | private: | |||
297 | /// CodeGen info about outer OpenMP region. | |||
298 | CodeGenFunction::CGCapturedStmtInfo *OldCSI; | |||
299 | CGOpenMPRegionInfo *OuterRegionInfo; | |||
300 | }; | |||
301 | ||||
302 | /// API for captured statement code generation in OpenMP target | |||
303 | /// constructs. For this captures, implicit parameters are used instead of the | |||
304 | /// captured fields. The name of the target region has to be unique in a given | |||
305 | /// application so it is provided by the client, because only the client has | |||
306 | /// the information to generate that. | |||
307 | class CGOpenMPTargetRegionInfo final : public CGOpenMPRegionInfo { | |||
308 | public: | |||
309 | CGOpenMPTargetRegionInfo(const CapturedStmt &CS, | |||
310 | const RegionCodeGenTy &CodeGen, StringRef HelperName) | |||
311 | : CGOpenMPRegionInfo(CS, TargetRegion, CodeGen, OMPD_target, | |||
312 | /*HasCancel=*/false), | |||
313 | HelperName(HelperName) {} | |||
314 | ||||
315 | /// This is unused for target regions because each starts executing | |||
316 | /// with a single thread. | |||
317 | const VarDecl *getThreadIDVariable() const override { return nullptr; } | |||
318 | ||||
319 | /// Get the name of the capture helper. | |||
320 | StringRef getHelperName() const override { return HelperName; } | |||
321 | ||||
322 | static bool classof(const CGCapturedStmtInfo *Info) { | |||
323 | return CGOpenMPRegionInfo::classof(Info) && | |||
324 | cast<CGOpenMPRegionInfo>(Info)->getRegionKind() == TargetRegion; | |||
325 | } | |||
326 | ||||
327 | private: | |||
328 | StringRef HelperName; | |||
329 | }; | |||
330 | ||||
331 | static void EmptyCodeGen(CodeGenFunction &, PrePostActionTy &) { | |||
332 | llvm_unreachable("No codegen for expressions")::llvm::llvm_unreachable_internal("No codegen for expressions" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 332); | |||
333 | } | |||
334 | /// API for generation of expressions captured in a innermost OpenMP | |||
335 | /// region. | |||
336 | class CGOpenMPInnerExprInfo final : public CGOpenMPInlinedRegionInfo { | |||
337 | public: | |||
338 | CGOpenMPInnerExprInfo(CodeGenFunction &CGF, const CapturedStmt &CS) | |||
339 | : CGOpenMPInlinedRegionInfo(CGF.CapturedStmtInfo, EmptyCodeGen, | |||
340 | OMPD_unknown, | |||
341 | /*HasCancel=*/false), | |||
342 | PrivScope(CGF) { | |||
343 | // Make sure the globals captured in the provided statement are local by | |||
344 | // using the privatization logic. We assume the same variable is not | |||
345 | // captured more than once. | |||
346 | for (const auto &C : CS.captures()) { | |||
347 | if (!C.capturesVariable() && !C.capturesVariableByCopy()) | |||
348 | continue; | |||
349 | ||||
350 | const VarDecl *VD = C.getCapturedVar(); | |||
351 | if (VD->isLocalVarDeclOrParm()) | |||
352 | continue; | |||
353 | ||||
354 | DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(VD), | |||
355 | /*RefersToEnclosingVariableOrCapture=*/false, | |||
356 | VD->getType().getNonReferenceType(), VK_LValue, | |||
357 | C.getLocation()); | |||
358 | PrivScope.addPrivate( | |||
359 | VD, [&CGF, &DRE]() { return CGF.EmitLValue(&DRE).getAddress(); }); | |||
360 | } | |||
361 | (void)PrivScope.Privatize(); | |||
362 | } | |||
363 | ||||
364 | /// Lookup the captured field decl for a variable. | |||
365 | const FieldDecl *lookup(const VarDecl *VD) const override { | |||
366 | if (const FieldDecl *FD = CGOpenMPInlinedRegionInfo::lookup(VD)) | |||
367 | return FD; | |||
368 | return nullptr; | |||
369 | } | |||
370 | ||||
371 | /// Emit the captured statement body. | |||
372 | void EmitBody(CodeGenFunction &CGF, const Stmt *S) override { | |||
373 | llvm_unreachable("No body for expressions")::llvm::llvm_unreachable_internal("No body for expressions", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 373); | |||
374 | } | |||
375 | ||||
376 | /// Get a variable or parameter for storing global thread id | |||
377 | /// inside OpenMP construct. | |||
378 | const VarDecl *getThreadIDVariable() const override { | |||
379 | llvm_unreachable("No thread id for expressions")::llvm::llvm_unreachable_internal("No thread id for expressions" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 379); | |||
380 | } | |||
381 | ||||
382 | /// Get the name of the capture helper. | |||
383 | StringRef getHelperName() const override { | |||
384 | llvm_unreachable("No helper name for expressions")::llvm::llvm_unreachable_internal("No helper name for expressions" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 384); | |||
385 | } | |||
386 | ||||
387 | static bool classof(const CGCapturedStmtInfo *Info) { return false; } | |||
388 | ||||
389 | private: | |||
390 | /// Private scope to capture global variables. | |||
391 | CodeGenFunction::OMPPrivateScope PrivScope; | |||
392 | }; | |||
393 | ||||
394 | /// RAII for emitting code of OpenMP constructs. | |||
395 | class InlinedOpenMPRegionRAII { | |||
396 | CodeGenFunction &CGF; | |||
397 | llvm::DenseMap<const VarDecl *, FieldDecl *> LambdaCaptureFields; | |||
398 | FieldDecl *LambdaThisCaptureField = nullptr; | |||
399 | const CodeGen::CGBlockInfo *BlockInfo = nullptr; | |||
400 | ||||
401 | public: | |||
402 | /// Constructs region for combined constructs. | |||
403 | /// \param CodeGen Code generation sequence for combined directives. Includes | |||
404 | /// a list of functions used for code generation of implicitly inlined | |||
405 | /// regions. | |||
406 | InlinedOpenMPRegionRAII(CodeGenFunction &CGF, const RegionCodeGenTy &CodeGen, | |||
407 | OpenMPDirectiveKind Kind, bool HasCancel) | |||
408 | : CGF(CGF) { | |||
409 | // Start emission for the construct. | |||
410 | CGF.CapturedStmtInfo = new CGOpenMPInlinedRegionInfo( | |||
411 | CGF.CapturedStmtInfo, CodeGen, Kind, HasCancel); | |||
412 | std::swap(CGF.LambdaCaptureFields, LambdaCaptureFields); | |||
413 | LambdaThisCaptureField = CGF.LambdaThisCaptureField; | |||
414 | CGF.LambdaThisCaptureField = nullptr; | |||
415 | BlockInfo = CGF.BlockInfo; | |||
416 | CGF.BlockInfo = nullptr; | |||
417 | } | |||
418 | ||||
419 | ~InlinedOpenMPRegionRAII() { | |||
420 | // Restore original CapturedStmtInfo only if we're done with code emission. | |||
421 | auto *OldCSI = | |||
422 | cast<CGOpenMPInlinedRegionInfo>(CGF.CapturedStmtInfo)->getOldCSI(); | |||
423 | delete CGF.CapturedStmtInfo; | |||
424 | CGF.CapturedStmtInfo = OldCSI; | |||
425 | std::swap(CGF.LambdaCaptureFields, LambdaCaptureFields); | |||
426 | CGF.LambdaThisCaptureField = LambdaThisCaptureField; | |||
427 | CGF.BlockInfo = BlockInfo; | |||
428 | } | |||
429 | }; | |||
430 | ||||
431 | /// Values for bit flags used in the ident_t to describe the fields. | |||
432 | /// All enumeric elements are named and described in accordance with the code | |||
433 | /// from https://github.com/llvm/llvm-project/blob/master/openmp/runtime/src/kmp.h | |||
434 | enum OpenMPLocationFlags : unsigned { | |||
435 | /// Use trampoline for internal microtask. | |||
436 | OMP_IDENT_IMD = 0x01, | |||
437 | /// Use c-style ident structure. | |||
438 | OMP_IDENT_KMPC = 0x02, | |||
439 | /// Atomic reduction option for kmpc_reduce. | |||
440 | OMP_ATOMIC_REDUCE = 0x10, | |||
441 | /// Explicit 'barrier' directive. | |||
442 | OMP_IDENT_BARRIER_EXPL = 0x20, | |||
443 | /// Implicit barrier in code. | |||
444 | OMP_IDENT_BARRIER_IMPL = 0x40, | |||
445 | /// Implicit barrier in 'for' directive. | |||
446 | OMP_IDENT_BARRIER_IMPL_FOR = 0x40, | |||
447 | /// Implicit barrier in 'sections' directive. | |||
448 | OMP_IDENT_BARRIER_IMPL_SECTIONS = 0xC0, | |||
449 | /// Implicit barrier in 'single' directive. | |||
450 | OMP_IDENT_BARRIER_IMPL_SINGLE = 0x140, | |||
451 | /// Call of __kmp_for_static_init for static loop. | |||
452 | OMP_IDENT_WORK_LOOP = 0x200, | |||
453 | /// Call of __kmp_for_static_init for sections. | |||
454 | OMP_IDENT_WORK_SECTIONS = 0x400, | |||
455 | /// Call of __kmp_for_static_init for distribute. | |||
456 | OMP_IDENT_WORK_DISTRIBUTE = 0x800, | |||
457 | LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/OMP_IDENT_WORK_DISTRIBUTE)LLVM_BITMASK_LARGEST_ENUMERATOR = OMP_IDENT_WORK_DISTRIBUTE | |||
458 | }; | |||
459 | ||||
460 | namespace { | |||
461 | LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE()using ::llvm::BitmaskEnumDetail::operator~; using ::llvm::BitmaskEnumDetail ::operator|; using ::llvm::BitmaskEnumDetail::operator&; using ::llvm::BitmaskEnumDetail::operator^; using ::llvm::BitmaskEnumDetail ::operator|=; using ::llvm::BitmaskEnumDetail::operator&= ; using ::llvm::BitmaskEnumDetail::operator^=; | |||
462 | /// Values for bit flags for marking which requires clauses have been used. | |||
463 | enum OpenMPOffloadingRequiresDirFlags : int64_t { | |||
464 | /// flag undefined. | |||
465 | OMP_REQ_UNDEFINED = 0x000, | |||
466 | /// no requires clause present. | |||
467 | OMP_REQ_NONE = 0x001, | |||
468 | /// reverse_offload clause. | |||
469 | OMP_REQ_REVERSE_OFFLOAD = 0x002, | |||
470 | /// unified_address clause. | |||
471 | OMP_REQ_UNIFIED_ADDRESS = 0x004, | |||
472 | /// unified_shared_memory clause. | |||
473 | OMP_REQ_UNIFIED_SHARED_MEMORY = 0x008, | |||
474 | /// dynamic_allocators clause. | |||
475 | OMP_REQ_DYNAMIC_ALLOCATORS = 0x010, | |||
476 | LLVM_MARK_AS_BITMASK_ENUM(/*LargestValue=*/OMP_REQ_DYNAMIC_ALLOCATORS)LLVM_BITMASK_LARGEST_ENUMERATOR = OMP_REQ_DYNAMIC_ALLOCATORS | |||
477 | }; | |||
478 | } // anonymous namespace | |||
479 | ||||
480 | /// Describes ident structure that describes a source location. | |||
481 | /// All descriptions are taken from | |||
482 | /// https://github.com/llvm/llvm-project/blob/master/openmp/runtime/src/kmp.h | |||
483 | /// Original structure: | |||
484 | /// typedef struct ident { | |||
485 | /// kmp_int32 reserved_1; /**< might be used in Fortran; | |||
486 | /// see above */ | |||
487 | /// kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags; | |||
488 | /// KMP_IDENT_KMPC identifies this union | |||
489 | /// member */ | |||
490 | /// kmp_int32 reserved_2; /**< not really used in Fortran any more; | |||
491 | /// see above */ | |||
492 | ///#if USE_ITT_BUILD | |||
493 | /// /* but currently used for storing | |||
494 | /// region-specific ITT */ | |||
495 | /// /* contextual information. */ | |||
496 | ///#endif /* USE_ITT_BUILD */ | |||
497 | /// kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for | |||
498 | /// C++ */ | |||
499 | /// char const *psource; /**< String describing the source location. | |||
500 | /// The string is composed of semi-colon separated | |||
501 | // fields which describe the source file, | |||
502 | /// the function and a pair of line numbers that | |||
503 | /// delimit the construct. | |||
504 | /// */ | |||
505 | /// } ident_t; | |||
506 | enum IdentFieldIndex { | |||
507 | /// might be used in Fortran | |||
508 | IdentField_Reserved_1, | |||
509 | /// OMP_IDENT_xxx flags; OMP_IDENT_KMPC identifies this union member. | |||
510 | IdentField_Flags, | |||
511 | /// Not really used in Fortran any more | |||
512 | IdentField_Reserved_2, | |||
513 | /// Source[4] in Fortran, do not use for C++ | |||
514 | IdentField_Reserved_3, | |||
515 | /// String describing the source location. The string is composed of | |||
516 | /// semi-colon separated fields which describe the source file, the function | |||
517 | /// and a pair of line numbers that delimit the construct. | |||
518 | IdentField_PSource | |||
519 | }; | |||
520 | ||||
521 | /// Schedule types for 'omp for' loops (these enumerators are taken from | |||
522 | /// the enum sched_type in kmp.h). | |||
523 | enum OpenMPSchedType { | |||
524 | /// Lower bound for default (unordered) versions. | |||
525 | OMP_sch_lower = 32, | |||
526 | OMP_sch_static_chunked = 33, | |||
527 | OMP_sch_static = 34, | |||
528 | OMP_sch_dynamic_chunked = 35, | |||
529 | OMP_sch_guided_chunked = 36, | |||
530 | OMP_sch_runtime = 37, | |||
531 | OMP_sch_auto = 38, | |||
532 | /// static with chunk adjustment (e.g., simd) | |||
533 | OMP_sch_static_balanced_chunked = 45, | |||
534 | /// Lower bound for 'ordered' versions. | |||
535 | OMP_ord_lower = 64, | |||
536 | OMP_ord_static_chunked = 65, | |||
537 | OMP_ord_static = 66, | |||
538 | OMP_ord_dynamic_chunked = 67, | |||
539 | OMP_ord_guided_chunked = 68, | |||
540 | OMP_ord_runtime = 69, | |||
541 | OMP_ord_auto = 70, | |||
542 | OMP_sch_default = OMP_sch_static, | |||
543 | /// dist_schedule types | |||
544 | OMP_dist_sch_static_chunked = 91, | |||
545 | OMP_dist_sch_static = 92, | |||
546 | /// Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. | |||
547 | /// Set if the monotonic schedule modifier was present. | |||
548 | OMP_sch_modifier_monotonic = (1 << 29), | |||
549 | /// Set if the nonmonotonic schedule modifier was present. | |||
550 | OMP_sch_modifier_nonmonotonic = (1 << 30), | |||
551 | }; | |||
552 | ||||
553 | enum OpenMPRTLFunction { | |||
554 | /// Call to void __kmpc_fork_call(ident_t *loc, kmp_int32 argc, | |||
555 | /// kmpc_micro microtask, ...); | |||
556 | OMPRTL__kmpc_fork_call, | |||
557 | /// Call to void *__kmpc_threadprivate_cached(ident_t *loc, | |||
558 | /// kmp_int32 global_tid, void *data, size_t size, void ***cache); | |||
559 | OMPRTL__kmpc_threadprivate_cached, | |||
560 | /// Call to void __kmpc_threadprivate_register( ident_t *, | |||
561 | /// void *data, kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor); | |||
562 | OMPRTL__kmpc_threadprivate_register, | |||
563 | // Call to __kmpc_int32 kmpc_global_thread_num(ident_t *loc); | |||
564 | OMPRTL__kmpc_global_thread_num, | |||
565 | // Call to void __kmpc_critical(ident_t *loc, kmp_int32 global_tid, | |||
566 | // kmp_critical_name *crit); | |||
567 | OMPRTL__kmpc_critical, | |||
568 | // Call to void __kmpc_critical_with_hint(ident_t *loc, kmp_int32 | |||
569 | // global_tid, kmp_critical_name *crit, uintptr_t hint); | |||
570 | OMPRTL__kmpc_critical_with_hint, | |||
571 | // Call to void __kmpc_end_critical(ident_t *loc, kmp_int32 global_tid, | |||
572 | // kmp_critical_name *crit); | |||
573 | OMPRTL__kmpc_end_critical, | |||
574 | // Call to kmp_int32 __kmpc_cancel_barrier(ident_t *loc, kmp_int32 | |||
575 | // global_tid); | |||
576 | OMPRTL__kmpc_cancel_barrier, | |||
577 | // Call to void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid); | |||
578 | OMPRTL__kmpc_barrier, | |||
579 | // Call to void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid); | |||
580 | OMPRTL__kmpc_for_static_fini, | |||
581 | // Call to void __kmpc_serialized_parallel(ident_t *loc, kmp_int32 | |||
582 | // global_tid); | |||
583 | OMPRTL__kmpc_serialized_parallel, | |||
584 | // Call to void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32 | |||
585 | // global_tid); | |||
586 | OMPRTL__kmpc_end_serialized_parallel, | |||
587 | // Call to void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, | |||
588 | // kmp_int32 num_threads); | |||
589 | OMPRTL__kmpc_push_num_threads, | |||
590 | // Call to void __kmpc_flush(ident_t *loc); | |||
591 | OMPRTL__kmpc_flush, | |||
592 | // Call to kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid); | |||
593 | OMPRTL__kmpc_master, | |||
594 | // Call to void __kmpc_end_master(ident_t *, kmp_int32 global_tid); | |||
595 | OMPRTL__kmpc_end_master, | |||
596 | // Call to kmp_int32 __kmpc_omp_taskyield(ident_t *, kmp_int32 global_tid, | |||
597 | // int end_part); | |||
598 | OMPRTL__kmpc_omp_taskyield, | |||
599 | // Call to kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid); | |||
600 | OMPRTL__kmpc_single, | |||
601 | // Call to void __kmpc_end_single(ident_t *, kmp_int32 global_tid); | |||
602 | OMPRTL__kmpc_end_single, | |||
603 | // Call to kmp_task_t * __kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, | |||
604 | // kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, | |||
605 | // kmp_routine_entry_t *task_entry); | |||
606 | OMPRTL__kmpc_omp_task_alloc, | |||
607 | // Call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t * | |||
608 | // new_task); | |||
609 | OMPRTL__kmpc_omp_task, | |||
610 | // Call to void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid, | |||
611 | // size_t cpy_size, void *cpy_data, void(*cpy_func)(void *, void *), | |||
612 | // kmp_int32 didit); | |||
613 | OMPRTL__kmpc_copyprivate, | |||
614 | // Call to kmp_int32 __kmpc_reduce(ident_t *loc, kmp_int32 global_tid, | |||
615 | // kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void | |||
616 | // (*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck); | |||
617 | OMPRTL__kmpc_reduce, | |||
618 | // Call to kmp_int32 __kmpc_reduce_nowait(ident_t *loc, kmp_int32 | |||
619 | // global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, | |||
620 | // void (*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name | |||
621 | // *lck); | |||
622 | OMPRTL__kmpc_reduce_nowait, | |||
623 | // Call to void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, | |||
624 | // kmp_critical_name *lck); | |||
625 | OMPRTL__kmpc_end_reduce, | |||
626 | // Call to void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, | |||
627 | // kmp_critical_name *lck); | |||
628 | OMPRTL__kmpc_end_reduce_nowait, | |||
629 | // Call to void __kmpc_omp_task_begin_if0(ident_t *, kmp_int32 gtid, | |||
630 | // kmp_task_t * new_task); | |||
631 | OMPRTL__kmpc_omp_task_begin_if0, | |||
632 | // Call to void __kmpc_omp_task_complete_if0(ident_t *, kmp_int32 gtid, | |||
633 | // kmp_task_t * new_task); | |||
634 | OMPRTL__kmpc_omp_task_complete_if0, | |||
635 | // Call to void __kmpc_ordered(ident_t *loc, kmp_int32 global_tid); | |||
636 | OMPRTL__kmpc_ordered, | |||
637 | // Call to void __kmpc_end_ordered(ident_t *loc, kmp_int32 global_tid); | |||
638 | OMPRTL__kmpc_end_ordered, | |||
639 | // Call to kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32 | |||
640 | // global_tid); | |||
641 | OMPRTL__kmpc_omp_taskwait, | |||
642 | // Call to void __kmpc_taskgroup(ident_t *loc, kmp_int32 global_tid); | |||
643 | OMPRTL__kmpc_taskgroup, | |||
644 | // Call to void __kmpc_end_taskgroup(ident_t *loc, kmp_int32 global_tid); | |||
645 | OMPRTL__kmpc_end_taskgroup, | |||
646 | // Call to void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid, | |||
647 | // int proc_bind); | |||
648 | OMPRTL__kmpc_push_proc_bind, | |||
649 | // Call to kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 | |||
650 | // gtid, kmp_task_t * new_task, kmp_int32 ndeps, kmp_depend_info_t | |||
651 | // *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list); | |||
652 | OMPRTL__kmpc_omp_task_with_deps, | |||
653 | // Call to void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 | |||
654 | // gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 | |||
655 | // ndeps_noalias, kmp_depend_info_t *noalias_dep_list); | |||
656 | OMPRTL__kmpc_omp_wait_deps, | |||
657 | // Call to kmp_int32 __kmpc_cancellationpoint(ident_t *loc, kmp_int32 | |||
658 | // global_tid, kmp_int32 cncl_kind); | |||
659 | OMPRTL__kmpc_cancellationpoint, | |||
660 | // Call to kmp_int32 __kmpc_cancel(ident_t *loc, kmp_int32 global_tid, | |||
661 | // kmp_int32 cncl_kind); | |||
662 | OMPRTL__kmpc_cancel, | |||
663 | // Call to void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, | |||
664 | // kmp_int32 num_teams, kmp_int32 thread_limit); | |||
665 | OMPRTL__kmpc_push_num_teams, | |||
666 | // Call to void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro | |||
667 | // microtask, ...); | |||
668 | OMPRTL__kmpc_fork_teams, | |||
669 | // Call to void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int | |||
670 | // if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int | |||
671 | // sched, kmp_uint64 grainsize, void *task_dup); | |||
672 | OMPRTL__kmpc_taskloop, | |||
673 | // Call to void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid, kmp_int32 | |||
674 | // num_dims, struct kmp_dim *dims); | |||
675 | OMPRTL__kmpc_doacross_init, | |||
676 | // Call to void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid); | |||
677 | OMPRTL__kmpc_doacross_fini, | |||
678 | // Call to void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid, kmp_int64 | |||
679 | // *vec); | |||
680 | OMPRTL__kmpc_doacross_post, | |||
681 | // Call to void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid, kmp_int64 | |||
682 | // *vec); | |||
683 | OMPRTL__kmpc_doacross_wait, | |||
684 | // Call to void *__kmpc_task_reduction_init(int gtid, int num_data, void | |||
685 | // *data); | |||
686 | OMPRTL__kmpc_task_reduction_init, | |||
687 | // Call to void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void | |||
688 | // *d); | |||
689 | OMPRTL__kmpc_task_reduction_get_th_data, | |||
690 | // Call to void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al); | |||
691 | OMPRTL__kmpc_alloc, | |||
692 | // Call to void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al); | |||
693 | OMPRTL__kmpc_free, | |||
694 | ||||
695 | // | |||
696 | // Offloading related calls | |||
697 | // | |||
698 | // Call to void __kmpc_push_target_tripcount(int64_t device_id, kmp_uint64 | |||
699 | // size); | |||
700 | OMPRTL__kmpc_push_target_tripcount, | |||
701 | // Call to int32_t __tgt_target(int64_t device_id, void *host_ptr, int32_t | |||
702 | // arg_num, void** args_base, void **args, size_t *arg_sizes, int64_t | |||
703 | // *arg_types); | |||
704 | OMPRTL__tgt_target, | |||
705 | // Call to int32_t __tgt_target_nowait(int64_t device_id, void *host_ptr, | |||
706 | // int32_t arg_num, void** args_base, void **args, size_t *arg_sizes, int64_t | |||
707 | // *arg_types); | |||
708 | OMPRTL__tgt_target_nowait, | |||
709 | // Call to int32_t __tgt_target_teams(int64_t device_id, void *host_ptr, | |||
710 | // int32_t arg_num, void** args_base, void **args, size_t *arg_sizes, int64_t | |||
711 | // *arg_types, int32_t num_teams, int32_t thread_limit); | |||
712 | OMPRTL__tgt_target_teams, | |||
713 | // Call to int32_t __tgt_target_teams_nowait(int64_t device_id, void | |||
714 | // *host_ptr, int32_t arg_num, void** args_base, void **args, size_t | |||
715 | // *arg_sizes, int64_t *arg_types, int32_t num_teams, int32_t thread_limit); | |||
716 | OMPRTL__tgt_target_teams_nowait, | |||
717 | // Call to void __tgt_register_requires(int64_t flags); | |||
718 | OMPRTL__tgt_register_requires, | |||
719 | // Call to void __tgt_register_lib(__tgt_bin_desc *desc); | |||
720 | OMPRTL__tgt_register_lib, | |||
721 | // Call to void __tgt_unregister_lib(__tgt_bin_desc *desc); | |||
722 | OMPRTL__tgt_unregister_lib, | |||
723 | // Call to void __tgt_target_data_begin(int64_t device_id, int32_t arg_num, | |||
724 | // void** args_base, void **args, size_t *arg_sizes, int64_t *arg_types); | |||
725 | OMPRTL__tgt_target_data_begin, | |||
726 | // Call to void __tgt_target_data_begin_nowait(int64_t device_id, int32_t | |||
727 | // arg_num, void** args_base, void **args, size_t *arg_sizes, int64_t | |||
728 | // *arg_types); | |||
729 | OMPRTL__tgt_target_data_begin_nowait, | |||
730 | // Call to void __tgt_target_data_end(int64_t device_id, int32_t arg_num, | |||
731 | // void** args_base, void **args, size_t *arg_sizes, int64_t *arg_types); | |||
732 | OMPRTL__tgt_target_data_end, | |||
733 | // Call to void __tgt_target_data_end_nowait(int64_t device_id, int32_t | |||
734 | // arg_num, void** args_base, void **args, size_t *arg_sizes, int64_t | |||
735 | // *arg_types); | |||
736 | OMPRTL__tgt_target_data_end_nowait, | |||
737 | // Call to void __tgt_target_data_update(int64_t device_id, int32_t arg_num, | |||
738 | // void** args_base, void **args, size_t *arg_sizes, int64_t *arg_types); | |||
739 | OMPRTL__tgt_target_data_update, | |||
740 | // Call to void __tgt_target_data_update_nowait(int64_t device_id, int32_t | |||
741 | // arg_num, void** args_base, void **args, size_t *arg_sizes, int64_t | |||
742 | // *arg_types); | |||
743 | OMPRTL__tgt_target_data_update_nowait, | |||
744 | }; | |||
745 | ||||
746 | /// A basic class for pre|post-action for advanced codegen sequence for OpenMP | |||
747 | /// region. | |||
748 | class CleanupTy final : public EHScopeStack::Cleanup { | |||
749 | PrePostActionTy *Action; | |||
750 | ||||
751 | public: | |||
752 | explicit CleanupTy(PrePostActionTy *Action) : Action(Action) {} | |||
753 | void Emit(CodeGenFunction &CGF, Flags /*flags*/) override { | |||
754 | if (!CGF.HaveInsertPoint()) | |||
755 | return; | |||
756 | Action->Exit(CGF); | |||
757 | } | |||
758 | }; | |||
759 | ||||
760 | } // anonymous namespace | |||
761 | ||||
762 | void RegionCodeGenTy::operator()(CodeGenFunction &CGF) const { | |||
763 | CodeGenFunction::RunCleanupsScope Scope(CGF); | |||
764 | if (PrePostAction) { | |||
765 | CGF.EHStack.pushCleanup<CleanupTy>(NormalAndEHCleanup, PrePostAction); | |||
766 | Callback(CodeGen, CGF, *PrePostAction); | |||
767 | } else { | |||
768 | PrePostActionTy Action; | |||
769 | Callback(CodeGen, CGF, Action); | |||
770 | } | |||
771 | } | |||
772 | ||||
773 | /// Check if the combiner is a call to UDR combiner and if it is so return the | |||
774 | /// UDR decl used for reduction. | |||
775 | static const OMPDeclareReductionDecl * | |||
776 | getReductionInit(const Expr *ReductionOp) { | |||
777 | if (const auto *CE = dyn_cast<CallExpr>(ReductionOp)) | |||
778 | if (const auto *OVE = dyn_cast<OpaqueValueExpr>(CE->getCallee())) | |||
779 | if (const auto *DRE = | |||
780 | dyn_cast<DeclRefExpr>(OVE->getSourceExpr()->IgnoreImpCasts())) | |||
781 | if (const auto *DRD = dyn_cast<OMPDeclareReductionDecl>(DRE->getDecl())) | |||
782 | return DRD; | |||
783 | return nullptr; | |||
784 | } | |||
785 | ||||
786 | static void emitInitWithReductionInitializer(CodeGenFunction &CGF, | |||
787 | const OMPDeclareReductionDecl *DRD, | |||
788 | const Expr *InitOp, | |||
789 | Address Private, Address Original, | |||
790 | QualType Ty) { | |||
791 | if (DRD->getInitializer()) { | |||
792 | std::pair<llvm::Function *, llvm::Function *> Reduction = | |||
793 | CGF.CGM.getOpenMPRuntime().getUserDefinedReduction(DRD); | |||
794 | const auto *CE = cast<CallExpr>(InitOp); | |||
795 | const auto *OVE = cast<OpaqueValueExpr>(CE->getCallee()); | |||
796 | const Expr *LHS = CE->getArg(/*Arg=*/0)->IgnoreParenImpCasts(); | |||
797 | const Expr *RHS = CE->getArg(/*Arg=*/1)->IgnoreParenImpCasts(); | |||
798 | const auto *LHSDRE = | |||
799 | cast<DeclRefExpr>(cast<UnaryOperator>(LHS)->getSubExpr()); | |||
800 | const auto *RHSDRE = | |||
801 | cast<DeclRefExpr>(cast<UnaryOperator>(RHS)->getSubExpr()); | |||
802 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); | |||
803 | PrivateScope.addPrivate(cast<VarDecl>(LHSDRE->getDecl()), | |||
804 | [=]() { return Private; }); | |||
805 | PrivateScope.addPrivate(cast<VarDecl>(RHSDRE->getDecl()), | |||
806 | [=]() { return Original; }); | |||
807 | (void)PrivateScope.Privatize(); | |||
808 | RValue Func = RValue::get(Reduction.second); | |||
809 | CodeGenFunction::OpaqueValueMapping Map(CGF, OVE, Func); | |||
810 | CGF.EmitIgnoredExpr(InitOp); | |||
811 | } else { | |||
812 | llvm::Constant *Init = CGF.CGM.EmitNullConstant(Ty); | |||
813 | std::string Name = CGF.CGM.getOpenMPRuntime().getName({"init"}); | |||
814 | auto *GV = new llvm::GlobalVariable( | |||
815 | CGF.CGM.getModule(), Init->getType(), /*isConstant=*/true, | |||
816 | llvm::GlobalValue::PrivateLinkage, Init, Name); | |||
817 | LValue LV = CGF.MakeNaturalAlignAddrLValue(GV, Ty); | |||
818 | RValue InitRVal; | |||
819 | switch (CGF.getEvaluationKind(Ty)) { | |||
820 | case TEK_Scalar: | |||
821 | InitRVal = CGF.EmitLoadOfLValue(LV, DRD->getLocation()); | |||
822 | break; | |||
823 | case TEK_Complex: | |||
824 | InitRVal = | |||
825 | RValue::getComplex(CGF.EmitLoadOfComplex(LV, DRD->getLocation())); | |||
826 | break; | |||
827 | case TEK_Aggregate: | |||
828 | InitRVal = RValue::getAggregate(LV.getAddress()); | |||
829 | break; | |||
830 | } | |||
831 | OpaqueValueExpr OVE(DRD->getLocation(), Ty, VK_RValue); | |||
832 | CodeGenFunction::OpaqueValueMapping OpaqueMap(CGF, &OVE, InitRVal); | |||
833 | CGF.EmitAnyExprToMem(&OVE, Private, Ty.getQualifiers(), | |||
834 | /*IsInitializer=*/false); | |||
835 | } | |||
836 | } | |||
837 | ||||
838 | /// Emit initialization of arrays of complex types. | |||
839 | /// \param DestAddr Address of the array. | |||
840 | /// \param Type Type of array. | |||
841 | /// \param Init Initial expression of array. | |||
842 | /// \param SrcAddr Address of the original array. | |||
843 | static void EmitOMPAggregateInit(CodeGenFunction &CGF, Address DestAddr, | |||
844 | QualType Type, bool EmitDeclareReductionInit, | |||
845 | const Expr *Init, | |||
846 | const OMPDeclareReductionDecl *DRD, | |||
847 | Address SrcAddr = Address::invalid()) { | |||
848 | // Perform element-by-element initialization. | |||
849 | QualType ElementTy; | |||
850 | ||||
851 | // Drill down to the base element type on both arrays. | |||
852 | const ArrayType *ArrayTy = Type->getAsArrayTypeUnsafe(); | |||
853 | llvm::Value *NumElements = CGF.emitArrayLength(ArrayTy, ElementTy, DestAddr); | |||
854 | DestAddr = | |||
855 | CGF.Builder.CreateElementBitCast(DestAddr, DestAddr.getElementType()); | |||
856 | if (DRD) | |||
857 | SrcAddr = | |||
858 | CGF.Builder.CreateElementBitCast(SrcAddr, DestAddr.getElementType()); | |||
859 | ||||
860 | llvm::Value *SrcBegin = nullptr; | |||
861 | if (DRD) | |||
862 | SrcBegin = SrcAddr.getPointer(); | |||
863 | llvm::Value *DestBegin = DestAddr.getPointer(); | |||
864 | // Cast from pointer to array type to pointer to single element. | |||
865 | llvm::Value *DestEnd = CGF.Builder.CreateGEP(DestBegin, NumElements); | |||
866 | // The basic structure here is a while-do loop. | |||
867 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.arrayinit.body"); | |||
868 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock("omp.arrayinit.done"); | |||
869 | llvm::Value *IsEmpty = | |||
870 | CGF.Builder.CreateICmpEQ(DestBegin, DestEnd, "omp.arrayinit.isempty"); | |||
871 | CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB); | |||
872 | ||||
873 | // Enter the loop body, making that address the current address. | |||
874 | llvm::BasicBlock *EntryBB = CGF.Builder.GetInsertBlock(); | |||
875 | CGF.EmitBlock(BodyBB); | |||
876 | ||||
877 | CharUnits ElementSize = CGF.getContext().getTypeSizeInChars(ElementTy); | |||
878 | ||||
879 | llvm::PHINode *SrcElementPHI = nullptr; | |||
880 | Address SrcElementCurrent = Address::invalid(); | |||
881 | if (DRD) { | |||
882 | SrcElementPHI = CGF.Builder.CreatePHI(SrcBegin->getType(), 2, | |||
883 | "omp.arraycpy.srcElementPast"); | |||
884 | SrcElementPHI->addIncoming(SrcBegin, EntryBB); | |||
885 | SrcElementCurrent = | |||
886 | Address(SrcElementPHI, | |||
887 | SrcAddr.getAlignment().alignmentOfArrayElement(ElementSize)); | |||
888 | } | |||
889 | llvm::PHINode *DestElementPHI = CGF.Builder.CreatePHI( | |||
890 | DestBegin->getType(), 2, "omp.arraycpy.destElementPast"); | |||
891 | DestElementPHI->addIncoming(DestBegin, EntryBB); | |||
892 | Address DestElementCurrent = | |||
893 | Address(DestElementPHI, | |||
894 | DestAddr.getAlignment().alignmentOfArrayElement(ElementSize)); | |||
895 | ||||
896 | // Emit copy. | |||
897 | { | |||
898 | CodeGenFunction::RunCleanupsScope InitScope(CGF); | |||
899 | if (EmitDeclareReductionInit) { | |||
900 | emitInitWithReductionInitializer(CGF, DRD, Init, DestElementCurrent, | |||
901 | SrcElementCurrent, ElementTy); | |||
902 | } else | |||
903 | CGF.EmitAnyExprToMem(Init, DestElementCurrent, ElementTy.getQualifiers(), | |||
904 | /*IsInitializer=*/false); | |||
905 | } | |||
906 | ||||
907 | if (DRD) { | |||
908 | // Shift the address forward by one element. | |||
909 | llvm::Value *SrcElementNext = CGF.Builder.CreateConstGEP1_32( | |||
910 | SrcElementPHI, /*Idx0=*/1, "omp.arraycpy.dest.element"); | |||
911 | SrcElementPHI->addIncoming(SrcElementNext, CGF.Builder.GetInsertBlock()); | |||
912 | } | |||
913 | ||||
914 | // Shift the address forward by one element. | |||
915 | llvm::Value *DestElementNext = CGF.Builder.CreateConstGEP1_32( | |||
916 | DestElementPHI, /*Idx0=*/1, "omp.arraycpy.dest.element"); | |||
917 | // Check whether we've reached the end. | |||
918 | llvm::Value *Done = | |||
919 | CGF.Builder.CreateICmpEQ(DestElementNext, DestEnd, "omp.arraycpy.done"); | |||
920 | CGF.Builder.CreateCondBr(Done, DoneBB, BodyBB); | |||
921 | DestElementPHI->addIncoming(DestElementNext, CGF.Builder.GetInsertBlock()); | |||
922 | ||||
923 | // Done. | |||
924 | CGF.EmitBlock(DoneBB, /*IsFinished=*/true); | |||
925 | } | |||
926 | ||||
927 | LValue ReductionCodeGen::emitSharedLValue(CodeGenFunction &CGF, const Expr *E) { | |||
928 | return CGF.EmitOMPSharedLValue(E); | |||
929 | } | |||
930 | ||||
931 | LValue ReductionCodeGen::emitSharedLValueUB(CodeGenFunction &CGF, | |||
932 | const Expr *E) { | |||
933 | if (const auto *OASE = dyn_cast<OMPArraySectionExpr>(E)) | |||
934 | return CGF.EmitOMPArraySectionExpr(OASE, /*IsLowerBound=*/false); | |||
935 | return LValue(); | |||
936 | } | |||
937 | ||||
938 | void ReductionCodeGen::emitAggregateInitialization( | |||
939 | CodeGenFunction &CGF, unsigned N, Address PrivateAddr, LValue SharedLVal, | |||
940 | const OMPDeclareReductionDecl *DRD) { | |||
941 | // Emit VarDecl with copy init for arrays. | |||
942 | // Get the address of the original variable captured in current | |||
943 | // captured region. | |||
944 | const auto *PrivateVD = | |||
945 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | |||
946 | bool EmitDeclareReductionInit = | |||
947 | DRD && (DRD->getInitializer() || !PrivateVD->hasInit()); | |||
948 | EmitOMPAggregateInit(CGF, PrivateAddr, PrivateVD->getType(), | |||
949 | EmitDeclareReductionInit, | |||
950 | EmitDeclareReductionInit ? ClausesData[N].ReductionOp | |||
951 | : PrivateVD->getInit(), | |||
952 | DRD, SharedLVal.getAddress()); | |||
953 | } | |||
954 | ||||
955 | ReductionCodeGen::ReductionCodeGen(ArrayRef<const Expr *> Shareds, | |||
956 | ArrayRef<const Expr *> Privates, | |||
957 | ArrayRef<const Expr *> ReductionOps) { | |||
958 | ClausesData.reserve(Shareds.size()); | |||
959 | SharedAddresses.reserve(Shareds.size()); | |||
960 | Sizes.reserve(Shareds.size()); | |||
961 | BaseDecls.reserve(Shareds.size()); | |||
962 | auto IPriv = Privates.begin(); | |||
963 | auto IRed = ReductionOps.begin(); | |||
964 | for (const Expr *Ref : Shareds) { | |||
965 | ClausesData.emplace_back(Ref, *IPriv, *IRed); | |||
966 | std::advance(IPriv, 1); | |||
967 | std::advance(IRed, 1); | |||
968 | } | |||
969 | } | |||
970 | ||||
971 | void ReductionCodeGen::emitSharedLValue(CodeGenFunction &CGF, unsigned N) { | |||
972 | assert(SharedAddresses.size() == N &&((SharedAddresses.size() == N && "Number of generated lvalues must be exactly N." ) ? static_cast<void> (0) : __assert_fail ("SharedAddresses.size() == N && \"Number of generated lvalues must be exactly N.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 973, __PRETTY_FUNCTION__)) | |||
973 | "Number of generated lvalues must be exactly N.")((SharedAddresses.size() == N && "Number of generated lvalues must be exactly N." ) ? static_cast<void> (0) : __assert_fail ("SharedAddresses.size() == N && \"Number of generated lvalues must be exactly N.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 973, __PRETTY_FUNCTION__)); | |||
974 | LValue First = emitSharedLValue(CGF, ClausesData[N].Ref); | |||
975 | LValue Second = emitSharedLValueUB(CGF, ClausesData[N].Ref); | |||
976 | SharedAddresses.emplace_back(First, Second); | |||
977 | } | |||
978 | ||||
979 | void ReductionCodeGen::emitAggregateType(CodeGenFunction &CGF, unsigned N) { | |||
980 | const auto *PrivateVD = | |||
981 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | |||
982 | QualType PrivateType = PrivateVD->getType(); | |||
983 | bool AsArraySection = isa<OMPArraySectionExpr>(ClausesData[N].Ref); | |||
984 | if (!PrivateType->isVariablyModifiedType()) { | |||
985 | Sizes.emplace_back( | |||
986 | CGF.getTypeSize( | |||
987 | SharedAddresses[N].first.getType().getNonReferenceType()), | |||
988 | nullptr); | |||
989 | return; | |||
990 | } | |||
991 | llvm::Value *Size; | |||
992 | llvm::Value *SizeInChars; | |||
993 | auto *ElemType = | |||
994 | cast<llvm::PointerType>(SharedAddresses[N].first.getPointer()->getType()) | |||
995 | ->getElementType(); | |||
996 | auto *ElemSizeOf = llvm::ConstantExpr::getSizeOf(ElemType); | |||
997 | if (AsArraySection) { | |||
998 | Size = CGF.Builder.CreatePtrDiff(SharedAddresses[N].second.getPointer(), | |||
999 | SharedAddresses[N].first.getPointer()); | |||
1000 | Size = CGF.Builder.CreateNUWAdd( | |||
1001 | Size, llvm::ConstantInt::get(Size->getType(), /*V=*/1)); | |||
1002 | SizeInChars = CGF.Builder.CreateNUWMul(Size, ElemSizeOf); | |||
1003 | } else { | |||
1004 | SizeInChars = CGF.getTypeSize( | |||
1005 | SharedAddresses[N].first.getType().getNonReferenceType()); | |||
1006 | Size = CGF.Builder.CreateExactUDiv(SizeInChars, ElemSizeOf); | |||
1007 | } | |||
1008 | Sizes.emplace_back(SizeInChars, Size); | |||
1009 | CodeGenFunction::OpaqueValueMapping OpaqueMap( | |||
1010 | CGF, | |||
1011 | cast<OpaqueValueExpr>( | |||
1012 | CGF.getContext().getAsVariableArrayType(PrivateType)->getSizeExpr()), | |||
1013 | RValue::get(Size)); | |||
1014 | CGF.EmitVariablyModifiedType(PrivateType); | |||
1015 | } | |||
1016 | ||||
1017 | void ReductionCodeGen::emitAggregateType(CodeGenFunction &CGF, unsigned N, | |||
1018 | llvm::Value *Size) { | |||
1019 | const auto *PrivateVD = | |||
1020 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | |||
1021 | QualType PrivateType = PrivateVD->getType(); | |||
1022 | if (!PrivateType->isVariablyModifiedType()) { | |||
1023 | assert(!Size && !Sizes[N].second &&((!Size && !Sizes[N].second && "Size should be nullptr for non-variably modified reduction " "items.") ? static_cast<void> (0) : __assert_fail ("!Size && !Sizes[N].second && \"Size should be nullptr for non-variably modified reduction \" \"items.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1025, __PRETTY_FUNCTION__)) | |||
1024 | "Size should be nullptr for non-variably modified reduction "((!Size && !Sizes[N].second && "Size should be nullptr for non-variably modified reduction " "items.") ? static_cast<void> (0) : __assert_fail ("!Size && !Sizes[N].second && \"Size should be nullptr for non-variably modified reduction \" \"items.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1025, __PRETTY_FUNCTION__)) | |||
1025 | "items.")((!Size && !Sizes[N].second && "Size should be nullptr for non-variably modified reduction " "items.") ? static_cast<void> (0) : __assert_fail ("!Size && !Sizes[N].second && \"Size should be nullptr for non-variably modified reduction \" \"items.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1025, __PRETTY_FUNCTION__)); | |||
1026 | return; | |||
1027 | } | |||
1028 | CodeGenFunction::OpaqueValueMapping OpaqueMap( | |||
1029 | CGF, | |||
1030 | cast<OpaqueValueExpr>( | |||
1031 | CGF.getContext().getAsVariableArrayType(PrivateType)->getSizeExpr()), | |||
1032 | RValue::get(Size)); | |||
1033 | CGF.EmitVariablyModifiedType(PrivateType); | |||
1034 | } | |||
1035 | ||||
1036 | void ReductionCodeGen::emitInitialization( | |||
1037 | CodeGenFunction &CGF, unsigned N, Address PrivateAddr, LValue SharedLVal, | |||
1038 | llvm::function_ref<bool(CodeGenFunction &)> DefaultInit) { | |||
1039 | assert(SharedAddresses.size() > N && "No variable was generated")((SharedAddresses.size() > N && "No variable was generated" ) ? static_cast<void> (0) : __assert_fail ("SharedAddresses.size() > N && \"No variable was generated\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1039, __PRETTY_FUNCTION__)); | |||
1040 | const auto *PrivateVD = | |||
1041 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | |||
1042 | const OMPDeclareReductionDecl *DRD = | |||
1043 | getReductionInit(ClausesData[N].ReductionOp); | |||
1044 | QualType PrivateType = PrivateVD->getType(); | |||
1045 | PrivateAddr = CGF.Builder.CreateElementBitCast( | |||
1046 | PrivateAddr, CGF.ConvertTypeForMem(PrivateType)); | |||
1047 | QualType SharedType = SharedAddresses[N].first.getType(); | |||
1048 | SharedLVal = CGF.MakeAddrLValue( | |||
1049 | CGF.Builder.CreateElementBitCast(SharedLVal.getAddress(), | |||
1050 | CGF.ConvertTypeForMem(SharedType)), | |||
1051 | SharedType, SharedAddresses[N].first.getBaseInfo(), | |||
1052 | CGF.CGM.getTBAAInfoForSubobject(SharedAddresses[N].first, SharedType)); | |||
1053 | if (CGF.getContext().getAsArrayType(PrivateVD->getType())) { | |||
1054 | emitAggregateInitialization(CGF, N, PrivateAddr, SharedLVal, DRD); | |||
1055 | } else if (DRD && (DRD->getInitializer() || !PrivateVD->hasInit())) { | |||
1056 | emitInitWithReductionInitializer(CGF, DRD, ClausesData[N].ReductionOp, | |||
1057 | PrivateAddr, SharedLVal.getAddress(), | |||
1058 | SharedLVal.getType()); | |||
1059 | } else if (!DefaultInit(CGF) && PrivateVD->hasInit() && | |||
1060 | !CGF.isTrivialInitializer(PrivateVD->getInit())) { | |||
1061 | CGF.EmitAnyExprToMem(PrivateVD->getInit(), PrivateAddr, | |||
1062 | PrivateVD->getType().getQualifiers(), | |||
1063 | /*IsInitializer=*/false); | |||
1064 | } | |||
1065 | } | |||
1066 | ||||
1067 | bool ReductionCodeGen::needCleanups(unsigned N) { | |||
1068 | const auto *PrivateVD = | |||
1069 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | |||
1070 | QualType PrivateType = PrivateVD->getType(); | |||
1071 | QualType::DestructionKind DTorKind = PrivateType.isDestructedType(); | |||
1072 | return DTorKind != QualType::DK_none; | |||
1073 | } | |||
1074 | ||||
1075 | void ReductionCodeGen::emitCleanups(CodeGenFunction &CGF, unsigned N, | |||
1076 | Address PrivateAddr) { | |||
1077 | const auto *PrivateVD = | |||
1078 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Private)->getDecl()); | |||
1079 | QualType PrivateType = PrivateVD->getType(); | |||
1080 | QualType::DestructionKind DTorKind = PrivateType.isDestructedType(); | |||
1081 | if (needCleanups(N)) { | |||
1082 | PrivateAddr = CGF.Builder.CreateElementBitCast( | |||
1083 | PrivateAddr, CGF.ConvertTypeForMem(PrivateType)); | |||
1084 | CGF.pushDestroy(DTorKind, PrivateAddr, PrivateType); | |||
1085 | } | |||
1086 | } | |||
1087 | ||||
1088 | static LValue loadToBegin(CodeGenFunction &CGF, QualType BaseTy, QualType ElTy, | |||
1089 | LValue BaseLV) { | |||
1090 | BaseTy = BaseTy.getNonReferenceType(); | |||
1091 | while ((BaseTy->isPointerType() || BaseTy->isReferenceType()) && | |||
1092 | !CGF.getContext().hasSameType(BaseTy, ElTy)) { | |||
1093 | if (const auto *PtrTy = BaseTy->getAs<PointerType>()) { | |||
1094 | BaseLV = CGF.EmitLoadOfPointerLValue(BaseLV.getAddress(), PtrTy); | |||
1095 | } else { | |||
1096 | LValue RefLVal = CGF.MakeAddrLValue(BaseLV.getAddress(), BaseTy); | |||
1097 | BaseLV = CGF.EmitLoadOfReferenceLValue(RefLVal); | |||
1098 | } | |||
1099 | BaseTy = BaseTy->getPointeeType(); | |||
1100 | } | |||
1101 | return CGF.MakeAddrLValue( | |||
1102 | CGF.Builder.CreateElementBitCast(BaseLV.getAddress(), | |||
1103 | CGF.ConvertTypeForMem(ElTy)), | |||
1104 | BaseLV.getType(), BaseLV.getBaseInfo(), | |||
1105 | CGF.CGM.getTBAAInfoForSubobject(BaseLV, BaseLV.getType())); | |||
1106 | } | |||
1107 | ||||
1108 | static Address castToBase(CodeGenFunction &CGF, QualType BaseTy, QualType ElTy, | |||
1109 | llvm::Type *BaseLVType, CharUnits BaseLVAlignment, | |||
1110 | llvm::Value *Addr) { | |||
1111 | Address Tmp = Address::invalid(); | |||
1112 | Address TopTmp = Address::invalid(); | |||
1113 | Address MostTopTmp = Address::invalid(); | |||
1114 | BaseTy = BaseTy.getNonReferenceType(); | |||
1115 | while ((BaseTy->isPointerType() || BaseTy->isReferenceType()) && | |||
1116 | !CGF.getContext().hasSameType(BaseTy, ElTy)) { | |||
1117 | Tmp = CGF.CreateMemTemp(BaseTy); | |||
1118 | if (TopTmp.isValid()) | |||
1119 | CGF.Builder.CreateStore(Tmp.getPointer(), TopTmp); | |||
1120 | else | |||
1121 | MostTopTmp = Tmp; | |||
1122 | TopTmp = Tmp; | |||
1123 | BaseTy = BaseTy->getPointeeType(); | |||
1124 | } | |||
1125 | llvm::Type *Ty = BaseLVType; | |||
1126 | if (Tmp.isValid()) | |||
1127 | Ty = Tmp.getElementType(); | |||
1128 | Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(Addr, Ty); | |||
1129 | if (Tmp.isValid()) { | |||
1130 | CGF.Builder.CreateStore(Addr, Tmp); | |||
1131 | return MostTopTmp; | |||
1132 | } | |||
1133 | return Address(Addr, BaseLVAlignment); | |||
1134 | } | |||
1135 | ||||
1136 | static const VarDecl *getBaseDecl(const Expr *Ref, const DeclRefExpr *&DE) { | |||
1137 | const VarDecl *OrigVD = nullptr; | |||
1138 | if (const auto *OASE = dyn_cast<OMPArraySectionExpr>(Ref)) { | |||
1139 | const Expr *Base = OASE->getBase()->IgnoreParenImpCasts(); | |||
1140 | while (const auto *TempOASE = dyn_cast<OMPArraySectionExpr>(Base)) | |||
1141 | Base = TempOASE->getBase()->IgnoreParenImpCasts(); | |||
1142 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) | |||
1143 | Base = TempASE->getBase()->IgnoreParenImpCasts(); | |||
1144 | DE = cast<DeclRefExpr>(Base); | |||
1145 | OrigVD = cast<VarDecl>(DE->getDecl()); | |||
1146 | } else if (const auto *ASE = dyn_cast<ArraySubscriptExpr>(Ref)) { | |||
1147 | const Expr *Base = ASE->getBase()->IgnoreParenImpCasts(); | |||
1148 | while (const auto *TempASE = dyn_cast<ArraySubscriptExpr>(Base)) | |||
1149 | Base = TempASE->getBase()->IgnoreParenImpCasts(); | |||
1150 | DE = cast<DeclRefExpr>(Base); | |||
1151 | OrigVD = cast<VarDecl>(DE->getDecl()); | |||
1152 | } | |||
1153 | return OrigVD; | |||
1154 | } | |||
1155 | ||||
1156 | Address ReductionCodeGen::adjustPrivateAddress(CodeGenFunction &CGF, unsigned N, | |||
1157 | Address PrivateAddr) { | |||
1158 | const DeclRefExpr *DE; | |||
1159 | if (const VarDecl *OrigVD = ::getBaseDecl(ClausesData[N].Ref, DE)) { | |||
1160 | BaseDecls.emplace_back(OrigVD); | |||
1161 | LValue OriginalBaseLValue = CGF.EmitLValue(DE); | |||
1162 | LValue BaseLValue = | |||
1163 | loadToBegin(CGF, OrigVD->getType(), SharedAddresses[N].first.getType(), | |||
1164 | OriginalBaseLValue); | |||
1165 | llvm::Value *Adjustment = CGF.Builder.CreatePtrDiff( | |||
1166 | BaseLValue.getPointer(), SharedAddresses[N].first.getPointer()); | |||
1167 | llvm::Value *PrivatePointer = | |||
1168 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
1169 | PrivateAddr.getPointer(), | |||
1170 | SharedAddresses[N].first.getAddress().getType()); | |||
1171 | llvm::Value *Ptr = CGF.Builder.CreateGEP(PrivatePointer, Adjustment); | |||
1172 | return castToBase(CGF, OrigVD->getType(), | |||
1173 | SharedAddresses[N].first.getType(), | |||
1174 | OriginalBaseLValue.getAddress().getType(), | |||
1175 | OriginalBaseLValue.getAlignment(), Ptr); | |||
1176 | } | |||
1177 | BaseDecls.emplace_back( | |||
1178 | cast<VarDecl>(cast<DeclRefExpr>(ClausesData[N].Ref)->getDecl())); | |||
1179 | return PrivateAddr; | |||
1180 | } | |||
1181 | ||||
1182 | bool ReductionCodeGen::usesReductionInitializer(unsigned N) const { | |||
1183 | const OMPDeclareReductionDecl *DRD = | |||
1184 | getReductionInit(ClausesData[N].ReductionOp); | |||
1185 | return DRD && DRD->getInitializer(); | |||
1186 | } | |||
1187 | ||||
1188 | LValue CGOpenMPRegionInfo::getThreadIDVariableLValue(CodeGenFunction &CGF) { | |||
1189 | return CGF.EmitLoadOfPointerLValue( | |||
1190 | CGF.GetAddrOfLocalVar(getThreadIDVariable()), | |||
1191 | getThreadIDVariable()->getType()->castAs<PointerType>()); | |||
1192 | } | |||
1193 | ||||
1194 | void CGOpenMPRegionInfo::EmitBody(CodeGenFunction &CGF, const Stmt * /*S*/) { | |||
1195 | if (!CGF.HaveInsertPoint()) | |||
1196 | return; | |||
1197 | // 1.2.2 OpenMP Language Terminology | |||
1198 | // Structured block - An executable statement with a single entry at the | |||
1199 | // top and a single exit at the bottom. | |||
1200 | // The point of exit cannot be a branch out of the structured block. | |||
1201 | // longjmp() and throw() must not violate the entry/exit criteria. | |||
1202 | CGF.EHStack.pushTerminate(); | |||
1203 | CodeGen(CGF); | |||
1204 | CGF.EHStack.popTerminate(); | |||
1205 | } | |||
1206 | ||||
1207 | LValue CGOpenMPTaskOutlinedRegionInfo::getThreadIDVariableLValue( | |||
1208 | CodeGenFunction &CGF) { | |||
1209 | return CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(getThreadIDVariable()), | |||
1210 | getThreadIDVariable()->getType(), | |||
1211 | AlignmentSource::Decl); | |||
1212 | } | |||
1213 | ||||
1214 | static FieldDecl *addFieldToRecordDecl(ASTContext &C, DeclContext *DC, | |||
1215 | QualType FieldTy) { | |||
1216 | auto *Field = FieldDecl::Create( | |||
1217 | C, DC, SourceLocation(), SourceLocation(), /*Id=*/nullptr, FieldTy, | |||
1218 | C.getTrivialTypeSourceInfo(FieldTy, SourceLocation()), | |||
1219 | /*BW=*/nullptr, /*Mutable=*/false, /*InitStyle=*/ICIS_NoInit); | |||
1220 | Field->setAccess(AS_public); | |||
1221 | DC->addDecl(Field); | |||
1222 | return Field; | |||
1223 | } | |||
1224 | ||||
1225 | CGOpenMPRuntime::CGOpenMPRuntime(CodeGenModule &CGM, StringRef FirstSeparator, | |||
1226 | StringRef Separator) | |||
1227 | : CGM(CGM), FirstSeparator(FirstSeparator), Separator(Separator), | |||
1228 | OffloadEntriesInfoManager(CGM) { | |||
1229 | ASTContext &C = CGM.getContext(); | |||
1230 | RecordDecl *RD = C.buildImplicitRecord("ident_t"); | |||
1231 | QualType KmpInt32Ty = C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1); | |||
1232 | RD->startDefinition(); | |||
1233 | // reserved_1 | |||
1234 | addFieldToRecordDecl(C, RD, KmpInt32Ty); | |||
1235 | // flags | |||
1236 | addFieldToRecordDecl(C, RD, KmpInt32Ty); | |||
1237 | // reserved_2 | |||
1238 | addFieldToRecordDecl(C, RD, KmpInt32Ty); | |||
1239 | // reserved_3 | |||
1240 | addFieldToRecordDecl(C, RD, KmpInt32Ty); | |||
1241 | // psource | |||
1242 | addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
1243 | RD->completeDefinition(); | |||
1244 | IdentQTy = C.getRecordType(RD); | |||
1245 | IdentTy = CGM.getTypes().ConvertRecordDeclType(RD); | |||
1246 | KmpCriticalNameTy = llvm::ArrayType::get(CGM.Int32Ty, /*NumElements*/ 8); | |||
1247 | ||||
1248 | loadOffloadInfoMetadata(); | |||
1249 | } | |||
1250 | ||||
1251 | void CGOpenMPRuntime::clear() { | |||
1252 | InternalVars.clear(); | |||
1253 | // Clean non-target variable declarations possibly used only in debug info. | |||
1254 | for (const auto &Data : EmittedNonTargetVariables) { | |||
1255 | if (!Data.getValue().pointsToAliveValue()) | |||
1256 | continue; | |||
1257 | auto *GV = dyn_cast<llvm::GlobalVariable>(Data.getValue()); | |||
1258 | if (!GV) | |||
1259 | continue; | |||
1260 | if (!GV->isDeclaration() || GV->getNumUses() > 0) | |||
1261 | continue; | |||
1262 | GV->eraseFromParent(); | |||
1263 | } | |||
1264 | } | |||
1265 | ||||
1266 | std::string CGOpenMPRuntime::getName(ArrayRef<StringRef> Parts) const { | |||
1267 | SmallString<128> Buffer; | |||
1268 | llvm::raw_svector_ostream OS(Buffer); | |||
1269 | StringRef Sep = FirstSeparator; | |||
1270 | for (StringRef Part : Parts) { | |||
1271 | OS << Sep << Part; | |||
1272 | Sep = Separator; | |||
1273 | } | |||
1274 | return OS.str(); | |||
1275 | } | |||
1276 | ||||
1277 | static llvm::Function * | |||
1278 | emitCombinerOrInitializer(CodeGenModule &CGM, QualType Ty, | |||
1279 | const Expr *CombinerInitializer, const VarDecl *In, | |||
1280 | const VarDecl *Out, bool IsCombiner) { | |||
1281 | // void .omp_combiner.(Ty *in, Ty *out); | |||
1282 | ASTContext &C = CGM.getContext(); | |||
1283 | QualType PtrTy = C.getPointerType(Ty).withRestrict(); | |||
1284 | FunctionArgList Args; | |||
1285 | ImplicitParamDecl OmpOutParm(C, /*DC=*/nullptr, Out->getLocation(), | |||
1286 | /*Id=*/nullptr, PtrTy, ImplicitParamDecl::Other); | |||
1287 | ImplicitParamDecl OmpInParm(C, /*DC=*/nullptr, In->getLocation(), | |||
1288 | /*Id=*/nullptr, PtrTy, ImplicitParamDecl::Other); | |||
1289 | Args.push_back(&OmpOutParm); | |||
1290 | Args.push_back(&OmpInParm); | |||
1291 | const CGFunctionInfo &FnInfo = | |||
1292 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
1293 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | |||
1294 | std::string Name = CGM.getOpenMPRuntime().getName( | |||
1295 | {IsCombiner ? "omp_combiner" : "omp_initializer", ""}); | |||
1296 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | |||
1297 | Name, &CGM.getModule()); | |||
1298 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | |||
1299 | if (CGM.getLangOpts().Optimize) { | |||
1300 | Fn->removeFnAttr(llvm::Attribute::NoInline); | |||
1301 | Fn->removeFnAttr(llvm::Attribute::OptimizeNone); | |||
1302 | Fn->addFnAttr(llvm::Attribute::AlwaysInline); | |||
1303 | } | |||
1304 | CodeGenFunction CGF(CGM); | |||
1305 | // Map "T omp_in;" variable to "*omp_in_parm" value in all expressions. | |||
1306 | // Map "T omp_out;" variable to "*omp_out_parm" value in all expressions. | |||
1307 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, In->getLocation(), | |||
1308 | Out->getLocation()); | |||
1309 | CodeGenFunction::OMPPrivateScope Scope(CGF); | |||
1310 | Address AddrIn = CGF.GetAddrOfLocalVar(&OmpInParm); | |||
1311 | Scope.addPrivate(In, [&CGF, AddrIn, PtrTy]() { | |||
1312 | return CGF.EmitLoadOfPointerLValue(AddrIn, PtrTy->castAs<PointerType>()) | |||
1313 | .getAddress(); | |||
1314 | }); | |||
1315 | Address AddrOut = CGF.GetAddrOfLocalVar(&OmpOutParm); | |||
1316 | Scope.addPrivate(Out, [&CGF, AddrOut, PtrTy]() { | |||
1317 | return CGF.EmitLoadOfPointerLValue(AddrOut, PtrTy->castAs<PointerType>()) | |||
1318 | .getAddress(); | |||
1319 | }); | |||
1320 | (void)Scope.Privatize(); | |||
1321 | if (!IsCombiner && Out->hasInit() && | |||
1322 | !CGF.isTrivialInitializer(Out->getInit())) { | |||
1323 | CGF.EmitAnyExprToMem(Out->getInit(), CGF.GetAddrOfLocalVar(Out), | |||
1324 | Out->getType().getQualifiers(), | |||
1325 | /*IsInitializer=*/true); | |||
1326 | } | |||
1327 | if (CombinerInitializer) | |||
1328 | CGF.EmitIgnoredExpr(CombinerInitializer); | |||
1329 | Scope.ForceCleanup(); | |||
1330 | CGF.FinishFunction(); | |||
1331 | return Fn; | |||
1332 | } | |||
1333 | ||||
1334 | void CGOpenMPRuntime::emitUserDefinedReduction( | |||
1335 | CodeGenFunction *CGF, const OMPDeclareReductionDecl *D) { | |||
1336 | if (UDRMap.count(D) > 0) | |||
1337 | return; | |||
1338 | llvm::Function *Combiner = emitCombinerOrInitializer( | |||
1339 | CGM, D->getType(), D->getCombiner(), | |||
1340 | cast<VarDecl>(cast<DeclRefExpr>(D->getCombinerIn())->getDecl()), | |||
1341 | cast<VarDecl>(cast<DeclRefExpr>(D->getCombinerOut())->getDecl()), | |||
1342 | /*IsCombiner=*/true); | |||
1343 | llvm::Function *Initializer = nullptr; | |||
1344 | if (const Expr *Init = D->getInitializer()) { | |||
1345 | Initializer = emitCombinerOrInitializer( | |||
1346 | CGM, D->getType(), | |||
1347 | D->getInitializerKind() == OMPDeclareReductionDecl::CallInit ? Init | |||
1348 | : nullptr, | |||
1349 | cast<VarDecl>(cast<DeclRefExpr>(D->getInitOrig())->getDecl()), | |||
1350 | cast<VarDecl>(cast<DeclRefExpr>(D->getInitPriv())->getDecl()), | |||
1351 | /*IsCombiner=*/false); | |||
1352 | } | |||
1353 | UDRMap.try_emplace(D, Combiner, Initializer); | |||
1354 | if (CGF) { | |||
1355 | auto &Decls = FunctionUDRMap.FindAndConstruct(CGF->CurFn); | |||
1356 | Decls.second.push_back(D); | |||
1357 | } | |||
1358 | } | |||
1359 | ||||
1360 | std::pair<llvm::Function *, llvm::Function *> | |||
1361 | CGOpenMPRuntime::getUserDefinedReduction(const OMPDeclareReductionDecl *D) { | |||
1362 | auto I = UDRMap.find(D); | |||
1363 | if (I != UDRMap.end()) | |||
1364 | return I->second; | |||
1365 | emitUserDefinedReduction(/*CGF=*/nullptr, D); | |||
1366 | return UDRMap.lookup(D); | |||
1367 | } | |||
1368 | ||||
1369 | static llvm::Function *emitParallelOrTeamsOutlinedFunction( | |||
1370 | CodeGenModule &CGM, const OMPExecutableDirective &D, const CapturedStmt *CS, | |||
1371 | const VarDecl *ThreadIDVar, OpenMPDirectiveKind InnermostKind, | |||
1372 | const StringRef OutlinedHelperName, const RegionCodeGenTy &CodeGen) { | |||
1373 | assert(ThreadIDVar->getType()->isPointerType() &&((ThreadIDVar->getType()->isPointerType() && "thread id variable must be of type kmp_int32 *" ) ? static_cast<void> (0) : __assert_fail ("ThreadIDVar->getType()->isPointerType() && \"thread id variable must be of type kmp_int32 *\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1374, __PRETTY_FUNCTION__)) | |||
1374 | "thread id variable must be of type kmp_int32 *")((ThreadIDVar->getType()->isPointerType() && "thread id variable must be of type kmp_int32 *" ) ? static_cast<void> (0) : __assert_fail ("ThreadIDVar->getType()->isPointerType() && \"thread id variable must be of type kmp_int32 *\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1374, __PRETTY_FUNCTION__)); | |||
1375 | CodeGenFunction CGF(CGM, true); | |||
1376 | bool HasCancel = false; | |||
1377 | if (const auto *OPD = dyn_cast<OMPParallelDirective>(&D)) | |||
1378 | HasCancel = OPD->hasCancel(); | |||
1379 | else if (const auto *OPSD = dyn_cast<OMPParallelSectionsDirective>(&D)) | |||
1380 | HasCancel = OPSD->hasCancel(); | |||
1381 | else if (const auto *OPFD = dyn_cast<OMPParallelForDirective>(&D)) | |||
1382 | HasCancel = OPFD->hasCancel(); | |||
1383 | else if (const auto *OPFD = dyn_cast<OMPTargetParallelForDirective>(&D)) | |||
1384 | HasCancel = OPFD->hasCancel(); | |||
1385 | else if (const auto *OPFD = dyn_cast<OMPDistributeParallelForDirective>(&D)) | |||
1386 | HasCancel = OPFD->hasCancel(); | |||
1387 | else if (const auto *OPFD = | |||
1388 | dyn_cast<OMPTeamsDistributeParallelForDirective>(&D)) | |||
1389 | HasCancel = OPFD->hasCancel(); | |||
1390 | else if (const auto *OPFD = | |||
1391 | dyn_cast<OMPTargetTeamsDistributeParallelForDirective>(&D)) | |||
1392 | HasCancel = OPFD->hasCancel(); | |||
1393 | CGOpenMPOutlinedRegionInfo CGInfo(*CS, ThreadIDVar, CodeGen, InnermostKind, | |||
1394 | HasCancel, OutlinedHelperName); | |||
1395 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
1396 | return CGF.GenerateOpenMPCapturedStmtFunction(*CS); | |||
1397 | } | |||
1398 | ||||
1399 | llvm::Function *CGOpenMPRuntime::emitParallelOutlinedFunction( | |||
1400 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
1401 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | |||
1402 | const CapturedStmt *CS = D.getCapturedStmt(OMPD_parallel); | |||
1403 | return emitParallelOrTeamsOutlinedFunction( | |||
1404 | CGM, D, CS, ThreadIDVar, InnermostKind, getOutlinedHelperName(), CodeGen); | |||
1405 | } | |||
1406 | ||||
1407 | llvm::Function *CGOpenMPRuntime::emitTeamsOutlinedFunction( | |||
1408 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
1409 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | |||
1410 | const CapturedStmt *CS = D.getCapturedStmt(OMPD_teams); | |||
1411 | return emitParallelOrTeamsOutlinedFunction( | |||
1412 | CGM, D, CS, ThreadIDVar, InnermostKind, getOutlinedHelperName(), CodeGen); | |||
1413 | } | |||
1414 | ||||
1415 | llvm::Function *CGOpenMPRuntime::emitTaskOutlinedFunction( | |||
1416 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
1417 | const VarDecl *PartIDVar, const VarDecl *TaskTVar, | |||
1418 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen, | |||
1419 | bool Tied, unsigned &NumberOfParts) { | |||
1420 | auto &&UntiedCodeGen = [this, &D, TaskTVar](CodeGenFunction &CGF, | |||
1421 | PrePostActionTy &) { | |||
1422 | llvm::Value *ThreadID = getThreadID(CGF, D.getBeginLoc()); | |||
1423 | llvm::Value *UpLoc = emitUpdateLocation(CGF, D.getBeginLoc()); | |||
1424 | llvm::Value *TaskArgs[] = { | |||
1425 | UpLoc, ThreadID, | |||
1426 | CGF.EmitLoadOfPointerLValue(CGF.GetAddrOfLocalVar(TaskTVar), | |||
1427 | TaskTVar->getType()->castAs<PointerType>()) | |||
1428 | .getPointer()}; | |||
1429 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_omp_task), TaskArgs); | |||
1430 | }; | |||
1431 | CGOpenMPTaskOutlinedRegionInfo::UntiedTaskActionTy Action(Tied, PartIDVar, | |||
1432 | UntiedCodeGen); | |||
1433 | CodeGen.setAction(Action); | |||
1434 | assert(!ThreadIDVar->getType()->isPointerType() &&((!ThreadIDVar->getType()->isPointerType() && "thread id variable must be of type kmp_int32 for tasks" ) ? static_cast<void> (0) : __assert_fail ("!ThreadIDVar->getType()->isPointerType() && \"thread id variable must be of type kmp_int32 for tasks\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1435, __PRETTY_FUNCTION__)) | |||
1435 | "thread id variable must be of type kmp_int32 for tasks")((!ThreadIDVar->getType()->isPointerType() && "thread id variable must be of type kmp_int32 for tasks" ) ? static_cast<void> (0) : __assert_fail ("!ThreadIDVar->getType()->isPointerType() && \"thread id variable must be of type kmp_int32 for tasks\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1435, __PRETTY_FUNCTION__)); | |||
1436 | const OpenMPDirectiveKind Region = | |||
1437 | isOpenMPTaskLoopDirective(D.getDirectiveKind()) ? OMPD_taskloop | |||
1438 | : OMPD_task; | |||
1439 | const CapturedStmt *CS = D.getCapturedStmt(Region); | |||
1440 | const auto *TD = dyn_cast<OMPTaskDirective>(&D); | |||
1441 | CodeGenFunction CGF(CGM, true); | |||
1442 | CGOpenMPTaskOutlinedRegionInfo CGInfo(*CS, ThreadIDVar, CodeGen, | |||
1443 | InnermostKind, | |||
1444 | TD ? TD->hasCancel() : false, Action); | |||
1445 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
1446 | llvm::Function *Res = CGF.GenerateCapturedStmtFunction(*CS); | |||
1447 | if (!Tied) | |||
1448 | NumberOfParts = Action.getNumberOfParts(); | |||
1449 | return Res; | |||
1450 | } | |||
1451 | ||||
1452 | static void buildStructValue(ConstantStructBuilder &Fields, CodeGenModule &CGM, | |||
1453 | const RecordDecl *RD, const CGRecordLayout &RL, | |||
1454 | ArrayRef<llvm::Constant *> Data) { | |||
1455 | llvm::StructType *StructTy = RL.getLLVMType(); | |||
1456 | unsigned PrevIdx = 0; | |||
1457 | ConstantInitBuilder CIBuilder(CGM); | |||
1458 | auto DI = Data.begin(); | |||
1459 | for (const FieldDecl *FD : RD->fields()) { | |||
1460 | unsigned Idx = RL.getLLVMFieldNo(FD); | |||
1461 | // Fill the alignment. | |||
1462 | for (unsigned I = PrevIdx; I < Idx; ++I) | |||
1463 | Fields.add(llvm::Constant::getNullValue(StructTy->getElementType(I))); | |||
1464 | PrevIdx = Idx + 1; | |||
1465 | Fields.add(*DI); | |||
1466 | ++DI; | |||
1467 | } | |||
1468 | } | |||
1469 | ||||
1470 | template <class... As> | |||
1471 | static llvm::GlobalVariable * | |||
1472 | createGlobalStruct(CodeGenModule &CGM, QualType Ty, bool IsConstant, | |||
1473 | ArrayRef<llvm::Constant *> Data, const Twine &Name, | |||
1474 | As &&... Args) { | |||
1475 | const auto *RD = cast<RecordDecl>(Ty->getAsTagDecl()); | |||
1476 | const CGRecordLayout &RL = CGM.getTypes().getCGRecordLayout(RD); | |||
1477 | ConstantInitBuilder CIBuilder(CGM); | |||
1478 | ConstantStructBuilder Fields = CIBuilder.beginStruct(RL.getLLVMType()); | |||
1479 | buildStructValue(Fields, CGM, RD, RL, Data); | |||
1480 | return Fields.finishAndCreateGlobal( | |||
1481 | Name, CGM.getContext().getAlignOfGlobalVarInChars(Ty), IsConstant, | |||
1482 | std::forward<As>(Args)...); | |||
1483 | } | |||
1484 | ||||
1485 | template <typename T> | |||
1486 | static void | |||
1487 | createConstantGlobalStructAndAddToParent(CodeGenModule &CGM, QualType Ty, | |||
1488 | ArrayRef<llvm::Constant *> Data, | |||
1489 | T &Parent) { | |||
1490 | const auto *RD = cast<RecordDecl>(Ty->getAsTagDecl()); | |||
1491 | const CGRecordLayout &RL = CGM.getTypes().getCGRecordLayout(RD); | |||
1492 | ConstantStructBuilder Fields = Parent.beginStruct(RL.getLLVMType()); | |||
1493 | buildStructValue(Fields, CGM, RD, RL, Data); | |||
1494 | Fields.finishAndAddTo(Parent); | |||
1495 | } | |||
1496 | ||||
1497 | Address CGOpenMPRuntime::getOrCreateDefaultLocation(unsigned Flags) { | |||
1498 | CharUnits Align = CGM.getContext().getTypeAlignInChars(IdentQTy); | |||
1499 | unsigned Reserved2Flags = getDefaultLocationReserved2Flags(); | |||
1500 | FlagsTy FlagsKey(Flags, Reserved2Flags); | |||
1501 | llvm::Value *Entry = OpenMPDefaultLocMap.lookup(FlagsKey); | |||
1502 | if (!Entry) { | |||
1503 | if (!DefaultOpenMPPSource) { | |||
1504 | // Initialize default location for psource field of ident_t structure of | |||
1505 | // all ident_t objects. Format is ";file;function;line;column;;". | |||
1506 | // Taken from | |||
1507 | // https://github.com/llvm/llvm-project/blob/master/openmp/runtime/src/kmp_str.cpp | |||
1508 | DefaultOpenMPPSource = | |||
1509 | CGM.GetAddrOfConstantCString(";unknown;unknown;0;0;;").getPointer(); | |||
1510 | DefaultOpenMPPSource = | |||
1511 | llvm::ConstantExpr::getBitCast(DefaultOpenMPPSource, CGM.Int8PtrTy); | |||
1512 | } | |||
1513 | ||||
1514 | llvm::Constant *Data[] = { | |||
1515 | llvm::ConstantInt::getNullValue(CGM.Int32Ty), | |||
1516 | llvm::ConstantInt::get(CGM.Int32Ty, Flags), | |||
1517 | llvm::ConstantInt::get(CGM.Int32Ty, Reserved2Flags), | |||
1518 | llvm::ConstantInt::getNullValue(CGM.Int32Ty), DefaultOpenMPPSource}; | |||
1519 | llvm::GlobalValue *DefaultOpenMPLocation = | |||
1520 | createGlobalStruct(CGM, IdentQTy, isDefaultLocationConstant(), Data, "", | |||
1521 | llvm::GlobalValue::PrivateLinkage); | |||
1522 | DefaultOpenMPLocation->setUnnamedAddr( | |||
1523 | llvm::GlobalValue::UnnamedAddr::Global); | |||
1524 | ||||
1525 | OpenMPDefaultLocMap[FlagsKey] = Entry = DefaultOpenMPLocation; | |||
1526 | } | |||
1527 | return Address(Entry, Align); | |||
1528 | } | |||
1529 | ||||
1530 | void CGOpenMPRuntime::setLocThreadIdInsertPt(CodeGenFunction &CGF, | |||
1531 | bool AtCurrentPoint) { | |||
1532 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | |||
1533 | assert(!Elem.second.ServiceInsertPt && "Insert point is set already.")((!Elem.second.ServiceInsertPt && "Insert point is set already." ) ? static_cast<void> (0) : __assert_fail ("!Elem.second.ServiceInsertPt && \"Insert point is set already.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1533, __PRETTY_FUNCTION__)); | |||
1534 | ||||
1535 | llvm::Value *Undef = llvm::UndefValue::get(CGF.Int32Ty); | |||
1536 | if (AtCurrentPoint) { | |||
1537 | Elem.second.ServiceInsertPt = new llvm::BitCastInst( | |||
1538 | Undef, CGF.Int32Ty, "svcpt", CGF.Builder.GetInsertBlock()); | |||
1539 | } else { | |||
1540 | Elem.second.ServiceInsertPt = | |||
1541 | new llvm::BitCastInst(Undef, CGF.Int32Ty, "svcpt"); | |||
1542 | Elem.second.ServiceInsertPt->insertAfter(CGF.AllocaInsertPt); | |||
1543 | } | |||
1544 | } | |||
1545 | ||||
1546 | void CGOpenMPRuntime::clearLocThreadIdInsertPt(CodeGenFunction &CGF) { | |||
1547 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | |||
1548 | if (Elem.second.ServiceInsertPt) { | |||
1549 | llvm::Instruction *Ptr = Elem.second.ServiceInsertPt; | |||
1550 | Elem.second.ServiceInsertPt = nullptr; | |||
1551 | Ptr->eraseFromParent(); | |||
1552 | } | |||
1553 | } | |||
1554 | ||||
1555 | llvm::Value *CGOpenMPRuntime::emitUpdateLocation(CodeGenFunction &CGF, | |||
1556 | SourceLocation Loc, | |||
1557 | unsigned Flags) { | |||
1558 | Flags |= OMP_IDENT_KMPC; | |||
1559 | // If no debug info is generated - return global default location. | |||
1560 | if (CGM.getCodeGenOpts().getDebugInfo() == codegenoptions::NoDebugInfo || | |||
1561 | Loc.isInvalid()) | |||
1562 | return getOrCreateDefaultLocation(Flags).getPointer(); | |||
1563 | ||||
1564 | assert(CGF.CurFn && "No function in current CodeGenFunction.")((CGF.CurFn && "No function in current CodeGenFunction." ) ? static_cast<void> (0) : __assert_fail ("CGF.CurFn && \"No function in current CodeGenFunction.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1564, __PRETTY_FUNCTION__)); | |||
1565 | ||||
1566 | CharUnits Align = CGM.getContext().getTypeAlignInChars(IdentQTy); | |||
1567 | Address LocValue = Address::invalid(); | |||
1568 | auto I = OpenMPLocThreadIDMap.find(CGF.CurFn); | |||
1569 | if (I != OpenMPLocThreadIDMap.end()) | |||
1570 | LocValue = Address(I->second.DebugLoc, Align); | |||
1571 | ||||
1572 | // OpenMPLocThreadIDMap may have null DebugLoc and non-null ThreadID, if | |||
1573 | // GetOpenMPThreadID was called before this routine. | |||
1574 | if (!LocValue.isValid()) { | |||
1575 | // Generate "ident_t .kmpc_loc.addr;" | |||
1576 | Address AI = CGF.CreateMemTemp(IdentQTy, ".kmpc_loc.addr"); | |||
1577 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | |||
1578 | Elem.second.DebugLoc = AI.getPointer(); | |||
1579 | LocValue = AI; | |||
1580 | ||||
1581 | if (!Elem.second.ServiceInsertPt) | |||
1582 | setLocThreadIdInsertPt(CGF); | |||
1583 | CGBuilderTy::InsertPointGuard IPG(CGF.Builder); | |||
1584 | CGF.Builder.SetInsertPoint(Elem.second.ServiceInsertPt); | |||
1585 | CGF.Builder.CreateMemCpy(LocValue, getOrCreateDefaultLocation(Flags), | |||
1586 | CGF.getTypeSize(IdentQTy)); | |||
1587 | } | |||
1588 | ||||
1589 | // char **psource = &.kmpc_loc_<flags>.addr.psource; | |||
1590 | LValue Base = CGF.MakeAddrLValue(LocValue, IdentQTy); | |||
1591 | auto Fields = cast<RecordDecl>(IdentQTy->getAsTagDecl())->field_begin(); | |||
1592 | LValue PSource = | |||
1593 | CGF.EmitLValueForField(Base, *std::next(Fields, IdentField_PSource)); | |||
1594 | ||||
1595 | llvm::Value *OMPDebugLoc = OpenMPDebugLocMap.lookup(Loc.getRawEncoding()); | |||
1596 | if (OMPDebugLoc == nullptr) { | |||
1597 | SmallString<128> Buffer2; | |||
1598 | llvm::raw_svector_ostream OS2(Buffer2); | |||
1599 | // Build debug location | |||
1600 | PresumedLoc PLoc = CGF.getContext().getSourceManager().getPresumedLoc(Loc); | |||
1601 | OS2 << ";" << PLoc.getFilename() << ";"; | |||
1602 | if (const auto *FD = dyn_cast_or_null<FunctionDecl>(CGF.CurFuncDecl)) | |||
1603 | OS2 << FD->getQualifiedNameAsString(); | |||
1604 | OS2 << ";" << PLoc.getLine() << ";" << PLoc.getColumn() << ";;"; | |||
1605 | OMPDebugLoc = CGF.Builder.CreateGlobalStringPtr(OS2.str()); | |||
1606 | OpenMPDebugLocMap[Loc.getRawEncoding()] = OMPDebugLoc; | |||
1607 | } | |||
1608 | // *psource = ";<File>;<Function>;<Line>;<Column>;;"; | |||
1609 | CGF.EmitStoreOfScalar(OMPDebugLoc, PSource); | |||
1610 | ||||
1611 | // Our callers always pass this to a runtime function, so for | |||
1612 | // convenience, go ahead and return a naked pointer. | |||
1613 | return LocValue.getPointer(); | |||
1614 | } | |||
1615 | ||||
1616 | llvm::Value *CGOpenMPRuntime::getThreadID(CodeGenFunction &CGF, | |||
1617 | SourceLocation Loc) { | |||
1618 | assert(CGF.CurFn && "No function in current CodeGenFunction.")((CGF.CurFn && "No function in current CodeGenFunction." ) ? static_cast<void> (0) : __assert_fail ("CGF.CurFn && \"No function in current CodeGenFunction.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1618, __PRETTY_FUNCTION__)); | |||
1619 | ||||
1620 | llvm::Value *ThreadID = nullptr; | |||
1621 | // Check whether we've already cached a load of the thread id in this | |||
1622 | // function. | |||
1623 | auto I = OpenMPLocThreadIDMap.find(CGF.CurFn); | |||
1624 | if (I != OpenMPLocThreadIDMap.end()) { | |||
1625 | ThreadID = I->second.ThreadID; | |||
1626 | if (ThreadID != nullptr) | |||
1627 | return ThreadID; | |||
1628 | } | |||
1629 | // If exceptions are enabled, do not use parameter to avoid possible crash. | |||
1630 | if (!CGF.EHStack.requiresLandingPad() || !CGF.getLangOpts().Exceptions || | |||
1631 | !CGF.getLangOpts().CXXExceptions || | |||
1632 | CGF.Builder.GetInsertBlock() == CGF.AllocaInsertPt->getParent()) { | |||
1633 | if (auto *OMPRegionInfo = | |||
1634 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { | |||
1635 | if (OMPRegionInfo->getThreadIDVariable()) { | |||
1636 | // Check if this an outlined function with thread id passed as argument. | |||
1637 | LValue LVal = OMPRegionInfo->getThreadIDVariableLValue(CGF); | |||
1638 | ThreadID = CGF.EmitLoadOfScalar(LVal, Loc); | |||
1639 | // If value loaded in entry block, cache it and use it everywhere in | |||
1640 | // function. | |||
1641 | if (CGF.Builder.GetInsertBlock() == CGF.AllocaInsertPt->getParent()) { | |||
1642 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | |||
1643 | Elem.second.ThreadID = ThreadID; | |||
1644 | } | |||
1645 | return ThreadID; | |||
1646 | } | |||
1647 | } | |||
1648 | } | |||
1649 | ||||
1650 | // This is not an outlined function region - need to call __kmpc_int32 | |||
1651 | // kmpc_global_thread_num(ident_t *loc). | |||
1652 | // Generate thread id value and cache this value for use across the | |||
1653 | // function. | |||
1654 | auto &Elem = OpenMPLocThreadIDMap.FindAndConstruct(CGF.CurFn); | |||
1655 | if (!Elem.second.ServiceInsertPt) | |||
1656 | setLocThreadIdInsertPt(CGF); | |||
1657 | CGBuilderTy::InsertPointGuard IPG(CGF.Builder); | |||
1658 | CGF.Builder.SetInsertPoint(Elem.second.ServiceInsertPt); | |||
1659 | llvm::CallInst *Call = CGF.Builder.CreateCall( | |||
1660 | createRuntimeFunction(OMPRTL__kmpc_global_thread_num), | |||
1661 | emitUpdateLocation(CGF, Loc)); | |||
1662 | Call->setCallingConv(CGF.getRuntimeCC()); | |||
1663 | Elem.second.ThreadID = Call; | |||
1664 | return Call; | |||
1665 | } | |||
1666 | ||||
1667 | void CGOpenMPRuntime::functionFinished(CodeGenFunction &CGF) { | |||
1668 | assert(CGF.CurFn && "No function in current CodeGenFunction.")((CGF.CurFn && "No function in current CodeGenFunction." ) ? static_cast<void> (0) : __assert_fail ("CGF.CurFn && \"No function in current CodeGenFunction.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1668, __PRETTY_FUNCTION__)); | |||
1669 | if (OpenMPLocThreadIDMap.count(CGF.CurFn)) { | |||
1670 | clearLocThreadIdInsertPt(CGF); | |||
1671 | OpenMPLocThreadIDMap.erase(CGF.CurFn); | |||
1672 | } | |||
1673 | if (FunctionUDRMap.count(CGF.CurFn) > 0) { | |||
1674 | for(auto *D : FunctionUDRMap[CGF.CurFn]) | |||
1675 | UDRMap.erase(D); | |||
1676 | FunctionUDRMap.erase(CGF.CurFn); | |||
1677 | } | |||
1678 | } | |||
1679 | ||||
1680 | llvm::Type *CGOpenMPRuntime::getIdentTyPointerTy() { | |||
1681 | return IdentTy->getPointerTo(); | |||
1682 | } | |||
1683 | ||||
1684 | llvm::Type *CGOpenMPRuntime::getKmpc_MicroPointerTy() { | |||
1685 | if (!Kmpc_MicroTy) { | |||
1686 | // Build void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...) | |||
1687 | llvm::Type *MicroParams[] = {llvm::PointerType::getUnqual(CGM.Int32Ty), | |||
1688 | llvm::PointerType::getUnqual(CGM.Int32Ty)}; | |||
1689 | Kmpc_MicroTy = llvm::FunctionType::get(CGM.VoidTy, MicroParams, true); | |||
1690 | } | |||
1691 | return llvm::PointerType::getUnqual(Kmpc_MicroTy); | |||
1692 | } | |||
1693 | ||||
1694 | llvm::FunctionCallee CGOpenMPRuntime::createRuntimeFunction(unsigned Function) { | |||
1695 | llvm::FunctionCallee RTLFn = nullptr; | |||
1696 | switch (static_cast<OpenMPRTLFunction>(Function)) { | |||
1697 | case OMPRTL__kmpc_fork_call: { | |||
1698 | // Build void __kmpc_fork_call(ident_t *loc, kmp_int32 argc, kmpc_micro | |||
1699 | // microtask, ...); | |||
1700 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, | |||
1701 | getKmpc_MicroPointerTy()}; | |||
1702 | auto *FnTy = | |||
1703 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ true); | |||
1704 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_fork_call"); | |||
1705 | if (auto *F = dyn_cast<llvm::Function>(RTLFn.getCallee())) { | |||
1706 | if (!F->hasMetadata(llvm::LLVMContext::MD_callback)) { | |||
1707 | llvm::LLVMContext &Ctx = F->getContext(); | |||
1708 | llvm::MDBuilder MDB(Ctx); | |||
1709 | // Annotate the callback behavior of the __kmpc_fork_call: | |||
1710 | // - The callback callee is argument number 2 (microtask). | |||
1711 | // - The first two arguments of the callback callee are unknown (-1). | |||
1712 | // - All variadic arguments to the __kmpc_fork_call are passed to the | |||
1713 | // callback callee. | |||
1714 | F->addMetadata( | |||
1715 | llvm::LLVMContext::MD_callback, | |||
1716 | *llvm::MDNode::get(Ctx, {MDB.createCallbackEncoding( | |||
1717 | 2, {-1, -1}, | |||
1718 | /* VarArgsArePassed */ true)})); | |||
1719 | } | |||
1720 | } | |||
1721 | break; | |||
1722 | } | |||
1723 | case OMPRTL__kmpc_global_thread_num: { | |||
1724 | // Build kmp_int32 __kmpc_global_thread_num(ident_t *loc); | |||
1725 | llvm::Type *TypeParams[] = {getIdentTyPointerTy()}; | |||
1726 | auto *FnTy = | |||
1727 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
1728 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_global_thread_num"); | |||
1729 | break; | |||
1730 | } | |||
1731 | case OMPRTL__kmpc_threadprivate_cached: { | |||
1732 | // Build void *__kmpc_threadprivate_cached(ident_t *loc, | |||
1733 | // kmp_int32 global_tid, void *data, size_t size, void ***cache); | |||
1734 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, | |||
1735 | CGM.VoidPtrTy, CGM.SizeTy, | |||
1736 | CGM.VoidPtrTy->getPointerTo()->getPointerTo()}; | |||
1737 | auto *FnTy = | |||
1738 | llvm::FunctionType::get(CGM.VoidPtrTy, TypeParams, /*isVarArg*/ false); | |||
1739 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_threadprivate_cached"); | |||
1740 | break; | |||
1741 | } | |||
1742 | case OMPRTL__kmpc_critical: { | |||
1743 | // Build void __kmpc_critical(ident_t *loc, kmp_int32 global_tid, | |||
1744 | // kmp_critical_name *crit); | |||
1745 | llvm::Type *TypeParams[] = { | |||
1746 | getIdentTyPointerTy(), CGM.Int32Ty, | |||
1747 | llvm::PointerType::getUnqual(KmpCriticalNameTy)}; | |||
1748 | auto *FnTy = | |||
1749 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1750 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_critical"); | |||
1751 | break; | |||
1752 | } | |||
1753 | case OMPRTL__kmpc_critical_with_hint: { | |||
1754 | // Build void __kmpc_critical_with_hint(ident_t *loc, kmp_int32 global_tid, | |||
1755 | // kmp_critical_name *crit, uintptr_t hint); | |||
1756 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, | |||
1757 | llvm::PointerType::getUnqual(KmpCriticalNameTy), | |||
1758 | CGM.IntPtrTy}; | |||
1759 | auto *FnTy = | |||
1760 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1761 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_critical_with_hint"); | |||
1762 | break; | |||
1763 | } | |||
1764 | case OMPRTL__kmpc_threadprivate_register: { | |||
1765 | // Build void __kmpc_threadprivate_register(ident_t *, void *data, | |||
1766 | // kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor); | |||
1767 | // typedef void *(*kmpc_ctor)(void *); | |||
1768 | auto *KmpcCtorTy = | |||
1769 | llvm::FunctionType::get(CGM.VoidPtrTy, CGM.VoidPtrTy, | |||
1770 | /*isVarArg*/ false)->getPointerTo(); | |||
1771 | // typedef void *(*kmpc_cctor)(void *, void *); | |||
1772 | llvm::Type *KmpcCopyCtorTyArgs[] = {CGM.VoidPtrTy, CGM.VoidPtrTy}; | |||
1773 | auto *KmpcCopyCtorTy = | |||
1774 | llvm::FunctionType::get(CGM.VoidPtrTy, KmpcCopyCtorTyArgs, | |||
1775 | /*isVarArg*/ false) | |||
1776 | ->getPointerTo(); | |||
1777 | // typedef void (*kmpc_dtor)(void *); | |||
1778 | auto *KmpcDtorTy = | |||
1779 | llvm::FunctionType::get(CGM.VoidTy, CGM.VoidPtrTy, /*isVarArg*/ false) | |||
1780 | ->getPointerTo(); | |||
1781 | llvm::Type *FnTyArgs[] = {getIdentTyPointerTy(), CGM.VoidPtrTy, KmpcCtorTy, | |||
1782 | KmpcCopyCtorTy, KmpcDtorTy}; | |||
1783 | auto *FnTy = llvm::FunctionType::get(CGM.VoidTy, FnTyArgs, | |||
1784 | /*isVarArg*/ false); | |||
1785 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_threadprivate_register"); | |||
1786 | break; | |||
1787 | } | |||
1788 | case OMPRTL__kmpc_end_critical: { | |||
1789 | // Build void __kmpc_end_critical(ident_t *loc, kmp_int32 global_tid, | |||
1790 | // kmp_critical_name *crit); | |||
1791 | llvm::Type *TypeParams[] = { | |||
1792 | getIdentTyPointerTy(), CGM.Int32Ty, | |||
1793 | llvm::PointerType::getUnqual(KmpCriticalNameTy)}; | |||
1794 | auto *FnTy = | |||
1795 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1796 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_critical"); | |||
1797 | break; | |||
1798 | } | |||
1799 | case OMPRTL__kmpc_cancel_barrier: { | |||
1800 | // Build kmp_int32 __kmpc_cancel_barrier(ident_t *loc, kmp_int32 | |||
1801 | // global_tid); | |||
1802 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1803 | auto *FnTy = | |||
1804 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
1805 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name*/ "__kmpc_cancel_barrier"); | |||
1806 | break; | |||
1807 | } | |||
1808 | case OMPRTL__kmpc_barrier: { | |||
1809 | // Build void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid); | |||
1810 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1811 | auto *FnTy = | |||
1812 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1813 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name*/ "__kmpc_barrier"); | |||
1814 | break; | |||
1815 | } | |||
1816 | case OMPRTL__kmpc_for_static_fini: { | |||
1817 | // Build void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid); | |||
1818 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1819 | auto *FnTy = | |||
1820 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1821 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_for_static_fini"); | |||
1822 | break; | |||
1823 | } | |||
1824 | case OMPRTL__kmpc_push_num_threads: { | |||
1825 | // Build void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, | |||
1826 | // kmp_int32 num_threads) | |||
1827 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, | |||
1828 | CGM.Int32Ty}; | |||
1829 | auto *FnTy = | |||
1830 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1831 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_push_num_threads"); | |||
1832 | break; | |||
1833 | } | |||
1834 | case OMPRTL__kmpc_serialized_parallel: { | |||
1835 | // Build void __kmpc_serialized_parallel(ident_t *loc, kmp_int32 | |||
1836 | // global_tid); | |||
1837 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1838 | auto *FnTy = | |||
1839 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1840 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_serialized_parallel"); | |||
1841 | break; | |||
1842 | } | |||
1843 | case OMPRTL__kmpc_end_serialized_parallel: { | |||
1844 | // Build void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32 | |||
1845 | // global_tid); | |||
1846 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1847 | auto *FnTy = | |||
1848 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1849 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_serialized_parallel"); | |||
1850 | break; | |||
1851 | } | |||
1852 | case OMPRTL__kmpc_flush: { | |||
1853 | // Build void __kmpc_flush(ident_t *loc); | |||
1854 | llvm::Type *TypeParams[] = {getIdentTyPointerTy()}; | |||
1855 | auto *FnTy = | |||
1856 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
1857 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_flush"); | |||
1858 | break; | |||
1859 | } | |||
1860 | case OMPRTL__kmpc_master: { | |||
1861 | // Build kmp_int32 __kmpc_master(ident_t *loc, kmp_int32 global_tid); | |||
1862 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1863 | auto *FnTy = | |||
1864 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false); | |||
1865 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_master"); | |||
1866 | break; | |||
1867 | } | |||
1868 | case OMPRTL__kmpc_end_master: { | |||
1869 | // Build void __kmpc_end_master(ident_t *loc, kmp_int32 global_tid); | |||
1870 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1871 | auto *FnTy = | |||
1872 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
1873 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_end_master"); | |||
1874 | break; | |||
1875 | } | |||
1876 | case OMPRTL__kmpc_omp_taskyield: { | |||
1877 | // Build kmp_int32 __kmpc_omp_taskyield(ident_t *, kmp_int32 global_tid, | |||
1878 | // int end_part); | |||
1879 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.IntTy}; | |||
1880 | auto *FnTy = | |||
1881 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false); | |||
1882 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_omp_taskyield"); | |||
1883 | break; | |||
1884 | } | |||
1885 | case OMPRTL__kmpc_single: { | |||
1886 | // Build kmp_int32 __kmpc_single(ident_t *loc, kmp_int32 global_tid); | |||
1887 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1888 | auto *FnTy = | |||
1889 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false); | |||
1890 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_single"); | |||
1891 | break; | |||
1892 | } | |||
1893 | case OMPRTL__kmpc_end_single: { | |||
1894 | // Build void __kmpc_end_single(ident_t *loc, kmp_int32 global_tid); | |||
1895 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
1896 | auto *FnTy = | |||
1897 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
1898 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_end_single"); | |||
1899 | break; | |||
1900 | } | |||
1901 | case OMPRTL__kmpc_omp_task_alloc: { | |||
1902 | // Build kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, | |||
1903 | // kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, | |||
1904 | // kmp_routine_entry_t *task_entry); | |||
1905 | assert(KmpRoutineEntryPtrTy != nullptr &&((KmpRoutineEntryPtrTy != nullptr && "Type kmp_routine_entry_t must be created." ) ? static_cast<void> (0) : __assert_fail ("KmpRoutineEntryPtrTy != nullptr && \"Type kmp_routine_entry_t must be created.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1906, __PRETTY_FUNCTION__)) | |||
1906 | "Type kmp_routine_entry_t must be created.")((KmpRoutineEntryPtrTy != nullptr && "Type kmp_routine_entry_t must be created." ) ? static_cast<void> (0) : __assert_fail ("KmpRoutineEntryPtrTy != nullptr && \"Type kmp_routine_entry_t must be created.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 1906, __PRETTY_FUNCTION__)); | |||
1907 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.Int32Ty, | |||
1908 | CGM.SizeTy, CGM.SizeTy, KmpRoutineEntryPtrTy}; | |||
1909 | // Return void * and then cast to particular kmp_task_t type. | |||
1910 | auto *FnTy = | |||
1911 | llvm::FunctionType::get(CGM.VoidPtrTy, TypeParams, /*isVarArg=*/false); | |||
1912 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_omp_task_alloc"); | |||
1913 | break; | |||
1914 | } | |||
1915 | case OMPRTL__kmpc_omp_task: { | |||
1916 | // Build kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t | |||
1917 | // *new_task); | |||
1918 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, | |||
1919 | CGM.VoidPtrTy}; | |||
1920 | auto *FnTy = | |||
1921 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false); | |||
1922 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_omp_task"); | |||
1923 | break; | |||
1924 | } | |||
1925 | case OMPRTL__kmpc_copyprivate: { | |||
1926 | // Build void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid, | |||
1927 | // size_t cpy_size, void *cpy_data, void(*cpy_func)(void *, void *), | |||
1928 | // kmp_int32 didit); | |||
1929 | llvm::Type *CpyTypeParams[] = {CGM.VoidPtrTy, CGM.VoidPtrTy}; | |||
1930 | auto *CpyFnTy = | |||
1931 | llvm::FunctionType::get(CGM.VoidTy, CpyTypeParams, /*isVarArg=*/false); | |||
1932 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.SizeTy, | |||
1933 | CGM.VoidPtrTy, CpyFnTy->getPointerTo(), | |||
1934 | CGM.Int32Ty}; | |||
1935 | auto *FnTy = | |||
1936 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
1937 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_copyprivate"); | |||
1938 | break; | |||
1939 | } | |||
1940 | case OMPRTL__kmpc_reduce: { | |||
1941 | // Build kmp_int32 __kmpc_reduce(ident_t *loc, kmp_int32 global_tid, | |||
1942 | // kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void | |||
1943 | // (*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck); | |||
1944 | llvm::Type *ReduceTypeParams[] = {CGM.VoidPtrTy, CGM.VoidPtrTy}; | |||
1945 | auto *ReduceFnTy = llvm::FunctionType::get(CGM.VoidTy, ReduceTypeParams, | |||
1946 | /*isVarArg=*/false); | |||
1947 | llvm::Type *TypeParams[] = { | |||
1948 | getIdentTyPointerTy(), CGM.Int32Ty, CGM.Int32Ty, CGM.SizeTy, | |||
1949 | CGM.VoidPtrTy, ReduceFnTy->getPointerTo(), | |||
1950 | llvm::PointerType::getUnqual(KmpCriticalNameTy)}; | |||
1951 | auto *FnTy = | |||
1952 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false); | |||
1953 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_reduce"); | |||
1954 | break; | |||
1955 | } | |||
1956 | case OMPRTL__kmpc_reduce_nowait: { | |||
1957 | // Build kmp_int32 __kmpc_reduce_nowait(ident_t *loc, kmp_int32 | |||
1958 | // global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, | |||
1959 | // void (*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name | |||
1960 | // *lck); | |||
1961 | llvm::Type *ReduceTypeParams[] = {CGM.VoidPtrTy, CGM.VoidPtrTy}; | |||
1962 | auto *ReduceFnTy = llvm::FunctionType::get(CGM.VoidTy, ReduceTypeParams, | |||
1963 | /*isVarArg=*/false); | |||
1964 | llvm::Type *TypeParams[] = { | |||
1965 | getIdentTyPointerTy(), CGM.Int32Ty, CGM.Int32Ty, CGM.SizeTy, | |||
1966 | CGM.VoidPtrTy, ReduceFnTy->getPointerTo(), | |||
1967 | llvm::PointerType::getUnqual(KmpCriticalNameTy)}; | |||
1968 | auto *FnTy = | |||
1969 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false); | |||
1970 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_reduce_nowait"); | |||
1971 | break; | |||
1972 | } | |||
1973 | case OMPRTL__kmpc_end_reduce: { | |||
1974 | // Build void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, | |||
1975 | // kmp_critical_name *lck); | |||
1976 | llvm::Type *TypeParams[] = { | |||
1977 | getIdentTyPointerTy(), CGM.Int32Ty, | |||
1978 | llvm::PointerType::getUnqual(KmpCriticalNameTy)}; | |||
1979 | auto *FnTy = | |||
1980 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
1981 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_end_reduce"); | |||
1982 | break; | |||
1983 | } | |||
1984 | case OMPRTL__kmpc_end_reduce_nowait: { | |||
1985 | // Build __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, | |||
1986 | // kmp_critical_name *lck); | |||
1987 | llvm::Type *TypeParams[] = { | |||
1988 | getIdentTyPointerTy(), CGM.Int32Ty, | |||
1989 | llvm::PointerType::getUnqual(KmpCriticalNameTy)}; | |||
1990 | auto *FnTy = | |||
1991 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
1992 | RTLFn = | |||
1993 | CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_end_reduce_nowait"); | |||
1994 | break; | |||
1995 | } | |||
1996 | case OMPRTL__kmpc_omp_task_begin_if0: { | |||
1997 | // Build void __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t | |||
1998 | // *new_task); | |||
1999 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, | |||
2000 | CGM.VoidPtrTy}; | |||
2001 | auto *FnTy = | |||
2002 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2003 | RTLFn = | |||
2004 | CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_omp_task_begin_if0"); | |||
2005 | break; | |||
2006 | } | |||
2007 | case OMPRTL__kmpc_omp_task_complete_if0: { | |||
2008 | // Build void __kmpc_omp_task(ident_t *, kmp_int32 gtid, kmp_task_t | |||
2009 | // *new_task); | |||
2010 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, | |||
2011 | CGM.VoidPtrTy}; | |||
2012 | auto *FnTy = | |||
2013 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2014 | RTLFn = CGM.CreateRuntimeFunction(FnTy, | |||
2015 | /*Name=*/"__kmpc_omp_task_complete_if0"); | |||
2016 | break; | |||
2017 | } | |||
2018 | case OMPRTL__kmpc_ordered: { | |||
2019 | // Build void __kmpc_ordered(ident_t *loc, kmp_int32 global_tid); | |||
2020 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
2021 | auto *FnTy = | |||
2022 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2023 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_ordered"); | |||
2024 | break; | |||
2025 | } | |||
2026 | case OMPRTL__kmpc_end_ordered: { | |||
2027 | // Build void __kmpc_end_ordered(ident_t *loc, kmp_int32 global_tid); | |||
2028 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
2029 | auto *FnTy = | |||
2030 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2031 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_ordered"); | |||
2032 | break; | |||
2033 | } | |||
2034 | case OMPRTL__kmpc_omp_taskwait: { | |||
2035 | // Build kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32 global_tid); | |||
2036 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
2037 | auto *FnTy = | |||
2038 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false); | |||
2039 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_omp_taskwait"); | |||
2040 | break; | |||
2041 | } | |||
2042 | case OMPRTL__kmpc_taskgroup: { | |||
2043 | // Build void __kmpc_taskgroup(ident_t *loc, kmp_int32 global_tid); | |||
2044 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
2045 | auto *FnTy = | |||
2046 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2047 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_taskgroup"); | |||
2048 | break; | |||
2049 | } | |||
2050 | case OMPRTL__kmpc_end_taskgroup: { | |||
2051 | // Build void __kmpc_end_taskgroup(ident_t *loc, kmp_int32 global_tid); | |||
2052 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
2053 | auto *FnTy = | |||
2054 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2055 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_end_taskgroup"); | |||
2056 | break; | |||
2057 | } | |||
2058 | case OMPRTL__kmpc_push_proc_bind: { | |||
2059 | // Build void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid, | |||
2060 | // int proc_bind) | |||
2061 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.IntTy}; | |||
2062 | auto *FnTy = | |||
2063 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
2064 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_push_proc_bind"); | |||
2065 | break; | |||
2066 | } | |||
2067 | case OMPRTL__kmpc_omp_task_with_deps: { | |||
2068 | // Build kmp_int32 __kmpc_omp_task_with_deps(ident_t *, kmp_int32 gtid, | |||
2069 | // kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list, | |||
2070 | // kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list); | |||
2071 | llvm::Type *TypeParams[] = { | |||
2072 | getIdentTyPointerTy(), CGM.Int32Ty, CGM.VoidPtrTy, CGM.Int32Ty, | |||
2073 | CGM.VoidPtrTy, CGM.Int32Ty, CGM.VoidPtrTy}; | |||
2074 | auto *FnTy = | |||
2075 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg=*/false); | |||
2076 | RTLFn = | |||
2077 | CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_omp_task_with_deps"); | |||
2078 | break; | |||
2079 | } | |||
2080 | case OMPRTL__kmpc_omp_wait_deps: { | |||
2081 | // Build void __kmpc_omp_wait_deps(ident_t *, kmp_int32 gtid, | |||
2082 | // kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, | |||
2083 | // kmp_depend_info_t *noalias_dep_list); | |||
2084 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, | |||
2085 | CGM.Int32Ty, CGM.VoidPtrTy, | |||
2086 | CGM.Int32Ty, CGM.VoidPtrTy}; | |||
2087 | auto *FnTy = | |||
2088 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2089 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_omp_wait_deps"); | |||
2090 | break; | |||
2091 | } | |||
2092 | case OMPRTL__kmpc_cancellationpoint: { | |||
2093 | // Build kmp_int32 __kmpc_cancellationpoint(ident_t *loc, kmp_int32 | |||
2094 | // global_tid, kmp_int32 cncl_kind) | |||
2095 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.IntTy}; | |||
2096 | auto *FnTy = | |||
2097 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
2098 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_cancellationpoint"); | |||
2099 | break; | |||
2100 | } | |||
2101 | case OMPRTL__kmpc_cancel: { | |||
2102 | // Build kmp_int32 __kmpc_cancel(ident_t *loc, kmp_int32 global_tid, | |||
2103 | // kmp_int32 cncl_kind) | |||
2104 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.IntTy}; | |||
2105 | auto *FnTy = | |||
2106 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
2107 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_cancel"); | |||
2108 | break; | |||
2109 | } | |||
2110 | case OMPRTL__kmpc_push_num_teams: { | |||
2111 | // Build void kmpc_push_num_teams (ident_t loc, kmp_int32 global_tid, | |||
2112 | // kmp_int32 num_teams, kmp_int32 num_threads) | |||
2113 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, CGM.Int32Ty, | |||
2114 | CGM.Int32Ty}; | |||
2115 | auto *FnTy = | |||
2116 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
2117 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_push_num_teams"); | |||
2118 | break; | |||
2119 | } | |||
2120 | case OMPRTL__kmpc_fork_teams: { | |||
2121 | // Build void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro | |||
2122 | // microtask, ...); | |||
2123 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, | |||
2124 | getKmpc_MicroPointerTy()}; | |||
2125 | auto *FnTy = | |||
2126 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ true); | |||
2127 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_fork_teams"); | |||
2128 | if (auto *F = dyn_cast<llvm::Function>(RTLFn.getCallee())) { | |||
2129 | if (!F->hasMetadata(llvm::LLVMContext::MD_callback)) { | |||
2130 | llvm::LLVMContext &Ctx = F->getContext(); | |||
2131 | llvm::MDBuilder MDB(Ctx); | |||
2132 | // Annotate the callback behavior of the __kmpc_fork_teams: | |||
2133 | // - The callback callee is argument number 2 (microtask). | |||
2134 | // - The first two arguments of the callback callee are unknown (-1). | |||
2135 | // - All variadic arguments to the __kmpc_fork_teams are passed to the | |||
2136 | // callback callee. | |||
2137 | F->addMetadata( | |||
2138 | llvm::LLVMContext::MD_callback, | |||
2139 | *llvm::MDNode::get(Ctx, {MDB.createCallbackEncoding( | |||
2140 | 2, {-1, -1}, | |||
2141 | /* VarArgsArePassed */ true)})); | |||
2142 | } | |||
2143 | } | |||
2144 | break; | |||
2145 | } | |||
2146 | case OMPRTL__kmpc_taskloop: { | |||
2147 | // Build void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int | |||
2148 | // if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int | |||
2149 | // sched, kmp_uint64 grainsize, void *task_dup); | |||
2150 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), | |||
2151 | CGM.IntTy, | |||
2152 | CGM.VoidPtrTy, | |||
2153 | CGM.IntTy, | |||
2154 | CGM.Int64Ty->getPointerTo(), | |||
2155 | CGM.Int64Ty->getPointerTo(), | |||
2156 | CGM.Int64Ty, | |||
2157 | CGM.IntTy, | |||
2158 | CGM.IntTy, | |||
2159 | CGM.Int64Ty, | |||
2160 | CGM.VoidPtrTy}; | |||
2161 | auto *FnTy = | |||
2162 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2163 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_taskloop"); | |||
2164 | break; | |||
2165 | } | |||
2166 | case OMPRTL__kmpc_doacross_init: { | |||
2167 | // Build void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid, kmp_int32 | |||
2168 | // num_dims, struct kmp_dim *dims); | |||
2169 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), | |||
2170 | CGM.Int32Ty, | |||
2171 | CGM.Int32Ty, | |||
2172 | CGM.VoidPtrTy}; | |||
2173 | auto *FnTy = | |||
2174 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2175 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_doacross_init"); | |||
2176 | break; | |||
2177 | } | |||
2178 | case OMPRTL__kmpc_doacross_fini: { | |||
2179 | // Build void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid); | |||
2180 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty}; | |||
2181 | auto *FnTy = | |||
2182 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2183 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_doacross_fini"); | |||
2184 | break; | |||
2185 | } | |||
2186 | case OMPRTL__kmpc_doacross_post: { | |||
2187 | // Build void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid, kmp_int64 | |||
2188 | // *vec); | |||
2189 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, | |||
2190 | CGM.Int64Ty->getPointerTo()}; | |||
2191 | auto *FnTy = | |||
2192 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2193 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_doacross_post"); | |||
2194 | break; | |||
2195 | } | |||
2196 | case OMPRTL__kmpc_doacross_wait: { | |||
2197 | // Build void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid, kmp_int64 | |||
2198 | // *vec); | |||
2199 | llvm::Type *TypeParams[] = {getIdentTyPointerTy(), CGM.Int32Ty, | |||
2200 | CGM.Int64Ty->getPointerTo()}; | |||
2201 | auto *FnTy = | |||
2202 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2203 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_doacross_wait"); | |||
2204 | break; | |||
2205 | } | |||
2206 | case OMPRTL__kmpc_task_reduction_init: { | |||
2207 | // Build void *__kmpc_task_reduction_init(int gtid, int num_data, void | |||
2208 | // *data); | |||
2209 | llvm::Type *TypeParams[] = {CGM.IntTy, CGM.IntTy, CGM.VoidPtrTy}; | |||
2210 | auto *FnTy = | |||
2211 | llvm::FunctionType::get(CGM.VoidPtrTy, TypeParams, /*isVarArg=*/false); | |||
2212 | RTLFn = | |||
2213 | CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_task_reduction_init"); | |||
2214 | break; | |||
2215 | } | |||
2216 | case OMPRTL__kmpc_task_reduction_get_th_data: { | |||
2217 | // Build void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void | |||
2218 | // *d); | |||
2219 | llvm::Type *TypeParams[] = {CGM.IntTy, CGM.VoidPtrTy, CGM.VoidPtrTy}; | |||
2220 | auto *FnTy = | |||
2221 | llvm::FunctionType::get(CGM.VoidPtrTy, TypeParams, /*isVarArg=*/false); | |||
2222 | RTLFn = CGM.CreateRuntimeFunction( | |||
2223 | FnTy, /*Name=*/"__kmpc_task_reduction_get_th_data"); | |||
2224 | break; | |||
2225 | } | |||
2226 | case OMPRTL__kmpc_alloc: { | |||
2227 | // Build to void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t | |||
2228 | // al); omp_allocator_handle_t type is void *. | |||
2229 | llvm::Type *TypeParams[] = {CGM.IntTy, CGM.SizeTy, CGM.VoidPtrTy}; | |||
2230 | auto *FnTy = | |||
2231 | llvm::FunctionType::get(CGM.VoidPtrTy, TypeParams, /*isVarArg=*/false); | |||
2232 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_alloc"); | |||
2233 | break; | |||
2234 | } | |||
2235 | case OMPRTL__kmpc_free: { | |||
2236 | // Build to void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t | |||
2237 | // al); omp_allocator_handle_t type is void *. | |||
2238 | llvm::Type *TypeParams[] = {CGM.IntTy, CGM.VoidPtrTy, CGM.VoidPtrTy}; | |||
2239 | auto *FnTy = | |||
2240 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2241 | RTLFn = CGM.CreateRuntimeFunction(FnTy, /*Name=*/"__kmpc_free"); | |||
2242 | break; | |||
2243 | } | |||
2244 | case OMPRTL__kmpc_push_target_tripcount: { | |||
2245 | // Build void __kmpc_push_target_tripcount(int64_t device_id, kmp_uint64 | |||
2246 | // size); | |||
2247 | llvm::Type *TypeParams[] = {CGM.Int64Ty, CGM.Int64Ty}; | |||
2248 | llvm::FunctionType *FnTy = | |||
2249 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2250 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__kmpc_push_target_tripcount"); | |||
2251 | break; | |||
2252 | } | |||
2253 | case OMPRTL__tgt_target: { | |||
2254 | // Build int32_t __tgt_target(int64_t device_id, void *host_ptr, int32_t | |||
2255 | // arg_num, void** args_base, void **args, size_t *arg_sizes, int64_t | |||
2256 | // *arg_types); | |||
2257 | llvm::Type *TypeParams[] = {CGM.Int64Ty, | |||
2258 | CGM.VoidPtrTy, | |||
2259 | CGM.Int32Ty, | |||
2260 | CGM.VoidPtrPtrTy, | |||
2261 | CGM.VoidPtrPtrTy, | |||
2262 | CGM.SizeTy->getPointerTo(), | |||
2263 | CGM.Int64Ty->getPointerTo()}; | |||
2264 | auto *FnTy = | |||
2265 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
2266 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target"); | |||
2267 | break; | |||
2268 | } | |||
2269 | case OMPRTL__tgt_target_nowait: { | |||
2270 | // Build int32_t __tgt_target_nowait(int64_t device_id, void *host_ptr, | |||
2271 | // int32_t arg_num, void** args_base, void **args, size_t *arg_sizes, | |||
2272 | // int64_t *arg_types); | |||
2273 | llvm::Type *TypeParams[] = {CGM.Int64Ty, | |||
2274 | CGM.VoidPtrTy, | |||
2275 | CGM.Int32Ty, | |||
2276 | CGM.VoidPtrPtrTy, | |||
2277 | CGM.VoidPtrPtrTy, | |||
2278 | CGM.SizeTy->getPointerTo(), | |||
2279 | CGM.Int64Ty->getPointerTo()}; | |||
2280 | auto *FnTy = | |||
2281 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
2282 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_nowait"); | |||
2283 | break; | |||
2284 | } | |||
2285 | case OMPRTL__tgt_target_teams: { | |||
2286 | // Build int32_t __tgt_target_teams(int64_t device_id, void *host_ptr, | |||
2287 | // int32_t arg_num, void** args_base, void **args, size_t *arg_sizes, | |||
2288 | // int64_t *arg_types, int32_t num_teams, int32_t thread_limit); | |||
2289 | llvm::Type *TypeParams[] = {CGM.Int64Ty, | |||
2290 | CGM.VoidPtrTy, | |||
2291 | CGM.Int32Ty, | |||
2292 | CGM.VoidPtrPtrTy, | |||
2293 | CGM.VoidPtrPtrTy, | |||
2294 | CGM.SizeTy->getPointerTo(), | |||
2295 | CGM.Int64Ty->getPointerTo(), | |||
2296 | CGM.Int32Ty, | |||
2297 | CGM.Int32Ty}; | |||
2298 | auto *FnTy = | |||
2299 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
2300 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_teams"); | |||
2301 | break; | |||
2302 | } | |||
2303 | case OMPRTL__tgt_target_teams_nowait: { | |||
2304 | // Build int32_t __tgt_target_teams_nowait(int64_t device_id, void | |||
2305 | // *host_ptr, int32_t arg_num, void** args_base, void **args, size_t | |||
2306 | // *arg_sizes, int64_t *arg_types, int32_t num_teams, int32_t thread_limit); | |||
2307 | llvm::Type *TypeParams[] = {CGM.Int64Ty, | |||
2308 | CGM.VoidPtrTy, | |||
2309 | CGM.Int32Ty, | |||
2310 | CGM.VoidPtrPtrTy, | |||
2311 | CGM.VoidPtrPtrTy, | |||
2312 | CGM.SizeTy->getPointerTo(), | |||
2313 | CGM.Int64Ty->getPointerTo(), | |||
2314 | CGM.Int32Ty, | |||
2315 | CGM.Int32Ty}; | |||
2316 | auto *FnTy = | |||
2317 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
2318 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_teams_nowait"); | |||
2319 | break; | |||
2320 | } | |||
2321 | case OMPRTL__tgt_register_requires: { | |||
2322 | // Build void __tgt_register_requires(int64_t flags); | |||
2323 | llvm::Type *TypeParams[] = {CGM.Int64Ty}; | |||
2324 | auto *FnTy = | |||
2325 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
2326 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_register_requires"); | |||
2327 | break; | |||
2328 | } | |||
2329 | case OMPRTL__tgt_register_lib: { | |||
2330 | // Build void __tgt_register_lib(__tgt_bin_desc *desc); | |||
2331 | QualType ParamTy = | |||
2332 | CGM.getContext().getPointerType(getTgtBinaryDescriptorQTy()); | |||
2333 | llvm::Type *TypeParams[] = {CGM.getTypes().ConvertTypeForMem(ParamTy)}; | |||
2334 | auto *FnTy = | |||
2335 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
2336 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_register_lib"); | |||
2337 | break; | |||
2338 | } | |||
2339 | case OMPRTL__tgt_unregister_lib: { | |||
2340 | // Build void __tgt_unregister_lib(__tgt_bin_desc *desc); | |||
2341 | QualType ParamTy = | |||
2342 | CGM.getContext().getPointerType(getTgtBinaryDescriptorQTy()); | |||
2343 | llvm::Type *TypeParams[] = {CGM.getTypes().ConvertTypeForMem(ParamTy)}; | |||
2344 | auto *FnTy = | |||
2345 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
2346 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_unregister_lib"); | |||
2347 | break; | |||
2348 | } | |||
2349 | case OMPRTL__tgt_target_data_begin: { | |||
2350 | // Build void __tgt_target_data_begin(int64_t device_id, int32_t arg_num, | |||
2351 | // void** args_base, void **args, size_t *arg_sizes, int64_t *arg_types); | |||
2352 | llvm::Type *TypeParams[] = {CGM.Int64Ty, | |||
2353 | CGM.Int32Ty, | |||
2354 | CGM.VoidPtrPtrTy, | |||
2355 | CGM.VoidPtrPtrTy, | |||
2356 | CGM.SizeTy->getPointerTo(), | |||
2357 | CGM.Int64Ty->getPointerTo()}; | |||
2358 | auto *FnTy = | |||
2359 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
2360 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_data_begin"); | |||
2361 | break; | |||
2362 | } | |||
2363 | case OMPRTL__tgt_target_data_begin_nowait: { | |||
2364 | // Build void __tgt_target_data_begin_nowait(int64_t device_id, int32_t | |||
2365 | // arg_num, void** args_base, void **args, size_t *arg_sizes, int64_t | |||
2366 | // *arg_types); | |||
2367 | llvm::Type *TypeParams[] = {CGM.Int64Ty, | |||
2368 | CGM.Int32Ty, | |||
2369 | CGM.VoidPtrPtrTy, | |||
2370 | CGM.VoidPtrPtrTy, | |||
2371 | CGM.SizeTy->getPointerTo(), | |||
2372 | CGM.Int64Ty->getPointerTo()}; | |||
2373 | auto *FnTy = | |||
2374 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2375 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_data_begin_nowait"); | |||
2376 | break; | |||
2377 | } | |||
2378 | case OMPRTL__tgt_target_data_end: { | |||
2379 | // Build void __tgt_target_data_end(int64_t device_id, int32_t arg_num, | |||
2380 | // void** args_base, void **args, size_t *arg_sizes, int64_t *arg_types); | |||
2381 | llvm::Type *TypeParams[] = {CGM.Int64Ty, | |||
2382 | CGM.Int32Ty, | |||
2383 | CGM.VoidPtrPtrTy, | |||
2384 | CGM.VoidPtrPtrTy, | |||
2385 | CGM.SizeTy->getPointerTo(), | |||
2386 | CGM.Int64Ty->getPointerTo()}; | |||
2387 | auto *FnTy = | |||
2388 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
2389 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_data_end"); | |||
2390 | break; | |||
2391 | } | |||
2392 | case OMPRTL__tgt_target_data_end_nowait: { | |||
2393 | // Build void __tgt_target_data_end_nowait(int64_t device_id, int32_t | |||
2394 | // arg_num, void** args_base, void **args, size_t *arg_sizes, int64_t | |||
2395 | // *arg_types); | |||
2396 | llvm::Type *TypeParams[] = {CGM.Int64Ty, | |||
2397 | CGM.Int32Ty, | |||
2398 | CGM.VoidPtrPtrTy, | |||
2399 | CGM.VoidPtrPtrTy, | |||
2400 | CGM.SizeTy->getPointerTo(), | |||
2401 | CGM.Int64Ty->getPointerTo()}; | |||
2402 | auto *FnTy = | |||
2403 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2404 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_data_end_nowait"); | |||
2405 | break; | |||
2406 | } | |||
2407 | case OMPRTL__tgt_target_data_update: { | |||
2408 | // Build void __tgt_target_data_update(int64_t device_id, int32_t arg_num, | |||
2409 | // void** args_base, void **args, size_t *arg_sizes, int64_t *arg_types); | |||
2410 | llvm::Type *TypeParams[] = {CGM.Int64Ty, | |||
2411 | CGM.Int32Ty, | |||
2412 | CGM.VoidPtrPtrTy, | |||
2413 | CGM.VoidPtrPtrTy, | |||
2414 | CGM.SizeTy->getPointerTo(), | |||
2415 | CGM.Int64Ty->getPointerTo()}; | |||
2416 | auto *FnTy = | |||
2417 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
2418 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_data_update"); | |||
2419 | break; | |||
2420 | } | |||
2421 | case OMPRTL__tgt_target_data_update_nowait: { | |||
2422 | // Build void __tgt_target_data_update_nowait(int64_t device_id, int32_t | |||
2423 | // arg_num, void** args_base, void **args, size_t *arg_sizes, int64_t | |||
2424 | // *arg_types); | |||
2425 | llvm::Type *TypeParams[] = {CGM.Int64Ty, | |||
2426 | CGM.Int32Ty, | |||
2427 | CGM.VoidPtrPtrTy, | |||
2428 | CGM.VoidPtrPtrTy, | |||
2429 | CGM.SizeTy->getPointerTo(), | |||
2430 | CGM.Int64Ty->getPointerTo()}; | |||
2431 | auto *FnTy = | |||
2432 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2433 | RTLFn = CGM.CreateRuntimeFunction(FnTy, "__tgt_target_data_update_nowait"); | |||
2434 | break; | |||
2435 | } | |||
2436 | } | |||
2437 | assert(RTLFn && "Unable to find OpenMP runtime function")((RTLFn && "Unable to find OpenMP runtime function") ? static_cast<void> (0) : __assert_fail ("RTLFn && \"Unable to find OpenMP runtime function\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2437, __PRETTY_FUNCTION__)); | |||
2438 | return RTLFn; | |||
2439 | } | |||
2440 | ||||
2441 | llvm::FunctionCallee | |||
2442 | CGOpenMPRuntime::createForStaticInitFunction(unsigned IVSize, bool IVSigned) { | |||
2443 | assert((IVSize == 32 || IVSize == 64) &&(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2444, __PRETTY_FUNCTION__)) | |||
2444 | "IV size is not compatible with the omp runtime")(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2444, __PRETTY_FUNCTION__)); | |||
2445 | StringRef Name = IVSize == 32 ? (IVSigned ? "__kmpc_for_static_init_4" | |||
2446 | : "__kmpc_for_static_init_4u") | |||
2447 | : (IVSigned ? "__kmpc_for_static_init_8" | |||
2448 | : "__kmpc_for_static_init_8u"); | |||
2449 | llvm::Type *ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty; | |||
2450 | auto *PtrTy = llvm::PointerType::getUnqual(ITy); | |||
2451 | llvm::Type *TypeParams[] = { | |||
2452 | getIdentTyPointerTy(), // loc | |||
2453 | CGM.Int32Ty, // tid | |||
2454 | CGM.Int32Ty, // schedtype | |||
2455 | llvm::PointerType::getUnqual(CGM.Int32Ty), // p_lastiter | |||
2456 | PtrTy, // p_lower | |||
2457 | PtrTy, // p_upper | |||
2458 | PtrTy, // p_stride | |||
2459 | ITy, // incr | |||
2460 | ITy // chunk | |||
2461 | }; | |||
2462 | auto *FnTy = | |||
2463 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
2464 | return CGM.CreateRuntimeFunction(FnTy, Name); | |||
2465 | } | |||
2466 | ||||
2467 | llvm::FunctionCallee | |||
2468 | CGOpenMPRuntime::createDispatchInitFunction(unsigned IVSize, bool IVSigned) { | |||
2469 | assert((IVSize == 32 || IVSize == 64) &&(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2470, __PRETTY_FUNCTION__)) | |||
2470 | "IV size is not compatible with the omp runtime")(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2470, __PRETTY_FUNCTION__)); | |||
2471 | StringRef Name = | |||
2472 | IVSize == 32 | |||
2473 | ? (IVSigned ? "__kmpc_dispatch_init_4" : "__kmpc_dispatch_init_4u") | |||
2474 | : (IVSigned ? "__kmpc_dispatch_init_8" : "__kmpc_dispatch_init_8u"); | |||
2475 | llvm::Type *ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty; | |||
2476 | llvm::Type *TypeParams[] = { getIdentTyPointerTy(), // loc | |||
2477 | CGM.Int32Ty, // tid | |||
2478 | CGM.Int32Ty, // schedtype | |||
2479 | ITy, // lower | |||
2480 | ITy, // upper | |||
2481 | ITy, // stride | |||
2482 | ITy // chunk | |||
2483 | }; | |||
2484 | auto *FnTy = | |||
2485 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg*/ false); | |||
2486 | return CGM.CreateRuntimeFunction(FnTy, Name); | |||
2487 | } | |||
2488 | ||||
2489 | llvm::FunctionCallee | |||
2490 | CGOpenMPRuntime::createDispatchFiniFunction(unsigned IVSize, bool IVSigned) { | |||
2491 | assert((IVSize == 32 || IVSize == 64) &&(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2492, __PRETTY_FUNCTION__)) | |||
2492 | "IV size is not compatible with the omp runtime")(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2492, __PRETTY_FUNCTION__)); | |||
2493 | StringRef Name = | |||
2494 | IVSize == 32 | |||
2495 | ? (IVSigned ? "__kmpc_dispatch_fini_4" : "__kmpc_dispatch_fini_4u") | |||
2496 | : (IVSigned ? "__kmpc_dispatch_fini_8" : "__kmpc_dispatch_fini_8u"); | |||
2497 | llvm::Type *TypeParams[] = { | |||
2498 | getIdentTyPointerTy(), // loc | |||
2499 | CGM.Int32Ty, // tid | |||
2500 | }; | |||
2501 | auto *FnTy = | |||
2502 | llvm::FunctionType::get(CGM.VoidTy, TypeParams, /*isVarArg=*/false); | |||
2503 | return CGM.CreateRuntimeFunction(FnTy, Name); | |||
2504 | } | |||
2505 | ||||
2506 | llvm::FunctionCallee | |||
2507 | CGOpenMPRuntime::createDispatchNextFunction(unsigned IVSize, bool IVSigned) { | |||
2508 | assert((IVSize == 32 || IVSize == 64) &&(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2509, __PRETTY_FUNCTION__)) | |||
2509 | "IV size is not compatible with the omp runtime")(((IVSize == 32 || IVSize == 64) && "IV size is not compatible with the omp runtime" ) ? static_cast<void> (0) : __assert_fail ("(IVSize == 32 || IVSize == 64) && \"IV size is not compatible with the omp runtime\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2509, __PRETTY_FUNCTION__)); | |||
2510 | StringRef Name = | |||
2511 | IVSize == 32 | |||
2512 | ? (IVSigned ? "__kmpc_dispatch_next_4" : "__kmpc_dispatch_next_4u") | |||
2513 | : (IVSigned ? "__kmpc_dispatch_next_8" : "__kmpc_dispatch_next_8u"); | |||
2514 | llvm::Type *ITy = IVSize == 32 ? CGM.Int32Ty : CGM.Int64Ty; | |||
2515 | auto *PtrTy = llvm::PointerType::getUnqual(ITy); | |||
2516 | llvm::Type *TypeParams[] = { | |||
2517 | getIdentTyPointerTy(), // loc | |||
2518 | CGM.Int32Ty, // tid | |||
2519 | llvm::PointerType::getUnqual(CGM.Int32Ty), // p_lastiter | |||
2520 | PtrTy, // p_lower | |||
2521 | PtrTy, // p_upper | |||
2522 | PtrTy // p_stride | |||
2523 | }; | |||
2524 | auto *FnTy = | |||
2525 | llvm::FunctionType::get(CGM.Int32Ty, TypeParams, /*isVarArg*/ false); | |||
2526 | return CGM.CreateRuntimeFunction(FnTy, Name); | |||
2527 | } | |||
2528 | ||||
2529 | Address CGOpenMPRuntime::getAddrOfDeclareTargetLink(const VarDecl *VD) { | |||
2530 | if (CGM.getLangOpts().OpenMPSimd) | |||
2531 | return Address::invalid(); | |||
2532 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
2533 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | |||
2534 | if (Res && *Res == OMPDeclareTargetDeclAttr::MT_Link) { | |||
2535 | SmallString<64> PtrName; | |||
2536 | { | |||
2537 | llvm::raw_svector_ostream OS(PtrName); | |||
2538 | OS << CGM.getMangledName(GlobalDecl(VD)) << "_decl_tgt_link_ptr"; | |||
2539 | } | |||
2540 | llvm::Value *Ptr = CGM.getModule().getNamedValue(PtrName); | |||
2541 | if (!Ptr) { | |||
2542 | QualType PtrTy = CGM.getContext().getPointerType(VD->getType()); | |||
2543 | Ptr = getOrCreateInternalVariable(CGM.getTypes().ConvertTypeForMem(PtrTy), | |||
2544 | PtrName); | |||
2545 | if (!CGM.getLangOpts().OpenMPIsDevice) { | |||
2546 | auto *GV = cast<llvm::GlobalVariable>(Ptr); | |||
2547 | GV->setLinkage(llvm::GlobalValue::ExternalLinkage); | |||
2548 | GV->setInitializer(CGM.GetAddrOfGlobal(VD)); | |||
2549 | } | |||
2550 | CGM.addUsedGlobal(cast<llvm::GlobalValue>(Ptr)); | |||
2551 | registerTargetGlobalVariable(VD, cast<llvm::Constant>(Ptr)); | |||
2552 | } | |||
2553 | return Address(Ptr, CGM.getContext().getDeclAlign(VD)); | |||
2554 | } | |||
2555 | return Address::invalid(); | |||
2556 | } | |||
2557 | ||||
2558 | llvm::Constant * | |||
2559 | CGOpenMPRuntime::getOrCreateThreadPrivateCache(const VarDecl *VD) { | |||
2560 | assert(!CGM.getLangOpts().OpenMPUseTLS ||((!CGM.getLangOpts().OpenMPUseTLS || !CGM.getContext().getTargetInfo ().isTLSSupported()) ? static_cast<void> (0) : __assert_fail ("!CGM.getLangOpts().OpenMPUseTLS || !CGM.getContext().getTargetInfo().isTLSSupported()" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2561, __PRETTY_FUNCTION__)) | |||
2561 | !CGM.getContext().getTargetInfo().isTLSSupported())((!CGM.getLangOpts().OpenMPUseTLS || !CGM.getContext().getTargetInfo ().isTLSSupported()) ? static_cast<void> (0) : __assert_fail ("!CGM.getLangOpts().OpenMPUseTLS || !CGM.getContext().getTargetInfo().isTLSSupported()" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2561, __PRETTY_FUNCTION__)); | |||
2562 | // Lookup the entry, lazily creating it if necessary. | |||
2563 | std::string Suffix = getName({"cache", ""}); | |||
2564 | return getOrCreateInternalVariable( | |||
2565 | CGM.Int8PtrPtrTy, Twine(CGM.getMangledName(VD)).concat(Suffix)); | |||
2566 | } | |||
2567 | ||||
2568 | Address CGOpenMPRuntime::getAddrOfThreadPrivate(CodeGenFunction &CGF, | |||
2569 | const VarDecl *VD, | |||
2570 | Address VDAddr, | |||
2571 | SourceLocation Loc) { | |||
2572 | if (CGM.getLangOpts().OpenMPUseTLS && | |||
2573 | CGM.getContext().getTargetInfo().isTLSSupported()) | |||
2574 | return VDAddr; | |||
2575 | ||||
2576 | llvm::Type *VarTy = VDAddr.getElementType(); | |||
2577 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
2578 | CGF.Builder.CreatePointerCast(VDAddr.getPointer(), | |||
2579 | CGM.Int8PtrTy), | |||
2580 | CGM.getSize(CGM.GetTargetTypeStoreSize(VarTy)), | |||
2581 | getOrCreateThreadPrivateCache(VD)}; | |||
2582 | return Address(CGF.EmitRuntimeCall( | |||
2583 | createRuntimeFunction(OMPRTL__kmpc_threadprivate_cached), Args), | |||
2584 | VDAddr.getAlignment()); | |||
2585 | } | |||
2586 | ||||
2587 | void CGOpenMPRuntime::emitThreadPrivateVarInit( | |||
2588 | CodeGenFunction &CGF, Address VDAddr, llvm::Value *Ctor, | |||
2589 | llvm::Value *CopyCtor, llvm::Value *Dtor, SourceLocation Loc) { | |||
2590 | // Call kmp_int32 __kmpc_global_thread_num(&loc) to init OpenMP runtime | |||
2591 | // library. | |||
2592 | llvm::Value *OMPLoc = emitUpdateLocation(CGF, Loc); | |||
2593 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_global_thread_num), | |||
2594 | OMPLoc); | |||
2595 | // Call __kmpc_threadprivate_register(&loc, &var, ctor, cctor/*NULL*/, dtor) | |||
2596 | // to register constructor/destructor for variable. | |||
2597 | llvm::Value *Args[] = { | |||
2598 | OMPLoc, CGF.Builder.CreatePointerCast(VDAddr.getPointer(), CGM.VoidPtrTy), | |||
2599 | Ctor, CopyCtor, Dtor}; | |||
2600 | CGF.EmitRuntimeCall( | |||
2601 | createRuntimeFunction(OMPRTL__kmpc_threadprivate_register), Args); | |||
2602 | } | |||
2603 | ||||
2604 | llvm::Function *CGOpenMPRuntime::emitThreadPrivateVarDefinition( | |||
2605 | const VarDecl *VD, Address VDAddr, SourceLocation Loc, | |||
2606 | bool PerformInit, CodeGenFunction *CGF) { | |||
2607 | if (CGM.getLangOpts().OpenMPUseTLS && | |||
2608 | CGM.getContext().getTargetInfo().isTLSSupported()) | |||
2609 | return nullptr; | |||
2610 | ||||
2611 | VD = VD->getDefinition(CGM.getContext()); | |||
2612 | if (VD && ThreadPrivateWithDefinition.insert(CGM.getMangledName(VD)).second) { | |||
2613 | QualType ASTTy = VD->getType(); | |||
2614 | ||||
2615 | llvm::Value *Ctor = nullptr, *CopyCtor = nullptr, *Dtor = nullptr; | |||
2616 | const Expr *Init = VD->getAnyInitializer(); | |||
2617 | if (CGM.getLangOpts().CPlusPlus && PerformInit) { | |||
2618 | // Generate function that re-emits the declaration's initializer into the | |||
2619 | // threadprivate copy of the variable VD | |||
2620 | CodeGenFunction CtorCGF(CGM); | |||
2621 | FunctionArgList Args; | |||
2622 | ImplicitParamDecl Dst(CGM.getContext(), /*DC=*/nullptr, Loc, | |||
2623 | /*Id=*/nullptr, CGM.getContext().VoidPtrTy, | |||
2624 | ImplicitParamDecl::Other); | |||
2625 | Args.push_back(&Dst); | |||
2626 | ||||
2627 | const auto &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration( | |||
2628 | CGM.getContext().VoidPtrTy, Args); | |||
2629 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | |||
2630 | std::string Name = getName({"__kmpc_global_ctor_", ""}); | |||
2631 | llvm::Function *Fn = | |||
2632 | CGM.CreateGlobalInitOrDestructFunction(FTy, Name, FI, Loc); | |||
2633 | CtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidPtrTy, Fn, FI, | |||
2634 | Args, Loc, Loc); | |||
2635 | llvm::Value *ArgVal = CtorCGF.EmitLoadOfScalar( | |||
2636 | CtorCGF.GetAddrOfLocalVar(&Dst), /*Volatile=*/false, | |||
2637 | CGM.getContext().VoidPtrTy, Dst.getLocation()); | |||
2638 | Address Arg = Address(ArgVal, VDAddr.getAlignment()); | |||
2639 | Arg = CtorCGF.Builder.CreateElementBitCast( | |||
2640 | Arg, CtorCGF.ConvertTypeForMem(ASTTy)); | |||
2641 | CtorCGF.EmitAnyExprToMem(Init, Arg, Init->getType().getQualifiers(), | |||
2642 | /*IsInitializer=*/true); | |||
2643 | ArgVal = CtorCGF.EmitLoadOfScalar( | |||
2644 | CtorCGF.GetAddrOfLocalVar(&Dst), /*Volatile=*/false, | |||
2645 | CGM.getContext().VoidPtrTy, Dst.getLocation()); | |||
2646 | CtorCGF.Builder.CreateStore(ArgVal, CtorCGF.ReturnValue); | |||
2647 | CtorCGF.FinishFunction(); | |||
2648 | Ctor = Fn; | |||
2649 | } | |||
2650 | if (VD->getType().isDestructedType() != QualType::DK_none) { | |||
2651 | // Generate function that emits destructor call for the threadprivate copy | |||
2652 | // of the variable VD | |||
2653 | CodeGenFunction DtorCGF(CGM); | |||
2654 | FunctionArgList Args; | |||
2655 | ImplicitParamDecl Dst(CGM.getContext(), /*DC=*/nullptr, Loc, | |||
2656 | /*Id=*/nullptr, CGM.getContext().VoidPtrTy, | |||
2657 | ImplicitParamDecl::Other); | |||
2658 | Args.push_back(&Dst); | |||
2659 | ||||
2660 | const auto &FI = CGM.getTypes().arrangeBuiltinFunctionDeclaration( | |||
2661 | CGM.getContext().VoidTy, Args); | |||
2662 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | |||
2663 | std::string Name = getName({"__kmpc_global_dtor_", ""}); | |||
2664 | llvm::Function *Fn = | |||
2665 | CGM.CreateGlobalInitOrDestructFunction(FTy, Name, FI, Loc); | |||
2666 | auto NL = ApplyDebugLocation::CreateEmpty(DtorCGF); | |||
2667 | DtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI, Args, | |||
2668 | Loc, Loc); | |||
2669 | // Create a scope with an artificial location for the body of this function. | |||
2670 | auto AL = ApplyDebugLocation::CreateArtificial(DtorCGF); | |||
2671 | llvm::Value *ArgVal = DtorCGF.EmitLoadOfScalar( | |||
2672 | DtorCGF.GetAddrOfLocalVar(&Dst), | |||
2673 | /*Volatile=*/false, CGM.getContext().VoidPtrTy, Dst.getLocation()); | |||
2674 | DtorCGF.emitDestroy(Address(ArgVal, VDAddr.getAlignment()), ASTTy, | |||
2675 | DtorCGF.getDestroyer(ASTTy.isDestructedType()), | |||
2676 | DtorCGF.needsEHCleanup(ASTTy.isDestructedType())); | |||
2677 | DtorCGF.FinishFunction(); | |||
2678 | Dtor = Fn; | |||
2679 | } | |||
2680 | // Do not emit init function if it is not required. | |||
2681 | if (!Ctor && !Dtor) | |||
2682 | return nullptr; | |||
2683 | ||||
2684 | llvm::Type *CopyCtorTyArgs[] = {CGM.VoidPtrTy, CGM.VoidPtrTy}; | |||
2685 | auto *CopyCtorTy = llvm::FunctionType::get(CGM.VoidPtrTy, CopyCtorTyArgs, | |||
2686 | /*isVarArg=*/false) | |||
2687 | ->getPointerTo(); | |||
2688 | // Copying constructor for the threadprivate variable. | |||
2689 | // Must be NULL - reserved by runtime, but currently it requires that this | |||
2690 | // parameter is always NULL. Otherwise it fires assertion. | |||
2691 | CopyCtor = llvm::Constant::getNullValue(CopyCtorTy); | |||
2692 | if (Ctor == nullptr) { | |||
2693 | auto *CtorTy = llvm::FunctionType::get(CGM.VoidPtrTy, CGM.VoidPtrTy, | |||
2694 | /*isVarArg=*/false) | |||
2695 | ->getPointerTo(); | |||
2696 | Ctor = llvm::Constant::getNullValue(CtorTy); | |||
2697 | } | |||
2698 | if (Dtor == nullptr) { | |||
2699 | auto *DtorTy = llvm::FunctionType::get(CGM.VoidTy, CGM.VoidPtrTy, | |||
2700 | /*isVarArg=*/false) | |||
2701 | ->getPointerTo(); | |||
2702 | Dtor = llvm::Constant::getNullValue(DtorTy); | |||
2703 | } | |||
2704 | if (!CGF) { | |||
2705 | auto *InitFunctionTy = | |||
2706 | llvm::FunctionType::get(CGM.VoidTy, /*isVarArg*/ false); | |||
2707 | std::string Name = getName({"__omp_threadprivate_init_", ""}); | |||
2708 | llvm::Function *InitFunction = CGM.CreateGlobalInitOrDestructFunction( | |||
2709 | InitFunctionTy, Name, CGM.getTypes().arrangeNullaryFunction()); | |||
2710 | CodeGenFunction InitCGF(CGM); | |||
2711 | FunctionArgList ArgList; | |||
2712 | InitCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, InitFunction, | |||
2713 | CGM.getTypes().arrangeNullaryFunction(), ArgList, | |||
2714 | Loc, Loc); | |||
2715 | emitThreadPrivateVarInit(InitCGF, VDAddr, Ctor, CopyCtor, Dtor, Loc); | |||
2716 | InitCGF.FinishFunction(); | |||
2717 | return InitFunction; | |||
2718 | } | |||
2719 | emitThreadPrivateVarInit(*CGF, VDAddr, Ctor, CopyCtor, Dtor, Loc); | |||
2720 | } | |||
2721 | return nullptr; | |||
2722 | } | |||
2723 | ||||
2724 | /// Obtain information that uniquely identifies a target entry. This | |||
2725 | /// consists of the file and device IDs as well as line number associated with | |||
2726 | /// the relevant entry source location. | |||
2727 | static void getTargetEntryUniqueInfo(ASTContext &C, SourceLocation Loc, | |||
2728 | unsigned &DeviceID, unsigned &FileID, | |||
2729 | unsigned &LineNum) { | |||
2730 | SourceManager &SM = C.getSourceManager(); | |||
2731 | ||||
2732 | // The loc should be always valid and have a file ID (the user cannot use | |||
2733 | // #pragma directives in macros) | |||
2734 | ||||
2735 | assert(Loc.isValid() && "Source location is expected to be always valid.")((Loc.isValid() && "Source location is expected to be always valid." ) ? static_cast<void> (0) : __assert_fail ("Loc.isValid() && \"Source location is expected to be always valid.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2735, __PRETTY_FUNCTION__)); | |||
2736 | ||||
2737 | PresumedLoc PLoc = SM.getPresumedLoc(Loc); | |||
2738 | assert(PLoc.isValid() && "Source location is expected to be always valid.")((PLoc.isValid() && "Source location is expected to be always valid." ) ? static_cast<void> (0) : __assert_fail ("PLoc.isValid() && \"Source location is expected to be always valid.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 2738, __PRETTY_FUNCTION__)); | |||
2739 | ||||
2740 | llvm::sys::fs::UniqueID ID; | |||
2741 | if (auto EC = llvm::sys::fs::getUniqueID(PLoc.getFilename(), ID)) | |||
2742 | SM.getDiagnostics().Report(diag::err_cannot_open_file) | |||
2743 | << PLoc.getFilename() << EC.message(); | |||
2744 | ||||
2745 | DeviceID = ID.getDevice(); | |||
2746 | FileID = ID.getFile(); | |||
2747 | LineNum = PLoc.getLine(); | |||
2748 | } | |||
2749 | ||||
2750 | bool CGOpenMPRuntime::emitDeclareTargetVarDefinition(const VarDecl *VD, | |||
2751 | llvm::GlobalVariable *Addr, | |||
2752 | bool PerformInit) { | |||
2753 | Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
2754 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | |||
2755 | if (!Res || *Res == OMPDeclareTargetDeclAttr::MT_Link) | |||
2756 | return CGM.getLangOpts().OpenMPIsDevice; | |||
2757 | VD = VD->getDefinition(CGM.getContext()); | |||
2758 | if (VD && !DeclareTargetWithDefinition.insert(CGM.getMangledName(VD)).second) | |||
2759 | return CGM.getLangOpts().OpenMPIsDevice; | |||
2760 | ||||
2761 | QualType ASTTy = VD->getType(); | |||
2762 | ||||
2763 | SourceLocation Loc = VD->getCanonicalDecl()->getBeginLoc(); | |||
2764 | // Produce the unique prefix to identify the new target regions. We use | |||
2765 | // the source location of the variable declaration which we know to not | |||
2766 | // conflict with any target region. | |||
2767 | unsigned DeviceID; | |||
2768 | unsigned FileID; | |||
2769 | unsigned Line; | |||
2770 | getTargetEntryUniqueInfo(CGM.getContext(), Loc, DeviceID, FileID, Line); | |||
2771 | SmallString<128> Buffer, Out; | |||
2772 | { | |||
2773 | llvm::raw_svector_ostream OS(Buffer); | |||
2774 | OS << "__omp_offloading_" << llvm::format("_%x", DeviceID) | |||
2775 | << llvm::format("_%x_", FileID) << VD->getName() << "_l" << Line; | |||
2776 | } | |||
2777 | ||||
2778 | const Expr *Init = VD->getAnyInitializer(); | |||
2779 | if (CGM.getLangOpts().CPlusPlus && PerformInit) { | |||
2780 | llvm::Constant *Ctor; | |||
2781 | llvm::Constant *ID; | |||
2782 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
2783 | // Generate function that re-emits the declaration's initializer into | |||
2784 | // the threadprivate copy of the variable VD | |||
2785 | CodeGenFunction CtorCGF(CGM); | |||
2786 | ||||
2787 | const CGFunctionInfo &FI = CGM.getTypes().arrangeNullaryFunction(); | |||
2788 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | |||
2789 | llvm::Function *Fn = CGM.CreateGlobalInitOrDestructFunction( | |||
2790 | FTy, Twine(Buffer, "_ctor"), FI, Loc); | |||
2791 | auto NL = ApplyDebugLocation::CreateEmpty(CtorCGF); | |||
2792 | CtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI, | |||
2793 | FunctionArgList(), Loc, Loc); | |||
2794 | auto AL = ApplyDebugLocation::CreateArtificial(CtorCGF); | |||
2795 | CtorCGF.EmitAnyExprToMem(Init, | |||
2796 | Address(Addr, CGM.getContext().getDeclAlign(VD)), | |||
2797 | Init->getType().getQualifiers(), | |||
2798 | /*IsInitializer=*/true); | |||
2799 | CtorCGF.FinishFunction(); | |||
2800 | Ctor = Fn; | |||
2801 | ID = llvm::ConstantExpr::getBitCast(Fn, CGM.Int8PtrTy); | |||
2802 | CGM.addUsedGlobal(cast<llvm::GlobalValue>(Ctor)); | |||
2803 | } else { | |||
2804 | Ctor = new llvm::GlobalVariable( | |||
2805 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, | |||
2806 | llvm::GlobalValue::PrivateLinkage, | |||
2807 | llvm::Constant::getNullValue(CGM.Int8Ty), Twine(Buffer, "_ctor")); | |||
2808 | ID = Ctor; | |||
2809 | } | |||
2810 | ||||
2811 | // Register the information for the entry associated with the constructor. | |||
2812 | Out.clear(); | |||
2813 | OffloadEntriesInfoManager.registerTargetRegionEntryInfo( | |||
2814 | DeviceID, FileID, Twine(Buffer, "_ctor").toStringRef(Out), Line, Ctor, | |||
2815 | ID, OffloadEntriesInfoManagerTy::OMPTargetRegionEntryCtor); | |||
2816 | } | |||
2817 | if (VD->getType().isDestructedType() != QualType::DK_none) { | |||
2818 | llvm::Constant *Dtor; | |||
2819 | llvm::Constant *ID; | |||
2820 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
2821 | // Generate function that emits destructor call for the threadprivate | |||
2822 | // copy of the variable VD | |||
2823 | CodeGenFunction DtorCGF(CGM); | |||
2824 | ||||
2825 | const CGFunctionInfo &FI = CGM.getTypes().arrangeNullaryFunction(); | |||
2826 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | |||
2827 | llvm::Function *Fn = CGM.CreateGlobalInitOrDestructFunction( | |||
2828 | FTy, Twine(Buffer, "_dtor"), FI, Loc); | |||
2829 | auto NL = ApplyDebugLocation::CreateEmpty(DtorCGF); | |||
2830 | DtorCGF.StartFunction(GlobalDecl(), CGM.getContext().VoidTy, Fn, FI, | |||
2831 | FunctionArgList(), Loc, Loc); | |||
2832 | // Create a scope with an artificial location for the body of this | |||
2833 | // function. | |||
2834 | auto AL = ApplyDebugLocation::CreateArtificial(DtorCGF); | |||
2835 | DtorCGF.emitDestroy(Address(Addr, CGM.getContext().getDeclAlign(VD)), | |||
2836 | ASTTy, DtorCGF.getDestroyer(ASTTy.isDestructedType()), | |||
2837 | DtorCGF.needsEHCleanup(ASTTy.isDestructedType())); | |||
2838 | DtorCGF.FinishFunction(); | |||
2839 | Dtor = Fn; | |||
2840 | ID = llvm::ConstantExpr::getBitCast(Fn, CGM.Int8PtrTy); | |||
2841 | CGM.addUsedGlobal(cast<llvm::GlobalValue>(Dtor)); | |||
2842 | } else { | |||
2843 | Dtor = new llvm::GlobalVariable( | |||
2844 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, | |||
2845 | llvm::GlobalValue::PrivateLinkage, | |||
2846 | llvm::Constant::getNullValue(CGM.Int8Ty), Twine(Buffer, "_dtor")); | |||
2847 | ID = Dtor; | |||
2848 | } | |||
2849 | // Register the information for the entry associated with the destructor. | |||
2850 | Out.clear(); | |||
2851 | OffloadEntriesInfoManager.registerTargetRegionEntryInfo( | |||
2852 | DeviceID, FileID, Twine(Buffer, "_dtor").toStringRef(Out), Line, Dtor, | |||
2853 | ID, OffloadEntriesInfoManagerTy::OMPTargetRegionEntryDtor); | |||
2854 | } | |||
2855 | return CGM.getLangOpts().OpenMPIsDevice; | |||
2856 | } | |||
2857 | ||||
2858 | Address CGOpenMPRuntime::getAddrOfArtificialThreadPrivate(CodeGenFunction &CGF, | |||
2859 | QualType VarType, | |||
2860 | StringRef Name) { | |||
2861 | std::string Suffix = getName({"artificial", ""}); | |||
2862 | std::string CacheSuffix = getName({"cache", ""}); | |||
2863 | llvm::Type *VarLVType = CGF.ConvertTypeForMem(VarType); | |||
2864 | llvm::Value *GAddr = | |||
2865 | getOrCreateInternalVariable(VarLVType, Twine(Name).concat(Suffix)); | |||
2866 | llvm::Value *Args[] = { | |||
2867 | emitUpdateLocation(CGF, SourceLocation()), | |||
2868 | getThreadID(CGF, SourceLocation()), | |||
2869 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(GAddr, CGM.VoidPtrTy), | |||
2870 | CGF.Builder.CreateIntCast(CGF.getTypeSize(VarType), CGM.SizeTy, | |||
2871 | /*IsSigned=*/false), | |||
2872 | getOrCreateInternalVariable( | |||
2873 | CGM.VoidPtrPtrTy, Twine(Name).concat(Suffix).concat(CacheSuffix))}; | |||
2874 | return Address( | |||
2875 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
2876 | CGF.EmitRuntimeCall( | |||
2877 | createRuntimeFunction(OMPRTL__kmpc_threadprivate_cached), Args), | |||
2878 | VarLVType->getPointerTo(/*AddrSpace=*/0)), | |||
2879 | CGM.getPointerAlign()); | |||
2880 | } | |||
2881 | ||||
2882 | void CGOpenMPRuntime::emitOMPIfClause(CodeGenFunction &CGF, const Expr *Cond, | |||
2883 | const RegionCodeGenTy &ThenGen, | |||
2884 | const RegionCodeGenTy &ElseGen) { | |||
2885 | CodeGenFunction::LexicalScope ConditionScope(CGF, Cond->getSourceRange()); | |||
2886 | ||||
2887 | // If the condition constant folds and can be elided, try to avoid emitting | |||
2888 | // the condition and the dead arm of the if/else. | |||
2889 | bool CondConstant; | |||
2890 | if (CGF.ConstantFoldsToSimpleInteger(Cond, CondConstant)) { | |||
2891 | if (CondConstant) | |||
2892 | ThenGen(CGF); | |||
2893 | else | |||
2894 | ElseGen(CGF); | |||
2895 | return; | |||
2896 | } | |||
2897 | ||||
2898 | // Otherwise, the condition did not fold, or we couldn't elide it. Just | |||
2899 | // emit the conditional branch. | |||
2900 | llvm::BasicBlock *ThenBlock = CGF.createBasicBlock("omp_if.then"); | |||
2901 | llvm::BasicBlock *ElseBlock = CGF.createBasicBlock("omp_if.else"); | |||
2902 | llvm::BasicBlock *ContBlock = CGF.createBasicBlock("omp_if.end"); | |||
2903 | CGF.EmitBranchOnBoolExpr(Cond, ThenBlock, ElseBlock, /*TrueCount=*/0); | |||
2904 | ||||
2905 | // Emit the 'then' code. | |||
2906 | CGF.EmitBlock(ThenBlock); | |||
2907 | ThenGen(CGF); | |||
2908 | CGF.EmitBranch(ContBlock); | |||
2909 | // Emit the 'else' code if present. | |||
2910 | // There is no need to emit line number for unconditional branch. | |||
2911 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
2912 | CGF.EmitBlock(ElseBlock); | |||
2913 | ElseGen(CGF); | |||
2914 | // There is no need to emit line number for unconditional branch. | |||
2915 | (void)ApplyDebugLocation::CreateEmpty(CGF); | |||
2916 | CGF.EmitBranch(ContBlock); | |||
2917 | // Emit the continuation block for code after the if. | |||
2918 | CGF.EmitBlock(ContBlock, /*IsFinished=*/true); | |||
2919 | } | |||
2920 | ||||
2921 | void CGOpenMPRuntime::emitParallelCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
2922 | llvm::Function *OutlinedFn, | |||
2923 | ArrayRef<llvm::Value *> CapturedVars, | |||
2924 | const Expr *IfCond) { | |||
2925 | if (!CGF.HaveInsertPoint()) | |||
2926 | return; | |||
2927 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | |||
2928 | auto &&ThenGen = [OutlinedFn, CapturedVars, RTLoc](CodeGenFunction &CGF, | |||
2929 | PrePostActionTy &) { | |||
2930 | // Build call __kmpc_fork_call(loc, n, microtask, var1, .., varn); | |||
2931 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | |||
2932 | llvm::Value *Args[] = { | |||
2933 | RTLoc, | |||
2934 | CGF.Builder.getInt32(CapturedVars.size()), // Number of captured vars | |||
2935 | CGF.Builder.CreateBitCast(OutlinedFn, RT.getKmpc_MicroPointerTy())}; | |||
2936 | llvm::SmallVector<llvm::Value *, 16> RealArgs; | |||
2937 | RealArgs.append(std::begin(Args), std::end(Args)); | |||
2938 | RealArgs.append(CapturedVars.begin(), CapturedVars.end()); | |||
2939 | ||||
2940 | llvm::FunctionCallee RTLFn = | |||
2941 | RT.createRuntimeFunction(OMPRTL__kmpc_fork_call); | |||
2942 | CGF.EmitRuntimeCall(RTLFn, RealArgs); | |||
2943 | }; | |||
2944 | auto &&ElseGen = [OutlinedFn, CapturedVars, RTLoc, Loc](CodeGenFunction &CGF, | |||
2945 | PrePostActionTy &) { | |||
2946 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | |||
2947 | llvm::Value *ThreadID = RT.getThreadID(CGF, Loc); | |||
2948 | // Build calls: | |||
2949 | // __kmpc_serialized_parallel(&Loc, GTid); | |||
2950 | llvm::Value *Args[] = {RTLoc, ThreadID}; | |||
2951 | CGF.EmitRuntimeCall( | |||
2952 | RT.createRuntimeFunction(OMPRTL__kmpc_serialized_parallel), Args); | |||
2953 | ||||
2954 | // OutlinedFn(>id, &zero, CapturedStruct); | |||
2955 | Address ZeroAddr = CGF.CreateDefaultAlignTempAlloca(CGF.Int32Ty, | |||
2956 | /*Name*/ ".zero.addr"); | |||
2957 | CGF.InitTempAlloca(ZeroAddr, CGF.Builder.getInt32(/*C*/ 0)); | |||
2958 | llvm::SmallVector<llvm::Value *, 16> OutlinedFnArgs; | |||
2959 | // ThreadId for serialized parallels is 0. | |||
2960 | OutlinedFnArgs.push_back(ZeroAddr.getPointer()); | |||
2961 | OutlinedFnArgs.push_back(ZeroAddr.getPointer()); | |||
2962 | OutlinedFnArgs.append(CapturedVars.begin(), CapturedVars.end()); | |||
2963 | RT.emitOutlinedFunctionCall(CGF, Loc, OutlinedFn, OutlinedFnArgs); | |||
2964 | ||||
2965 | // __kmpc_end_serialized_parallel(&Loc, GTid); | |||
2966 | llvm::Value *EndArgs[] = {RT.emitUpdateLocation(CGF, Loc), ThreadID}; | |||
2967 | CGF.EmitRuntimeCall( | |||
2968 | RT.createRuntimeFunction(OMPRTL__kmpc_end_serialized_parallel), | |||
2969 | EndArgs); | |||
2970 | }; | |||
2971 | if (IfCond) { | |||
2972 | emitOMPIfClause(CGF, IfCond, ThenGen, ElseGen); | |||
2973 | } else { | |||
2974 | RegionCodeGenTy ThenRCG(ThenGen); | |||
2975 | ThenRCG(CGF); | |||
2976 | } | |||
2977 | } | |||
2978 | ||||
2979 | // If we're inside an (outlined) parallel region, use the region info's | |||
2980 | // thread-ID variable (it is passed in a first argument of the outlined function | |||
2981 | // as "kmp_int32 *gtid"). Otherwise, if we're not inside parallel region, but in | |||
2982 | // regular serial code region, get thread ID by calling kmp_int32 | |||
2983 | // kmpc_global_thread_num(ident_t *loc), stash this thread ID in a temporary and | |||
2984 | // return the address of that temp. | |||
2985 | Address CGOpenMPRuntime::emitThreadIDAddress(CodeGenFunction &CGF, | |||
2986 | SourceLocation Loc) { | |||
2987 | if (auto *OMPRegionInfo = | |||
2988 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | |||
2989 | if (OMPRegionInfo->getThreadIDVariable()) | |||
2990 | return OMPRegionInfo->getThreadIDVariableLValue(CGF).getAddress(); | |||
2991 | ||||
2992 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
2993 | QualType Int32Ty = | |||
2994 | CGF.getContext().getIntTypeForBitwidth(/*DestWidth*/ 32, /*Signed*/ true); | |||
2995 | Address ThreadIDTemp = CGF.CreateMemTemp(Int32Ty, /*Name*/ ".threadid_temp."); | |||
2996 | CGF.EmitStoreOfScalar(ThreadID, | |||
2997 | CGF.MakeAddrLValue(ThreadIDTemp, Int32Ty)); | |||
2998 | ||||
2999 | return ThreadIDTemp; | |||
3000 | } | |||
3001 | ||||
3002 | llvm::Constant *CGOpenMPRuntime::getOrCreateInternalVariable( | |||
3003 | llvm::Type *Ty, const llvm::Twine &Name, unsigned AddressSpace) { | |||
3004 | SmallString<256> Buffer; | |||
3005 | llvm::raw_svector_ostream Out(Buffer); | |||
3006 | Out << Name; | |||
3007 | StringRef RuntimeName = Out.str(); | |||
3008 | auto &Elem = *InternalVars.try_emplace(RuntimeName, nullptr).first; | |||
3009 | if (Elem.second) { | |||
3010 | assert(Elem.second->getType()->getPointerElementType() == Ty &&((Elem.second->getType()->getPointerElementType() == Ty && "OMP internal variable has different type than requested" ) ? static_cast<void> (0) : __assert_fail ("Elem.second->getType()->getPointerElementType() == Ty && \"OMP internal variable has different type than requested\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3011, __PRETTY_FUNCTION__)) | |||
3011 | "OMP internal variable has different type than requested")((Elem.second->getType()->getPointerElementType() == Ty && "OMP internal variable has different type than requested" ) ? static_cast<void> (0) : __assert_fail ("Elem.second->getType()->getPointerElementType() == Ty && \"OMP internal variable has different type than requested\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3011, __PRETTY_FUNCTION__)); | |||
3012 | return &*Elem.second; | |||
3013 | } | |||
3014 | ||||
3015 | return Elem.second = new llvm::GlobalVariable( | |||
3016 | CGM.getModule(), Ty, /*IsConstant*/ false, | |||
3017 | llvm::GlobalValue::CommonLinkage, llvm::Constant::getNullValue(Ty), | |||
3018 | Elem.first(), /*InsertBefore=*/nullptr, | |||
3019 | llvm::GlobalValue::NotThreadLocal, AddressSpace); | |||
3020 | } | |||
3021 | ||||
3022 | llvm::Value *CGOpenMPRuntime::getCriticalRegionLock(StringRef CriticalName) { | |||
3023 | std::string Prefix = Twine("gomp_critical_user_", CriticalName).str(); | |||
3024 | std::string Name = getName({Prefix, "var"}); | |||
3025 | return getOrCreateInternalVariable(KmpCriticalNameTy, Name); | |||
3026 | } | |||
3027 | ||||
3028 | namespace { | |||
3029 | /// Common pre(post)-action for different OpenMP constructs. | |||
3030 | class CommonActionTy final : public PrePostActionTy { | |||
3031 | llvm::FunctionCallee EnterCallee; | |||
3032 | ArrayRef<llvm::Value *> EnterArgs; | |||
3033 | llvm::FunctionCallee ExitCallee; | |||
3034 | ArrayRef<llvm::Value *> ExitArgs; | |||
3035 | bool Conditional; | |||
3036 | llvm::BasicBlock *ContBlock = nullptr; | |||
3037 | ||||
3038 | public: | |||
3039 | CommonActionTy(llvm::FunctionCallee EnterCallee, | |||
3040 | ArrayRef<llvm::Value *> EnterArgs, | |||
3041 | llvm::FunctionCallee ExitCallee, | |||
3042 | ArrayRef<llvm::Value *> ExitArgs, bool Conditional = false) | |||
3043 | : EnterCallee(EnterCallee), EnterArgs(EnterArgs), ExitCallee(ExitCallee), | |||
3044 | ExitArgs(ExitArgs), Conditional(Conditional) {} | |||
3045 | void Enter(CodeGenFunction &CGF) override { | |||
3046 | llvm::Value *EnterRes = CGF.EmitRuntimeCall(EnterCallee, EnterArgs); | |||
3047 | if (Conditional) { | |||
3048 | llvm::Value *CallBool = CGF.Builder.CreateIsNotNull(EnterRes); | |||
3049 | auto *ThenBlock = CGF.createBasicBlock("omp_if.then"); | |||
3050 | ContBlock = CGF.createBasicBlock("omp_if.end"); | |||
3051 | // Generate the branch (If-stmt) | |||
3052 | CGF.Builder.CreateCondBr(CallBool, ThenBlock, ContBlock); | |||
3053 | CGF.EmitBlock(ThenBlock); | |||
3054 | } | |||
3055 | } | |||
3056 | void Done(CodeGenFunction &CGF) { | |||
3057 | // Emit the rest of blocks/branches | |||
3058 | CGF.EmitBranch(ContBlock); | |||
3059 | CGF.EmitBlock(ContBlock, true); | |||
3060 | } | |||
3061 | void Exit(CodeGenFunction &CGF) override { | |||
3062 | CGF.EmitRuntimeCall(ExitCallee, ExitArgs); | |||
3063 | } | |||
3064 | }; | |||
3065 | } // anonymous namespace | |||
3066 | ||||
3067 | void CGOpenMPRuntime::emitCriticalRegion(CodeGenFunction &CGF, | |||
3068 | StringRef CriticalName, | |||
3069 | const RegionCodeGenTy &CriticalOpGen, | |||
3070 | SourceLocation Loc, const Expr *Hint) { | |||
3071 | // __kmpc_critical[_with_hint](ident_t *, gtid, Lock[, hint]); | |||
3072 | // CriticalOpGen(); | |||
3073 | // __kmpc_end_critical(ident_t *, gtid, Lock); | |||
3074 | // Prepare arguments and build a call to __kmpc_critical | |||
3075 | if (!CGF.HaveInsertPoint()) | |||
3076 | return; | |||
3077 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
3078 | getCriticalRegionLock(CriticalName)}; | |||
3079 | llvm::SmallVector<llvm::Value *, 4> EnterArgs(std::begin(Args), | |||
3080 | std::end(Args)); | |||
3081 | if (Hint) { | |||
3082 | EnterArgs.push_back(CGF.Builder.CreateIntCast( | |||
3083 | CGF.EmitScalarExpr(Hint), CGM.IntPtrTy, /*isSigned=*/false)); | |||
3084 | } | |||
3085 | CommonActionTy Action( | |||
3086 | createRuntimeFunction(Hint ? OMPRTL__kmpc_critical_with_hint | |||
3087 | : OMPRTL__kmpc_critical), | |||
3088 | EnterArgs, createRuntimeFunction(OMPRTL__kmpc_end_critical), Args); | |||
3089 | CriticalOpGen.setAction(Action); | |||
3090 | emitInlinedDirective(CGF, OMPD_critical, CriticalOpGen); | |||
3091 | } | |||
3092 | ||||
3093 | void CGOpenMPRuntime::emitMasterRegion(CodeGenFunction &CGF, | |||
3094 | const RegionCodeGenTy &MasterOpGen, | |||
3095 | SourceLocation Loc) { | |||
3096 | if (!CGF.HaveInsertPoint()) | |||
3097 | return; | |||
3098 | // if(__kmpc_master(ident_t *, gtid)) { | |||
3099 | // MasterOpGen(); | |||
3100 | // __kmpc_end_master(ident_t *, gtid); | |||
3101 | // } | |||
3102 | // Prepare arguments and build a call to __kmpc_master | |||
3103 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | |||
3104 | CommonActionTy Action(createRuntimeFunction(OMPRTL__kmpc_master), Args, | |||
3105 | createRuntimeFunction(OMPRTL__kmpc_end_master), Args, | |||
3106 | /*Conditional=*/true); | |||
3107 | MasterOpGen.setAction(Action); | |||
3108 | emitInlinedDirective(CGF, OMPD_master, MasterOpGen); | |||
3109 | Action.Done(CGF); | |||
3110 | } | |||
3111 | ||||
3112 | void CGOpenMPRuntime::emitTaskyieldCall(CodeGenFunction &CGF, | |||
3113 | SourceLocation Loc) { | |||
3114 | if (!CGF.HaveInsertPoint()) | |||
3115 | return; | |||
3116 | // Build call __kmpc_omp_taskyield(loc, thread_id, 0); | |||
3117 | llvm::Value *Args[] = { | |||
3118 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
3119 | llvm::ConstantInt::get(CGM.IntTy, /*V=*/0, /*isSigned=*/true)}; | |||
3120 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_omp_taskyield), Args); | |||
3121 | if (auto *Region = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | |||
3122 | Region->emitUntiedSwitch(CGF); | |||
3123 | } | |||
3124 | ||||
3125 | void CGOpenMPRuntime::emitTaskgroupRegion(CodeGenFunction &CGF, | |||
3126 | const RegionCodeGenTy &TaskgroupOpGen, | |||
3127 | SourceLocation Loc) { | |||
3128 | if (!CGF.HaveInsertPoint()) | |||
3129 | return; | |||
3130 | // __kmpc_taskgroup(ident_t *, gtid); | |||
3131 | // TaskgroupOpGen(); | |||
3132 | // __kmpc_end_taskgroup(ident_t *, gtid); | |||
3133 | // Prepare arguments and build a call to __kmpc_taskgroup | |||
3134 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | |||
3135 | CommonActionTy Action(createRuntimeFunction(OMPRTL__kmpc_taskgroup), Args, | |||
3136 | createRuntimeFunction(OMPRTL__kmpc_end_taskgroup), | |||
3137 | Args); | |||
3138 | TaskgroupOpGen.setAction(Action); | |||
3139 | emitInlinedDirective(CGF, OMPD_taskgroup, TaskgroupOpGen); | |||
3140 | } | |||
3141 | ||||
3142 | /// Given an array of pointers to variables, project the address of a | |||
3143 | /// given variable. | |||
3144 | static Address emitAddrOfVarFromArray(CodeGenFunction &CGF, Address Array, | |||
3145 | unsigned Index, const VarDecl *Var) { | |||
3146 | // Pull out the pointer to the variable. | |||
3147 | Address PtrAddr = CGF.Builder.CreateConstArrayGEP(Array, Index); | |||
3148 | llvm::Value *Ptr = CGF.Builder.CreateLoad(PtrAddr); | |||
3149 | ||||
3150 | Address Addr = Address(Ptr, CGF.getContext().getDeclAlign(Var)); | |||
3151 | Addr = CGF.Builder.CreateElementBitCast( | |||
3152 | Addr, CGF.ConvertTypeForMem(Var->getType())); | |||
3153 | return Addr; | |||
3154 | } | |||
3155 | ||||
3156 | static llvm::Value *emitCopyprivateCopyFunction( | |||
3157 | CodeGenModule &CGM, llvm::Type *ArgsType, | |||
3158 | ArrayRef<const Expr *> CopyprivateVars, ArrayRef<const Expr *> DestExprs, | |||
3159 | ArrayRef<const Expr *> SrcExprs, ArrayRef<const Expr *> AssignmentOps, | |||
3160 | SourceLocation Loc) { | |||
3161 | ASTContext &C = CGM.getContext(); | |||
3162 | // void copy_func(void *LHSArg, void *RHSArg); | |||
3163 | FunctionArgList Args; | |||
3164 | ImplicitParamDecl LHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
3165 | ImplicitParamDecl::Other); | |||
3166 | ImplicitParamDecl RHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
3167 | ImplicitParamDecl::Other); | |||
3168 | Args.push_back(&LHSArg); | |||
3169 | Args.push_back(&RHSArg); | |||
3170 | const auto &CGFI = | |||
3171 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
3172 | std::string Name = | |||
3173 | CGM.getOpenMPRuntime().getName({"omp", "copyprivate", "copy_func"}); | |||
3174 | auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI), | |||
3175 | llvm::GlobalValue::InternalLinkage, Name, | |||
3176 | &CGM.getModule()); | |||
3177 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | |||
3178 | Fn->setDoesNotRecurse(); | |||
3179 | CodeGenFunction CGF(CGM); | |||
3180 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); | |||
3181 | // Dest = (void*[n])(LHSArg); | |||
3182 | // Src = (void*[n])(RHSArg); | |||
3183 | Address LHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
3184 | CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&LHSArg)), | |||
3185 | ArgsType), CGF.getPointerAlign()); | |||
3186 | Address RHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
3187 | CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&RHSArg)), | |||
3188 | ArgsType), CGF.getPointerAlign()); | |||
3189 | // *(Type0*)Dst[0] = *(Type0*)Src[0]; | |||
3190 | // *(Type1*)Dst[1] = *(Type1*)Src[1]; | |||
3191 | // ... | |||
3192 | // *(Typen*)Dst[n] = *(Typen*)Src[n]; | |||
3193 | for (unsigned I = 0, E = AssignmentOps.size(); I < E; ++I) { | |||
3194 | const auto *DestVar = | |||
3195 | cast<VarDecl>(cast<DeclRefExpr>(DestExprs[I])->getDecl()); | |||
3196 | Address DestAddr = emitAddrOfVarFromArray(CGF, LHS, I, DestVar); | |||
3197 | ||||
3198 | const auto *SrcVar = | |||
3199 | cast<VarDecl>(cast<DeclRefExpr>(SrcExprs[I])->getDecl()); | |||
3200 | Address SrcAddr = emitAddrOfVarFromArray(CGF, RHS, I, SrcVar); | |||
3201 | ||||
3202 | const auto *VD = cast<DeclRefExpr>(CopyprivateVars[I])->getDecl(); | |||
3203 | QualType Type = VD->getType(); | |||
3204 | CGF.EmitOMPCopy(Type, DestAddr, SrcAddr, DestVar, SrcVar, AssignmentOps[I]); | |||
3205 | } | |||
3206 | CGF.FinishFunction(); | |||
3207 | return Fn; | |||
3208 | } | |||
3209 | ||||
3210 | void CGOpenMPRuntime::emitSingleRegion(CodeGenFunction &CGF, | |||
3211 | const RegionCodeGenTy &SingleOpGen, | |||
3212 | SourceLocation Loc, | |||
3213 | ArrayRef<const Expr *> CopyprivateVars, | |||
3214 | ArrayRef<const Expr *> SrcExprs, | |||
3215 | ArrayRef<const Expr *> DstExprs, | |||
3216 | ArrayRef<const Expr *> AssignmentOps) { | |||
3217 | if (!CGF.HaveInsertPoint()) | |||
3218 | return; | |||
3219 | assert(CopyprivateVars.size() == SrcExprs.size() &&((CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars .size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()) ? static_cast<void> (0) : __assert_fail ("CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars.size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3221, __PRETTY_FUNCTION__)) | |||
3220 | CopyprivateVars.size() == DstExprs.size() &&((CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars .size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()) ? static_cast<void> (0) : __assert_fail ("CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars.size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3221, __PRETTY_FUNCTION__)) | |||
3221 | CopyprivateVars.size() == AssignmentOps.size())((CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars .size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()) ? static_cast<void> (0) : __assert_fail ("CopyprivateVars.size() == SrcExprs.size() && CopyprivateVars.size() == DstExprs.size() && CopyprivateVars.size() == AssignmentOps.size()" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3221, __PRETTY_FUNCTION__)); | |||
3222 | ASTContext &C = CGM.getContext(); | |||
3223 | // int32 did_it = 0; | |||
3224 | // if(__kmpc_single(ident_t *, gtid)) { | |||
3225 | // SingleOpGen(); | |||
3226 | // __kmpc_end_single(ident_t *, gtid); | |||
3227 | // did_it = 1; | |||
3228 | // } | |||
3229 | // call __kmpc_copyprivate(ident_t *, gtid, <buf_size>, <copyprivate list>, | |||
3230 | // <copy_func>, did_it); | |||
3231 | ||||
3232 | Address DidIt = Address::invalid(); | |||
3233 | if (!CopyprivateVars.empty()) { | |||
3234 | // int32 did_it = 0; | |||
3235 | QualType KmpInt32Ty = | |||
3236 | C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1); | |||
3237 | DidIt = CGF.CreateMemTemp(KmpInt32Ty, ".omp.copyprivate.did_it"); | |||
3238 | CGF.Builder.CreateStore(CGF.Builder.getInt32(0), DidIt); | |||
3239 | } | |||
3240 | // Prepare arguments and build a call to __kmpc_single | |||
3241 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | |||
3242 | CommonActionTy Action(createRuntimeFunction(OMPRTL__kmpc_single), Args, | |||
3243 | createRuntimeFunction(OMPRTL__kmpc_end_single), Args, | |||
3244 | /*Conditional=*/true); | |||
3245 | SingleOpGen.setAction(Action); | |||
3246 | emitInlinedDirective(CGF, OMPD_single, SingleOpGen); | |||
3247 | if (DidIt.isValid()) { | |||
3248 | // did_it = 1; | |||
3249 | CGF.Builder.CreateStore(CGF.Builder.getInt32(1), DidIt); | |||
3250 | } | |||
3251 | Action.Done(CGF); | |||
3252 | // call __kmpc_copyprivate(ident_t *, gtid, <buf_size>, <copyprivate list>, | |||
3253 | // <copy_func>, did_it); | |||
3254 | if (DidIt.isValid()) { | |||
3255 | llvm::APInt ArraySize(/*unsigned int numBits=*/32, CopyprivateVars.size()); | |||
3256 | QualType CopyprivateArrayTy = | |||
3257 | C.getConstantArrayType(C.VoidPtrTy, ArraySize, ArrayType::Normal, | |||
3258 | /*IndexTypeQuals=*/0); | |||
3259 | // Create a list of all private variables for copyprivate. | |||
3260 | Address CopyprivateList = | |||
3261 | CGF.CreateMemTemp(CopyprivateArrayTy, ".omp.copyprivate.cpr_list"); | |||
3262 | for (unsigned I = 0, E = CopyprivateVars.size(); I < E; ++I) { | |||
3263 | Address Elem = CGF.Builder.CreateConstArrayGEP(CopyprivateList, I); | |||
3264 | CGF.Builder.CreateStore( | |||
3265 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
3266 | CGF.EmitLValue(CopyprivateVars[I]).getPointer(), CGF.VoidPtrTy), | |||
3267 | Elem); | |||
3268 | } | |||
3269 | // Build function that copies private values from single region to all other | |||
3270 | // threads in the corresponding parallel region. | |||
3271 | llvm::Value *CpyFn = emitCopyprivateCopyFunction( | |||
3272 | CGM, CGF.ConvertTypeForMem(CopyprivateArrayTy)->getPointerTo(), | |||
3273 | CopyprivateVars, SrcExprs, DstExprs, AssignmentOps, Loc); | |||
3274 | llvm::Value *BufSize = CGF.getTypeSize(CopyprivateArrayTy); | |||
3275 | Address CL = | |||
3276 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(CopyprivateList, | |||
3277 | CGF.VoidPtrTy); | |||
3278 | llvm::Value *DidItVal = CGF.Builder.CreateLoad(DidIt); | |||
3279 | llvm::Value *Args[] = { | |||
3280 | emitUpdateLocation(CGF, Loc), // ident_t *<loc> | |||
3281 | getThreadID(CGF, Loc), // i32 <gtid> | |||
3282 | BufSize, // size_t <buf_size> | |||
3283 | CL.getPointer(), // void *<copyprivate list> | |||
3284 | CpyFn, // void (*) (void *, void *) <copy_func> | |||
3285 | DidItVal // i32 did_it | |||
3286 | }; | |||
3287 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_copyprivate), Args); | |||
3288 | } | |||
3289 | } | |||
3290 | ||||
3291 | void CGOpenMPRuntime::emitOrderedRegion(CodeGenFunction &CGF, | |||
3292 | const RegionCodeGenTy &OrderedOpGen, | |||
3293 | SourceLocation Loc, bool IsThreads) { | |||
3294 | if (!CGF.HaveInsertPoint()) | |||
3295 | return; | |||
3296 | // __kmpc_ordered(ident_t *, gtid); | |||
3297 | // OrderedOpGen(); | |||
3298 | // __kmpc_end_ordered(ident_t *, gtid); | |||
3299 | // Prepare arguments and build a call to __kmpc_ordered | |||
3300 | if (IsThreads) { | |||
3301 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | |||
3302 | CommonActionTy Action(createRuntimeFunction(OMPRTL__kmpc_ordered), Args, | |||
3303 | createRuntimeFunction(OMPRTL__kmpc_end_ordered), | |||
3304 | Args); | |||
3305 | OrderedOpGen.setAction(Action); | |||
3306 | emitInlinedDirective(CGF, OMPD_ordered, OrderedOpGen); | |||
3307 | return; | |||
3308 | } | |||
3309 | emitInlinedDirective(CGF, OMPD_ordered, OrderedOpGen); | |||
3310 | } | |||
3311 | ||||
3312 | unsigned CGOpenMPRuntime::getDefaultFlagsForBarriers(OpenMPDirectiveKind Kind) { | |||
3313 | unsigned Flags; | |||
3314 | if (Kind == OMPD_for) | |||
3315 | Flags = OMP_IDENT_BARRIER_IMPL_FOR; | |||
3316 | else if (Kind == OMPD_sections) | |||
3317 | Flags = OMP_IDENT_BARRIER_IMPL_SECTIONS; | |||
3318 | else if (Kind == OMPD_single) | |||
3319 | Flags = OMP_IDENT_BARRIER_IMPL_SINGLE; | |||
3320 | else if (Kind == OMPD_barrier) | |||
3321 | Flags = OMP_IDENT_BARRIER_EXPL; | |||
3322 | else | |||
3323 | Flags = OMP_IDENT_BARRIER_IMPL; | |||
3324 | return Flags; | |||
3325 | } | |||
3326 | ||||
3327 | void CGOpenMPRuntime::getDefaultScheduleAndChunk( | |||
3328 | CodeGenFunction &CGF, const OMPLoopDirective &S, | |||
3329 | OpenMPScheduleClauseKind &ScheduleKind, const Expr *&ChunkExpr) const { | |||
3330 | // Check if the loop directive is actually a doacross loop directive. In this | |||
3331 | // case choose static, 1 schedule. | |||
3332 | if (llvm::any_of( | |||
3333 | S.getClausesOfKind<OMPOrderedClause>(), | |||
3334 | [](const OMPOrderedClause *C) { return C->getNumForLoops(); })) { | |||
3335 | ScheduleKind = OMPC_SCHEDULE_static; | |||
3336 | // Chunk size is 1 in this case. | |||
3337 | llvm::APInt ChunkSize(32, 1); | |||
3338 | ChunkExpr = IntegerLiteral::Create( | |||
3339 | CGF.getContext(), ChunkSize, | |||
3340 | CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/0), | |||
3341 | SourceLocation()); | |||
3342 | } | |||
3343 | } | |||
3344 | ||||
3345 | void CGOpenMPRuntime::emitBarrierCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
3346 | OpenMPDirectiveKind Kind, bool EmitChecks, | |||
3347 | bool ForceSimpleCall) { | |||
3348 | if (!CGF.HaveInsertPoint()) | |||
3349 | return; | |||
3350 | // Build call __kmpc_cancel_barrier(loc, thread_id); | |||
3351 | // Build call __kmpc_barrier(loc, thread_id); | |||
3352 | unsigned Flags = getDefaultFlagsForBarriers(Kind); | |||
3353 | // Build call __kmpc_cancel_barrier(loc, thread_id) or __kmpc_barrier(loc, | |||
3354 | // thread_id); | |||
3355 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags), | |||
3356 | getThreadID(CGF, Loc)}; | |||
3357 | if (auto *OMPRegionInfo = | |||
3358 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { | |||
3359 | if (!ForceSimpleCall && OMPRegionInfo->hasCancel()) { | |||
3360 | llvm::Value *Result = CGF.EmitRuntimeCall( | |||
3361 | createRuntimeFunction(OMPRTL__kmpc_cancel_barrier), Args); | |||
3362 | if (EmitChecks) { | |||
3363 | // if (__kmpc_cancel_barrier()) { | |||
3364 | // exit from construct; | |||
3365 | // } | |||
3366 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".cancel.exit"); | |||
3367 | llvm::BasicBlock *ContBB = CGF.createBasicBlock(".cancel.continue"); | |||
3368 | llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Result); | |||
3369 | CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB); | |||
3370 | CGF.EmitBlock(ExitBB); | |||
3371 | // exit from construct; | |||
3372 | CodeGenFunction::JumpDest CancelDestination = | |||
3373 | CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind()); | |||
3374 | CGF.EmitBranchThroughCleanup(CancelDestination); | |||
3375 | CGF.EmitBlock(ContBB, /*IsFinished=*/true); | |||
3376 | } | |||
3377 | return; | |||
3378 | } | |||
3379 | } | |||
3380 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_barrier), Args); | |||
3381 | } | |||
3382 | ||||
3383 | /// Map the OpenMP loop schedule to the runtime enumeration. | |||
3384 | static OpenMPSchedType getRuntimeSchedule(OpenMPScheduleClauseKind ScheduleKind, | |||
3385 | bool Chunked, bool Ordered) { | |||
3386 | switch (ScheduleKind) { | |||
3387 | case OMPC_SCHEDULE_static: | |||
3388 | return Chunked ? (Ordered ? OMP_ord_static_chunked : OMP_sch_static_chunked) | |||
3389 | : (Ordered ? OMP_ord_static : OMP_sch_static); | |||
3390 | case OMPC_SCHEDULE_dynamic: | |||
3391 | return Ordered ? OMP_ord_dynamic_chunked : OMP_sch_dynamic_chunked; | |||
3392 | case OMPC_SCHEDULE_guided: | |||
3393 | return Ordered ? OMP_ord_guided_chunked : OMP_sch_guided_chunked; | |||
3394 | case OMPC_SCHEDULE_runtime: | |||
3395 | return Ordered ? OMP_ord_runtime : OMP_sch_runtime; | |||
3396 | case OMPC_SCHEDULE_auto: | |||
3397 | return Ordered ? OMP_ord_auto : OMP_sch_auto; | |||
3398 | case OMPC_SCHEDULE_unknown: | |||
3399 | assert(!Chunked && "chunk was specified but schedule kind not known")((!Chunked && "chunk was specified but schedule kind not known" ) ? static_cast<void> (0) : __assert_fail ("!Chunked && \"chunk was specified but schedule kind not known\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3399, __PRETTY_FUNCTION__)); | |||
3400 | return Ordered ? OMP_ord_static : OMP_sch_static; | |||
3401 | } | |||
3402 | llvm_unreachable("Unexpected runtime schedule")::llvm::llvm_unreachable_internal("Unexpected runtime schedule" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3402); | |||
3403 | } | |||
3404 | ||||
3405 | /// Map the OpenMP distribute schedule to the runtime enumeration. | |||
3406 | static OpenMPSchedType | |||
3407 | getRuntimeSchedule(OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) { | |||
3408 | // only static is allowed for dist_schedule | |||
3409 | return Chunked ? OMP_dist_sch_static_chunked : OMP_dist_sch_static; | |||
3410 | } | |||
3411 | ||||
3412 | bool CGOpenMPRuntime::isStaticNonchunked(OpenMPScheduleClauseKind ScheduleKind, | |||
3413 | bool Chunked) const { | |||
3414 | OpenMPSchedType Schedule = | |||
3415 | getRuntimeSchedule(ScheduleKind, Chunked, /*Ordered=*/false); | |||
3416 | return Schedule == OMP_sch_static; | |||
3417 | } | |||
3418 | ||||
3419 | bool CGOpenMPRuntime::isStaticNonchunked( | |||
3420 | OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) const { | |||
3421 | OpenMPSchedType Schedule = getRuntimeSchedule(ScheduleKind, Chunked); | |||
3422 | return Schedule == OMP_dist_sch_static; | |||
3423 | } | |||
3424 | ||||
3425 | bool CGOpenMPRuntime::isStaticChunked(OpenMPScheduleClauseKind ScheduleKind, | |||
3426 | bool Chunked) const { | |||
3427 | OpenMPSchedType Schedule = | |||
3428 | getRuntimeSchedule(ScheduleKind, Chunked, /*Ordered=*/false); | |||
3429 | return Schedule == OMP_sch_static_chunked; | |||
3430 | } | |||
3431 | ||||
3432 | bool CGOpenMPRuntime::isStaticChunked( | |||
3433 | OpenMPDistScheduleClauseKind ScheduleKind, bool Chunked) const { | |||
3434 | OpenMPSchedType Schedule = getRuntimeSchedule(ScheduleKind, Chunked); | |||
3435 | return Schedule == OMP_dist_sch_static_chunked; | |||
3436 | } | |||
3437 | ||||
3438 | bool CGOpenMPRuntime::isDynamic(OpenMPScheduleClauseKind ScheduleKind) const { | |||
3439 | OpenMPSchedType Schedule = | |||
3440 | getRuntimeSchedule(ScheduleKind, /*Chunked=*/false, /*Ordered=*/false); | |||
3441 | assert(Schedule != OMP_sch_static_chunked && "cannot be chunked here")((Schedule != OMP_sch_static_chunked && "cannot be chunked here" ) ? static_cast<void> (0) : __assert_fail ("Schedule != OMP_sch_static_chunked && \"cannot be chunked here\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3441, __PRETTY_FUNCTION__)); | |||
3442 | return Schedule != OMP_sch_static; | |||
3443 | } | |||
3444 | ||||
3445 | static int addMonoNonMonoModifier(OpenMPSchedType Schedule, | |||
3446 | OpenMPScheduleClauseModifier M1, | |||
3447 | OpenMPScheduleClauseModifier M2) { | |||
3448 | int Modifier = 0; | |||
3449 | switch (M1) { | |||
3450 | case OMPC_SCHEDULE_MODIFIER_monotonic: | |||
3451 | Modifier = OMP_sch_modifier_monotonic; | |||
3452 | break; | |||
3453 | case OMPC_SCHEDULE_MODIFIER_nonmonotonic: | |||
3454 | Modifier = OMP_sch_modifier_nonmonotonic; | |||
3455 | break; | |||
3456 | case OMPC_SCHEDULE_MODIFIER_simd: | |||
3457 | if (Schedule == OMP_sch_static_chunked) | |||
3458 | Schedule = OMP_sch_static_balanced_chunked; | |||
3459 | break; | |||
3460 | case OMPC_SCHEDULE_MODIFIER_last: | |||
3461 | case OMPC_SCHEDULE_MODIFIER_unknown: | |||
3462 | break; | |||
3463 | } | |||
3464 | switch (M2) { | |||
3465 | case OMPC_SCHEDULE_MODIFIER_monotonic: | |||
3466 | Modifier = OMP_sch_modifier_monotonic; | |||
3467 | break; | |||
3468 | case OMPC_SCHEDULE_MODIFIER_nonmonotonic: | |||
3469 | Modifier = OMP_sch_modifier_nonmonotonic; | |||
3470 | break; | |||
3471 | case OMPC_SCHEDULE_MODIFIER_simd: | |||
3472 | if (Schedule == OMP_sch_static_chunked) | |||
3473 | Schedule = OMP_sch_static_balanced_chunked; | |||
3474 | break; | |||
3475 | case OMPC_SCHEDULE_MODIFIER_last: | |||
3476 | case OMPC_SCHEDULE_MODIFIER_unknown: | |||
3477 | break; | |||
3478 | } | |||
3479 | return Schedule | Modifier; | |||
3480 | } | |||
3481 | ||||
3482 | void CGOpenMPRuntime::emitForDispatchInit( | |||
3483 | CodeGenFunction &CGF, SourceLocation Loc, | |||
3484 | const OpenMPScheduleTy &ScheduleKind, unsigned IVSize, bool IVSigned, | |||
3485 | bool Ordered, const DispatchRTInput &DispatchValues) { | |||
3486 | if (!CGF.HaveInsertPoint()) | |||
3487 | return; | |||
3488 | OpenMPSchedType Schedule = getRuntimeSchedule( | |||
3489 | ScheduleKind.Schedule, DispatchValues.Chunk != nullptr, Ordered); | |||
3490 | assert(Ordered ||((Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)) ? static_cast<void> (0) : __assert_fail ("Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3493, __PRETTY_FUNCTION__)) | |||
3491 | (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked &&((Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)) ? static_cast<void> (0) : __assert_fail ("Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3493, __PRETTY_FUNCTION__)) | |||
3492 | Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked &&((Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)) ? static_cast<void> (0) : __assert_fail ("Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3493, __PRETTY_FUNCTION__)) | |||
3493 | Schedule != OMP_sch_static_balanced_chunked))((Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)) ? static_cast<void> (0) : __assert_fail ("Ordered || (Schedule != OMP_sch_static && Schedule != OMP_sch_static_chunked && Schedule != OMP_ord_static && Schedule != OMP_ord_static_chunked && Schedule != OMP_sch_static_balanced_chunked)" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3493, __PRETTY_FUNCTION__)); | |||
3494 | // Call __kmpc_dispatch_init( | |||
3495 | // ident_t *loc, kmp_int32 tid, kmp_int32 schedule, | |||
3496 | // kmp_int[32|64] lower, kmp_int[32|64] upper, | |||
3497 | // kmp_int[32|64] stride, kmp_int[32|64] chunk); | |||
3498 | ||||
3499 | // If the Chunk was not specified in the clause - use default value 1. | |||
3500 | llvm::Value *Chunk = DispatchValues.Chunk ? DispatchValues.Chunk | |||
3501 | : CGF.Builder.getIntN(IVSize, 1); | |||
3502 | llvm::Value *Args[] = { | |||
3503 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
3504 | CGF.Builder.getInt32(addMonoNonMonoModifier( | |||
3505 | Schedule, ScheduleKind.M1, ScheduleKind.M2)), // Schedule type | |||
3506 | DispatchValues.LB, // Lower | |||
3507 | DispatchValues.UB, // Upper | |||
3508 | CGF.Builder.getIntN(IVSize, 1), // Stride | |||
3509 | Chunk // Chunk | |||
3510 | }; | |||
3511 | CGF.EmitRuntimeCall(createDispatchInitFunction(IVSize, IVSigned), Args); | |||
3512 | } | |||
3513 | ||||
3514 | static void emitForStaticInitCall( | |||
3515 | CodeGenFunction &CGF, llvm::Value *UpdateLocation, llvm::Value *ThreadId, | |||
3516 | llvm::FunctionCallee ForStaticInitFunction, OpenMPSchedType Schedule, | |||
3517 | OpenMPScheduleClauseModifier M1, OpenMPScheduleClauseModifier M2, | |||
3518 | const CGOpenMPRuntime::StaticRTInput &Values) { | |||
3519 | if (!CGF.HaveInsertPoint()) | |||
3520 | return; | |||
3521 | ||||
3522 | assert(!Values.Ordered)((!Values.Ordered) ? static_cast<void> (0) : __assert_fail ("!Values.Ordered", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3522, __PRETTY_FUNCTION__)); | |||
3523 | assert(Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked ||((Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked ) ? static_cast<void> (0) : __assert_fail ("Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3527, __PRETTY_FUNCTION__)) | |||
3524 | Schedule == OMP_sch_static_balanced_chunked ||((Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked ) ? static_cast<void> (0) : __assert_fail ("Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3527, __PRETTY_FUNCTION__)) | |||
3525 | Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked ||((Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked ) ? static_cast<void> (0) : __assert_fail ("Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3527, __PRETTY_FUNCTION__)) | |||
3526 | Schedule == OMP_dist_sch_static ||((Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked ) ? static_cast<void> (0) : __assert_fail ("Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3527, __PRETTY_FUNCTION__)) | |||
3527 | Schedule == OMP_dist_sch_static_chunked)((Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked ) ? static_cast<void> (0) : __assert_fail ("Schedule == OMP_sch_static || Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static || Schedule == OMP_dist_sch_static_chunked" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3527, __PRETTY_FUNCTION__)); | |||
3528 | ||||
3529 | // Call __kmpc_for_static_init( | |||
3530 | // ident_t *loc, kmp_int32 tid, kmp_int32 schedtype, | |||
3531 | // kmp_int32 *p_lastiter, kmp_int[32|64] *p_lower, | |||
3532 | // kmp_int[32|64] *p_upper, kmp_int[32|64] *p_stride, | |||
3533 | // kmp_int[32|64] incr, kmp_int[32|64] chunk); | |||
3534 | llvm::Value *Chunk = Values.Chunk; | |||
3535 | if (Chunk == nullptr) { | |||
3536 | assert((Schedule == OMP_sch_static || Schedule == OMP_ord_static ||(((Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && "expected static non-chunked schedule" ) ? static_cast<void> (0) : __assert_fail ("(Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && \"expected static non-chunked schedule\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3538, __PRETTY_FUNCTION__)) | |||
3537 | Schedule == OMP_dist_sch_static) &&(((Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && "expected static non-chunked schedule" ) ? static_cast<void> (0) : __assert_fail ("(Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && \"expected static non-chunked schedule\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3538, __PRETTY_FUNCTION__)) | |||
3538 | "expected static non-chunked schedule")(((Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && "expected static non-chunked schedule" ) ? static_cast<void> (0) : __assert_fail ("(Schedule == OMP_sch_static || Schedule == OMP_ord_static || Schedule == OMP_dist_sch_static) && \"expected static non-chunked schedule\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3538, __PRETTY_FUNCTION__)); | |||
3539 | // If the Chunk was not specified in the clause - use default value 1. | |||
3540 | Chunk = CGF.Builder.getIntN(Values.IVSize, 1); | |||
3541 | } else { | |||
3542 | assert((Schedule == OMP_sch_static_chunked ||(((Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked ) && "expected static chunked schedule") ? static_cast <void> (0) : __assert_fail ("(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked) && \"expected static chunked schedule\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3546, __PRETTY_FUNCTION__)) | |||
3543 | Schedule == OMP_sch_static_balanced_chunked ||(((Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked ) && "expected static chunked schedule") ? static_cast <void> (0) : __assert_fail ("(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked) && \"expected static chunked schedule\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3546, __PRETTY_FUNCTION__)) | |||
3544 | Schedule == OMP_ord_static_chunked ||(((Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked ) && "expected static chunked schedule") ? static_cast <void> (0) : __assert_fail ("(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked) && \"expected static chunked schedule\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3546, __PRETTY_FUNCTION__)) | |||
3545 | Schedule == OMP_dist_sch_static_chunked) &&(((Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked ) && "expected static chunked schedule") ? static_cast <void> (0) : __assert_fail ("(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked) && \"expected static chunked schedule\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3546, __PRETTY_FUNCTION__)) | |||
3546 | "expected static chunked schedule")(((Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked ) && "expected static chunked schedule") ? static_cast <void> (0) : __assert_fail ("(Schedule == OMP_sch_static_chunked || Schedule == OMP_sch_static_balanced_chunked || Schedule == OMP_ord_static_chunked || Schedule == OMP_dist_sch_static_chunked) && \"expected static chunked schedule\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3546, __PRETTY_FUNCTION__)); | |||
3547 | } | |||
3548 | llvm::Value *Args[] = { | |||
3549 | UpdateLocation, | |||
3550 | ThreadId, | |||
3551 | CGF.Builder.getInt32(addMonoNonMonoModifier(Schedule, M1, | |||
3552 | M2)), // Schedule type | |||
3553 | Values.IL.getPointer(), // &isLastIter | |||
3554 | Values.LB.getPointer(), // &LB | |||
3555 | Values.UB.getPointer(), // &UB | |||
3556 | Values.ST.getPointer(), // &Stride | |||
3557 | CGF.Builder.getIntN(Values.IVSize, 1), // Incr | |||
3558 | Chunk // Chunk | |||
3559 | }; | |||
3560 | CGF.EmitRuntimeCall(ForStaticInitFunction, Args); | |||
3561 | } | |||
3562 | ||||
3563 | void CGOpenMPRuntime::emitForStaticInit(CodeGenFunction &CGF, | |||
3564 | SourceLocation Loc, | |||
3565 | OpenMPDirectiveKind DKind, | |||
3566 | const OpenMPScheduleTy &ScheduleKind, | |||
3567 | const StaticRTInput &Values) { | |||
3568 | OpenMPSchedType ScheduleNum = getRuntimeSchedule( | |||
3569 | ScheduleKind.Schedule, Values.Chunk != nullptr, Values.Ordered); | |||
3570 | assert(isOpenMPWorksharingDirective(DKind) &&((isOpenMPWorksharingDirective(DKind) && "Expected loop-based or sections-based directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPWorksharingDirective(DKind) && \"Expected loop-based or sections-based directive.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3571, __PRETTY_FUNCTION__)) | |||
3571 | "Expected loop-based or sections-based directive.")((isOpenMPWorksharingDirective(DKind) && "Expected loop-based or sections-based directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPWorksharingDirective(DKind) && \"Expected loop-based or sections-based directive.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3571, __PRETTY_FUNCTION__)); | |||
3572 | llvm::Value *UpdatedLocation = emitUpdateLocation(CGF, Loc, | |||
3573 | isOpenMPLoopDirective(DKind) | |||
3574 | ? OMP_IDENT_WORK_LOOP | |||
3575 | : OMP_IDENT_WORK_SECTIONS); | |||
3576 | llvm::Value *ThreadId = getThreadID(CGF, Loc); | |||
3577 | llvm::FunctionCallee StaticInitFunction = | |||
3578 | createForStaticInitFunction(Values.IVSize, Values.IVSigned); | |||
3579 | emitForStaticInitCall(CGF, UpdatedLocation, ThreadId, StaticInitFunction, | |||
3580 | ScheduleNum, ScheduleKind.M1, ScheduleKind.M2, Values); | |||
3581 | } | |||
3582 | ||||
3583 | void CGOpenMPRuntime::emitDistributeStaticInit( | |||
3584 | CodeGenFunction &CGF, SourceLocation Loc, | |||
3585 | OpenMPDistScheduleClauseKind SchedKind, | |||
3586 | const CGOpenMPRuntime::StaticRTInput &Values) { | |||
3587 | OpenMPSchedType ScheduleNum = | |||
3588 | getRuntimeSchedule(SchedKind, Values.Chunk != nullptr); | |||
3589 | llvm::Value *UpdatedLocation = | |||
3590 | emitUpdateLocation(CGF, Loc, OMP_IDENT_WORK_DISTRIBUTE); | |||
3591 | llvm::Value *ThreadId = getThreadID(CGF, Loc); | |||
3592 | llvm::FunctionCallee StaticInitFunction = | |||
3593 | createForStaticInitFunction(Values.IVSize, Values.IVSigned); | |||
3594 | emitForStaticInitCall(CGF, UpdatedLocation, ThreadId, StaticInitFunction, | |||
3595 | ScheduleNum, OMPC_SCHEDULE_MODIFIER_unknown, | |||
3596 | OMPC_SCHEDULE_MODIFIER_unknown, Values); | |||
3597 | } | |||
3598 | ||||
3599 | void CGOpenMPRuntime::emitForStaticFinish(CodeGenFunction &CGF, | |||
3600 | SourceLocation Loc, | |||
3601 | OpenMPDirectiveKind DKind) { | |||
3602 | if (!CGF.HaveInsertPoint()) | |||
3603 | return; | |||
3604 | // Call __kmpc_for_static_fini(ident_t *loc, kmp_int32 tid); | |||
3605 | llvm::Value *Args[] = { | |||
3606 | emitUpdateLocation(CGF, Loc, | |||
3607 | isOpenMPDistributeDirective(DKind) | |||
3608 | ? OMP_IDENT_WORK_DISTRIBUTE | |||
3609 | : isOpenMPLoopDirective(DKind) | |||
3610 | ? OMP_IDENT_WORK_LOOP | |||
3611 | : OMP_IDENT_WORK_SECTIONS), | |||
3612 | getThreadID(CGF, Loc)}; | |||
3613 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_for_static_fini), | |||
3614 | Args); | |||
3615 | } | |||
3616 | ||||
3617 | void CGOpenMPRuntime::emitForOrderedIterationEnd(CodeGenFunction &CGF, | |||
3618 | SourceLocation Loc, | |||
3619 | unsigned IVSize, | |||
3620 | bool IVSigned) { | |||
3621 | if (!CGF.HaveInsertPoint()) | |||
3622 | return; | |||
3623 | // Call __kmpc_for_dynamic_fini_(4|8)[u](ident_t *loc, kmp_int32 tid); | |||
3624 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | |||
3625 | CGF.EmitRuntimeCall(createDispatchFiniFunction(IVSize, IVSigned), Args); | |||
3626 | } | |||
3627 | ||||
3628 | llvm::Value *CGOpenMPRuntime::emitForNext(CodeGenFunction &CGF, | |||
3629 | SourceLocation Loc, unsigned IVSize, | |||
3630 | bool IVSigned, Address IL, | |||
3631 | Address LB, Address UB, | |||
3632 | Address ST) { | |||
3633 | // Call __kmpc_dispatch_next( | |||
3634 | // ident_t *loc, kmp_int32 tid, kmp_int32 *p_lastiter, | |||
3635 | // kmp_int[32|64] *p_lower, kmp_int[32|64] *p_upper, | |||
3636 | // kmp_int[32|64] *p_stride); | |||
3637 | llvm::Value *Args[] = { | |||
3638 | emitUpdateLocation(CGF, Loc), | |||
3639 | getThreadID(CGF, Loc), | |||
3640 | IL.getPointer(), // &isLastIter | |||
3641 | LB.getPointer(), // &Lower | |||
3642 | UB.getPointer(), // &Upper | |||
3643 | ST.getPointer() // &Stride | |||
3644 | }; | |||
3645 | llvm::Value *Call = | |||
3646 | CGF.EmitRuntimeCall(createDispatchNextFunction(IVSize, IVSigned), Args); | |||
3647 | return CGF.EmitScalarConversion( | |||
3648 | Call, CGF.getContext().getIntTypeForBitwidth(32, /*Signed=*/1), | |||
3649 | CGF.getContext().BoolTy, Loc); | |||
3650 | } | |||
3651 | ||||
3652 | void CGOpenMPRuntime::emitNumThreadsClause(CodeGenFunction &CGF, | |||
3653 | llvm::Value *NumThreads, | |||
3654 | SourceLocation Loc) { | |||
3655 | if (!CGF.HaveInsertPoint()) | |||
3656 | return; | |||
3657 | // Build call __kmpc_push_num_threads(&loc, global_tid, num_threads) | |||
3658 | llvm::Value *Args[] = { | |||
3659 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
3660 | CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty, /*isSigned*/ true)}; | |||
3661 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_push_num_threads), | |||
3662 | Args); | |||
3663 | } | |||
3664 | ||||
3665 | void CGOpenMPRuntime::emitProcBindClause(CodeGenFunction &CGF, | |||
3666 | OpenMPProcBindClauseKind ProcBind, | |||
3667 | SourceLocation Loc) { | |||
3668 | if (!CGF.HaveInsertPoint()) | |||
3669 | return; | |||
3670 | // Constants for proc bind value accepted by the runtime. | |||
3671 | enum ProcBindTy { | |||
3672 | ProcBindFalse = 0, | |||
3673 | ProcBindTrue, | |||
3674 | ProcBindMaster, | |||
3675 | ProcBindClose, | |||
3676 | ProcBindSpread, | |||
3677 | ProcBindIntel, | |||
3678 | ProcBindDefault | |||
3679 | } RuntimeProcBind; | |||
3680 | switch (ProcBind) { | |||
3681 | case OMPC_PROC_BIND_master: | |||
3682 | RuntimeProcBind = ProcBindMaster; | |||
3683 | break; | |||
3684 | case OMPC_PROC_BIND_close: | |||
3685 | RuntimeProcBind = ProcBindClose; | |||
3686 | break; | |||
3687 | case OMPC_PROC_BIND_spread: | |||
3688 | RuntimeProcBind = ProcBindSpread; | |||
3689 | break; | |||
3690 | case OMPC_PROC_BIND_unknown: | |||
3691 | llvm_unreachable("Unsupported proc_bind value.")::llvm::llvm_unreachable_internal("Unsupported proc_bind value." , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3691); | |||
3692 | } | |||
3693 | // Build call __kmpc_push_proc_bind(&loc, global_tid, proc_bind) | |||
3694 | llvm::Value *Args[] = { | |||
3695 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
3696 | llvm::ConstantInt::get(CGM.IntTy, RuntimeProcBind, /*isSigned=*/true)}; | |||
3697 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_push_proc_bind), Args); | |||
3698 | } | |||
3699 | ||||
3700 | void CGOpenMPRuntime::emitFlush(CodeGenFunction &CGF, ArrayRef<const Expr *>, | |||
3701 | SourceLocation Loc) { | |||
3702 | if (!CGF.HaveInsertPoint()) | |||
3703 | return; | |||
3704 | // Build call void __kmpc_flush(ident_t *loc) | |||
3705 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_flush), | |||
3706 | emitUpdateLocation(CGF, Loc)); | |||
3707 | } | |||
3708 | ||||
3709 | namespace { | |||
3710 | /// Indexes of fields for type kmp_task_t. | |||
3711 | enum KmpTaskTFields { | |||
3712 | /// List of shared variables. | |||
3713 | KmpTaskTShareds, | |||
3714 | /// Task routine. | |||
3715 | KmpTaskTRoutine, | |||
3716 | /// Partition id for the untied tasks. | |||
3717 | KmpTaskTPartId, | |||
3718 | /// Function with call of destructors for private variables. | |||
3719 | Data1, | |||
3720 | /// Task priority. | |||
3721 | Data2, | |||
3722 | /// (Taskloops only) Lower bound. | |||
3723 | KmpTaskTLowerBound, | |||
3724 | /// (Taskloops only) Upper bound. | |||
3725 | KmpTaskTUpperBound, | |||
3726 | /// (Taskloops only) Stride. | |||
3727 | KmpTaskTStride, | |||
3728 | /// (Taskloops only) Is last iteration flag. | |||
3729 | KmpTaskTLastIter, | |||
3730 | /// (Taskloops only) Reduction data. | |||
3731 | KmpTaskTReductions, | |||
3732 | }; | |||
3733 | } // anonymous namespace | |||
3734 | ||||
3735 | bool CGOpenMPRuntime::OffloadEntriesInfoManagerTy::empty() const { | |||
3736 | return OffloadEntriesTargetRegion.empty() && | |||
3737 | OffloadEntriesDeviceGlobalVar.empty(); | |||
3738 | } | |||
3739 | ||||
3740 | /// Initialize target region entry. | |||
3741 | void CGOpenMPRuntime::OffloadEntriesInfoManagerTy:: | |||
3742 | initializeTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID, | |||
3743 | StringRef ParentName, unsigned LineNum, | |||
3744 | unsigned Order) { | |||
3745 | assert(CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is "((CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is " "only required for the device " "code generation.") ? static_cast <void> (0) : __assert_fail ("CGM.getLangOpts().OpenMPIsDevice && \"Initialization of entries is \" \"only required for the device \" \"code generation.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3747, __PRETTY_FUNCTION__)) | |||
3746 | "only required for the device "((CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is " "only required for the device " "code generation.") ? static_cast <void> (0) : __assert_fail ("CGM.getLangOpts().OpenMPIsDevice && \"Initialization of entries is \" \"only required for the device \" \"code generation.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3747, __PRETTY_FUNCTION__)) | |||
3747 | "code generation.")((CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is " "only required for the device " "code generation.") ? static_cast <void> (0) : __assert_fail ("CGM.getLangOpts().OpenMPIsDevice && \"Initialization of entries is \" \"only required for the device \" \"code generation.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3747, __PRETTY_FUNCTION__)); | |||
3748 | OffloadEntriesTargetRegion[DeviceID][FileID][ParentName][LineNum] = | |||
3749 | OffloadEntryInfoTargetRegion(Order, /*Addr=*/nullptr, /*ID=*/nullptr, | |||
3750 | OMPTargetRegionEntryTargetRegion); | |||
3751 | ++OffloadingEntriesNum; | |||
3752 | } | |||
3753 | ||||
3754 | void CGOpenMPRuntime::OffloadEntriesInfoManagerTy:: | |||
3755 | registerTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID, | |||
3756 | StringRef ParentName, unsigned LineNum, | |||
3757 | llvm::Constant *Addr, llvm::Constant *ID, | |||
3758 | OMPTargetRegionEntryKind Flags) { | |||
3759 | // If we are emitting code for a target, the entry is already initialized, | |||
3760 | // only has to be registered. | |||
3761 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
3762 | if (!hasTargetRegionEntryInfo(DeviceID, FileID, ParentName, LineNum)) { | |||
3763 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
3764 | DiagnosticsEngine::Error, | |||
3765 | "Unable to find target region on line '%0' in the device code."); | |||
3766 | CGM.getDiags().Report(DiagID) << LineNum; | |||
3767 | return; | |||
3768 | } | |||
3769 | auto &Entry = | |||
3770 | OffloadEntriesTargetRegion[DeviceID][FileID][ParentName][LineNum]; | |||
3771 | assert(Entry.isValid() && "Entry not initialized!")((Entry.isValid() && "Entry not initialized!") ? static_cast <void> (0) : __assert_fail ("Entry.isValid() && \"Entry not initialized!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3771, __PRETTY_FUNCTION__)); | |||
3772 | Entry.setAddress(Addr); | |||
3773 | Entry.setID(ID); | |||
3774 | Entry.setFlags(Flags); | |||
3775 | } else { | |||
3776 | OffloadEntryInfoTargetRegion Entry(OffloadingEntriesNum, Addr, ID, Flags); | |||
3777 | OffloadEntriesTargetRegion[DeviceID][FileID][ParentName][LineNum] = Entry; | |||
3778 | ++OffloadingEntriesNum; | |||
3779 | } | |||
3780 | } | |||
3781 | ||||
3782 | bool CGOpenMPRuntime::OffloadEntriesInfoManagerTy::hasTargetRegionEntryInfo( | |||
3783 | unsigned DeviceID, unsigned FileID, StringRef ParentName, | |||
3784 | unsigned LineNum) const { | |||
3785 | auto PerDevice = OffloadEntriesTargetRegion.find(DeviceID); | |||
3786 | if (PerDevice == OffloadEntriesTargetRegion.end()) | |||
3787 | return false; | |||
3788 | auto PerFile = PerDevice->second.find(FileID); | |||
3789 | if (PerFile == PerDevice->second.end()) | |||
3790 | return false; | |||
3791 | auto PerParentName = PerFile->second.find(ParentName); | |||
3792 | if (PerParentName == PerFile->second.end()) | |||
3793 | return false; | |||
3794 | auto PerLine = PerParentName->second.find(LineNum); | |||
3795 | if (PerLine == PerParentName->second.end()) | |||
3796 | return false; | |||
3797 | // Fail if this entry is already registered. | |||
3798 | if (PerLine->second.getAddress() || PerLine->second.getID()) | |||
3799 | return false; | |||
3800 | return true; | |||
3801 | } | |||
3802 | ||||
3803 | void CGOpenMPRuntime::OffloadEntriesInfoManagerTy::actOnTargetRegionEntriesInfo( | |||
3804 | const OffloadTargetRegionEntryInfoActTy &Action) { | |||
3805 | // Scan all target region entries and perform the provided action. | |||
3806 | for (const auto &D : OffloadEntriesTargetRegion) | |||
3807 | for (const auto &F : D.second) | |||
3808 | for (const auto &P : F.second) | |||
3809 | for (const auto &L : P.second) | |||
3810 | Action(D.first, F.first, P.first(), L.first, L.second); | |||
3811 | } | |||
3812 | ||||
3813 | void CGOpenMPRuntime::OffloadEntriesInfoManagerTy:: | |||
3814 | initializeDeviceGlobalVarEntryInfo(StringRef Name, | |||
3815 | OMPTargetGlobalVarEntryKind Flags, | |||
3816 | unsigned Order) { | |||
3817 | assert(CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is "((CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is " "only required for the device " "code generation.") ? static_cast <void> (0) : __assert_fail ("CGM.getLangOpts().OpenMPIsDevice && \"Initialization of entries is \" \"only required for the device \" \"code generation.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3819, __PRETTY_FUNCTION__)) | |||
3818 | "only required for the device "((CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is " "only required for the device " "code generation.") ? static_cast <void> (0) : __assert_fail ("CGM.getLangOpts().OpenMPIsDevice && \"Initialization of entries is \" \"only required for the device \" \"code generation.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3819, __PRETTY_FUNCTION__)) | |||
3819 | "code generation.")((CGM.getLangOpts().OpenMPIsDevice && "Initialization of entries is " "only required for the device " "code generation.") ? static_cast <void> (0) : __assert_fail ("CGM.getLangOpts().OpenMPIsDevice && \"Initialization of entries is \" \"only required for the device \" \"code generation.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3819, __PRETTY_FUNCTION__)); | |||
3820 | OffloadEntriesDeviceGlobalVar.try_emplace(Name, Order, Flags); | |||
3821 | ++OffloadingEntriesNum; | |||
3822 | } | |||
3823 | ||||
3824 | void CGOpenMPRuntime::OffloadEntriesInfoManagerTy:: | |||
3825 | registerDeviceGlobalVarEntryInfo(StringRef VarName, llvm::Constant *Addr, | |||
3826 | CharUnits VarSize, | |||
3827 | OMPTargetGlobalVarEntryKind Flags, | |||
3828 | llvm::GlobalValue::LinkageTypes Linkage) { | |||
3829 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
3830 | auto &Entry = OffloadEntriesDeviceGlobalVar[VarName]; | |||
3831 | assert(Entry.isValid() && Entry.getFlags() == Flags &&((Entry.isValid() && Entry.getFlags() == Flags && "Entry not initialized!") ? static_cast<void> (0) : __assert_fail ("Entry.isValid() && Entry.getFlags() == Flags && \"Entry not initialized!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3832, __PRETTY_FUNCTION__)) | |||
3832 | "Entry not initialized!")((Entry.isValid() && Entry.getFlags() == Flags && "Entry not initialized!") ? static_cast<void> (0) : __assert_fail ("Entry.isValid() && Entry.getFlags() == Flags && \"Entry not initialized!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3832, __PRETTY_FUNCTION__)); | |||
3833 | assert((!Entry.getAddress() || Entry.getAddress() == Addr) &&(((!Entry.getAddress() || Entry.getAddress() == Addr) && "Resetting with the new address.") ? static_cast<void> (0) : __assert_fail ("(!Entry.getAddress() || Entry.getAddress() == Addr) && \"Resetting with the new address.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3834, __PRETTY_FUNCTION__)) | |||
3834 | "Resetting with the new address.")(((!Entry.getAddress() || Entry.getAddress() == Addr) && "Resetting with the new address.") ? static_cast<void> (0) : __assert_fail ("(!Entry.getAddress() || Entry.getAddress() == Addr) && \"Resetting with the new address.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3834, __PRETTY_FUNCTION__)); | |||
3835 | if (Entry.getAddress() && hasDeviceGlobalVarEntryInfo(VarName)) { | |||
3836 | if (Entry.getVarSize().isZero()) { | |||
3837 | Entry.setVarSize(VarSize); | |||
3838 | Entry.setLinkage(Linkage); | |||
3839 | } | |||
3840 | return; | |||
3841 | } | |||
3842 | Entry.setVarSize(VarSize); | |||
3843 | Entry.setLinkage(Linkage); | |||
3844 | Entry.setAddress(Addr); | |||
3845 | } else { | |||
3846 | if (hasDeviceGlobalVarEntryInfo(VarName)) { | |||
3847 | auto &Entry = OffloadEntriesDeviceGlobalVar[VarName]; | |||
3848 | assert(Entry.isValid() && Entry.getFlags() == Flags &&((Entry.isValid() && Entry.getFlags() == Flags && "Entry not initialized!") ? static_cast<void> (0) : __assert_fail ("Entry.isValid() && Entry.getFlags() == Flags && \"Entry not initialized!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3849, __PRETTY_FUNCTION__)) | |||
3849 | "Entry not initialized!")((Entry.isValid() && Entry.getFlags() == Flags && "Entry not initialized!") ? static_cast<void> (0) : __assert_fail ("Entry.isValid() && Entry.getFlags() == Flags && \"Entry not initialized!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3849, __PRETTY_FUNCTION__)); | |||
3850 | assert((!Entry.getAddress() || Entry.getAddress() == Addr) &&(((!Entry.getAddress() || Entry.getAddress() == Addr) && "Resetting with the new address.") ? static_cast<void> (0) : __assert_fail ("(!Entry.getAddress() || Entry.getAddress() == Addr) && \"Resetting with the new address.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3851, __PRETTY_FUNCTION__)) | |||
3851 | "Resetting with the new address.")(((!Entry.getAddress() || Entry.getAddress() == Addr) && "Resetting with the new address.") ? static_cast<void> (0) : __assert_fail ("(!Entry.getAddress() || Entry.getAddress() == Addr) && \"Resetting with the new address.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3851, __PRETTY_FUNCTION__)); | |||
3852 | if (Entry.getVarSize().isZero()) { | |||
3853 | Entry.setVarSize(VarSize); | |||
3854 | Entry.setLinkage(Linkage); | |||
3855 | } | |||
3856 | return; | |||
3857 | } | |||
3858 | OffloadEntriesDeviceGlobalVar.try_emplace( | |||
3859 | VarName, OffloadingEntriesNum, Addr, VarSize, Flags, Linkage); | |||
3860 | ++OffloadingEntriesNum; | |||
3861 | } | |||
3862 | } | |||
3863 | ||||
3864 | void CGOpenMPRuntime::OffloadEntriesInfoManagerTy:: | |||
3865 | actOnDeviceGlobalVarEntriesInfo( | |||
3866 | const OffloadDeviceGlobalVarEntryInfoActTy &Action) { | |||
3867 | // Scan all target region entries and perform the provided action. | |||
3868 | for (const auto &E : OffloadEntriesDeviceGlobalVar) | |||
3869 | Action(E.getKey(), E.getValue()); | |||
3870 | } | |||
3871 | ||||
3872 | llvm::Function * | |||
3873 | CGOpenMPRuntime::createOffloadingBinaryDescriptorRegistration() { | |||
3874 | // If we don't have entries or if we are emitting code for the device, we | |||
3875 | // don't need to do anything. | |||
3876 | if (CGM.getLangOpts().OpenMPIsDevice || OffloadEntriesInfoManager.empty()) | |||
3877 | return nullptr; | |||
3878 | ||||
3879 | llvm::Module &M = CGM.getModule(); | |||
3880 | ASTContext &C = CGM.getContext(); | |||
3881 | ||||
3882 | // Get list of devices we care about | |||
3883 | const std::vector<llvm::Triple> &Devices = CGM.getLangOpts().OMPTargetTriples; | |||
3884 | ||||
3885 | // We should be creating an offloading descriptor only if there are devices | |||
3886 | // specified. | |||
3887 | assert(!Devices.empty() && "No OpenMP offloading devices??")((!Devices.empty() && "No OpenMP offloading devices??" ) ? static_cast<void> (0) : __assert_fail ("!Devices.empty() && \"No OpenMP offloading devices??\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 3887, __PRETTY_FUNCTION__)); | |||
3888 | ||||
3889 | // Create the external variables that will point to the begin and end of the | |||
3890 | // host entries section. These will be defined by the linker. | |||
3891 | llvm::Type *OffloadEntryTy = | |||
3892 | CGM.getTypes().ConvertTypeForMem(getTgtOffloadEntryQTy()); | |||
3893 | std::string EntriesBeginName = getName({"omp_offloading", "entries_begin"}); | |||
3894 | auto *HostEntriesBegin = new llvm::GlobalVariable( | |||
3895 | M, OffloadEntryTy, /*isConstant=*/true, | |||
3896 | llvm::GlobalValue::ExternalLinkage, /*Initializer=*/nullptr, | |||
3897 | EntriesBeginName); | |||
3898 | std::string EntriesEndName = getName({"omp_offloading", "entries_end"}); | |||
3899 | auto *HostEntriesEnd = | |||
3900 | new llvm::GlobalVariable(M, OffloadEntryTy, /*isConstant=*/true, | |||
3901 | llvm::GlobalValue::ExternalLinkage, | |||
3902 | /*Initializer=*/nullptr, EntriesEndName); | |||
3903 | ||||
3904 | // Create all device images | |||
3905 | auto *DeviceImageTy = cast<llvm::StructType>( | |||
3906 | CGM.getTypes().ConvertTypeForMem(getTgtDeviceImageQTy())); | |||
3907 | ConstantInitBuilder DeviceImagesBuilder(CGM); | |||
3908 | ConstantArrayBuilder DeviceImagesEntries = | |||
3909 | DeviceImagesBuilder.beginArray(DeviceImageTy); | |||
3910 | ||||
3911 | for (const llvm::Triple &Device : Devices) { | |||
3912 | StringRef T = Device.getTriple(); | |||
3913 | std::string BeginName = getName({"omp_offloading", "img_start", ""}); | |||
3914 | auto *ImgBegin = new llvm::GlobalVariable( | |||
3915 | M, CGM.Int8Ty, /*isConstant=*/true, | |||
3916 | llvm::GlobalValue::ExternalWeakLinkage, | |||
3917 | /*Initializer=*/nullptr, Twine(BeginName).concat(T)); | |||
3918 | std::string EndName = getName({"omp_offloading", "img_end", ""}); | |||
3919 | auto *ImgEnd = new llvm::GlobalVariable( | |||
3920 | M, CGM.Int8Ty, /*isConstant=*/true, | |||
3921 | llvm::GlobalValue::ExternalWeakLinkage, | |||
3922 | /*Initializer=*/nullptr, Twine(EndName).concat(T)); | |||
3923 | ||||
3924 | llvm::Constant *Data[] = {ImgBegin, ImgEnd, HostEntriesBegin, | |||
3925 | HostEntriesEnd}; | |||
3926 | createConstantGlobalStructAndAddToParent(CGM, getTgtDeviceImageQTy(), Data, | |||
3927 | DeviceImagesEntries); | |||
3928 | } | |||
3929 | ||||
3930 | // Create device images global array. | |||
3931 | std::string ImagesName = getName({"omp_offloading", "device_images"}); | |||
3932 | llvm::GlobalVariable *DeviceImages = | |||
3933 | DeviceImagesEntries.finishAndCreateGlobal(ImagesName, | |||
3934 | CGM.getPointerAlign(), | |||
3935 | /*isConstant=*/true); | |||
3936 | DeviceImages->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); | |||
3937 | ||||
3938 | // This is a Zero array to be used in the creation of the constant expressions | |||
3939 | llvm::Constant *Index[] = {llvm::Constant::getNullValue(CGM.Int32Ty), | |||
3940 | llvm::Constant::getNullValue(CGM.Int32Ty)}; | |||
3941 | ||||
3942 | // Create the target region descriptor. | |||
3943 | llvm::Constant *Data[] = { | |||
3944 | llvm::ConstantInt::get(CGM.Int32Ty, Devices.size()), | |||
3945 | llvm::ConstantExpr::getGetElementPtr(DeviceImages->getValueType(), | |||
3946 | DeviceImages, Index), | |||
3947 | HostEntriesBegin, HostEntriesEnd}; | |||
3948 | std::string Descriptor = getName({"omp_offloading", "descriptor"}); | |||
3949 | llvm::GlobalVariable *Desc = createGlobalStruct( | |||
3950 | CGM, getTgtBinaryDescriptorQTy(), /*IsConstant=*/true, Data, Descriptor); | |||
3951 | ||||
3952 | // Emit code to register or unregister the descriptor at execution | |||
3953 | // startup or closing, respectively. | |||
3954 | ||||
3955 | llvm::Function *UnRegFn; | |||
3956 | { | |||
3957 | FunctionArgList Args; | |||
3958 | ImplicitParamDecl DummyPtr(C, C.VoidPtrTy, ImplicitParamDecl::Other); | |||
3959 | Args.push_back(&DummyPtr); | |||
3960 | ||||
3961 | CodeGenFunction CGF(CGM); | |||
3962 | // Disable debug info for global (de-)initializer because they are not part | |||
3963 | // of some particular construct. | |||
3964 | CGF.disableDebugInfo(); | |||
3965 | const auto &FI = | |||
3966 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
3967 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | |||
3968 | std::string UnregName = getName({"omp_offloading", "descriptor_unreg"}); | |||
3969 | UnRegFn = CGM.CreateGlobalInitOrDestructFunction(FTy, UnregName, FI); | |||
3970 | CGF.StartFunction(GlobalDecl(), C.VoidTy, UnRegFn, FI, Args); | |||
3971 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__tgt_unregister_lib), | |||
3972 | Desc); | |||
3973 | CGF.FinishFunction(); | |||
3974 | } | |||
3975 | llvm::Function *RegFn; | |||
3976 | { | |||
3977 | CodeGenFunction CGF(CGM); | |||
3978 | // Disable debug info for global (de-)initializer because they are not part | |||
3979 | // of some particular construct. | |||
3980 | CGF.disableDebugInfo(); | |||
3981 | const auto &FI = CGM.getTypes().arrangeNullaryFunction(); | |||
3982 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | |||
3983 | ||||
3984 | // Encode offload target triples into the registration function name. It | |||
3985 | // will serve as a comdat key for the registration/unregistration code for | |||
3986 | // this particular combination of offloading targets. | |||
3987 | SmallVector<StringRef, 4U> RegFnNameParts(Devices.size() + 2U); | |||
3988 | RegFnNameParts[0] = "omp_offloading"; | |||
3989 | RegFnNameParts[1] = "descriptor_reg"; | |||
3990 | llvm::transform(Devices, std::next(RegFnNameParts.begin(), 2), | |||
3991 | [](const llvm::Triple &T) -> const std::string& { | |||
3992 | return T.getTriple(); | |||
3993 | }); | |||
3994 | llvm::sort(std::next(RegFnNameParts.begin(), 2), RegFnNameParts.end()); | |||
3995 | std::string Descriptor = getName(RegFnNameParts); | |||
3996 | RegFn = CGM.CreateGlobalInitOrDestructFunction(FTy, Descriptor, FI); | |||
3997 | CGF.StartFunction(GlobalDecl(), C.VoidTy, RegFn, FI, FunctionArgList()); | |||
3998 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__tgt_register_lib), Desc); | |||
3999 | // Create a variable to drive the registration and unregistration of the | |||
4000 | // descriptor, so we can reuse the logic that emits Ctors and Dtors. | |||
4001 | ImplicitParamDecl RegUnregVar(C, C.getTranslationUnitDecl(), | |||
4002 | SourceLocation(), nullptr, C.CharTy, | |||
4003 | ImplicitParamDecl::Other); | |||
4004 | CGM.getCXXABI().registerGlobalDtor(CGF, RegUnregVar, UnRegFn, Desc); | |||
4005 | CGF.FinishFunction(); | |||
4006 | } | |||
4007 | if (CGM.supportsCOMDAT()) { | |||
4008 | // It is sufficient to call registration function only once, so create a | |||
4009 | // COMDAT group for registration/unregistration functions and associated | |||
4010 | // data. That would reduce startup time and code size. Registration | |||
4011 | // function serves as a COMDAT group key. | |||
4012 | llvm::Comdat *ComdatKey = M.getOrInsertComdat(RegFn->getName()); | |||
4013 | RegFn->setLinkage(llvm::GlobalValue::LinkOnceAnyLinkage); | |||
4014 | RegFn->setVisibility(llvm::GlobalValue::HiddenVisibility); | |||
4015 | RegFn->setComdat(ComdatKey); | |||
4016 | UnRegFn->setComdat(ComdatKey); | |||
4017 | DeviceImages->setComdat(ComdatKey); | |||
4018 | Desc->setComdat(ComdatKey); | |||
4019 | } | |||
4020 | return RegFn; | |||
4021 | } | |||
4022 | ||||
4023 | void CGOpenMPRuntime::createOffloadEntry( | |||
4024 | llvm::Constant *ID, llvm::Constant *Addr, uint64_t Size, int32_t Flags, | |||
4025 | llvm::GlobalValue::LinkageTypes Linkage) { | |||
4026 | StringRef Name = Addr->getName(); | |||
4027 | llvm::Module &M = CGM.getModule(); | |||
4028 | llvm::LLVMContext &C = M.getContext(); | |||
4029 | ||||
4030 | // Create constant string with the name. | |||
4031 | llvm::Constant *StrPtrInit = llvm::ConstantDataArray::getString(C, Name); | |||
4032 | ||||
4033 | std::string StringName = getName({"omp_offloading", "entry_name"}); | |||
4034 | auto *Str = new llvm::GlobalVariable( | |||
4035 | M, StrPtrInit->getType(), /*isConstant=*/true, | |||
4036 | llvm::GlobalValue::InternalLinkage, StrPtrInit, StringName); | |||
4037 | Str->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); | |||
4038 | ||||
4039 | llvm::Constant *Data[] = {llvm::ConstantExpr::getBitCast(ID, CGM.VoidPtrTy), | |||
4040 | llvm::ConstantExpr::getBitCast(Str, CGM.Int8PtrTy), | |||
4041 | llvm::ConstantInt::get(CGM.SizeTy, Size), | |||
4042 | llvm::ConstantInt::get(CGM.Int32Ty, Flags), | |||
4043 | llvm::ConstantInt::get(CGM.Int32Ty, 0)}; | |||
4044 | std::string EntryName = getName({"omp_offloading", "entry", ""}); | |||
4045 | llvm::GlobalVariable *Entry = createGlobalStruct( | |||
4046 | CGM, getTgtOffloadEntryQTy(), /*IsConstant=*/true, Data, | |||
4047 | Twine(EntryName).concat(Name), llvm::GlobalValue::WeakAnyLinkage); | |||
4048 | ||||
4049 | // The entry has to be created in the section the linker expects it to be. | |||
4050 | std::string Section = getName({"omp_offloading", "entries"}); | |||
4051 | Entry->setSection(Section); | |||
4052 | } | |||
4053 | ||||
4054 | void CGOpenMPRuntime::createOffloadEntriesAndInfoMetadata() { | |||
4055 | // Emit the offloading entries and metadata so that the device codegen side | |||
4056 | // can easily figure out what to emit. The produced metadata looks like | |||
4057 | // this: | |||
4058 | // | |||
4059 | // !omp_offload.info = !{!1, ...} | |||
4060 | // | |||
4061 | // Right now we only generate metadata for function that contain target | |||
4062 | // regions. | |||
4063 | ||||
4064 | // If we do not have entries, we don't need to do anything. | |||
4065 | if (OffloadEntriesInfoManager.empty()) | |||
4066 | return; | |||
4067 | ||||
4068 | llvm::Module &M = CGM.getModule(); | |||
4069 | llvm::LLVMContext &C = M.getContext(); | |||
4070 | SmallVector<const OffloadEntriesInfoManagerTy::OffloadEntryInfo *, 16> | |||
4071 | OrderedEntries(OffloadEntriesInfoManager.size()); | |||
4072 | llvm::SmallVector<StringRef, 16> ParentFunctions( | |||
4073 | OffloadEntriesInfoManager.size()); | |||
4074 | ||||
4075 | // Auxiliary methods to create metadata values and strings. | |||
4076 | auto &&GetMDInt = [this](unsigned V) { | |||
4077 | return llvm::ConstantAsMetadata::get( | |||
4078 | llvm::ConstantInt::get(CGM.Int32Ty, V)); | |||
4079 | }; | |||
4080 | ||||
4081 | auto &&GetMDString = [&C](StringRef V) { return llvm::MDString::get(C, V); }; | |||
4082 | ||||
4083 | // Create the offloading info metadata node. | |||
4084 | llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("omp_offload.info"); | |||
4085 | ||||
4086 | // Create function that emits metadata for each target region entry; | |||
4087 | auto &&TargetRegionMetadataEmitter = | |||
4088 | [&C, MD, &OrderedEntries, &ParentFunctions, &GetMDInt, &GetMDString]( | |||
4089 | unsigned DeviceID, unsigned FileID, StringRef ParentName, | |||
4090 | unsigned Line, | |||
4091 | const OffloadEntriesInfoManagerTy::OffloadEntryInfoTargetRegion &E) { | |||
4092 | // Generate metadata for target regions. Each entry of this metadata | |||
4093 | // contains: | |||
4094 | // - Entry 0 -> Kind of this type of metadata (0). | |||
4095 | // - Entry 1 -> Device ID of the file where the entry was identified. | |||
4096 | // - Entry 2 -> File ID of the file where the entry was identified. | |||
4097 | // - Entry 3 -> Mangled name of the function where the entry was | |||
4098 | // identified. | |||
4099 | // - Entry 4 -> Line in the file where the entry was identified. | |||
4100 | // - Entry 5 -> Order the entry was created. | |||
4101 | // The first element of the metadata node is the kind. | |||
4102 | llvm::Metadata *Ops[] = {GetMDInt(E.getKind()), GetMDInt(DeviceID), | |||
4103 | GetMDInt(FileID), GetMDString(ParentName), | |||
4104 | GetMDInt(Line), GetMDInt(E.getOrder())}; | |||
4105 | ||||
4106 | // Save this entry in the right position of the ordered entries array. | |||
4107 | OrderedEntries[E.getOrder()] = &E; | |||
4108 | ParentFunctions[E.getOrder()] = ParentName; | |||
4109 | ||||
4110 | // Add metadata to the named metadata node. | |||
4111 | MD->addOperand(llvm::MDNode::get(C, Ops)); | |||
4112 | }; | |||
4113 | ||||
4114 | OffloadEntriesInfoManager.actOnTargetRegionEntriesInfo( | |||
4115 | TargetRegionMetadataEmitter); | |||
4116 | ||||
4117 | // Create function that emits metadata for each device global variable entry; | |||
4118 | auto &&DeviceGlobalVarMetadataEmitter = | |||
4119 | [&C, &OrderedEntries, &GetMDInt, &GetMDString, | |||
4120 | MD](StringRef MangledName, | |||
4121 | const OffloadEntriesInfoManagerTy::OffloadEntryInfoDeviceGlobalVar | |||
4122 | &E) { | |||
4123 | // Generate metadata for global variables. Each entry of this metadata | |||
4124 | // contains: | |||
4125 | // - Entry 0 -> Kind of this type of metadata (1). | |||
4126 | // - Entry 1 -> Mangled name of the variable. | |||
4127 | // - Entry 2 -> Declare target kind. | |||
4128 | // - Entry 3 -> Order the entry was created. | |||
4129 | // The first element of the metadata node is the kind. | |||
4130 | llvm::Metadata *Ops[] = { | |||
4131 | GetMDInt(E.getKind()), GetMDString(MangledName), | |||
4132 | GetMDInt(E.getFlags()), GetMDInt(E.getOrder())}; | |||
4133 | ||||
4134 | // Save this entry in the right position of the ordered entries array. | |||
4135 | OrderedEntries[E.getOrder()] = &E; | |||
4136 | ||||
4137 | // Add metadata to the named metadata node. | |||
4138 | MD->addOperand(llvm::MDNode::get(C, Ops)); | |||
4139 | }; | |||
4140 | ||||
4141 | OffloadEntriesInfoManager.actOnDeviceGlobalVarEntriesInfo( | |||
4142 | DeviceGlobalVarMetadataEmitter); | |||
4143 | ||||
4144 | for (const auto *E : OrderedEntries) { | |||
4145 | assert(E && "All ordered entries must exist!")((E && "All ordered entries must exist!") ? static_cast <void> (0) : __assert_fail ("E && \"All ordered entries must exist!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4145, __PRETTY_FUNCTION__)); | |||
4146 | if (const auto *CE = | |||
4147 | dyn_cast<OffloadEntriesInfoManagerTy::OffloadEntryInfoTargetRegion>( | |||
4148 | E)) { | |||
4149 | if (!CE->getID() || !CE->getAddress()) { | |||
4150 | // Do not blame the entry if the parent funtion is not emitted. | |||
4151 | StringRef FnName = ParentFunctions[CE->getOrder()]; | |||
4152 | if (!CGM.GetGlobalValue(FnName)) | |||
4153 | continue; | |||
4154 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
4155 | DiagnosticsEngine::Error, | |||
4156 | "Offloading entry for target region is incorrect: either the " | |||
4157 | "address or the ID is invalid."); | |||
4158 | CGM.getDiags().Report(DiagID); | |||
4159 | continue; | |||
4160 | } | |||
4161 | createOffloadEntry(CE->getID(), CE->getAddress(), /*Size=*/0, | |||
4162 | CE->getFlags(), llvm::GlobalValue::WeakAnyLinkage); | |||
4163 | } else if (const auto *CE = | |||
4164 | dyn_cast<OffloadEntriesInfoManagerTy:: | |||
4165 | OffloadEntryInfoDeviceGlobalVar>(E)) { | |||
4166 | OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryKind Flags = | |||
4167 | static_cast<OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryKind>( | |||
4168 | CE->getFlags()); | |||
4169 | switch (Flags) { | |||
4170 | case OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryTo: { | |||
4171 | if (!CE->getAddress()) { | |||
4172 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
4173 | DiagnosticsEngine::Error, | |||
4174 | "Offloading entry for declare target variable is incorrect: the " | |||
4175 | "address is invalid."); | |||
4176 | CGM.getDiags().Report(DiagID); | |||
4177 | continue; | |||
4178 | } | |||
4179 | // The vaiable has no definition - no need to add the entry. | |||
4180 | if (CE->getVarSize().isZero()) | |||
4181 | continue; | |||
4182 | break; | |||
4183 | } | |||
4184 | case OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryLink: | |||
4185 | assert(((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress()) ||((((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress ()) || (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress ())) && "Declaret target link address is set.") ? static_cast <void> (0) : __assert_fail ("((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress()) || (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress())) && \"Declaret target link address is set.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4187, __PRETTY_FUNCTION__)) | |||
4186 | (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress())) &&((((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress ()) || (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress ())) && "Declaret target link address is set.") ? static_cast <void> (0) : __assert_fail ("((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress()) || (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress())) && \"Declaret target link address is set.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4187, __PRETTY_FUNCTION__)) | |||
4187 | "Declaret target link address is set.")((((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress ()) || (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress ())) && "Declaret target link address is set.") ? static_cast <void> (0) : __assert_fail ("((CGM.getLangOpts().OpenMPIsDevice && !CE->getAddress()) || (!CGM.getLangOpts().OpenMPIsDevice && CE->getAddress())) && \"Declaret target link address is set.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4187, __PRETTY_FUNCTION__)); | |||
4188 | if (CGM.getLangOpts().OpenMPIsDevice) | |||
4189 | continue; | |||
4190 | if (!CE->getAddress()) { | |||
4191 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
4192 | DiagnosticsEngine::Error, | |||
4193 | "Offloading entry for declare target variable is incorrect: the " | |||
4194 | "address is invalid."); | |||
4195 | CGM.getDiags().Report(DiagID); | |||
4196 | continue; | |||
4197 | } | |||
4198 | break; | |||
4199 | } | |||
4200 | createOffloadEntry(CE->getAddress(), CE->getAddress(), | |||
4201 | CE->getVarSize().getQuantity(), Flags, | |||
4202 | CE->getLinkage()); | |||
4203 | } else { | |||
4204 | llvm_unreachable("Unsupported entry kind.")::llvm::llvm_unreachable_internal("Unsupported entry kind.", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4204); | |||
4205 | } | |||
4206 | } | |||
4207 | } | |||
4208 | ||||
4209 | /// Loads all the offload entries information from the host IR | |||
4210 | /// metadata. | |||
4211 | void CGOpenMPRuntime::loadOffloadInfoMetadata() { | |||
4212 | // If we are in target mode, load the metadata from the host IR. This code has | |||
4213 | // to match the metadaata creation in createOffloadEntriesAndInfoMetadata(). | |||
4214 | ||||
4215 | if (!CGM.getLangOpts().OpenMPIsDevice) | |||
4216 | return; | |||
4217 | ||||
4218 | if (CGM.getLangOpts().OMPHostIRFile.empty()) | |||
4219 | return; | |||
4220 | ||||
4221 | auto Buf = llvm::MemoryBuffer::getFile(CGM.getLangOpts().OMPHostIRFile); | |||
4222 | if (auto EC = Buf.getError()) { | |||
4223 | CGM.getDiags().Report(diag::err_cannot_open_file) | |||
4224 | << CGM.getLangOpts().OMPHostIRFile << EC.message(); | |||
4225 | return; | |||
4226 | } | |||
4227 | ||||
4228 | llvm::LLVMContext C; | |||
4229 | auto ME = expectedToErrorOrAndEmitErrors( | |||
4230 | C, llvm::parseBitcodeFile(Buf.get()->getMemBufferRef(), C)); | |||
4231 | ||||
4232 | if (auto EC = ME.getError()) { | |||
4233 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
4234 | DiagnosticsEngine::Error, "Unable to parse host IR file '%0':'%1'"); | |||
4235 | CGM.getDiags().Report(DiagID) | |||
4236 | << CGM.getLangOpts().OMPHostIRFile << EC.message(); | |||
4237 | return; | |||
4238 | } | |||
4239 | ||||
4240 | llvm::NamedMDNode *MD = ME.get()->getNamedMetadata("omp_offload.info"); | |||
4241 | if (!MD) | |||
4242 | return; | |||
4243 | ||||
4244 | for (llvm::MDNode *MN : MD->operands()) { | |||
4245 | auto &&GetMDInt = [MN](unsigned Idx) { | |||
4246 | auto *V = cast<llvm::ConstantAsMetadata>(MN->getOperand(Idx)); | |||
4247 | return cast<llvm::ConstantInt>(V->getValue())->getZExtValue(); | |||
4248 | }; | |||
4249 | ||||
4250 | auto &&GetMDString = [MN](unsigned Idx) { | |||
4251 | auto *V = cast<llvm::MDString>(MN->getOperand(Idx)); | |||
4252 | return V->getString(); | |||
4253 | }; | |||
4254 | ||||
4255 | switch (GetMDInt(0)) { | |||
4256 | default: | |||
4257 | llvm_unreachable("Unexpected metadata!")::llvm::llvm_unreachable_internal("Unexpected metadata!", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4257); | |||
4258 | break; | |||
4259 | case OffloadEntriesInfoManagerTy::OffloadEntryInfo:: | |||
4260 | OffloadingEntryInfoTargetRegion: | |||
4261 | OffloadEntriesInfoManager.initializeTargetRegionEntryInfo( | |||
4262 | /*DeviceID=*/GetMDInt(1), /*FileID=*/GetMDInt(2), | |||
4263 | /*ParentName=*/GetMDString(3), /*Line=*/GetMDInt(4), | |||
4264 | /*Order=*/GetMDInt(5)); | |||
4265 | break; | |||
4266 | case OffloadEntriesInfoManagerTy::OffloadEntryInfo:: | |||
4267 | OffloadingEntryInfoDeviceGlobalVar: | |||
4268 | OffloadEntriesInfoManager.initializeDeviceGlobalVarEntryInfo( | |||
4269 | /*MangledName=*/GetMDString(1), | |||
4270 | static_cast<OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryKind>( | |||
4271 | /*Flags=*/GetMDInt(2)), | |||
4272 | /*Order=*/GetMDInt(3)); | |||
4273 | break; | |||
4274 | } | |||
4275 | } | |||
4276 | } | |||
4277 | ||||
4278 | void CGOpenMPRuntime::emitKmpRoutineEntryT(QualType KmpInt32Ty) { | |||
4279 | if (!KmpRoutineEntryPtrTy) { | |||
4280 | // Build typedef kmp_int32 (* kmp_routine_entry_t)(kmp_int32, void *); type. | |||
4281 | ASTContext &C = CGM.getContext(); | |||
4282 | QualType KmpRoutineEntryTyArgs[] = {KmpInt32Ty, C.VoidPtrTy}; | |||
4283 | FunctionProtoType::ExtProtoInfo EPI; | |||
4284 | KmpRoutineEntryPtrQTy = C.getPointerType( | |||
4285 | C.getFunctionType(KmpInt32Ty, KmpRoutineEntryTyArgs, EPI)); | |||
4286 | KmpRoutineEntryPtrTy = CGM.getTypes().ConvertType(KmpRoutineEntryPtrQTy); | |||
4287 | } | |||
4288 | } | |||
4289 | ||||
4290 | QualType CGOpenMPRuntime::getTgtOffloadEntryQTy() { | |||
4291 | // Make sure the type of the entry is already created. This is the type we | |||
4292 | // have to create: | |||
4293 | // struct __tgt_offload_entry{ | |||
4294 | // void *addr; // Pointer to the offload entry info. | |||
4295 | // // (function or global) | |||
4296 | // char *name; // Name of the function or global. | |||
4297 | // size_t size; // Size of the entry info (0 if it a function). | |||
4298 | // int32_t flags; // Flags associated with the entry, e.g. 'link'. | |||
4299 | // int32_t reserved; // Reserved, to use by the runtime library. | |||
4300 | // }; | |||
4301 | if (TgtOffloadEntryQTy.isNull()) { | |||
4302 | ASTContext &C = CGM.getContext(); | |||
4303 | RecordDecl *RD = C.buildImplicitRecord("__tgt_offload_entry"); | |||
4304 | RD->startDefinition(); | |||
4305 | addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
4306 | addFieldToRecordDecl(C, RD, C.getPointerType(C.CharTy)); | |||
4307 | addFieldToRecordDecl(C, RD, C.getSizeType()); | |||
4308 | addFieldToRecordDecl( | |||
4309 | C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true)); | |||
4310 | addFieldToRecordDecl( | |||
4311 | C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true)); | |||
4312 | RD->completeDefinition(); | |||
4313 | RD->addAttr(PackedAttr::CreateImplicit(C)); | |||
4314 | TgtOffloadEntryQTy = C.getRecordType(RD); | |||
4315 | } | |||
4316 | return TgtOffloadEntryQTy; | |||
4317 | } | |||
4318 | ||||
4319 | QualType CGOpenMPRuntime::getTgtDeviceImageQTy() { | |||
4320 | // These are the types we need to build: | |||
4321 | // struct __tgt_device_image{ | |||
4322 | // void *ImageStart; // Pointer to the target code start. | |||
4323 | // void *ImageEnd; // Pointer to the target code end. | |||
4324 | // // We also add the host entries to the device image, as it may be useful | |||
4325 | // // for the target runtime to have access to that information. | |||
4326 | // __tgt_offload_entry *EntriesBegin; // Begin of the table with all | |||
4327 | // // the entries. | |||
4328 | // __tgt_offload_entry *EntriesEnd; // End of the table with all the | |||
4329 | // // entries (non inclusive). | |||
4330 | // }; | |||
4331 | if (TgtDeviceImageQTy.isNull()) { | |||
4332 | ASTContext &C = CGM.getContext(); | |||
4333 | RecordDecl *RD = C.buildImplicitRecord("__tgt_device_image"); | |||
4334 | RD->startDefinition(); | |||
4335 | addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
4336 | addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
4337 | addFieldToRecordDecl(C, RD, C.getPointerType(getTgtOffloadEntryQTy())); | |||
4338 | addFieldToRecordDecl(C, RD, C.getPointerType(getTgtOffloadEntryQTy())); | |||
4339 | RD->completeDefinition(); | |||
4340 | TgtDeviceImageQTy = C.getRecordType(RD); | |||
4341 | } | |||
4342 | return TgtDeviceImageQTy; | |||
4343 | } | |||
4344 | ||||
4345 | QualType CGOpenMPRuntime::getTgtBinaryDescriptorQTy() { | |||
4346 | // struct __tgt_bin_desc{ | |||
4347 | // int32_t NumDevices; // Number of devices supported. | |||
4348 | // __tgt_device_image *DeviceImages; // Arrays of device images | |||
4349 | // // (one per device). | |||
4350 | // __tgt_offload_entry *EntriesBegin; // Begin of the table with all the | |||
4351 | // // entries. | |||
4352 | // __tgt_offload_entry *EntriesEnd; // End of the table with all the | |||
4353 | // // entries (non inclusive). | |||
4354 | // }; | |||
4355 | if (TgtBinaryDescriptorQTy.isNull()) { | |||
4356 | ASTContext &C = CGM.getContext(); | |||
4357 | RecordDecl *RD = C.buildImplicitRecord("__tgt_bin_desc"); | |||
4358 | RD->startDefinition(); | |||
4359 | addFieldToRecordDecl( | |||
4360 | C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/true)); | |||
4361 | addFieldToRecordDecl(C, RD, C.getPointerType(getTgtDeviceImageQTy())); | |||
4362 | addFieldToRecordDecl(C, RD, C.getPointerType(getTgtOffloadEntryQTy())); | |||
4363 | addFieldToRecordDecl(C, RD, C.getPointerType(getTgtOffloadEntryQTy())); | |||
4364 | RD->completeDefinition(); | |||
4365 | TgtBinaryDescriptorQTy = C.getRecordType(RD); | |||
4366 | } | |||
4367 | return TgtBinaryDescriptorQTy; | |||
4368 | } | |||
4369 | ||||
4370 | namespace { | |||
4371 | struct PrivateHelpersTy { | |||
4372 | PrivateHelpersTy(const VarDecl *Original, const VarDecl *PrivateCopy, | |||
4373 | const VarDecl *PrivateElemInit) | |||
4374 | : Original(Original), PrivateCopy(PrivateCopy), | |||
4375 | PrivateElemInit(PrivateElemInit) {} | |||
4376 | const VarDecl *Original; | |||
4377 | const VarDecl *PrivateCopy; | |||
4378 | const VarDecl *PrivateElemInit; | |||
4379 | }; | |||
4380 | typedef std::pair<CharUnits /*Align*/, PrivateHelpersTy> PrivateDataTy; | |||
4381 | } // anonymous namespace | |||
4382 | ||||
4383 | static RecordDecl * | |||
4384 | createPrivatesRecordDecl(CodeGenModule &CGM, ArrayRef<PrivateDataTy> Privates) { | |||
4385 | if (!Privates.empty()) { | |||
4386 | ASTContext &C = CGM.getContext(); | |||
4387 | // Build struct .kmp_privates_t. { | |||
4388 | // /* private vars */ | |||
4389 | // }; | |||
4390 | RecordDecl *RD = C.buildImplicitRecord(".kmp_privates.t"); | |||
4391 | RD->startDefinition(); | |||
4392 | for (const auto &Pair : Privates) { | |||
4393 | const VarDecl *VD = Pair.second.Original; | |||
4394 | QualType Type = VD->getType().getNonReferenceType(); | |||
4395 | FieldDecl *FD = addFieldToRecordDecl(C, RD, Type); | |||
4396 | if (VD->hasAttrs()) { | |||
4397 | for (specific_attr_iterator<AlignedAttr> I(VD->getAttrs().begin()), | |||
4398 | E(VD->getAttrs().end()); | |||
4399 | I != E; ++I) | |||
4400 | FD->addAttr(*I); | |||
4401 | } | |||
4402 | } | |||
4403 | RD->completeDefinition(); | |||
4404 | return RD; | |||
4405 | } | |||
4406 | return nullptr; | |||
4407 | } | |||
4408 | ||||
4409 | static RecordDecl * | |||
4410 | createKmpTaskTRecordDecl(CodeGenModule &CGM, OpenMPDirectiveKind Kind, | |||
4411 | QualType KmpInt32Ty, | |||
4412 | QualType KmpRoutineEntryPointerQTy) { | |||
4413 | ASTContext &C = CGM.getContext(); | |||
4414 | // Build struct kmp_task_t { | |||
4415 | // void * shareds; | |||
4416 | // kmp_routine_entry_t routine; | |||
4417 | // kmp_int32 part_id; | |||
4418 | // kmp_cmplrdata_t data1; | |||
4419 | // kmp_cmplrdata_t data2; | |||
4420 | // For taskloops additional fields: | |||
4421 | // kmp_uint64 lb; | |||
4422 | // kmp_uint64 ub; | |||
4423 | // kmp_int64 st; | |||
4424 | // kmp_int32 liter; | |||
4425 | // void * reductions; | |||
4426 | // }; | |||
4427 | RecordDecl *UD = C.buildImplicitRecord("kmp_cmplrdata_t", TTK_Union); | |||
4428 | UD->startDefinition(); | |||
4429 | addFieldToRecordDecl(C, UD, KmpInt32Ty); | |||
4430 | addFieldToRecordDecl(C, UD, KmpRoutineEntryPointerQTy); | |||
4431 | UD->completeDefinition(); | |||
4432 | QualType KmpCmplrdataTy = C.getRecordType(UD); | |||
4433 | RecordDecl *RD = C.buildImplicitRecord("kmp_task_t"); | |||
4434 | RD->startDefinition(); | |||
4435 | addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
4436 | addFieldToRecordDecl(C, RD, KmpRoutineEntryPointerQTy); | |||
4437 | addFieldToRecordDecl(C, RD, KmpInt32Ty); | |||
4438 | addFieldToRecordDecl(C, RD, KmpCmplrdataTy); | |||
4439 | addFieldToRecordDecl(C, RD, KmpCmplrdataTy); | |||
4440 | if (isOpenMPTaskLoopDirective(Kind)) { | |||
4441 | QualType KmpUInt64Ty = | |||
4442 | CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/0); | |||
4443 | QualType KmpInt64Ty = | |||
4444 | CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1); | |||
4445 | addFieldToRecordDecl(C, RD, KmpUInt64Ty); | |||
4446 | addFieldToRecordDecl(C, RD, KmpUInt64Ty); | |||
4447 | addFieldToRecordDecl(C, RD, KmpInt64Ty); | |||
4448 | addFieldToRecordDecl(C, RD, KmpInt32Ty); | |||
4449 | addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
4450 | } | |||
4451 | RD->completeDefinition(); | |||
4452 | return RD; | |||
4453 | } | |||
4454 | ||||
4455 | static RecordDecl * | |||
4456 | createKmpTaskTWithPrivatesRecordDecl(CodeGenModule &CGM, QualType KmpTaskTQTy, | |||
4457 | ArrayRef<PrivateDataTy> Privates) { | |||
4458 | ASTContext &C = CGM.getContext(); | |||
4459 | // Build struct kmp_task_t_with_privates { | |||
4460 | // kmp_task_t task_data; | |||
4461 | // .kmp_privates_t. privates; | |||
4462 | // }; | |||
4463 | RecordDecl *RD = C.buildImplicitRecord("kmp_task_t_with_privates"); | |||
4464 | RD->startDefinition(); | |||
4465 | addFieldToRecordDecl(C, RD, KmpTaskTQTy); | |||
4466 | if (const RecordDecl *PrivateRD = createPrivatesRecordDecl(CGM, Privates)) | |||
4467 | addFieldToRecordDecl(C, RD, C.getRecordType(PrivateRD)); | |||
4468 | RD->completeDefinition(); | |||
4469 | return RD; | |||
4470 | } | |||
4471 | ||||
4472 | /// Emit a proxy function which accepts kmp_task_t as the second | |||
4473 | /// argument. | |||
4474 | /// \code | |||
4475 | /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { | |||
4476 | /// TaskFunction(gtid, tt->part_id, &tt->privates, task_privates_map, tt, | |||
4477 | /// For taskloops: | |||
4478 | /// tt->task_data.lb, tt->task_data.ub, tt->task_data.st, tt->task_data.liter, | |||
4479 | /// tt->reductions, tt->shareds); | |||
4480 | /// return 0; | |||
4481 | /// } | |||
4482 | /// \endcode | |||
4483 | static llvm::Function * | |||
4484 | emitProxyTaskFunction(CodeGenModule &CGM, SourceLocation Loc, | |||
4485 | OpenMPDirectiveKind Kind, QualType KmpInt32Ty, | |||
4486 | QualType KmpTaskTWithPrivatesPtrQTy, | |||
4487 | QualType KmpTaskTWithPrivatesQTy, QualType KmpTaskTQTy, | |||
4488 | QualType SharedsPtrTy, llvm::Function *TaskFunction, | |||
4489 | llvm::Value *TaskPrivatesMap) { | |||
4490 | ASTContext &C = CGM.getContext(); | |||
4491 | FunctionArgList Args; | |||
4492 | ImplicitParamDecl GtidArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, KmpInt32Ty, | |||
4493 | ImplicitParamDecl::Other); | |||
4494 | ImplicitParamDecl TaskTypeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
4495 | KmpTaskTWithPrivatesPtrQTy.withRestrict(), | |||
4496 | ImplicitParamDecl::Other); | |||
4497 | Args.push_back(&GtidArg); | |||
4498 | Args.push_back(&TaskTypeArg); | |||
4499 | const auto &TaskEntryFnInfo = | |||
4500 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(KmpInt32Ty, Args); | |||
4501 | llvm::FunctionType *TaskEntryTy = | |||
4502 | CGM.getTypes().GetFunctionType(TaskEntryFnInfo); | |||
4503 | std::string Name = CGM.getOpenMPRuntime().getName({"omp_task_entry", ""}); | |||
4504 | auto *TaskEntry = llvm::Function::Create( | |||
4505 | TaskEntryTy, llvm::GlobalValue::InternalLinkage, Name, &CGM.getModule()); | |||
4506 | CGM.SetInternalFunctionAttributes(GlobalDecl(), TaskEntry, TaskEntryFnInfo); | |||
4507 | TaskEntry->setDoesNotRecurse(); | |||
4508 | CodeGenFunction CGF(CGM); | |||
4509 | CGF.StartFunction(GlobalDecl(), KmpInt32Ty, TaskEntry, TaskEntryFnInfo, Args, | |||
4510 | Loc, Loc); | |||
4511 | ||||
4512 | // TaskFunction(gtid, tt->task_data.part_id, &tt->privates, task_privates_map, | |||
4513 | // tt, | |||
4514 | // For taskloops: | |||
4515 | // tt->task_data.lb, tt->task_data.ub, tt->task_data.st, tt->task_data.liter, | |||
4516 | // tt->task_data.shareds); | |||
4517 | llvm::Value *GtidParam = CGF.EmitLoadOfScalar( | |||
4518 | CGF.GetAddrOfLocalVar(&GtidArg), /*Volatile=*/false, KmpInt32Ty, Loc); | |||
4519 | LValue TDBase = CGF.EmitLoadOfPointerLValue( | |||
4520 | CGF.GetAddrOfLocalVar(&TaskTypeArg), | |||
4521 | KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>()); | |||
4522 | const auto *KmpTaskTWithPrivatesQTyRD = | |||
4523 | cast<RecordDecl>(KmpTaskTWithPrivatesQTy->getAsTagDecl()); | |||
4524 | LValue Base = | |||
4525 | CGF.EmitLValueForField(TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
4526 | const auto *KmpTaskTQTyRD = cast<RecordDecl>(KmpTaskTQTy->getAsTagDecl()); | |||
4527 | auto PartIdFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTPartId); | |||
4528 | LValue PartIdLVal = CGF.EmitLValueForField(Base, *PartIdFI); | |||
4529 | llvm::Value *PartidParam = PartIdLVal.getPointer(); | |||
4530 | ||||
4531 | auto SharedsFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTShareds); | |||
4532 | LValue SharedsLVal = CGF.EmitLValueForField(Base, *SharedsFI); | |||
4533 | llvm::Value *SharedsParam = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4534 | CGF.EmitLoadOfScalar(SharedsLVal, Loc), | |||
4535 | CGF.ConvertTypeForMem(SharedsPtrTy)); | |||
4536 | ||||
4537 | auto PrivatesFI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin(), 1); | |||
4538 | llvm::Value *PrivatesParam; | |||
4539 | if (PrivatesFI != KmpTaskTWithPrivatesQTyRD->field_end()) { | |||
4540 | LValue PrivatesLVal = CGF.EmitLValueForField(TDBase, *PrivatesFI); | |||
4541 | PrivatesParam = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4542 | PrivatesLVal.getPointer(), CGF.VoidPtrTy); | |||
4543 | } else { | |||
4544 | PrivatesParam = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | |||
4545 | } | |||
4546 | ||||
4547 | llvm::Value *CommonArgs[] = {GtidParam, PartidParam, PrivatesParam, | |||
4548 | TaskPrivatesMap, | |||
4549 | CGF.Builder | |||
4550 | .CreatePointerBitCastOrAddrSpaceCast( | |||
4551 | TDBase.getAddress(), CGF.VoidPtrTy) | |||
4552 | .getPointer()}; | |||
4553 | SmallVector<llvm::Value *, 16> CallArgs(std::begin(CommonArgs), | |||
4554 | std::end(CommonArgs)); | |||
4555 | if (isOpenMPTaskLoopDirective(Kind)) { | |||
4556 | auto LBFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLowerBound); | |||
4557 | LValue LBLVal = CGF.EmitLValueForField(Base, *LBFI); | |||
4558 | llvm::Value *LBParam = CGF.EmitLoadOfScalar(LBLVal, Loc); | |||
4559 | auto UBFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTUpperBound); | |||
4560 | LValue UBLVal = CGF.EmitLValueForField(Base, *UBFI); | |||
4561 | llvm::Value *UBParam = CGF.EmitLoadOfScalar(UBLVal, Loc); | |||
4562 | auto StFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTStride); | |||
4563 | LValue StLVal = CGF.EmitLValueForField(Base, *StFI); | |||
4564 | llvm::Value *StParam = CGF.EmitLoadOfScalar(StLVal, Loc); | |||
4565 | auto LIFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLastIter); | |||
4566 | LValue LILVal = CGF.EmitLValueForField(Base, *LIFI); | |||
4567 | llvm::Value *LIParam = CGF.EmitLoadOfScalar(LILVal, Loc); | |||
4568 | auto RFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTReductions); | |||
4569 | LValue RLVal = CGF.EmitLValueForField(Base, *RFI); | |||
4570 | llvm::Value *RParam = CGF.EmitLoadOfScalar(RLVal, Loc); | |||
4571 | CallArgs.push_back(LBParam); | |||
4572 | CallArgs.push_back(UBParam); | |||
4573 | CallArgs.push_back(StParam); | |||
4574 | CallArgs.push_back(LIParam); | |||
4575 | CallArgs.push_back(RParam); | |||
4576 | } | |||
4577 | CallArgs.push_back(SharedsParam); | |||
4578 | ||||
4579 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, Loc, TaskFunction, | |||
4580 | CallArgs); | |||
4581 | CGF.EmitStoreThroughLValue(RValue::get(CGF.Builder.getInt32(/*C=*/0)), | |||
4582 | CGF.MakeAddrLValue(CGF.ReturnValue, KmpInt32Ty)); | |||
4583 | CGF.FinishFunction(); | |||
4584 | return TaskEntry; | |||
4585 | } | |||
4586 | ||||
4587 | static llvm::Value *emitDestructorsFunction(CodeGenModule &CGM, | |||
4588 | SourceLocation Loc, | |||
4589 | QualType KmpInt32Ty, | |||
4590 | QualType KmpTaskTWithPrivatesPtrQTy, | |||
4591 | QualType KmpTaskTWithPrivatesQTy) { | |||
4592 | ASTContext &C = CGM.getContext(); | |||
4593 | FunctionArgList Args; | |||
4594 | ImplicitParamDecl GtidArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, KmpInt32Ty, | |||
4595 | ImplicitParamDecl::Other); | |||
4596 | ImplicitParamDecl TaskTypeArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
4597 | KmpTaskTWithPrivatesPtrQTy.withRestrict(), | |||
4598 | ImplicitParamDecl::Other); | |||
4599 | Args.push_back(&GtidArg); | |||
4600 | Args.push_back(&TaskTypeArg); | |||
4601 | const auto &DestructorFnInfo = | |||
4602 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(KmpInt32Ty, Args); | |||
4603 | llvm::FunctionType *DestructorFnTy = | |||
4604 | CGM.getTypes().GetFunctionType(DestructorFnInfo); | |||
4605 | std::string Name = | |||
4606 | CGM.getOpenMPRuntime().getName({"omp_task_destructor", ""}); | |||
4607 | auto *DestructorFn = | |||
4608 | llvm::Function::Create(DestructorFnTy, llvm::GlobalValue::InternalLinkage, | |||
4609 | Name, &CGM.getModule()); | |||
4610 | CGM.SetInternalFunctionAttributes(GlobalDecl(), DestructorFn, | |||
4611 | DestructorFnInfo); | |||
4612 | DestructorFn->setDoesNotRecurse(); | |||
4613 | CodeGenFunction CGF(CGM); | |||
4614 | CGF.StartFunction(GlobalDecl(), KmpInt32Ty, DestructorFn, DestructorFnInfo, | |||
4615 | Args, Loc, Loc); | |||
4616 | ||||
4617 | LValue Base = CGF.EmitLoadOfPointerLValue( | |||
4618 | CGF.GetAddrOfLocalVar(&TaskTypeArg), | |||
4619 | KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>()); | |||
4620 | const auto *KmpTaskTWithPrivatesQTyRD = | |||
4621 | cast<RecordDecl>(KmpTaskTWithPrivatesQTy->getAsTagDecl()); | |||
4622 | auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
4623 | Base = CGF.EmitLValueForField(Base, *FI); | |||
4624 | for (const auto *Field : | |||
4625 | cast<RecordDecl>(FI->getType()->getAsTagDecl())->fields()) { | |||
4626 | if (QualType::DestructionKind DtorKind = | |||
4627 | Field->getType().isDestructedType()) { | |||
4628 | LValue FieldLValue = CGF.EmitLValueForField(Base, Field); | |||
4629 | CGF.pushDestroy(DtorKind, FieldLValue.getAddress(), Field->getType()); | |||
4630 | } | |||
4631 | } | |||
4632 | CGF.FinishFunction(); | |||
4633 | return DestructorFn; | |||
4634 | } | |||
4635 | ||||
4636 | /// Emit a privates mapping function for correct handling of private and | |||
4637 | /// firstprivate variables. | |||
4638 | /// \code | |||
4639 | /// void .omp_task_privates_map.(const .privates. *noalias privs, <ty1> | |||
4640 | /// **noalias priv1,..., <tyn> **noalias privn) { | |||
4641 | /// *priv1 = &.privates.priv1; | |||
4642 | /// ...; | |||
4643 | /// *privn = &.privates.privn; | |||
4644 | /// } | |||
4645 | /// \endcode | |||
4646 | static llvm::Value * | |||
4647 | emitTaskPrivateMappingFunction(CodeGenModule &CGM, SourceLocation Loc, | |||
4648 | ArrayRef<const Expr *> PrivateVars, | |||
4649 | ArrayRef<const Expr *> FirstprivateVars, | |||
4650 | ArrayRef<const Expr *> LastprivateVars, | |||
4651 | QualType PrivatesQTy, | |||
4652 | ArrayRef<PrivateDataTy> Privates) { | |||
4653 | ASTContext &C = CGM.getContext(); | |||
4654 | FunctionArgList Args; | |||
4655 | ImplicitParamDecl TaskPrivatesArg( | |||
4656 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
4657 | C.getPointerType(PrivatesQTy).withConst().withRestrict(), | |||
4658 | ImplicitParamDecl::Other); | |||
4659 | Args.push_back(&TaskPrivatesArg); | |||
4660 | llvm::DenseMap<const VarDecl *, unsigned> PrivateVarsPos; | |||
4661 | unsigned Counter = 1; | |||
4662 | for (const Expr *E : PrivateVars) { | |||
4663 | Args.push_back(ImplicitParamDecl::Create( | |||
4664 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
4665 | C.getPointerType(C.getPointerType(E->getType())) | |||
4666 | .withConst() | |||
4667 | .withRestrict(), | |||
4668 | ImplicitParamDecl::Other)); | |||
4669 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | |||
4670 | PrivateVarsPos[VD] = Counter; | |||
4671 | ++Counter; | |||
4672 | } | |||
4673 | for (const Expr *E : FirstprivateVars) { | |||
4674 | Args.push_back(ImplicitParamDecl::Create( | |||
4675 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
4676 | C.getPointerType(C.getPointerType(E->getType())) | |||
4677 | .withConst() | |||
4678 | .withRestrict(), | |||
4679 | ImplicitParamDecl::Other)); | |||
4680 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | |||
4681 | PrivateVarsPos[VD] = Counter; | |||
4682 | ++Counter; | |||
4683 | } | |||
4684 | for (const Expr *E : LastprivateVars) { | |||
4685 | Args.push_back(ImplicitParamDecl::Create( | |||
4686 | C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
4687 | C.getPointerType(C.getPointerType(E->getType())) | |||
4688 | .withConst() | |||
4689 | .withRestrict(), | |||
4690 | ImplicitParamDecl::Other)); | |||
4691 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | |||
4692 | PrivateVarsPos[VD] = Counter; | |||
4693 | ++Counter; | |||
4694 | } | |||
4695 | const auto &TaskPrivatesMapFnInfo = | |||
4696 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
4697 | llvm::FunctionType *TaskPrivatesMapTy = | |||
4698 | CGM.getTypes().GetFunctionType(TaskPrivatesMapFnInfo); | |||
4699 | std::string Name = | |||
4700 | CGM.getOpenMPRuntime().getName({"omp_task_privates_map", ""}); | |||
4701 | auto *TaskPrivatesMap = llvm::Function::Create( | |||
4702 | TaskPrivatesMapTy, llvm::GlobalValue::InternalLinkage, Name, | |||
4703 | &CGM.getModule()); | |||
4704 | CGM.SetInternalFunctionAttributes(GlobalDecl(), TaskPrivatesMap, | |||
4705 | TaskPrivatesMapFnInfo); | |||
4706 | if (CGM.getLangOpts().Optimize) { | |||
4707 | TaskPrivatesMap->removeFnAttr(llvm::Attribute::NoInline); | |||
4708 | TaskPrivatesMap->removeFnAttr(llvm::Attribute::OptimizeNone); | |||
4709 | TaskPrivatesMap->addFnAttr(llvm::Attribute::AlwaysInline); | |||
4710 | } | |||
4711 | CodeGenFunction CGF(CGM); | |||
4712 | CGF.StartFunction(GlobalDecl(), C.VoidTy, TaskPrivatesMap, | |||
4713 | TaskPrivatesMapFnInfo, Args, Loc, Loc); | |||
4714 | ||||
4715 | // *privi = &.privates.privi; | |||
4716 | LValue Base = CGF.EmitLoadOfPointerLValue( | |||
4717 | CGF.GetAddrOfLocalVar(&TaskPrivatesArg), | |||
4718 | TaskPrivatesArg.getType()->castAs<PointerType>()); | |||
4719 | const auto *PrivatesQTyRD = cast<RecordDecl>(PrivatesQTy->getAsTagDecl()); | |||
4720 | Counter = 0; | |||
4721 | for (const FieldDecl *Field : PrivatesQTyRD->fields()) { | |||
4722 | LValue FieldLVal = CGF.EmitLValueForField(Base, Field); | |||
4723 | const VarDecl *VD = Args[PrivateVarsPos[Privates[Counter].second.Original]]; | |||
4724 | LValue RefLVal = | |||
4725 | CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(VD), VD->getType()); | |||
4726 | LValue RefLoadLVal = CGF.EmitLoadOfPointerLValue( | |||
4727 | RefLVal.getAddress(), RefLVal.getType()->castAs<PointerType>()); | |||
4728 | CGF.EmitStoreOfScalar(FieldLVal.getPointer(), RefLoadLVal); | |||
4729 | ++Counter; | |||
4730 | } | |||
4731 | CGF.FinishFunction(); | |||
4732 | return TaskPrivatesMap; | |||
4733 | } | |||
4734 | ||||
4735 | /// Emit initialization for private variables in task-based directives. | |||
4736 | static void emitPrivatesInit(CodeGenFunction &CGF, | |||
4737 | const OMPExecutableDirective &D, | |||
4738 | Address KmpTaskSharedsPtr, LValue TDBase, | |||
4739 | const RecordDecl *KmpTaskTWithPrivatesQTyRD, | |||
4740 | QualType SharedsTy, QualType SharedsPtrTy, | |||
4741 | const OMPTaskDataTy &Data, | |||
4742 | ArrayRef<PrivateDataTy> Privates, bool ForDup) { | |||
4743 | ASTContext &C = CGF.getContext(); | |||
4744 | auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
4745 | LValue PrivatesBase = CGF.EmitLValueForField(TDBase, *FI); | |||
4746 | OpenMPDirectiveKind Kind = isOpenMPTaskLoopDirective(D.getDirectiveKind()) | |||
4747 | ? OMPD_taskloop | |||
4748 | : OMPD_task; | |||
4749 | const CapturedStmt &CS = *D.getCapturedStmt(Kind); | |||
4750 | CodeGenFunction::CGCapturedStmtInfo CapturesInfo(CS); | |||
4751 | LValue SrcBase; | |||
4752 | bool IsTargetTask = | |||
4753 | isOpenMPTargetDataManagementDirective(D.getDirectiveKind()) || | |||
4754 | isOpenMPTargetExecutionDirective(D.getDirectiveKind()); | |||
4755 | // For target-based directives skip 3 firstprivate arrays BasePointersArray, | |||
4756 | // PointersArray and SizesArray. The original variables for these arrays are | |||
4757 | // not captured and we get their addresses explicitly. | |||
4758 | if ((!IsTargetTask && !Data.FirstprivateVars.empty()) || | |||
4759 | (IsTargetTask && KmpTaskSharedsPtr.isValid())) { | |||
4760 | SrcBase = CGF.MakeAddrLValue( | |||
4761 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
4762 | KmpTaskSharedsPtr, CGF.ConvertTypeForMem(SharedsPtrTy)), | |||
4763 | SharedsTy); | |||
4764 | } | |||
4765 | FI = cast<RecordDecl>(FI->getType()->getAsTagDecl())->field_begin(); | |||
4766 | for (const PrivateDataTy &Pair : Privates) { | |||
4767 | const VarDecl *VD = Pair.second.PrivateCopy; | |||
4768 | const Expr *Init = VD->getAnyInitializer(); | |||
4769 | if (Init && (!ForDup || (isa<CXXConstructExpr>(Init) && | |||
4770 | !CGF.isTrivialInitializer(Init)))) { | |||
4771 | LValue PrivateLValue = CGF.EmitLValueForField(PrivatesBase, *FI); | |||
4772 | if (const VarDecl *Elem = Pair.second.PrivateElemInit) { | |||
4773 | const VarDecl *OriginalVD = Pair.second.Original; | |||
4774 | // Check if the variable is the target-based BasePointersArray, | |||
4775 | // PointersArray or SizesArray. | |||
4776 | LValue SharedRefLValue; | |||
4777 | QualType Type = PrivateLValue.getType(); | |||
4778 | const FieldDecl *SharedField = CapturesInfo.lookup(OriginalVD); | |||
4779 | if (IsTargetTask && !SharedField) { | |||
4780 | assert(isa<ImplicitParamDecl>(OriginalVD) &&((isa<ImplicitParamDecl>(OriginalVD) && isa< CapturedDecl>(OriginalVD->getDeclContext()) && cast <CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams () == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl >(OriginalVD->getDeclContext()) ->getDeclContext()) && "Expected artificial target data variable.") ? static_cast< void> (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4787, __PRETTY_FUNCTION__)) | |||
4781 | isa<CapturedDecl>(OriginalVD->getDeclContext()) &&((isa<ImplicitParamDecl>(OriginalVD) && isa< CapturedDecl>(OriginalVD->getDeclContext()) && cast <CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams () == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl >(OriginalVD->getDeclContext()) ->getDeclContext()) && "Expected artificial target data variable.") ? static_cast< void> (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4787, __PRETTY_FUNCTION__)) | |||
4782 | cast<CapturedDecl>(OriginalVD->getDeclContext())((isa<ImplicitParamDecl>(OriginalVD) && isa< CapturedDecl>(OriginalVD->getDeclContext()) && cast <CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams () == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl >(OriginalVD->getDeclContext()) ->getDeclContext()) && "Expected artificial target data variable.") ? static_cast< void> (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4787, __PRETTY_FUNCTION__)) | |||
4783 | ->getNumParams() == 0 &&((isa<ImplicitParamDecl>(OriginalVD) && isa< CapturedDecl>(OriginalVD->getDeclContext()) && cast <CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams () == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl >(OriginalVD->getDeclContext()) ->getDeclContext()) && "Expected artificial target data variable.") ? static_cast< void> (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4787, __PRETTY_FUNCTION__)) | |||
4784 | isa<TranslationUnitDecl>(((isa<ImplicitParamDecl>(OriginalVD) && isa< CapturedDecl>(OriginalVD->getDeclContext()) && cast <CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams () == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl >(OriginalVD->getDeclContext()) ->getDeclContext()) && "Expected artificial target data variable.") ? static_cast< void> (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4787, __PRETTY_FUNCTION__)) | |||
4785 | cast<CapturedDecl>(OriginalVD->getDeclContext())((isa<ImplicitParamDecl>(OriginalVD) && isa< CapturedDecl>(OriginalVD->getDeclContext()) && cast <CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams () == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl >(OriginalVD->getDeclContext()) ->getDeclContext()) && "Expected artificial target data variable.") ? static_cast< void> (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4787, __PRETTY_FUNCTION__)) | |||
4786 | ->getDeclContext()) &&((isa<ImplicitParamDecl>(OriginalVD) && isa< CapturedDecl>(OriginalVD->getDeclContext()) && cast <CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams () == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl >(OriginalVD->getDeclContext()) ->getDeclContext()) && "Expected artificial target data variable.") ? static_cast< void> (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4787, __PRETTY_FUNCTION__)) | |||
4787 | "Expected artificial target data variable.")((isa<ImplicitParamDecl>(OriginalVD) && isa< CapturedDecl>(OriginalVD->getDeclContext()) && cast <CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams () == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl >(OriginalVD->getDeclContext()) ->getDeclContext()) && "Expected artificial target data variable.") ? static_cast< void> (0) : __assert_fail ("isa<ImplicitParamDecl>(OriginalVD) && isa<CapturedDecl>(OriginalVD->getDeclContext()) && cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getNumParams() == 0 && isa<TranslationUnitDecl>( cast<CapturedDecl>(OriginalVD->getDeclContext()) ->getDeclContext()) && \"Expected artificial target data variable.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4787, __PRETTY_FUNCTION__)); | |||
4788 | SharedRefLValue = | |||
4789 | CGF.MakeAddrLValue(CGF.GetAddrOfLocalVar(OriginalVD), Type); | |||
4790 | } else { | |||
4791 | SharedRefLValue = CGF.EmitLValueForField(SrcBase, SharedField); | |||
4792 | SharedRefLValue = CGF.MakeAddrLValue( | |||
4793 | Address(SharedRefLValue.getPointer(), C.getDeclAlign(OriginalVD)), | |||
4794 | SharedRefLValue.getType(), LValueBaseInfo(AlignmentSource::Decl), | |||
4795 | SharedRefLValue.getTBAAInfo()); | |||
4796 | } | |||
4797 | if (Type->isArrayType()) { | |||
4798 | // Initialize firstprivate array. | |||
4799 | if (!isa<CXXConstructExpr>(Init) || CGF.isTrivialInitializer(Init)) { | |||
4800 | // Perform simple memcpy. | |||
4801 | CGF.EmitAggregateAssign(PrivateLValue, SharedRefLValue, Type); | |||
4802 | } else { | |||
4803 | // Initialize firstprivate array using element-by-element | |||
4804 | // initialization. | |||
4805 | CGF.EmitOMPAggregateAssign( | |||
4806 | PrivateLValue.getAddress(), SharedRefLValue.getAddress(), Type, | |||
4807 | [&CGF, Elem, Init, &CapturesInfo](Address DestElement, | |||
4808 | Address SrcElement) { | |||
4809 | // Clean up any temporaries needed by the initialization. | |||
4810 | CodeGenFunction::OMPPrivateScope InitScope(CGF); | |||
4811 | InitScope.addPrivate( | |||
4812 | Elem, [SrcElement]() -> Address { return SrcElement; }); | |||
4813 | (void)InitScope.Privatize(); | |||
4814 | // Emit initialization for single element. | |||
4815 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII( | |||
4816 | CGF, &CapturesInfo); | |||
4817 | CGF.EmitAnyExprToMem(Init, DestElement, | |||
4818 | Init->getType().getQualifiers(), | |||
4819 | /*IsInitializer=*/false); | |||
4820 | }); | |||
4821 | } | |||
4822 | } else { | |||
4823 | CodeGenFunction::OMPPrivateScope InitScope(CGF); | |||
4824 | InitScope.addPrivate(Elem, [SharedRefLValue]() -> Address { | |||
4825 | return SharedRefLValue.getAddress(); | |||
4826 | }); | |||
4827 | (void)InitScope.Privatize(); | |||
4828 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CapturesInfo); | |||
4829 | CGF.EmitExprAsInit(Init, VD, PrivateLValue, | |||
4830 | /*capturedByInit=*/false); | |||
4831 | } | |||
4832 | } else { | |||
4833 | CGF.EmitExprAsInit(Init, VD, PrivateLValue, /*capturedByInit=*/false); | |||
4834 | } | |||
4835 | } | |||
4836 | ++FI; | |||
4837 | } | |||
4838 | } | |||
4839 | ||||
4840 | /// Check if duplication function is required for taskloops. | |||
4841 | static bool checkInitIsRequired(CodeGenFunction &CGF, | |||
4842 | ArrayRef<PrivateDataTy> Privates) { | |||
4843 | bool InitRequired = false; | |||
4844 | for (const PrivateDataTy &Pair : Privates) { | |||
4845 | const VarDecl *VD = Pair.second.PrivateCopy; | |||
4846 | const Expr *Init = VD->getAnyInitializer(); | |||
4847 | InitRequired = InitRequired || (Init && isa<CXXConstructExpr>(Init) && | |||
4848 | !CGF.isTrivialInitializer(Init)); | |||
4849 | if (InitRequired) | |||
4850 | break; | |||
4851 | } | |||
4852 | return InitRequired; | |||
4853 | } | |||
4854 | ||||
4855 | ||||
4856 | /// Emit task_dup function (for initialization of | |||
4857 | /// private/firstprivate/lastprivate vars and last_iter flag) | |||
4858 | /// \code | |||
4859 | /// void __task_dup_entry(kmp_task_t *task_dst, const kmp_task_t *task_src, int | |||
4860 | /// lastpriv) { | |||
4861 | /// // setup lastprivate flag | |||
4862 | /// task_dst->last = lastpriv; | |||
4863 | /// // could be constructor calls here... | |||
4864 | /// } | |||
4865 | /// \endcode | |||
4866 | static llvm::Value * | |||
4867 | emitTaskDupFunction(CodeGenModule &CGM, SourceLocation Loc, | |||
4868 | const OMPExecutableDirective &D, | |||
4869 | QualType KmpTaskTWithPrivatesPtrQTy, | |||
4870 | const RecordDecl *KmpTaskTWithPrivatesQTyRD, | |||
4871 | const RecordDecl *KmpTaskTQTyRD, QualType SharedsTy, | |||
4872 | QualType SharedsPtrTy, const OMPTaskDataTy &Data, | |||
4873 | ArrayRef<PrivateDataTy> Privates, bool WithLastIter) { | |||
4874 | ASTContext &C = CGM.getContext(); | |||
4875 | FunctionArgList Args; | |||
4876 | ImplicitParamDecl DstArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
4877 | KmpTaskTWithPrivatesPtrQTy, | |||
4878 | ImplicitParamDecl::Other); | |||
4879 | ImplicitParamDecl SrcArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
4880 | KmpTaskTWithPrivatesPtrQTy, | |||
4881 | ImplicitParamDecl::Other); | |||
4882 | ImplicitParamDecl LastprivArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.IntTy, | |||
4883 | ImplicitParamDecl::Other); | |||
4884 | Args.push_back(&DstArg); | |||
4885 | Args.push_back(&SrcArg); | |||
4886 | Args.push_back(&LastprivArg); | |||
4887 | const auto &TaskDupFnInfo = | |||
4888 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
4889 | llvm::FunctionType *TaskDupTy = CGM.getTypes().GetFunctionType(TaskDupFnInfo); | |||
4890 | std::string Name = CGM.getOpenMPRuntime().getName({"omp_task_dup", ""}); | |||
4891 | auto *TaskDup = llvm::Function::Create( | |||
4892 | TaskDupTy, llvm::GlobalValue::InternalLinkage, Name, &CGM.getModule()); | |||
4893 | CGM.SetInternalFunctionAttributes(GlobalDecl(), TaskDup, TaskDupFnInfo); | |||
4894 | TaskDup->setDoesNotRecurse(); | |||
4895 | CodeGenFunction CGF(CGM); | |||
4896 | CGF.StartFunction(GlobalDecl(), C.VoidTy, TaskDup, TaskDupFnInfo, Args, Loc, | |||
4897 | Loc); | |||
4898 | ||||
4899 | LValue TDBase = CGF.EmitLoadOfPointerLValue( | |||
4900 | CGF.GetAddrOfLocalVar(&DstArg), | |||
4901 | KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>()); | |||
4902 | // task_dst->liter = lastpriv; | |||
4903 | if (WithLastIter) { | |||
4904 | auto LIFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTLastIter); | |||
4905 | LValue Base = CGF.EmitLValueForField( | |||
4906 | TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
4907 | LValue LILVal = CGF.EmitLValueForField(Base, *LIFI); | |||
4908 | llvm::Value *Lastpriv = CGF.EmitLoadOfScalar( | |||
4909 | CGF.GetAddrOfLocalVar(&LastprivArg), /*Volatile=*/false, C.IntTy, Loc); | |||
4910 | CGF.EmitStoreOfScalar(Lastpriv, LILVal); | |||
4911 | } | |||
4912 | ||||
4913 | // Emit initial values for private copies (if any). | |||
4914 | assert(!Privates.empty())((!Privates.empty()) ? static_cast<void> (0) : __assert_fail ("!Privates.empty()", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 4914, __PRETTY_FUNCTION__)); | |||
4915 | Address KmpTaskSharedsPtr = Address::invalid(); | |||
4916 | if (!Data.FirstprivateVars.empty()) { | |||
4917 | LValue TDBase = CGF.EmitLoadOfPointerLValue( | |||
4918 | CGF.GetAddrOfLocalVar(&SrcArg), | |||
4919 | KmpTaskTWithPrivatesPtrQTy->castAs<PointerType>()); | |||
4920 | LValue Base = CGF.EmitLValueForField( | |||
4921 | TDBase, *KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
4922 | KmpTaskSharedsPtr = Address( | |||
4923 | CGF.EmitLoadOfScalar(CGF.EmitLValueForField( | |||
4924 | Base, *std::next(KmpTaskTQTyRD->field_begin(), | |||
4925 | KmpTaskTShareds)), | |||
4926 | Loc), | |||
4927 | CGF.getNaturalTypeAlignment(SharedsTy)); | |||
4928 | } | |||
4929 | emitPrivatesInit(CGF, D, KmpTaskSharedsPtr, TDBase, KmpTaskTWithPrivatesQTyRD, | |||
4930 | SharedsTy, SharedsPtrTy, Data, Privates, /*ForDup=*/true); | |||
4931 | CGF.FinishFunction(); | |||
4932 | return TaskDup; | |||
4933 | } | |||
4934 | ||||
4935 | /// Checks if destructor function is required to be generated. | |||
4936 | /// \return true if cleanups are required, false otherwise. | |||
4937 | static bool | |||
4938 | checkDestructorsRequired(const RecordDecl *KmpTaskTWithPrivatesQTyRD) { | |||
4939 | bool NeedsCleanup = false; | |||
4940 | auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin(), 1); | |||
4941 | const auto *PrivateRD = cast<RecordDecl>(FI->getType()->getAsTagDecl()); | |||
4942 | for (const FieldDecl *FD : PrivateRD->fields()) { | |||
4943 | NeedsCleanup = NeedsCleanup || FD->getType().isDestructedType(); | |||
4944 | if (NeedsCleanup) | |||
4945 | break; | |||
4946 | } | |||
4947 | return NeedsCleanup; | |||
4948 | } | |||
4949 | ||||
4950 | CGOpenMPRuntime::TaskResultTy | |||
4951 | CGOpenMPRuntime::emitTaskInit(CodeGenFunction &CGF, SourceLocation Loc, | |||
4952 | const OMPExecutableDirective &D, | |||
4953 | llvm::Function *TaskFunction, QualType SharedsTy, | |||
4954 | Address Shareds, const OMPTaskDataTy &Data) { | |||
4955 | ASTContext &C = CGM.getContext(); | |||
4956 | llvm::SmallVector<PrivateDataTy, 4> Privates; | |||
4957 | // Aggregate privates and sort them by the alignment. | |||
4958 | auto I = Data.PrivateCopies.begin(); | |||
4959 | for (const Expr *E : Data.PrivateVars) { | |||
4960 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | |||
4961 | Privates.emplace_back( | |||
4962 | C.getDeclAlign(VD), | |||
4963 | PrivateHelpersTy(VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()), | |||
4964 | /*PrivateElemInit=*/nullptr)); | |||
4965 | ++I; | |||
4966 | } | |||
4967 | I = Data.FirstprivateCopies.begin(); | |||
4968 | auto IElemInitRef = Data.FirstprivateInits.begin(); | |||
4969 | for (const Expr *E : Data.FirstprivateVars) { | |||
4970 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | |||
4971 | Privates.emplace_back( | |||
4972 | C.getDeclAlign(VD), | |||
4973 | PrivateHelpersTy( | |||
4974 | VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()), | |||
4975 | cast<VarDecl>(cast<DeclRefExpr>(*IElemInitRef)->getDecl()))); | |||
4976 | ++I; | |||
4977 | ++IElemInitRef; | |||
4978 | } | |||
4979 | I = Data.LastprivateCopies.begin(); | |||
4980 | for (const Expr *E : Data.LastprivateVars) { | |||
4981 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); | |||
4982 | Privates.emplace_back( | |||
4983 | C.getDeclAlign(VD), | |||
4984 | PrivateHelpersTy(VD, cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()), | |||
4985 | /*PrivateElemInit=*/nullptr)); | |||
4986 | ++I; | |||
4987 | } | |||
4988 | llvm::stable_sort(Privates, [](PrivateDataTy L, PrivateDataTy R) { | |||
4989 | return L.first > R.first; | |||
4990 | }); | |||
4991 | QualType KmpInt32Ty = C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1); | |||
4992 | // Build type kmp_routine_entry_t (if not built yet). | |||
4993 | emitKmpRoutineEntryT(KmpInt32Ty); | |||
4994 | // Build type kmp_task_t (if not built yet). | |||
4995 | if (isOpenMPTaskLoopDirective(D.getDirectiveKind())) { | |||
4996 | if (SavedKmpTaskloopTQTy.isNull()) { | |||
4997 | SavedKmpTaskloopTQTy = C.getRecordType(createKmpTaskTRecordDecl( | |||
4998 | CGM, D.getDirectiveKind(), KmpInt32Ty, KmpRoutineEntryPtrQTy)); | |||
4999 | } | |||
5000 | KmpTaskTQTy = SavedKmpTaskloopTQTy; | |||
5001 | } else { | |||
5002 | assert((D.getDirectiveKind() == OMPD_task ||(((D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective (D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective (D.getDirectiveKind())) && "Expected taskloop, task or target directive" ) ? static_cast<void> (0) : __assert_fail ("(D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && \"Expected taskloop, task or target directive\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 5005, __PRETTY_FUNCTION__)) | |||
5003 | isOpenMPTargetExecutionDirective(D.getDirectiveKind()) ||(((D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective (D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective (D.getDirectiveKind())) && "Expected taskloop, task or target directive" ) ? static_cast<void> (0) : __assert_fail ("(D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && \"Expected taskloop, task or target directive\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 5005, __PRETTY_FUNCTION__)) | |||
5004 | isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) &&(((D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective (D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective (D.getDirectiveKind())) && "Expected taskloop, task or target directive" ) ? static_cast<void> (0) : __assert_fail ("(D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && \"Expected taskloop, task or target directive\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 5005, __PRETTY_FUNCTION__)) | |||
5005 | "Expected taskloop, task or target directive")(((D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective (D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective (D.getDirectiveKind())) && "Expected taskloop, task or target directive" ) ? static_cast<void> (0) : __assert_fail ("(D.getDirectiveKind() == OMPD_task || isOpenMPTargetExecutionDirective(D.getDirectiveKind()) || isOpenMPTargetDataManagementDirective(D.getDirectiveKind())) && \"Expected taskloop, task or target directive\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 5005, __PRETTY_FUNCTION__)); | |||
5006 | if (SavedKmpTaskTQTy.isNull()) { | |||
5007 | SavedKmpTaskTQTy = C.getRecordType(createKmpTaskTRecordDecl( | |||
5008 | CGM, D.getDirectiveKind(), KmpInt32Ty, KmpRoutineEntryPtrQTy)); | |||
5009 | } | |||
5010 | KmpTaskTQTy = SavedKmpTaskTQTy; | |||
5011 | } | |||
5012 | const auto *KmpTaskTQTyRD = cast<RecordDecl>(KmpTaskTQTy->getAsTagDecl()); | |||
5013 | // Build particular struct kmp_task_t for the given task. | |||
5014 | const RecordDecl *KmpTaskTWithPrivatesQTyRD = | |||
5015 | createKmpTaskTWithPrivatesRecordDecl(CGM, KmpTaskTQTy, Privates); | |||
5016 | QualType KmpTaskTWithPrivatesQTy = C.getRecordType(KmpTaskTWithPrivatesQTyRD); | |||
5017 | QualType KmpTaskTWithPrivatesPtrQTy = | |||
5018 | C.getPointerType(KmpTaskTWithPrivatesQTy); | |||
5019 | llvm::Type *KmpTaskTWithPrivatesTy = CGF.ConvertType(KmpTaskTWithPrivatesQTy); | |||
5020 | llvm::Type *KmpTaskTWithPrivatesPtrTy = | |||
5021 | KmpTaskTWithPrivatesTy->getPointerTo(); | |||
5022 | llvm::Value *KmpTaskTWithPrivatesTySize = | |||
5023 | CGF.getTypeSize(KmpTaskTWithPrivatesQTy); | |||
5024 | QualType SharedsPtrTy = C.getPointerType(SharedsTy); | |||
5025 | ||||
5026 | // Emit initial values for private copies (if any). | |||
5027 | llvm::Value *TaskPrivatesMap = nullptr; | |||
5028 | llvm::Type *TaskPrivatesMapTy = | |||
5029 | std::next(TaskFunction->arg_begin(), 3)->getType(); | |||
5030 | if (!Privates.empty()) { | |||
5031 | auto FI = std::next(KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
5032 | TaskPrivatesMap = emitTaskPrivateMappingFunction( | |||
5033 | CGM, Loc, Data.PrivateVars, Data.FirstprivateVars, Data.LastprivateVars, | |||
5034 | FI->getType(), Privates); | |||
5035 | TaskPrivatesMap = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5036 | TaskPrivatesMap, TaskPrivatesMapTy); | |||
5037 | } else { | |||
5038 | TaskPrivatesMap = llvm::ConstantPointerNull::get( | |||
5039 | cast<llvm::PointerType>(TaskPrivatesMapTy)); | |||
5040 | } | |||
5041 | // Build a proxy function kmp_int32 .omp_task_entry.(kmp_int32 gtid, | |||
5042 | // kmp_task_t *tt); | |||
5043 | llvm::Function *TaskEntry = emitProxyTaskFunction( | |||
5044 | CGM, Loc, D.getDirectiveKind(), KmpInt32Ty, KmpTaskTWithPrivatesPtrQTy, | |||
5045 | KmpTaskTWithPrivatesQTy, KmpTaskTQTy, SharedsPtrTy, TaskFunction, | |||
5046 | TaskPrivatesMap); | |||
5047 | ||||
5048 | // Build call kmp_task_t * __kmpc_omp_task_alloc(ident_t *, kmp_int32 gtid, | |||
5049 | // kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, | |||
5050 | // kmp_routine_entry_t *task_entry); | |||
5051 | // Task flags. Format is taken from | |||
5052 | // https://github.com/llvm/llvm-project/blob/master/openmp/runtime/src/kmp.h, | |||
5053 | // description of kmp_tasking_flags struct. | |||
5054 | enum { | |||
5055 | TiedFlag = 0x1, | |||
5056 | FinalFlag = 0x2, | |||
5057 | DestructorsFlag = 0x8, | |||
5058 | PriorityFlag = 0x20 | |||
5059 | }; | |||
5060 | unsigned Flags = Data.Tied ? TiedFlag : 0; | |||
5061 | bool NeedsCleanup = false; | |||
5062 | if (!Privates.empty()) { | |||
5063 | NeedsCleanup = checkDestructorsRequired(KmpTaskTWithPrivatesQTyRD); | |||
5064 | if (NeedsCleanup) | |||
5065 | Flags = Flags | DestructorsFlag; | |||
5066 | } | |||
5067 | if (Data.Priority.getInt()) | |||
5068 | Flags = Flags | PriorityFlag; | |||
5069 | llvm::Value *TaskFlags = | |||
5070 | Data.Final.getPointer() | |||
5071 | ? CGF.Builder.CreateSelect(Data.Final.getPointer(), | |||
5072 | CGF.Builder.getInt32(FinalFlag), | |||
5073 | CGF.Builder.getInt32(/*C=*/0)) | |||
5074 | : CGF.Builder.getInt32(Data.Final.getInt() ? FinalFlag : 0); | |||
5075 | TaskFlags = CGF.Builder.CreateOr(TaskFlags, CGF.Builder.getInt32(Flags)); | |||
5076 | llvm::Value *SharedsSize = CGM.getSize(C.getTypeSizeInChars(SharedsTy)); | |||
5077 | llvm::Value *AllocArgs[] = {emitUpdateLocation(CGF, Loc), | |||
5078 | getThreadID(CGF, Loc), TaskFlags, | |||
5079 | KmpTaskTWithPrivatesTySize, SharedsSize, | |||
5080 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5081 | TaskEntry, KmpRoutineEntryPtrTy)}; | |||
5082 | llvm::Value *NewTask = CGF.EmitRuntimeCall( | |||
5083 | createRuntimeFunction(OMPRTL__kmpc_omp_task_alloc), AllocArgs); | |||
5084 | llvm::Value *NewTaskNewTaskTTy = | |||
5085 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5086 | NewTask, KmpTaskTWithPrivatesPtrTy); | |||
5087 | LValue Base = CGF.MakeNaturalAlignAddrLValue(NewTaskNewTaskTTy, | |||
5088 | KmpTaskTWithPrivatesQTy); | |||
5089 | LValue TDBase = | |||
5090 | CGF.EmitLValueForField(Base, *KmpTaskTWithPrivatesQTyRD->field_begin()); | |||
5091 | // Fill the data in the resulting kmp_task_t record. | |||
5092 | // Copy shareds if there are any. | |||
5093 | Address KmpTaskSharedsPtr = Address::invalid(); | |||
5094 | if (!SharedsTy->getAsStructureType()->getDecl()->field_empty()) { | |||
5095 | KmpTaskSharedsPtr = | |||
5096 | Address(CGF.EmitLoadOfScalar( | |||
5097 | CGF.EmitLValueForField( | |||
5098 | TDBase, *std::next(KmpTaskTQTyRD->field_begin(), | |||
5099 | KmpTaskTShareds)), | |||
5100 | Loc), | |||
5101 | CGF.getNaturalTypeAlignment(SharedsTy)); | |||
5102 | LValue Dest = CGF.MakeAddrLValue(KmpTaskSharedsPtr, SharedsTy); | |||
5103 | LValue Src = CGF.MakeAddrLValue(Shareds, SharedsTy); | |||
5104 | CGF.EmitAggregateCopy(Dest, Src, SharedsTy, AggValueSlot::DoesNotOverlap); | |||
5105 | } | |||
5106 | // Emit initial values for private copies (if any). | |||
5107 | TaskResultTy Result; | |||
5108 | if (!Privates.empty()) { | |||
5109 | emitPrivatesInit(CGF, D, KmpTaskSharedsPtr, Base, KmpTaskTWithPrivatesQTyRD, | |||
5110 | SharedsTy, SharedsPtrTy, Data, Privates, | |||
5111 | /*ForDup=*/false); | |||
5112 | if (isOpenMPTaskLoopDirective(D.getDirectiveKind()) && | |||
5113 | (!Data.LastprivateVars.empty() || checkInitIsRequired(CGF, Privates))) { | |||
5114 | Result.TaskDupFn = emitTaskDupFunction( | |||
5115 | CGM, Loc, D, KmpTaskTWithPrivatesPtrQTy, KmpTaskTWithPrivatesQTyRD, | |||
5116 | KmpTaskTQTyRD, SharedsTy, SharedsPtrTy, Data, Privates, | |||
5117 | /*WithLastIter=*/!Data.LastprivateVars.empty()); | |||
5118 | } | |||
5119 | } | |||
5120 | // Fields of union "kmp_cmplrdata_t" for destructors and priority. | |||
5121 | enum { Priority = 0, Destructors = 1 }; | |||
5122 | // Provide pointer to function with destructors for privates. | |||
5123 | auto FI = std::next(KmpTaskTQTyRD->field_begin(), Data1); | |||
5124 | const RecordDecl *KmpCmplrdataUD = | |||
5125 | (*FI)->getType()->getAsUnionType()->getDecl(); | |||
5126 | if (NeedsCleanup) { | |||
5127 | llvm::Value *DestructorFn = emitDestructorsFunction( | |||
5128 | CGM, Loc, KmpInt32Ty, KmpTaskTWithPrivatesPtrQTy, | |||
5129 | KmpTaskTWithPrivatesQTy); | |||
5130 | LValue Data1LV = CGF.EmitLValueForField(TDBase, *FI); | |||
5131 | LValue DestructorsLV = CGF.EmitLValueForField( | |||
5132 | Data1LV, *std::next(KmpCmplrdataUD->field_begin(), Destructors)); | |||
5133 | CGF.EmitStoreOfScalar(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5134 | DestructorFn, KmpRoutineEntryPtrTy), | |||
5135 | DestructorsLV); | |||
5136 | } | |||
5137 | // Set priority. | |||
5138 | if (Data.Priority.getInt()) { | |||
5139 | LValue Data2LV = CGF.EmitLValueForField( | |||
5140 | TDBase, *std::next(KmpTaskTQTyRD->field_begin(), Data2)); | |||
5141 | LValue PriorityLV = CGF.EmitLValueForField( | |||
5142 | Data2LV, *std::next(KmpCmplrdataUD->field_begin(), Priority)); | |||
5143 | CGF.EmitStoreOfScalar(Data.Priority.getPointer(), PriorityLV); | |||
5144 | } | |||
5145 | Result.NewTask = NewTask; | |||
5146 | Result.TaskEntry = TaskEntry; | |||
5147 | Result.NewTaskNewTaskTTy = NewTaskNewTaskTTy; | |||
5148 | Result.TDBase = TDBase; | |||
5149 | Result.KmpTaskTQTyRD = KmpTaskTQTyRD; | |||
5150 | return Result; | |||
5151 | } | |||
5152 | ||||
5153 | void CGOpenMPRuntime::emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
5154 | const OMPExecutableDirective &D, | |||
5155 | llvm::Function *TaskFunction, | |||
5156 | QualType SharedsTy, Address Shareds, | |||
5157 | const Expr *IfCond, | |||
5158 | const OMPTaskDataTy &Data) { | |||
5159 | if (!CGF.HaveInsertPoint()) | |||
5160 | return; | |||
5161 | ||||
5162 | TaskResultTy Result = | |||
5163 | emitTaskInit(CGF, Loc, D, TaskFunction, SharedsTy, Shareds, Data); | |||
5164 | llvm::Value *NewTask = Result.NewTask; | |||
5165 | llvm::Function *TaskEntry = Result.TaskEntry; | |||
5166 | llvm::Value *NewTaskNewTaskTTy = Result.NewTaskNewTaskTTy; | |||
5167 | LValue TDBase = Result.TDBase; | |||
5168 | const RecordDecl *KmpTaskTQTyRD = Result.KmpTaskTQTyRD; | |||
5169 | ASTContext &C = CGM.getContext(); | |||
5170 | // Process list of dependences. | |||
5171 | Address DependenciesArray = Address::invalid(); | |||
5172 | unsigned NumDependencies = Data.Dependences.size(); | |||
5173 | if (NumDependencies) { | |||
5174 | // Dependence kind for RTL. | |||
5175 | enum RTLDependenceKindTy { DepIn = 0x01, DepInOut = 0x3, DepMutexInOutSet = 0x4 }; | |||
5176 | enum RTLDependInfoFieldsTy { BaseAddr, Len, Flags }; | |||
5177 | RecordDecl *KmpDependInfoRD; | |||
5178 | QualType FlagsTy = | |||
5179 | C.getIntTypeForBitwidth(C.getTypeSize(C.BoolTy), /*Signed=*/false); | |||
5180 | llvm::Type *LLVMFlagsTy = CGF.ConvertTypeForMem(FlagsTy); | |||
5181 | if (KmpDependInfoTy.isNull()) { | |||
5182 | KmpDependInfoRD = C.buildImplicitRecord("kmp_depend_info"); | |||
5183 | KmpDependInfoRD->startDefinition(); | |||
5184 | addFieldToRecordDecl(C, KmpDependInfoRD, C.getIntPtrType()); | |||
5185 | addFieldToRecordDecl(C, KmpDependInfoRD, C.getSizeType()); | |||
5186 | addFieldToRecordDecl(C, KmpDependInfoRD, FlagsTy); | |||
5187 | KmpDependInfoRD->completeDefinition(); | |||
5188 | KmpDependInfoTy = C.getRecordType(KmpDependInfoRD); | |||
5189 | } else { | |||
5190 | KmpDependInfoRD = cast<RecordDecl>(KmpDependInfoTy->getAsTagDecl()); | |||
5191 | } | |||
5192 | // Define type kmp_depend_info[<Dependences.size()>]; | |||
5193 | QualType KmpDependInfoArrayTy = C.getConstantArrayType( | |||
5194 | KmpDependInfoTy, llvm::APInt(/*numBits=*/64, NumDependencies), | |||
5195 | ArrayType::Normal, /*IndexTypeQuals=*/0); | |||
5196 | // kmp_depend_info[<Dependences.size()>] deps; | |||
5197 | DependenciesArray = | |||
5198 | CGF.CreateMemTemp(KmpDependInfoArrayTy, ".dep.arr.addr"); | |||
5199 | for (unsigned I = 0; I < NumDependencies; ++I) { | |||
5200 | const Expr *E = Data.Dependences[I].second; | |||
5201 | LValue Addr = CGF.EmitLValue(E); | |||
5202 | llvm::Value *Size; | |||
5203 | QualType Ty = E->getType(); | |||
5204 | if (const auto *ASE = | |||
5205 | dyn_cast<OMPArraySectionExpr>(E->IgnoreParenImpCasts())) { | |||
5206 | LValue UpAddrLVal = | |||
5207 | CGF.EmitOMPArraySectionExpr(ASE, /*LowerBound=*/false); | |||
5208 | llvm::Value *UpAddr = | |||
5209 | CGF.Builder.CreateConstGEP1_32(UpAddrLVal.getPointer(), /*Idx0=*/1); | |||
5210 | llvm::Value *LowIntPtr = | |||
5211 | CGF.Builder.CreatePtrToInt(Addr.getPointer(), CGM.SizeTy); | |||
5212 | llvm::Value *UpIntPtr = CGF.Builder.CreatePtrToInt(UpAddr, CGM.SizeTy); | |||
5213 | Size = CGF.Builder.CreateNUWSub(UpIntPtr, LowIntPtr); | |||
5214 | } else { | |||
5215 | Size = CGF.getTypeSize(Ty); | |||
5216 | } | |||
5217 | LValue Base = CGF.MakeAddrLValue( | |||
5218 | CGF.Builder.CreateConstArrayGEP(DependenciesArray, I), | |||
5219 | KmpDependInfoTy); | |||
5220 | // deps[i].base_addr = &<Dependences[i].second>; | |||
5221 | LValue BaseAddrLVal = CGF.EmitLValueForField( | |||
5222 | Base, *std::next(KmpDependInfoRD->field_begin(), BaseAddr)); | |||
5223 | CGF.EmitStoreOfScalar( | |||
5224 | CGF.Builder.CreatePtrToInt(Addr.getPointer(), CGF.IntPtrTy), | |||
5225 | BaseAddrLVal); | |||
5226 | // deps[i].len = sizeof(<Dependences[i].second>); | |||
5227 | LValue LenLVal = CGF.EmitLValueForField( | |||
5228 | Base, *std::next(KmpDependInfoRD->field_begin(), Len)); | |||
5229 | CGF.EmitStoreOfScalar(Size, LenLVal); | |||
5230 | // deps[i].flags = <Dependences[i].first>; | |||
5231 | RTLDependenceKindTy DepKind; | |||
5232 | switch (Data.Dependences[I].first) { | |||
5233 | case OMPC_DEPEND_in: | |||
5234 | DepKind = DepIn; | |||
5235 | break; | |||
5236 | // Out and InOut dependencies must use the same code. | |||
5237 | case OMPC_DEPEND_out: | |||
5238 | case OMPC_DEPEND_inout: | |||
5239 | DepKind = DepInOut; | |||
5240 | break; | |||
5241 | case OMPC_DEPEND_mutexinoutset: | |||
5242 | DepKind = DepMutexInOutSet; | |||
5243 | break; | |||
5244 | case OMPC_DEPEND_source: | |||
5245 | case OMPC_DEPEND_sink: | |||
5246 | case OMPC_DEPEND_unknown: | |||
5247 | llvm_unreachable("Unknown task dependence type")::llvm::llvm_unreachable_internal("Unknown task dependence type" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 5247); | |||
5248 | } | |||
5249 | LValue FlagsLVal = CGF.EmitLValueForField( | |||
5250 | Base, *std::next(KmpDependInfoRD->field_begin(), Flags)); | |||
5251 | CGF.EmitStoreOfScalar(llvm::ConstantInt::get(LLVMFlagsTy, DepKind), | |||
5252 | FlagsLVal); | |||
5253 | } | |||
5254 | DependenciesArray = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5255 | CGF.Builder.CreateConstArrayGEP(DependenciesArray, 0), CGF.VoidPtrTy); | |||
5256 | } | |||
5257 | ||||
5258 | // NOTE: routine and part_id fields are initialized by __kmpc_omp_task_alloc() | |||
5259 | // libcall. | |||
5260 | // Build kmp_int32 __kmpc_omp_task_with_deps(ident_t *, kmp_int32 gtid, | |||
5261 | // kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list, | |||
5262 | // kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list) if dependence | |||
5263 | // list is not empty | |||
5264 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
5265 | llvm::Value *UpLoc = emitUpdateLocation(CGF, Loc); | |||
5266 | llvm::Value *TaskArgs[] = { UpLoc, ThreadID, NewTask }; | |||
5267 | llvm::Value *DepTaskArgs[7]; | |||
5268 | if (NumDependencies) { | |||
5269 | DepTaskArgs[0] = UpLoc; | |||
5270 | DepTaskArgs[1] = ThreadID; | |||
5271 | DepTaskArgs[2] = NewTask; | |||
5272 | DepTaskArgs[3] = CGF.Builder.getInt32(NumDependencies); | |||
5273 | DepTaskArgs[4] = DependenciesArray.getPointer(); | |||
5274 | DepTaskArgs[5] = CGF.Builder.getInt32(0); | |||
5275 | DepTaskArgs[6] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | |||
5276 | } | |||
5277 | auto &&ThenCodeGen = [this, &Data, TDBase, KmpTaskTQTyRD, NumDependencies, | |||
5278 | &TaskArgs, | |||
5279 | &DepTaskArgs](CodeGenFunction &CGF, PrePostActionTy &) { | |||
5280 | if (!Data.Tied) { | |||
5281 | auto PartIdFI = std::next(KmpTaskTQTyRD->field_begin(), KmpTaskTPartId); | |||
5282 | LValue PartIdLVal = CGF.EmitLValueForField(TDBase, *PartIdFI); | |||
5283 | CGF.EmitStoreOfScalar(CGF.Builder.getInt32(0), PartIdLVal); | |||
5284 | } | |||
5285 | if (NumDependencies) { | |||
5286 | CGF.EmitRuntimeCall( | |||
5287 | createRuntimeFunction(OMPRTL__kmpc_omp_task_with_deps), DepTaskArgs); | |||
5288 | } else { | |||
5289 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_omp_task), | |||
5290 | TaskArgs); | |||
5291 | } | |||
5292 | // Check if parent region is untied and build return for untied task; | |||
5293 | if (auto *Region = | |||
5294 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | |||
5295 | Region->emitUntiedSwitch(CGF); | |||
5296 | }; | |||
5297 | ||||
5298 | llvm::Value *DepWaitTaskArgs[6]; | |||
5299 | if (NumDependencies) { | |||
5300 | DepWaitTaskArgs[0] = UpLoc; | |||
5301 | DepWaitTaskArgs[1] = ThreadID; | |||
5302 | DepWaitTaskArgs[2] = CGF.Builder.getInt32(NumDependencies); | |||
5303 | DepWaitTaskArgs[3] = DependenciesArray.getPointer(); | |||
5304 | DepWaitTaskArgs[4] = CGF.Builder.getInt32(0); | |||
5305 | DepWaitTaskArgs[5] = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); | |||
5306 | } | |||
5307 | auto &&ElseCodeGen = [&TaskArgs, ThreadID, NewTaskNewTaskTTy, TaskEntry, | |||
5308 | NumDependencies, &DepWaitTaskArgs, | |||
5309 | Loc](CodeGenFunction &CGF, PrePostActionTy &) { | |||
5310 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | |||
5311 | CodeGenFunction::RunCleanupsScope LocalScope(CGF); | |||
5312 | // Build void __kmpc_omp_wait_deps(ident_t *, kmp_int32 gtid, | |||
5313 | // kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 | |||
5314 | // ndeps_noalias, kmp_depend_info_t *noalias_dep_list); if dependence info | |||
5315 | // is specified. | |||
5316 | if (NumDependencies) | |||
5317 | CGF.EmitRuntimeCall(RT.createRuntimeFunction(OMPRTL__kmpc_omp_wait_deps), | |||
5318 | DepWaitTaskArgs); | |||
5319 | // Call proxy_task_entry(gtid, new_task); | |||
5320 | auto &&CodeGen = [TaskEntry, ThreadID, NewTaskNewTaskTTy, | |||
5321 | Loc](CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
5322 | Action.Enter(CGF); | |||
5323 | llvm::Value *OutlinedFnArgs[] = {ThreadID, NewTaskNewTaskTTy}; | |||
5324 | CGF.CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, Loc, TaskEntry, | |||
5325 | OutlinedFnArgs); | |||
5326 | }; | |||
5327 | ||||
5328 | // Build void __kmpc_omp_task_begin_if0(ident_t *, kmp_int32 gtid, | |||
5329 | // kmp_task_t *new_task); | |||
5330 | // Build void __kmpc_omp_task_complete_if0(ident_t *, kmp_int32 gtid, | |||
5331 | // kmp_task_t *new_task); | |||
5332 | RegionCodeGenTy RCG(CodeGen); | |||
5333 | CommonActionTy Action( | |||
5334 | RT.createRuntimeFunction(OMPRTL__kmpc_omp_task_begin_if0), TaskArgs, | |||
5335 | RT.createRuntimeFunction(OMPRTL__kmpc_omp_task_complete_if0), TaskArgs); | |||
5336 | RCG.setAction(Action); | |||
5337 | RCG(CGF); | |||
5338 | }; | |||
5339 | ||||
5340 | if (IfCond) { | |||
5341 | emitOMPIfClause(CGF, IfCond, ThenCodeGen, ElseCodeGen); | |||
5342 | } else { | |||
5343 | RegionCodeGenTy ThenRCG(ThenCodeGen); | |||
5344 | ThenRCG(CGF); | |||
5345 | } | |||
5346 | } | |||
5347 | ||||
5348 | void CGOpenMPRuntime::emitTaskLoopCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
5349 | const OMPLoopDirective &D, | |||
5350 | llvm::Function *TaskFunction, | |||
5351 | QualType SharedsTy, Address Shareds, | |||
5352 | const Expr *IfCond, | |||
5353 | const OMPTaskDataTy &Data) { | |||
5354 | if (!CGF.HaveInsertPoint()) | |||
5355 | return; | |||
5356 | TaskResultTy Result = | |||
5357 | emitTaskInit(CGF, Loc, D, TaskFunction, SharedsTy, Shareds, Data); | |||
5358 | // NOTE: routine and part_id fields are initialized by __kmpc_omp_task_alloc() | |||
5359 | // libcall. | |||
5360 | // Call to void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int | |||
5361 | // if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int | |||
5362 | // sched, kmp_uint64 grainsize, void *task_dup); | |||
5363 | llvm::Value *ThreadID = getThreadID(CGF, Loc); | |||
5364 | llvm::Value *UpLoc = emitUpdateLocation(CGF, Loc); | |||
5365 | llvm::Value *IfVal; | |||
5366 | if (IfCond) { | |||
5367 | IfVal = CGF.Builder.CreateIntCast(CGF.EvaluateExprAsBool(IfCond), CGF.IntTy, | |||
5368 | /*isSigned=*/true); | |||
5369 | } else { | |||
5370 | IfVal = llvm::ConstantInt::getSigned(CGF.IntTy, /*V=*/1); | |||
5371 | } | |||
5372 | ||||
5373 | LValue LBLVal = CGF.EmitLValueForField( | |||
5374 | Result.TDBase, | |||
5375 | *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTLowerBound)); | |||
5376 | const auto *LBVar = | |||
5377 | cast<VarDecl>(cast<DeclRefExpr>(D.getLowerBoundVariable())->getDecl()); | |||
5378 | CGF.EmitAnyExprToMem(LBVar->getInit(), LBLVal.getAddress(), LBLVal.getQuals(), | |||
5379 | /*IsInitializer=*/true); | |||
5380 | LValue UBLVal = CGF.EmitLValueForField( | |||
5381 | Result.TDBase, | |||
5382 | *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTUpperBound)); | |||
5383 | const auto *UBVar = | |||
5384 | cast<VarDecl>(cast<DeclRefExpr>(D.getUpperBoundVariable())->getDecl()); | |||
5385 | CGF.EmitAnyExprToMem(UBVar->getInit(), UBLVal.getAddress(), UBLVal.getQuals(), | |||
5386 | /*IsInitializer=*/true); | |||
5387 | LValue StLVal = CGF.EmitLValueForField( | |||
5388 | Result.TDBase, | |||
5389 | *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTStride)); | |||
5390 | const auto *StVar = | |||
5391 | cast<VarDecl>(cast<DeclRefExpr>(D.getStrideVariable())->getDecl()); | |||
5392 | CGF.EmitAnyExprToMem(StVar->getInit(), StLVal.getAddress(), StLVal.getQuals(), | |||
5393 | /*IsInitializer=*/true); | |||
5394 | // Store reductions address. | |||
5395 | LValue RedLVal = CGF.EmitLValueForField( | |||
5396 | Result.TDBase, | |||
5397 | *std::next(Result.KmpTaskTQTyRD->field_begin(), KmpTaskTReductions)); | |||
5398 | if (Data.Reductions) { | |||
5399 | CGF.EmitStoreOfScalar(Data.Reductions, RedLVal); | |||
5400 | } else { | |||
5401 | CGF.EmitNullInitialization(RedLVal.getAddress(), | |||
5402 | CGF.getContext().VoidPtrTy); | |||
5403 | } | |||
5404 | enum { NoSchedule = 0, Grainsize = 1, NumTasks = 2 }; | |||
5405 | llvm::Value *TaskArgs[] = { | |||
5406 | UpLoc, | |||
5407 | ThreadID, | |||
5408 | Result.NewTask, | |||
5409 | IfVal, | |||
5410 | LBLVal.getPointer(), | |||
5411 | UBLVal.getPointer(), | |||
5412 | CGF.EmitLoadOfScalar(StLVal, Loc), | |||
5413 | llvm::ConstantInt::getSigned( | |||
5414 | CGF.IntTy, 1), // Always 1 because taskgroup emitted by the compiler | |||
5415 | llvm::ConstantInt::getSigned( | |||
5416 | CGF.IntTy, Data.Schedule.getPointer() | |||
5417 | ? Data.Schedule.getInt() ? NumTasks : Grainsize | |||
5418 | : NoSchedule), | |||
5419 | Data.Schedule.getPointer() | |||
5420 | ? CGF.Builder.CreateIntCast(Data.Schedule.getPointer(), CGF.Int64Ty, | |||
5421 | /*isSigned=*/false) | |||
5422 | : llvm::ConstantInt::get(CGF.Int64Ty, /*V=*/0), | |||
5423 | Result.TaskDupFn ? CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5424 | Result.TaskDupFn, CGF.VoidPtrTy) | |||
5425 | : llvm::ConstantPointerNull::get(CGF.VoidPtrTy)}; | |||
5426 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_taskloop), TaskArgs); | |||
5427 | } | |||
5428 | ||||
5429 | /// Emit reduction operation for each element of array (required for | |||
5430 | /// array sections) LHS op = RHS. | |||
5431 | /// \param Type Type of array. | |||
5432 | /// \param LHSVar Variable on the left side of the reduction operation | |||
5433 | /// (references element of array in original variable). | |||
5434 | /// \param RHSVar Variable on the right side of the reduction operation | |||
5435 | /// (references element of array in original variable). | |||
5436 | /// \param RedOpGen Generator of reduction operation with use of LHSVar and | |||
5437 | /// RHSVar. | |||
5438 | static void EmitOMPAggregateReduction( | |||
5439 | CodeGenFunction &CGF, QualType Type, const VarDecl *LHSVar, | |||
5440 | const VarDecl *RHSVar, | |||
5441 | const llvm::function_ref<void(CodeGenFunction &CGF, const Expr *, | |||
5442 | const Expr *, const Expr *)> &RedOpGen, | |||
5443 | const Expr *XExpr = nullptr, const Expr *EExpr = nullptr, | |||
5444 | const Expr *UpExpr = nullptr) { | |||
5445 | // Perform element-by-element initialization. | |||
5446 | QualType ElementTy; | |||
5447 | Address LHSAddr = CGF.GetAddrOfLocalVar(LHSVar); | |||
5448 | Address RHSAddr = CGF.GetAddrOfLocalVar(RHSVar); | |||
5449 | ||||
5450 | // Drill down to the base element type on both arrays. | |||
5451 | const ArrayType *ArrayTy = Type->getAsArrayTypeUnsafe(); | |||
5452 | llvm::Value *NumElements = CGF.emitArrayLength(ArrayTy, ElementTy, LHSAddr); | |||
5453 | ||||
5454 | llvm::Value *RHSBegin = RHSAddr.getPointer(); | |||
5455 | llvm::Value *LHSBegin = LHSAddr.getPointer(); | |||
5456 | // Cast from pointer to array type to pointer to single element. | |||
5457 | llvm::Value *LHSEnd = CGF.Builder.CreateGEP(LHSBegin, NumElements); | |||
5458 | // The basic structure here is a while-do loop. | |||
5459 | llvm::BasicBlock *BodyBB = CGF.createBasicBlock("omp.arraycpy.body"); | |||
5460 | llvm::BasicBlock *DoneBB = CGF.createBasicBlock("omp.arraycpy.done"); | |||
5461 | llvm::Value *IsEmpty = | |||
5462 | CGF.Builder.CreateICmpEQ(LHSBegin, LHSEnd, "omp.arraycpy.isempty"); | |||
5463 | CGF.Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB); | |||
5464 | ||||
5465 | // Enter the loop body, making that address the current address. | |||
5466 | llvm::BasicBlock *EntryBB = CGF.Builder.GetInsertBlock(); | |||
5467 | CGF.EmitBlock(BodyBB); | |||
5468 | ||||
5469 | CharUnits ElementSize = CGF.getContext().getTypeSizeInChars(ElementTy); | |||
5470 | ||||
5471 | llvm::PHINode *RHSElementPHI = CGF.Builder.CreatePHI( | |||
5472 | RHSBegin->getType(), 2, "omp.arraycpy.srcElementPast"); | |||
5473 | RHSElementPHI->addIncoming(RHSBegin, EntryBB); | |||
5474 | Address RHSElementCurrent = | |||
5475 | Address(RHSElementPHI, | |||
5476 | RHSAddr.getAlignment().alignmentOfArrayElement(ElementSize)); | |||
5477 | ||||
5478 | llvm::PHINode *LHSElementPHI = CGF.Builder.CreatePHI( | |||
5479 | LHSBegin->getType(), 2, "omp.arraycpy.destElementPast"); | |||
5480 | LHSElementPHI->addIncoming(LHSBegin, EntryBB); | |||
5481 | Address LHSElementCurrent = | |||
5482 | Address(LHSElementPHI, | |||
5483 | LHSAddr.getAlignment().alignmentOfArrayElement(ElementSize)); | |||
5484 | ||||
5485 | // Emit copy. | |||
5486 | CodeGenFunction::OMPPrivateScope Scope(CGF); | |||
5487 | Scope.addPrivate(LHSVar, [=]() { return LHSElementCurrent; }); | |||
5488 | Scope.addPrivate(RHSVar, [=]() { return RHSElementCurrent; }); | |||
5489 | Scope.Privatize(); | |||
5490 | RedOpGen(CGF, XExpr, EExpr, UpExpr); | |||
5491 | Scope.ForceCleanup(); | |||
5492 | ||||
5493 | // Shift the address forward by one element. | |||
5494 | llvm::Value *LHSElementNext = CGF.Builder.CreateConstGEP1_32( | |||
5495 | LHSElementPHI, /*Idx0=*/1, "omp.arraycpy.dest.element"); | |||
5496 | llvm::Value *RHSElementNext = CGF.Builder.CreateConstGEP1_32( | |||
5497 | RHSElementPHI, /*Idx0=*/1, "omp.arraycpy.src.element"); | |||
5498 | // Check whether we've reached the end. | |||
5499 | llvm::Value *Done = | |||
5500 | CGF.Builder.CreateICmpEQ(LHSElementNext, LHSEnd, "omp.arraycpy.done"); | |||
5501 | CGF.Builder.CreateCondBr(Done, DoneBB, BodyBB); | |||
5502 | LHSElementPHI->addIncoming(LHSElementNext, CGF.Builder.GetInsertBlock()); | |||
5503 | RHSElementPHI->addIncoming(RHSElementNext, CGF.Builder.GetInsertBlock()); | |||
5504 | ||||
5505 | // Done. | |||
5506 | CGF.EmitBlock(DoneBB, /*IsFinished=*/true); | |||
5507 | } | |||
5508 | ||||
5509 | /// Emit reduction combiner. If the combiner is a simple expression emit it as | |||
5510 | /// is, otherwise consider it as combiner of UDR decl and emit it as a call of | |||
5511 | /// UDR combiner function. | |||
5512 | static void emitReductionCombiner(CodeGenFunction &CGF, | |||
5513 | const Expr *ReductionOp) { | |||
5514 | if (const auto *CE = dyn_cast<CallExpr>(ReductionOp)) | |||
5515 | if (const auto *OVE = dyn_cast<OpaqueValueExpr>(CE->getCallee())) | |||
5516 | if (const auto *DRE = | |||
5517 | dyn_cast<DeclRefExpr>(OVE->getSourceExpr()->IgnoreImpCasts())) | |||
5518 | if (const auto *DRD = | |||
5519 | dyn_cast<OMPDeclareReductionDecl>(DRE->getDecl())) { | |||
5520 | std::pair<llvm::Function *, llvm::Function *> Reduction = | |||
5521 | CGF.CGM.getOpenMPRuntime().getUserDefinedReduction(DRD); | |||
5522 | RValue Func = RValue::get(Reduction.first); | |||
5523 | CodeGenFunction::OpaqueValueMapping Map(CGF, OVE, Func); | |||
5524 | CGF.EmitIgnoredExpr(ReductionOp); | |||
5525 | return; | |||
5526 | } | |||
5527 | CGF.EmitIgnoredExpr(ReductionOp); | |||
5528 | } | |||
5529 | ||||
5530 | llvm::Function *CGOpenMPRuntime::emitReductionFunction( | |||
5531 | SourceLocation Loc, llvm::Type *ArgsType, ArrayRef<const Expr *> Privates, | |||
5532 | ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs, | |||
5533 | ArrayRef<const Expr *> ReductionOps) { | |||
5534 | ASTContext &C = CGM.getContext(); | |||
5535 | ||||
5536 | // void reduction_func(void *LHSArg, void *RHSArg); | |||
5537 | FunctionArgList Args; | |||
5538 | ImplicitParamDecl LHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
5539 | ImplicitParamDecl::Other); | |||
5540 | ImplicitParamDecl RHSArg(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
5541 | ImplicitParamDecl::Other); | |||
5542 | Args.push_back(&LHSArg); | |||
5543 | Args.push_back(&RHSArg); | |||
5544 | const auto &CGFI = | |||
5545 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
5546 | std::string Name = getName({"omp", "reduction", "reduction_func"}); | |||
5547 | auto *Fn = llvm::Function::Create(CGM.getTypes().GetFunctionType(CGFI), | |||
5548 | llvm::GlobalValue::InternalLinkage, Name, | |||
5549 | &CGM.getModule()); | |||
5550 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, CGFI); | |||
5551 | Fn->setDoesNotRecurse(); | |||
5552 | CodeGenFunction CGF(CGM); | |||
5553 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, CGFI, Args, Loc, Loc); | |||
5554 | ||||
5555 | // Dst = (void*[n])(LHSArg); | |||
5556 | // Src = (void*[n])(RHSArg); | |||
5557 | Address LHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5558 | CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&LHSArg)), | |||
5559 | ArgsType), CGF.getPointerAlign()); | |||
5560 | Address RHS(CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5561 | CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar(&RHSArg)), | |||
5562 | ArgsType), CGF.getPointerAlign()); | |||
5563 | ||||
5564 | // ... | |||
5565 | // *(Type<i>*)lhs[i] = RedOp<i>(*(Type<i>*)lhs[i], *(Type<i>*)rhs[i]); | |||
5566 | // ... | |||
5567 | CodeGenFunction::OMPPrivateScope Scope(CGF); | |||
5568 | auto IPriv = Privates.begin(); | |||
5569 | unsigned Idx = 0; | |||
5570 | for (unsigned I = 0, E = ReductionOps.size(); I < E; ++I, ++IPriv, ++Idx) { | |||
5571 | const auto *RHSVar = | |||
5572 | cast<VarDecl>(cast<DeclRefExpr>(RHSExprs[I])->getDecl()); | |||
5573 | Scope.addPrivate(RHSVar, [&CGF, RHS, Idx, RHSVar]() { | |||
5574 | return emitAddrOfVarFromArray(CGF, RHS, Idx, RHSVar); | |||
5575 | }); | |||
5576 | const auto *LHSVar = | |||
5577 | cast<VarDecl>(cast<DeclRefExpr>(LHSExprs[I])->getDecl()); | |||
5578 | Scope.addPrivate(LHSVar, [&CGF, LHS, Idx, LHSVar]() { | |||
5579 | return emitAddrOfVarFromArray(CGF, LHS, Idx, LHSVar); | |||
5580 | }); | |||
5581 | QualType PrivTy = (*IPriv)->getType(); | |||
5582 | if (PrivTy->isVariablyModifiedType()) { | |||
5583 | // Get array size and emit VLA type. | |||
5584 | ++Idx; | |||
5585 | Address Elem = CGF.Builder.CreateConstArrayGEP(LHS, Idx); | |||
5586 | llvm::Value *Ptr = CGF.Builder.CreateLoad(Elem); | |||
5587 | const VariableArrayType *VLA = | |||
5588 | CGF.getContext().getAsVariableArrayType(PrivTy); | |||
5589 | const auto *OVE = cast<OpaqueValueExpr>(VLA->getSizeExpr()); | |||
5590 | CodeGenFunction::OpaqueValueMapping OpaqueMap( | |||
5591 | CGF, OVE, RValue::get(CGF.Builder.CreatePtrToInt(Ptr, CGF.SizeTy))); | |||
5592 | CGF.EmitVariablyModifiedType(PrivTy); | |||
5593 | } | |||
5594 | } | |||
5595 | Scope.Privatize(); | |||
5596 | IPriv = Privates.begin(); | |||
5597 | auto ILHS = LHSExprs.begin(); | |||
5598 | auto IRHS = RHSExprs.begin(); | |||
5599 | for (const Expr *E : ReductionOps) { | |||
5600 | if ((*IPriv)->getType()->isArrayType()) { | |||
5601 | // Emit reduction for array section. | |||
5602 | const auto *LHSVar = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl()); | |||
5603 | const auto *RHSVar = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl()); | |||
5604 | EmitOMPAggregateReduction( | |||
5605 | CGF, (*IPriv)->getType(), LHSVar, RHSVar, | |||
5606 | [=](CodeGenFunction &CGF, const Expr *, const Expr *, const Expr *) { | |||
5607 | emitReductionCombiner(CGF, E); | |||
5608 | }); | |||
5609 | } else { | |||
5610 | // Emit reduction for array subscript or single variable. | |||
5611 | emitReductionCombiner(CGF, E); | |||
5612 | } | |||
5613 | ++IPriv; | |||
5614 | ++ILHS; | |||
5615 | ++IRHS; | |||
5616 | } | |||
5617 | Scope.ForceCleanup(); | |||
5618 | CGF.FinishFunction(); | |||
5619 | return Fn; | |||
5620 | } | |||
5621 | ||||
5622 | void CGOpenMPRuntime::emitSingleReductionCombiner(CodeGenFunction &CGF, | |||
5623 | const Expr *ReductionOp, | |||
5624 | const Expr *PrivateRef, | |||
5625 | const DeclRefExpr *LHS, | |||
5626 | const DeclRefExpr *RHS) { | |||
5627 | if (PrivateRef->getType()->isArrayType()) { | |||
5628 | // Emit reduction for array section. | |||
5629 | const auto *LHSVar = cast<VarDecl>(LHS->getDecl()); | |||
5630 | const auto *RHSVar = cast<VarDecl>(RHS->getDecl()); | |||
5631 | EmitOMPAggregateReduction( | |||
5632 | CGF, PrivateRef->getType(), LHSVar, RHSVar, | |||
5633 | [=](CodeGenFunction &CGF, const Expr *, const Expr *, const Expr *) { | |||
5634 | emitReductionCombiner(CGF, ReductionOp); | |||
5635 | }); | |||
5636 | } else { | |||
5637 | // Emit reduction for array subscript or single variable. | |||
5638 | emitReductionCombiner(CGF, ReductionOp); | |||
5639 | } | |||
5640 | } | |||
5641 | ||||
5642 | void CGOpenMPRuntime::emitReduction(CodeGenFunction &CGF, SourceLocation Loc, | |||
5643 | ArrayRef<const Expr *> Privates, | |||
5644 | ArrayRef<const Expr *> LHSExprs, | |||
5645 | ArrayRef<const Expr *> RHSExprs, | |||
5646 | ArrayRef<const Expr *> ReductionOps, | |||
5647 | ReductionOptionsTy Options) { | |||
5648 | if (!CGF.HaveInsertPoint()) | |||
5649 | return; | |||
5650 | ||||
5651 | bool WithNowait = Options.WithNowait; | |||
5652 | bool SimpleReduction = Options.SimpleReduction; | |||
5653 | ||||
5654 | // Next code should be emitted for reduction: | |||
5655 | // | |||
5656 | // static kmp_critical_name lock = { 0 }; | |||
5657 | // | |||
5658 | // void reduce_func(void *lhs[<n>], void *rhs[<n>]) { | |||
5659 | // *(Type0*)lhs[0] = ReductionOperation0(*(Type0*)lhs[0], *(Type0*)rhs[0]); | |||
5660 | // ... | |||
5661 | // *(Type<n>-1*)lhs[<n>-1] = ReductionOperation<n>-1(*(Type<n>-1*)lhs[<n>-1], | |||
5662 | // *(Type<n>-1*)rhs[<n>-1]); | |||
5663 | // } | |||
5664 | // | |||
5665 | // ... | |||
5666 | // void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; | |||
5667 | // switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), | |||
5668 | // RedList, reduce_func, &<lock>)) { | |||
5669 | // case 1: | |||
5670 | // ... | |||
5671 | // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]); | |||
5672 | // ... | |||
5673 | // __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); | |||
5674 | // break; | |||
5675 | // case 2: | |||
5676 | // ... | |||
5677 | // Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i])); | |||
5678 | // ... | |||
5679 | // [__kmpc_end_reduce(<loc>, <gtid>, &<lock>);] | |||
5680 | // break; | |||
5681 | // default:; | |||
5682 | // } | |||
5683 | // | |||
5684 | // if SimpleReduction is true, only the next code is generated: | |||
5685 | // ... | |||
5686 | // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]); | |||
5687 | // ... | |||
5688 | ||||
5689 | ASTContext &C = CGM.getContext(); | |||
5690 | ||||
5691 | if (SimpleReduction) { | |||
5692 | CodeGenFunction::RunCleanupsScope Scope(CGF); | |||
5693 | auto IPriv = Privates.begin(); | |||
5694 | auto ILHS = LHSExprs.begin(); | |||
5695 | auto IRHS = RHSExprs.begin(); | |||
5696 | for (const Expr *E : ReductionOps) { | |||
5697 | emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS), | |||
5698 | cast<DeclRefExpr>(*IRHS)); | |||
5699 | ++IPriv; | |||
5700 | ++ILHS; | |||
5701 | ++IRHS; | |||
5702 | } | |||
5703 | return; | |||
5704 | } | |||
5705 | ||||
5706 | // 1. Build a list of reduction variables. | |||
5707 | // void *RedList[<n>] = {<ReductionVars>[0], ..., <ReductionVars>[<n>-1]}; | |||
5708 | auto Size = RHSExprs.size(); | |||
5709 | for (const Expr *E : Privates) { | |||
5710 | if (E->getType()->isVariablyModifiedType()) | |||
5711 | // Reserve place for array size. | |||
5712 | ++Size; | |||
5713 | } | |||
5714 | llvm::APInt ArraySize(/*unsigned int numBits=*/32, Size); | |||
5715 | QualType ReductionArrayTy = | |||
5716 | C.getConstantArrayType(C.VoidPtrTy, ArraySize, ArrayType::Normal, | |||
5717 | /*IndexTypeQuals=*/0); | |||
5718 | Address ReductionList = | |||
5719 | CGF.CreateMemTemp(ReductionArrayTy, ".omp.reduction.red_list"); | |||
5720 | auto IPriv = Privates.begin(); | |||
5721 | unsigned Idx = 0; | |||
5722 | for (unsigned I = 0, E = RHSExprs.size(); I < E; ++I, ++IPriv, ++Idx) { | |||
5723 | Address Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); | |||
5724 | CGF.Builder.CreateStore( | |||
5725 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5726 | CGF.EmitLValue(RHSExprs[I]).getPointer(), CGF.VoidPtrTy), | |||
5727 | Elem); | |||
5728 | if ((*IPriv)->getType()->isVariablyModifiedType()) { | |||
5729 | // Store array size. | |||
5730 | ++Idx; | |||
5731 | Elem = CGF.Builder.CreateConstArrayGEP(ReductionList, Idx); | |||
5732 | llvm::Value *Size = CGF.Builder.CreateIntCast( | |||
5733 | CGF.getVLASize( | |||
5734 | CGF.getContext().getAsVariableArrayType((*IPriv)->getType())) | |||
5735 | .NumElts, | |||
5736 | CGF.SizeTy, /*isSigned=*/false); | |||
5737 | CGF.Builder.CreateStore(CGF.Builder.CreateIntToPtr(Size, CGF.VoidPtrTy), | |||
5738 | Elem); | |||
5739 | } | |||
5740 | } | |||
5741 | ||||
5742 | // 2. Emit reduce_func(). | |||
5743 | llvm::Function *ReductionFn = emitReductionFunction( | |||
5744 | Loc, CGF.ConvertTypeForMem(ReductionArrayTy)->getPointerTo(), Privates, | |||
5745 | LHSExprs, RHSExprs, ReductionOps); | |||
5746 | ||||
5747 | // 3. Create static kmp_critical_name lock = { 0 }; | |||
5748 | std::string Name = getName({"reduction"}); | |||
5749 | llvm::Value *Lock = getCriticalRegionLock(Name); | |||
5750 | ||||
5751 | // 4. Build res = __kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), | |||
5752 | // RedList, reduce_func, &<lock>); | |||
5753 | llvm::Value *IdentTLoc = emitUpdateLocation(CGF, Loc, OMP_ATOMIC_REDUCE); | |||
5754 | llvm::Value *ThreadId = getThreadID(CGF, Loc); | |||
5755 | llvm::Value *ReductionArrayTySize = CGF.getTypeSize(ReductionArrayTy); | |||
5756 | llvm::Value *RL = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
5757 | ReductionList.getPointer(), CGF.VoidPtrTy); | |||
5758 | llvm::Value *Args[] = { | |||
5759 | IdentTLoc, // ident_t *<loc> | |||
5760 | ThreadId, // i32 <gtid> | |||
5761 | CGF.Builder.getInt32(RHSExprs.size()), // i32 <n> | |||
5762 | ReductionArrayTySize, // size_type sizeof(RedList) | |||
5763 | RL, // void *RedList | |||
5764 | ReductionFn, // void (*) (void *, void *) <reduce_func> | |||
5765 | Lock // kmp_critical_name *&<lock> | |||
5766 | }; | |||
5767 | llvm::Value *Res = CGF.EmitRuntimeCall( | |||
5768 | createRuntimeFunction(WithNowait ? OMPRTL__kmpc_reduce_nowait | |||
5769 | : OMPRTL__kmpc_reduce), | |||
5770 | Args); | |||
5771 | ||||
5772 | // 5. Build switch(res) | |||
5773 | llvm::BasicBlock *DefaultBB = CGF.createBasicBlock(".omp.reduction.default"); | |||
5774 | llvm::SwitchInst *SwInst = | |||
5775 | CGF.Builder.CreateSwitch(Res, DefaultBB, /*NumCases=*/2); | |||
5776 | ||||
5777 | // 6. Build case 1: | |||
5778 | // ... | |||
5779 | // <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]); | |||
5780 | // ... | |||
5781 | // __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); | |||
5782 | // break; | |||
5783 | llvm::BasicBlock *Case1BB = CGF.createBasicBlock(".omp.reduction.case1"); | |||
5784 | SwInst->addCase(CGF.Builder.getInt32(1), Case1BB); | |||
5785 | CGF.EmitBlock(Case1BB); | |||
5786 | ||||
5787 | // Add emission of __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); | |||
5788 | llvm::Value *EndArgs[] = { | |||
5789 | IdentTLoc, // ident_t *<loc> | |||
5790 | ThreadId, // i32 <gtid> | |||
5791 | Lock // kmp_critical_name *&<lock> | |||
5792 | }; | |||
5793 | auto &&CodeGen = [Privates, LHSExprs, RHSExprs, ReductionOps]( | |||
5794 | CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
5795 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | |||
5796 | auto IPriv = Privates.begin(); | |||
5797 | auto ILHS = LHSExprs.begin(); | |||
5798 | auto IRHS = RHSExprs.begin(); | |||
5799 | for (const Expr *E : ReductionOps) { | |||
5800 | RT.emitSingleReductionCombiner(CGF, E, *IPriv, cast<DeclRefExpr>(*ILHS), | |||
5801 | cast<DeclRefExpr>(*IRHS)); | |||
5802 | ++IPriv; | |||
5803 | ++ILHS; | |||
5804 | ++IRHS; | |||
5805 | } | |||
5806 | }; | |||
5807 | RegionCodeGenTy RCG(CodeGen); | |||
5808 | CommonActionTy Action( | |||
5809 | nullptr, llvm::None, | |||
5810 | createRuntimeFunction(WithNowait ? OMPRTL__kmpc_end_reduce_nowait | |||
5811 | : OMPRTL__kmpc_end_reduce), | |||
5812 | EndArgs); | |||
5813 | RCG.setAction(Action); | |||
5814 | RCG(CGF); | |||
5815 | ||||
5816 | CGF.EmitBranch(DefaultBB); | |||
5817 | ||||
5818 | // 7. Build case 2: | |||
5819 | // ... | |||
5820 | // Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i])); | |||
5821 | // ... | |||
5822 | // break; | |||
5823 | llvm::BasicBlock *Case2BB = CGF.createBasicBlock(".omp.reduction.case2"); | |||
5824 | SwInst->addCase(CGF.Builder.getInt32(2), Case2BB); | |||
5825 | CGF.EmitBlock(Case2BB); | |||
5826 | ||||
5827 | auto &&AtomicCodeGen = [Loc, Privates, LHSExprs, RHSExprs, ReductionOps]( | |||
5828 | CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
5829 | auto ILHS = LHSExprs.begin(); | |||
5830 | auto IRHS = RHSExprs.begin(); | |||
5831 | auto IPriv = Privates.begin(); | |||
5832 | for (const Expr *E : ReductionOps) { | |||
5833 | const Expr *XExpr = nullptr; | |||
5834 | const Expr *EExpr = nullptr; | |||
5835 | const Expr *UpExpr = nullptr; | |||
5836 | BinaryOperatorKind BO = BO_Comma; | |||
5837 | if (const auto *BO = dyn_cast<BinaryOperator>(E)) { | |||
5838 | if (BO->getOpcode() == BO_Assign) { | |||
5839 | XExpr = BO->getLHS(); | |||
5840 | UpExpr = BO->getRHS(); | |||
5841 | } | |||
5842 | } | |||
5843 | // Try to emit update expression as a simple atomic. | |||
5844 | const Expr *RHSExpr = UpExpr; | |||
5845 | if (RHSExpr) { | |||
5846 | // Analyze RHS part of the whole expression. | |||
5847 | if (const auto *ACO = dyn_cast<AbstractConditionalOperator>( | |||
5848 | RHSExpr->IgnoreParenImpCasts())) { | |||
5849 | // If this is a conditional operator, analyze its condition for | |||
5850 | // min/max reduction operator. | |||
5851 | RHSExpr = ACO->getCond(); | |||
5852 | } | |||
5853 | if (const auto *BORHS = | |||
5854 | dyn_cast<BinaryOperator>(RHSExpr->IgnoreParenImpCasts())) { | |||
5855 | EExpr = BORHS->getRHS(); | |||
5856 | BO = BORHS->getOpcode(); | |||
5857 | } | |||
5858 | } | |||
5859 | if (XExpr) { | |||
5860 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl()); | |||
5861 | auto &&AtomicRedGen = [BO, VD, | |||
5862 | Loc](CodeGenFunction &CGF, const Expr *XExpr, | |||
5863 | const Expr *EExpr, const Expr *UpExpr) { | |||
5864 | LValue X = CGF.EmitLValue(XExpr); | |||
5865 | RValue E; | |||
5866 | if (EExpr) | |||
5867 | E = CGF.EmitAnyExpr(EExpr); | |||
5868 | CGF.EmitOMPAtomicSimpleUpdateExpr( | |||
5869 | X, E, BO, /*IsXLHSInRHSPart=*/true, | |||
5870 | llvm::AtomicOrdering::Monotonic, Loc, | |||
5871 | [&CGF, UpExpr, VD, Loc](RValue XRValue) { | |||
5872 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); | |||
5873 | PrivateScope.addPrivate( | |||
5874 | VD, [&CGF, VD, XRValue, Loc]() { | |||
5875 | Address LHSTemp = CGF.CreateMemTemp(VD->getType()); | |||
5876 | CGF.emitOMPSimpleStore( | |||
5877 | CGF.MakeAddrLValue(LHSTemp, VD->getType()), XRValue, | |||
5878 | VD->getType().getNonReferenceType(), Loc); | |||
5879 | return LHSTemp; | |||
5880 | }); | |||
5881 | (void)PrivateScope.Privatize(); | |||
5882 | return CGF.EmitAnyExpr(UpExpr); | |||
5883 | }); | |||
5884 | }; | |||
5885 | if ((*IPriv)->getType()->isArrayType()) { | |||
5886 | // Emit atomic reduction for array section. | |||
5887 | const auto *RHSVar = | |||
5888 | cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl()); | |||
5889 | EmitOMPAggregateReduction(CGF, (*IPriv)->getType(), VD, RHSVar, | |||
5890 | AtomicRedGen, XExpr, EExpr, UpExpr); | |||
5891 | } else { | |||
5892 | // Emit atomic reduction for array subscript or single variable. | |||
5893 | AtomicRedGen(CGF, XExpr, EExpr, UpExpr); | |||
5894 | } | |||
5895 | } else { | |||
5896 | // Emit as a critical region. | |||
5897 | auto &&CritRedGen = [E, Loc](CodeGenFunction &CGF, const Expr *, | |||
5898 | const Expr *, const Expr *) { | |||
5899 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | |||
5900 | std::string Name = RT.getName({"atomic_reduction"}); | |||
5901 | RT.emitCriticalRegion( | |||
5902 | CGF, Name, | |||
5903 | [=](CodeGenFunction &CGF, PrePostActionTy &Action) { | |||
5904 | Action.Enter(CGF); | |||
5905 | emitReductionCombiner(CGF, E); | |||
5906 | }, | |||
5907 | Loc); | |||
5908 | }; | |||
5909 | if ((*IPriv)->getType()->isArrayType()) { | |||
5910 | const auto *LHSVar = | |||
5911 | cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl()); | |||
5912 | const auto *RHSVar = | |||
5913 | cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl()); | |||
5914 | EmitOMPAggregateReduction(CGF, (*IPriv)->getType(), LHSVar, RHSVar, | |||
5915 | CritRedGen); | |||
5916 | } else { | |||
5917 | CritRedGen(CGF, nullptr, nullptr, nullptr); | |||
5918 | } | |||
5919 | } | |||
5920 | ++ILHS; | |||
5921 | ++IRHS; | |||
5922 | ++IPriv; | |||
5923 | } | |||
5924 | }; | |||
5925 | RegionCodeGenTy AtomicRCG(AtomicCodeGen); | |||
5926 | if (!WithNowait) { | |||
5927 | // Add emission of __kmpc_end_reduce(<loc>, <gtid>, &<lock>); | |||
5928 | llvm::Value *EndArgs[] = { | |||
5929 | IdentTLoc, // ident_t *<loc> | |||
5930 | ThreadId, // i32 <gtid> | |||
5931 | Lock // kmp_critical_name *&<lock> | |||
5932 | }; | |||
5933 | CommonActionTy Action(nullptr, llvm::None, | |||
5934 | createRuntimeFunction(OMPRTL__kmpc_end_reduce), | |||
5935 | EndArgs); | |||
5936 | AtomicRCG.setAction(Action); | |||
5937 | AtomicRCG(CGF); | |||
5938 | } else { | |||
5939 | AtomicRCG(CGF); | |||
5940 | } | |||
5941 | ||||
5942 | CGF.EmitBranch(DefaultBB); | |||
5943 | CGF.EmitBlock(DefaultBB, /*IsFinished=*/true); | |||
5944 | } | |||
5945 | ||||
5946 | /// Generates unique name for artificial threadprivate variables. | |||
5947 | /// Format is: <Prefix> "." <Decl_mangled_name> "_" "<Decl_start_loc_raw_enc>" | |||
5948 | static std::string generateUniqueName(CodeGenModule &CGM, StringRef Prefix, | |||
5949 | const Expr *Ref) { | |||
5950 | SmallString<256> Buffer; | |||
5951 | llvm::raw_svector_ostream Out(Buffer); | |||
5952 | const clang::DeclRefExpr *DE; | |||
5953 | const VarDecl *D = ::getBaseDecl(Ref, DE); | |||
5954 | if (!D) | |||
5955 | D = cast<VarDecl>(cast<DeclRefExpr>(Ref)->getDecl()); | |||
5956 | D = D->getCanonicalDecl(); | |||
5957 | std::string Name = CGM.getOpenMPRuntime().getName( | |||
5958 | {D->isLocalVarDeclOrParm() ? D->getName() : CGM.getMangledName(D)}); | |||
5959 | Out << Prefix << Name << "_" | |||
5960 | << D->getCanonicalDecl()->getBeginLoc().getRawEncoding(); | |||
5961 | return Out.str(); | |||
5962 | } | |||
5963 | ||||
5964 | /// Emits reduction initializer function: | |||
5965 | /// \code | |||
5966 | /// void @.red_init(void* %arg) { | |||
5967 | /// %0 = bitcast void* %arg to <type>* | |||
5968 | /// store <type> <init>, <type>* %0 | |||
5969 | /// ret void | |||
5970 | /// } | |||
5971 | /// \endcode | |||
5972 | static llvm::Value *emitReduceInitFunction(CodeGenModule &CGM, | |||
5973 | SourceLocation Loc, | |||
5974 | ReductionCodeGen &RCG, unsigned N) { | |||
5975 | ASTContext &C = CGM.getContext(); | |||
5976 | FunctionArgList Args; | |||
5977 | ImplicitParamDecl Param(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
5978 | ImplicitParamDecl::Other); | |||
5979 | Args.emplace_back(&Param); | |||
5980 | const auto &FnInfo = | |||
5981 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
5982 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | |||
5983 | std::string Name = CGM.getOpenMPRuntime().getName({"red_init", ""}); | |||
5984 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | |||
5985 | Name, &CGM.getModule()); | |||
5986 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | |||
5987 | Fn->setDoesNotRecurse(); | |||
5988 | CodeGenFunction CGF(CGM); | |||
5989 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc); | |||
5990 | Address PrivateAddr = CGF.EmitLoadOfPointer( | |||
5991 | CGF.GetAddrOfLocalVar(&Param), | |||
5992 | C.getPointerType(C.VoidPtrTy).castAs<PointerType>()); | |||
5993 | llvm::Value *Size = nullptr; | |||
5994 | // If the size of the reduction item is non-constant, load it from global | |||
5995 | // threadprivate variable. | |||
5996 | if (RCG.getSizes(N).second) { | |||
5997 | Address SizeAddr = CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate( | |||
5998 | CGF, CGM.getContext().getSizeType(), | |||
5999 | generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N))); | |||
6000 | Size = CGF.EmitLoadOfScalar(SizeAddr, /*Volatile=*/false, | |||
6001 | CGM.getContext().getSizeType(), Loc); | |||
6002 | } | |||
6003 | RCG.emitAggregateType(CGF, N, Size); | |||
6004 | LValue SharedLVal; | |||
6005 | // If initializer uses initializer from declare reduction construct, emit a | |||
6006 | // pointer to the address of the original reduction item (reuired by reduction | |||
6007 | // initializer) | |||
6008 | if (RCG.usesReductionInitializer(N)) { | |||
6009 | Address SharedAddr = | |||
6010 | CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate( | |||
6011 | CGF, CGM.getContext().VoidPtrTy, | |||
6012 | generateUniqueName(CGM, "reduction", RCG.getRefExpr(N))); | |||
6013 | SharedAddr = CGF.EmitLoadOfPointer( | |||
6014 | SharedAddr, | |||
6015 | CGM.getContext().VoidPtrTy.castAs<PointerType>()->getTypePtr()); | |||
6016 | SharedLVal = CGF.MakeAddrLValue(SharedAddr, CGM.getContext().VoidPtrTy); | |||
6017 | } else { | |||
6018 | SharedLVal = CGF.MakeNaturalAlignAddrLValue( | |||
6019 | llvm::ConstantPointerNull::get(CGM.VoidPtrTy), | |||
6020 | CGM.getContext().VoidPtrTy); | |||
6021 | } | |||
6022 | // Emit the initializer: | |||
6023 | // %0 = bitcast void* %arg to <type>* | |||
6024 | // store <type> <init>, <type>* %0 | |||
6025 | RCG.emitInitialization(CGF, N, PrivateAddr, SharedLVal, | |||
6026 | [](CodeGenFunction &) { return false; }); | |||
6027 | CGF.FinishFunction(); | |||
6028 | return Fn; | |||
6029 | } | |||
6030 | ||||
6031 | /// Emits reduction combiner function: | |||
6032 | /// \code | |||
6033 | /// void @.red_comb(void* %arg0, void* %arg1) { | |||
6034 | /// %lhs = bitcast void* %arg0 to <type>* | |||
6035 | /// %rhs = bitcast void* %arg1 to <type>* | |||
6036 | /// %2 = <ReductionOp>(<type>* %lhs, <type>* %rhs) | |||
6037 | /// store <type> %2, <type>* %lhs | |||
6038 | /// ret void | |||
6039 | /// } | |||
6040 | /// \endcode | |||
6041 | static llvm::Value *emitReduceCombFunction(CodeGenModule &CGM, | |||
6042 | SourceLocation Loc, | |||
6043 | ReductionCodeGen &RCG, unsigned N, | |||
6044 | const Expr *ReductionOp, | |||
6045 | const Expr *LHS, const Expr *RHS, | |||
6046 | const Expr *PrivateRef) { | |||
6047 | ASTContext &C = CGM.getContext(); | |||
6048 | const auto *LHSVD = cast<VarDecl>(cast<DeclRefExpr>(LHS)->getDecl()); | |||
6049 | const auto *RHSVD = cast<VarDecl>(cast<DeclRefExpr>(RHS)->getDecl()); | |||
6050 | FunctionArgList Args; | |||
6051 | ImplicitParamDecl ParamInOut(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, | |||
6052 | C.VoidPtrTy, ImplicitParamDecl::Other); | |||
6053 | ImplicitParamDecl ParamIn(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
6054 | ImplicitParamDecl::Other); | |||
6055 | Args.emplace_back(&ParamInOut); | |||
6056 | Args.emplace_back(&ParamIn); | |||
6057 | const auto &FnInfo = | |||
6058 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
6059 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | |||
6060 | std::string Name = CGM.getOpenMPRuntime().getName({"red_comb", ""}); | |||
6061 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | |||
6062 | Name, &CGM.getModule()); | |||
6063 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | |||
6064 | Fn->setDoesNotRecurse(); | |||
6065 | CodeGenFunction CGF(CGM); | |||
6066 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc); | |||
6067 | llvm::Value *Size = nullptr; | |||
6068 | // If the size of the reduction item is non-constant, load it from global | |||
6069 | // threadprivate variable. | |||
6070 | if (RCG.getSizes(N).second) { | |||
6071 | Address SizeAddr = CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate( | |||
6072 | CGF, CGM.getContext().getSizeType(), | |||
6073 | generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N))); | |||
6074 | Size = CGF.EmitLoadOfScalar(SizeAddr, /*Volatile=*/false, | |||
6075 | CGM.getContext().getSizeType(), Loc); | |||
6076 | } | |||
6077 | RCG.emitAggregateType(CGF, N, Size); | |||
6078 | // Remap lhs and rhs variables to the addresses of the function arguments. | |||
6079 | // %lhs = bitcast void* %arg0 to <type>* | |||
6080 | // %rhs = bitcast void* %arg1 to <type>* | |||
6081 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); | |||
6082 | PrivateScope.addPrivate(LHSVD, [&C, &CGF, &ParamInOut, LHSVD]() { | |||
6083 | // Pull out the pointer to the variable. | |||
6084 | Address PtrAddr = CGF.EmitLoadOfPointer( | |||
6085 | CGF.GetAddrOfLocalVar(&ParamInOut), | |||
6086 | C.getPointerType(C.VoidPtrTy).castAs<PointerType>()); | |||
6087 | return CGF.Builder.CreateElementBitCast( | |||
6088 | PtrAddr, CGF.ConvertTypeForMem(LHSVD->getType())); | |||
6089 | }); | |||
6090 | PrivateScope.addPrivate(RHSVD, [&C, &CGF, &ParamIn, RHSVD]() { | |||
6091 | // Pull out the pointer to the variable. | |||
6092 | Address PtrAddr = CGF.EmitLoadOfPointer( | |||
6093 | CGF.GetAddrOfLocalVar(&ParamIn), | |||
6094 | C.getPointerType(C.VoidPtrTy).castAs<PointerType>()); | |||
6095 | return CGF.Builder.CreateElementBitCast( | |||
6096 | PtrAddr, CGF.ConvertTypeForMem(RHSVD->getType())); | |||
6097 | }); | |||
6098 | PrivateScope.Privatize(); | |||
6099 | // Emit the combiner body: | |||
6100 | // %2 = <ReductionOp>(<type> *%lhs, <type> *%rhs) | |||
6101 | // store <type> %2, <type>* %lhs | |||
6102 | CGM.getOpenMPRuntime().emitSingleReductionCombiner( | |||
6103 | CGF, ReductionOp, PrivateRef, cast<DeclRefExpr>(LHS), | |||
6104 | cast<DeclRefExpr>(RHS)); | |||
6105 | CGF.FinishFunction(); | |||
6106 | return Fn; | |||
6107 | } | |||
6108 | ||||
6109 | /// Emits reduction finalizer function: | |||
6110 | /// \code | |||
6111 | /// void @.red_fini(void* %arg) { | |||
6112 | /// %0 = bitcast void* %arg to <type>* | |||
6113 | /// <destroy>(<type>* %0) | |||
6114 | /// ret void | |||
6115 | /// } | |||
6116 | /// \endcode | |||
6117 | static llvm::Value *emitReduceFiniFunction(CodeGenModule &CGM, | |||
6118 | SourceLocation Loc, | |||
6119 | ReductionCodeGen &RCG, unsigned N) { | |||
6120 | if (!RCG.needCleanups(N)) | |||
6121 | return nullptr; | |||
6122 | ASTContext &C = CGM.getContext(); | |||
6123 | FunctionArgList Args; | |||
6124 | ImplicitParamDecl Param(C, /*DC=*/nullptr, Loc, /*Id=*/nullptr, C.VoidPtrTy, | |||
6125 | ImplicitParamDecl::Other); | |||
6126 | Args.emplace_back(&Param); | |||
6127 | const auto &FnInfo = | |||
6128 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(C.VoidTy, Args); | |||
6129 | llvm::FunctionType *FnTy = CGM.getTypes().GetFunctionType(FnInfo); | |||
6130 | std::string Name = CGM.getOpenMPRuntime().getName({"red_fini", ""}); | |||
6131 | auto *Fn = llvm::Function::Create(FnTy, llvm::GlobalValue::InternalLinkage, | |||
6132 | Name, &CGM.getModule()); | |||
6133 | CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FnInfo); | |||
6134 | Fn->setDoesNotRecurse(); | |||
6135 | CodeGenFunction CGF(CGM); | |||
6136 | CGF.StartFunction(GlobalDecl(), C.VoidTy, Fn, FnInfo, Args, Loc, Loc); | |||
6137 | Address PrivateAddr = CGF.EmitLoadOfPointer( | |||
6138 | CGF.GetAddrOfLocalVar(&Param), | |||
6139 | C.getPointerType(C.VoidPtrTy).castAs<PointerType>()); | |||
6140 | llvm::Value *Size = nullptr; | |||
6141 | // If the size of the reduction item is non-constant, load it from global | |||
6142 | // threadprivate variable. | |||
6143 | if (RCG.getSizes(N).second) { | |||
6144 | Address SizeAddr = CGM.getOpenMPRuntime().getAddrOfArtificialThreadPrivate( | |||
6145 | CGF, CGM.getContext().getSizeType(), | |||
6146 | generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N))); | |||
6147 | Size = CGF.EmitLoadOfScalar(SizeAddr, /*Volatile=*/false, | |||
6148 | CGM.getContext().getSizeType(), Loc); | |||
6149 | } | |||
6150 | RCG.emitAggregateType(CGF, N, Size); | |||
6151 | // Emit the finalizer body: | |||
6152 | // <destroy>(<type>* %0) | |||
6153 | RCG.emitCleanups(CGF, N, PrivateAddr); | |||
6154 | CGF.FinishFunction(); | |||
6155 | return Fn; | |||
6156 | } | |||
6157 | ||||
6158 | llvm::Value *CGOpenMPRuntime::emitTaskReductionInit( | |||
6159 | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> LHSExprs, | |||
6160 | ArrayRef<const Expr *> RHSExprs, const OMPTaskDataTy &Data) { | |||
6161 | if (!CGF.HaveInsertPoint() || Data.ReductionVars.empty()) | |||
6162 | return nullptr; | |||
6163 | ||||
6164 | // Build typedef struct: | |||
6165 | // kmp_task_red_input { | |||
6166 | // void *reduce_shar; // shared reduction item | |||
6167 | // size_t reduce_size; // size of data item | |||
6168 | // void *reduce_init; // data initialization routine | |||
6169 | // void *reduce_fini; // data finalization routine | |||
6170 | // void *reduce_comb; // data combiner routine | |||
6171 | // kmp_task_red_flags_t flags; // flags for additional info from compiler | |||
6172 | // } kmp_task_red_input_t; | |||
6173 | ASTContext &C = CGM.getContext(); | |||
6174 | RecordDecl *RD = C.buildImplicitRecord("kmp_task_red_input_t"); | |||
6175 | RD->startDefinition(); | |||
6176 | const FieldDecl *SharedFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
6177 | const FieldDecl *SizeFD = addFieldToRecordDecl(C, RD, C.getSizeType()); | |||
6178 | const FieldDecl *InitFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
6179 | const FieldDecl *FiniFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
6180 | const FieldDecl *CombFD = addFieldToRecordDecl(C, RD, C.VoidPtrTy); | |||
6181 | const FieldDecl *FlagsFD = addFieldToRecordDecl( | |||
6182 | C, RD, C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/false)); | |||
6183 | RD->completeDefinition(); | |||
6184 | QualType RDType = C.getRecordType(RD); | |||
6185 | unsigned Size = Data.ReductionVars.size(); | |||
6186 | llvm::APInt ArraySize(/*numBits=*/64, Size); | |||
6187 | QualType ArrayRDType = C.getConstantArrayType( | |||
6188 | RDType, ArraySize, ArrayType::Normal, /*IndexTypeQuals=*/0); | |||
6189 | // kmp_task_red_input_t .rd_input.[Size]; | |||
6190 | Address TaskRedInput = CGF.CreateMemTemp(ArrayRDType, ".rd_input."); | |||
6191 | ReductionCodeGen RCG(Data.ReductionVars, Data.ReductionCopies, | |||
6192 | Data.ReductionOps); | |||
6193 | for (unsigned Cnt = 0; Cnt < Size; ++Cnt) { | |||
6194 | // kmp_task_red_input_t &ElemLVal = .rd_input.[Cnt]; | |||
6195 | llvm::Value *Idxs[] = {llvm::ConstantInt::get(CGM.SizeTy, /*V=*/0), | |||
6196 | llvm::ConstantInt::get(CGM.SizeTy, Cnt)}; | |||
6197 | llvm::Value *GEP = CGF.EmitCheckedInBoundsGEP( | |||
6198 | TaskRedInput.getPointer(), Idxs, | |||
6199 | /*SignedIndices=*/false, /*IsSubtraction=*/false, Loc, | |||
6200 | ".rd_input.gep."); | |||
6201 | LValue ElemLVal = CGF.MakeNaturalAlignAddrLValue(GEP, RDType); | |||
6202 | // ElemLVal.reduce_shar = &Shareds[Cnt]; | |||
6203 | LValue SharedLVal = CGF.EmitLValueForField(ElemLVal, SharedFD); | |||
6204 | RCG.emitSharedLValue(CGF, Cnt); | |||
6205 | llvm::Value *CastedShared = | |||
6206 | CGF.EmitCastToVoidPtr(RCG.getSharedLValue(Cnt).getPointer()); | |||
6207 | CGF.EmitStoreOfScalar(CastedShared, SharedLVal); | |||
6208 | RCG.emitAggregateType(CGF, Cnt); | |||
6209 | llvm::Value *SizeValInChars; | |||
6210 | llvm::Value *SizeVal; | |||
6211 | std::tie(SizeValInChars, SizeVal) = RCG.getSizes(Cnt); | |||
6212 | // We use delayed creation/initialization for VLAs, array sections and | |||
6213 | // custom reduction initializations. It is required because runtime does not | |||
6214 | // provide the way to pass the sizes of VLAs/array sections to | |||
6215 | // initializer/combiner/finalizer functions and does not pass the pointer to | |||
6216 | // original reduction item to the initializer. Instead threadprivate global | |||
6217 | // variables are used to store these values and use them in the functions. | |||
6218 | bool DelayedCreation = !!SizeVal; | |||
6219 | SizeValInChars = CGF.Builder.CreateIntCast(SizeValInChars, CGM.SizeTy, | |||
6220 | /*isSigned=*/false); | |||
6221 | LValue SizeLVal = CGF.EmitLValueForField(ElemLVal, SizeFD); | |||
6222 | CGF.EmitStoreOfScalar(SizeValInChars, SizeLVal); | |||
6223 | // ElemLVal.reduce_init = init; | |||
6224 | LValue InitLVal = CGF.EmitLValueForField(ElemLVal, InitFD); | |||
6225 | llvm::Value *InitAddr = | |||
6226 | CGF.EmitCastToVoidPtr(emitReduceInitFunction(CGM, Loc, RCG, Cnt)); | |||
6227 | CGF.EmitStoreOfScalar(InitAddr, InitLVal); | |||
6228 | DelayedCreation = DelayedCreation || RCG.usesReductionInitializer(Cnt); | |||
6229 | // ElemLVal.reduce_fini = fini; | |||
6230 | LValue FiniLVal = CGF.EmitLValueForField(ElemLVal, FiniFD); | |||
6231 | llvm::Value *Fini = emitReduceFiniFunction(CGM, Loc, RCG, Cnt); | |||
6232 | llvm::Value *FiniAddr = Fini | |||
6233 | ? CGF.EmitCastToVoidPtr(Fini) | |||
6234 | : llvm::ConstantPointerNull::get(CGM.VoidPtrTy); | |||
6235 | CGF.EmitStoreOfScalar(FiniAddr, FiniLVal); | |||
6236 | // ElemLVal.reduce_comb = comb; | |||
6237 | LValue CombLVal = CGF.EmitLValueForField(ElemLVal, CombFD); | |||
6238 | llvm::Value *CombAddr = CGF.EmitCastToVoidPtr(emitReduceCombFunction( | |||
6239 | CGM, Loc, RCG, Cnt, Data.ReductionOps[Cnt], LHSExprs[Cnt], | |||
6240 | RHSExprs[Cnt], Data.ReductionCopies[Cnt])); | |||
6241 | CGF.EmitStoreOfScalar(CombAddr, CombLVal); | |||
6242 | // ElemLVal.flags = 0; | |||
6243 | LValue FlagsLVal = CGF.EmitLValueForField(ElemLVal, FlagsFD); | |||
6244 | if (DelayedCreation) { | |||
6245 | CGF.EmitStoreOfScalar( | |||
6246 | llvm::ConstantInt::get(CGM.Int32Ty, /*V=*/1, /*IsSigned=*/true), | |||
6247 | FlagsLVal); | |||
6248 | } else | |||
6249 | CGF.EmitNullInitialization(FlagsLVal.getAddress(), FlagsLVal.getType()); | |||
6250 | } | |||
6251 | // Build call void *__kmpc_task_reduction_init(int gtid, int num_data, void | |||
6252 | // *data); | |||
6253 | llvm::Value *Args[] = { | |||
6254 | CGF.Builder.CreateIntCast(getThreadID(CGF, Loc), CGM.IntTy, | |||
6255 | /*isSigned=*/true), | |||
6256 | llvm::ConstantInt::get(CGM.IntTy, Size, /*isSigned=*/true), | |||
6257 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(TaskRedInput.getPointer(), | |||
6258 | CGM.VoidPtrTy)}; | |||
6259 | return CGF.EmitRuntimeCall( | |||
6260 | createRuntimeFunction(OMPRTL__kmpc_task_reduction_init), Args); | |||
6261 | } | |||
6262 | ||||
6263 | void CGOpenMPRuntime::emitTaskReductionFixups(CodeGenFunction &CGF, | |||
6264 | SourceLocation Loc, | |||
6265 | ReductionCodeGen &RCG, | |||
6266 | unsigned N) { | |||
6267 | auto Sizes = RCG.getSizes(N); | |||
6268 | // Emit threadprivate global variable if the type is non-constant | |||
6269 | // (Sizes.second = nullptr). | |||
6270 | if (Sizes.second) { | |||
6271 | llvm::Value *SizeVal = CGF.Builder.CreateIntCast(Sizes.second, CGM.SizeTy, | |||
6272 | /*isSigned=*/false); | |||
6273 | Address SizeAddr = getAddrOfArtificialThreadPrivate( | |||
6274 | CGF, CGM.getContext().getSizeType(), | |||
6275 | generateUniqueName(CGM, "reduction_size", RCG.getRefExpr(N))); | |||
6276 | CGF.Builder.CreateStore(SizeVal, SizeAddr, /*IsVolatile=*/false); | |||
6277 | } | |||
6278 | // Store address of the original reduction item if custom initializer is used. | |||
6279 | if (RCG.usesReductionInitializer(N)) { | |||
6280 | Address SharedAddr = getAddrOfArtificialThreadPrivate( | |||
6281 | CGF, CGM.getContext().VoidPtrTy, | |||
6282 | generateUniqueName(CGM, "reduction", RCG.getRefExpr(N))); | |||
6283 | CGF.Builder.CreateStore( | |||
6284 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
6285 | RCG.getSharedLValue(N).getPointer(), CGM.VoidPtrTy), | |||
6286 | SharedAddr, /*IsVolatile=*/false); | |||
6287 | } | |||
6288 | } | |||
6289 | ||||
6290 | Address CGOpenMPRuntime::getTaskReductionItem(CodeGenFunction &CGF, | |||
6291 | SourceLocation Loc, | |||
6292 | llvm::Value *ReductionsPtr, | |||
6293 | LValue SharedLVal) { | |||
6294 | // Build call void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void | |||
6295 | // *d); | |||
6296 | llvm::Value *Args[] = { | |||
6297 | CGF.Builder.CreateIntCast(getThreadID(CGF, Loc), CGM.IntTy, | |||
6298 | /*isSigned=*/true), | |||
6299 | ReductionsPtr, | |||
6300 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(SharedLVal.getPointer(), | |||
6301 | CGM.VoidPtrTy)}; | |||
6302 | return Address( | |||
6303 | CGF.EmitRuntimeCall( | |||
6304 | createRuntimeFunction(OMPRTL__kmpc_task_reduction_get_th_data), Args), | |||
6305 | SharedLVal.getAlignment()); | |||
6306 | } | |||
6307 | ||||
6308 | void CGOpenMPRuntime::emitTaskwaitCall(CodeGenFunction &CGF, | |||
6309 | SourceLocation Loc) { | |||
6310 | if (!CGF.HaveInsertPoint()) | |||
6311 | return; | |||
6312 | // Build call kmp_int32 __kmpc_omp_taskwait(ident_t *loc, kmp_int32 | |||
6313 | // global_tid); | |||
6314 | llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc)}; | |||
6315 | // Ignore return result until untied tasks are supported. | |||
6316 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_omp_taskwait), Args); | |||
6317 | if (auto *Region = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) | |||
6318 | Region->emitUntiedSwitch(CGF); | |||
6319 | } | |||
6320 | ||||
6321 | void CGOpenMPRuntime::emitInlinedDirective(CodeGenFunction &CGF, | |||
6322 | OpenMPDirectiveKind InnerKind, | |||
6323 | const RegionCodeGenTy &CodeGen, | |||
6324 | bool HasCancel) { | |||
6325 | if (!CGF.HaveInsertPoint()) | |||
6326 | return; | |||
6327 | InlinedOpenMPRegionRAII Region(CGF, CodeGen, InnerKind, HasCancel); | |||
6328 | CGF.CapturedStmtInfo->EmitBody(CGF, /*S=*/nullptr); | |||
6329 | } | |||
6330 | ||||
6331 | namespace { | |||
6332 | enum RTCancelKind { | |||
6333 | CancelNoreq = 0, | |||
6334 | CancelParallel = 1, | |||
6335 | CancelLoop = 2, | |||
6336 | CancelSections = 3, | |||
6337 | CancelTaskgroup = 4 | |||
6338 | }; | |||
6339 | } // anonymous namespace | |||
6340 | ||||
6341 | static RTCancelKind getCancellationKind(OpenMPDirectiveKind CancelRegion) { | |||
6342 | RTCancelKind CancelKind = CancelNoreq; | |||
6343 | if (CancelRegion == OMPD_parallel) | |||
6344 | CancelKind = CancelParallel; | |||
6345 | else if (CancelRegion == OMPD_for) | |||
6346 | CancelKind = CancelLoop; | |||
6347 | else if (CancelRegion == OMPD_sections) | |||
6348 | CancelKind = CancelSections; | |||
6349 | else { | |||
6350 | assert(CancelRegion == OMPD_taskgroup)((CancelRegion == OMPD_taskgroup) ? static_cast<void> ( 0) : __assert_fail ("CancelRegion == OMPD_taskgroup", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6350, __PRETTY_FUNCTION__)); | |||
6351 | CancelKind = CancelTaskgroup; | |||
6352 | } | |||
6353 | return CancelKind; | |||
6354 | } | |||
6355 | ||||
6356 | void CGOpenMPRuntime::emitCancellationPointCall( | |||
6357 | CodeGenFunction &CGF, SourceLocation Loc, | |||
6358 | OpenMPDirectiveKind CancelRegion) { | |||
6359 | if (!CGF.HaveInsertPoint()) | |||
6360 | return; | |||
6361 | // Build call kmp_int32 __kmpc_cancellationpoint(ident_t *loc, kmp_int32 | |||
6362 | // global_tid, kmp_int32 cncl_kind); | |||
6363 | if (auto *OMPRegionInfo = | |||
6364 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { | |||
6365 | // For 'cancellation point taskgroup', the task region info may not have a | |||
6366 | // cancel. This may instead happen in another adjacent task. | |||
6367 | if (CancelRegion == OMPD_taskgroup || OMPRegionInfo->hasCancel()) { | |||
6368 | llvm::Value *Args[] = { | |||
6369 | emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), | |||
6370 | CGF.Builder.getInt32(getCancellationKind(CancelRegion))}; | |||
6371 | // Ignore return result until untied tasks are supported. | |||
6372 | llvm::Value *Result = CGF.EmitRuntimeCall( | |||
6373 | createRuntimeFunction(OMPRTL__kmpc_cancellationpoint), Args); | |||
6374 | // if (__kmpc_cancellationpoint()) { | |||
6375 | // exit from construct; | |||
6376 | // } | |||
6377 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".cancel.exit"); | |||
6378 | llvm::BasicBlock *ContBB = CGF.createBasicBlock(".cancel.continue"); | |||
6379 | llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Result); | |||
6380 | CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB); | |||
6381 | CGF.EmitBlock(ExitBB); | |||
6382 | // exit from construct; | |||
6383 | CodeGenFunction::JumpDest CancelDest = | |||
6384 | CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind()); | |||
6385 | CGF.EmitBranchThroughCleanup(CancelDest); | |||
6386 | CGF.EmitBlock(ContBB, /*IsFinished=*/true); | |||
6387 | } | |||
6388 | } | |||
6389 | } | |||
6390 | ||||
6391 | void CGOpenMPRuntime::emitCancelCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
6392 | const Expr *IfCond, | |||
6393 | OpenMPDirectiveKind CancelRegion) { | |||
6394 | if (!CGF.HaveInsertPoint()) | |||
6395 | return; | |||
6396 | // Build call kmp_int32 __kmpc_cancel(ident_t *loc, kmp_int32 global_tid, | |||
6397 | // kmp_int32 cncl_kind); | |||
6398 | if (auto *OMPRegionInfo = | |||
6399 | dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { | |||
6400 | auto &&ThenGen = [Loc, CancelRegion, OMPRegionInfo](CodeGenFunction &CGF, | |||
6401 | PrePostActionTy &) { | |||
6402 | CGOpenMPRuntime &RT = CGF.CGM.getOpenMPRuntime(); | |||
6403 | llvm::Value *Args[] = { | |||
6404 | RT.emitUpdateLocation(CGF, Loc), RT.getThreadID(CGF, Loc), | |||
6405 | CGF.Builder.getInt32(getCancellationKind(CancelRegion))}; | |||
6406 | // Ignore return result until untied tasks are supported. | |||
6407 | llvm::Value *Result = CGF.EmitRuntimeCall( | |||
6408 | RT.createRuntimeFunction(OMPRTL__kmpc_cancel), Args); | |||
6409 | // if (__kmpc_cancel()) { | |||
6410 | // exit from construct; | |||
6411 | // } | |||
6412 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".cancel.exit"); | |||
6413 | llvm::BasicBlock *ContBB = CGF.createBasicBlock(".cancel.continue"); | |||
6414 | llvm::Value *Cmp = CGF.Builder.CreateIsNotNull(Result); | |||
6415 | CGF.Builder.CreateCondBr(Cmp, ExitBB, ContBB); | |||
6416 | CGF.EmitBlock(ExitBB); | |||
6417 | // exit from construct; | |||
6418 | CodeGenFunction::JumpDest CancelDest = | |||
6419 | CGF.getOMPCancelDestination(OMPRegionInfo->getDirectiveKind()); | |||
6420 | CGF.EmitBranchThroughCleanup(CancelDest); | |||
6421 | CGF.EmitBlock(ContBB, /*IsFinished=*/true); | |||
6422 | }; | |||
6423 | if (IfCond) { | |||
6424 | emitOMPIfClause(CGF, IfCond, ThenGen, | |||
6425 | [](CodeGenFunction &, PrePostActionTy &) {}); | |||
6426 | } else { | |||
6427 | RegionCodeGenTy ThenRCG(ThenGen); | |||
6428 | ThenRCG(CGF); | |||
6429 | } | |||
6430 | } | |||
6431 | } | |||
6432 | ||||
6433 | void CGOpenMPRuntime::emitTargetOutlinedFunction( | |||
6434 | const OMPExecutableDirective &D, StringRef ParentName, | |||
6435 | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, | |||
6436 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { | |||
6437 | assert(!ParentName.empty() && "Invalid target region parent name!")((!ParentName.empty() && "Invalid target region parent name!" ) ? static_cast<void> (0) : __assert_fail ("!ParentName.empty() && \"Invalid target region parent name!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6437, __PRETTY_FUNCTION__)); | |||
6438 | HasEmittedTargetRegion = true; | |||
6439 | emitTargetOutlinedFunctionHelper(D, ParentName, OutlinedFn, OutlinedFnID, | |||
6440 | IsOffloadEntry, CodeGen); | |||
6441 | } | |||
6442 | ||||
6443 | void CGOpenMPRuntime::emitTargetOutlinedFunctionHelper( | |||
6444 | const OMPExecutableDirective &D, StringRef ParentName, | |||
6445 | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, | |||
6446 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { | |||
6447 | // Create a unique name for the entry function using the source location | |||
6448 | // information of the current target region. The name will be something like: | |||
6449 | // | |||
6450 | // __omp_offloading_DD_FFFF_PP_lBB | |||
6451 | // | |||
6452 | // where DD_FFFF is an ID unique to the file (device and file IDs), PP is the | |||
6453 | // mangled name of the function that encloses the target region and BB is the | |||
6454 | // line number of the target region. | |||
6455 | ||||
6456 | unsigned DeviceID; | |||
6457 | unsigned FileID; | |||
6458 | unsigned Line; | |||
6459 | getTargetEntryUniqueInfo(CGM.getContext(), D.getBeginLoc(), DeviceID, FileID, | |||
6460 | Line); | |||
6461 | SmallString<64> EntryFnName; | |||
6462 | { | |||
6463 | llvm::raw_svector_ostream OS(EntryFnName); | |||
6464 | OS << "__omp_offloading" << llvm::format("_%x", DeviceID) | |||
6465 | << llvm::format("_%x_", FileID) << ParentName << "_l" << Line; | |||
6466 | } | |||
6467 | ||||
6468 | const CapturedStmt &CS = *D.getCapturedStmt(OMPD_target); | |||
6469 | ||||
6470 | CodeGenFunction CGF(CGM, true); | |||
6471 | CGOpenMPTargetRegionInfo CGInfo(CS, CodeGen, EntryFnName); | |||
6472 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
6473 | ||||
6474 | OutlinedFn = CGF.GenerateOpenMPCapturedStmtFunction(CS); | |||
6475 | ||||
6476 | // If this target outline function is not an offload entry, we don't need to | |||
6477 | // register it. | |||
6478 | if (!IsOffloadEntry) | |||
6479 | return; | |||
6480 | ||||
6481 | // The target region ID is used by the runtime library to identify the current | |||
6482 | // target region, so it only has to be unique and not necessarily point to | |||
6483 | // anything. It could be the pointer to the outlined function that implements | |||
6484 | // the target region, but we aren't using that so that the compiler doesn't | |||
6485 | // need to keep that, and could therefore inline the host function if proven | |||
6486 | // worthwhile during optimization. In the other hand, if emitting code for the | |||
6487 | // device, the ID has to be the function address so that it can retrieved from | |||
6488 | // the offloading entry and launched by the runtime library. We also mark the | |||
6489 | // outlined function to have external linkage in case we are emitting code for | |||
6490 | // the device, because these functions will be entry points to the device. | |||
6491 | ||||
6492 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
6493 | OutlinedFnID = llvm::ConstantExpr::getBitCast(OutlinedFn, CGM.Int8PtrTy); | |||
6494 | OutlinedFn->setLinkage(llvm::GlobalValue::WeakAnyLinkage); | |||
6495 | OutlinedFn->setDSOLocal(false); | |||
6496 | } else { | |||
6497 | std::string Name = getName({EntryFnName, "region_id"}); | |||
6498 | OutlinedFnID = new llvm::GlobalVariable( | |||
6499 | CGM.getModule(), CGM.Int8Ty, /*isConstant=*/true, | |||
6500 | llvm::GlobalValue::WeakAnyLinkage, | |||
6501 | llvm::Constant::getNullValue(CGM.Int8Ty), Name); | |||
6502 | } | |||
6503 | ||||
6504 | // Register the information for the entry associated with this target region. | |||
6505 | OffloadEntriesInfoManager.registerTargetRegionEntryInfo( | |||
6506 | DeviceID, FileID, ParentName, Line, OutlinedFn, OutlinedFnID, | |||
6507 | OffloadEntriesInfoManagerTy::OMPTargetRegionEntryTargetRegion); | |||
6508 | } | |||
6509 | ||||
6510 | /// Checks if the expression is constant or does not have non-trivial function | |||
6511 | /// calls. | |||
6512 | static bool isTrivial(ASTContext &Ctx, const Expr * E) { | |||
6513 | // We can skip constant expressions. | |||
6514 | // We can skip expressions with trivial calls or simple expressions. | |||
6515 | return (E->isEvaluatable(Ctx, Expr::SE_AllowUndefinedBehavior) || | |||
6516 | !E->hasNonTrivialCall(Ctx)) && | |||
6517 | !E->HasSideEffects(Ctx, /*IncludePossibleEffects=*/true); | |||
6518 | } | |||
6519 | ||||
6520 | const Stmt *CGOpenMPRuntime::getSingleCompoundChild(ASTContext &Ctx, | |||
6521 | const Stmt *Body) { | |||
6522 | const Stmt *Child = Body->IgnoreContainers(); | |||
6523 | while (const auto *C = dyn_cast_or_null<CompoundStmt>(Child)) { | |||
6524 | Child = nullptr; | |||
6525 | for (const Stmt *S : C->body()) { | |||
6526 | if (const auto *E = dyn_cast<Expr>(S)) { | |||
6527 | if (isTrivial(Ctx, E)) | |||
6528 | continue; | |||
6529 | } | |||
6530 | // Some of the statements can be ignored. | |||
6531 | if (isa<AsmStmt>(S) || isa<NullStmt>(S) || isa<OMPFlushDirective>(S) || | |||
6532 | isa<OMPBarrierDirective>(S) || isa<OMPTaskyieldDirective>(S)) | |||
6533 | continue; | |||
6534 | // Analyze declarations. | |||
6535 | if (const auto *DS = dyn_cast<DeclStmt>(S)) { | |||
6536 | if (llvm::all_of(DS->decls(), [&Ctx](const Decl *D) { | |||
6537 | if (isa<EmptyDecl>(D) || isa<DeclContext>(D) || | |||
6538 | isa<TypeDecl>(D) || isa<PragmaCommentDecl>(D) || | |||
6539 | isa<PragmaDetectMismatchDecl>(D) || isa<UsingDecl>(D) || | |||
6540 | isa<UsingDirectiveDecl>(D) || | |||
6541 | isa<OMPDeclareReductionDecl>(D) || | |||
6542 | isa<OMPThreadPrivateDecl>(D) || isa<OMPAllocateDecl>(D)) | |||
6543 | return true; | |||
6544 | const auto *VD = dyn_cast<VarDecl>(D); | |||
6545 | if (!VD) | |||
6546 | return false; | |||
6547 | return VD->isConstexpr() || | |||
6548 | ((VD->getType().isTrivialType(Ctx) || | |||
6549 | VD->getType()->isReferenceType()) && | |||
6550 | (!VD->hasInit() || isTrivial(Ctx, VD->getInit()))); | |||
6551 | })) | |||
6552 | continue; | |||
6553 | } | |||
6554 | // Found multiple children - cannot get the one child only. | |||
6555 | if (Child) | |||
6556 | return nullptr; | |||
6557 | Child = S; | |||
6558 | } | |||
6559 | if (Child) | |||
6560 | Child = Child->IgnoreContainers(); | |||
6561 | } | |||
6562 | return Child; | |||
6563 | } | |||
6564 | ||||
6565 | /// Emit the number of teams for a target directive. Inspect the num_teams | |||
6566 | /// clause associated with a teams construct combined or closely nested | |||
6567 | /// with the target directive. | |||
6568 | /// | |||
6569 | /// Emit a team of size one for directives such as 'target parallel' that | |||
6570 | /// have no associated teams construct. | |||
6571 | /// | |||
6572 | /// Otherwise, return nullptr. | |||
6573 | static llvm::Value * | |||
6574 | emitNumTeamsForTargetDirective(CodeGenFunction &CGF, | |||
6575 | const OMPExecutableDirective &D) { | |||
6576 | assert(!CGF.getLangOpts().OpenMPIsDevice &&((!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? static_cast<void> (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6578, __PRETTY_FUNCTION__)) | |||
6577 | "Clauses associated with the teams directive expected to be emitted "((!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? static_cast<void> (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6578, __PRETTY_FUNCTION__)) | |||
6578 | "only for the host!")((!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? static_cast<void> (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6578, __PRETTY_FUNCTION__)); | |||
6579 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); | |||
6580 | assert(isOpenMPTargetExecutionDirective(DirectiveKind) &&((isOpenMPTargetExecutionDirective(DirectiveKind) && "Expected target-based executable directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6581, __PRETTY_FUNCTION__)) | |||
6581 | "Expected target-based executable directive.")((isOpenMPTargetExecutionDirective(DirectiveKind) && "Expected target-based executable directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6581, __PRETTY_FUNCTION__)); | |||
6582 | CGBuilderTy &Bld = CGF.Builder; | |||
6583 | switch (DirectiveKind) { | |||
6584 | case OMPD_target: { | |||
6585 | const auto *CS = D.getInnermostCapturedStmt(); | |||
6586 | const auto *Body = | |||
6587 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); | |||
6588 | const Stmt *ChildStmt = | |||
6589 | CGOpenMPRuntime::getSingleCompoundChild(CGF.getContext(), Body); | |||
6590 | if (const auto *NestedDir = | |||
6591 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
6592 | if (isOpenMPTeamsDirective(NestedDir->getDirectiveKind())) { | |||
6593 | if (NestedDir->hasClausesOfKind<OMPNumTeamsClause>()) { | |||
6594 | CGOpenMPInnerExprInfo CGInfo(CGF, *CS); | |||
6595 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
6596 | const Expr *NumTeams = | |||
6597 | NestedDir->getSingleClause<OMPNumTeamsClause>()->getNumTeams(); | |||
6598 | llvm::Value *NumTeamsVal = | |||
6599 | CGF.EmitScalarExpr(NumTeams, | |||
6600 | /*IgnoreResultAssign*/ true); | |||
6601 | return Bld.CreateIntCast(NumTeamsVal, CGF.Int32Ty, | |||
6602 | /*IsSigned=*/true); | |||
6603 | } | |||
6604 | return Bld.getInt32(0); | |||
6605 | } | |||
6606 | if (isOpenMPParallelDirective(NestedDir->getDirectiveKind()) || | |||
6607 | isOpenMPSimdDirective(NestedDir->getDirectiveKind())) | |||
6608 | return Bld.getInt32(1); | |||
6609 | return Bld.getInt32(0); | |||
6610 | } | |||
6611 | return nullptr; | |||
6612 | } | |||
6613 | case OMPD_target_teams: | |||
6614 | case OMPD_target_teams_distribute: | |||
6615 | case OMPD_target_teams_distribute_simd: | |||
6616 | case OMPD_target_teams_distribute_parallel_for: | |||
6617 | case OMPD_target_teams_distribute_parallel_for_simd: { | |||
6618 | if (D.hasClausesOfKind<OMPNumTeamsClause>()) { | |||
6619 | CodeGenFunction::RunCleanupsScope NumTeamsScope(CGF); | |||
6620 | const Expr *NumTeams = | |||
6621 | D.getSingleClause<OMPNumTeamsClause>()->getNumTeams(); | |||
6622 | llvm::Value *NumTeamsVal = | |||
6623 | CGF.EmitScalarExpr(NumTeams, | |||
6624 | /*IgnoreResultAssign*/ true); | |||
6625 | return Bld.CreateIntCast(NumTeamsVal, CGF.Int32Ty, | |||
6626 | /*IsSigned=*/true); | |||
6627 | } | |||
6628 | return Bld.getInt32(0); | |||
6629 | } | |||
6630 | case OMPD_target_parallel: | |||
6631 | case OMPD_target_parallel_for: | |||
6632 | case OMPD_target_parallel_for_simd: | |||
6633 | case OMPD_target_simd: | |||
6634 | return Bld.getInt32(1); | |||
6635 | case OMPD_parallel: | |||
6636 | case OMPD_for: | |||
6637 | case OMPD_parallel_for: | |||
6638 | case OMPD_parallel_sections: | |||
6639 | case OMPD_for_simd: | |||
6640 | case OMPD_parallel_for_simd: | |||
6641 | case OMPD_cancel: | |||
6642 | case OMPD_cancellation_point: | |||
6643 | case OMPD_ordered: | |||
6644 | case OMPD_threadprivate: | |||
6645 | case OMPD_allocate: | |||
6646 | case OMPD_task: | |||
6647 | case OMPD_simd: | |||
6648 | case OMPD_sections: | |||
6649 | case OMPD_section: | |||
6650 | case OMPD_single: | |||
6651 | case OMPD_master: | |||
6652 | case OMPD_critical: | |||
6653 | case OMPD_taskyield: | |||
6654 | case OMPD_barrier: | |||
6655 | case OMPD_taskwait: | |||
6656 | case OMPD_taskgroup: | |||
6657 | case OMPD_atomic: | |||
6658 | case OMPD_flush: | |||
6659 | case OMPD_teams: | |||
6660 | case OMPD_target_data: | |||
6661 | case OMPD_target_exit_data: | |||
6662 | case OMPD_target_enter_data: | |||
6663 | case OMPD_distribute: | |||
6664 | case OMPD_distribute_simd: | |||
6665 | case OMPD_distribute_parallel_for: | |||
6666 | case OMPD_distribute_parallel_for_simd: | |||
6667 | case OMPD_teams_distribute: | |||
6668 | case OMPD_teams_distribute_simd: | |||
6669 | case OMPD_teams_distribute_parallel_for: | |||
6670 | case OMPD_teams_distribute_parallel_for_simd: | |||
6671 | case OMPD_target_update: | |||
6672 | case OMPD_declare_simd: | |||
6673 | case OMPD_declare_target: | |||
6674 | case OMPD_end_declare_target: | |||
6675 | case OMPD_declare_reduction: | |||
6676 | case OMPD_declare_mapper: | |||
6677 | case OMPD_taskloop: | |||
6678 | case OMPD_taskloop_simd: | |||
6679 | case OMPD_requires: | |||
6680 | case OMPD_unknown: | |||
6681 | break; | |||
6682 | } | |||
6683 | llvm_unreachable("Unexpected directive kind.")::llvm::llvm_unreachable_internal("Unexpected directive kind." , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6683); | |||
6684 | } | |||
6685 | ||||
6686 | static llvm::Value *getNumThreads(CodeGenFunction &CGF, const CapturedStmt *CS, | |||
6687 | llvm::Value *DefaultThreadLimitVal) { | |||
6688 | const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild( | |||
6689 | CGF.getContext(), CS->getCapturedStmt()); | |||
6690 | if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) { | |||
6691 | if (isOpenMPParallelDirective(Dir->getDirectiveKind())) { | |||
6692 | llvm::Value *NumThreads = nullptr; | |||
6693 | llvm::Value *CondVal = nullptr; | |||
6694 | // Handle if clause. If if clause present, the number of threads is | |||
6695 | // calculated as <cond> ? (<numthreads> ? <numthreads> : 0 ) : 1. | |||
6696 | if (Dir->hasClausesOfKind<OMPIfClause>()) { | |||
6697 | CGOpenMPInnerExprInfo CGInfo(CGF, *CS); | |||
6698 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
6699 | const OMPIfClause *IfClause = nullptr; | |||
6700 | for (const auto *C : Dir->getClausesOfKind<OMPIfClause>()) { | |||
6701 | if (C->getNameModifier() == OMPD_unknown || | |||
6702 | C->getNameModifier() == OMPD_parallel) { | |||
6703 | IfClause = C; | |||
6704 | break; | |||
6705 | } | |||
6706 | } | |||
6707 | if (IfClause) { | |||
6708 | const Expr *Cond = IfClause->getCondition(); | |||
6709 | bool Result; | |||
6710 | if (Cond->EvaluateAsBooleanCondition(Result, CGF.getContext())) { | |||
6711 | if (!Result) | |||
6712 | return CGF.Builder.getInt32(1); | |||
6713 | } else { | |||
6714 | CodeGenFunction::LexicalScope Scope(CGF, Cond->getSourceRange()); | |||
6715 | if (const auto *PreInit = | |||
6716 | cast_or_null<DeclStmt>(IfClause->getPreInitStmt())) { | |||
6717 | for (const auto *I : PreInit->decls()) { | |||
6718 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { | |||
6719 | CGF.EmitVarDecl(cast<VarDecl>(*I)); | |||
6720 | } else { | |||
6721 | CodeGenFunction::AutoVarEmission Emission = | |||
6722 | CGF.EmitAutoVarAlloca(cast<VarDecl>(*I)); | |||
6723 | CGF.EmitAutoVarCleanups(Emission); | |||
6724 | } | |||
6725 | } | |||
6726 | } | |||
6727 | CondVal = CGF.EvaluateExprAsBool(Cond); | |||
6728 | } | |||
6729 | } | |||
6730 | } | |||
6731 | // Check the value of num_threads clause iff if clause was not specified | |||
6732 | // or is not evaluated to false. | |||
6733 | if (Dir->hasClausesOfKind<OMPNumThreadsClause>()) { | |||
6734 | CGOpenMPInnerExprInfo CGInfo(CGF, *CS); | |||
6735 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
6736 | const auto *NumThreadsClause = | |||
6737 | Dir->getSingleClause<OMPNumThreadsClause>(); | |||
6738 | CodeGenFunction::LexicalScope Scope( | |||
6739 | CGF, NumThreadsClause->getNumThreads()->getSourceRange()); | |||
6740 | if (const auto *PreInit = | |||
6741 | cast_or_null<DeclStmt>(NumThreadsClause->getPreInitStmt())) { | |||
6742 | for (const auto *I : PreInit->decls()) { | |||
6743 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { | |||
6744 | CGF.EmitVarDecl(cast<VarDecl>(*I)); | |||
6745 | } else { | |||
6746 | CodeGenFunction::AutoVarEmission Emission = | |||
6747 | CGF.EmitAutoVarAlloca(cast<VarDecl>(*I)); | |||
6748 | CGF.EmitAutoVarCleanups(Emission); | |||
6749 | } | |||
6750 | } | |||
6751 | } | |||
6752 | NumThreads = CGF.EmitScalarExpr(NumThreadsClause->getNumThreads()); | |||
6753 | NumThreads = CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty, | |||
6754 | /*IsSigned=*/false); | |||
6755 | if (DefaultThreadLimitVal) | |||
6756 | NumThreads = CGF.Builder.CreateSelect( | |||
6757 | CGF.Builder.CreateICmpULT(DefaultThreadLimitVal, NumThreads), | |||
6758 | DefaultThreadLimitVal, NumThreads); | |||
6759 | } else { | |||
6760 | NumThreads = DefaultThreadLimitVal ? DefaultThreadLimitVal | |||
6761 | : CGF.Builder.getInt32(0); | |||
6762 | } | |||
6763 | // Process condition of the if clause. | |||
6764 | if (CondVal) { | |||
6765 | NumThreads = CGF.Builder.CreateSelect(CondVal, NumThreads, | |||
6766 | CGF.Builder.getInt32(1)); | |||
6767 | } | |||
6768 | return NumThreads; | |||
6769 | } | |||
6770 | if (isOpenMPSimdDirective(Dir->getDirectiveKind())) | |||
6771 | return CGF.Builder.getInt32(1); | |||
6772 | return DefaultThreadLimitVal; | |||
6773 | } | |||
6774 | return DefaultThreadLimitVal ? DefaultThreadLimitVal | |||
6775 | : CGF.Builder.getInt32(0); | |||
6776 | } | |||
6777 | ||||
6778 | /// Emit the number of threads for a target directive. Inspect the | |||
6779 | /// thread_limit clause associated with a teams construct combined or closely | |||
6780 | /// nested with the target directive. | |||
6781 | /// | |||
6782 | /// Emit the num_threads clause for directives such as 'target parallel' that | |||
6783 | /// have no associated teams construct. | |||
6784 | /// | |||
6785 | /// Otherwise, return nullptr. | |||
6786 | static llvm::Value * | |||
6787 | emitNumThreadsForTargetDirective(CodeGenFunction &CGF, | |||
6788 | const OMPExecutableDirective &D) { | |||
6789 | assert(!CGF.getLangOpts().OpenMPIsDevice &&((!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? static_cast<void> (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6791, __PRETTY_FUNCTION__)) | |||
6790 | "Clauses associated with the teams directive expected to be emitted "((!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? static_cast<void> (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6791, __PRETTY_FUNCTION__)) | |||
6791 | "only for the host!")((!CGF.getLangOpts().OpenMPIsDevice && "Clauses associated with the teams directive expected to be emitted " "only for the host!") ? static_cast<void> (0) : __assert_fail ("!CGF.getLangOpts().OpenMPIsDevice && \"Clauses associated with the teams directive expected to be emitted \" \"only for the host!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6791, __PRETTY_FUNCTION__)); | |||
6792 | OpenMPDirectiveKind DirectiveKind = D.getDirectiveKind(); | |||
6793 | assert(isOpenMPTargetExecutionDirective(DirectiveKind) &&((isOpenMPTargetExecutionDirective(DirectiveKind) && "Expected target-based executable directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6794, __PRETTY_FUNCTION__)) | |||
6794 | "Expected target-based executable directive.")((isOpenMPTargetExecutionDirective(DirectiveKind) && "Expected target-based executable directive." ) ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(DirectiveKind) && \"Expected target-based executable directive.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6794, __PRETTY_FUNCTION__)); | |||
6795 | CGBuilderTy &Bld = CGF.Builder; | |||
6796 | llvm::Value *ThreadLimitVal = nullptr; | |||
6797 | llvm::Value *NumThreadsVal = nullptr; | |||
6798 | switch (DirectiveKind) { | |||
6799 | case OMPD_target: { | |||
6800 | const CapturedStmt *CS = D.getInnermostCapturedStmt(); | |||
6801 | if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal)) | |||
6802 | return NumThreads; | |||
6803 | const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild( | |||
6804 | CGF.getContext(), CS->getCapturedStmt()); | |||
6805 | if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) { | |||
6806 | if (Dir->hasClausesOfKind<OMPThreadLimitClause>()) { | |||
6807 | CGOpenMPInnerExprInfo CGInfo(CGF, *CS); | |||
6808 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo); | |||
6809 | const auto *ThreadLimitClause = | |||
6810 | Dir->getSingleClause<OMPThreadLimitClause>(); | |||
6811 | CodeGenFunction::LexicalScope Scope( | |||
6812 | CGF, ThreadLimitClause->getThreadLimit()->getSourceRange()); | |||
6813 | if (const auto *PreInit = | |||
6814 | cast_or_null<DeclStmt>(ThreadLimitClause->getPreInitStmt())) { | |||
6815 | for (const auto *I : PreInit->decls()) { | |||
6816 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { | |||
6817 | CGF.EmitVarDecl(cast<VarDecl>(*I)); | |||
6818 | } else { | |||
6819 | CodeGenFunction::AutoVarEmission Emission = | |||
6820 | CGF.EmitAutoVarAlloca(cast<VarDecl>(*I)); | |||
6821 | CGF.EmitAutoVarCleanups(Emission); | |||
6822 | } | |||
6823 | } | |||
6824 | } | |||
6825 | llvm::Value *ThreadLimit = CGF.EmitScalarExpr( | |||
6826 | ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true); | |||
6827 | ThreadLimitVal = | |||
6828 | Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*IsSigned=*/false); | |||
6829 | } | |||
6830 | if (isOpenMPTeamsDirective(Dir->getDirectiveKind()) && | |||
6831 | !isOpenMPDistributeDirective(Dir->getDirectiveKind())) { | |||
6832 | CS = Dir->getInnermostCapturedStmt(); | |||
6833 | const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild( | |||
6834 | CGF.getContext(), CS->getCapturedStmt()); | |||
6835 | Dir = dyn_cast_or_null<OMPExecutableDirective>(Child); | |||
6836 | } | |||
6837 | if (Dir && isOpenMPDistributeDirective(Dir->getDirectiveKind()) && | |||
6838 | !isOpenMPSimdDirective(Dir->getDirectiveKind())) { | |||
6839 | CS = Dir->getInnermostCapturedStmt(); | |||
6840 | if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal)) | |||
6841 | return NumThreads; | |||
6842 | } | |||
6843 | if (Dir && isOpenMPSimdDirective(Dir->getDirectiveKind())) | |||
6844 | return Bld.getInt32(1); | |||
6845 | } | |||
6846 | return ThreadLimitVal ? ThreadLimitVal : Bld.getInt32(0); | |||
6847 | } | |||
6848 | case OMPD_target_teams: { | |||
6849 | if (D.hasClausesOfKind<OMPThreadLimitClause>()) { | |||
6850 | CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF); | |||
6851 | const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>(); | |||
6852 | llvm::Value *ThreadLimit = CGF.EmitScalarExpr( | |||
6853 | ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true); | |||
6854 | ThreadLimitVal = | |||
6855 | Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*IsSigned=*/false); | |||
6856 | } | |||
6857 | const CapturedStmt *CS = D.getInnermostCapturedStmt(); | |||
6858 | if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal)) | |||
6859 | return NumThreads; | |||
6860 | const Stmt *Child = CGOpenMPRuntime::getSingleCompoundChild( | |||
6861 | CGF.getContext(), CS->getCapturedStmt()); | |||
6862 | if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) { | |||
6863 | if (Dir->getDirectiveKind() == OMPD_distribute) { | |||
6864 | CS = Dir->getInnermostCapturedStmt(); | |||
6865 | if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal)) | |||
6866 | return NumThreads; | |||
6867 | } | |||
6868 | } | |||
6869 | return ThreadLimitVal ? ThreadLimitVal : Bld.getInt32(0); | |||
6870 | } | |||
6871 | case OMPD_target_teams_distribute: | |||
6872 | if (D.hasClausesOfKind<OMPThreadLimitClause>()) { | |||
6873 | CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF); | |||
6874 | const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>(); | |||
6875 | llvm::Value *ThreadLimit = CGF.EmitScalarExpr( | |||
6876 | ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true); | |||
6877 | ThreadLimitVal = | |||
6878 | Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*IsSigned=*/false); | |||
6879 | } | |||
6880 | return getNumThreads(CGF, D.getInnermostCapturedStmt(), ThreadLimitVal); | |||
6881 | case OMPD_target_parallel: | |||
6882 | case OMPD_target_parallel_for: | |||
6883 | case OMPD_target_parallel_for_simd: | |||
6884 | case OMPD_target_teams_distribute_parallel_for: | |||
6885 | case OMPD_target_teams_distribute_parallel_for_simd: { | |||
6886 | llvm::Value *CondVal = nullptr; | |||
6887 | // Handle if clause. If if clause present, the number of threads is | |||
6888 | // calculated as <cond> ? (<numthreads> ? <numthreads> : 0 ) : 1. | |||
6889 | if (D.hasClausesOfKind<OMPIfClause>()) { | |||
6890 | const OMPIfClause *IfClause = nullptr; | |||
6891 | for (const auto *C : D.getClausesOfKind<OMPIfClause>()) { | |||
6892 | if (C->getNameModifier() == OMPD_unknown || | |||
6893 | C->getNameModifier() == OMPD_parallel) { | |||
6894 | IfClause = C; | |||
6895 | break; | |||
6896 | } | |||
6897 | } | |||
6898 | if (IfClause) { | |||
6899 | const Expr *Cond = IfClause->getCondition(); | |||
6900 | bool Result; | |||
6901 | if (Cond->EvaluateAsBooleanCondition(Result, CGF.getContext())) { | |||
6902 | if (!Result) | |||
6903 | return Bld.getInt32(1); | |||
6904 | } else { | |||
6905 | CodeGenFunction::RunCleanupsScope Scope(CGF); | |||
6906 | CondVal = CGF.EvaluateExprAsBool(Cond); | |||
6907 | } | |||
6908 | } | |||
6909 | } | |||
6910 | if (D.hasClausesOfKind<OMPThreadLimitClause>()) { | |||
6911 | CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF); | |||
6912 | const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>(); | |||
6913 | llvm::Value *ThreadLimit = CGF.EmitScalarExpr( | |||
6914 | ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true); | |||
6915 | ThreadLimitVal = | |||
6916 | Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*IsSigned=*/false); | |||
6917 | } | |||
6918 | if (D.hasClausesOfKind<OMPNumThreadsClause>()) { | |||
6919 | CodeGenFunction::RunCleanupsScope NumThreadsScope(CGF); | |||
6920 | const auto *NumThreadsClause = D.getSingleClause<OMPNumThreadsClause>(); | |||
6921 | llvm::Value *NumThreads = CGF.EmitScalarExpr( | |||
6922 | NumThreadsClause->getNumThreads(), /*IgnoreResultAssign=*/true); | |||
6923 | NumThreadsVal = | |||
6924 | Bld.CreateIntCast(NumThreads, CGF.Int32Ty, /*IsSigned=*/false); | |||
6925 | ThreadLimitVal = ThreadLimitVal | |||
6926 | ? Bld.CreateSelect(Bld.CreateICmpULT(NumThreadsVal, | |||
6927 | ThreadLimitVal), | |||
6928 | NumThreadsVal, ThreadLimitVal) | |||
6929 | : NumThreadsVal; | |||
6930 | } | |||
6931 | if (!ThreadLimitVal) | |||
6932 | ThreadLimitVal = Bld.getInt32(0); | |||
6933 | if (CondVal) | |||
6934 | return Bld.CreateSelect(CondVal, ThreadLimitVal, Bld.getInt32(1)); | |||
6935 | return ThreadLimitVal; | |||
6936 | } | |||
6937 | case OMPD_target_teams_distribute_simd: | |||
6938 | case OMPD_target_simd: | |||
6939 | return Bld.getInt32(1); | |||
6940 | case OMPD_parallel: | |||
6941 | case OMPD_for: | |||
6942 | case OMPD_parallel_for: | |||
6943 | case OMPD_parallel_sections: | |||
6944 | case OMPD_for_simd: | |||
6945 | case OMPD_parallel_for_simd: | |||
6946 | case OMPD_cancel: | |||
6947 | case OMPD_cancellation_point: | |||
6948 | case OMPD_ordered: | |||
6949 | case OMPD_threadprivate: | |||
6950 | case OMPD_allocate: | |||
6951 | case OMPD_task: | |||
6952 | case OMPD_simd: | |||
6953 | case OMPD_sections: | |||
6954 | case OMPD_section: | |||
6955 | case OMPD_single: | |||
6956 | case OMPD_master: | |||
6957 | case OMPD_critical: | |||
6958 | case OMPD_taskyield: | |||
6959 | case OMPD_barrier: | |||
6960 | case OMPD_taskwait: | |||
6961 | case OMPD_taskgroup: | |||
6962 | case OMPD_atomic: | |||
6963 | case OMPD_flush: | |||
6964 | case OMPD_teams: | |||
6965 | case OMPD_target_data: | |||
6966 | case OMPD_target_exit_data: | |||
6967 | case OMPD_target_enter_data: | |||
6968 | case OMPD_distribute: | |||
6969 | case OMPD_distribute_simd: | |||
6970 | case OMPD_distribute_parallel_for: | |||
6971 | case OMPD_distribute_parallel_for_simd: | |||
6972 | case OMPD_teams_distribute: | |||
6973 | case OMPD_teams_distribute_simd: | |||
6974 | case OMPD_teams_distribute_parallel_for: | |||
6975 | case OMPD_teams_distribute_parallel_for_simd: | |||
6976 | case OMPD_target_update: | |||
6977 | case OMPD_declare_simd: | |||
6978 | case OMPD_declare_target: | |||
6979 | case OMPD_end_declare_target: | |||
6980 | case OMPD_declare_reduction: | |||
6981 | case OMPD_declare_mapper: | |||
6982 | case OMPD_taskloop: | |||
6983 | case OMPD_taskloop_simd: | |||
6984 | case OMPD_requires: | |||
6985 | case OMPD_unknown: | |||
6986 | break; | |||
6987 | } | |||
6988 | llvm_unreachable("Unsupported directive kind.")::llvm::llvm_unreachable_internal("Unsupported directive kind." , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 6988); | |||
6989 | } | |||
6990 | ||||
6991 | namespace { | |||
6992 | LLVM_ENABLE_BITMASK_ENUMS_IN_NAMESPACE()using ::llvm::BitmaskEnumDetail::operator~; using ::llvm::BitmaskEnumDetail ::operator|; using ::llvm::BitmaskEnumDetail::operator&; using ::llvm::BitmaskEnumDetail::operator^; using ::llvm::BitmaskEnumDetail ::operator|=; using ::llvm::BitmaskEnumDetail::operator&= ; using ::llvm::BitmaskEnumDetail::operator^=; | |||
6993 | ||||
6994 | // Utility to handle information from clauses associated with a given | |||
6995 | // construct that use mappable expressions (e.g. 'map' clause, 'to' clause). | |||
6996 | // It provides a convenient interface to obtain the information and generate | |||
6997 | // code for that information. | |||
6998 | class MappableExprsHandler { | |||
6999 | public: | |||
7000 | /// Values for bit flags used to specify the mapping type for | |||
7001 | /// offloading. | |||
7002 | enum OpenMPOffloadMappingFlags : uint64_t { | |||
7003 | /// No flags | |||
7004 | OMP_MAP_NONE = 0x0, | |||
7005 | /// Allocate memory on the device and move data from host to device. | |||
7006 | OMP_MAP_TO = 0x01, | |||
7007 | /// Allocate memory on the device and move data from device to host. | |||
7008 | OMP_MAP_FROM = 0x02, | |||
7009 | /// Always perform the requested mapping action on the element, even | |||
7010 | /// if it was already mapped before. | |||
7011 | OMP_MAP_ALWAYS = 0x04, | |||
7012 | /// Delete the element from the device environment, ignoring the | |||
7013 | /// current reference count associated with the element. | |||
7014 | OMP_MAP_DELETE = 0x08, | |||
7015 | /// The element being mapped is a pointer-pointee pair; both the | |||
7016 | /// pointer and the pointee should be mapped. | |||
7017 | OMP_MAP_PTR_AND_OBJ = 0x10, | |||
7018 | /// This flags signals that the base address of an entry should be | |||
7019 | /// passed to the target kernel as an argument. | |||
7020 | OMP_MAP_TARGET_PARAM = 0x20, | |||
7021 | /// Signal that the runtime library has to return the device pointer | |||
7022 | /// in the current position for the data being mapped. Used when we have the | |||
7023 | /// use_device_ptr clause. | |||
7024 | OMP_MAP_RETURN_PARAM = 0x40, | |||
7025 | /// This flag signals that the reference being passed is a pointer to | |||
7026 | /// private data. | |||
7027 | OMP_MAP_PRIVATE = 0x80, | |||
7028 | /// Pass the element to the device by value. | |||
7029 | OMP_MAP_LITERAL = 0x100, | |||
7030 | /// Implicit map | |||
7031 | OMP_MAP_IMPLICIT = 0x200, | |||
7032 | /// The 16 MSBs of the flags indicate whether the entry is member of some | |||
7033 | /// struct/class. | |||
7034 | OMP_MAP_MEMBER_OF = 0xffff000000000000, | |||
7035 | LLVM_MARK_AS_BITMASK_ENUM(/* LargestFlag = */ OMP_MAP_MEMBER_OF)LLVM_BITMASK_LARGEST_ENUMERATOR = OMP_MAP_MEMBER_OF, | |||
7036 | }; | |||
7037 | ||||
7038 | /// Class that associates information with a base pointer to be passed to the | |||
7039 | /// runtime library. | |||
7040 | class BasePointerInfo { | |||
7041 | /// The base pointer. | |||
7042 | llvm::Value *Ptr = nullptr; | |||
7043 | /// The base declaration that refers to this device pointer, or null if | |||
7044 | /// there is none. | |||
7045 | const ValueDecl *DevPtrDecl = nullptr; | |||
7046 | ||||
7047 | public: | |||
7048 | BasePointerInfo(llvm::Value *Ptr, const ValueDecl *DevPtrDecl = nullptr) | |||
7049 | : Ptr(Ptr), DevPtrDecl(DevPtrDecl) {} | |||
7050 | llvm::Value *operator*() const { return Ptr; } | |||
7051 | const ValueDecl *getDevicePtrDecl() const { return DevPtrDecl; } | |||
7052 | void setDevicePtrDecl(const ValueDecl *D) { DevPtrDecl = D; } | |||
7053 | }; | |||
7054 | ||||
7055 | using MapBaseValuesArrayTy = SmallVector<BasePointerInfo, 4>; | |||
7056 | using MapValuesArrayTy = SmallVector<llvm::Value *, 4>; | |||
7057 | using MapFlagsArrayTy = SmallVector<OpenMPOffloadMappingFlags, 4>; | |||
7058 | ||||
7059 | /// Map between a struct and the its lowest & highest elements which have been | |||
7060 | /// mapped. | |||
7061 | /// [ValueDecl *] --> {LE(FieldIndex, Pointer), | |||
7062 | /// HE(FieldIndex, Pointer)} | |||
7063 | struct StructRangeInfoTy { | |||
7064 | std::pair<unsigned /*FieldIndex*/, Address /*Pointer*/> LowestElem = { | |||
7065 | 0, Address::invalid()}; | |||
7066 | std::pair<unsigned /*FieldIndex*/, Address /*Pointer*/> HighestElem = { | |||
7067 | 0, Address::invalid()}; | |||
7068 | Address Base = Address::invalid(); | |||
7069 | }; | |||
7070 | ||||
7071 | private: | |||
7072 | /// Kind that defines how a device pointer has to be returned. | |||
7073 | struct MapInfo { | |||
7074 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | |||
7075 | OpenMPMapClauseKind MapType = OMPC_MAP_unknown; | |||
7076 | ArrayRef<OpenMPMapModifierKind> MapModifiers; | |||
7077 | bool ReturnDevicePointer = false; | |||
7078 | bool IsImplicit = false; | |||
7079 | ||||
7080 | MapInfo() = default; | |||
7081 | MapInfo( | |||
7082 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components, | |||
7083 | OpenMPMapClauseKind MapType, | |||
7084 | ArrayRef<OpenMPMapModifierKind> MapModifiers, | |||
7085 | bool ReturnDevicePointer, bool IsImplicit) | |||
7086 | : Components(Components), MapType(MapType), MapModifiers(MapModifiers), | |||
7087 | ReturnDevicePointer(ReturnDevicePointer), IsImplicit(IsImplicit) {} | |||
7088 | }; | |||
7089 | ||||
7090 | /// If use_device_ptr is used on a pointer which is a struct member and there | |||
7091 | /// is no map information about it, then emission of that entry is deferred | |||
7092 | /// until the whole struct has been processed. | |||
7093 | struct DeferredDevicePtrEntryTy { | |||
7094 | const Expr *IE = nullptr; | |||
7095 | const ValueDecl *VD = nullptr; | |||
7096 | ||||
7097 | DeferredDevicePtrEntryTy(const Expr *IE, const ValueDecl *VD) | |||
7098 | : IE(IE), VD(VD) {} | |||
7099 | }; | |||
7100 | ||||
7101 | /// Directive from where the map clauses were extracted. | |||
7102 | const OMPExecutableDirective &CurDir; | |||
7103 | ||||
7104 | /// Function the directive is being generated for. | |||
7105 | CodeGenFunction &CGF; | |||
7106 | ||||
7107 | /// Set of all first private variables in the current directive. | |||
7108 | llvm::SmallPtrSet<const VarDecl *, 8> FirstPrivateDecls; | |||
7109 | ||||
7110 | /// Map between device pointer declarations and their expression components. | |||
7111 | /// The key value for declarations in 'this' is null. | |||
7112 | llvm::DenseMap< | |||
7113 | const ValueDecl *, | |||
7114 | SmallVector<OMPClauseMappableExprCommon::MappableExprComponentListRef, 4>> | |||
7115 | DevPointersMap; | |||
7116 | ||||
7117 | llvm::Value *getExprTypeSize(const Expr *E) const { | |||
7118 | QualType ExprTy = E->getType().getCanonicalType(); | |||
7119 | ||||
7120 | // Reference types are ignored for mapping purposes. | |||
7121 | if (const auto *RefTy = ExprTy->getAs<ReferenceType>()) | |||
7122 | ExprTy = RefTy->getPointeeType().getCanonicalType(); | |||
7123 | ||||
7124 | // Given that an array section is considered a built-in type, we need to | |||
7125 | // do the calculation based on the length of the section instead of relying | |||
7126 | // on CGF.getTypeSize(E->getType()). | |||
7127 | if (const auto *OAE = dyn_cast<OMPArraySectionExpr>(E)) { | |||
7128 | QualType BaseTy = OMPArraySectionExpr::getBaseOriginalType( | |||
7129 | OAE->getBase()->IgnoreParenImpCasts()) | |||
7130 | .getCanonicalType(); | |||
7131 | ||||
7132 | // If there is no length associated with the expression, that means we | |||
7133 | // are using the whole length of the base. | |||
7134 | if (!OAE->getLength() && OAE->getColonLoc().isValid()) | |||
7135 | return CGF.getTypeSize(BaseTy); | |||
7136 | ||||
7137 | llvm::Value *ElemSize; | |||
7138 | if (const auto *PTy = BaseTy->getAs<PointerType>()) { | |||
7139 | ElemSize = CGF.getTypeSize(PTy->getPointeeType().getCanonicalType()); | |||
7140 | } else { | |||
7141 | const auto *ATy = cast<ArrayType>(BaseTy.getTypePtr()); | |||
7142 | assert(ATy && "Expecting array type if not a pointer type.")((ATy && "Expecting array type if not a pointer type." ) ? static_cast<void> (0) : __assert_fail ("ATy && \"Expecting array type if not a pointer type.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7142, __PRETTY_FUNCTION__)); | |||
7143 | ElemSize = CGF.getTypeSize(ATy->getElementType().getCanonicalType()); | |||
7144 | } | |||
7145 | ||||
7146 | // If we don't have a length at this point, that is because we have an | |||
7147 | // array section with a single element. | |||
7148 | if (!OAE->getLength()) | |||
7149 | return ElemSize; | |||
7150 | ||||
7151 | llvm::Value *LengthVal = CGF.EmitScalarExpr(OAE->getLength()); | |||
7152 | LengthVal = | |||
7153 | CGF.Builder.CreateIntCast(LengthVal, CGF.SizeTy, /*isSigned=*/false); | |||
7154 | return CGF.Builder.CreateNUWMul(LengthVal, ElemSize); | |||
7155 | } | |||
7156 | return CGF.getTypeSize(ExprTy); | |||
7157 | } | |||
7158 | ||||
7159 | /// Return the corresponding bits for a given map clause modifier. Add | |||
7160 | /// a flag marking the map as a pointer if requested. Add a flag marking the | |||
7161 | /// map as the first one of a series of maps that relate to the same map | |||
7162 | /// expression. | |||
7163 | OpenMPOffloadMappingFlags getMapTypeBits( | |||
7164 | OpenMPMapClauseKind MapType, ArrayRef<OpenMPMapModifierKind> MapModifiers, | |||
7165 | bool IsImplicit, bool AddPtrFlag, bool AddIsTargetParamFlag) const { | |||
7166 | OpenMPOffloadMappingFlags Bits = | |||
7167 | IsImplicit ? OMP_MAP_IMPLICIT : OMP_MAP_NONE; | |||
7168 | switch (MapType) { | |||
7169 | case OMPC_MAP_alloc: | |||
7170 | case OMPC_MAP_release: | |||
7171 | // alloc and release is the default behavior in the runtime library, i.e. | |||
7172 | // if we don't pass any bits alloc/release that is what the runtime is | |||
7173 | // going to do. Therefore, we don't need to signal anything for these two | |||
7174 | // type modifiers. | |||
7175 | break; | |||
7176 | case OMPC_MAP_to: | |||
7177 | Bits |= OMP_MAP_TO; | |||
7178 | break; | |||
7179 | case OMPC_MAP_from: | |||
7180 | Bits |= OMP_MAP_FROM; | |||
7181 | break; | |||
7182 | case OMPC_MAP_tofrom: | |||
7183 | Bits |= OMP_MAP_TO | OMP_MAP_FROM; | |||
7184 | break; | |||
7185 | case OMPC_MAP_delete: | |||
7186 | Bits |= OMP_MAP_DELETE; | |||
7187 | break; | |||
7188 | case OMPC_MAP_unknown: | |||
7189 | llvm_unreachable("Unexpected map type!")::llvm::llvm_unreachable_internal("Unexpected map type!", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7189); | |||
7190 | } | |||
7191 | if (AddPtrFlag) | |||
7192 | Bits |= OMP_MAP_PTR_AND_OBJ; | |||
7193 | if (AddIsTargetParamFlag) | |||
7194 | Bits |= OMP_MAP_TARGET_PARAM; | |||
7195 | if (llvm::find(MapModifiers, OMPC_MAP_MODIFIER_always) | |||
7196 | != MapModifiers.end()) | |||
7197 | Bits |= OMP_MAP_ALWAYS; | |||
7198 | return Bits; | |||
7199 | } | |||
7200 | ||||
7201 | /// Return true if the provided expression is a final array section. A | |||
7202 | /// final array section, is one whose length can't be proved to be one. | |||
7203 | bool isFinalArraySectionExpression(const Expr *E) const { | |||
7204 | const auto *OASE = dyn_cast<OMPArraySectionExpr>(E); | |||
7205 | ||||
7206 | // It is not an array section and therefore not a unity-size one. | |||
7207 | if (!OASE) | |||
7208 | return false; | |||
7209 | ||||
7210 | // An array section with no colon always refer to a single element. | |||
7211 | if (OASE->getColonLoc().isInvalid()) | |||
7212 | return false; | |||
7213 | ||||
7214 | const Expr *Length = OASE->getLength(); | |||
7215 | ||||
7216 | // If we don't have a length we have to check if the array has size 1 | |||
7217 | // for this dimension. Also, we should always expect a length if the | |||
7218 | // base type is pointer. | |||
7219 | if (!Length) { | |||
7220 | QualType BaseQTy = OMPArraySectionExpr::getBaseOriginalType( | |||
7221 | OASE->getBase()->IgnoreParenImpCasts()) | |||
7222 | .getCanonicalType(); | |||
7223 | if (const auto *ATy = dyn_cast<ConstantArrayType>(BaseQTy.getTypePtr())) | |||
7224 | return ATy->getSize().getSExtValue() != 1; | |||
7225 | // If we don't have a constant dimension length, we have to consider | |||
7226 | // the current section as having any size, so it is not necessarily | |||
7227 | // unitary. If it happen to be unity size, that's user fault. | |||
7228 | return true; | |||
7229 | } | |||
7230 | ||||
7231 | // Check if the length evaluates to 1. | |||
7232 | Expr::EvalResult Result; | |||
7233 | if (!Length->EvaluateAsInt(Result, CGF.getContext())) | |||
7234 | return true; // Can have more that size 1. | |||
7235 | ||||
7236 | llvm::APSInt ConstLength = Result.Val.getInt(); | |||
7237 | return ConstLength.getSExtValue() != 1; | |||
7238 | } | |||
7239 | ||||
7240 | /// Generate the base pointers, section pointers, sizes and map type | |||
7241 | /// bits for the provided map type, map modifier, and expression components. | |||
7242 | /// \a IsFirstComponent should be set to true if the provided set of | |||
7243 | /// components is the first associated with a capture. | |||
7244 | void generateInfoForComponentList( | |||
7245 | OpenMPMapClauseKind MapType, | |||
7246 | ArrayRef<OpenMPMapModifierKind> MapModifiers, | |||
7247 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components, | |||
7248 | MapBaseValuesArrayTy &BasePointers, MapValuesArrayTy &Pointers, | |||
7249 | MapValuesArrayTy &Sizes, MapFlagsArrayTy &Types, | |||
7250 | StructRangeInfoTy &PartialStruct, bool IsFirstComponentList, | |||
7251 | bool IsImplicit, | |||
7252 | ArrayRef<OMPClauseMappableExprCommon::MappableExprComponentListRef> | |||
7253 | OverlappedElements = llvm::None) const { | |||
7254 | // The following summarizes what has to be generated for each map and the | |||
7255 | // types below. The generated information is expressed in this order: | |||
7256 | // base pointer, section pointer, size, flags | |||
7257 | // (to add to the ones that come from the map type and modifier). | |||
7258 | // | |||
7259 | // double d; | |||
7260 | // int i[100]; | |||
7261 | // float *p; | |||
7262 | // | |||
7263 | // struct S1 { | |||
7264 | // int i; | |||
7265 | // float f[50]; | |||
7266 | // } | |||
7267 | // struct S2 { | |||
7268 | // int i; | |||
7269 | // float f[50]; | |||
7270 | // S1 s; | |||
7271 | // double *p; | |||
7272 | // struct S2 *ps; | |||
7273 | // } | |||
7274 | // S2 s; | |||
7275 | // S2 *ps; | |||
7276 | // | |||
7277 | // map(d) | |||
7278 | // &d, &d, sizeof(double), TARGET_PARAM | TO | FROM | |||
7279 | // | |||
7280 | // map(i) | |||
7281 | // &i, &i, 100*sizeof(int), TARGET_PARAM | TO | FROM | |||
7282 | // | |||
7283 | // map(i[1:23]) | |||
7284 | // &i(=&i[0]), &i[1], 23*sizeof(int), TARGET_PARAM | TO | FROM | |||
7285 | // | |||
7286 | // map(p) | |||
7287 | // &p, &p, sizeof(float*), TARGET_PARAM | TO | FROM | |||
7288 | // | |||
7289 | // map(p[1:24]) | |||
7290 | // p, &p[1], 24*sizeof(float), TARGET_PARAM | TO | FROM | |||
7291 | // | |||
7292 | // map(s) | |||
7293 | // &s, &s, sizeof(S2), TARGET_PARAM | TO | FROM | |||
7294 | // | |||
7295 | // map(s.i) | |||
7296 | // &s, &(s.i), sizeof(int), TARGET_PARAM | TO | FROM | |||
7297 | // | |||
7298 | // map(s.s.f) | |||
7299 | // &s, &(s.s.f[0]), 50*sizeof(float), TARGET_PARAM | TO | FROM | |||
7300 | // | |||
7301 | // map(s.p) | |||
7302 | // &s, &(s.p), sizeof(double*), TARGET_PARAM | TO | FROM | |||
7303 | // | |||
7304 | // map(to: s.p[:22]) | |||
7305 | // &s, &(s.p), sizeof(double*), TARGET_PARAM (*) | |||
7306 | // &s, &(s.p), sizeof(double*), MEMBER_OF(1) (**) | |||
7307 | // &(s.p), &(s.p[0]), 22*sizeof(double), | |||
7308 | // MEMBER_OF(1) | PTR_AND_OBJ | TO (***) | |||
7309 | // (*) alloc space for struct members, only this is a target parameter | |||
7310 | // (**) map the pointer (nothing to be mapped in this example) (the compiler | |||
7311 | // optimizes this entry out, same in the examples below) | |||
7312 | // (***) map the pointee (map: to) | |||
7313 | // | |||
7314 | // map(s.ps) | |||
7315 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | TO | FROM | |||
7316 | // | |||
7317 | // map(from: s.ps->s.i) | |||
7318 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | |||
7319 | // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1) | |||
7320 | // &(s.ps), &(s.ps->s.i), sizeof(int), MEMBER_OF(1) | PTR_AND_OBJ | FROM | |||
7321 | // | |||
7322 | // map(to: s.ps->ps) | |||
7323 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | |||
7324 | // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1) | |||
7325 | // &(s.ps), &(s.ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | TO | |||
7326 | // | |||
7327 | // map(s.ps->ps->ps) | |||
7328 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | |||
7329 | // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1) | |||
7330 | // &(s.ps), &(s.ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | |||
7331 | // &(s.ps->ps), &(s.ps->ps->ps), sizeof(S2*), PTR_AND_OBJ | TO | FROM | |||
7332 | // | |||
7333 | // map(to: s.ps->ps->s.f[:22]) | |||
7334 | // &s, &(s.ps), sizeof(S2*), TARGET_PARAM | |||
7335 | // &s, &(s.ps), sizeof(S2*), MEMBER_OF(1) | |||
7336 | // &(s.ps), &(s.ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | |||
7337 | // &(s.ps->ps), &(s.ps->ps->s.f[0]), 22*sizeof(float), PTR_AND_OBJ | TO | |||
7338 | // | |||
7339 | // map(ps) | |||
7340 | // &ps, &ps, sizeof(S2*), TARGET_PARAM | TO | FROM | |||
7341 | // | |||
7342 | // map(ps->i) | |||
7343 | // ps, &(ps->i), sizeof(int), TARGET_PARAM | TO | FROM | |||
7344 | // | |||
7345 | // map(ps->s.f) | |||
7346 | // ps, &(ps->s.f[0]), 50*sizeof(float), TARGET_PARAM | TO | FROM | |||
7347 | // | |||
7348 | // map(from: ps->p) | |||
7349 | // ps, &(ps->p), sizeof(double*), TARGET_PARAM | FROM | |||
7350 | // | |||
7351 | // map(to: ps->p[:22]) | |||
7352 | // ps, &(ps->p), sizeof(double*), TARGET_PARAM | |||
7353 | // ps, &(ps->p), sizeof(double*), MEMBER_OF(1) | |||
7354 | // &(ps->p), &(ps->p[0]), 22*sizeof(double), MEMBER_OF(1) | PTR_AND_OBJ | TO | |||
7355 | // | |||
7356 | // map(ps->ps) | |||
7357 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | TO | FROM | |||
7358 | // | |||
7359 | // map(from: ps->ps->s.i) | |||
7360 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | |||
7361 | // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1) | |||
7362 | // &(ps->ps), &(ps->ps->s.i), sizeof(int), MEMBER_OF(1) | PTR_AND_OBJ | FROM | |||
7363 | // | |||
7364 | // map(from: ps->ps->ps) | |||
7365 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | |||
7366 | // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1) | |||
7367 | // &(ps->ps), &(ps->ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | FROM | |||
7368 | // | |||
7369 | // map(ps->ps->ps->ps) | |||
7370 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | |||
7371 | // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1) | |||
7372 | // &(ps->ps), &(ps->ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | |||
7373 | // &(ps->ps->ps), &(ps->ps->ps->ps), sizeof(S2*), PTR_AND_OBJ | TO | FROM | |||
7374 | // | |||
7375 | // map(to: ps->ps->ps->s.f[:22]) | |||
7376 | // ps, &(ps->ps), sizeof(S2*), TARGET_PARAM | |||
7377 | // ps, &(ps->ps), sizeof(S2*), MEMBER_OF(1) | |||
7378 | // &(ps->ps), &(ps->ps->ps), sizeof(S2*), MEMBER_OF(1) | PTR_AND_OBJ | |||
7379 | // &(ps->ps->ps), &(ps->ps->ps->s.f[0]), 22*sizeof(float), PTR_AND_OBJ | TO | |||
7380 | // | |||
7381 | // map(to: s.f[:22]) map(from: s.p[:33]) | |||
7382 | // &s, &(s.f[0]), 50*sizeof(float) + sizeof(struct S1) + | |||
7383 | // sizeof(double*) (**), TARGET_PARAM | |||
7384 | // &s, &(s.f[0]), 22*sizeof(float), MEMBER_OF(1) | TO | |||
7385 | // &s, &(s.p), sizeof(double*), MEMBER_OF(1) | |||
7386 | // &(s.p), &(s.p[0]), 33*sizeof(double), MEMBER_OF(1) | PTR_AND_OBJ | FROM | |||
7387 | // (*) allocate contiguous space needed to fit all mapped members even if | |||
7388 | // we allocate space for members not mapped (in this example, | |||
7389 | // s.f[22..49] and s.s are not mapped, yet we must allocate space for | |||
7390 | // them as well because they fall between &s.f[0] and &s.p) | |||
7391 | // | |||
7392 | // map(from: s.f[:22]) map(to: ps->p[:33]) | |||
7393 | // &s, &(s.f[0]), 22*sizeof(float), TARGET_PARAM | FROM | |||
7394 | // ps, &(ps->p), sizeof(S2*), TARGET_PARAM | |||
7395 | // ps, &(ps->p), sizeof(double*), MEMBER_OF(2) (*) | |||
7396 | // &(ps->p), &(ps->p[0]), 33*sizeof(double), MEMBER_OF(2) | PTR_AND_OBJ | TO | |||
7397 | // (*) the struct this entry pertains to is the 2nd element in the list of | |||
7398 | // arguments, hence MEMBER_OF(2) | |||
7399 | // | |||
7400 | // map(from: s.f[:22], s.s) map(to: ps->p[:33]) | |||
7401 | // &s, &(s.f[0]), 50*sizeof(float) + sizeof(struct S1), TARGET_PARAM | |||
7402 | // &s, &(s.f[0]), 22*sizeof(float), MEMBER_OF(1) | FROM | |||
7403 | // &s, &(s.s), sizeof(struct S1), MEMBER_OF(1) | FROM | |||
7404 | // ps, &(ps->p), sizeof(S2*), TARGET_PARAM | |||
7405 | // ps, &(ps->p), sizeof(double*), MEMBER_OF(4) (*) | |||
7406 | // &(ps->p), &(ps->p[0]), 33*sizeof(double), MEMBER_OF(4) | PTR_AND_OBJ | TO | |||
7407 | // (*) the struct this entry pertains to is the 4th element in the list | |||
7408 | // of arguments, hence MEMBER_OF(4) | |||
7409 | ||||
7410 | // Track if the map information being generated is the first for a capture. | |||
7411 | bool IsCaptureFirstInfo = IsFirstComponentList; | |||
7412 | bool IsLink = false; // Is this variable a "declare target link"? | |||
7413 | ||||
7414 | // Scan the components from the base to the complete expression. | |||
7415 | auto CI = Components.rbegin(); | |||
7416 | auto CE = Components.rend(); | |||
7417 | auto I = CI; | |||
7418 | ||||
7419 | // Track if the map information being generated is the first for a list of | |||
7420 | // components. | |||
7421 | bool IsExpressionFirstInfo = true; | |||
7422 | Address BP = Address::invalid(); | |||
7423 | const Expr *AssocExpr = I->getAssociatedExpression(); | |||
7424 | const auto *AE = dyn_cast<ArraySubscriptExpr>(AssocExpr); | |||
7425 | const auto *OASE = dyn_cast<OMPArraySectionExpr>(AssocExpr); | |||
7426 | ||||
7427 | if (isa<MemberExpr>(AssocExpr)) { | |||
7428 | // The base is the 'this' pointer. The content of the pointer is going | |||
7429 | // to be the base of the field being mapped. | |||
7430 | BP = CGF.LoadCXXThisAddress(); | |||
7431 | } else if ((AE && isa<CXXThisExpr>(AE->getBase()->IgnoreParenImpCasts())) || | |||
7432 | (OASE && | |||
7433 | isa<CXXThisExpr>(OASE->getBase()->IgnoreParenImpCasts()))) { | |||
7434 | BP = CGF.EmitOMPSharedLValue(AssocExpr).getAddress(); | |||
7435 | } else { | |||
7436 | // The base is the reference to the variable. | |||
7437 | // BP = &Var. | |||
7438 | BP = CGF.EmitOMPSharedLValue(AssocExpr).getAddress(); | |||
7439 | if (const auto *VD = | |||
7440 | dyn_cast_or_null<VarDecl>(I->getAssociatedDeclaration())) { | |||
7441 | if (llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
7442 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD)) | |||
7443 | if (*Res == OMPDeclareTargetDeclAttr::MT_Link) { | |||
7444 | IsLink = true; | |||
7445 | BP = CGF.CGM.getOpenMPRuntime().getAddrOfDeclareTargetLink(VD); | |||
7446 | } | |||
7447 | } | |||
7448 | ||||
7449 | // If the variable is a pointer and is being dereferenced (i.e. is not | |||
7450 | // the last component), the base has to be the pointer itself, not its | |||
7451 | // reference. References are ignored for mapping purposes. | |||
7452 | QualType Ty = | |||
7453 | I->getAssociatedDeclaration()->getType().getNonReferenceType(); | |||
7454 | if (Ty->isAnyPointerType() && std::next(I) != CE) { | |||
7455 | BP = CGF.EmitLoadOfPointer(BP, Ty->castAs<PointerType>()); | |||
7456 | ||||
7457 | // We do not need to generate individual map information for the | |||
7458 | // pointer, it can be associated with the combined storage. | |||
7459 | ++I; | |||
7460 | } | |||
7461 | } | |||
7462 | ||||
7463 | // Track whether a component of the list should be marked as MEMBER_OF some | |||
7464 | // combined entry (for partial structs). Only the first PTR_AND_OBJ entry | |||
7465 | // in a component list should be marked as MEMBER_OF, all subsequent entries | |||
7466 | // do not belong to the base struct. E.g. | |||
7467 | // struct S2 s; | |||
7468 | // s.ps->ps->ps->f[:] | |||
7469 | // (1) (2) (3) (4) | |||
7470 | // ps(1) is a member pointer, ps(2) is a pointee of ps(1), so it is a | |||
7471 | // PTR_AND_OBJ entry; the PTR is ps(1), so MEMBER_OF the base struct. ps(3) | |||
7472 | // is the pointee of ps(2) which is not member of struct s, so it should not | |||
7473 | // be marked as such (it is still PTR_AND_OBJ). | |||
7474 | // The variable is initialized to false so that PTR_AND_OBJ entries which | |||
7475 | // are not struct members are not considered (e.g. array of pointers to | |||
7476 | // data). | |||
7477 | bool ShouldBeMemberOf = false; | |||
7478 | ||||
7479 | // Variable keeping track of whether or not we have encountered a component | |||
7480 | // in the component list which is a member expression. Useful when we have a | |||
7481 | // pointer or a final array section, in which case it is the previous | |||
7482 | // component in the list which tells us whether we have a member expression. | |||
7483 | // E.g. X.f[:] | |||
7484 | // While processing the final array section "[:]" it is "f" which tells us | |||
7485 | // whether we are dealing with a member of a declared struct. | |||
7486 | const MemberExpr *EncounteredME = nullptr; | |||
7487 | ||||
7488 | for (; I != CE; ++I) { | |||
7489 | // If the current component is member of a struct (parent struct) mark it. | |||
7490 | if (!EncounteredME) { | |||
7491 | EncounteredME = dyn_cast<MemberExpr>(I->getAssociatedExpression()); | |||
7492 | // If we encounter a PTR_AND_OBJ entry from now on it should be marked | |||
7493 | // as MEMBER_OF the parent struct. | |||
7494 | if (EncounteredME) | |||
7495 | ShouldBeMemberOf = true; | |||
7496 | } | |||
7497 | ||||
7498 | auto Next = std::next(I); | |||
7499 | ||||
7500 | // We need to generate the addresses and sizes if this is the last | |||
7501 | // component, if the component is a pointer or if it is an array section | |||
7502 | // whose length can't be proved to be one. If this is a pointer, it | |||
7503 | // becomes the base address for the following components. | |||
7504 | ||||
7505 | // A final array section, is one whose length can't be proved to be one. | |||
7506 | bool IsFinalArraySection = | |||
7507 | isFinalArraySectionExpression(I->getAssociatedExpression()); | |||
7508 | ||||
7509 | // Get information on whether the element is a pointer. Have to do a | |||
7510 | // special treatment for array sections given that they are built-in | |||
7511 | // types. | |||
7512 | const auto *OASE = | |||
7513 | dyn_cast<OMPArraySectionExpr>(I->getAssociatedExpression()); | |||
7514 | bool IsPointer = | |||
7515 | (OASE && OMPArraySectionExpr::getBaseOriginalType(OASE) | |||
7516 | .getCanonicalType() | |||
7517 | ->isAnyPointerType()) || | |||
7518 | I->getAssociatedExpression()->getType()->isAnyPointerType(); | |||
7519 | ||||
7520 | if (Next == CE || IsPointer || IsFinalArraySection) { | |||
7521 | // If this is not the last component, we expect the pointer to be | |||
7522 | // associated with an array expression or member expression. | |||
7523 | assert((Next == CE ||(((Next == CE || isa<MemberExpr>(Next->getAssociatedExpression ()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression ()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression ())) && "Unexpected expression") ? static_cast<void > (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7527, __PRETTY_FUNCTION__)) | |||
7524 | isa<MemberExpr>(Next->getAssociatedExpression()) ||(((Next == CE || isa<MemberExpr>(Next->getAssociatedExpression ()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression ()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression ())) && "Unexpected expression") ? static_cast<void > (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7527, __PRETTY_FUNCTION__)) | |||
7525 | isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) ||(((Next == CE || isa<MemberExpr>(Next->getAssociatedExpression ()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression ()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression ())) && "Unexpected expression") ? static_cast<void > (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7527, __PRETTY_FUNCTION__)) | |||
7526 | isa<OMPArraySectionExpr>(Next->getAssociatedExpression())) &&(((Next == CE || isa<MemberExpr>(Next->getAssociatedExpression ()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression ()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression ())) && "Unexpected expression") ? static_cast<void > (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7527, __PRETTY_FUNCTION__)) | |||
7527 | "Unexpected expression")(((Next == CE || isa<MemberExpr>(Next->getAssociatedExpression ()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression ()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression ())) && "Unexpected expression") ? static_cast<void > (0) : __assert_fail ("(Next == CE || isa<MemberExpr>(Next->getAssociatedExpression()) || isa<ArraySubscriptExpr>(Next->getAssociatedExpression()) || isa<OMPArraySectionExpr>(Next->getAssociatedExpression())) && \"Unexpected expression\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7527, __PRETTY_FUNCTION__)); | |||
7528 | ||||
7529 | Address LB = | |||
7530 | CGF.EmitOMPSharedLValue(I->getAssociatedExpression()).getAddress(); | |||
7531 | ||||
7532 | // If this component is a pointer inside the base struct then we don't | |||
7533 | // need to create any entry for it - it will be combined with the object | |||
7534 | // it is pointing to into a single PTR_AND_OBJ entry. | |||
7535 | bool IsMemberPointer = | |||
7536 | IsPointer && EncounteredME && | |||
7537 | (dyn_cast<MemberExpr>(I->getAssociatedExpression()) == | |||
7538 | EncounteredME); | |||
7539 | if (!OverlappedElements.empty()) { | |||
7540 | // Handle base element with the info for overlapped elements. | |||
7541 | assert(!PartialStruct.Base.isValid() && "The base element is set.")((!PartialStruct.Base.isValid() && "The base element is set." ) ? static_cast<void> (0) : __assert_fail ("!PartialStruct.Base.isValid() && \"The base element is set.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7541, __PRETTY_FUNCTION__)); | |||
7542 | assert(Next == CE &&((Next == CE && "Expected last element for the overlapped elements." ) ? static_cast<void> (0) : __assert_fail ("Next == CE && \"Expected last element for the overlapped elements.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7543, __PRETTY_FUNCTION__)) | |||
7543 | "Expected last element for the overlapped elements.")((Next == CE && "Expected last element for the overlapped elements." ) ? static_cast<void> (0) : __assert_fail ("Next == CE && \"Expected last element for the overlapped elements.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7543, __PRETTY_FUNCTION__)); | |||
7544 | assert(!IsPointer &&((!IsPointer && "Unexpected base element with the pointer type." ) ? static_cast<void> (0) : __assert_fail ("!IsPointer && \"Unexpected base element with the pointer type.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7545, __PRETTY_FUNCTION__)) | |||
7545 | "Unexpected base element with the pointer type.")((!IsPointer && "Unexpected base element with the pointer type." ) ? static_cast<void> (0) : __assert_fail ("!IsPointer && \"Unexpected base element with the pointer type.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7545, __PRETTY_FUNCTION__)); | |||
7546 | // Mark the whole struct as the struct that requires allocation on the | |||
7547 | // device. | |||
7548 | PartialStruct.LowestElem = {0, LB}; | |||
7549 | CharUnits TypeSize = CGF.getContext().getTypeSizeInChars( | |||
7550 | I->getAssociatedExpression()->getType()); | |||
7551 | Address HB = CGF.Builder.CreateConstGEP( | |||
7552 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(LB, | |||
7553 | CGF.VoidPtrTy), | |||
7554 | TypeSize.getQuantity() - 1); | |||
7555 | PartialStruct.HighestElem = { | |||
7556 | std::numeric_limits<decltype( | |||
7557 | PartialStruct.HighestElem.first)>::max(), | |||
7558 | HB}; | |||
7559 | PartialStruct.Base = BP; | |||
7560 | // Emit data for non-overlapped data. | |||
7561 | OpenMPOffloadMappingFlags Flags = | |||
7562 | OMP_MAP_MEMBER_OF | | |||
7563 | getMapTypeBits(MapType, MapModifiers, IsImplicit, | |||
7564 | /*AddPtrFlag=*/false, | |||
7565 | /*AddIsTargetParamFlag=*/false); | |||
7566 | LB = BP; | |||
7567 | llvm::Value *Size = nullptr; | |||
7568 | // Do bitcopy of all non-overlapped structure elements. | |||
7569 | for (OMPClauseMappableExprCommon::MappableExprComponentListRef | |||
7570 | Component : OverlappedElements) { | |||
7571 | Address ComponentLB = Address::invalid(); | |||
7572 | for (const OMPClauseMappableExprCommon::MappableComponent &MC : | |||
7573 | Component) { | |||
7574 | if (MC.getAssociatedDeclaration()) { | |||
7575 | ComponentLB = | |||
7576 | CGF.EmitOMPSharedLValue(MC.getAssociatedExpression()) | |||
7577 | .getAddress(); | |||
7578 | Size = CGF.Builder.CreatePtrDiff( | |||
7579 | CGF.EmitCastToVoidPtr(ComponentLB.getPointer()), | |||
7580 | CGF.EmitCastToVoidPtr(LB.getPointer())); | |||
7581 | break; | |||
7582 | } | |||
7583 | } | |||
7584 | BasePointers.push_back(BP.getPointer()); | |||
7585 | Pointers.push_back(LB.getPointer()); | |||
7586 | Sizes.push_back(Size); | |||
7587 | Types.push_back(Flags); | |||
7588 | LB = CGF.Builder.CreateConstGEP(ComponentLB, 1); | |||
7589 | } | |||
7590 | BasePointers.push_back(BP.getPointer()); | |||
7591 | Pointers.push_back(LB.getPointer()); | |||
7592 | Size = CGF.Builder.CreatePtrDiff( | |||
7593 | CGF.EmitCastToVoidPtr( | |||
7594 | CGF.Builder.CreateConstGEP(HB, 1).getPointer()), | |||
7595 | CGF.EmitCastToVoidPtr(LB.getPointer())); | |||
7596 | Sizes.push_back(Size); | |||
7597 | Types.push_back(Flags); | |||
7598 | break; | |||
7599 | } | |||
7600 | llvm::Value *Size = getExprTypeSize(I->getAssociatedExpression()); | |||
7601 | if (!IsMemberPointer) { | |||
7602 | BasePointers.push_back(BP.getPointer()); | |||
7603 | Pointers.push_back(LB.getPointer()); | |||
7604 | Sizes.push_back(Size); | |||
7605 | ||||
7606 | // We need to add a pointer flag for each map that comes from the | |||
7607 | // same expression except for the first one. We also need to signal | |||
7608 | // this map is the first one that relates with the current capture | |||
7609 | // (there is a set of entries for each capture). | |||
7610 | OpenMPOffloadMappingFlags Flags = getMapTypeBits( | |||
7611 | MapType, MapModifiers, IsImplicit, | |||
7612 | !IsExpressionFirstInfo || IsLink, IsCaptureFirstInfo && !IsLink); | |||
7613 | ||||
7614 | if (!IsExpressionFirstInfo) { | |||
7615 | // If we have a PTR_AND_OBJ pair where the OBJ is a pointer as well, | |||
7616 | // then we reset the TO/FROM/ALWAYS/DELETE flags. | |||
7617 | if (IsPointer) | |||
7618 | Flags &= ~(OMP_MAP_TO | OMP_MAP_FROM | OMP_MAP_ALWAYS | | |||
7619 | OMP_MAP_DELETE); | |||
7620 | ||||
7621 | if (ShouldBeMemberOf) { | |||
7622 | // Set placeholder value MEMBER_OF=FFFF to indicate that the flag | |||
7623 | // should be later updated with the correct value of MEMBER_OF. | |||
7624 | Flags |= OMP_MAP_MEMBER_OF; | |||
7625 | // From now on, all subsequent PTR_AND_OBJ entries should not be | |||
7626 | // marked as MEMBER_OF. | |||
7627 | ShouldBeMemberOf = false; | |||
7628 | } | |||
7629 | } | |||
7630 | ||||
7631 | Types.push_back(Flags); | |||
7632 | } | |||
7633 | ||||
7634 | // If we have encountered a member expression so far, keep track of the | |||
7635 | // mapped member. If the parent is "*this", then the value declaration | |||
7636 | // is nullptr. | |||
7637 | if (EncounteredME) { | |||
7638 | const auto *FD = dyn_cast<FieldDecl>(EncounteredME->getMemberDecl()); | |||
7639 | unsigned FieldIndex = FD->getFieldIndex(); | |||
7640 | ||||
7641 | // Update info about the lowest and highest elements for this struct | |||
7642 | if (!PartialStruct.Base.isValid()) { | |||
7643 | PartialStruct.LowestElem = {FieldIndex, LB}; | |||
7644 | PartialStruct.HighestElem = {FieldIndex, LB}; | |||
7645 | PartialStruct.Base = BP; | |||
7646 | } else if (FieldIndex < PartialStruct.LowestElem.first) { | |||
7647 | PartialStruct.LowestElem = {FieldIndex, LB}; | |||
7648 | } else if (FieldIndex > PartialStruct.HighestElem.first) { | |||
7649 | PartialStruct.HighestElem = {FieldIndex, LB}; | |||
7650 | } | |||
7651 | } | |||
7652 | ||||
7653 | // If we have a final array section, we are done with this expression. | |||
7654 | if (IsFinalArraySection) | |||
7655 | break; | |||
7656 | ||||
7657 | // The pointer becomes the base for the next element. | |||
7658 | if (Next != CE) | |||
7659 | BP = LB; | |||
7660 | ||||
7661 | IsExpressionFirstInfo = false; | |||
7662 | IsCaptureFirstInfo = false; | |||
7663 | } | |||
7664 | } | |||
7665 | } | |||
7666 | ||||
7667 | /// Return the adjusted map modifiers if the declaration a capture refers to | |||
7668 | /// appears in a first-private clause. This is expected to be used only with | |||
7669 | /// directives that start with 'target'. | |||
7670 | MappableExprsHandler::OpenMPOffloadMappingFlags | |||
7671 | getMapModifiersForPrivateClauses(const CapturedStmt::Capture &Cap) const { | |||
7672 | assert(Cap.capturesVariable() && "Expected capture by reference only!")((Cap.capturesVariable() && "Expected capture by reference only!" ) ? static_cast<void> (0) : __assert_fail ("Cap.capturesVariable() && \"Expected capture by reference only!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7672, __PRETTY_FUNCTION__)); | |||
7673 | ||||
7674 | // A first private variable captured by reference will use only the | |||
7675 | // 'private ptr' and 'map to' flag. Return the right flags if the captured | |||
7676 | // declaration is known as first-private in this handler. | |||
7677 | if (FirstPrivateDecls.count(Cap.getCapturedVar())) { | |||
7678 | if (Cap.getCapturedVar()->getType().isConstant(CGF.getContext()) && | |||
7679 | Cap.getCaptureKind() == CapturedStmt::VCK_ByRef) | |||
7680 | return MappableExprsHandler::OMP_MAP_ALWAYS | | |||
7681 | MappableExprsHandler::OMP_MAP_TO; | |||
7682 | if (Cap.getCapturedVar()->getType()->isAnyPointerType()) | |||
7683 | return MappableExprsHandler::OMP_MAP_TO | | |||
7684 | MappableExprsHandler::OMP_MAP_PTR_AND_OBJ; | |||
7685 | return MappableExprsHandler::OMP_MAP_PRIVATE | | |||
7686 | MappableExprsHandler::OMP_MAP_TO; | |||
7687 | } | |||
7688 | return MappableExprsHandler::OMP_MAP_TO | | |||
7689 | MappableExprsHandler::OMP_MAP_FROM; | |||
7690 | } | |||
7691 | ||||
7692 | static OpenMPOffloadMappingFlags getMemberOfFlag(unsigned Position) { | |||
7693 | // Member of is given by the 16 MSB of the flag, so rotate by 48 bits. | |||
7694 | return static_cast<OpenMPOffloadMappingFlags>(((uint64_t)Position + 1) | |||
7695 | << 48); | |||
7696 | } | |||
7697 | ||||
7698 | static void setCorrectMemberOfFlag(OpenMPOffloadMappingFlags &Flags, | |||
7699 | OpenMPOffloadMappingFlags MemberOfFlag) { | |||
7700 | // If the entry is PTR_AND_OBJ but has not been marked with the special | |||
7701 | // placeholder value 0xFFFF in the MEMBER_OF field, then it should not be | |||
7702 | // marked as MEMBER_OF. | |||
7703 | if ((Flags & OMP_MAP_PTR_AND_OBJ) && | |||
7704 | ((Flags & OMP_MAP_MEMBER_OF) != OMP_MAP_MEMBER_OF)) | |||
7705 | return; | |||
7706 | ||||
7707 | // Reset the placeholder value to prepare the flag for the assignment of the | |||
7708 | // proper MEMBER_OF value. | |||
7709 | Flags &= ~OMP_MAP_MEMBER_OF; | |||
7710 | Flags |= MemberOfFlag; | |||
7711 | } | |||
7712 | ||||
7713 | void getPlainLayout(const CXXRecordDecl *RD, | |||
7714 | llvm::SmallVectorImpl<const FieldDecl *> &Layout, | |||
7715 | bool AsBase) const { | |||
7716 | const CGRecordLayout &RL = CGF.getTypes().getCGRecordLayout(RD); | |||
7717 | ||||
7718 | llvm::StructType *St = | |||
7719 | AsBase ? RL.getBaseSubobjectLLVMType() : RL.getLLVMType(); | |||
7720 | ||||
7721 | unsigned NumElements = St->getNumElements(); | |||
7722 | llvm::SmallVector< | |||
7723 | llvm::PointerUnion<const CXXRecordDecl *, const FieldDecl *>, 4> | |||
7724 | RecordLayout(NumElements); | |||
7725 | ||||
7726 | // Fill bases. | |||
7727 | for (const auto &I : RD->bases()) { | |||
7728 | if (I.isVirtual()) | |||
7729 | continue; | |||
7730 | const auto *Base = I.getType()->getAsCXXRecordDecl(); | |||
7731 | // Ignore empty bases. | |||
7732 | if (Base->isEmpty() || CGF.getContext() | |||
7733 | .getASTRecordLayout(Base) | |||
7734 | .getNonVirtualSize() | |||
7735 | .isZero()) | |||
7736 | continue; | |||
7737 | ||||
7738 | unsigned FieldIndex = RL.getNonVirtualBaseLLVMFieldNo(Base); | |||
7739 | RecordLayout[FieldIndex] = Base; | |||
7740 | } | |||
7741 | // Fill in virtual bases. | |||
7742 | for (const auto &I : RD->vbases()) { | |||
7743 | const auto *Base = I.getType()->getAsCXXRecordDecl(); | |||
7744 | // Ignore empty bases. | |||
7745 | if (Base->isEmpty()) | |||
7746 | continue; | |||
7747 | unsigned FieldIndex = RL.getVirtualBaseIndex(Base); | |||
7748 | if (RecordLayout[FieldIndex]) | |||
7749 | continue; | |||
7750 | RecordLayout[FieldIndex] = Base; | |||
7751 | } | |||
7752 | // Fill in all the fields. | |||
7753 | assert(!RD->isUnion() && "Unexpected union.")((!RD->isUnion() && "Unexpected union.") ? static_cast <void> (0) : __assert_fail ("!RD->isUnion() && \"Unexpected union.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7753, __PRETTY_FUNCTION__)); | |||
7754 | for (const auto *Field : RD->fields()) { | |||
7755 | // Fill in non-bitfields. (Bitfields always use a zero pattern, which we | |||
7756 | // will fill in later.) | |||
7757 | if (!Field->isBitField()) { | |||
7758 | unsigned FieldIndex = RL.getLLVMFieldNo(Field); | |||
7759 | RecordLayout[FieldIndex] = Field; | |||
7760 | } | |||
7761 | } | |||
7762 | for (const llvm::PointerUnion<const CXXRecordDecl *, const FieldDecl *> | |||
7763 | &Data : RecordLayout) { | |||
7764 | if (Data.isNull()) | |||
7765 | continue; | |||
7766 | if (const auto *Base = Data.dyn_cast<const CXXRecordDecl *>()) | |||
7767 | getPlainLayout(Base, Layout, /*AsBase=*/true); | |||
7768 | else | |||
7769 | Layout.push_back(Data.get<const FieldDecl *>()); | |||
7770 | } | |||
7771 | } | |||
7772 | ||||
7773 | public: | |||
7774 | MappableExprsHandler(const OMPExecutableDirective &Dir, CodeGenFunction &CGF) | |||
7775 | : CurDir(Dir), CGF(CGF) { | |||
7776 | // Extract firstprivate clause information. | |||
7777 | for (const auto *C : Dir.getClausesOfKind<OMPFirstprivateClause>()) | |||
7778 | for (const auto *D : C->varlists()) | |||
7779 | FirstPrivateDecls.insert( | |||
7780 | cast<VarDecl>(cast<DeclRefExpr>(D)->getDecl())->getCanonicalDecl()); | |||
7781 | // Extract device pointer clause information. | |||
7782 | for (const auto *C : Dir.getClausesOfKind<OMPIsDevicePtrClause>()) | |||
7783 | for (auto L : C->component_lists()) | |||
7784 | DevPointersMap[L.first].push_back(L.second); | |||
7785 | } | |||
7786 | ||||
7787 | /// Generate code for the combined entry if we have a partially mapped struct | |||
7788 | /// and take care of the mapping flags of the arguments corresponding to | |||
7789 | /// individual struct members. | |||
7790 | void emitCombinedEntry(MapBaseValuesArrayTy &BasePointers, | |||
7791 | MapValuesArrayTy &Pointers, MapValuesArrayTy &Sizes, | |||
7792 | MapFlagsArrayTy &Types, MapFlagsArrayTy &CurTypes, | |||
7793 | const StructRangeInfoTy &PartialStruct) const { | |||
7794 | // Base is the base of the struct | |||
7795 | BasePointers.push_back(PartialStruct.Base.getPointer()); | |||
7796 | // Pointer is the address of the lowest element | |||
7797 | llvm::Value *LB = PartialStruct.LowestElem.second.getPointer(); | |||
7798 | Pointers.push_back(LB); | |||
7799 | // Size is (addr of {highest+1} element) - (addr of lowest element) | |||
7800 | llvm::Value *HB = PartialStruct.HighestElem.second.getPointer(); | |||
7801 | llvm::Value *HAddr = CGF.Builder.CreateConstGEP1_32(HB, /*Idx0=*/1); | |||
7802 | llvm::Value *CLAddr = CGF.Builder.CreatePointerCast(LB, CGF.VoidPtrTy); | |||
7803 | llvm::Value *CHAddr = CGF.Builder.CreatePointerCast(HAddr, CGF.VoidPtrTy); | |||
7804 | llvm::Value *Diff = CGF.Builder.CreatePtrDiff(CHAddr, CLAddr); | |||
7805 | llvm::Value *Size = CGF.Builder.CreateIntCast(Diff, CGF.SizeTy, | |||
7806 | /*isSinged=*/false); | |||
7807 | Sizes.push_back(Size); | |||
7808 | // Map type is always TARGET_PARAM | |||
7809 | Types.push_back(OMP_MAP_TARGET_PARAM); | |||
7810 | // Remove TARGET_PARAM flag from the first element | |||
7811 | (*CurTypes.begin()) &= ~OMP_MAP_TARGET_PARAM; | |||
7812 | ||||
7813 | // All other current entries will be MEMBER_OF the combined entry | |||
7814 | // (except for PTR_AND_OBJ entries which do not have a placeholder value | |||
7815 | // 0xFFFF in the MEMBER_OF field). | |||
7816 | OpenMPOffloadMappingFlags MemberOfFlag = | |||
7817 | getMemberOfFlag(BasePointers.size() - 1); | |||
7818 | for (auto &M : CurTypes) | |||
7819 | setCorrectMemberOfFlag(M, MemberOfFlag); | |||
7820 | } | |||
7821 | ||||
7822 | /// Generate all the base pointers, section pointers, sizes and map | |||
7823 | /// types for the extracted mappable expressions. Also, for each item that | |||
7824 | /// relates with a device pointer, a pair of the relevant declaration and | |||
7825 | /// index where it occurs is appended to the device pointers info array. | |||
7826 | void generateAllInfo(MapBaseValuesArrayTy &BasePointers, | |||
7827 | MapValuesArrayTy &Pointers, MapValuesArrayTy &Sizes, | |||
7828 | MapFlagsArrayTy &Types) const { | |||
7829 | // We have to process the component lists that relate with the same | |||
7830 | // declaration in a single chunk so that we can generate the map flags | |||
7831 | // correctly. Therefore, we organize all lists in a map. | |||
7832 | llvm::MapVector<const ValueDecl *, SmallVector<MapInfo, 8>> Info; | |||
7833 | ||||
7834 | // Helper function to fill the information map for the different supported | |||
7835 | // clauses. | |||
7836 | auto &&InfoGen = [&Info]( | |||
7837 | const ValueDecl *D, | |||
7838 | OMPClauseMappableExprCommon::MappableExprComponentListRef L, | |||
7839 | OpenMPMapClauseKind MapType, | |||
7840 | ArrayRef<OpenMPMapModifierKind> MapModifiers, | |||
7841 | bool ReturnDevicePointer, bool IsImplicit) { | |||
7842 | const ValueDecl *VD = | |||
7843 | D ? cast<ValueDecl>(D->getCanonicalDecl()) : nullptr; | |||
7844 | Info[VD].emplace_back(L, MapType, MapModifiers, ReturnDevicePointer, | |||
7845 | IsImplicit); | |||
7846 | }; | |||
7847 | ||||
7848 | // FIXME: MSVC 2013 seems to require this-> to find member CurDir. | |||
7849 | for (const auto *C : this->CurDir.getClausesOfKind<OMPMapClause>()) | |||
7850 | for (const auto &L : C->component_lists()) { | |||
7851 | InfoGen(L.first, L.second, C->getMapType(), C->getMapTypeModifiers(), | |||
7852 | /*ReturnDevicePointer=*/false, C->isImplicit()); | |||
7853 | } | |||
7854 | for (const auto *C : this->CurDir.getClausesOfKind<OMPToClause>()) | |||
7855 | for (const auto &L : C->component_lists()) { | |||
7856 | InfoGen(L.first, L.second, OMPC_MAP_to, llvm::None, | |||
7857 | /*ReturnDevicePointer=*/false, C->isImplicit()); | |||
7858 | } | |||
7859 | for (const auto *C : this->CurDir.getClausesOfKind<OMPFromClause>()) | |||
7860 | for (const auto &L : C->component_lists()) { | |||
7861 | InfoGen(L.first, L.second, OMPC_MAP_from, llvm::None, | |||
7862 | /*ReturnDevicePointer=*/false, C->isImplicit()); | |||
7863 | } | |||
7864 | ||||
7865 | // Look at the use_device_ptr clause information and mark the existing map | |||
7866 | // entries as such. If there is no map information for an entry in the | |||
7867 | // use_device_ptr list, we create one with map type 'alloc' and zero size | |||
7868 | // section. It is the user fault if that was not mapped before. If there is | |||
7869 | // no map information and the pointer is a struct member, then we defer the | |||
7870 | // emission of that entry until the whole struct has been processed. | |||
7871 | llvm::MapVector<const ValueDecl *, SmallVector<DeferredDevicePtrEntryTy, 4>> | |||
7872 | DeferredInfo; | |||
7873 | ||||
7874 | // FIXME: MSVC 2013 seems to require this-> to find member CurDir. | |||
7875 | for (const auto *C : | |||
7876 | this->CurDir.getClausesOfKind<OMPUseDevicePtrClause>()) { | |||
7877 | for (const auto &L : C->component_lists()) { | |||
7878 | assert(!L.second.empty() && "Not expecting empty list of components!")((!L.second.empty() && "Not expecting empty list of components!" ) ? static_cast<void> (0) : __assert_fail ("!L.second.empty() && \"Not expecting empty list of components!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7878, __PRETTY_FUNCTION__)); | |||
7879 | const ValueDecl *VD = L.second.back().getAssociatedDeclaration(); | |||
7880 | VD = cast<ValueDecl>(VD->getCanonicalDecl()); | |||
7881 | const Expr *IE = L.second.back().getAssociatedExpression(); | |||
7882 | // If the first component is a member expression, we have to look into | |||
7883 | // 'this', which maps to null in the map of map information. Otherwise | |||
7884 | // look directly for the information. | |||
7885 | auto It = Info.find(isa<MemberExpr>(IE) ? nullptr : VD); | |||
7886 | ||||
7887 | // We potentially have map information for this declaration already. | |||
7888 | // Look for the first set of components that refer to it. | |||
7889 | if (It != Info.end()) { | |||
7890 | auto CI = std::find_if( | |||
7891 | It->second.begin(), It->second.end(), [VD](const MapInfo &MI) { | |||
7892 | return MI.Components.back().getAssociatedDeclaration() == VD; | |||
7893 | }); | |||
7894 | // If we found a map entry, signal that the pointer has to be returned | |||
7895 | // and move on to the next declaration. | |||
7896 | if (CI != It->second.end()) { | |||
7897 | CI->ReturnDevicePointer = true; | |||
7898 | continue; | |||
7899 | } | |||
7900 | } | |||
7901 | ||||
7902 | // We didn't find any match in our map information - generate a zero | |||
7903 | // size array section - if the pointer is a struct member we defer this | |||
7904 | // action until the whole struct has been processed. | |||
7905 | // FIXME: MSVC 2013 seems to require this-> to find member CGF. | |||
7906 | if (isa<MemberExpr>(IE)) { | |||
7907 | // Insert the pointer into Info to be processed by | |||
7908 | // generateInfoForComponentList. Because it is a member pointer | |||
7909 | // without a pointee, no entry will be generated for it, therefore | |||
7910 | // we need to generate one after the whole struct has been processed. | |||
7911 | // Nonetheless, generateInfoForComponentList must be called to take | |||
7912 | // the pointer into account for the calculation of the range of the | |||
7913 | // partial struct. | |||
7914 | InfoGen(nullptr, L.second, OMPC_MAP_unknown, llvm::None, | |||
7915 | /*ReturnDevicePointer=*/false, C->isImplicit()); | |||
7916 | DeferredInfo[nullptr].emplace_back(IE, VD); | |||
7917 | } else { | |||
7918 | llvm::Value *Ptr = this->CGF.EmitLoadOfScalar( | |||
7919 | this->CGF.EmitLValue(IE), IE->getExprLoc()); | |||
7920 | BasePointers.emplace_back(Ptr, VD); | |||
7921 | Pointers.push_back(Ptr); | |||
7922 | Sizes.push_back(llvm::Constant::getNullValue(this->CGF.SizeTy)); | |||
7923 | Types.push_back(OMP_MAP_RETURN_PARAM | OMP_MAP_TARGET_PARAM); | |||
7924 | } | |||
7925 | } | |||
7926 | } | |||
7927 | ||||
7928 | for (const auto &M : Info) { | |||
7929 | // We need to know when we generate information for the first component | |||
7930 | // associated with a capture, because the mapping flags depend on it. | |||
7931 | bool IsFirstComponentList = true; | |||
7932 | ||||
7933 | // Temporary versions of arrays | |||
7934 | MapBaseValuesArrayTy CurBasePointers; | |||
7935 | MapValuesArrayTy CurPointers; | |||
7936 | MapValuesArrayTy CurSizes; | |||
7937 | MapFlagsArrayTy CurTypes; | |||
7938 | StructRangeInfoTy PartialStruct; | |||
7939 | ||||
7940 | for (const MapInfo &L : M.second) { | |||
7941 | assert(!L.Components.empty() &&((!L.Components.empty() && "Not expecting declaration with no component lists." ) ? static_cast<void> (0) : __assert_fail ("!L.Components.empty() && \"Not expecting declaration with no component lists.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7942, __PRETTY_FUNCTION__)) | |||
7942 | "Not expecting declaration with no component lists.")((!L.Components.empty() && "Not expecting declaration with no component lists." ) ? static_cast<void> (0) : __assert_fail ("!L.Components.empty() && \"Not expecting declaration with no component lists.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7942, __PRETTY_FUNCTION__)); | |||
7943 | ||||
7944 | // Remember the current base pointer index. | |||
7945 | unsigned CurrentBasePointersIdx = CurBasePointers.size(); | |||
7946 | // FIXME: MSVC 2013 seems to require this-> to find the member method. | |||
7947 | this->generateInfoForComponentList( | |||
7948 | L.MapType, L.MapModifiers, L.Components, CurBasePointers, | |||
7949 | CurPointers, CurSizes, CurTypes, PartialStruct, | |||
7950 | IsFirstComponentList, L.IsImplicit); | |||
7951 | ||||
7952 | // If this entry relates with a device pointer, set the relevant | |||
7953 | // declaration and add the 'return pointer' flag. | |||
7954 | if (L.ReturnDevicePointer) { | |||
7955 | assert(CurBasePointers.size() > CurrentBasePointersIdx &&((CurBasePointers.size() > CurrentBasePointersIdx && "Unexpected number of mapped base pointers.") ? static_cast< void> (0) : __assert_fail ("CurBasePointers.size() > CurrentBasePointersIdx && \"Unexpected number of mapped base pointers.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7956, __PRETTY_FUNCTION__)) | |||
7956 | "Unexpected number of mapped base pointers.")((CurBasePointers.size() > CurrentBasePointersIdx && "Unexpected number of mapped base pointers.") ? static_cast< void> (0) : __assert_fail ("CurBasePointers.size() > CurrentBasePointersIdx && \"Unexpected number of mapped base pointers.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7956, __PRETTY_FUNCTION__)); | |||
7957 | ||||
7958 | const ValueDecl *RelevantVD = | |||
7959 | L.Components.back().getAssociatedDeclaration(); | |||
7960 | assert(RelevantVD &&((RelevantVD && "No relevant declaration related with device pointer??" ) ? static_cast<void> (0) : __assert_fail ("RelevantVD && \"No relevant declaration related with device pointer??\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7961, __PRETTY_FUNCTION__)) | |||
7961 | "No relevant declaration related with device pointer??")((RelevantVD && "No relevant declaration related with device pointer??" ) ? static_cast<void> (0) : __assert_fail ("RelevantVD && \"No relevant declaration related with device pointer??\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 7961, __PRETTY_FUNCTION__)); | |||
7962 | ||||
7963 | CurBasePointers[CurrentBasePointersIdx].setDevicePtrDecl(RelevantVD); | |||
7964 | CurTypes[CurrentBasePointersIdx] |= OMP_MAP_RETURN_PARAM; | |||
7965 | } | |||
7966 | IsFirstComponentList = false; | |||
7967 | } | |||
7968 | ||||
7969 | // Append any pending zero-length pointers which are struct members and | |||
7970 | // used with use_device_ptr. | |||
7971 | auto CI = DeferredInfo.find(M.first); | |||
7972 | if (CI != DeferredInfo.end()) { | |||
7973 | for (const DeferredDevicePtrEntryTy &L : CI->second) { | |||
7974 | llvm::Value *BasePtr = this->CGF.EmitLValue(L.IE).getPointer(); | |||
7975 | llvm::Value *Ptr = this->CGF.EmitLoadOfScalar( | |||
7976 | this->CGF.EmitLValue(L.IE), L.IE->getExprLoc()); | |||
7977 | CurBasePointers.emplace_back(BasePtr, L.VD); | |||
7978 | CurPointers.push_back(Ptr); | |||
7979 | CurSizes.push_back(llvm::Constant::getNullValue(this->CGF.SizeTy)); | |||
7980 | // Entry is PTR_AND_OBJ and RETURN_PARAM. Also, set the placeholder | |||
7981 | // value MEMBER_OF=FFFF so that the entry is later updated with the | |||
7982 | // correct value of MEMBER_OF. | |||
7983 | CurTypes.push_back(OMP_MAP_PTR_AND_OBJ | OMP_MAP_RETURN_PARAM | | |||
7984 | OMP_MAP_MEMBER_OF); | |||
7985 | } | |||
7986 | } | |||
7987 | ||||
7988 | // If there is an entry in PartialStruct it means we have a struct with | |||
7989 | // individual members mapped. Emit an extra combined entry. | |||
7990 | if (PartialStruct.Base.isValid()) | |||
7991 | emitCombinedEntry(BasePointers, Pointers, Sizes, Types, CurTypes, | |||
7992 | PartialStruct); | |||
7993 | ||||
7994 | // We need to append the results of this capture to what we already have. | |||
7995 | BasePointers.append(CurBasePointers.begin(), CurBasePointers.end()); | |||
7996 | Pointers.append(CurPointers.begin(), CurPointers.end()); | |||
7997 | Sizes.append(CurSizes.begin(), CurSizes.end()); | |||
7998 | Types.append(CurTypes.begin(), CurTypes.end()); | |||
7999 | } | |||
8000 | } | |||
8001 | ||||
8002 | /// Emit capture info for lambdas for variables captured by reference. | |||
8003 | void generateInfoForLambdaCaptures( | |||
8004 | const ValueDecl *VD, llvm::Value *Arg, MapBaseValuesArrayTy &BasePointers, | |||
8005 | MapValuesArrayTy &Pointers, MapValuesArrayTy &Sizes, | |||
8006 | MapFlagsArrayTy &Types, | |||
8007 | llvm::DenseMap<llvm::Value *, llvm::Value *> &LambdaPointers) const { | |||
8008 | const auto *RD = VD->getType() | |||
8009 | .getCanonicalType() | |||
8010 | .getNonReferenceType() | |||
8011 | ->getAsCXXRecordDecl(); | |||
8012 | if (!RD || !RD->isLambda()) | |||
8013 | return; | |||
8014 | Address VDAddr = Address(Arg, CGF.getContext().getDeclAlign(VD)); | |||
8015 | LValue VDLVal = CGF.MakeAddrLValue( | |||
8016 | VDAddr, VD->getType().getCanonicalType().getNonReferenceType()); | |||
8017 | llvm::DenseMap<const VarDecl *, FieldDecl *> Captures; | |||
8018 | FieldDecl *ThisCapture = nullptr; | |||
8019 | RD->getCaptureFields(Captures, ThisCapture); | |||
8020 | if (ThisCapture) { | |||
8021 | LValue ThisLVal = | |||
8022 | CGF.EmitLValueForFieldInitialization(VDLVal, ThisCapture); | |||
8023 | LValue ThisLValVal = CGF.EmitLValueForField(VDLVal, ThisCapture); | |||
8024 | LambdaPointers.try_emplace(ThisLVal.getPointer(), VDLVal.getPointer()); | |||
8025 | BasePointers.push_back(ThisLVal.getPointer()); | |||
8026 | Pointers.push_back(ThisLValVal.getPointer()); | |||
8027 | Sizes.push_back(CGF.getTypeSize(CGF.getContext().VoidPtrTy)); | |||
8028 | Types.push_back(OMP_MAP_PTR_AND_OBJ | OMP_MAP_LITERAL | | |||
8029 | OMP_MAP_MEMBER_OF | OMP_MAP_IMPLICIT); | |||
8030 | } | |||
8031 | for (const LambdaCapture &LC : RD->captures()) { | |||
8032 | if (LC.getCaptureKind() != LCK_ByRef) | |||
8033 | continue; | |||
8034 | const VarDecl *VD = LC.getCapturedVar(); | |||
8035 | auto It = Captures.find(VD); | |||
8036 | assert(It != Captures.end() && "Found lambda capture without field.")((It != Captures.end() && "Found lambda capture without field." ) ? static_cast<void> (0) : __assert_fail ("It != Captures.end() && \"Found lambda capture without field.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8036, __PRETTY_FUNCTION__)); | |||
8037 | LValue VarLVal = CGF.EmitLValueForFieldInitialization(VDLVal, It->second); | |||
8038 | LValue VarLValVal = CGF.EmitLValueForField(VDLVal, It->second); | |||
8039 | LambdaPointers.try_emplace(VarLVal.getPointer(), VDLVal.getPointer()); | |||
8040 | BasePointers.push_back(VarLVal.getPointer()); | |||
8041 | Pointers.push_back(VarLValVal.getPointer()); | |||
8042 | Sizes.push_back(CGF.getTypeSize( | |||
8043 | VD->getType().getCanonicalType().getNonReferenceType())); | |||
8044 | Types.push_back(OMP_MAP_PTR_AND_OBJ | OMP_MAP_LITERAL | | |||
8045 | OMP_MAP_MEMBER_OF | OMP_MAP_IMPLICIT); | |||
8046 | } | |||
8047 | } | |||
8048 | ||||
8049 | /// Set correct indices for lambdas captures. | |||
8050 | void adjustMemberOfForLambdaCaptures( | |||
8051 | const llvm::DenseMap<llvm::Value *, llvm::Value *> &LambdaPointers, | |||
8052 | MapBaseValuesArrayTy &BasePointers, MapValuesArrayTy &Pointers, | |||
8053 | MapFlagsArrayTy &Types) const { | |||
8054 | for (unsigned I = 0, E = Types.size(); I < E; ++I) { | |||
8055 | // Set correct member_of idx for all implicit lambda captures. | |||
8056 | if (Types[I] != (OMP_MAP_PTR_AND_OBJ | OMP_MAP_LITERAL | | |||
8057 | OMP_MAP_MEMBER_OF | OMP_MAP_IMPLICIT)) | |||
8058 | continue; | |||
8059 | llvm::Value *BasePtr = LambdaPointers.lookup(*BasePointers[I]); | |||
8060 | assert(BasePtr && "Unable to find base lambda address.")((BasePtr && "Unable to find base lambda address.") ? static_cast<void> (0) : __assert_fail ("BasePtr && \"Unable to find base lambda address.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8060, __PRETTY_FUNCTION__)); | |||
8061 | int TgtIdx = -1; | |||
8062 | for (unsigned J = I; J > 0; --J) { | |||
8063 | unsigned Idx = J - 1; | |||
8064 | if (Pointers[Idx] != BasePtr) | |||
8065 | continue; | |||
8066 | TgtIdx = Idx; | |||
8067 | break; | |||
8068 | } | |||
8069 | assert(TgtIdx != -1 && "Unable to find parent lambda.")((TgtIdx != -1 && "Unable to find parent lambda.") ? static_cast <void> (0) : __assert_fail ("TgtIdx != -1 && \"Unable to find parent lambda.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8069, __PRETTY_FUNCTION__)); | |||
8070 | // All other current entries will be MEMBER_OF the combined entry | |||
8071 | // (except for PTR_AND_OBJ entries which do not have a placeholder value | |||
8072 | // 0xFFFF in the MEMBER_OF field). | |||
8073 | OpenMPOffloadMappingFlags MemberOfFlag = getMemberOfFlag(TgtIdx); | |||
8074 | setCorrectMemberOfFlag(Types[I], MemberOfFlag); | |||
8075 | } | |||
8076 | } | |||
8077 | ||||
8078 | /// Generate the base pointers, section pointers, sizes and map types | |||
8079 | /// associated to a given capture. | |||
8080 | void generateInfoForCapture(const CapturedStmt::Capture *Cap, | |||
8081 | llvm::Value *Arg, | |||
8082 | MapBaseValuesArrayTy &BasePointers, | |||
8083 | MapValuesArrayTy &Pointers, | |||
8084 | MapValuesArrayTy &Sizes, MapFlagsArrayTy &Types, | |||
8085 | StructRangeInfoTy &PartialStruct) const { | |||
8086 | assert(!Cap->capturesVariableArrayType() &&((!Cap->capturesVariableArrayType() && "Not expecting to generate map info for a variable array type!" ) ? static_cast<void> (0) : __assert_fail ("!Cap->capturesVariableArrayType() && \"Not expecting to generate map info for a variable array type!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8087, __PRETTY_FUNCTION__)) | |||
8087 | "Not expecting to generate map info for a variable array type!")((!Cap->capturesVariableArrayType() && "Not expecting to generate map info for a variable array type!" ) ? static_cast<void> (0) : __assert_fail ("!Cap->capturesVariableArrayType() && \"Not expecting to generate map info for a variable array type!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8087, __PRETTY_FUNCTION__)); | |||
8088 | ||||
8089 | // We need to know when we generating information for the first component | |||
8090 | const ValueDecl *VD = Cap->capturesThis() | |||
8091 | ? nullptr | |||
8092 | : Cap->getCapturedVar()->getCanonicalDecl(); | |||
8093 | ||||
8094 | // If this declaration appears in a is_device_ptr clause we just have to | |||
8095 | // pass the pointer by value. If it is a reference to a declaration, we just | |||
8096 | // pass its value. | |||
8097 | if (DevPointersMap.count(VD)) { | |||
8098 | BasePointers.emplace_back(Arg, VD); | |||
8099 | Pointers.push_back(Arg); | |||
8100 | Sizes.push_back(CGF.getTypeSize(CGF.getContext().VoidPtrTy)); | |||
8101 | Types.push_back(OMP_MAP_LITERAL | OMP_MAP_TARGET_PARAM); | |||
8102 | return; | |||
8103 | } | |||
8104 | ||||
8105 | using MapData = | |||
8106 | std::tuple<OMPClauseMappableExprCommon::MappableExprComponentListRef, | |||
8107 | OpenMPMapClauseKind, ArrayRef<OpenMPMapModifierKind>, bool>; | |||
8108 | SmallVector<MapData, 4> DeclComponentLists; | |||
8109 | // FIXME: MSVC 2013 seems to require this-> to find member CurDir. | |||
8110 | for (const auto *C : this->CurDir.getClausesOfKind<OMPMapClause>()) { | |||
8111 | for (const auto &L : C->decl_component_lists(VD)) { | |||
8112 | assert(L.first == VD &&((L.first == VD && "We got information for the wrong declaration??" ) ? static_cast<void> (0) : __assert_fail ("L.first == VD && \"We got information for the wrong declaration??\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8113, __PRETTY_FUNCTION__)) | |||
8113 | "We got information for the wrong declaration??")((L.first == VD && "We got information for the wrong declaration??" ) ? static_cast<void> (0) : __assert_fail ("L.first == VD && \"We got information for the wrong declaration??\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8113, __PRETTY_FUNCTION__)); | |||
8114 | assert(!L.second.empty() &&((!L.second.empty() && "Not expecting declaration with no component lists." ) ? static_cast<void> (0) : __assert_fail ("!L.second.empty() && \"Not expecting declaration with no component lists.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8115, __PRETTY_FUNCTION__)) | |||
8115 | "Not expecting declaration with no component lists.")((!L.second.empty() && "Not expecting declaration with no component lists." ) ? static_cast<void> (0) : __assert_fail ("!L.second.empty() && \"Not expecting declaration with no component lists.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8115, __PRETTY_FUNCTION__)); | |||
8116 | DeclComponentLists.emplace_back(L.second, C->getMapType(), | |||
8117 | C->getMapTypeModifiers(), | |||
8118 | C->isImplicit()); | |||
8119 | } | |||
8120 | } | |||
8121 | ||||
8122 | // Find overlapping elements (including the offset from the base element). | |||
8123 | llvm::SmallDenseMap< | |||
8124 | const MapData *, | |||
8125 | llvm::SmallVector< | |||
8126 | OMPClauseMappableExprCommon::MappableExprComponentListRef, 4>, | |||
8127 | 4> | |||
8128 | OverlappedData; | |||
8129 | size_t Count = 0; | |||
8130 | for (const MapData &L : DeclComponentLists) { | |||
8131 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | |||
8132 | OpenMPMapClauseKind MapType; | |||
8133 | ArrayRef<OpenMPMapModifierKind> MapModifiers; | |||
8134 | bool IsImplicit; | |||
8135 | std::tie(Components, MapType, MapModifiers, IsImplicit) = L; | |||
8136 | ++Count; | |||
8137 | for (const MapData &L1 : makeArrayRef(DeclComponentLists).slice(Count)) { | |||
8138 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components1; | |||
8139 | std::tie(Components1, MapType, MapModifiers, IsImplicit) = L1; | |||
8140 | auto CI = Components.rbegin(); | |||
8141 | auto CE = Components.rend(); | |||
8142 | auto SI = Components1.rbegin(); | |||
8143 | auto SE = Components1.rend(); | |||
8144 | for (; CI != CE && SI != SE; ++CI, ++SI) { | |||
8145 | if (CI->getAssociatedExpression()->getStmtClass() != | |||
8146 | SI->getAssociatedExpression()->getStmtClass()) | |||
8147 | break; | |||
8148 | // Are we dealing with different variables/fields? | |||
8149 | if (CI->getAssociatedDeclaration() != SI->getAssociatedDeclaration()) | |||
8150 | break; | |||
8151 | } | |||
8152 | // Found overlapping if, at least for one component, reached the head of | |||
8153 | // the components list. | |||
8154 | if (CI == CE || SI == SE) { | |||
8155 | assert((CI != CE || SI != SE) &&(((CI != CE || SI != SE) && "Unexpected full match of the mapping components." ) ? static_cast<void> (0) : __assert_fail ("(CI != CE || SI != SE) && \"Unexpected full match of the mapping components.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8156, __PRETTY_FUNCTION__)) | |||
8156 | "Unexpected full match of the mapping components.")(((CI != CE || SI != SE) && "Unexpected full match of the mapping components." ) ? static_cast<void> (0) : __assert_fail ("(CI != CE || SI != SE) && \"Unexpected full match of the mapping components.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8156, __PRETTY_FUNCTION__)); | |||
8157 | const MapData &BaseData = CI == CE ? L : L1; | |||
8158 | OMPClauseMappableExprCommon::MappableExprComponentListRef SubData = | |||
8159 | SI == SE ? Components : Components1; | |||
8160 | auto &OverlappedElements = OverlappedData.FindAndConstruct(&BaseData); | |||
8161 | OverlappedElements.getSecond().push_back(SubData); | |||
8162 | } | |||
8163 | } | |||
8164 | } | |||
8165 | // Sort the overlapped elements for each item. | |||
8166 | llvm::SmallVector<const FieldDecl *, 4> Layout; | |||
8167 | if (!OverlappedData.empty()) { | |||
8168 | if (const auto *CRD = | |||
8169 | VD->getType().getCanonicalType()->getAsCXXRecordDecl()) | |||
| ||||
8170 | getPlainLayout(CRD, Layout, /*AsBase=*/false); | |||
8171 | else { | |||
8172 | const auto *RD = VD->getType().getCanonicalType()->getAsRecordDecl(); | |||
8173 | Layout.append(RD->field_begin(), RD->field_end()); | |||
8174 | } | |||
8175 | } | |||
8176 | for (auto &Pair : OverlappedData) { | |||
8177 | llvm::sort( | |||
8178 | Pair.getSecond(), | |||
8179 | [&Layout]( | |||
8180 | OMPClauseMappableExprCommon::MappableExprComponentListRef First, | |||
8181 | OMPClauseMappableExprCommon::MappableExprComponentListRef | |||
8182 | Second) { | |||
8183 | auto CI = First.rbegin(); | |||
8184 | auto CE = First.rend(); | |||
8185 | auto SI = Second.rbegin(); | |||
8186 | auto SE = Second.rend(); | |||
8187 | for (; CI != CE && SI != SE; ++CI, ++SI) { | |||
8188 | if (CI->getAssociatedExpression()->getStmtClass() != | |||
8189 | SI->getAssociatedExpression()->getStmtClass()) | |||
8190 | break; | |||
8191 | // Are we dealing with different variables/fields? | |||
8192 | if (CI->getAssociatedDeclaration() != | |||
8193 | SI->getAssociatedDeclaration()) | |||
8194 | break; | |||
8195 | } | |||
8196 | ||||
8197 | // Lists contain the same elements. | |||
8198 | if (CI == CE && SI == SE) | |||
8199 | return false; | |||
8200 | ||||
8201 | // List with less elements is less than list with more elements. | |||
8202 | if (CI == CE || SI == SE) | |||
8203 | return CI == CE; | |||
8204 | ||||
8205 | const auto *FD1 = cast<FieldDecl>(CI->getAssociatedDeclaration()); | |||
8206 | const auto *FD2 = cast<FieldDecl>(SI->getAssociatedDeclaration()); | |||
8207 | if (FD1->getParent() == FD2->getParent()) | |||
8208 | return FD1->getFieldIndex() < FD2->getFieldIndex(); | |||
8209 | const auto It = | |||
8210 | llvm::find_if(Layout, [FD1, FD2](const FieldDecl *FD) { | |||
8211 | return FD == FD1 || FD == FD2; | |||
8212 | }); | |||
8213 | return *It == FD1; | |||
8214 | }); | |||
8215 | } | |||
8216 | ||||
8217 | // Associated with a capture, because the mapping flags depend on it. | |||
8218 | // Go through all of the elements with the overlapped elements. | |||
8219 | for (const auto &Pair : OverlappedData) { | |||
8220 | const MapData &L = *Pair.getFirst(); | |||
8221 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | |||
8222 | OpenMPMapClauseKind MapType; | |||
8223 | ArrayRef<OpenMPMapModifierKind> MapModifiers; | |||
8224 | bool IsImplicit; | |||
8225 | std::tie(Components, MapType, MapModifiers, IsImplicit) = L; | |||
8226 | ArrayRef<OMPClauseMappableExprCommon::MappableExprComponentListRef> | |||
8227 | OverlappedComponents = Pair.getSecond(); | |||
8228 | bool IsFirstComponentList = true; | |||
8229 | generateInfoForComponentList(MapType, MapModifiers, Components, | |||
8230 | BasePointers, Pointers, Sizes, Types, | |||
8231 | PartialStruct, IsFirstComponentList, | |||
8232 | IsImplicit, OverlappedComponents); | |||
8233 | } | |||
8234 | // Go through other elements without overlapped elements. | |||
8235 | bool IsFirstComponentList = OverlappedData.empty(); | |||
8236 | for (const MapData &L : DeclComponentLists) { | |||
8237 | OMPClauseMappableExprCommon::MappableExprComponentListRef Components; | |||
8238 | OpenMPMapClauseKind MapType; | |||
8239 | ArrayRef<OpenMPMapModifierKind> MapModifiers; | |||
8240 | bool IsImplicit; | |||
8241 | std::tie(Components, MapType, MapModifiers, IsImplicit) = L; | |||
8242 | auto It = OverlappedData.find(&L); | |||
8243 | if (It == OverlappedData.end()) | |||
8244 | generateInfoForComponentList(MapType, MapModifiers, Components, | |||
8245 | BasePointers, Pointers, Sizes, Types, | |||
8246 | PartialStruct, IsFirstComponentList, | |||
8247 | IsImplicit); | |||
8248 | IsFirstComponentList = false; | |||
8249 | } | |||
8250 | } | |||
8251 | ||||
8252 | /// Generate the base pointers, section pointers, sizes and map types | |||
8253 | /// associated with the declare target link variables. | |||
8254 | void generateInfoForDeclareTargetLink(MapBaseValuesArrayTy &BasePointers, | |||
8255 | MapValuesArrayTy &Pointers, | |||
8256 | MapValuesArrayTy &Sizes, | |||
8257 | MapFlagsArrayTy &Types) const { | |||
8258 | // Map other list items in the map clause which are not captured variables | |||
8259 | // but "declare target link" global variables. | |||
8260 | for (const auto *C : this->CurDir.getClausesOfKind<OMPMapClause>()) { | |||
8261 | for (const auto &L : C->component_lists()) { | |||
8262 | if (!L.first) | |||
8263 | continue; | |||
8264 | const auto *VD = dyn_cast<VarDecl>(L.first); | |||
8265 | if (!VD) | |||
8266 | continue; | |||
8267 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
8268 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | |||
8269 | if (!Res || *Res != OMPDeclareTargetDeclAttr::MT_Link) | |||
8270 | continue; | |||
8271 | StructRangeInfoTy PartialStruct; | |||
8272 | generateInfoForComponentList( | |||
8273 | C->getMapType(), C->getMapTypeModifiers(), L.second, BasePointers, | |||
8274 | Pointers, Sizes, Types, PartialStruct, | |||
8275 | /*IsFirstComponentList=*/true, C->isImplicit()); | |||
8276 | assert(!PartialStruct.Base.isValid() &&((!PartialStruct.Base.isValid() && "No partial structs for declare target link expected." ) ? static_cast<void> (0) : __assert_fail ("!PartialStruct.Base.isValid() && \"No partial structs for declare target link expected.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8277, __PRETTY_FUNCTION__)) | |||
8277 | "No partial structs for declare target link expected.")((!PartialStruct.Base.isValid() && "No partial structs for declare target link expected." ) ? static_cast<void> (0) : __assert_fail ("!PartialStruct.Base.isValid() && \"No partial structs for declare target link expected.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8277, __PRETTY_FUNCTION__)); | |||
8278 | } | |||
8279 | } | |||
8280 | } | |||
8281 | ||||
8282 | /// Generate the default map information for a given capture \a CI, | |||
8283 | /// record field declaration \a RI and captured value \a CV. | |||
8284 | void generateDefaultMapInfo(const CapturedStmt::Capture &CI, | |||
8285 | const FieldDecl &RI, llvm::Value *CV, | |||
8286 | MapBaseValuesArrayTy &CurBasePointers, | |||
8287 | MapValuesArrayTy &CurPointers, | |||
8288 | MapValuesArrayTy &CurSizes, | |||
8289 | MapFlagsArrayTy &CurMapTypes) const { | |||
8290 | // Do the default mapping. | |||
8291 | if (CI.capturesThis()) { | |||
8292 | CurBasePointers.push_back(CV); | |||
8293 | CurPointers.push_back(CV); | |||
8294 | const auto *PtrTy = cast<PointerType>(RI.getType().getTypePtr()); | |||
8295 | CurSizes.push_back(CGF.getTypeSize(PtrTy->getPointeeType())); | |||
8296 | // Default map type. | |||
8297 | CurMapTypes.push_back(OMP_MAP_TO | OMP_MAP_FROM); | |||
8298 | } else if (CI.capturesVariableByCopy()) { | |||
8299 | CurBasePointers.push_back(CV); | |||
8300 | CurPointers.push_back(CV); | |||
8301 | if (!RI.getType()->isAnyPointerType()) { | |||
8302 | // We have to signal to the runtime captures passed by value that are | |||
8303 | // not pointers. | |||
8304 | CurMapTypes.push_back(OMP_MAP_LITERAL); | |||
8305 | CurSizes.push_back(CGF.getTypeSize(RI.getType())); | |||
8306 | } else { | |||
8307 | // Pointers are implicitly mapped with a zero size and no flags | |||
8308 | // (other than first map that is added for all implicit maps). | |||
8309 | CurMapTypes.push_back(OMP_MAP_NONE); | |||
8310 | CurSizes.push_back(llvm::Constant::getNullValue(CGF.SizeTy)); | |||
8311 | } | |||
8312 | } else { | |||
8313 | assert(CI.capturesVariable() && "Expected captured reference.")((CI.capturesVariable() && "Expected captured reference." ) ? static_cast<void> (0) : __assert_fail ("CI.capturesVariable() && \"Expected captured reference.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8313, __PRETTY_FUNCTION__)); | |||
8314 | const auto *PtrTy = cast<ReferenceType>(RI.getType().getTypePtr()); | |||
8315 | QualType ElementType = PtrTy->getPointeeType(); | |||
8316 | CurSizes.push_back(CGF.getTypeSize(ElementType)); | |||
8317 | // The default map type for a scalar/complex type is 'to' because by | |||
8318 | // default the value doesn't have to be retrieved. For an aggregate | |||
8319 | // type, the default is 'tofrom'. | |||
8320 | CurMapTypes.push_back(getMapModifiersForPrivateClauses(CI)); | |||
8321 | const VarDecl *VD = CI.getCapturedVar(); | |||
8322 | if (FirstPrivateDecls.count(VD) && | |||
8323 | VD->getType().isConstant(CGF.getContext())) { | |||
8324 | llvm::Constant *Addr = | |||
8325 | CGF.CGM.getOpenMPRuntime().registerTargetFirstprivateCopy(CGF, VD); | |||
8326 | // Copy the value of the original variable to the new global copy. | |||
8327 | CGF.Builder.CreateMemCpy( | |||
8328 | CGF.MakeNaturalAlignAddrLValue(Addr, ElementType).getAddress(), | |||
8329 | Address(CV, CGF.getContext().getTypeAlignInChars(ElementType)), | |||
8330 | CurSizes.back(), /*isVolatile=*/false); | |||
8331 | // Use new global variable as the base pointers. | |||
8332 | CurBasePointers.push_back(Addr); | |||
8333 | CurPointers.push_back(Addr); | |||
8334 | } else { | |||
8335 | CurBasePointers.push_back(CV); | |||
8336 | if (FirstPrivateDecls.count(VD) && ElementType->isAnyPointerType()) { | |||
8337 | Address PtrAddr = CGF.EmitLoadOfReference(CGF.MakeAddrLValue( | |||
8338 | CV, ElementType, CGF.getContext().getDeclAlign(VD), | |||
8339 | AlignmentSource::Decl)); | |||
8340 | CurPointers.push_back(PtrAddr.getPointer()); | |||
8341 | } else { | |||
8342 | CurPointers.push_back(CV); | |||
8343 | } | |||
8344 | } | |||
8345 | } | |||
8346 | // Every default map produces a single argument which is a target parameter. | |||
8347 | CurMapTypes.back() |= OMP_MAP_TARGET_PARAM; | |||
8348 | ||||
8349 | // Add flag stating this is an implicit map. | |||
8350 | CurMapTypes.back() |= OMP_MAP_IMPLICIT; | |||
8351 | } | |||
8352 | }; | |||
8353 | ||||
8354 | enum OpenMPOffloadingReservedDeviceIDs { | |||
8355 | /// Device ID if the device was not defined, runtime should get it | |||
8356 | /// from environment variables in the spec. | |||
8357 | OMP_DEVICEID_UNDEF = -1, | |||
8358 | }; | |||
8359 | } // anonymous namespace | |||
8360 | ||||
8361 | /// Emit the arrays used to pass the captures and map information to the | |||
8362 | /// offloading runtime library. If there is no map or capture information, | |||
8363 | /// return nullptr by reference. | |||
8364 | static void | |||
8365 | emitOffloadingArrays(CodeGenFunction &CGF, | |||
8366 | MappableExprsHandler::MapBaseValuesArrayTy &BasePointers, | |||
8367 | MappableExprsHandler::MapValuesArrayTy &Pointers, | |||
8368 | MappableExprsHandler::MapValuesArrayTy &Sizes, | |||
8369 | MappableExprsHandler::MapFlagsArrayTy &MapTypes, | |||
8370 | CGOpenMPRuntime::TargetDataInfo &Info) { | |||
8371 | CodeGenModule &CGM = CGF.CGM; | |||
8372 | ASTContext &Ctx = CGF.getContext(); | |||
8373 | ||||
8374 | // Reset the array information. | |||
8375 | Info.clearArrayInfo(); | |||
8376 | Info.NumberOfPtrs = BasePointers.size(); | |||
8377 | ||||
8378 | if (Info.NumberOfPtrs) { | |||
8379 | // Detect if we have any capture size requiring runtime evaluation of the | |||
8380 | // size so that a constant array could be eventually used. | |||
8381 | bool hasRuntimeEvaluationCaptureSize = false; | |||
8382 | for (llvm::Value *S : Sizes) | |||
8383 | if (!isa<llvm::Constant>(S)) { | |||
8384 | hasRuntimeEvaluationCaptureSize = true; | |||
8385 | break; | |||
8386 | } | |||
8387 | ||||
8388 | llvm::APInt PointerNumAP(32, Info.NumberOfPtrs, /*isSigned=*/true); | |||
8389 | QualType PointerArrayType = | |||
8390 | Ctx.getConstantArrayType(Ctx.VoidPtrTy, PointerNumAP, ArrayType::Normal, | |||
8391 | /*IndexTypeQuals=*/0); | |||
8392 | ||||
8393 | Info.BasePointersArray = | |||
8394 | CGF.CreateMemTemp(PointerArrayType, ".offload_baseptrs").getPointer(); | |||
8395 | Info.PointersArray = | |||
8396 | CGF.CreateMemTemp(PointerArrayType, ".offload_ptrs").getPointer(); | |||
8397 | ||||
8398 | // If we don't have any VLA types or other types that require runtime | |||
8399 | // evaluation, we can use a constant array for the map sizes, otherwise we | |||
8400 | // need to fill up the arrays as we do for the pointers. | |||
8401 | if (hasRuntimeEvaluationCaptureSize) { | |||
8402 | QualType SizeArrayType = Ctx.getConstantArrayType( | |||
8403 | Ctx.getSizeType(), PointerNumAP, ArrayType::Normal, | |||
8404 | /*IndexTypeQuals=*/0); | |||
8405 | Info.SizesArray = | |||
8406 | CGF.CreateMemTemp(SizeArrayType, ".offload_sizes").getPointer(); | |||
8407 | } else { | |||
8408 | // We expect all the sizes to be constant, so we collect them to create | |||
8409 | // a constant array. | |||
8410 | SmallVector<llvm::Constant *, 16> ConstSizes; | |||
8411 | for (llvm::Value *S : Sizes) | |||
8412 | ConstSizes.push_back(cast<llvm::Constant>(S)); | |||
8413 | ||||
8414 | auto *SizesArrayInit = llvm::ConstantArray::get( | |||
8415 | llvm::ArrayType::get(CGM.SizeTy, ConstSizes.size()), ConstSizes); | |||
8416 | std::string Name = CGM.getOpenMPRuntime().getName({"offload_sizes"}); | |||
8417 | auto *SizesArrayGbl = new llvm::GlobalVariable( | |||
8418 | CGM.getModule(), SizesArrayInit->getType(), | |||
8419 | /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, | |||
8420 | SizesArrayInit, Name); | |||
8421 | SizesArrayGbl->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); | |||
8422 | Info.SizesArray = SizesArrayGbl; | |||
8423 | } | |||
8424 | ||||
8425 | // The map types are always constant so we don't need to generate code to | |||
8426 | // fill arrays. Instead, we create an array constant. | |||
8427 | SmallVector<uint64_t, 4> Mapping(MapTypes.size(), 0); | |||
8428 | llvm::copy(MapTypes, Mapping.begin()); | |||
8429 | llvm::Constant *MapTypesArrayInit = | |||
8430 | llvm::ConstantDataArray::get(CGF.Builder.getContext(), Mapping); | |||
8431 | std::string MaptypesName = | |||
8432 | CGM.getOpenMPRuntime().getName({"offload_maptypes"}); | |||
8433 | auto *MapTypesArrayGbl = new llvm::GlobalVariable( | |||
8434 | CGM.getModule(), MapTypesArrayInit->getType(), | |||
8435 | /*isConstant=*/true, llvm::GlobalValue::PrivateLinkage, | |||
8436 | MapTypesArrayInit, MaptypesName); | |||
8437 | MapTypesArrayGbl->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); | |||
8438 | Info.MapTypesArray = MapTypesArrayGbl; | |||
8439 | ||||
8440 | for (unsigned I = 0; I < Info.NumberOfPtrs; ++I) { | |||
8441 | llvm::Value *BPVal = *BasePointers[I]; | |||
8442 | llvm::Value *BP = CGF.Builder.CreateConstInBoundsGEP2_32( | |||
8443 | llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs), | |||
8444 | Info.BasePointersArray, 0, I); | |||
8445 | BP = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
8446 | BP, BPVal->getType()->getPointerTo(/*AddrSpace=*/0)); | |||
8447 | Address BPAddr(BP, Ctx.getTypeAlignInChars(Ctx.VoidPtrTy)); | |||
8448 | CGF.Builder.CreateStore(BPVal, BPAddr); | |||
8449 | ||||
8450 | if (Info.requiresDevicePointerInfo()) | |||
8451 | if (const ValueDecl *DevVD = BasePointers[I].getDevicePtrDecl()) | |||
8452 | Info.CaptureDeviceAddrMap.try_emplace(DevVD, BPAddr); | |||
8453 | ||||
8454 | llvm::Value *PVal = Pointers[I]; | |||
8455 | llvm::Value *P = CGF.Builder.CreateConstInBoundsGEP2_32( | |||
8456 | llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs), | |||
8457 | Info.PointersArray, 0, I); | |||
8458 | P = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
8459 | P, PVal->getType()->getPointerTo(/*AddrSpace=*/0)); | |||
8460 | Address PAddr(P, Ctx.getTypeAlignInChars(Ctx.VoidPtrTy)); | |||
8461 | CGF.Builder.CreateStore(PVal, PAddr); | |||
8462 | ||||
8463 | if (hasRuntimeEvaluationCaptureSize) { | |||
8464 | llvm::Value *S = CGF.Builder.CreateConstInBoundsGEP2_32( | |||
8465 | llvm::ArrayType::get(CGM.SizeTy, Info.NumberOfPtrs), | |||
8466 | Info.SizesArray, | |||
8467 | /*Idx0=*/0, | |||
8468 | /*Idx1=*/I); | |||
8469 | Address SAddr(S, Ctx.getTypeAlignInChars(Ctx.getSizeType())); | |||
8470 | CGF.Builder.CreateStore( | |||
8471 | CGF.Builder.CreateIntCast(Sizes[I], CGM.SizeTy, /*isSigned=*/true), | |||
8472 | SAddr); | |||
8473 | } | |||
8474 | } | |||
8475 | } | |||
8476 | } | |||
8477 | /// Emit the arguments to be passed to the runtime library based on the | |||
8478 | /// arrays of pointers, sizes and map types. | |||
8479 | static void emitOffloadingArraysArgument( | |||
8480 | CodeGenFunction &CGF, llvm::Value *&BasePointersArrayArg, | |||
8481 | llvm::Value *&PointersArrayArg, llvm::Value *&SizesArrayArg, | |||
8482 | llvm::Value *&MapTypesArrayArg, CGOpenMPRuntime::TargetDataInfo &Info) { | |||
8483 | CodeGenModule &CGM = CGF.CGM; | |||
8484 | if (Info.NumberOfPtrs) { | |||
8485 | BasePointersArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32( | |||
8486 | llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs), | |||
8487 | Info.BasePointersArray, | |||
8488 | /*Idx0=*/0, /*Idx1=*/0); | |||
8489 | PointersArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32( | |||
8490 | llvm::ArrayType::get(CGM.VoidPtrTy, Info.NumberOfPtrs), | |||
8491 | Info.PointersArray, | |||
8492 | /*Idx0=*/0, | |||
8493 | /*Idx1=*/0); | |||
8494 | SizesArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32( | |||
8495 | llvm::ArrayType::get(CGM.SizeTy, Info.NumberOfPtrs), Info.SizesArray, | |||
8496 | /*Idx0=*/0, /*Idx1=*/0); | |||
8497 | MapTypesArrayArg = CGF.Builder.CreateConstInBoundsGEP2_32( | |||
8498 | llvm::ArrayType::get(CGM.Int64Ty, Info.NumberOfPtrs), | |||
8499 | Info.MapTypesArray, | |||
8500 | /*Idx0=*/0, | |||
8501 | /*Idx1=*/0); | |||
8502 | } else { | |||
8503 | BasePointersArrayArg = llvm::ConstantPointerNull::get(CGM.VoidPtrPtrTy); | |||
8504 | PointersArrayArg = llvm::ConstantPointerNull::get(CGM.VoidPtrPtrTy); | |||
8505 | SizesArrayArg = llvm::ConstantPointerNull::get(CGM.SizeTy->getPointerTo()); | |||
8506 | MapTypesArrayArg = | |||
8507 | llvm::ConstantPointerNull::get(CGM.Int64Ty->getPointerTo()); | |||
8508 | } | |||
8509 | } | |||
8510 | ||||
8511 | /// Check for inner distribute directive. | |||
8512 | static const OMPExecutableDirective * | |||
8513 | getNestedDistributeDirective(ASTContext &Ctx, const OMPExecutableDirective &D) { | |||
8514 | const auto *CS = D.getInnermostCapturedStmt(); | |||
8515 | const auto *Body = | |||
8516 | CS->getCapturedStmt()->IgnoreContainers(/*IgnoreCaptured=*/true); | |||
8517 | const Stmt *ChildStmt = | |||
8518 | CGOpenMPSIMDRuntime::getSingleCompoundChild(Ctx, Body); | |||
8519 | ||||
8520 | if (const auto *NestedDir = | |||
8521 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
8522 | OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind(); | |||
8523 | switch (D.getDirectiveKind()) { | |||
8524 | case OMPD_target: | |||
8525 | if (isOpenMPDistributeDirective(DKind)) | |||
8526 | return NestedDir; | |||
8527 | if (DKind == OMPD_teams) { | |||
8528 | Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers( | |||
8529 | /*IgnoreCaptured=*/true); | |||
8530 | if (!Body) | |||
8531 | return nullptr; | |||
8532 | ChildStmt = CGOpenMPSIMDRuntime::getSingleCompoundChild(Ctx, Body); | |||
8533 | if (const auto *NND = | |||
8534 | dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) { | |||
8535 | DKind = NND->getDirectiveKind(); | |||
8536 | if (isOpenMPDistributeDirective(DKind)) | |||
8537 | return NND; | |||
8538 | } | |||
8539 | } | |||
8540 | return nullptr; | |||
8541 | case OMPD_target_teams: | |||
8542 | if (isOpenMPDistributeDirective(DKind)) | |||
8543 | return NestedDir; | |||
8544 | return nullptr; | |||
8545 | case OMPD_target_parallel: | |||
8546 | case OMPD_target_simd: | |||
8547 | case OMPD_target_parallel_for: | |||
8548 | case OMPD_target_parallel_for_simd: | |||
8549 | return nullptr; | |||
8550 | case OMPD_target_teams_distribute: | |||
8551 | case OMPD_target_teams_distribute_simd: | |||
8552 | case OMPD_target_teams_distribute_parallel_for: | |||
8553 | case OMPD_target_teams_distribute_parallel_for_simd: | |||
8554 | case OMPD_parallel: | |||
8555 | case OMPD_for: | |||
8556 | case OMPD_parallel_for: | |||
8557 | case OMPD_parallel_sections: | |||
8558 | case OMPD_for_simd: | |||
8559 | case OMPD_parallel_for_simd: | |||
8560 | case OMPD_cancel: | |||
8561 | case OMPD_cancellation_point: | |||
8562 | case OMPD_ordered: | |||
8563 | case OMPD_threadprivate: | |||
8564 | case OMPD_allocate: | |||
8565 | case OMPD_task: | |||
8566 | case OMPD_simd: | |||
8567 | case OMPD_sections: | |||
8568 | case OMPD_section: | |||
8569 | case OMPD_single: | |||
8570 | case OMPD_master: | |||
8571 | case OMPD_critical: | |||
8572 | case OMPD_taskyield: | |||
8573 | case OMPD_barrier: | |||
8574 | case OMPD_taskwait: | |||
8575 | case OMPD_taskgroup: | |||
8576 | case OMPD_atomic: | |||
8577 | case OMPD_flush: | |||
8578 | case OMPD_teams: | |||
8579 | case OMPD_target_data: | |||
8580 | case OMPD_target_exit_data: | |||
8581 | case OMPD_target_enter_data: | |||
8582 | case OMPD_distribute: | |||
8583 | case OMPD_distribute_simd: | |||
8584 | case OMPD_distribute_parallel_for: | |||
8585 | case OMPD_distribute_parallel_for_simd: | |||
8586 | case OMPD_teams_distribute: | |||
8587 | case OMPD_teams_distribute_simd: | |||
8588 | case OMPD_teams_distribute_parallel_for: | |||
8589 | case OMPD_teams_distribute_parallel_for_simd: | |||
8590 | case OMPD_target_update: | |||
8591 | case OMPD_declare_simd: | |||
8592 | case OMPD_declare_target: | |||
8593 | case OMPD_end_declare_target: | |||
8594 | case OMPD_declare_reduction: | |||
8595 | case OMPD_declare_mapper: | |||
8596 | case OMPD_taskloop: | |||
8597 | case OMPD_taskloop_simd: | |||
8598 | case OMPD_requires: | |||
8599 | case OMPD_unknown: | |||
8600 | llvm_unreachable("Unexpected directive.")::llvm::llvm_unreachable_internal("Unexpected directive.", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8600); | |||
8601 | } | |||
8602 | } | |||
8603 | ||||
8604 | return nullptr; | |||
8605 | } | |||
8606 | ||||
8607 | void CGOpenMPRuntime::emitTargetNumIterationsCall( | |||
8608 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *Device, | |||
8609 | const llvm::function_ref<llvm::Value *( | |||
8610 | CodeGenFunction &CGF, const OMPLoopDirective &D)> &SizeEmitter) { | |||
8611 | OpenMPDirectiveKind Kind = D.getDirectiveKind(); | |||
8612 | const OMPExecutableDirective *TD = &D; | |||
8613 | // Get nested teams distribute kind directive, if any. | |||
8614 | if (!isOpenMPDistributeDirective(Kind) || !isOpenMPTeamsDirective(Kind)) | |||
8615 | TD = getNestedDistributeDirective(CGM.getContext(), D); | |||
8616 | if (!TD) | |||
8617 | return; | |||
8618 | const auto *LD = cast<OMPLoopDirective>(TD); | |||
8619 | auto &&CodeGen = [LD, &Device, &SizeEmitter, this](CodeGenFunction &CGF, | |||
8620 | PrePostActionTy &) { | |||
8621 | llvm::Value *NumIterations = SizeEmitter(CGF, *LD); | |||
8622 | ||||
8623 | // Emit device ID if any. | |||
8624 | llvm::Value *DeviceID; | |||
8625 | if (Device) | |||
8626 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | |||
8627 | CGF.Int64Ty, /*isSigned=*/true); | |||
8628 | else | |||
8629 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | |||
8630 | ||||
8631 | llvm::Value *Args[] = {DeviceID, NumIterations}; | |||
8632 | CGF.EmitRuntimeCall( | |||
8633 | createRuntimeFunction(OMPRTL__kmpc_push_target_tripcount), Args); | |||
8634 | }; | |||
8635 | emitInlinedDirective(CGF, OMPD_unknown, CodeGen); | |||
8636 | } | |||
8637 | ||||
8638 | void CGOpenMPRuntime::emitTargetCall(CodeGenFunction &CGF, | |||
8639 | const OMPExecutableDirective &D, | |||
8640 | llvm::Function *OutlinedFn, | |||
8641 | llvm::Value *OutlinedFnID, | |||
8642 | const Expr *IfCond, const Expr *Device) { | |||
8643 | if (!CGF.HaveInsertPoint()) | |||
| ||||
8644 | return; | |||
8645 | ||||
8646 | assert(OutlinedFn && "Invalid outlined function!")((OutlinedFn && "Invalid outlined function!") ? static_cast <void> (0) : __assert_fail ("OutlinedFn && \"Invalid outlined function!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8646, __PRETTY_FUNCTION__)); | |||
8647 | ||||
8648 | const bool RequiresOuterTask = D.hasClausesOfKind<OMPDependClause>(); | |||
8649 | llvm::SmallVector<llvm::Value *, 16> CapturedVars; | |||
8650 | const CapturedStmt &CS = *D.getCapturedStmt(OMPD_target); | |||
8651 | auto &&ArgsCodegen = [&CS, &CapturedVars](CodeGenFunction &CGF, | |||
8652 | PrePostActionTy &) { | |||
8653 | CGF.GenerateOpenMPCapturedVars(CS, CapturedVars); | |||
8654 | }; | |||
8655 | emitInlinedDirective(CGF, OMPD_unknown, ArgsCodegen); | |||
8656 | ||||
8657 | CodeGenFunction::OMPTargetDataInfo InputInfo; | |||
8658 | llvm::Value *MapTypesArray = nullptr; | |||
8659 | // Fill up the pointer arrays and transfer execution to the device. | |||
8660 | auto &&ThenGen = [this, Device, OutlinedFn, OutlinedFnID, &D, &InputInfo, | |||
8661 | &MapTypesArray, &CS, RequiresOuterTask, | |||
8662 | &CapturedVars](CodeGenFunction &CGF, PrePostActionTy &) { | |||
8663 | // On top of the arrays that were filled up, the target offloading call | |||
8664 | // takes as arguments the device id as well as the host pointer. The host | |||
8665 | // pointer is used by the runtime library to identify the current target | |||
8666 | // region, so it only has to be unique and not necessarily point to | |||
8667 | // anything. It could be the pointer to the outlined function that | |||
8668 | // implements the target region, but we aren't using that so that the | |||
8669 | // compiler doesn't need to keep that, and could therefore inline the host | |||
8670 | // function if proven worthwhile during optimization. | |||
8671 | ||||
8672 | // From this point on, we need to have an ID of the target region defined. | |||
8673 | assert(OutlinedFnID && "Invalid outlined function ID!")((OutlinedFnID && "Invalid outlined function ID!") ? static_cast <void> (0) : __assert_fail ("OutlinedFnID && \"Invalid outlined function ID!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8673, __PRETTY_FUNCTION__)); | |||
8674 | ||||
8675 | // Emit device ID if any. | |||
8676 | llvm::Value *DeviceID; | |||
8677 | if (Device) { | |||
8678 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | |||
8679 | CGF.Int64Ty, /*isSigned=*/true); | |||
8680 | } else { | |||
8681 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | |||
8682 | } | |||
8683 | ||||
8684 | // Emit the number of elements in the offloading arrays. | |||
8685 | llvm::Value *PointerNum = | |||
8686 | CGF.Builder.getInt32(InputInfo.NumberOfTargetItems); | |||
8687 | ||||
8688 | // Return value of the runtime offloading call. | |||
8689 | llvm::Value *Return; | |||
8690 | ||||
8691 | llvm::Value *NumTeams = emitNumTeamsForTargetDirective(CGF, D); | |||
8692 | llvm::Value *NumThreads = emitNumThreadsForTargetDirective(CGF, D); | |||
8693 | ||||
8694 | bool HasNowait = D.hasClausesOfKind<OMPNowaitClause>(); | |||
8695 | // The target region is an outlined function launched by the runtime | |||
8696 | // via calls __tgt_target() or __tgt_target_teams(). | |||
8697 | // | |||
8698 | // __tgt_target() launches a target region with one team and one thread, | |||
8699 | // executing a serial region. This master thread may in turn launch | |||
8700 | // more threads within its team upon encountering a parallel region, | |||
8701 | // however, no additional teams can be launched on the device. | |||
8702 | // | |||
8703 | // __tgt_target_teams() launches a target region with one or more teams, | |||
8704 | // each with one or more threads. This call is required for target | |||
8705 | // constructs such as: | |||
8706 | // 'target teams' | |||
8707 | // 'target' / 'teams' | |||
8708 | // 'target teams distribute parallel for' | |||
8709 | // 'target parallel' | |||
8710 | // and so on. | |||
8711 | // | |||
8712 | // Note that on the host and CPU targets, the runtime implementation of | |||
8713 | // these calls simply call the outlined function without forking threads. | |||
8714 | // The outlined functions themselves have runtime calls to | |||
8715 | // __kmpc_fork_teams() and __kmpc_fork() for this purpose, codegen'd by | |||
8716 | // the compiler in emitTeamsCall() and emitParallelCall(). | |||
8717 | // | |||
8718 | // In contrast, on the NVPTX target, the implementation of | |||
8719 | // __tgt_target_teams() launches a GPU kernel with the requested number | |||
8720 | // of teams and threads so no additional calls to the runtime are required. | |||
8721 | if (NumTeams) { | |||
8722 | // If we have NumTeams defined this means that we have an enclosed teams | |||
8723 | // region. Therefore we also expect to have NumThreads defined. These two | |||
8724 | // values should be defined in the presence of a teams directive, | |||
8725 | // regardless of having any clauses associated. If the user is using teams | |||
8726 | // but no clauses, these two values will be the default that should be | |||
8727 | // passed to the runtime library - a 32-bit integer with the value zero. | |||
8728 | assert(NumThreads && "Thread limit expression should be available along "((NumThreads && "Thread limit expression should be available along " "with number of teams.") ? static_cast<void> (0) : __assert_fail ("NumThreads && \"Thread limit expression should be available along \" \"with number of teams.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8729, __PRETTY_FUNCTION__)) | |||
8729 | "with number of teams.")((NumThreads && "Thread limit expression should be available along " "with number of teams.") ? static_cast<void> (0) : __assert_fail ("NumThreads && \"Thread limit expression should be available along \" \"with number of teams.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8729, __PRETTY_FUNCTION__)); | |||
8730 | llvm::Value *OffloadingArgs[] = {DeviceID, | |||
8731 | OutlinedFnID, | |||
8732 | PointerNum, | |||
8733 | InputInfo.BasePointersArray.getPointer(), | |||
8734 | InputInfo.PointersArray.getPointer(), | |||
8735 | InputInfo.SizesArray.getPointer(), | |||
8736 | MapTypesArray, | |||
8737 | NumTeams, | |||
8738 | NumThreads}; | |||
8739 | Return = CGF.EmitRuntimeCall( | |||
8740 | createRuntimeFunction(HasNowait ? OMPRTL__tgt_target_teams_nowait | |||
8741 | : OMPRTL__tgt_target_teams), | |||
8742 | OffloadingArgs); | |||
8743 | } else { | |||
8744 | llvm::Value *OffloadingArgs[] = {DeviceID, | |||
8745 | OutlinedFnID, | |||
8746 | PointerNum, | |||
8747 | InputInfo.BasePointersArray.getPointer(), | |||
8748 | InputInfo.PointersArray.getPointer(), | |||
8749 | InputInfo.SizesArray.getPointer(), | |||
8750 | MapTypesArray}; | |||
8751 | Return = CGF.EmitRuntimeCall( | |||
8752 | createRuntimeFunction(HasNowait ? OMPRTL__tgt_target_nowait | |||
8753 | : OMPRTL__tgt_target), | |||
8754 | OffloadingArgs); | |||
8755 | } | |||
8756 | ||||
8757 | // Check the error code and execute the host version if required. | |||
8758 | llvm::BasicBlock *OffloadFailedBlock = | |||
8759 | CGF.createBasicBlock("omp_offload.failed"); | |||
8760 | llvm::BasicBlock *OffloadContBlock = | |||
8761 | CGF.createBasicBlock("omp_offload.cont"); | |||
8762 | llvm::Value *Failed = CGF.Builder.CreateIsNotNull(Return); | |||
8763 | CGF.Builder.CreateCondBr(Failed, OffloadFailedBlock, OffloadContBlock); | |||
8764 | ||||
8765 | CGF.EmitBlock(OffloadFailedBlock); | |||
8766 | if (RequiresOuterTask) { | |||
8767 | CapturedVars.clear(); | |||
8768 | CGF.GenerateOpenMPCapturedVars(CS, CapturedVars); | |||
8769 | } | |||
8770 | emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedFn, CapturedVars); | |||
8771 | CGF.EmitBranch(OffloadContBlock); | |||
8772 | ||||
8773 | CGF.EmitBlock(OffloadContBlock, /*IsFinished=*/true); | |||
8774 | }; | |||
8775 | ||||
8776 | // Notify that the host version must be executed. | |||
8777 | auto &&ElseGen = [this, &D, OutlinedFn, &CS, &CapturedVars, | |||
8778 | RequiresOuterTask](CodeGenFunction &CGF, | |||
8779 | PrePostActionTy &) { | |||
8780 | if (RequiresOuterTask) { | |||
8781 | CapturedVars.clear(); | |||
8782 | CGF.GenerateOpenMPCapturedVars(CS, CapturedVars); | |||
8783 | } | |||
8784 | emitOutlinedFunctionCall(CGF, D.getBeginLoc(), OutlinedFn, CapturedVars); | |||
8785 | }; | |||
8786 | ||||
8787 | auto &&TargetThenGen = [this, &ThenGen, &D, &InputInfo, &MapTypesArray, | |||
8788 | &CapturedVars, RequiresOuterTask, | |||
8789 | &CS](CodeGenFunction &CGF, PrePostActionTy &) { | |||
8790 | // Fill up the arrays with all the captured variables. | |||
8791 | MappableExprsHandler::MapBaseValuesArrayTy BasePointers; | |||
8792 | MappableExprsHandler::MapValuesArrayTy Pointers; | |||
8793 | MappableExprsHandler::MapValuesArrayTy Sizes; | |||
8794 | MappableExprsHandler::MapFlagsArrayTy MapTypes; | |||
8795 | ||||
8796 | // Get mappable expression information. | |||
8797 | MappableExprsHandler MEHandler(D, CGF); | |||
8798 | llvm::DenseMap<llvm::Value *, llvm::Value *> LambdaPointers; | |||
8799 | ||||
8800 | auto RI = CS.getCapturedRecordDecl()->field_begin(); | |||
8801 | auto CV = CapturedVars.begin(); | |||
8802 | for (CapturedStmt::const_capture_iterator CI = CS.capture_begin(), | |||
8803 | CE = CS.capture_end(); | |||
8804 | CI != CE; ++CI, ++RI, ++CV) { | |||
8805 | MappableExprsHandler::MapBaseValuesArrayTy CurBasePointers; | |||
8806 | MappableExprsHandler::MapValuesArrayTy CurPointers; | |||
8807 | MappableExprsHandler::MapValuesArrayTy CurSizes; | |||
8808 | MappableExprsHandler::MapFlagsArrayTy CurMapTypes; | |||
8809 | MappableExprsHandler::StructRangeInfoTy PartialStruct; | |||
8810 | ||||
8811 | // VLA sizes are passed to the outlined region by copy and do not have map | |||
8812 | // information associated. | |||
8813 | if (CI->capturesVariableArrayType()) { | |||
8814 | CurBasePointers.push_back(*CV); | |||
8815 | CurPointers.push_back(*CV); | |||
8816 | CurSizes.push_back(CGF.getTypeSize(RI->getType())); | |||
8817 | // Copy to the device as an argument. No need to retrieve it. | |||
8818 | CurMapTypes.push_back(MappableExprsHandler::OMP_MAP_LITERAL | | |||
8819 | MappableExprsHandler::OMP_MAP_TARGET_PARAM); | |||
8820 | } else { | |||
8821 | // If we have any information in the map clause, we use it, otherwise we | |||
8822 | // just do a default mapping. | |||
8823 | MEHandler.generateInfoForCapture(CI, *CV, CurBasePointers, CurPointers, | |||
8824 | CurSizes, CurMapTypes, PartialStruct); | |||
8825 | if (CurBasePointers.empty()) | |||
8826 | MEHandler.generateDefaultMapInfo(*CI, **RI, *CV, CurBasePointers, | |||
8827 | CurPointers, CurSizes, CurMapTypes); | |||
8828 | // Generate correct mapping for variables captured by reference in | |||
8829 | // lambdas. | |||
8830 | if (CI->capturesVariable()) | |||
8831 | MEHandler.generateInfoForLambdaCaptures( | |||
8832 | CI->getCapturedVar(), *CV, CurBasePointers, CurPointers, CurSizes, | |||
8833 | CurMapTypes, LambdaPointers); | |||
8834 | } | |||
8835 | // We expect to have at least an element of information for this capture. | |||
8836 | assert(!CurBasePointers.empty() &&((!CurBasePointers.empty() && "Non-existing map pointer for capture!" ) ? static_cast<void> (0) : __assert_fail ("!CurBasePointers.empty() && \"Non-existing map pointer for capture!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8837, __PRETTY_FUNCTION__)) | |||
8837 | "Non-existing map pointer for capture!")((!CurBasePointers.empty() && "Non-existing map pointer for capture!" ) ? static_cast<void> (0) : __assert_fail ("!CurBasePointers.empty() && \"Non-existing map pointer for capture!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8837, __PRETTY_FUNCTION__)); | |||
8838 | assert(CurBasePointers.size() == CurPointers.size() &&((CurBasePointers.size() == CurPointers.size() && CurBasePointers .size() == CurSizes.size() && CurBasePointers.size() == CurMapTypes.size() && "Inconsistent map information sizes!" ) ? static_cast<void> (0) : __assert_fail ("CurBasePointers.size() == CurPointers.size() && CurBasePointers.size() == CurSizes.size() && CurBasePointers.size() == CurMapTypes.size() && \"Inconsistent map information sizes!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8841, __PRETTY_FUNCTION__)) | |||
8839 | CurBasePointers.size() == CurSizes.size() &&((CurBasePointers.size() == CurPointers.size() && CurBasePointers .size() == CurSizes.size() && CurBasePointers.size() == CurMapTypes.size() && "Inconsistent map information sizes!" ) ? static_cast<void> (0) : __assert_fail ("CurBasePointers.size() == CurPointers.size() && CurBasePointers.size() == CurSizes.size() && CurBasePointers.size() == CurMapTypes.size() && \"Inconsistent map information sizes!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8841, __PRETTY_FUNCTION__)) | |||
8840 | CurBasePointers.size() == CurMapTypes.size() &&((CurBasePointers.size() == CurPointers.size() && CurBasePointers .size() == CurSizes.size() && CurBasePointers.size() == CurMapTypes.size() && "Inconsistent map information sizes!" ) ? static_cast<void> (0) : __assert_fail ("CurBasePointers.size() == CurPointers.size() && CurBasePointers.size() == CurSizes.size() && CurBasePointers.size() == CurMapTypes.size() && \"Inconsistent map information sizes!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8841, __PRETTY_FUNCTION__)) | |||
8841 | "Inconsistent map information sizes!")((CurBasePointers.size() == CurPointers.size() && CurBasePointers .size() == CurSizes.size() && CurBasePointers.size() == CurMapTypes.size() && "Inconsistent map information sizes!" ) ? static_cast<void> (0) : __assert_fail ("CurBasePointers.size() == CurPointers.size() && CurBasePointers.size() == CurSizes.size() && CurBasePointers.size() == CurMapTypes.size() && \"Inconsistent map information sizes!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 8841, __PRETTY_FUNCTION__)); | |||
8842 | ||||
8843 | // If there is an entry in PartialStruct it means we have a struct with | |||
8844 | // individual members mapped. Emit an extra combined entry. | |||
8845 | if (PartialStruct.Base.isValid()) | |||
8846 | MEHandler.emitCombinedEntry(BasePointers, Pointers, Sizes, MapTypes, | |||
8847 | CurMapTypes, PartialStruct); | |||
8848 | ||||
8849 | // We need to append the results of this capture to what we already have. | |||
8850 | BasePointers.append(CurBasePointers.begin(), CurBasePointers.end()); | |||
8851 | Pointers.append(CurPointers.begin(), CurPointers.end()); | |||
8852 | Sizes.append(CurSizes.begin(), CurSizes.end()); | |||
8853 | MapTypes.append(CurMapTypes.begin(), CurMapTypes.end()); | |||
8854 | } | |||
8855 | // Adjust MEMBER_OF flags for the lambdas captures. | |||
8856 | MEHandler.adjustMemberOfForLambdaCaptures(LambdaPointers, BasePointers, | |||
8857 | Pointers, MapTypes); | |||
8858 | // Map other list items in the map clause which are not captured variables | |||
8859 | // but "declare target link" global variables. | |||
8860 | MEHandler.generateInfoForDeclareTargetLink(BasePointers, Pointers, Sizes, | |||
8861 | MapTypes); | |||
8862 | ||||
8863 | TargetDataInfo Info; | |||
8864 | // Fill up the arrays and create the arguments. | |||
8865 | emitOffloadingArrays(CGF, BasePointers, Pointers, Sizes, MapTypes, Info); | |||
8866 | emitOffloadingArraysArgument(CGF, Info.BasePointersArray, | |||
8867 | Info.PointersArray, Info.SizesArray, | |||
8868 | Info.MapTypesArray, Info); | |||
8869 | InputInfo.NumberOfTargetItems = Info.NumberOfPtrs; | |||
8870 | InputInfo.BasePointersArray = | |||
8871 | Address(Info.BasePointersArray, CGM.getPointerAlign()); | |||
8872 | InputInfo.PointersArray = | |||
8873 | Address(Info.PointersArray, CGM.getPointerAlign()); | |||
8874 | InputInfo.SizesArray = Address(Info.SizesArray, CGM.getPointerAlign()); | |||
8875 | MapTypesArray = Info.MapTypesArray; | |||
8876 | if (RequiresOuterTask) | |||
8877 | CGF.EmitOMPTargetTaskBasedDirective(D, ThenGen, InputInfo); | |||
8878 | else | |||
8879 | emitInlinedDirective(CGF, D.getDirectiveKind(), ThenGen); | |||
8880 | }; | |||
8881 | ||||
8882 | auto &&TargetElseGen = [this, &ElseGen, &D, RequiresOuterTask]( | |||
8883 | CodeGenFunction &CGF, PrePostActionTy &) { | |||
8884 | if (RequiresOuterTask) { | |||
8885 | CodeGenFunction::OMPTargetDataInfo InputInfo; | |||
8886 | CGF.EmitOMPTargetTaskBasedDirective(D, ElseGen, InputInfo); | |||
8887 | } else { | |||
8888 | emitInlinedDirective(CGF, D.getDirectiveKind(), ElseGen); | |||
8889 | } | |||
8890 | }; | |||
8891 | ||||
8892 | // If we have a target function ID it means that we need to support | |||
8893 | // offloading, otherwise, just execute on the host. We need to execute on host | |||
8894 | // regardless of the conditional in the if clause if, e.g., the user do not | |||
8895 | // specify target triples. | |||
8896 | if (OutlinedFnID) { | |||
8897 | if (IfCond) { | |||
8898 | emitOMPIfClause(CGF, IfCond, TargetThenGen, TargetElseGen); | |||
8899 | } else { | |||
8900 | RegionCodeGenTy ThenRCG(TargetThenGen); | |||
8901 | ThenRCG(CGF); | |||
8902 | } | |||
8903 | } else { | |||
8904 | RegionCodeGenTy ElseRCG(TargetElseGen); | |||
8905 | ElseRCG(CGF); | |||
8906 | } | |||
8907 | } | |||
8908 | ||||
8909 | void CGOpenMPRuntime::scanForTargetRegionsFunctions(const Stmt *S, | |||
8910 | StringRef ParentName) { | |||
8911 | if (!S) | |||
8912 | return; | |||
8913 | ||||
8914 | // Codegen OMP target directives that offload compute to the device. | |||
8915 | bool RequiresDeviceCodegen = | |||
8916 | isa<OMPExecutableDirective>(S) && | |||
8917 | isOpenMPTargetExecutionDirective( | |||
8918 | cast<OMPExecutableDirective>(S)->getDirectiveKind()); | |||
8919 | ||||
8920 | if (RequiresDeviceCodegen) { | |||
8921 | const auto &E = *cast<OMPExecutableDirective>(S); | |||
8922 | unsigned DeviceID; | |||
8923 | unsigned FileID; | |||
8924 | unsigned Line; | |||
8925 | getTargetEntryUniqueInfo(CGM.getContext(), E.getBeginLoc(), DeviceID, | |||
8926 | FileID, Line); | |||
8927 | ||||
8928 | // Is this a target region that should not be emitted as an entry point? If | |||
8929 | // so just signal we are done with this target region. | |||
8930 | if (!OffloadEntriesInfoManager.hasTargetRegionEntryInfo(DeviceID, FileID, | |||
8931 | ParentName, Line)) | |||
8932 | return; | |||
8933 | ||||
8934 | switch (E.getDirectiveKind()) { | |||
8935 | case OMPD_target: | |||
8936 | CodeGenFunction::EmitOMPTargetDeviceFunction(CGM, ParentName, | |||
8937 | cast<OMPTargetDirective>(E)); | |||
8938 | break; | |||
8939 | case OMPD_target_parallel: | |||
8940 | CodeGenFunction::EmitOMPTargetParallelDeviceFunction( | |||
8941 | CGM, ParentName, cast<OMPTargetParallelDirective>(E)); | |||
8942 | break; | |||
8943 | case OMPD_target_teams: | |||
8944 | CodeGenFunction::EmitOMPTargetTeamsDeviceFunction( | |||
8945 | CGM, ParentName, cast<OMPTargetTeamsDirective>(E)); | |||
8946 | break; | |||
8947 | case OMPD_target_teams_distribute: | |||
8948 | CodeGenFunction::EmitOMPTargetTeamsDistributeDeviceFunction( | |||
8949 | CGM, ParentName, cast<OMPTargetTeamsDistributeDirective>(E)); | |||
8950 | break; | |||
8951 | case OMPD_target_teams_distribute_simd: | |||
8952 | CodeGenFunction::EmitOMPTargetTeamsDistributeSimdDeviceFunction( | |||
8953 | CGM, ParentName, cast<OMPTargetTeamsDistributeSimdDirective>(E)); | |||
8954 | break; | |||
8955 | case OMPD_target_parallel_for: | |||
8956 | CodeGenFunction::EmitOMPTargetParallelForDeviceFunction( | |||
8957 | CGM, ParentName, cast<OMPTargetParallelForDirective>(E)); | |||
8958 | break; | |||
8959 | case OMPD_target_parallel_for_simd: | |||
8960 | CodeGenFunction::EmitOMPTargetParallelForSimdDeviceFunction( | |||
8961 | CGM, ParentName, cast<OMPTargetParallelForSimdDirective>(E)); | |||
8962 | break; | |||
8963 | case OMPD_target_simd: | |||
8964 | CodeGenFunction::EmitOMPTargetSimdDeviceFunction( | |||
8965 | CGM, ParentName, cast<OMPTargetSimdDirective>(E)); | |||
8966 | break; | |||
8967 | case OMPD_target_teams_distribute_parallel_for: | |||
8968 | CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForDeviceFunction( | |||
8969 | CGM, ParentName, | |||
8970 | cast<OMPTargetTeamsDistributeParallelForDirective>(E)); | |||
8971 | break; | |||
8972 | case OMPD_target_teams_distribute_parallel_for_simd: | |||
8973 | CodeGenFunction:: | |||
8974 | EmitOMPTargetTeamsDistributeParallelForSimdDeviceFunction( | |||
8975 | CGM, ParentName, | |||
8976 | cast<OMPTargetTeamsDistributeParallelForSimdDirective>(E)); | |||
8977 | break; | |||
8978 | case OMPD_parallel: | |||
8979 | case OMPD_for: | |||
8980 | case OMPD_parallel_for: | |||
8981 | case OMPD_parallel_sections: | |||
8982 | case OMPD_for_simd: | |||
8983 | case OMPD_parallel_for_simd: | |||
8984 | case OMPD_cancel: | |||
8985 | case OMPD_cancellation_point: | |||
8986 | case OMPD_ordered: | |||
8987 | case OMPD_threadprivate: | |||
8988 | case OMPD_allocate: | |||
8989 | case OMPD_task: | |||
8990 | case OMPD_simd: | |||
8991 | case OMPD_sections: | |||
8992 | case OMPD_section: | |||
8993 | case OMPD_single: | |||
8994 | case OMPD_master: | |||
8995 | case OMPD_critical: | |||
8996 | case OMPD_taskyield: | |||
8997 | case OMPD_barrier: | |||
8998 | case OMPD_taskwait: | |||
8999 | case OMPD_taskgroup: | |||
9000 | case OMPD_atomic: | |||
9001 | case OMPD_flush: | |||
9002 | case OMPD_teams: | |||
9003 | case OMPD_target_data: | |||
9004 | case OMPD_target_exit_data: | |||
9005 | case OMPD_target_enter_data: | |||
9006 | case OMPD_distribute: | |||
9007 | case OMPD_distribute_simd: | |||
9008 | case OMPD_distribute_parallel_for: | |||
9009 | case OMPD_distribute_parallel_for_simd: | |||
9010 | case OMPD_teams_distribute: | |||
9011 | case OMPD_teams_distribute_simd: | |||
9012 | case OMPD_teams_distribute_parallel_for: | |||
9013 | case OMPD_teams_distribute_parallel_for_simd: | |||
9014 | case OMPD_target_update: | |||
9015 | case OMPD_declare_simd: | |||
9016 | case OMPD_declare_target: | |||
9017 | case OMPD_end_declare_target: | |||
9018 | case OMPD_declare_reduction: | |||
9019 | case OMPD_declare_mapper: | |||
9020 | case OMPD_taskloop: | |||
9021 | case OMPD_taskloop_simd: | |||
9022 | case OMPD_requires: | |||
9023 | case OMPD_unknown: | |||
9024 | llvm_unreachable("Unknown target directive for OpenMP device codegen.")::llvm::llvm_unreachable_internal("Unknown target directive for OpenMP device codegen." , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9024); | |||
9025 | } | |||
9026 | return; | |||
9027 | } | |||
9028 | ||||
9029 | if (const auto *E = dyn_cast<OMPExecutableDirective>(S)) { | |||
9030 | if (!E->hasAssociatedStmt() || !E->getAssociatedStmt()) | |||
9031 | return; | |||
9032 | ||||
9033 | scanForTargetRegionsFunctions( | |||
9034 | E->getInnermostCapturedStmt()->getCapturedStmt(), ParentName); | |||
9035 | return; | |||
9036 | } | |||
9037 | ||||
9038 | // If this is a lambda function, look into its body. | |||
9039 | if (const auto *L = dyn_cast<LambdaExpr>(S)) | |||
9040 | S = L->getBody(); | |||
9041 | ||||
9042 | // Keep looking for target regions recursively. | |||
9043 | for (const Stmt *II : S->children()) | |||
9044 | scanForTargetRegionsFunctions(II, ParentName); | |||
9045 | } | |||
9046 | ||||
9047 | bool CGOpenMPRuntime::emitTargetFunctions(GlobalDecl GD) { | |||
9048 | // If emitting code for the host, we do not process FD here. Instead we do | |||
9049 | // the normal code generation. | |||
9050 | if (!CGM.getLangOpts().OpenMPIsDevice) | |||
9051 | return false; | |||
9052 | ||||
9053 | const ValueDecl *VD = cast<ValueDecl>(GD.getDecl()); | |||
9054 | StringRef Name = CGM.getMangledName(GD); | |||
9055 | // Try to detect target regions in the function. | |||
9056 | if (const auto *FD = dyn_cast<FunctionDecl>(VD)) | |||
9057 | scanForTargetRegionsFunctions(FD->getBody(), Name); | |||
9058 | ||||
9059 | // Do not to emit function if it is not marked as declare target. | |||
9060 | return !OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD) && | |||
9061 | AlreadyEmittedTargetFunctions.count(Name) == 0; | |||
9062 | } | |||
9063 | ||||
9064 | bool CGOpenMPRuntime::emitTargetGlobalVariable(GlobalDecl GD) { | |||
9065 | if (!CGM.getLangOpts().OpenMPIsDevice) | |||
9066 | return false; | |||
9067 | ||||
9068 | // Check if there are Ctors/Dtors in this declaration and look for target | |||
9069 | // regions in it. We use the complete variant to produce the kernel name | |||
9070 | // mangling. | |||
9071 | QualType RDTy = cast<VarDecl>(GD.getDecl())->getType(); | |||
9072 | if (const auto *RD = RDTy->getBaseElementTypeUnsafe()->getAsCXXRecordDecl()) { | |||
9073 | for (const CXXConstructorDecl *Ctor : RD->ctors()) { | |||
9074 | StringRef ParentName = | |||
9075 | CGM.getMangledName(GlobalDecl(Ctor, Ctor_Complete)); | |||
9076 | scanForTargetRegionsFunctions(Ctor->getBody(), ParentName); | |||
9077 | } | |||
9078 | if (const CXXDestructorDecl *Dtor = RD->getDestructor()) { | |||
9079 | StringRef ParentName = | |||
9080 | CGM.getMangledName(GlobalDecl(Dtor, Dtor_Complete)); | |||
9081 | scanForTargetRegionsFunctions(Dtor->getBody(), ParentName); | |||
9082 | } | |||
9083 | } | |||
9084 | ||||
9085 | // Do not to emit variable if it is not marked as declare target. | |||
9086 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
9087 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration( | |||
9088 | cast<VarDecl>(GD.getDecl())); | |||
9089 | if (!Res || *Res == OMPDeclareTargetDeclAttr::MT_Link) { | |||
9090 | DeferredGlobalVariables.insert(cast<VarDecl>(GD.getDecl())); | |||
9091 | return true; | |||
9092 | } | |||
9093 | return false; | |||
9094 | } | |||
9095 | ||||
9096 | llvm::Constant * | |||
9097 | CGOpenMPRuntime::registerTargetFirstprivateCopy(CodeGenFunction &CGF, | |||
9098 | const VarDecl *VD) { | |||
9099 | assert(VD->getType().isConstant(CGM.getContext()) &&((VD->getType().isConstant(CGM.getContext()) && "Expected constant variable." ) ? static_cast<void> (0) : __assert_fail ("VD->getType().isConstant(CGM.getContext()) && \"Expected constant variable.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9100, __PRETTY_FUNCTION__)) | |||
9100 | "Expected constant variable.")((VD->getType().isConstant(CGM.getContext()) && "Expected constant variable." ) ? static_cast<void> (0) : __assert_fail ("VD->getType().isConstant(CGM.getContext()) && \"Expected constant variable.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9100, __PRETTY_FUNCTION__)); | |||
9101 | StringRef VarName; | |||
9102 | llvm::Constant *Addr; | |||
9103 | llvm::GlobalValue::LinkageTypes Linkage; | |||
9104 | QualType Ty = VD->getType(); | |||
9105 | SmallString<128> Buffer; | |||
9106 | { | |||
9107 | unsigned DeviceID; | |||
9108 | unsigned FileID; | |||
9109 | unsigned Line; | |||
9110 | getTargetEntryUniqueInfo(CGM.getContext(), VD->getLocation(), DeviceID, | |||
9111 | FileID, Line); | |||
9112 | llvm::raw_svector_ostream OS(Buffer); | |||
9113 | OS << "__omp_offloading_firstprivate_" << llvm::format("_%x", DeviceID) | |||
9114 | << llvm::format("_%x_", FileID) << VD->getName() << "_l" << Line; | |||
9115 | VarName = OS.str(); | |||
9116 | } | |||
9117 | Linkage = llvm::GlobalValue::InternalLinkage; | |||
9118 | Addr = | |||
9119 | getOrCreateInternalVariable(CGM.getTypes().ConvertTypeForMem(Ty), VarName, | |||
9120 | getDefaultFirstprivateAddressSpace()); | |||
9121 | cast<llvm::GlobalValue>(Addr)->setLinkage(Linkage); | |||
9122 | CharUnits VarSize = CGM.getContext().getTypeSizeInChars(Ty); | |||
9123 | CGM.addCompilerUsedGlobal(cast<llvm::GlobalValue>(Addr)); | |||
9124 | OffloadEntriesInfoManager.registerDeviceGlobalVarEntryInfo( | |||
9125 | VarName, Addr, VarSize, | |||
9126 | OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryTo, Linkage); | |||
9127 | return Addr; | |||
9128 | } | |||
9129 | ||||
9130 | void CGOpenMPRuntime::registerTargetGlobalVariable(const VarDecl *VD, | |||
9131 | llvm::Constant *Addr) { | |||
9132 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
9133 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | |||
9134 | if (!Res) { | |||
9135 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
9136 | // Register non-target variables being emitted in device code (debug info | |||
9137 | // may cause this). | |||
9138 | StringRef VarName = CGM.getMangledName(VD); | |||
9139 | EmittedNonTargetVariables.try_emplace(VarName, Addr); | |||
9140 | } | |||
9141 | return; | |||
9142 | } | |||
9143 | // Register declare target variables. | |||
9144 | OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryKind Flags; | |||
9145 | StringRef VarName; | |||
9146 | CharUnits VarSize; | |||
9147 | llvm::GlobalValue::LinkageTypes Linkage; | |||
9148 | switch (*Res) { | |||
9149 | case OMPDeclareTargetDeclAttr::MT_To: | |||
9150 | Flags = OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryTo; | |||
9151 | VarName = CGM.getMangledName(VD); | |||
9152 | if (VD->hasDefinition(CGM.getContext()) != VarDecl::DeclarationOnly) { | |||
9153 | VarSize = CGM.getContext().getTypeSizeInChars(VD->getType()); | |||
9154 | assert(!VarSize.isZero() && "Expected non-zero size of the variable")((!VarSize.isZero() && "Expected non-zero size of the variable" ) ? static_cast<void> (0) : __assert_fail ("!VarSize.isZero() && \"Expected non-zero size of the variable\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9154, __PRETTY_FUNCTION__)); | |||
9155 | } else { | |||
9156 | VarSize = CharUnits::Zero(); | |||
9157 | } | |||
9158 | Linkage = CGM.getLLVMLinkageVarDefinition(VD, /*IsConstant=*/false); | |||
9159 | // Temp solution to prevent optimizations of the internal variables. | |||
9160 | if (CGM.getLangOpts().OpenMPIsDevice && !VD->isExternallyVisible()) { | |||
9161 | std::string RefName = getName({VarName, "ref"}); | |||
9162 | if (!CGM.GetGlobalValue(RefName)) { | |||
9163 | llvm::Constant *AddrRef = | |||
9164 | getOrCreateInternalVariable(Addr->getType(), RefName); | |||
9165 | auto *GVAddrRef = cast<llvm::GlobalVariable>(AddrRef); | |||
9166 | GVAddrRef->setConstant(/*Val=*/true); | |||
9167 | GVAddrRef->setLinkage(llvm::GlobalValue::InternalLinkage); | |||
9168 | GVAddrRef->setInitializer(Addr); | |||
9169 | CGM.addCompilerUsedGlobal(GVAddrRef); | |||
9170 | } | |||
9171 | } | |||
9172 | break; | |||
9173 | case OMPDeclareTargetDeclAttr::MT_Link: | |||
9174 | Flags = OffloadEntriesInfoManagerTy::OMPTargetGlobalVarEntryLink; | |||
9175 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
9176 | VarName = Addr->getName(); | |||
9177 | Addr = nullptr; | |||
9178 | } else { | |||
9179 | VarName = getAddrOfDeclareTargetLink(VD).getName(); | |||
9180 | Addr = cast<llvm::Constant>(getAddrOfDeclareTargetLink(VD).getPointer()); | |||
9181 | } | |||
9182 | VarSize = CGM.getPointerSize(); | |||
9183 | Linkage = llvm::GlobalValue::WeakAnyLinkage; | |||
9184 | break; | |||
9185 | } | |||
9186 | OffloadEntriesInfoManager.registerDeviceGlobalVarEntryInfo( | |||
9187 | VarName, Addr, VarSize, Flags, Linkage); | |||
9188 | } | |||
9189 | ||||
9190 | bool CGOpenMPRuntime::emitTargetGlobal(GlobalDecl GD) { | |||
9191 | if (isa<FunctionDecl>(GD.getDecl()) || | |||
9192 | isa<OMPDeclareReductionDecl>(GD.getDecl())) | |||
9193 | return emitTargetFunctions(GD); | |||
9194 | ||||
9195 | return emitTargetGlobalVariable(GD); | |||
9196 | } | |||
9197 | ||||
9198 | void CGOpenMPRuntime::emitDeferredTargetDecls() const { | |||
9199 | for (const VarDecl *VD : DeferredGlobalVariables) { | |||
9200 | llvm::Optional<OMPDeclareTargetDeclAttr::MapTypeTy> Res = | |||
9201 | OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(VD); | |||
9202 | if (!Res) | |||
9203 | continue; | |||
9204 | if (*Res == OMPDeclareTargetDeclAttr::MT_To) { | |||
9205 | CGM.EmitGlobal(VD); | |||
9206 | } else { | |||
9207 | assert(*Res == OMPDeclareTargetDeclAttr::MT_Link &&((*Res == OMPDeclareTargetDeclAttr::MT_Link && "Expected to or link clauses." ) ? static_cast<void> (0) : __assert_fail ("*Res == OMPDeclareTargetDeclAttr::MT_Link && \"Expected to or link clauses.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9208, __PRETTY_FUNCTION__)) | |||
9208 | "Expected to or link clauses.")((*Res == OMPDeclareTargetDeclAttr::MT_Link && "Expected to or link clauses." ) ? static_cast<void> (0) : __assert_fail ("*Res == OMPDeclareTargetDeclAttr::MT_Link && \"Expected to or link clauses.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9208, __PRETTY_FUNCTION__)); | |||
9209 | (void)CGM.getOpenMPRuntime().getAddrOfDeclareTargetLink(VD); | |||
9210 | } | |||
9211 | } | |||
9212 | } | |||
9213 | ||||
9214 | void CGOpenMPRuntime::adjustTargetSpecificDataForLambdas( | |||
9215 | CodeGenFunction &CGF, const OMPExecutableDirective &D) const { | |||
9216 | assert(isOpenMPTargetExecutionDirective(D.getDirectiveKind()) &&((isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && " Expected target-based directive.") ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9217, __PRETTY_FUNCTION__)) | |||
9217 | " Expected target-based directive.")((isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && " Expected target-based directive.") ? static_cast<void> (0) : __assert_fail ("isOpenMPTargetExecutionDirective(D.getDirectiveKind()) && \" Expected target-based directive.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9217, __PRETTY_FUNCTION__)); | |||
9218 | } | |||
9219 | ||||
9220 | void CGOpenMPRuntime::checkArchForUnifiedAddressing( | |||
9221 | const OMPRequiresDecl *D) { | |||
9222 | for (const OMPClause *Clause : D->clauselists()) { | |||
9223 | if (Clause->getClauseKind() == OMPC_unified_shared_memory) { | |||
9224 | HasRequiresUnifiedSharedMemory = true; | |||
9225 | break; | |||
9226 | } | |||
9227 | } | |||
9228 | } | |||
9229 | ||||
9230 | bool CGOpenMPRuntime::hasAllocateAttributeForGlobalVar(const VarDecl *VD, | |||
9231 | LangAS &AS) { | |||
9232 | if (!VD || !VD->hasAttr<OMPAllocateDeclAttr>()) | |||
9233 | return false; | |||
9234 | const auto *A = VD->getAttr<OMPAllocateDeclAttr>(); | |||
9235 | switch(A->getAllocatorType()) { | |||
9236 | case OMPAllocateDeclAttr::OMPDefaultMemAlloc: | |||
9237 | // Not supported, fallback to the default mem space. | |||
9238 | case OMPAllocateDeclAttr::OMPLargeCapMemAlloc: | |||
9239 | case OMPAllocateDeclAttr::OMPCGroupMemAlloc: | |||
9240 | case OMPAllocateDeclAttr::OMPHighBWMemAlloc: | |||
9241 | case OMPAllocateDeclAttr::OMPLowLatMemAlloc: | |||
9242 | case OMPAllocateDeclAttr::OMPThreadMemAlloc: | |||
9243 | case OMPAllocateDeclAttr::OMPConstMemAlloc: | |||
9244 | case OMPAllocateDeclAttr::OMPPTeamMemAlloc: | |||
9245 | AS = LangAS::Default; | |||
9246 | return true; | |||
9247 | case OMPAllocateDeclAttr::OMPUserDefinedMemAlloc: | |||
9248 | llvm_unreachable("Expected predefined allocator for the variables with the "::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the " "static storage.", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9249) | |||
9249 | "static storage.")::llvm::llvm_unreachable_internal("Expected predefined allocator for the variables with the " "static storage.", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9249); | |||
9250 | } | |||
9251 | return false; | |||
9252 | } | |||
9253 | ||||
9254 | CGOpenMPRuntime::DisableAutoDeclareTargetRAII::DisableAutoDeclareTargetRAII( | |||
9255 | CodeGenModule &CGM) | |||
9256 | : CGM(CGM) { | |||
9257 | if (CGM.getLangOpts().OpenMPIsDevice) { | |||
9258 | SavedShouldMarkAsGlobal = CGM.getOpenMPRuntime().ShouldMarkAsGlobal; | |||
9259 | CGM.getOpenMPRuntime().ShouldMarkAsGlobal = false; | |||
9260 | } | |||
9261 | } | |||
9262 | ||||
9263 | CGOpenMPRuntime::DisableAutoDeclareTargetRAII::~DisableAutoDeclareTargetRAII() { | |||
9264 | if (CGM.getLangOpts().OpenMPIsDevice) | |||
9265 | CGM.getOpenMPRuntime().ShouldMarkAsGlobal = SavedShouldMarkAsGlobal; | |||
9266 | } | |||
9267 | ||||
9268 | bool CGOpenMPRuntime::markAsGlobalTarget(GlobalDecl GD) { | |||
9269 | if (!CGM.getLangOpts().OpenMPIsDevice || !ShouldMarkAsGlobal) | |||
9270 | return true; | |||
9271 | ||||
9272 | StringRef Name = CGM.getMangledName(GD); | |||
9273 | const auto *D = cast<FunctionDecl>(GD.getDecl()); | |||
9274 | // Do not to emit function if it is marked as declare target as it was already | |||
9275 | // emitted. | |||
9276 | if (OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(D)) { | |||
9277 | if (D->hasBody() && AlreadyEmittedTargetFunctions.count(Name) == 0) { | |||
9278 | if (auto *F = dyn_cast_or_null<llvm::Function>(CGM.GetGlobalValue(Name))) | |||
9279 | return !F->isDeclaration(); | |||
9280 | return false; | |||
9281 | } | |||
9282 | return true; | |||
9283 | } | |||
9284 | ||||
9285 | return !AlreadyEmittedTargetFunctions.insert(Name).second; | |||
9286 | } | |||
9287 | ||||
9288 | llvm::Function *CGOpenMPRuntime::emitRequiresDirectiveRegFun() { | |||
9289 | // If we don't have entries or if we are emitting code for the device, we | |||
9290 | // don't need to do anything. | |||
9291 | if (CGM.getLangOpts().OMPTargetTriples.empty() || | |||
9292 | CGM.getLangOpts().OpenMPSimd || CGM.getLangOpts().OpenMPIsDevice || | |||
9293 | (OffloadEntriesInfoManager.empty() && | |||
9294 | !HasEmittedDeclareTargetRegion && | |||
9295 | !HasEmittedTargetRegion)) | |||
9296 | return nullptr; | |||
9297 | ||||
9298 | // Create and register the function that handles the requires directives. | |||
9299 | ASTContext &C = CGM.getContext(); | |||
9300 | ||||
9301 | llvm::Function *RequiresRegFn; | |||
9302 | { | |||
9303 | CodeGenFunction CGF(CGM); | |||
9304 | const auto &FI = CGM.getTypes().arrangeNullaryFunction(); | |||
9305 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(FI); | |||
9306 | std::string ReqName = getName({"omp_offloading", "requires_reg"}); | |||
9307 | RequiresRegFn = CGM.CreateGlobalInitOrDestructFunction(FTy, ReqName, FI); | |||
9308 | CGF.StartFunction(GlobalDecl(), C.VoidTy, RequiresRegFn, FI, {}); | |||
9309 | OpenMPOffloadingRequiresDirFlags Flags = OMP_REQ_NONE; | |||
9310 | // TODO: check for other requires clauses. | |||
9311 | // The requires directive takes effect only when a target region is | |||
9312 | // present in the compilation unit. Otherwise it is ignored and not | |||
9313 | // passed to the runtime. This avoids the runtime from throwing an error | |||
9314 | // for mismatching requires clauses across compilation units that don't | |||
9315 | // contain at least 1 target region. | |||
9316 | assert((HasEmittedTargetRegion ||(((HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && "Target or declare target region expected." ) ? static_cast<void> (0) : __assert_fail ("(HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && \"Target or declare target region expected.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9319, __PRETTY_FUNCTION__)) | |||
9317 | HasEmittedDeclareTargetRegion ||(((HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && "Target or declare target region expected." ) ? static_cast<void> (0) : __assert_fail ("(HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && \"Target or declare target region expected.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9319, __PRETTY_FUNCTION__)) | |||
9318 | !OffloadEntriesInfoManager.empty()) &&(((HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && "Target or declare target region expected." ) ? static_cast<void> (0) : __assert_fail ("(HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && \"Target or declare target region expected.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9319, __PRETTY_FUNCTION__)) | |||
9319 | "Target or declare target region expected.")(((HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && "Target or declare target region expected." ) ? static_cast<void> (0) : __assert_fail ("(HasEmittedTargetRegion || HasEmittedDeclareTargetRegion || !OffloadEntriesInfoManager.empty()) && \"Target or declare target region expected.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9319, __PRETTY_FUNCTION__)); | |||
9320 | if (HasRequiresUnifiedSharedMemory) | |||
9321 | Flags = OMP_REQ_UNIFIED_SHARED_MEMORY; | |||
9322 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__tgt_register_requires), | |||
9323 | llvm::ConstantInt::get(CGM.Int64Ty, Flags)); | |||
9324 | CGF.FinishFunction(); | |||
9325 | } | |||
9326 | return RequiresRegFn; | |||
9327 | } | |||
9328 | ||||
9329 | llvm::Function *CGOpenMPRuntime::emitRegistrationFunction() { | |||
9330 | // If we have offloading in the current module, we need to emit the entries | |||
9331 | // now and register the offloading descriptor. | |||
9332 | createOffloadEntriesAndInfoMetadata(); | |||
9333 | ||||
9334 | // Create and register the offloading binary descriptors. This is the main | |||
9335 | // entity that captures all the information about offloading in the current | |||
9336 | // compilation unit. | |||
9337 | return createOffloadingBinaryDescriptorRegistration(); | |||
9338 | } | |||
9339 | ||||
9340 | void CGOpenMPRuntime::emitTeamsCall(CodeGenFunction &CGF, | |||
9341 | const OMPExecutableDirective &D, | |||
9342 | SourceLocation Loc, | |||
9343 | llvm::Function *OutlinedFn, | |||
9344 | ArrayRef<llvm::Value *> CapturedVars) { | |||
9345 | if (!CGF.HaveInsertPoint()) | |||
9346 | return; | |||
9347 | ||||
9348 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | |||
9349 | CodeGenFunction::RunCleanupsScope Scope(CGF); | |||
9350 | ||||
9351 | // Build call __kmpc_fork_teams(loc, n, microtask, var1, .., varn); | |||
9352 | llvm::Value *Args[] = { | |||
9353 | RTLoc, | |||
9354 | CGF.Builder.getInt32(CapturedVars.size()), // Number of captured vars | |||
9355 | CGF.Builder.CreateBitCast(OutlinedFn, getKmpc_MicroPointerTy())}; | |||
9356 | llvm::SmallVector<llvm::Value *, 16> RealArgs; | |||
9357 | RealArgs.append(std::begin(Args), std::end(Args)); | |||
9358 | RealArgs.append(CapturedVars.begin(), CapturedVars.end()); | |||
9359 | ||||
9360 | llvm::FunctionCallee RTLFn = createRuntimeFunction(OMPRTL__kmpc_fork_teams); | |||
9361 | CGF.EmitRuntimeCall(RTLFn, RealArgs); | |||
9362 | } | |||
9363 | ||||
9364 | void CGOpenMPRuntime::emitNumTeamsClause(CodeGenFunction &CGF, | |||
9365 | const Expr *NumTeams, | |||
9366 | const Expr *ThreadLimit, | |||
9367 | SourceLocation Loc) { | |||
9368 | if (!CGF.HaveInsertPoint()) | |||
9369 | return; | |||
9370 | ||||
9371 | llvm::Value *RTLoc = emitUpdateLocation(CGF, Loc); | |||
9372 | ||||
9373 | llvm::Value *NumTeamsVal = | |||
9374 | NumTeams | |||
9375 | ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(NumTeams), | |||
9376 | CGF.CGM.Int32Ty, /* isSigned = */ true) | |||
9377 | : CGF.Builder.getInt32(0); | |||
9378 | ||||
9379 | llvm::Value *ThreadLimitVal = | |||
9380 | ThreadLimit | |||
9381 | ? CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(ThreadLimit), | |||
9382 | CGF.CGM.Int32Ty, /* isSigned = */ true) | |||
9383 | : CGF.Builder.getInt32(0); | |||
9384 | ||||
9385 | // Build call __kmpc_push_num_teamss(&loc, global_tid, num_teams, thread_limit) | |||
9386 | llvm::Value *PushNumTeamsArgs[] = {RTLoc, getThreadID(CGF, Loc), NumTeamsVal, | |||
9387 | ThreadLimitVal}; | |||
9388 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_push_num_teams), | |||
9389 | PushNumTeamsArgs); | |||
9390 | } | |||
9391 | ||||
9392 | void CGOpenMPRuntime::emitTargetDataCalls( | |||
9393 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond, | |||
9394 | const Expr *Device, const RegionCodeGenTy &CodeGen, TargetDataInfo &Info) { | |||
9395 | if (!CGF.HaveInsertPoint()) | |||
9396 | return; | |||
9397 | ||||
9398 | // Action used to replace the default codegen action and turn privatization | |||
9399 | // off. | |||
9400 | PrePostActionTy NoPrivAction; | |||
9401 | ||||
9402 | // Generate the code for the opening of the data environment. Capture all the | |||
9403 | // arguments of the runtime call by reference because they are used in the | |||
9404 | // closing of the region. | |||
9405 | auto &&BeginThenGen = [this, &D, Device, &Info, | |||
9406 | &CodeGen](CodeGenFunction &CGF, PrePostActionTy &) { | |||
9407 | // Fill up the arrays with all the mapped variables. | |||
9408 | MappableExprsHandler::MapBaseValuesArrayTy BasePointers; | |||
9409 | MappableExprsHandler::MapValuesArrayTy Pointers; | |||
9410 | MappableExprsHandler::MapValuesArrayTy Sizes; | |||
9411 | MappableExprsHandler::MapFlagsArrayTy MapTypes; | |||
9412 | ||||
9413 | // Get map clause information. | |||
9414 | MappableExprsHandler MCHandler(D, CGF); | |||
9415 | MCHandler.generateAllInfo(BasePointers, Pointers, Sizes, MapTypes); | |||
9416 | ||||
9417 | // Fill up the arrays and create the arguments. | |||
9418 | emitOffloadingArrays(CGF, BasePointers, Pointers, Sizes, MapTypes, Info); | |||
9419 | ||||
9420 | llvm::Value *BasePointersArrayArg = nullptr; | |||
9421 | llvm::Value *PointersArrayArg = nullptr; | |||
9422 | llvm::Value *SizesArrayArg = nullptr; | |||
9423 | llvm::Value *MapTypesArrayArg = nullptr; | |||
9424 | emitOffloadingArraysArgument(CGF, BasePointersArrayArg, PointersArrayArg, | |||
9425 | SizesArrayArg, MapTypesArrayArg, Info); | |||
9426 | ||||
9427 | // Emit device ID if any. | |||
9428 | llvm::Value *DeviceID = nullptr; | |||
9429 | if (Device) { | |||
9430 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | |||
9431 | CGF.Int64Ty, /*isSigned=*/true); | |||
9432 | } else { | |||
9433 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | |||
9434 | } | |||
9435 | ||||
9436 | // Emit the number of elements in the offloading arrays. | |||
9437 | llvm::Value *PointerNum = CGF.Builder.getInt32(Info.NumberOfPtrs); | |||
9438 | ||||
9439 | llvm::Value *OffloadingArgs[] = { | |||
9440 | DeviceID, PointerNum, BasePointersArrayArg, | |||
9441 | PointersArrayArg, SizesArrayArg, MapTypesArrayArg}; | |||
9442 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__tgt_target_data_begin), | |||
9443 | OffloadingArgs); | |||
9444 | ||||
9445 | // If device pointer privatization is required, emit the body of the region | |||
9446 | // here. It will have to be duplicated: with and without privatization. | |||
9447 | if (!Info.CaptureDeviceAddrMap.empty()) | |||
9448 | CodeGen(CGF); | |||
9449 | }; | |||
9450 | ||||
9451 | // Generate code for the closing of the data region. | |||
9452 | auto &&EndThenGen = [this, Device, &Info](CodeGenFunction &CGF, | |||
9453 | PrePostActionTy &) { | |||
9454 | assert(Info.isValid() && "Invalid data environment closing arguments.")((Info.isValid() && "Invalid data environment closing arguments." ) ? static_cast<void> (0) : __assert_fail ("Info.isValid() && \"Invalid data environment closing arguments.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9454, __PRETTY_FUNCTION__)); | |||
9455 | ||||
9456 | llvm::Value *BasePointersArrayArg = nullptr; | |||
9457 | llvm::Value *PointersArrayArg = nullptr; | |||
9458 | llvm::Value *SizesArrayArg = nullptr; | |||
9459 | llvm::Value *MapTypesArrayArg = nullptr; | |||
9460 | emitOffloadingArraysArgument(CGF, BasePointersArrayArg, PointersArrayArg, | |||
9461 | SizesArrayArg, MapTypesArrayArg, Info); | |||
9462 | ||||
9463 | // Emit device ID if any. | |||
9464 | llvm::Value *DeviceID = nullptr; | |||
9465 | if (Device) { | |||
9466 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | |||
9467 | CGF.Int64Ty, /*isSigned=*/true); | |||
9468 | } else { | |||
9469 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | |||
9470 | } | |||
9471 | ||||
9472 | // Emit the number of elements in the offloading arrays. | |||
9473 | llvm::Value *PointerNum = CGF.Builder.getInt32(Info.NumberOfPtrs); | |||
9474 | ||||
9475 | llvm::Value *OffloadingArgs[] = { | |||
9476 | DeviceID, PointerNum, BasePointersArrayArg, | |||
9477 | PointersArrayArg, SizesArrayArg, MapTypesArrayArg}; | |||
9478 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__tgt_target_data_end), | |||
9479 | OffloadingArgs); | |||
9480 | }; | |||
9481 | ||||
9482 | // If we need device pointer privatization, we need to emit the body of the | |||
9483 | // region with no privatization in the 'else' branch of the conditional. | |||
9484 | // Otherwise, we don't have to do anything. | |||
9485 | auto &&BeginElseGen = [&Info, &CodeGen, &NoPrivAction](CodeGenFunction &CGF, | |||
9486 | PrePostActionTy &) { | |||
9487 | if (!Info.CaptureDeviceAddrMap.empty()) { | |||
9488 | CodeGen.setAction(NoPrivAction); | |||
9489 | CodeGen(CGF); | |||
9490 | } | |||
9491 | }; | |||
9492 | ||||
9493 | // We don't have to do anything to close the region if the if clause evaluates | |||
9494 | // to false. | |||
9495 | auto &&EndElseGen = [](CodeGenFunction &CGF, PrePostActionTy &) {}; | |||
9496 | ||||
9497 | if (IfCond) { | |||
9498 | emitOMPIfClause(CGF, IfCond, BeginThenGen, BeginElseGen); | |||
9499 | } else { | |||
9500 | RegionCodeGenTy RCG(BeginThenGen); | |||
9501 | RCG(CGF); | |||
9502 | } | |||
9503 | ||||
9504 | // If we don't require privatization of device pointers, we emit the body in | |||
9505 | // between the runtime calls. This avoids duplicating the body code. | |||
9506 | if (Info.CaptureDeviceAddrMap.empty()) { | |||
9507 | CodeGen.setAction(NoPrivAction); | |||
9508 | CodeGen(CGF); | |||
9509 | } | |||
9510 | ||||
9511 | if (IfCond) { | |||
9512 | emitOMPIfClause(CGF, IfCond, EndThenGen, EndElseGen); | |||
9513 | } else { | |||
9514 | RegionCodeGenTy RCG(EndThenGen); | |||
9515 | RCG(CGF); | |||
9516 | } | |||
9517 | } | |||
9518 | ||||
9519 | void CGOpenMPRuntime::emitTargetDataStandAloneCall( | |||
9520 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond, | |||
9521 | const Expr *Device) { | |||
9522 | if (!CGF.HaveInsertPoint()) | |||
9523 | return; | |||
9524 | ||||
9525 | assert((isa<OMPTargetEnterDataDirective>(D) ||(((isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective >(D) || isa<OMPTargetUpdateDirective>(D)) && "Expecting either target enter, exit data, or update directives." ) ? static_cast<void> (0) : __assert_fail ("(isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective>(D) || isa<OMPTargetUpdateDirective>(D)) && \"Expecting either target enter, exit data, or update directives.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9528, __PRETTY_FUNCTION__)) | |||
9526 | isa<OMPTargetExitDataDirective>(D) ||(((isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective >(D) || isa<OMPTargetUpdateDirective>(D)) && "Expecting either target enter, exit data, or update directives." ) ? static_cast<void> (0) : __assert_fail ("(isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective>(D) || isa<OMPTargetUpdateDirective>(D)) && \"Expecting either target enter, exit data, or update directives.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9528, __PRETTY_FUNCTION__)) | |||
9527 | isa<OMPTargetUpdateDirective>(D)) &&(((isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective >(D) || isa<OMPTargetUpdateDirective>(D)) && "Expecting either target enter, exit data, or update directives." ) ? static_cast<void> (0) : __assert_fail ("(isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective>(D) || isa<OMPTargetUpdateDirective>(D)) && \"Expecting either target enter, exit data, or update directives.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9528, __PRETTY_FUNCTION__)) | |||
9528 | "Expecting either target enter, exit data, or update directives.")(((isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective >(D) || isa<OMPTargetUpdateDirective>(D)) && "Expecting either target enter, exit data, or update directives." ) ? static_cast<void> (0) : __assert_fail ("(isa<OMPTargetEnterDataDirective>(D) || isa<OMPTargetExitDataDirective>(D) || isa<OMPTargetUpdateDirective>(D)) && \"Expecting either target enter, exit data, or update directives.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9528, __PRETTY_FUNCTION__)); | |||
9529 | ||||
9530 | CodeGenFunction::OMPTargetDataInfo InputInfo; | |||
9531 | llvm::Value *MapTypesArray = nullptr; | |||
9532 | // Generate the code for the opening of the data environment. | |||
9533 | auto &&ThenGen = [this, &D, Device, &InputInfo, | |||
9534 | &MapTypesArray](CodeGenFunction &CGF, PrePostActionTy &) { | |||
9535 | // Emit device ID if any. | |||
9536 | llvm::Value *DeviceID = nullptr; | |||
9537 | if (Device) { | |||
9538 | DeviceID = CGF.Builder.CreateIntCast(CGF.EmitScalarExpr(Device), | |||
9539 | CGF.Int64Ty, /*isSigned=*/true); | |||
9540 | } else { | |||
9541 | DeviceID = CGF.Builder.getInt64(OMP_DEVICEID_UNDEF); | |||
9542 | } | |||
9543 | ||||
9544 | // Emit the number of elements in the offloading arrays. | |||
9545 | llvm::Constant *PointerNum = | |||
9546 | CGF.Builder.getInt32(InputInfo.NumberOfTargetItems); | |||
9547 | ||||
9548 | llvm::Value *OffloadingArgs[] = {DeviceID, | |||
9549 | PointerNum, | |||
9550 | InputInfo.BasePointersArray.getPointer(), | |||
9551 | InputInfo.PointersArray.getPointer(), | |||
9552 | InputInfo.SizesArray.getPointer(), | |||
9553 | MapTypesArray}; | |||
9554 | ||||
9555 | // Select the right runtime function call for each expected standalone | |||
9556 | // directive. | |||
9557 | const bool HasNowait = D.hasClausesOfKind<OMPNowaitClause>(); | |||
9558 | OpenMPRTLFunction RTLFn; | |||
9559 | switch (D.getDirectiveKind()) { | |||
9560 | case OMPD_target_enter_data: | |||
9561 | RTLFn = HasNowait ? OMPRTL__tgt_target_data_begin_nowait | |||
9562 | : OMPRTL__tgt_target_data_begin; | |||
9563 | break; | |||
9564 | case OMPD_target_exit_data: | |||
9565 | RTLFn = HasNowait ? OMPRTL__tgt_target_data_end_nowait | |||
9566 | : OMPRTL__tgt_target_data_end; | |||
9567 | break; | |||
9568 | case OMPD_target_update: | |||
9569 | RTLFn = HasNowait ? OMPRTL__tgt_target_data_update_nowait | |||
9570 | : OMPRTL__tgt_target_data_update; | |||
9571 | break; | |||
9572 | case OMPD_parallel: | |||
9573 | case OMPD_for: | |||
9574 | case OMPD_parallel_for: | |||
9575 | case OMPD_parallel_sections: | |||
9576 | case OMPD_for_simd: | |||
9577 | case OMPD_parallel_for_simd: | |||
9578 | case OMPD_cancel: | |||
9579 | case OMPD_cancellation_point: | |||
9580 | case OMPD_ordered: | |||
9581 | case OMPD_threadprivate: | |||
9582 | case OMPD_allocate: | |||
9583 | case OMPD_task: | |||
9584 | case OMPD_simd: | |||
9585 | case OMPD_sections: | |||
9586 | case OMPD_section: | |||
9587 | case OMPD_single: | |||
9588 | case OMPD_master: | |||
9589 | case OMPD_critical: | |||
9590 | case OMPD_taskyield: | |||
9591 | case OMPD_barrier: | |||
9592 | case OMPD_taskwait: | |||
9593 | case OMPD_taskgroup: | |||
9594 | case OMPD_atomic: | |||
9595 | case OMPD_flush: | |||
9596 | case OMPD_teams: | |||
9597 | case OMPD_target_data: | |||
9598 | case OMPD_distribute: | |||
9599 | case OMPD_distribute_simd: | |||
9600 | case OMPD_distribute_parallel_for: | |||
9601 | case OMPD_distribute_parallel_for_simd: | |||
9602 | case OMPD_teams_distribute: | |||
9603 | case OMPD_teams_distribute_simd: | |||
9604 | case OMPD_teams_distribute_parallel_for: | |||
9605 | case OMPD_teams_distribute_parallel_for_simd: | |||
9606 | case OMPD_declare_simd: | |||
9607 | case OMPD_declare_target: | |||
9608 | case OMPD_end_declare_target: | |||
9609 | case OMPD_declare_reduction: | |||
9610 | case OMPD_declare_mapper: | |||
9611 | case OMPD_taskloop: | |||
9612 | case OMPD_taskloop_simd: | |||
9613 | case OMPD_target: | |||
9614 | case OMPD_target_simd: | |||
9615 | case OMPD_target_teams_distribute: | |||
9616 | case OMPD_target_teams_distribute_simd: | |||
9617 | case OMPD_target_teams_distribute_parallel_for: | |||
9618 | case OMPD_target_teams_distribute_parallel_for_simd: | |||
9619 | case OMPD_target_teams: | |||
9620 | case OMPD_target_parallel: | |||
9621 | case OMPD_target_parallel_for: | |||
9622 | case OMPD_target_parallel_for_simd: | |||
9623 | case OMPD_requires: | |||
9624 | case OMPD_unknown: | |||
9625 | llvm_unreachable("Unexpected standalone target data directive.")::llvm::llvm_unreachable_internal("Unexpected standalone target data directive." , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9625); | |||
9626 | break; | |||
9627 | } | |||
9628 | CGF.EmitRuntimeCall(createRuntimeFunction(RTLFn), OffloadingArgs); | |||
9629 | }; | |||
9630 | ||||
9631 | auto &&TargetThenGen = [this, &ThenGen, &D, &InputInfo, &MapTypesArray]( | |||
9632 | CodeGenFunction &CGF, PrePostActionTy &) { | |||
9633 | // Fill up the arrays with all the mapped variables. | |||
9634 | MappableExprsHandler::MapBaseValuesArrayTy BasePointers; | |||
9635 | MappableExprsHandler::MapValuesArrayTy Pointers; | |||
9636 | MappableExprsHandler::MapValuesArrayTy Sizes; | |||
9637 | MappableExprsHandler::MapFlagsArrayTy MapTypes; | |||
9638 | ||||
9639 | // Get map clause information. | |||
9640 | MappableExprsHandler MEHandler(D, CGF); | |||
9641 | MEHandler.generateAllInfo(BasePointers, Pointers, Sizes, MapTypes); | |||
9642 | ||||
9643 | TargetDataInfo Info; | |||
9644 | // Fill up the arrays and create the arguments. | |||
9645 | emitOffloadingArrays(CGF, BasePointers, Pointers, Sizes, MapTypes, Info); | |||
9646 | emitOffloadingArraysArgument(CGF, Info.BasePointersArray, | |||
9647 | Info.PointersArray, Info.SizesArray, | |||
9648 | Info.MapTypesArray, Info); | |||
9649 | InputInfo.NumberOfTargetItems = Info.NumberOfPtrs; | |||
9650 | InputInfo.BasePointersArray = | |||
9651 | Address(Info.BasePointersArray, CGM.getPointerAlign()); | |||
9652 | InputInfo.PointersArray = | |||
9653 | Address(Info.PointersArray, CGM.getPointerAlign()); | |||
9654 | InputInfo.SizesArray = | |||
9655 | Address(Info.SizesArray, CGM.getPointerAlign()); | |||
9656 | MapTypesArray = Info.MapTypesArray; | |||
9657 | if (D.hasClausesOfKind<OMPDependClause>()) | |||
9658 | CGF.EmitOMPTargetTaskBasedDirective(D, ThenGen, InputInfo); | |||
9659 | else | |||
9660 | emitInlinedDirective(CGF, D.getDirectiveKind(), ThenGen); | |||
9661 | }; | |||
9662 | ||||
9663 | if (IfCond) { | |||
9664 | emitOMPIfClause(CGF, IfCond, TargetThenGen, | |||
9665 | [](CodeGenFunction &CGF, PrePostActionTy &) {}); | |||
9666 | } else { | |||
9667 | RegionCodeGenTy ThenRCG(TargetThenGen); | |||
9668 | ThenRCG(CGF); | |||
9669 | } | |||
9670 | } | |||
9671 | ||||
9672 | namespace { | |||
9673 | /// Kind of parameter in a function with 'declare simd' directive. | |||
9674 | enum ParamKindTy { LinearWithVarStride, Linear, Uniform, Vector }; | |||
9675 | /// Attribute set of the parameter. | |||
9676 | struct ParamAttrTy { | |||
9677 | ParamKindTy Kind = Vector; | |||
9678 | llvm::APSInt StrideOrArg; | |||
9679 | llvm::APSInt Alignment; | |||
9680 | }; | |||
9681 | } // namespace | |||
9682 | ||||
9683 | static unsigned evaluateCDTSize(const FunctionDecl *FD, | |||
9684 | ArrayRef<ParamAttrTy> ParamAttrs) { | |||
9685 | // Every vector variant of a SIMD-enabled function has a vector length (VLEN). | |||
9686 | // If OpenMP clause "simdlen" is used, the VLEN is the value of the argument | |||
9687 | // of that clause. The VLEN value must be power of 2. | |||
9688 | // In other case the notion of the function`s "characteristic data type" (CDT) | |||
9689 | // is used to compute the vector length. | |||
9690 | // CDT is defined in the following order: | |||
9691 | // a) For non-void function, the CDT is the return type. | |||
9692 | // b) If the function has any non-uniform, non-linear parameters, then the | |||
9693 | // CDT is the type of the first such parameter. | |||
9694 | // c) If the CDT determined by a) or b) above is struct, union, or class | |||
9695 | // type which is pass-by-value (except for the type that maps to the | |||
9696 | // built-in complex data type), the characteristic data type is int. | |||
9697 | // d) If none of the above three cases is applicable, the CDT is int. | |||
9698 | // The VLEN is then determined based on the CDT and the size of vector | |||
9699 | // register of that ISA for which current vector version is generated. The | |||
9700 | // VLEN is computed using the formula below: | |||
9701 | // VLEN = sizeof(vector_register) / sizeof(CDT), | |||
9702 | // where vector register size specified in section 3.2.1 Registers and the | |||
9703 | // Stack Frame of original AMD64 ABI document. | |||
9704 | QualType RetType = FD->getReturnType(); | |||
9705 | if (RetType.isNull()) | |||
9706 | return 0; | |||
9707 | ASTContext &C = FD->getASTContext(); | |||
9708 | QualType CDT; | |||
9709 | if (!RetType.isNull() && !RetType->isVoidType()) { | |||
9710 | CDT = RetType; | |||
9711 | } else { | |||
9712 | unsigned Offset = 0; | |||
9713 | if (const auto *MD = dyn_cast<CXXMethodDecl>(FD)) { | |||
9714 | if (ParamAttrs[Offset].Kind == Vector) | |||
9715 | CDT = C.getPointerType(C.getRecordType(MD->getParent())); | |||
9716 | ++Offset; | |||
9717 | } | |||
9718 | if (CDT.isNull()) { | |||
9719 | for (unsigned I = 0, E = FD->getNumParams(); I < E; ++I) { | |||
9720 | if (ParamAttrs[I + Offset].Kind == Vector) { | |||
9721 | CDT = FD->getParamDecl(I)->getType(); | |||
9722 | break; | |||
9723 | } | |||
9724 | } | |||
9725 | } | |||
9726 | } | |||
9727 | if (CDT.isNull()) | |||
9728 | CDT = C.IntTy; | |||
9729 | CDT = CDT->getCanonicalTypeUnqualified(); | |||
9730 | if (CDT->isRecordType() || CDT->isUnionType()) | |||
9731 | CDT = C.IntTy; | |||
9732 | return C.getTypeSize(CDT); | |||
9733 | } | |||
9734 | ||||
9735 | static void | |||
9736 | emitX86DeclareSimdFunction(const FunctionDecl *FD, llvm::Function *Fn, | |||
9737 | const llvm::APSInt &VLENVal, | |||
9738 | ArrayRef<ParamAttrTy> ParamAttrs, | |||
9739 | OMPDeclareSimdDeclAttr::BranchStateTy State) { | |||
9740 | struct ISADataTy { | |||
9741 | char ISA; | |||
9742 | unsigned VecRegSize; | |||
9743 | }; | |||
9744 | ISADataTy ISAData[] = { | |||
9745 | { | |||
9746 | 'b', 128 | |||
9747 | }, // SSE | |||
9748 | { | |||
9749 | 'c', 256 | |||
9750 | }, // AVX | |||
9751 | { | |||
9752 | 'd', 256 | |||
9753 | }, // AVX2 | |||
9754 | { | |||
9755 | 'e', 512 | |||
9756 | }, // AVX512 | |||
9757 | }; | |||
9758 | llvm::SmallVector<char, 2> Masked; | |||
9759 | switch (State) { | |||
9760 | case OMPDeclareSimdDeclAttr::BS_Undefined: | |||
9761 | Masked.push_back('N'); | |||
9762 | Masked.push_back('M'); | |||
9763 | break; | |||
9764 | case OMPDeclareSimdDeclAttr::BS_Notinbranch: | |||
9765 | Masked.push_back('N'); | |||
9766 | break; | |||
9767 | case OMPDeclareSimdDeclAttr::BS_Inbranch: | |||
9768 | Masked.push_back('M'); | |||
9769 | break; | |||
9770 | } | |||
9771 | for (char Mask : Masked) { | |||
9772 | for (const ISADataTy &Data : ISAData) { | |||
9773 | SmallString<256> Buffer; | |||
9774 | llvm::raw_svector_ostream Out(Buffer); | |||
9775 | Out << "_ZGV" << Data.ISA << Mask; | |||
9776 | if (!VLENVal) { | |||
9777 | unsigned NumElts = evaluateCDTSize(FD, ParamAttrs); | |||
9778 | assert(NumElts && "Non-zero simdlen/cdtsize expected")((NumElts && "Non-zero simdlen/cdtsize expected") ? static_cast <void> (0) : __assert_fail ("NumElts && \"Non-zero simdlen/cdtsize expected\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9778, __PRETTY_FUNCTION__)); | |||
9779 | Out << llvm::APSInt::getUnsigned(Data.VecRegSize / NumElts); | |||
9780 | } else { | |||
9781 | Out << VLENVal; | |||
9782 | } | |||
9783 | for (const ParamAttrTy &ParamAttr : ParamAttrs) { | |||
9784 | switch (ParamAttr.Kind){ | |||
9785 | case LinearWithVarStride: | |||
9786 | Out << 's' << ParamAttr.StrideOrArg; | |||
9787 | break; | |||
9788 | case Linear: | |||
9789 | Out << 'l'; | |||
9790 | if (!!ParamAttr.StrideOrArg) | |||
9791 | Out << ParamAttr.StrideOrArg; | |||
9792 | break; | |||
9793 | case Uniform: | |||
9794 | Out << 'u'; | |||
9795 | break; | |||
9796 | case Vector: | |||
9797 | Out << 'v'; | |||
9798 | break; | |||
9799 | } | |||
9800 | if (!!ParamAttr.Alignment) | |||
9801 | Out << 'a' << ParamAttr.Alignment; | |||
9802 | } | |||
9803 | Out << '_' << Fn->getName(); | |||
9804 | Fn->addFnAttr(Out.str()); | |||
9805 | } | |||
9806 | } | |||
9807 | } | |||
9808 | ||||
9809 | // This are the Functions that are needed to mangle the name of the | |||
9810 | // vector functions generated by the compiler, according to the rules | |||
9811 | // defined in the "Vector Function ABI specifications for AArch64", | |||
9812 | // available at | |||
9813 | // https://developer.arm.com/products/software-development-tools/hpc/arm-compiler-for-hpc/vector-function-abi. | |||
9814 | ||||
9815 | /// Maps To Vector (MTV), as defined in 3.1.1 of the AAVFABI. | |||
9816 | /// | |||
9817 | /// TODO: Need to implement the behavior for reference marked with a | |||
9818 | /// var or no linear modifiers (1.b in the section). For this, we | |||
9819 | /// need to extend ParamKindTy to support the linear modifiers. | |||
9820 | static bool getAArch64MTV(QualType QT, ParamKindTy Kind) { | |||
9821 | QT = QT.getCanonicalType(); | |||
9822 | ||||
9823 | if (QT->isVoidType()) | |||
9824 | return false; | |||
9825 | ||||
9826 | if (Kind == ParamKindTy::Uniform) | |||
9827 | return false; | |||
9828 | ||||
9829 | if (Kind == ParamKindTy::Linear) | |||
9830 | return false; | |||
9831 | ||||
9832 | // TODO: Handle linear references with modifiers | |||
9833 | ||||
9834 | if (Kind == ParamKindTy::LinearWithVarStride) | |||
9835 | return false; | |||
9836 | ||||
9837 | return true; | |||
9838 | } | |||
9839 | ||||
9840 | /// Pass By Value (PBV), as defined in 3.1.2 of the AAVFABI. | |||
9841 | static bool getAArch64PBV(QualType QT, ASTContext &C) { | |||
9842 | QT = QT.getCanonicalType(); | |||
9843 | unsigned Size = C.getTypeSize(QT); | |||
9844 | ||||
9845 | // Only scalars and complex within 16 bytes wide set PVB to true. | |||
9846 | if (Size != 8 && Size != 16 && Size != 32 && Size != 64 && Size != 128) | |||
9847 | return false; | |||
9848 | ||||
9849 | if (QT->isFloatingType()) | |||
9850 | return true; | |||
9851 | ||||
9852 | if (QT->isIntegerType()) | |||
9853 | return true; | |||
9854 | ||||
9855 | if (QT->isPointerType()) | |||
9856 | return true; | |||
9857 | ||||
9858 | // TODO: Add support for complex types (section 3.1.2, item 2). | |||
9859 | ||||
9860 | return false; | |||
9861 | } | |||
9862 | ||||
9863 | /// Computes the lane size (LS) of a return type or of an input parameter, | |||
9864 | /// as defined by `LS(P)` in 3.2.1 of the AAVFABI. | |||
9865 | /// TODO: Add support for references, section 3.2.1, item 1. | |||
9866 | static unsigned getAArch64LS(QualType QT, ParamKindTy Kind, ASTContext &C) { | |||
9867 | if (getAArch64MTV(QT, Kind) && QT.getCanonicalType()->isPointerType()) { | |||
9868 | QualType PTy = QT.getCanonicalType()->getPointeeType(); | |||
9869 | if (getAArch64PBV(PTy, C)) | |||
9870 | return C.getTypeSize(PTy); | |||
9871 | } | |||
9872 | if (getAArch64PBV(QT, C)) | |||
9873 | return C.getTypeSize(QT); | |||
9874 | ||||
9875 | return C.getTypeSize(C.getUIntPtrType()); | |||
9876 | } | |||
9877 | ||||
9878 | // Get Narrowest Data Size (NDS) and Widest Data Size (WDS) from the | |||
9879 | // signature of the scalar function, as defined in 3.2.2 of the | |||
9880 | // AAVFABI. | |||
9881 | static std::tuple<unsigned, unsigned, bool> | |||
9882 | getNDSWDS(const FunctionDecl *FD, ArrayRef<ParamAttrTy> ParamAttrs) { | |||
9883 | QualType RetType = FD->getReturnType().getCanonicalType(); | |||
9884 | ||||
9885 | ASTContext &C = FD->getASTContext(); | |||
9886 | ||||
9887 | bool OutputBecomesInput = false; | |||
9888 | ||||
9889 | llvm::SmallVector<unsigned, 8> Sizes; | |||
9890 | if (!RetType->isVoidType()) { | |||
9891 | Sizes.push_back(getAArch64LS(RetType, ParamKindTy::Vector, C)); | |||
9892 | if (!getAArch64PBV(RetType, C) && getAArch64MTV(RetType, {})) | |||
9893 | OutputBecomesInput = true; | |||
9894 | } | |||
9895 | for (unsigned I = 0, E = FD->getNumParams(); I < E; ++I) { | |||
9896 | QualType QT = FD->getParamDecl(I)->getType().getCanonicalType(); | |||
9897 | Sizes.push_back(getAArch64LS(QT, ParamAttrs[I].Kind, C)); | |||
9898 | } | |||
9899 | ||||
9900 | assert(!Sizes.empty() && "Unable to determine NDS and WDS.")((!Sizes.empty() && "Unable to determine NDS and WDS." ) ? static_cast<void> (0) : __assert_fail ("!Sizes.empty() && \"Unable to determine NDS and WDS.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9900, __PRETTY_FUNCTION__)); | |||
9901 | // The LS of a function parameter / return value can only be a power | |||
9902 | // of 2, starting from 8 bits, up to 128. | |||
9903 | assert(std::all_of(Sizes.begin(), Sizes.end(),((std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? static_cast< void> (0) : __assert_fail ("std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9908, __PRETTY_FUNCTION__)) | |||
9904 | [](unsigned Size) {((std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? static_cast< void> (0) : __assert_fail ("std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9908, __PRETTY_FUNCTION__)) | |||
9905 | return Size == 8 || Size == 16 || Size == 32 ||((std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? static_cast< void> (0) : __assert_fail ("std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9908, __PRETTY_FUNCTION__)) | |||
9906 | Size == 64 || Size == 128;((std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? static_cast< void> (0) : __assert_fail ("std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9908, __PRETTY_FUNCTION__)) | |||
9907 | }) &&((std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? static_cast< void> (0) : __assert_fail ("std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9908, __PRETTY_FUNCTION__)) | |||
9908 | "Invalid size")((std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && "Invalid size") ? static_cast< void> (0) : __assert_fail ("std::all_of(Sizes.begin(), Sizes.end(), [](unsigned Size) { return Size == 8 || Size == 16 || Size == 32 || Size == 64 || Size == 128; }) && \"Invalid size\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9908, __PRETTY_FUNCTION__)); | |||
9909 | ||||
9910 | return std::make_tuple(*std::min_element(std::begin(Sizes), std::end(Sizes)), | |||
9911 | *std::max_element(std::begin(Sizes), std::end(Sizes)), | |||
9912 | OutputBecomesInput); | |||
9913 | } | |||
9914 | ||||
9915 | /// Mangle the parameter part of the vector function name according to | |||
9916 | /// their OpenMP classification. The mangling function is defined in | |||
9917 | /// section 3.5 of the AAVFABI. | |||
9918 | static std::string mangleVectorParameters(ArrayRef<ParamAttrTy> ParamAttrs) { | |||
9919 | SmallString<256> Buffer; | |||
9920 | llvm::raw_svector_ostream Out(Buffer); | |||
9921 | for (const auto &ParamAttr : ParamAttrs) { | |||
9922 | switch (ParamAttr.Kind) { | |||
9923 | case LinearWithVarStride: | |||
9924 | Out << "ls" << ParamAttr.StrideOrArg; | |||
9925 | break; | |||
9926 | case Linear: | |||
9927 | Out << 'l'; | |||
9928 | // Don't print the step value if it is not present or if it is | |||
9929 | // equal to 1. | |||
9930 | if (!!ParamAttr.StrideOrArg && ParamAttr.StrideOrArg != 1) | |||
9931 | Out << ParamAttr.StrideOrArg; | |||
9932 | break; | |||
9933 | case Uniform: | |||
9934 | Out << 'u'; | |||
9935 | break; | |||
9936 | case Vector: | |||
9937 | Out << 'v'; | |||
9938 | break; | |||
9939 | } | |||
9940 | ||||
9941 | if (!!ParamAttr.Alignment) | |||
9942 | Out << 'a' << ParamAttr.Alignment; | |||
9943 | } | |||
9944 | ||||
9945 | return Out.str(); | |||
9946 | } | |||
9947 | ||||
9948 | // Function used to add the attribute. The parameter `VLEN` is | |||
9949 | // templated to allow the use of "x" when targeting scalable functions | |||
9950 | // for SVE. | |||
9951 | template <typename T> | |||
9952 | static void addAArch64VectorName(T VLEN, StringRef LMask, StringRef Prefix, | |||
9953 | char ISA, StringRef ParSeq, | |||
9954 | StringRef MangledName, bool OutputBecomesInput, | |||
9955 | llvm::Function *Fn) { | |||
9956 | SmallString<256> Buffer; | |||
9957 | llvm::raw_svector_ostream Out(Buffer); | |||
9958 | Out << Prefix << ISA << LMask << VLEN; | |||
9959 | if (OutputBecomesInput) | |||
9960 | Out << "v"; | |||
9961 | Out << ParSeq << "_" << MangledName; | |||
9962 | Fn->addFnAttr(Out.str()); | |||
9963 | } | |||
9964 | ||||
9965 | // Helper function to generate the Advanced SIMD names depending on | |||
9966 | // the value of the NDS when simdlen is not present. | |||
9967 | static void addAArch64AdvSIMDNDSNames(unsigned NDS, StringRef Mask, | |||
9968 | StringRef Prefix, char ISA, | |||
9969 | StringRef ParSeq, StringRef MangledName, | |||
9970 | bool OutputBecomesInput, | |||
9971 | llvm::Function *Fn) { | |||
9972 | switch (NDS) { | |||
9973 | case 8: | |||
9974 | addAArch64VectorName(8, Mask, Prefix, ISA, ParSeq, MangledName, | |||
9975 | OutputBecomesInput, Fn); | |||
9976 | addAArch64VectorName(16, Mask, Prefix, ISA, ParSeq, MangledName, | |||
9977 | OutputBecomesInput, Fn); | |||
9978 | break; | |||
9979 | case 16: | |||
9980 | addAArch64VectorName(4, Mask, Prefix, ISA, ParSeq, MangledName, | |||
9981 | OutputBecomesInput, Fn); | |||
9982 | addAArch64VectorName(8, Mask, Prefix, ISA, ParSeq, MangledName, | |||
9983 | OutputBecomesInput, Fn); | |||
9984 | break; | |||
9985 | case 32: | |||
9986 | addAArch64VectorName(2, Mask, Prefix, ISA, ParSeq, MangledName, | |||
9987 | OutputBecomesInput, Fn); | |||
9988 | addAArch64VectorName(4, Mask, Prefix, ISA, ParSeq, MangledName, | |||
9989 | OutputBecomesInput, Fn); | |||
9990 | break; | |||
9991 | case 64: | |||
9992 | case 128: | |||
9993 | addAArch64VectorName(2, Mask, Prefix, ISA, ParSeq, MangledName, | |||
9994 | OutputBecomesInput, Fn); | |||
9995 | break; | |||
9996 | default: | |||
9997 | llvm_unreachable("Scalar type is too wide.")::llvm::llvm_unreachable_internal("Scalar type is too wide.", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 9997); | |||
9998 | } | |||
9999 | } | |||
10000 | ||||
10001 | /// Emit vector function attributes for AArch64, as defined in the AAVFABI. | |||
10002 | static void emitAArch64DeclareSimdFunction( | |||
10003 | CodeGenModule &CGM, const FunctionDecl *FD, unsigned UserVLEN, | |||
10004 | ArrayRef<ParamAttrTy> ParamAttrs, | |||
10005 | OMPDeclareSimdDeclAttr::BranchStateTy State, StringRef MangledName, | |||
10006 | char ISA, unsigned VecRegSize, llvm::Function *Fn, SourceLocation SLoc) { | |||
10007 | ||||
10008 | // Get basic data for building the vector signature. | |||
10009 | const auto Data = getNDSWDS(FD, ParamAttrs); | |||
10010 | const unsigned NDS = std::get<0>(Data); | |||
10011 | const unsigned WDS = std::get<1>(Data); | |||
10012 | const bool OutputBecomesInput = std::get<2>(Data); | |||
10013 | ||||
10014 | // Check the values provided via `simdlen` by the user. | |||
10015 | // 1. A `simdlen(1)` doesn't produce vector signatures, | |||
10016 | if (UserVLEN == 1) { | |||
10017 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
10018 | DiagnosticsEngine::Warning, | |||
10019 | "The clause simdlen(1) has no effect when targeting aarch64."); | |||
10020 | CGM.getDiags().Report(SLoc, DiagID); | |||
10021 | return; | |||
10022 | } | |||
10023 | ||||
10024 | // 2. Section 3.3.1, item 1: user input must be a power of 2 for | |||
10025 | // Advanced SIMD output. | |||
10026 | if (ISA == 'n' && UserVLEN && !llvm::isPowerOf2_32(UserVLEN)) { | |||
10027 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
10028 | DiagnosticsEngine::Warning, "The value specified in simdlen must be a " | |||
10029 | "power of 2 when targeting Advanced SIMD."); | |||
10030 | CGM.getDiags().Report(SLoc, DiagID); | |||
10031 | return; | |||
10032 | } | |||
10033 | ||||
10034 | // 3. Section 3.4.1. SVE fixed lengh must obey the architectural | |||
10035 | // limits. | |||
10036 | if (ISA == 's' && UserVLEN != 0) { | |||
10037 | if ((UserVLEN * WDS > 2048) || (UserVLEN * WDS % 128 != 0)) { | |||
10038 | unsigned DiagID = CGM.getDiags().getCustomDiagID( | |||
10039 | DiagnosticsEngine::Warning, "The clause simdlen must fit the %0-bit " | |||
10040 | "lanes in the architectural constraints " | |||
10041 | "for SVE (min is 128-bit, max is " | |||
10042 | "2048-bit, by steps of 128-bit)"); | |||
10043 | CGM.getDiags().Report(SLoc, DiagID) << WDS; | |||
10044 | return; | |||
10045 | } | |||
10046 | } | |||
10047 | ||||
10048 | // Sort out parameter sequence. | |||
10049 | const std::string ParSeq = mangleVectorParameters(ParamAttrs); | |||
10050 | StringRef Prefix = "_ZGV"; | |||
10051 | // Generate simdlen from user input (if any). | |||
10052 | if (UserVLEN) { | |||
10053 | if (ISA == 's') { | |||
10054 | // SVE generates only a masked function. | |||
10055 | addAArch64VectorName(UserVLEN, "M", Prefix, ISA, ParSeq, MangledName, | |||
10056 | OutputBecomesInput, Fn); | |||
10057 | } else { | |||
10058 | assert(ISA == 'n' && "Expected ISA either 's' or 'n'.")((ISA == 'n' && "Expected ISA either 's' or 'n'.") ? static_cast <void> (0) : __assert_fail ("ISA == 'n' && \"Expected ISA either 's' or 'n'.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10058, __PRETTY_FUNCTION__)); | |||
10059 | // Advanced SIMD generates one or two functions, depending on | |||
10060 | // the `[not]inbranch` clause. | |||
10061 | switch (State) { | |||
10062 | case OMPDeclareSimdDeclAttr::BS_Undefined: | |||
10063 | addAArch64VectorName(UserVLEN, "N", Prefix, ISA, ParSeq, MangledName, | |||
10064 | OutputBecomesInput, Fn); | |||
10065 | addAArch64VectorName(UserVLEN, "M", Prefix, ISA, ParSeq, MangledName, | |||
10066 | OutputBecomesInput, Fn); | |||
10067 | break; | |||
10068 | case OMPDeclareSimdDeclAttr::BS_Notinbranch: | |||
10069 | addAArch64VectorName(UserVLEN, "N", Prefix, ISA, ParSeq, MangledName, | |||
10070 | OutputBecomesInput, Fn); | |||
10071 | break; | |||
10072 | case OMPDeclareSimdDeclAttr::BS_Inbranch: | |||
10073 | addAArch64VectorName(UserVLEN, "M", Prefix, ISA, ParSeq, MangledName, | |||
10074 | OutputBecomesInput, Fn); | |||
10075 | break; | |||
10076 | } | |||
10077 | } | |||
10078 | } else { | |||
10079 | // If no user simdlen is provided, follow the AAVFABI rules for | |||
10080 | // generating the vector length. | |||
10081 | if (ISA == 's') { | |||
10082 | // SVE, section 3.4.1, item 1. | |||
10083 | addAArch64VectorName("x", "M", Prefix, ISA, ParSeq, MangledName, | |||
10084 | OutputBecomesInput, Fn); | |||
10085 | } else { | |||
10086 | assert(ISA == 'n' && "Expected ISA either 's' or 'n'.")((ISA == 'n' && "Expected ISA either 's' or 'n'.") ? static_cast <void> (0) : __assert_fail ("ISA == 'n' && \"Expected ISA either 's' or 'n'.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10086, __PRETTY_FUNCTION__)); | |||
10087 | // Advanced SIMD, Section 3.3.1 of the AAVFABI, generates one or | |||
10088 | // two vector names depending on the use of the clause | |||
10089 | // `[not]inbranch`. | |||
10090 | switch (State) { | |||
10091 | case OMPDeclareSimdDeclAttr::BS_Undefined: | |||
10092 | addAArch64AdvSIMDNDSNames(NDS, "N", Prefix, ISA, ParSeq, MangledName, | |||
10093 | OutputBecomesInput, Fn); | |||
10094 | addAArch64AdvSIMDNDSNames(NDS, "M", Prefix, ISA, ParSeq, MangledName, | |||
10095 | OutputBecomesInput, Fn); | |||
10096 | break; | |||
10097 | case OMPDeclareSimdDeclAttr::BS_Notinbranch: | |||
10098 | addAArch64AdvSIMDNDSNames(NDS, "N", Prefix, ISA, ParSeq, MangledName, | |||
10099 | OutputBecomesInput, Fn); | |||
10100 | break; | |||
10101 | case OMPDeclareSimdDeclAttr::BS_Inbranch: | |||
10102 | addAArch64AdvSIMDNDSNames(NDS, "M", Prefix, ISA, ParSeq, MangledName, | |||
10103 | OutputBecomesInput, Fn); | |||
10104 | break; | |||
10105 | } | |||
10106 | } | |||
10107 | } | |||
10108 | } | |||
10109 | ||||
10110 | void CGOpenMPRuntime::emitDeclareSimdFunction(const FunctionDecl *FD, | |||
10111 | llvm::Function *Fn) { | |||
10112 | ASTContext &C = CGM.getContext(); | |||
10113 | FD = FD->getMostRecentDecl(); | |||
10114 | // Map params to their positions in function decl. | |||
10115 | llvm::DenseMap<const Decl *, unsigned> ParamPositions; | |||
10116 | if (isa<CXXMethodDecl>(FD)) | |||
10117 | ParamPositions.try_emplace(FD, 0); | |||
10118 | unsigned ParamPos = ParamPositions.size(); | |||
10119 | for (const ParmVarDecl *P : FD->parameters()) { | |||
10120 | ParamPositions.try_emplace(P->getCanonicalDecl(), ParamPos); | |||
10121 | ++ParamPos; | |||
10122 | } | |||
10123 | while (FD) { | |||
10124 | for (const auto *Attr : FD->specific_attrs<OMPDeclareSimdDeclAttr>()) { | |||
10125 | llvm::SmallVector<ParamAttrTy, 8> ParamAttrs(ParamPositions.size()); | |||
10126 | // Mark uniform parameters. | |||
10127 | for (const Expr *E : Attr->uniforms()) { | |||
10128 | E = E->IgnoreParenImpCasts(); | |||
10129 | unsigned Pos; | |||
10130 | if (isa<CXXThisExpr>(E)) { | |||
10131 | Pos = ParamPositions[FD]; | |||
10132 | } else { | |||
10133 | const auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl()) | |||
10134 | ->getCanonicalDecl(); | |||
10135 | Pos = ParamPositions[PVD]; | |||
10136 | } | |||
10137 | ParamAttrs[Pos].Kind = Uniform; | |||
10138 | } | |||
10139 | // Get alignment info. | |||
10140 | auto NI = Attr->alignments_begin(); | |||
10141 | for (const Expr *E : Attr->aligneds()) { | |||
10142 | E = E->IgnoreParenImpCasts(); | |||
10143 | unsigned Pos; | |||
10144 | QualType ParmTy; | |||
10145 | if (isa<CXXThisExpr>(E)) { | |||
10146 | Pos = ParamPositions[FD]; | |||
10147 | ParmTy = E->getType(); | |||
10148 | } else { | |||
10149 | const auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl()) | |||
10150 | ->getCanonicalDecl(); | |||
10151 | Pos = ParamPositions[PVD]; | |||
10152 | ParmTy = PVD->getType(); | |||
10153 | } | |||
10154 | ParamAttrs[Pos].Alignment = | |||
10155 | (*NI) | |||
10156 | ? (*NI)->EvaluateKnownConstInt(C) | |||
10157 | : llvm::APSInt::getUnsigned( | |||
10158 | C.toCharUnitsFromBits(C.getOpenMPDefaultSimdAlign(ParmTy)) | |||
10159 | .getQuantity()); | |||
10160 | ++NI; | |||
10161 | } | |||
10162 | // Mark linear parameters. | |||
10163 | auto SI = Attr->steps_begin(); | |||
10164 | auto MI = Attr->modifiers_begin(); | |||
10165 | for (const Expr *E : Attr->linears()) { | |||
10166 | E = E->IgnoreParenImpCasts(); | |||
10167 | unsigned Pos; | |||
10168 | if (isa<CXXThisExpr>(E)) { | |||
10169 | Pos = ParamPositions[FD]; | |||
10170 | } else { | |||
10171 | const auto *PVD = cast<ParmVarDecl>(cast<DeclRefExpr>(E)->getDecl()) | |||
10172 | ->getCanonicalDecl(); | |||
10173 | Pos = ParamPositions[PVD]; | |||
10174 | } | |||
10175 | ParamAttrTy &ParamAttr = ParamAttrs[Pos]; | |||
10176 | ParamAttr.Kind = Linear; | |||
10177 | if (*SI) { | |||
10178 | Expr::EvalResult Result; | |||
10179 | if (!(*SI)->EvaluateAsInt(Result, C, Expr::SE_AllowSideEffects)) { | |||
10180 | if (const auto *DRE = | |||
10181 | cast<DeclRefExpr>((*SI)->IgnoreParenImpCasts())) { | |||
10182 | if (const auto *StridePVD = cast<ParmVarDecl>(DRE->getDecl())) { | |||
10183 | ParamAttr.Kind = LinearWithVarStride; | |||
10184 | ParamAttr.StrideOrArg = llvm::APSInt::getUnsigned( | |||
10185 | ParamPositions[StridePVD->getCanonicalDecl()]); | |||
10186 | } | |||
10187 | } | |||
10188 | } else { | |||
10189 | ParamAttr.StrideOrArg = Result.Val.getInt(); | |||
10190 | } | |||
10191 | } | |||
10192 | ++SI; | |||
10193 | ++MI; | |||
10194 | } | |||
10195 | llvm::APSInt VLENVal; | |||
10196 | SourceLocation ExprLoc; | |||
10197 | const Expr *VLENExpr = Attr->getSimdlen(); | |||
10198 | if (VLENExpr) { | |||
10199 | VLENVal = VLENExpr->EvaluateKnownConstInt(C); | |||
10200 | ExprLoc = VLENExpr->getExprLoc(); | |||
10201 | } | |||
10202 | OMPDeclareSimdDeclAttr::BranchStateTy State = Attr->getBranchState(); | |||
10203 | if (CGM.getTriple().getArch() == llvm::Triple::x86 || | |||
10204 | CGM.getTriple().getArch() == llvm::Triple::x86_64) { | |||
10205 | emitX86DeclareSimdFunction(FD, Fn, VLENVal, ParamAttrs, State); | |||
10206 | } else if (CGM.getTriple().getArch() == llvm::Triple::aarch64) { | |||
10207 | unsigned VLEN = VLENVal.getExtValue(); | |||
10208 | StringRef MangledName = Fn->getName(); | |||
10209 | if (CGM.getTarget().hasFeature("sve")) | |||
10210 | emitAArch64DeclareSimdFunction(CGM, FD, VLEN, ParamAttrs, State, | |||
10211 | MangledName, 's', 128, Fn, ExprLoc); | |||
10212 | if (CGM.getTarget().hasFeature("neon")) | |||
10213 | emitAArch64DeclareSimdFunction(CGM, FD, VLEN, ParamAttrs, State, | |||
10214 | MangledName, 'n', 128, Fn, ExprLoc); | |||
10215 | } | |||
10216 | } | |||
10217 | FD = FD->getPreviousDecl(); | |||
10218 | } | |||
10219 | } | |||
10220 | ||||
10221 | namespace { | |||
10222 | /// Cleanup action for doacross support. | |||
10223 | class DoacrossCleanupTy final : public EHScopeStack::Cleanup { | |||
10224 | public: | |||
10225 | static const int DoacrossFinArgs = 2; | |||
10226 | ||||
10227 | private: | |||
10228 | llvm::FunctionCallee RTLFn; | |||
10229 | llvm::Value *Args[DoacrossFinArgs]; | |||
10230 | ||||
10231 | public: | |||
10232 | DoacrossCleanupTy(llvm::FunctionCallee RTLFn, | |||
10233 | ArrayRef<llvm::Value *> CallArgs) | |||
10234 | : RTLFn(RTLFn) { | |||
10235 | assert(CallArgs.size() == DoacrossFinArgs)((CallArgs.size() == DoacrossFinArgs) ? static_cast<void> (0) : __assert_fail ("CallArgs.size() == DoacrossFinArgs", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10235, __PRETTY_FUNCTION__)); | |||
10236 | std::copy(CallArgs.begin(), CallArgs.end(), std::begin(Args)); | |||
10237 | } | |||
10238 | void Emit(CodeGenFunction &CGF, Flags /*flags*/) override { | |||
10239 | if (!CGF.HaveInsertPoint()) | |||
10240 | return; | |||
10241 | CGF.EmitRuntimeCall(RTLFn, Args); | |||
10242 | } | |||
10243 | }; | |||
10244 | } // namespace | |||
10245 | ||||
10246 | void CGOpenMPRuntime::emitDoacrossInit(CodeGenFunction &CGF, | |||
10247 | const OMPLoopDirective &D, | |||
10248 | ArrayRef<Expr *> NumIterations) { | |||
10249 | if (!CGF.HaveInsertPoint()) | |||
10250 | return; | |||
10251 | ||||
10252 | ASTContext &C = CGM.getContext(); | |||
10253 | QualType Int64Ty = C.getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/true); | |||
10254 | RecordDecl *RD; | |||
10255 | if (KmpDimTy.isNull()) { | |||
10256 | // Build struct kmp_dim { // loop bounds info casted to kmp_int64 | |||
10257 | // kmp_int64 lo; // lower | |||
10258 | // kmp_int64 up; // upper | |||
10259 | // kmp_int64 st; // stride | |||
10260 | // }; | |||
10261 | RD = C.buildImplicitRecord("kmp_dim"); | |||
10262 | RD->startDefinition(); | |||
10263 | addFieldToRecordDecl(C, RD, Int64Ty); | |||
10264 | addFieldToRecordDecl(C, RD, Int64Ty); | |||
10265 | addFieldToRecordDecl(C, RD, Int64Ty); | |||
10266 | RD->completeDefinition(); | |||
10267 | KmpDimTy = C.getRecordType(RD); | |||
10268 | } else { | |||
10269 | RD = cast<RecordDecl>(KmpDimTy->getAsTagDecl()); | |||
10270 | } | |||
10271 | llvm::APInt Size(/*numBits=*/32, NumIterations.size()); | |||
10272 | QualType ArrayTy = | |||
10273 | C.getConstantArrayType(KmpDimTy, Size, ArrayType::Normal, 0); | |||
10274 | ||||
10275 | Address DimsAddr = CGF.CreateMemTemp(ArrayTy, "dims"); | |||
10276 | CGF.EmitNullInitialization(DimsAddr, ArrayTy); | |||
10277 | enum { LowerFD = 0, UpperFD, StrideFD }; | |||
10278 | // Fill dims with data. | |||
10279 | for (unsigned I = 0, E = NumIterations.size(); I < E; ++I) { | |||
10280 | LValue DimsLVal = CGF.MakeAddrLValue( | |||
10281 | CGF.Builder.CreateConstArrayGEP(DimsAddr, I), KmpDimTy); | |||
10282 | // dims.upper = num_iterations; | |||
10283 | LValue UpperLVal = CGF.EmitLValueForField( | |||
10284 | DimsLVal, *std::next(RD->field_begin(), UpperFD)); | |||
10285 | llvm::Value *NumIterVal = | |||
10286 | CGF.EmitScalarConversion(CGF.EmitScalarExpr(NumIterations[I]), | |||
10287 | D.getNumIterations()->getType(), Int64Ty, | |||
10288 | D.getNumIterations()->getExprLoc()); | |||
10289 | CGF.EmitStoreOfScalar(NumIterVal, UpperLVal); | |||
10290 | // dims.stride = 1; | |||
10291 | LValue StrideLVal = CGF.EmitLValueForField( | |||
10292 | DimsLVal, *std::next(RD->field_begin(), StrideFD)); | |||
10293 | CGF.EmitStoreOfScalar(llvm::ConstantInt::getSigned(CGM.Int64Ty, /*V=*/1), | |||
10294 | StrideLVal); | |||
10295 | } | |||
10296 | ||||
10297 | // Build call void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid, | |||
10298 | // kmp_int32 num_dims, struct kmp_dim * dims); | |||
10299 | llvm::Value *Args[] = { | |||
10300 | emitUpdateLocation(CGF, D.getBeginLoc()), | |||
10301 | getThreadID(CGF, D.getBeginLoc()), | |||
10302 | llvm::ConstantInt::getSigned(CGM.Int32Ty, NumIterations.size()), | |||
10303 | CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
10304 | CGF.Builder.CreateConstArrayGEP(DimsAddr, 0).getPointer(), | |||
10305 | CGM.VoidPtrTy)}; | |||
10306 | ||||
10307 | llvm::FunctionCallee RTLFn = | |||
10308 | createRuntimeFunction(OMPRTL__kmpc_doacross_init); | |||
10309 | CGF.EmitRuntimeCall(RTLFn, Args); | |||
10310 | llvm::Value *FiniArgs[DoacrossCleanupTy::DoacrossFinArgs] = { | |||
10311 | emitUpdateLocation(CGF, D.getEndLoc()), getThreadID(CGF, D.getEndLoc())}; | |||
10312 | llvm::FunctionCallee FiniRTLFn = | |||
10313 | createRuntimeFunction(OMPRTL__kmpc_doacross_fini); | |||
10314 | CGF.EHStack.pushCleanup<DoacrossCleanupTy>(NormalAndEHCleanup, FiniRTLFn, | |||
10315 | llvm::makeArrayRef(FiniArgs)); | |||
10316 | } | |||
10317 | ||||
10318 | void CGOpenMPRuntime::emitDoacrossOrdered(CodeGenFunction &CGF, | |||
10319 | const OMPDependClause *C) { | |||
10320 | QualType Int64Ty = | |||
10321 | CGM.getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1); | |||
10322 | llvm::APInt Size(/*numBits=*/32, C->getNumLoops()); | |||
10323 | QualType ArrayTy = CGM.getContext().getConstantArrayType( | |||
10324 | Int64Ty, Size, ArrayType::Normal, 0); | |||
10325 | Address CntAddr = CGF.CreateMemTemp(ArrayTy, ".cnt.addr"); | |||
10326 | for (unsigned I = 0, E = C->getNumLoops(); I < E; ++I) { | |||
10327 | const Expr *CounterVal = C->getLoopData(I); | |||
10328 | assert(CounterVal)((CounterVal) ? static_cast<void> (0) : __assert_fail ( "CounterVal", "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10328, __PRETTY_FUNCTION__)); | |||
10329 | llvm::Value *CntVal = CGF.EmitScalarConversion( | |||
10330 | CGF.EmitScalarExpr(CounterVal), CounterVal->getType(), Int64Ty, | |||
10331 | CounterVal->getExprLoc()); | |||
10332 | CGF.EmitStoreOfScalar(CntVal, CGF.Builder.CreateConstArrayGEP(CntAddr, I), | |||
10333 | /*Volatile=*/false, Int64Ty); | |||
10334 | } | |||
10335 | llvm::Value *Args[] = { | |||
10336 | emitUpdateLocation(CGF, C->getBeginLoc()), | |||
10337 | getThreadID(CGF, C->getBeginLoc()), | |||
10338 | CGF.Builder.CreateConstArrayGEP(CntAddr, 0).getPointer()}; | |||
10339 | llvm::FunctionCallee RTLFn; | |||
10340 | if (C->getDependencyKind() == OMPC_DEPEND_source) { | |||
10341 | RTLFn = createRuntimeFunction(OMPRTL__kmpc_doacross_post); | |||
10342 | } else { | |||
10343 | assert(C->getDependencyKind() == OMPC_DEPEND_sink)((C->getDependencyKind() == OMPC_DEPEND_sink) ? static_cast <void> (0) : __assert_fail ("C->getDependencyKind() == OMPC_DEPEND_sink" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10343, __PRETTY_FUNCTION__)); | |||
10344 | RTLFn = createRuntimeFunction(OMPRTL__kmpc_doacross_wait); | |||
10345 | } | |||
10346 | CGF.EmitRuntimeCall(RTLFn, Args); | |||
10347 | } | |||
10348 | ||||
10349 | void CGOpenMPRuntime::emitCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
10350 | llvm::FunctionCallee Callee, | |||
10351 | ArrayRef<llvm::Value *> Args) const { | |||
10352 | assert(Loc.isValid() && "Outlined function call location must be valid.")((Loc.isValid() && "Outlined function call location must be valid." ) ? static_cast<void> (0) : __assert_fail ("Loc.isValid() && \"Outlined function call location must be valid.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10352, __PRETTY_FUNCTION__)); | |||
10353 | auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF, Loc); | |||
10354 | ||||
10355 | if (auto *Fn = dyn_cast<llvm::Function>(Callee.getCallee())) { | |||
10356 | if (Fn->doesNotThrow()) { | |||
10357 | CGF.EmitNounwindRuntimeCall(Fn, Args); | |||
10358 | return; | |||
10359 | } | |||
10360 | } | |||
10361 | CGF.EmitRuntimeCall(Callee, Args); | |||
10362 | } | |||
10363 | ||||
10364 | void CGOpenMPRuntime::emitOutlinedFunctionCall( | |||
10365 | CodeGenFunction &CGF, SourceLocation Loc, llvm::FunctionCallee OutlinedFn, | |||
10366 | ArrayRef<llvm::Value *> Args) const { | |||
10367 | emitCall(CGF, Loc, OutlinedFn, Args); | |||
10368 | } | |||
10369 | ||||
10370 | void CGOpenMPRuntime::emitFunctionProlog(CodeGenFunction &CGF, const Decl *D) { | |||
10371 | if (const auto *FD = dyn_cast<FunctionDecl>(D)) | |||
10372 | if (OMPDeclareTargetDeclAttr::isDeclareTargetDeclaration(FD)) | |||
10373 | HasEmittedDeclareTargetRegion = true; | |||
10374 | } | |||
10375 | ||||
10376 | Address CGOpenMPRuntime::getParameterAddress(CodeGenFunction &CGF, | |||
10377 | const VarDecl *NativeParam, | |||
10378 | const VarDecl *TargetParam) const { | |||
10379 | return CGF.GetAddrOfLocalVar(NativeParam); | |||
10380 | } | |||
10381 | ||||
10382 | namespace { | |||
10383 | /// Cleanup action for allocate support. | |||
10384 | class OMPAllocateCleanupTy final : public EHScopeStack::Cleanup { | |||
10385 | public: | |||
10386 | static const int CleanupArgs = 3; | |||
10387 | ||||
10388 | private: | |||
10389 | llvm::FunctionCallee RTLFn; | |||
10390 | llvm::Value *Args[CleanupArgs]; | |||
10391 | ||||
10392 | public: | |||
10393 | OMPAllocateCleanupTy(llvm::FunctionCallee RTLFn, | |||
10394 | ArrayRef<llvm::Value *> CallArgs) | |||
10395 | : RTLFn(RTLFn) { | |||
10396 | assert(CallArgs.size() == CleanupArgs &&((CallArgs.size() == CleanupArgs && "Size of arguments does not match." ) ? static_cast<void> (0) : __assert_fail ("CallArgs.size() == CleanupArgs && \"Size of arguments does not match.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10397, __PRETTY_FUNCTION__)) | |||
10397 | "Size of arguments does not match.")((CallArgs.size() == CleanupArgs && "Size of arguments does not match." ) ? static_cast<void> (0) : __assert_fail ("CallArgs.size() == CleanupArgs && \"Size of arguments does not match.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10397, __PRETTY_FUNCTION__)); | |||
10398 | std::copy(CallArgs.begin(), CallArgs.end(), std::begin(Args)); | |||
10399 | } | |||
10400 | void Emit(CodeGenFunction &CGF, Flags /*flags*/) override { | |||
10401 | if (!CGF.HaveInsertPoint()) | |||
10402 | return; | |||
10403 | CGF.EmitRuntimeCall(RTLFn, Args); | |||
10404 | } | |||
10405 | }; | |||
10406 | } // namespace | |||
10407 | ||||
10408 | Address CGOpenMPRuntime::getAddressOfLocalVariable(CodeGenFunction &CGF, | |||
10409 | const VarDecl *VD) { | |||
10410 | if (!VD) | |||
10411 | return Address::invalid(); | |||
10412 | const VarDecl *CVD = VD->getCanonicalDecl(); | |||
10413 | if (!CVD->hasAttr<OMPAllocateDeclAttr>()) | |||
10414 | return Address::invalid(); | |||
10415 | const auto *AA = CVD->getAttr<OMPAllocateDeclAttr>(); | |||
10416 | // Use the default allocation. | |||
10417 | if (AA->getAllocatorType() == OMPAllocateDeclAttr::OMPDefaultMemAlloc && | |||
10418 | !AA->getAllocator()) | |||
10419 | return Address::invalid(); | |||
10420 | llvm::Value *Size; | |||
10421 | CharUnits Align = CGM.getContext().getDeclAlign(CVD); | |||
10422 | if (CVD->getType()->isVariablyModifiedType()) { | |||
10423 | Size = CGF.getTypeSize(CVD->getType()); | |||
10424 | // Align the size: ((size + align - 1) / align) * align | |||
10425 | Size = CGF.Builder.CreateNUWAdd( | |||
10426 | Size, CGM.getSize(Align - CharUnits::fromQuantity(1))); | |||
10427 | Size = CGF.Builder.CreateUDiv(Size, CGM.getSize(Align)); | |||
10428 | Size = CGF.Builder.CreateNUWMul(Size, CGM.getSize(Align)); | |||
10429 | } else { | |||
10430 | CharUnits Sz = CGM.getContext().getTypeSizeInChars(CVD->getType()); | |||
10431 | Size = CGM.getSize(Sz.alignTo(Align)); | |||
10432 | } | |||
10433 | llvm::Value *ThreadID = getThreadID(CGF, CVD->getBeginLoc()); | |||
10434 | assert(AA->getAllocator() &&((AA->getAllocator() && "Expected allocator expression for non-default allocator." ) ? static_cast<void> (0) : __assert_fail ("AA->getAllocator() && \"Expected allocator expression for non-default allocator.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10435, __PRETTY_FUNCTION__)) | |||
10435 | "Expected allocator expression for non-default allocator.")((AA->getAllocator() && "Expected allocator expression for non-default allocator." ) ? static_cast<void> (0) : __assert_fail ("AA->getAllocator() && \"Expected allocator expression for non-default allocator.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10435, __PRETTY_FUNCTION__)); | |||
10436 | llvm::Value *Allocator = CGF.EmitScalarExpr(AA->getAllocator()); | |||
10437 | // According to the standard, the original allocator type is a enum (integer). | |||
10438 | // Convert to pointer type, if required. | |||
10439 | if (Allocator->getType()->isIntegerTy()) | |||
10440 | Allocator = CGF.Builder.CreateIntToPtr(Allocator, CGM.VoidPtrTy); | |||
10441 | else if (Allocator->getType()->isPointerTy()) | |||
10442 | Allocator = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(Allocator, | |||
10443 | CGM.VoidPtrTy); | |||
10444 | llvm::Value *Args[] = {ThreadID, Size, Allocator}; | |||
10445 | ||||
10446 | llvm::Value *Addr = | |||
10447 | CGF.EmitRuntimeCall(createRuntimeFunction(OMPRTL__kmpc_alloc), Args, | |||
10448 | CVD->getName() + ".void.addr"); | |||
10449 | llvm::Value *FiniArgs[OMPAllocateCleanupTy::CleanupArgs] = {ThreadID, Addr, | |||
10450 | Allocator}; | |||
10451 | llvm::FunctionCallee FiniRTLFn = createRuntimeFunction(OMPRTL__kmpc_free); | |||
10452 | ||||
10453 | CGF.EHStack.pushCleanup<OMPAllocateCleanupTy>(NormalAndEHCleanup, FiniRTLFn, | |||
10454 | llvm::makeArrayRef(FiniArgs)); | |||
10455 | Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( | |||
10456 | Addr, | |||
10457 | CGF.ConvertTypeForMem(CGM.getContext().getPointerType(CVD->getType())), | |||
10458 | CVD->getName() + ".addr"); | |||
10459 | return Address(Addr, Align); | |||
10460 | } | |||
10461 | ||||
10462 | llvm::Function *CGOpenMPSIMDRuntime::emitParallelOutlinedFunction( | |||
10463 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
10464 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | |||
10465 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10465); | |||
10466 | } | |||
10467 | ||||
10468 | llvm::Function *CGOpenMPSIMDRuntime::emitTeamsOutlinedFunction( | |||
10469 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
10470 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen) { | |||
10471 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10471); | |||
10472 | } | |||
10473 | ||||
10474 | llvm::Function *CGOpenMPSIMDRuntime::emitTaskOutlinedFunction( | |||
10475 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, | |||
10476 | const VarDecl *PartIDVar, const VarDecl *TaskTVar, | |||
10477 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen, | |||
10478 | bool Tied, unsigned &NumberOfParts) { | |||
10479 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10479); | |||
10480 | } | |||
10481 | ||||
10482 | void CGOpenMPSIMDRuntime::emitParallelCall(CodeGenFunction &CGF, | |||
10483 | SourceLocation Loc, | |||
10484 | llvm::Function *OutlinedFn, | |||
10485 | ArrayRef<llvm::Value *> CapturedVars, | |||
10486 | const Expr *IfCond) { | |||
10487 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10487); | |||
10488 | } | |||
10489 | ||||
10490 | void CGOpenMPSIMDRuntime::emitCriticalRegion( | |||
10491 | CodeGenFunction &CGF, StringRef CriticalName, | |||
10492 | const RegionCodeGenTy &CriticalOpGen, SourceLocation Loc, | |||
10493 | const Expr *Hint) { | |||
10494 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10494); | |||
10495 | } | |||
10496 | ||||
10497 | void CGOpenMPSIMDRuntime::emitMasterRegion(CodeGenFunction &CGF, | |||
10498 | const RegionCodeGenTy &MasterOpGen, | |||
10499 | SourceLocation Loc) { | |||
10500 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10500); | |||
10501 | } | |||
10502 | ||||
10503 | void CGOpenMPSIMDRuntime::emitTaskyieldCall(CodeGenFunction &CGF, | |||
10504 | SourceLocation Loc) { | |||
10505 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10505); | |||
10506 | } | |||
10507 | ||||
10508 | void CGOpenMPSIMDRuntime::emitTaskgroupRegion( | |||
10509 | CodeGenFunction &CGF, const RegionCodeGenTy &TaskgroupOpGen, | |||
10510 | SourceLocation Loc) { | |||
10511 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10511); | |||
10512 | } | |||
10513 | ||||
10514 | void CGOpenMPSIMDRuntime::emitSingleRegion( | |||
10515 | CodeGenFunction &CGF, const RegionCodeGenTy &SingleOpGen, | |||
10516 | SourceLocation Loc, ArrayRef<const Expr *> CopyprivateVars, | |||
10517 | ArrayRef<const Expr *> DestExprs, ArrayRef<const Expr *> SrcExprs, | |||
10518 | ArrayRef<const Expr *> AssignmentOps) { | |||
10519 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10519); | |||
10520 | } | |||
10521 | ||||
10522 | void CGOpenMPSIMDRuntime::emitOrderedRegion(CodeGenFunction &CGF, | |||
10523 | const RegionCodeGenTy &OrderedOpGen, | |||
10524 | SourceLocation Loc, | |||
10525 | bool IsThreads) { | |||
10526 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10526); | |||
10527 | } | |||
10528 | ||||
10529 | void CGOpenMPSIMDRuntime::emitBarrierCall(CodeGenFunction &CGF, | |||
10530 | SourceLocation Loc, | |||
10531 | OpenMPDirectiveKind Kind, | |||
10532 | bool EmitChecks, | |||
10533 | bool ForceSimpleCall) { | |||
10534 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10534); | |||
10535 | } | |||
10536 | ||||
10537 | void CGOpenMPSIMDRuntime::emitForDispatchInit( | |||
10538 | CodeGenFunction &CGF, SourceLocation Loc, | |||
10539 | const OpenMPScheduleTy &ScheduleKind, unsigned IVSize, bool IVSigned, | |||
10540 | bool Ordered, const DispatchRTInput &DispatchValues) { | |||
10541 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10541); | |||
10542 | } | |||
10543 | ||||
10544 | void CGOpenMPSIMDRuntime::emitForStaticInit( | |||
10545 | CodeGenFunction &CGF, SourceLocation Loc, OpenMPDirectiveKind DKind, | |||
10546 | const OpenMPScheduleTy &ScheduleKind, const StaticRTInput &Values) { | |||
10547 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10547); | |||
10548 | } | |||
10549 | ||||
10550 | void CGOpenMPSIMDRuntime::emitDistributeStaticInit( | |||
10551 | CodeGenFunction &CGF, SourceLocation Loc, | |||
10552 | OpenMPDistScheduleClauseKind SchedKind, const StaticRTInput &Values) { | |||
10553 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10553); | |||
10554 | } | |||
10555 | ||||
10556 | void CGOpenMPSIMDRuntime::emitForOrderedIterationEnd(CodeGenFunction &CGF, | |||
10557 | SourceLocation Loc, | |||
10558 | unsigned IVSize, | |||
10559 | bool IVSigned) { | |||
10560 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10560); | |||
10561 | } | |||
10562 | ||||
10563 | void CGOpenMPSIMDRuntime::emitForStaticFinish(CodeGenFunction &CGF, | |||
10564 | SourceLocation Loc, | |||
10565 | OpenMPDirectiveKind DKind) { | |||
10566 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10566); | |||
10567 | } | |||
10568 | ||||
10569 | llvm::Value *CGOpenMPSIMDRuntime::emitForNext(CodeGenFunction &CGF, | |||
10570 | SourceLocation Loc, | |||
10571 | unsigned IVSize, bool IVSigned, | |||
10572 | Address IL, Address LB, | |||
10573 | Address UB, Address ST) { | |||
10574 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10574); | |||
10575 | } | |||
10576 | ||||
10577 | void CGOpenMPSIMDRuntime::emitNumThreadsClause(CodeGenFunction &CGF, | |||
10578 | llvm::Value *NumThreads, | |||
10579 | SourceLocation Loc) { | |||
10580 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10580); | |||
10581 | } | |||
10582 | ||||
10583 | void CGOpenMPSIMDRuntime::emitProcBindClause(CodeGenFunction &CGF, | |||
10584 | OpenMPProcBindClauseKind ProcBind, | |||
10585 | SourceLocation Loc) { | |||
10586 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10586); | |||
10587 | } | |||
10588 | ||||
10589 | Address CGOpenMPSIMDRuntime::getAddrOfThreadPrivate(CodeGenFunction &CGF, | |||
10590 | const VarDecl *VD, | |||
10591 | Address VDAddr, | |||
10592 | SourceLocation Loc) { | |||
10593 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10593); | |||
10594 | } | |||
10595 | ||||
10596 | llvm::Function *CGOpenMPSIMDRuntime::emitThreadPrivateVarDefinition( | |||
10597 | const VarDecl *VD, Address VDAddr, SourceLocation Loc, bool PerformInit, | |||
10598 | CodeGenFunction *CGF) { | |||
10599 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10599); | |||
10600 | } | |||
10601 | ||||
10602 | Address CGOpenMPSIMDRuntime::getAddrOfArtificialThreadPrivate( | |||
10603 | CodeGenFunction &CGF, QualType VarType, StringRef Name) { | |||
10604 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10604); | |||
10605 | } | |||
10606 | ||||
10607 | void CGOpenMPSIMDRuntime::emitFlush(CodeGenFunction &CGF, | |||
10608 | ArrayRef<const Expr *> Vars, | |||
10609 | SourceLocation Loc) { | |||
10610 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10610); | |||
10611 | } | |||
10612 | ||||
10613 | void CGOpenMPSIMDRuntime::emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc, | |||
10614 | const OMPExecutableDirective &D, | |||
10615 | llvm::Function *TaskFunction, | |||
10616 | QualType SharedsTy, Address Shareds, | |||
10617 | const Expr *IfCond, | |||
10618 | const OMPTaskDataTy &Data) { | |||
10619 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10619); | |||
10620 | } | |||
10621 | ||||
10622 | void CGOpenMPSIMDRuntime::emitTaskLoopCall( | |||
10623 | CodeGenFunction &CGF, SourceLocation Loc, const OMPLoopDirective &D, | |||
10624 | llvm::Function *TaskFunction, QualType SharedsTy, Address Shareds, | |||
10625 | const Expr *IfCond, const OMPTaskDataTy &Data) { | |||
10626 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10626); | |||
10627 | } | |||
10628 | ||||
10629 | void CGOpenMPSIMDRuntime::emitReduction( | |||
10630 | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> Privates, | |||
10631 | ArrayRef<const Expr *> LHSExprs, ArrayRef<const Expr *> RHSExprs, | |||
10632 | ArrayRef<const Expr *> ReductionOps, ReductionOptionsTy Options) { | |||
10633 | assert(Options.SimpleReduction && "Only simple reduction is expected.")((Options.SimpleReduction && "Only simple reduction is expected." ) ? static_cast<void> (0) : __assert_fail ("Options.SimpleReduction && \"Only simple reduction is expected.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10633, __PRETTY_FUNCTION__)); | |||
10634 | CGOpenMPRuntime::emitReduction(CGF, Loc, Privates, LHSExprs, RHSExprs, | |||
10635 | ReductionOps, Options); | |||
10636 | } | |||
10637 | ||||
10638 | llvm::Value *CGOpenMPSIMDRuntime::emitTaskReductionInit( | |||
10639 | CodeGenFunction &CGF, SourceLocation Loc, ArrayRef<const Expr *> LHSExprs, | |||
10640 | ArrayRef<const Expr *> RHSExprs, const OMPTaskDataTy &Data) { | |||
10641 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10641); | |||
10642 | } | |||
10643 | ||||
10644 | void CGOpenMPSIMDRuntime::emitTaskReductionFixups(CodeGenFunction &CGF, | |||
10645 | SourceLocation Loc, | |||
10646 | ReductionCodeGen &RCG, | |||
10647 | unsigned N) { | |||
10648 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10648); | |||
10649 | } | |||
10650 | ||||
10651 | Address CGOpenMPSIMDRuntime::getTaskReductionItem(CodeGenFunction &CGF, | |||
10652 | SourceLocation Loc, | |||
10653 | llvm::Value *ReductionsPtr, | |||
10654 | LValue SharedLVal) { | |||
10655 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10655); | |||
10656 | } | |||
10657 | ||||
10658 | void CGOpenMPSIMDRuntime::emitTaskwaitCall(CodeGenFunction &CGF, | |||
10659 | SourceLocation Loc) { | |||
10660 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10660); | |||
10661 | } | |||
10662 | ||||
10663 | void CGOpenMPSIMDRuntime::emitCancellationPointCall( | |||
10664 | CodeGenFunction &CGF, SourceLocation Loc, | |||
10665 | OpenMPDirectiveKind CancelRegion) { | |||
10666 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10666); | |||
10667 | } | |||
10668 | ||||
10669 | void CGOpenMPSIMDRuntime::emitCancelCall(CodeGenFunction &CGF, | |||
10670 | SourceLocation Loc, const Expr *IfCond, | |||
10671 | OpenMPDirectiveKind CancelRegion) { | |||
10672 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10672); | |||
10673 | } | |||
10674 | ||||
10675 | void CGOpenMPSIMDRuntime::emitTargetOutlinedFunction( | |||
10676 | const OMPExecutableDirective &D, StringRef ParentName, | |||
10677 | llvm::Function *&OutlinedFn, llvm::Constant *&OutlinedFnID, | |||
10678 | bool IsOffloadEntry, const RegionCodeGenTy &CodeGen) { | |||
10679 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10679); | |||
10680 | } | |||
10681 | ||||
10682 | void CGOpenMPSIMDRuntime::emitTargetCall(CodeGenFunction &CGF, | |||
10683 | const OMPExecutableDirective &D, | |||
10684 | llvm::Function *OutlinedFn, | |||
10685 | llvm::Value *OutlinedFnID, | |||
10686 | const Expr *IfCond, | |||
10687 | const Expr *Device) { | |||
10688 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10688); | |||
10689 | } | |||
10690 | ||||
10691 | bool CGOpenMPSIMDRuntime::emitTargetFunctions(GlobalDecl GD) { | |||
10692 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10692); | |||
10693 | } | |||
10694 | ||||
10695 | bool CGOpenMPSIMDRuntime::emitTargetGlobalVariable(GlobalDecl GD) { | |||
10696 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10696); | |||
10697 | } | |||
10698 | ||||
10699 | bool CGOpenMPSIMDRuntime::emitTargetGlobal(GlobalDecl GD) { | |||
10700 | return false; | |||
10701 | } | |||
10702 | ||||
10703 | llvm::Function *CGOpenMPSIMDRuntime::emitRegistrationFunction() { | |||
10704 | return nullptr; | |||
10705 | } | |||
10706 | ||||
10707 | void CGOpenMPSIMDRuntime::emitTeamsCall(CodeGenFunction &CGF, | |||
10708 | const OMPExecutableDirective &D, | |||
10709 | SourceLocation Loc, | |||
10710 | llvm::Function *OutlinedFn, | |||
10711 | ArrayRef<llvm::Value *> CapturedVars) { | |||
10712 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10712); | |||
10713 | } | |||
10714 | ||||
10715 | void CGOpenMPSIMDRuntime::emitNumTeamsClause(CodeGenFunction &CGF, | |||
10716 | const Expr *NumTeams, | |||
10717 | const Expr *ThreadLimit, | |||
10718 | SourceLocation Loc) { | |||
10719 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10719); | |||
10720 | } | |||
10721 | ||||
10722 | void CGOpenMPSIMDRuntime::emitTargetDataCalls( | |||
10723 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond, | |||
10724 | const Expr *Device, const RegionCodeGenTy &CodeGen, TargetDataInfo &Info) { | |||
10725 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10725); | |||
10726 | } | |||
10727 | ||||
10728 | void CGOpenMPSIMDRuntime::emitTargetDataStandAloneCall( | |||
10729 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *IfCond, | |||
10730 | const Expr *Device) { | |||
10731 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10731); | |||
10732 | } | |||
10733 | ||||
10734 | void CGOpenMPSIMDRuntime::emitDoacrossInit(CodeGenFunction &CGF, | |||
10735 | const OMPLoopDirective &D, | |||
10736 | ArrayRef<Expr *> NumIterations) { | |||
10737 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10737); | |||
10738 | } | |||
10739 | ||||
10740 | void CGOpenMPSIMDRuntime::emitDoacrossOrdered(CodeGenFunction &CGF, | |||
10741 | const OMPDependClause *C) { | |||
10742 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10742); | |||
10743 | } | |||
10744 | ||||
10745 | const VarDecl * | |||
10746 | CGOpenMPSIMDRuntime::translateParameter(const FieldDecl *FD, | |||
10747 | const VarDecl *NativeParam) const { | |||
10748 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10748); | |||
10749 | } | |||
10750 | ||||
10751 | Address | |||
10752 | CGOpenMPSIMDRuntime::getParameterAddress(CodeGenFunction &CGF, | |||
10753 | const VarDecl *NativeParam, | |||
10754 | const VarDecl *TargetParam) const { | |||
10755 | llvm_unreachable("Not supported in SIMD-only mode")::llvm::llvm_unreachable_internal("Not supported in SIMD-only mode" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.cpp" , 10755); | |||
10756 | } |
1 | //===----- CGOpenMPRuntime.h - Interface to OpenMP Runtimes -----*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This provides a class for OpenMP runtime code generation. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #ifndef LLVM_CLANG_LIB_CODEGEN_CGOPENMPRUNTIME_H |
14 | #define LLVM_CLANG_LIB_CODEGEN_CGOPENMPRUNTIME_H |
15 | |
16 | #include "CGValue.h" |
17 | #include "clang/AST/DeclOpenMP.h" |
18 | #include "clang/AST/Type.h" |
19 | #include "clang/Basic/OpenMPKinds.h" |
20 | #include "clang/Basic/SourceLocation.h" |
21 | #include "llvm/ADT/DenseMap.h" |
22 | #include "llvm/ADT/StringMap.h" |
23 | #include "llvm/ADT/StringSet.h" |
24 | #include "llvm/IR/Function.h" |
25 | #include "llvm/IR/ValueHandle.h" |
26 | |
27 | namespace llvm { |
28 | class ArrayType; |
29 | class Constant; |
30 | class FunctionType; |
31 | class GlobalVariable; |
32 | class StructType; |
33 | class Type; |
34 | class Value; |
35 | } // namespace llvm |
36 | |
37 | namespace clang { |
38 | class Expr; |
39 | class GlobalDecl; |
40 | class OMPDependClause; |
41 | class OMPExecutableDirective; |
42 | class OMPLoopDirective; |
43 | class VarDecl; |
44 | class OMPDeclareReductionDecl; |
45 | class IdentifierInfo; |
46 | |
47 | namespace CodeGen { |
48 | class Address; |
49 | class CodeGenFunction; |
50 | class CodeGenModule; |
51 | |
52 | /// A basic class for pre|post-action for advanced codegen sequence for OpenMP |
53 | /// region. |
54 | class PrePostActionTy { |
55 | public: |
56 | explicit PrePostActionTy() {} |
57 | virtual void Enter(CodeGenFunction &CGF) {} |
58 | virtual void Exit(CodeGenFunction &CGF) {} |
59 | virtual ~PrePostActionTy() {} |
60 | }; |
61 | |
62 | /// Class provides a way to call simple version of codegen for OpenMP region, or |
63 | /// an advanced with possible pre|post-actions in codegen. |
64 | class RegionCodeGenTy final { |
65 | intptr_t CodeGen; |
66 | typedef void (*CodeGenTy)(intptr_t, CodeGenFunction &, PrePostActionTy &); |
67 | CodeGenTy Callback; |
68 | mutable PrePostActionTy *PrePostAction; |
69 | RegionCodeGenTy() = delete; |
70 | RegionCodeGenTy &operator=(const RegionCodeGenTy &) = delete; |
71 | template <typename Callable> |
72 | static void CallbackFn(intptr_t CodeGen, CodeGenFunction &CGF, |
73 | PrePostActionTy &Action) { |
74 | return (*reinterpret_cast<Callable *>(CodeGen))(CGF, Action); |
75 | } |
76 | |
77 | public: |
78 | template <typename Callable> |
79 | RegionCodeGenTy( |
80 | Callable &&CodeGen, |
81 | typename std::enable_if< |
82 | !std::is_same<typename std::remove_reference<Callable>::type, |
83 | RegionCodeGenTy>::value>::type * = nullptr) |
84 | : CodeGen(reinterpret_cast<intptr_t>(&CodeGen)), |
85 | Callback(CallbackFn<typename std::remove_reference<Callable>::type>), |
86 | PrePostAction(nullptr) {} |
87 | void setAction(PrePostActionTy &Action) const { PrePostAction = &Action; } |
88 | void operator()(CodeGenFunction &CGF) const; |
89 | }; |
90 | |
91 | struct OMPTaskDataTy final { |
92 | SmallVector<const Expr *, 4> PrivateVars; |
93 | SmallVector<const Expr *, 4> PrivateCopies; |
94 | SmallVector<const Expr *, 4> FirstprivateVars; |
95 | SmallVector<const Expr *, 4> FirstprivateCopies; |
96 | SmallVector<const Expr *, 4> FirstprivateInits; |
97 | SmallVector<const Expr *, 4> LastprivateVars; |
98 | SmallVector<const Expr *, 4> LastprivateCopies; |
99 | SmallVector<const Expr *, 4> ReductionVars; |
100 | SmallVector<const Expr *, 4> ReductionCopies; |
101 | SmallVector<const Expr *, 4> ReductionOps; |
102 | SmallVector<std::pair<OpenMPDependClauseKind, const Expr *>, 4> Dependences; |
103 | llvm::PointerIntPair<llvm::Value *, 1, bool> Final; |
104 | llvm::PointerIntPair<llvm::Value *, 1, bool> Schedule; |
105 | llvm::PointerIntPair<llvm::Value *, 1, bool> Priority; |
106 | llvm::Value *Reductions = nullptr; |
107 | unsigned NumberOfParts = 0; |
108 | bool Tied = true; |
109 | bool Nogroup = false; |
110 | }; |
111 | |
112 | /// Class intended to support codegen of all kind of the reduction clauses. |
113 | class ReductionCodeGen { |
114 | private: |
115 | /// Data required for codegen of reduction clauses. |
116 | struct ReductionData { |
117 | /// Reference to the original shared item. |
118 | const Expr *Ref = nullptr; |
119 | /// Helper expression for generation of private copy. |
120 | const Expr *Private = nullptr; |
121 | /// Helper expression for generation reduction operation. |
122 | const Expr *ReductionOp = nullptr; |
123 | ReductionData(const Expr *Ref, const Expr *Private, const Expr *ReductionOp) |
124 | : Ref(Ref), Private(Private), ReductionOp(ReductionOp) {} |
125 | }; |
126 | /// List of reduction-based clauses. |
127 | SmallVector<ReductionData, 4> ClausesData; |
128 | |
129 | /// List of addresses of original shared variables/expressions. |
130 | SmallVector<std::pair<LValue, LValue>, 4> SharedAddresses; |
131 | /// Sizes of the reduction items in chars. |
132 | SmallVector<std::pair<llvm::Value *, llvm::Value *>, 4> Sizes; |
133 | /// Base declarations for the reduction items. |
134 | SmallVector<const VarDecl *, 4> BaseDecls; |
135 | |
136 | /// Emits lvalue for shared expression. |
137 | LValue emitSharedLValue(CodeGenFunction &CGF, const Expr *E); |
138 | /// Emits upper bound for shared expression (if array section). |
139 | LValue emitSharedLValueUB(CodeGenFunction &CGF, const Expr *E); |
140 | /// Performs aggregate initialization. |
141 | /// \param N Number of reduction item in the common list. |
142 | /// \param PrivateAddr Address of the corresponding private item. |
143 | /// \param SharedLVal Address of the original shared variable. |
144 | /// \param DRD Declare reduction construct used for reduction item. |
145 | void emitAggregateInitialization(CodeGenFunction &CGF, unsigned N, |
146 | Address PrivateAddr, LValue SharedLVal, |
147 | const OMPDeclareReductionDecl *DRD); |
148 | |
149 | public: |
150 | ReductionCodeGen(ArrayRef<const Expr *> Shareds, |
151 | ArrayRef<const Expr *> Privates, |
152 | ArrayRef<const Expr *> ReductionOps); |
153 | /// Emits lvalue for a reduction item. |
154 | /// \param N Number of the reduction item. |
155 | void emitSharedLValue(CodeGenFunction &CGF, unsigned N); |
156 | /// Emits the code for the variable-modified type, if required. |
157 | /// \param N Number of the reduction item. |
158 | void emitAggregateType(CodeGenFunction &CGF, unsigned N); |
159 | /// Emits the code for the variable-modified type, if required. |
160 | /// \param N Number of the reduction item. |
161 | /// \param Size Size of the type in chars. |
162 | void emitAggregateType(CodeGenFunction &CGF, unsigned N, llvm::Value *Size); |
163 | /// Performs initialization of the private copy for the reduction item. |
164 | /// \param N Number of the reduction item. |
165 | /// \param PrivateAddr Address of the corresponding private item. |
166 | /// \param DefaultInit Default initialization sequence that should be |
167 | /// performed if no reduction specific initialization is found. |
168 | /// \param SharedLVal Address of the original shared variable. |
169 | void |
170 | emitInitialization(CodeGenFunction &CGF, unsigned N, Address PrivateAddr, |
171 | LValue SharedLVal, |
172 | llvm::function_ref<bool(CodeGenFunction &)> DefaultInit); |
173 | /// Returns true if the private copy requires cleanups. |
174 | bool needCleanups(unsigned N); |
175 | /// Emits cleanup code for the reduction item. |
176 | /// \param N Number of the reduction item. |
177 | /// \param PrivateAddr Address of the corresponding private item. |
178 | void emitCleanups(CodeGenFunction &CGF, unsigned N, Address PrivateAddr); |
179 | /// Adjusts \p PrivatedAddr for using instead of the original variable |
180 | /// address in normal operations. |
181 | /// \param N Number of the reduction item. |
182 | /// \param PrivateAddr Address of the corresponding private item. |
183 | Address adjustPrivateAddress(CodeGenFunction &CGF, unsigned N, |
184 | Address PrivateAddr); |
185 | /// Returns LValue for the reduction item. |
186 | LValue getSharedLValue(unsigned N) const { return SharedAddresses[N].first; } |
187 | /// Returns the size of the reduction item (in chars and total number of |
188 | /// elements in the item), or nullptr, if the size is a constant. |
189 | std::pair<llvm::Value *, llvm::Value *> getSizes(unsigned N) const { |
190 | return Sizes[N]; |
191 | } |
192 | /// Returns the base declaration of the reduction item. |
193 | const VarDecl *getBaseDecl(unsigned N) const { return BaseDecls[N]; } |
194 | /// Returns the base declaration of the reduction item. |
195 | const Expr *getRefExpr(unsigned N) const { return ClausesData[N].Ref; } |
196 | /// Returns true if the initialization of the reduction item uses initializer |
197 | /// from declare reduction construct. |
198 | bool usesReductionInitializer(unsigned N) const; |
199 | }; |
200 | |
201 | class CGOpenMPRuntime { |
202 | public: |
203 | /// Allows to disable automatic handling of functions used in target regions |
204 | /// as those marked as `omp declare target`. |
205 | class DisableAutoDeclareTargetRAII { |
206 | CodeGenModule &CGM; |
207 | bool SavedShouldMarkAsGlobal; |
208 | |
209 | public: |
210 | DisableAutoDeclareTargetRAII(CodeGenModule &CGM); |
211 | ~DisableAutoDeclareTargetRAII(); |
212 | }; |
213 | |
214 | protected: |
215 | CodeGenModule &CGM; |
216 | StringRef FirstSeparator, Separator; |
217 | |
218 | /// Constructor allowing to redefine the name separator for the variables. |
219 | explicit CGOpenMPRuntime(CodeGenModule &CGM, StringRef FirstSeparator, |
220 | StringRef Separator); |
221 | |
222 | /// Creates offloading entry for the provided entry ID \a ID, |
223 | /// address \a Addr, size \a Size, and flags \a Flags. |
224 | virtual void createOffloadEntry(llvm::Constant *ID, llvm::Constant *Addr, |
225 | uint64_t Size, int32_t Flags, |
226 | llvm::GlobalValue::LinkageTypes Linkage); |
227 | |
228 | /// Helper to emit outlined function for 'target' directive. |
229 | /// \param D Directive to emit. |
230 | /// \param ParentName Name of the function that encloses the target region. |
231 | /// \param OutlinedFn Outlined function value to be defined by this call. |
232 | /// \param OutlinedFnID Outlined function ID value to be defined by this call. |
233 | /// \param IsOffloadEntry True if the outlined function is an offload entry. |
234 | /// \param CodeGen Lambda codegen specific to an accelerator device. |
235 | /// An outlined function may not be an entry if, e.g. the if clause always |
236 | /// evaluates to false. |
237 | virtual void emitTargetOutlinedFunctionHelper(const OMPExecutableDirective &D, |
238 | StringRef ParentName, |
239 | llvm::Function *&OutlinedFn, |
240 | llvm::Constant *&OutlinedFnID, |
241 | bool IsOffloadEntry, |
242 | const RegionCodeGenTy &CodeGen); |
243 | |
244 | /// Emits code for OpenMP 'if' clause using specified \a CodeGen |
245 | /// function. Here is the logic: |
246 | /// if (Cond) { |
247 | /// ThenGen(); |
248 | /// } else { |
249 | /// ElseGen(); |
250 | /// } |
251 | void emitOMPIfClause(CodeGenFunction &CGF, const Expr *Cond, |
252 | const RegionCodeGenTy &ThenGen, |
253 | const RegionCodeGenTy &ElseGen); |
254 | |
255 | /// Emits object of ident_t type with info for source location. |
256 | /// \param Flags Flags for OpenMP location. |
257 | /// |
258 | llvm::Value *emitUpdateLocation(CodeGenFunction &CGF, SourceLocation Loc, |
259 | unsigned Flags = 0); |
260 | |
261 | /// Returns pointer to ident_t type. |
262 | llvm::Type *getIdentTyPointerTy(); |
263 | |
264 | /// Gets thread id value for the current thread. |
265 | /// |
266 | llvm::Value *getThreadID(CodeGenFunction &CGF, SourceLocation Loc); |
267 | |
268 | /// Get the function name of an outlined region. |
269 | // The name can be customized depending on the target. |
270 | // |
271 | virtual StringRef getOutlinedHelperName() const { return ".omp_outlined."; } |
272 | |
273 | /// Emits \p Callee function call with arguments \p Args with location \p Loc. |
274 | void emitCall(CodeGenFunction &CGF, SourceLocation Loc, |
275 | llvm::FunctionCallee Callee, |
276 | ArrayRef<llvm::Value *> Args = llvm::None) const; |
277 | |
278 | /// Emits address of the word in a memory where current thread id is |
279 | /// stored. |
280 | virtual Address emitThreadIDAddress(CodeGenFunction &CGF, SourceLocation Loc); |
281 | |
282 | void setLocThreadIdInsertPt(CodeGenFunction &CGF, |
283 | bool AtCurrentPoint = false); |
284 | void clearLocThreadIdInsertPt(CodeGenFunction &CGF); |
285 | |
286 | /// Check if the default location must be constant. |
287 | /// Default is false to support OMPT/OMPD. |
288 | virtual bool isDefaultLocationConstant() const { return false; } |
289 | |
290 | /// Returns additional flags that can be stored in reserved_2 field of the |
291 | /// default location. |
292 | virtual unsigned getDefaultLocationReserved2Flags() const { return 0; } |
293 | |
294 | /// Returns default flags for the barriers depending on the directive, for |
295 | /// which this barier is going to be emitted. |
296 | static unsigned getDefaultFlagsForBarriers(OpenMPDirectiveKind Kind); |
297 | |
298 | /// Get the LLVM type for the critical name. |
299 | llvm::ArrayType *getKmpCriticalNameTy() const {return KmpCriticalNameTy;} |
300 | |
301 | /// Returns corresponding lock object for the specified critical region |
302 | /// name. If the lock object does not exist it is created, otherwise the |
303 | /// reference to the existing copy is returned. |
304 | /// \param CriticalName Name of the critical region. |
305 | /// |
306 | llvm::Value *getCriticalRegionLock(StringRef CriticalName); |
307 | |
308 | private: |
309 | /// Default const ident_t object used for initialization of all other |
310 | /// ident_t objects. |
311 | llvm::Constant *DefaultOpenMPPSource = nullptr; |
312 | using FlagsTy = std::pair<unsigned, unsigned>; |
313 | /// Map of flags and corresponding default locations. |
314 | using OpenMPDefaultLocMapTy = llvm::DenseMap<FlagsTy, llvm::Value *>; |
315 | OpenMPDefaultLocMapTy OpenMPDefaultLocMap; |
316 | Address getOrCreateDefaultLocation(unsigned Flags); |
317 | |
318 | QualType IdentQTy; |
319 | llvm::StructType *IdentTy = nullptr; |
320 | /// Map for SourceLocation and OpenMP runtime library debug locations. |
321 | typedef llvm::DenseMap<unsigned, llvm::Value *> OpenMPDebugLocMapTy; |
322 | OpenMPDebugLocMapTy OpenMPDebugLocMap; |
323 | /// The type for a microtask which gets passed to __kmpc_fork_call(). |
324 | /// Original representation is: |
325 | /// typedef void (kmpc_micro)(kmp_int32 global_tid, kmp_int32 bound_tid,...); |
326 | llvm::FunctionType *Kmpc_MicroTy = nullptr; |
327 | /// Stores debug location and ThreadID for the function. |
328 | struct DebugLocThreadIdTy { |
329 | llvm::Value *DebugLoc; |
330 | llvm::Value *ThreadID; |
331 | /// Insert point for the service instructions. |
332 | llvm::AssertingVH<llvm::Instruction> ServiceInsertPt = nullptr; |
333 | }; |
334 | /// Map of local debug location, ThreadId and functions. |
335 | typedef llvm::DenseMap<llvm::Function *, DebugLocThreadIdTy> |
336 | OpenMPLocThreadIDMapTy; |
337 | OpenMPLocThreadIDMapTy OpenMPLocThreadIDMap; |
338 | /// Map of UDRs and corresponding combiner/initializer. |
339 | typedef llvm::DenseMap<const OMPDeclareReductionDecl *, |
340 | std::pair<llvm::Function *, llvm::Function *>> |
341 | UDRMapTy; |
342 | UDRMapTy UDRMap; |
343 | /// Map of functions and locally defined UDRs. |
344 | typedef llvm::DenseMap<llvm::Function *, |
345 | SmallVector<const OMPDeclareReductionDecl *, 4>> |
346 | FunctionUDRMapTy; |
347 | FunctionUDRMapTy FunctionUDRMap; |
348 | /// Type kmp_critical_name, originally defined as typedef kmp_int32 |
349 | /// kmp_critical_name[8]; |
350 | llvm::ArrayType *KmpCriticalNameTy; |
351 | /// An ordered map of auto-generated variables to their unique names. |
352 | /// It stores variables with the following names: 1) ".gomp_critical_user_" + |
353 | /// <critical_section_name> + ".var" for "omp critical" directives; 2) |
354 | /// <mangled_name_for_global_var> + ".cache." for cache for threadprivate |
355 | /// variables. |
356 | llvm::StringMap<llvm::AssertingVH<llvm::Constant>, llvm::BumpPtrAllocator> |
357 | InternalVars; |
358 | /// Type typedef kmp_int32 (* kmp_routine_entry_t)(kmp_int32, void *); |
359 | llvm::Type *KmpRoutineEntryPtrTy = nullptr; |
360 | QualType KmpRoutineEntryPtrQTy; |
361 | /// Type typedef struct kmp_task { |
362 | /// void * shareds; /**< pointer to block of pointers to |
363 | /// shared vars */ |
364 | /// kmp_routine_entry_t routine; /**< pointer to routine to call for |
365 | /// executing task */ |
366 | /// kmp_int32 part_id; /**< part id for the task */ |
367 | /// kmp_routine_entry_t destructors; /* pointer to function to invoke |
368 | /// deconstructors of firstprivate C++ objects */ |
369 | /// } kmp_task_t; |
370 | QualType KmpTaskTQTy; |
371 | /// Saved kmp_task_t for task directive. |
372 | QualType SavedKmpTaskTQTy; |
373 | /// Saved kmp_task_t for taskloop-based directive. |
374 | QualType SavedKmpTaskloopTQTy; |
375 | /// Type typedef struct kmp_depend_info { |
376 | /// kmp_intptr_t base_addr; |
377 | /// size_t len; |
378 | /// struct { |
379 | /// bool in:1; |
380 | /// bool out:1; |
381 | /// } flags; |
382 | /// } kmp_depend_info_t; |
383 | QualType KmpDependInfoTy; |
384 | /// struct kmp_dim { // loop bounds info casted to kmp_int64 |
385 | /// kmp_int64 lo; // lower |
386 | /// kmp_int64 up; // upper |
387 | /// kmp_int64 st; // stride |
388 | /// }; |
389 | QualType KmpDimTy; |
390 | /// Type struct __tgt_offload_entry{ |
391 | /// void *addr; // Pointer to the offload entry info. |
392 | /// // (function or global) |
393 | /// char *name; // Name of the function or global. |
394 | /// size_t size; // Size of the entry info (0 if it a function). |
395 | /// }; |
396 | QualType TgtOffloadEntryQTy; |
397 | /// struct __tgt_device_image{ |
398 | /// void *ImageStart; // Pointer to the target code start. |
399 | /// void *ImageEnd; // Pointer to the target code end. |
400 | /// // We also add the host entries to the device image, as it may be useful |
401 | /// // for the target runtime to have access to that information. |
402 | /// __tgt_offload_entry *EntriesBegin; // Begin of the table with all |
403 | /// // the entries. |
404 | /// __tgt_offload_entry *EntriesEnd; // End of the table with all the |
405 | /// // entries (non inclusive). |
406 | /// }; |
407 | QualType TgtDeviceImageQTy; |
408 | /// struct __tgt_bin_desc{ |
409 | /// int32_t NumDevices; // Number of devices supported. |
410 | /// __tgt_device_image *DeviceImages; // Arrays of device images |
411 | /// // (one per device). |
412 | /// __tgt_offload_entry *EntriesBegin; // Begin of the table with all the |
413 | /// // entries. |
414 | /// __tgt_offload_entry *EntriesEnd; // End of the table with all the |
415 | /// // entries (non inclusive). |
416 | /// }; |
417 | QualType TgtBinaryDescriptorQTy; |
418 | /// Entity that registers the offloading constants that were emitted so |
419 | /// far. |
420 | class OffloadEntriesInfoManagerTy { |
421 | CodeGenModule &CGM; |
422 | |
423 | /// Number of entries registered so far. |
424 | unsigned OffloadingEntriesNum = 0; |
425 | |
426 | public: |
427 | /// Base class of the entries info. |
428 | class OffloadEntryInfo { |
429 | public: |
430 | /// Kind of a given entry. |
431 | enum OffloadingEntryInfoKinds : unsigned { |
432 | /// Entry is a target region. |
433 | OffloadingEntryInfoTargetRegion = 0, |
434 | /// Entry is a declare target variable. |
435 | OffloadingEntryInfoDeviceGlobalVar = 1, |
436 | /// Invalid entry info. |
437 | OffloadingEntryInfoInvalid = ~0u |
438 | }; |
439 | |
440 | protected: |
441 | OffloadEntryInfo() = delete; |
442 | explicit OffloadEntryInfo(OffloadingEntryInfoKinds Kind) : Kind(Kind) {} |
443 | explicit OffloadEntryInfo(OffloadingEntryInfoKinds Kind, unsigned Order, |
444 | uint32_t Flags) |
445 | : Flags(Flags), Order(Order), Kind(Kind) {} |
446 | ~OffloadEntryInfo() = default; |
447 | |
448 | public: |
449 | bool isValid() const { return Order != ~0u; } |
450 | unsigned getOrder() const { return Order; } |
451 | OffloadingEntryInfoKinds getKind() const { return Kind; } |
452 | uint32_t getFlags() const { return Flags; } |
453 | void setFlags(uint32_t NewFlags) { Flags = NewFlags; } |
454 | llvm::Constant *getAddress() const { |
455 | return cast_or_null<llvm::Constant>(Addr); |
456 | } |
457 | void setAddress(llvm::Constant *V) { |
458 | assert(!Addr.pointsToAliveValue() && "Address has been set before!")((!Addr.pointsToAliveValue() && "Address has been set before!" ) ? static_cast<void> (0) : __assert_fail ("!Addr.pointsToAliveValue() && \"Address has been set before!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.h" , 458, __PRETTY_FUNCTION__)); |
459 | Addr = V; |
460 | } |
461 | static bool classof(const OffloadEntryInfo *Info) { return true; } |
462 | |
463 | private: |
464 | /// Address of the entity that has to be mapped for offloading. |
465 | llvm::WeakTrackingVH Addr; |
466 | |
467 | /// Flags associated with the device global. |
468 | uint32_t Flags = 0u; |
469 | |
470 | /// Order this entry was emitted. |
471 | unsigned Order = ~0u; |
472 | |
473 | OffloadingEntryInfoKinds Kind = OffloadingEntryInfoInvalid; |
474 | }; |
475 | |
476 | /// Return true if a there are no entries defined. |
477 | bool empty() const; |
478 | /// Return number of entries defined so far. |
479 | unsigned size() const { return OffloadingEntriesNum; } |
480 | OffloadEntriesInfoManagerTy(CodeGenModule &CGM) : CGM(CGM) {} |
481 | |
482 | // |
483 | // Target region entries related. |
484 | // |
485 | |
486 | /// Kind of the target registry entry. |
487 | enum OMPTargetRegionEntryKind : uint32_t { |
488 | /// Mark the entry as target region. |
489 | OMPTargetRegionEntryTargetRegion = 0x0, |
490 | /// Mark the entry as a global constructor. |
491 | OMPTargetRegionEntryCtor = 0x02, |
492 | /// Mark the entry as a global destructor. |
493 | OMPTargetRegionEntryDtor = 0x04, |
494 | }; |
495 | |
496 | /// Target region entries info. |
497 | class OffloadEntryInfoTargetRegion final : public OffloadEntryInfo { |
498 | /// Address that can be used as the ID of the entry. |
499 | llvm::Constant *ID = nullptr; |
500 | |
501 | public: |
502 | OffloadEntryInfoTargetRegion() |
503 | : OffloadEntryInfo(OffloadingEntryInfoTargetRegion) {} |
504 | explicit OffloadEntryInfoTargetRegion(unsigned Order, |
505 | llvm::Constant *Addr, |
506 | llvm::Constant *ID, |
507 | OMPTargetRegionEntryKind Flags) |
508 | : OffloadEntryInfo(OffloadingEntryInfoTargetRegion, Order, Flags), |
509 | ID(ID) { |
510 | setAddress(Addr); |
511 | } |
512 | |
513 | llvm::Constant *getID() const { return ID; } |
514 | void setID(llvm::Constant *V) { |
515 | assert(!ID && "ID has been set before!")((!ID && "ID has been set before!") ? static_cast< void> (0) : __assert_fail ("!ID && \"ID has been set before!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/tools/clang/lib/CodeGen/CGOpenMPRuntime.h" , 515, __PRETTY_FUNCTION__)); |
516 | ID = V; |
517 | } |
518 | static bool classof(const OffloadEntryInfo *Info) { |
519 | return Info->getKind() == OffloadingEntryInfoTargetRegion; |
520 | } |
521 | }; |
522 | |
523 | /// Initialize target region entry. |
524 | void initializeTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID, |
525 | StringRef ParentName, unsigned LineNum, |
526 | unsigned Order); |
527 | /// Register target region entry. |
528 | void registerTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID, |
529 | StringRef ParentName, unsigned LineNum, |
530 | llvm::Constant *Addr, llvm::Constant *ID, |
531 | OMPTargetRegionEntryKind Flags); |
532 | /// Return true if a target region entry with the provided information |
533 | /// exists. |
534 | bool hasTargetRegionEntryInfo(unsigned DeviceID, unsigned FileID, |
535 | StringRef ParentName, unsigned LineNum) const; |
536 | /// brief Applies action \a Action on all registered entries. |
537 | typedef llvm::function_ref<void(unsigned, unsigned, StringRef, unsigned, |
538 | const OffloadEntryInfoTargetRegion &)> |
539 | OffloadTargetRegionEntryInfoActTy; |
540 | void actOnTargetRegionEntriesInfo( |
541 | const OffloadTargetRegionEntryInfoActTy &Action); |
542 | |
543 | // |
544 | // Device global variable entries related. |
545 | // |
546 | |
547 | /// Kind of the global variable entry.. |
548 | enum OMPTargetGlobalVarEntryKind : uint32_t { |
549 | /// Mark the entry as a to declare target. |
550 | OMPTargetGlobalVarEntryTo = 0x0, |
551 | /// Mark the entry as a to declare target link. |
552 | OMPTargetGlobalVarEntryLink = 0x1, |
553 | }; |
554 | |
555 | /// Device global variable entries info. |
556 | class OffloadEntryInfoDeviceGlobalVar final : public OffloadEntryInfo { |
557 | /// Type of the global variable. |
558 | CharUnits VarSize; |
559 | llvm::GlobalValue::LinkageTypes Linkage; |
560 | |
561 | public: |
562 | OffloadEntryInfoDeviceGlobalVar() |
563 | : OffloadEntryInfo(OffloadingEntryInfoDeviceGlobalVar) {} |
564 | explicit OffloadEntryInfoDeviceGlobalVar(unsigned Order, |
565 | OMPTargetGlobalVarEntryKind Flags) |
566 | : OffloadEntryInfo(OffloadingEntryInfoDeviceGlobalVar, Order, Flags) {} |
567 | explicit OffloadEntryInfoDeviceGlobalVar( |
568 | unsigned Order, llvm::Constant *Addr, CharUnits VarSize, |
569 | OMPTargetGlobalVarEntryKind Flags, |
570 | llvm::GlobalValue::LinkageTypes Linkage) |
571 | : OffloadEntryInfo(OffloadingEntryInfoDeviceGlobalVar, Order, Flags), |
572 | VarSize(VarSize), Linkage(Linkage) { |
573 | setAddress(Addr); |
574 | } |
575 | |
576 | CharUnits getVarSize() const { return VarSize; } |
577 | void setVarSize(CharUnits Size) { VarSize = Size; } |
578 | llvm::GlobalValue::LinkageTypes getLinkage() const { return Linkage; } |
579 | void setLinkage(llvm::GlobalValue::LinkageTypes LT) { Linkage = LT; } |
580 | static bool classof(const OffloadEntryInfo *Info) { |
581 | return Info->getKind() == OffloadingEntryInfoDeviceGlobalVar; |
582 | } |
583 | }; |
584 | |
585 | /// Initialize device global variable entry. |
586 | void initializeDeviceGlobalVarEntryInfo(StringRef Name, |
587 | OMPTargetGlobalVarEntryKind Flags, |
588 | unsigned Order); |
589 | |
590 | /// Register device global variable entry. |
591 | void |
592 | registerDeviceGlobalVarEntryInfo(StringRef VarName, llvm::Constant *Addr, |
593 | CharUnits VarSize, |
594 | OMPTargetGlobalVarEntryKind Flags, |
595 | llvm::GlobalValue::LinkageTypes Linkage); |
596 | /// Checks if the variable with the given name has been registered already. |
597 | bool hasDeviceGlobalVarEntryInfo(StringRef VarName) const { |
598 | return OffloadEntriesDeviceGlobalVar.count(VarName) > 0; |
599 | } |
600 | /// Applies action \a Action on all registered entries. |
601 | typedef llvm::function_ref<void(StringRef, |
602 | const OffloadEntryInfoDeviceGlobalVar &)> |
603 | OffloadDeviceGlobalVarEntryInfoActTy; |
604 | void actOnDeviceGlobalVarEntriesInfo( |
605 | const OffloadDeviceGlobalVarEntryInfoActTy &Action); |
606 | |
607 | private: |
608 | // Storage for target region entries kind. The storage is to be indexed by |
609 | // file ID, device ID, parent function name and line number. |
610 | typedef llvm::DenseMap<unsigned, OffloadEntryInfoTargetRegion> |
611 | OffloadEntriesTargetRegionPerLine; |
612 | typedef llvm::StringMap<OffloadEntriesTargetRegionPerLine> |
613 | OffloadEntriesTargetRegionPerParentName; |
614 | typedef llvm::DenseMap<unsigned, OffloadEntriesTargetRegionPerParentName> |
615 | OffloadEntriesTargetRegionPerFile; |
616 | typedef llvm::DenseMap<unsigned, OffloadEntriesTargetRegionPerFile> |
617 | OffloadEntriesTargetRegionPerDevice; |
618 | typedef OffloadEntriesTargetRegionPerDevice OffloadEntriesTargetRegionTy; |
619 | OffloadEntriesTargetRegionTy OffloadEntriesTargetRegion; |
620 | /// Storage for device global variable entries kind. The storage is to be |
621 | /// indexed by mangled name. |
622 | typedef llvm::StringMap<OffloadEntryInfoDeviceGlobalVar> |
623 | OffloadEntriesDeviceGlobalVarTy; |
624 | OffloadEntriesDeviceGlobalVarTy OffloadEntriesDeviceGlobalVar; |
625 | }; |
626 | OffloadEntriesInfoManagerTy OffloadEntriesInfoManager; |
627 | |
628 | bool ShouldMarkAsGlobal = true; |
629 | /// List of the emitted functions. |
630 | llvm::StringSet<> AlreadyEmittedTargetFunctions; |
631 | /// List of the global variables with their addresses that should not be |
632 | /// emitted for the target. |
633 | llvm::StringMap<llvm::WeakTrackingVH> EmittedNonTargetVariables; |
634 | |
635 | /// List of variables that can become declare target implicitly and, thus, |
636 | /// must be emitted. |
637 | llvm::SmallDenseSet<const VarDecl *> DeferredGlobalVariables; |
638 | |
639 | /// Flag for keeping track of weather a requires unified_shared_memory |
640 | /// directive is present. |
641 | bool HasRequiresUnifiedSharedMemory = false; |
642 | |
643 | /// Flag for keeping track of weather a target region has been emitted. |
644 | bool HasEmittedTargetRegion = false; |
645 | |
646 | /// Flag for keeping track of weather a device routine has been emitted. |
647 | /// Device routines are specific to the |
648 | bool HasEmittedDeclareTargetRegion = false; |
649 | |
650 | /// Creates and registers offloading binary descriptor for the current |
651 | /// compilation unit. The function that does the registration is returned. |
652 | llvm::Function *createOffloadingBinaryDescriptorRegistration(); |
653 | |
654 | /// Creates all the offload entries in the current compilation unit |
655 | /// along with the associated metadata. |
656 | void createOffloadEntriesAndInfoMetadata(); |
657 | |
658 | /// Loads all the offload entries information from the host IR |
659 | /// metadata. |
660 | void loadOffloadInfoMetadata(); |
661 | |
662 | /// Returns __tgt_offload_entry type. |
663 | QualType getTgtOffloadEntryQTy(); |
664 | |
665 | /// Returns __tgt_device_image type. |
666 | QualType getTgtDeviceImageQTy(); |
667 | |
668 | /// Returns __tgt_bin_desc type. |
669 | QualType getTgtBinaryDescriptorQTy(); |
670 | |
671 | /// Start scanning from statement \a S and and emit all target regions |
672 | /// found along the way. |
673 | /// \param S Starting statement. |
674 | /// \param ParentName Name of the function declaration that is being scanned. |
675 | void scanForTargetRegionsFunctions(const Stmt *S, StringRef ParentName); |
676 | |
677 | /// Build type kmp_routine_entry_t (if not built yet). |
678 | void emitKmpRoutineEntryT(QualType KmpInt32Ty); |
679 | |
680 | /// Returns pointer to kmpc_micro type. |
681 | llvm::Type *getKmpc_MicroPointerTy(); |
682 | |
683 | /// Returns specified OpenMP runtime function. |
684 | /// \param Function OpenMP runtime function. |
685 | /// \return Specified function. |
686 | llvm::FunctionCallee createRuntimeFunction(unsigned Function); |
687 | |
688 | /// Returns __kmpc_for_static_init_* runtime function for the specified |
689 | /// size \a IVSize and sign \a IVSigned. |
690 | llvm::FunctionCallee createForStaticInitFunction(unsigned IVSize, |
691 | bool IVSigned); |
692 | |
693 | /// Returns __kmpc_dispatch_init_* runtime function for the specified |
694 | /// size \a IVSize and sign \a IVSigned. |
695 | llvm::FunctionCallee createDispatchInitFunction(unsigned IVSize, |
696 | bool IVSigned); |
697 | |
698 | /// Returns __kmpc_dispatch_next_* runtime function for the specified |
699 | /// size \a IVSize and sign \a IVSigned. |
700 | llvm::FunctionCallee createDispatchNextFunction(unsigned IVSize, |
701 | bool IVSigned); |
702 | |
703 | /// Returns __kmpc_dispatch_fini_* runtime function for the specified |
704 | /// size \a IVSize and sign \a IVSigned. |
705 | llvm::FunctionCallee createDispatchFiniFunction(unsigned IVSize, |
706 | bool IVSigned); |
707 | |
708 | /// If the specified mangled name is not in the module, create and |
709 | /// return threadprivate cache object. This object is a pointer's worth of |
710 | /// storage that's reserved for use by the OpenMP runtime. |
711 | /// \param VD Threadprivate variable. |
712 | /// \return Cache variable for the specified threadprivate. |
713 | llvm::Constant *getOrCreateThreadPrivateCache(const VarDecl *VD); |
714 | |
715 | /// Gets (if variable with the given name already exist) or creates |
716 | /// internal global variable with the specified Name. The created variable has |
717 | /// linkage CommonLinkage by default and is initialized by null value. |
718 | /// \param Ty Type of the global variable. If it is exist already the type |
719 | /// must be the same. |
720 | /// \param Name Name of the variable. |
721 | llvm::Constant *getOrCreateInternalVariable(llvm::Type *Ty, |
722 | const llvm::Twine &Name, |
723 | unsigned AddressSpace = 0); |
724 | |
725 | /// Set of threadprivate variables with the generated initializer. |
726 | llvm::StringSet<> ThreadPrivateWithDefinition; |
727 | |
728 | /// Set of declare target variables with the generated initializer. |
729 | llvm::StringSet<> DeclareTargetWithDefinition; |
730 | |
731 | /// Emits initialization code for the threadprivate variables. |
732 | /// \param VDAddr Address of the global variable \a VD. |
733 | /// \param Ctor Pointer to a global init function for \a VD. |
734 | /// \param CopyCtor Pointer to a global copy function for \a VD. |
735 | /// \param Dtor Pointer to a global destructor function for \a VD. |
736 | /// \param Loc Location of threadprivate declaration. |
737 | void emitThreadPrivateVarInit(CodeGenFunction &CGF, Address VDAddr, |
738 | llvm::Value *Ctor, llvm::Value *CopyCtor, |
739 | llvm::Value *Dtor, SourceLocation Loc); |
740 | |
741 | struct TaskResultTy { |
742 | llvm::Value *NewTask = nullptr; |
743 | llvm::Function *TaskEntry = nullptr; |
744 | llvm::Value *NewTaskNewTaskTTy = nullptr; |
745 | LValue TDBase; |
746 | const RecordDecl *KmpTaskTQTyRD = nullptr; |
747 | llvm::Value *TaskDupFn = nullptr; |
748 | }; |
749 | /// Emit task region for the task directive. The task region is emitted in |
750 | /// several steps: |
751 | /// 1. Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 |
752 | /// gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, |
753 | /// kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the |
754 | /// function: |
755 | /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { |
756 | /// TaskFunction(gtid, tt->part_id, tt->shareds); |
757 | /// return 0; |
758 | /// } |
759 | /// 2. Copy a list of shared variables to field shareds of the resulting |
760 | /// structure kmp_task_t returned by the previous call (if any). |
761 | /// 3. Copy a pointer to destructions function to field destructions of the |
762 | /// resulting structure kmp_task_t. |
763 | /// \param D Current task directive. |
764 | /// \param TaskFunction An LLVM function with type void (*)(i32 /*gtid*/, i32 |
765 | /// /*part_id*/, captured_struct */*__context*/); |
766 | /// \param SharedsTy A type which contains references the shared variables. |
767 | /// \param Shareds Context with the list of shared variables from the \p |
768 | /// TaskFunction. |
769 | /// \param Data Additional data for task generation like tiednsee, final |
770 | /// state, list of privates etc. |
771 | TaskResultTy emitTaskInit(CodeGenFunction &CGF, SourceLocation Loc, |
772 | const OMPExecutableDirective &D, |
773 | llvm::Function *TaskFunction, QualType SharedsTy, |
774 | Address Shareds, const OMPTaskDataTy &Data); |
775 | |
776 | /// Returns default address space for the constant firstprivates, 0 by |
777 | /// default. |
778 | virtual unsigned getDefaultFirstprivateAddressSpace() const { return 0; } |
779 | |
780 | public: |
781 | explicit CGOpenMPRuntime(CodeGenModule &CGM) |
782 | : CGOpenMPRuntime(CGM, ".", ".") {} |
783 | virtual ~CGOpenMPRuntime() {} |
784 | virtual void clear(); |
785 | |
786 | /// Checks if the \p Body is the \a CompoundStmt and returns its child |
787 | /// statement iff there is only one that is not evaluatable at the compile |
788 | /// time. |
789 | static const Stmt *getSingleCompoundChild(ASTContext &Ctx, const Stmt *Body); |
790 | |
791 | /// Get the platform-specific name separator. |
792 | std::string getName(ArrayRef<StringRef> Parts) const; |
793 | |
794 | /// Emit code for the specified user defined reduction construct. |
795 | virtual void emitUserDefinedReduction(CodeGenFunction *CGF, |
796 | const OMPDeclareReductionDecl *D); |
797 | /// Get combiner/initializer for the specified user-defined reduction, if any. |
798 | virtual std::pair<llvm::Function *, llvm::Function *> |
799 | getUserDefinedReduction(const OMPDeclareReductionDecl *D); |
800 | |
801 | /// Emits outlined function for the specified OpenMP parallel directive |
802 | /// \a D. This outlined function has type void(*)(kmp_int32 *ThreadID, |
803 | /// kmp_int32 BoundID, struct context_vars*). |
804 | /// \param D OpenMP directive. |
805 | /// \param ThreadIDVar Variable for thread id in the current OpenMP region. |
806 | /// \param InnermostKind Kind of innermost directive (for simple directives it |
807 | /// is a directive itself, for combined - its innermost directive). |
808 | /// \param CodeGen Code generation sequence for the \a D directive. |
809 | virtual llvm::Function *emitParallelOutlinedFunction( |
810 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, |
811 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen); |
812 | |
813 | /// Emits outlined function for the specified OpenMP teams directive |
814 | /// \a D. This outlined function has type void(*)(kmp_int32 *ThreadID, |
815 | /// kmp_int32 BoundID, struct context_vars*). |
816 | /// \param D OpenMP directive. |
817 | /// \param ThreadIDVar Variable for thread id in the current OpenMP region. |
818 | /// \param InnermostKind Kind of innermost directive (for simple directives it |
819 | /// is a directive itself, for combined - its innermost directive). |
820 | /// \param CodeGen Code generation sequence for the \a D directive. |
821 | virtual llvm::Function *emitTeamsOutlinedFunction( |
822 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, |
823 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen); |
824 | |
825 | /// Emits outlined function for the OpenMP task directive \a D. This |
826 | /// outlined function has type void(*)(kmp_int32 ThreadID, struct task_t* |
827 | /// TaskT). |
828 | /// \param D OpenMP directive. |
829 | /// \param ThreadIDVar Variable for thread id in the current OpenMP region. |
830 | /// \param PartIDVar Variable for partition id in the current OpenMP untied |
831 | /// task region. |
832 | /// \param TaskTVar Variable for task_t argument. |
833 | /// \param InnermostKind Kind of innermost directive (for simple directives it |
834 | /// is a directive itself, for combined - its innermost directive). |
835 | /// \param CodeGen Code generation sequence for the \a D directive. |
836 | /// \param Tied true if task is generated for tied task, false otherwise. |
837 | /// \param NumberOfParts Number of parts in untied task. Ignored for tied |
838 | /// tasks. |
839 | /// |
840 | virtual llvm::Function *emitTaskOutlinedFunction( |
841 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, |
842 | const VarDecl *PartIDVar, const VarDecl *TaskTVar, |
843 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen, |
844 | bool Tied, unsigned &NumberOfParts); |
845 | |
846 | /// Cleans up references to the objects in finished function. |
847 | /// |
848 | virtual void functionFinished(CodeGenFunction &CGF); |
849 | |
850 | /// Emits code for parallel or serial call of the \a OutlinedFn with |
851 | /// variables captured in a record which address is stored in \a |
852 | /// CapturedStruct. |
853 | /// \param OutlinedFn Outlined function to be run in parallel threads. Type of |
854 | /// this function is void(*)(kmp_int32 *, kmp_int32, struct context_vars*). |
855 | /// \param CapturedVars A pointer to the record with the references to |
856 | /// variables used in \a OutlinedFn function. |
857 | /// \param IfCond Condition in the associated 'if' clause, if it was |
858 | /// specified, nullptr otherwise. |
859 | /// |
860 | virtual void emitParallelCall(CodeGenFunction &CGF, SourceLocation Loc, |
861 | llvm::Function *OutlinedFn, |
862 | ArrayRef<llvm::Value *> CapturedVars, |
863 | const Expr *IfCond); |
864 | |
865 | /// Emits a critical region. |
866 | /// \param CriticalName Name of the critical region. |
867 | /// \param CriticalOpGen Generator for the statement associated with the given |
868 | /// critical region. |
869 | /// \param Hint Value of the 'hint' clause (optional). |
870 | virtual void emitCriticalRegion(CodeGenFunction &CGF, StringRef CriticalName, |
871 | const RegionCodeGenTy &CriticalOpGen, |
872 | SourceLocation Loc, |
873 | const Expr *Hint = nullptr); |
874 | |
875 | /// Emits a master region. |
876 | /// \param MasterOpGen Generator for the statement associated with the given |
877 | /// master region. |
878 | virtual void emitMasterRegion(CodeGenFunction &CGF, |
879 | const RegionCodeGenTy &MasterOpGen, |
880 | SourceLocation Loc); |
881 | |
882 | /// Emits code for a taskyield directive. |
883 | virtual void emitTaskyieldCall(CodeGenFunction &CGF, SourceLocation Loc); |
884 | |
885 | /// Emit a taskgroup region. |
886 | /// \param TaskgroupOpGen Generator for the statement associated with the |
887 | /// given taskgroup region. |
888 | virtual void emitTaskgroupRegion(CodeGenFunction &CGF, |
889 | const RegionCodeGenTy &TaskgroupOpGen, |
890 | SourceLocation Loc); |
891 | |
892 | /// Emits a single region. |
893 | /// \param SingleOpGen Generator for the statement associated with the given |
894 | /// single region. |
895 | virtual void emitSingleRegion(CodeGenFunction &CGF, |
896 | const RegionCodeGenTy &SingleOpGen, |
897 | SourceLocation Loc, |
898 | ArrayRef<const Expr *> CopyprivateVars, |
899 | ArrayRef<const Expr *> DestExprs, |
900 | ArrayRef<const Expr *> SrcExprs, |
901 | ArrayRef<const Expr *> AssignmentOps); |
902 | |
903 | /// Emit an ordered region. |
904 | /// \param OrderedOpGen Generator for the statement associated with the given |
905 | /// ordered region. |
906 | virtual void emitOrderedRegion(CodeGenFunction &CGF, |
907 | const RegionCodeGenTy &OrderedOpGen, |
908 | SourceLocation Loc, bool IsThreads); |
909 | |
910 | /// Emit an implicit/explicit barrier for OpenMP threads. |
911 | /// \param Kind Directive for which this implicit barrier call must be |
912 | /// generated. Must be OMPD_barrier for explicit barrier generation. |
913 | /// \param EmitChecks true if need to emit checks for cancellation barriers. |
914 | /// \param ForceSimpleCall true simple barrier call must be emitted, false if |
915 | /// runtime class decides which one to emit (simple or with cancellation |
916 | /// checks). |
917 | /// |
918 | virtual void emitBarrierCall(CodeGenFunction &CGF, SourceLocation Loc, |
919 | OpenMPDirectiveKind Kind, |
920 | bool EmitChecks = true, |
921 | bool ForceSimpleCall = false); |
922 | |
923 | /// Check if the specified \a ScheduleKind is static non-chunked. |
924 | /// This kind of worksharing directive is emitted without outer loop. |
925 | /// \param ScheduleKind Schedule kind specified in the 'schedule' clause. |
926 | /// \param Chunked True if chunk is specified in the clause. |
927 | /// |
928 | virtual bool isStaticNonchunked(OpenMPScheduleClauseKind ScheduleKind, |
929 | bool Chunked) const; |
930 | |
931 | /// Check if the specified \a ScheduleKind is static non-chunked. |
932 | /// This kind of distribute directive is emitted without outer loop. |
933 | /// \param ScheduleKind Schedule kind specified in the 'dist_schedule' clause. |
934 | /// \param Chunked True if chunk is specified in the clause. |
935 | /// |
936 | virtual bool isStaticNonchunked(OpenMPDistScheduleClauseKind ScheduleKind, |
937 | bool Chunked) const; |
938 | |
939 | /// Check if the specified \a ScheduleKind is static chunked. |
940 | /// \param ScheduleKind Schedule kind specified in the 'schedule' clause. |
941 | /// \param Chunked True if chunk is specified in the clause. |
942 | /// |
943 | virtual bool isStaticChunked(OpenMPScheduleClauseKind ScheduleKind, |
944 | bool Chunked) const; |
945 | |
946 | /// Check if the specified \a ScheduleKind is static non-chunked. |
947 | /// \param ScheduleKind Schedule kind specified in the 'dist_schedule' clause. |
948 | /// \param Chunked True if chunk is specified in the clause. |
949 | /// |
950 | virtual bool isStaticChunked(OpenMPDistScheduleClauseKind ScheduleKind, |
951 | bool Chunked) const; |
952 | |
953 | /// Check if the specified \a ScheduleKind is dynamic. |
954 | /// This kind of worksharing directive is emitted without outer loop. |
955 | /// \param ScheduleKind Schedule Kind specified in the 'schedule' clause. |
956 | /// |
957 | virtual bool isDynamic(OpenMPScheduleClauseKind ScheduleKind) const; |
958 | |
959 | /// struct with the values to be passed to the dispatch runtime function |
960 | struct DispatchRTInput { |
961 | /// Loop lower bound |
962 | llvm::Value *LB = nullptr; |
963 | /// Loop upper bound |
964 | llvm::Value *UB = nullptr; |
965 | /// Chunk size specified using 'schedule' clause (nullptr if chunk |
966 | /// was not specified) |
967 | llvm::Value *Chunk = nullptr; |
968 | DispatchRTInput() = default; |
969 | DispatchRTInput(llvm::Value *LB, llvm::Value *UB, llvm::Value *Chunk) |
970 | : LB(LB), UB(UB), Chunk(Chunk) {} |
971 | }; |
972 | |
973 | /// Call the appropriate runtime routine to initialize it before start |
974 | /// of loop. |
975 | |
976 | /// This is used for non static scheduled types and when the ordered |
977 | /// clause is present on the loop construct. |
978 | /// Depending on the loop schedule, it is necessary to call some runtime |
979 | /// routine before start of the OpenMP loop to get the loop upper / lower |
980 | /// bounds \a LB and \a UB and stride \a ST. |
981 | /// |
982 | /// \param CGF Reference to current CodeGenFunction. |
983 | /// \param Loc Clang source location. |
984 | /// \param ScheduleKind Schedule kind, specified by the 'schedule' clause. |
985 | /// \param IVSize Size of the iteration variable in bits. |
986 | /// \param IVSigned Sign of the iteration variable. |
987 | /// \param Ordered true if loop is ordered, false otherwise. |
988 | /// \param DispatchValues struct containing llvm values for lower bound, upper |
989 | /// bound, and chunk expression. |
990 | /// For the default (nullptr) value, the chunk 1 will be used. |
991 | /// |
992 | virtual void emitForDispatchInit(CodeGenFunction &CGF, SourceLocation Loc, |
993 | const OpenMPScheduleTy &ScheduleKind, |
994 | unsigned IVSize, bool IVSigned, bool Ordered, |
995 | const DispatchRTInput &DispatchValues); |
996 | |
997 | /// Struct with the values to be passed to the static runtime function |
998 | struct StaticRTInput { |
999 | /// Size of the iteration variable in bits. |
1000 | unsigned IVSize = 0; |
1001 | /// Sign of the iteration variable. |
1002 | bool IVSigned = false; |
1003 | /// true if loop is ordered, false otherwise. |
1004 | bool Ordered = false; |
1005 | /// Address of the output variable in which the flag of the last iteration |
1006 | /// is returned. |
1007 | Address IL = Address::invalid(); |
1008 | /// Address of the output variable in which the lower iteration number is |
1009 | /// returned. |
1010 | Address LB = Address::invalid(); |
1011 | /// Address of the output variable in which the upper iteration number is |
1012 | /// returned. |
1013 | Address UB = Address::invalid(); |
1014 | /// Address of the output variable in which the stride value is returned |
1015 | /// necessary to generated the static_chunked scheduled loop. |
1016 | Address ST = Address::invalid(); |
1017 | /// Value of the chunk for the static_chunked scheduled loop. For the |
1018 | /// default (nullptr) value, the chunk 1 will be used. |
1019 | llvm::Value *Chunk = nullptr; |
1020 | StaticRTInput(unsigned IVSize, bool IVSigned, bool Ordered, Address IL, |
1021 | Address LB, Address UB, Address ST, |
1022 | llvm::Value *Chunk = nullptr) |
1023 | : IVSize(IVSize), IVSigned(IVSigned), Ordered(Ordered), IL(IL), LB(LB), |
1024 | UB(UB), ST(ST), Chunk(Chunk) {} |
1025 | }; |
1026 | /// Call the appropriate runtime routine to initialize it before start |
1027 | /// of loop. |
1028 | /// |
1029 | /// This is used only in case of static schedule, when the user did not |
1030 | /// specify a ordered clause on the loop construct. |
1031 | /// Depending on the loop schedule, it is necessary to call some runtime |
1032 | /// routine before start of the OpenMP loop to get the loop upper / lower |
1033 | /// bounds LB and UB and stride ST. |
1034 | /// |
1035 | /// \param CGF Reference to current CodeGenFunction. |
1036 | /// \param Loc Clang source location. |
1037 | /// \param DKind Kind of the directive. |
1038 | /// \param ScheduleKind Schedule kind, specified by the 'schedule' clause. |
1039 | /// \param Values Input arguments for the construct. |
1040 | /// |
1041 | virtual void emitForStaticInit(CodeGenFunction &CGF, SourceLocation Loc, |
1042 | OpenMPDirectiveKind DKind, |
1043 | const OpenMPScheduleTy &ScheduleKind, |
1044 | const StaticRTInput &Values); |
1045 | |
1046 | /// |
1047 | /// \param CGF Reference to current CodeGenFunction. |
1048 | /// \param Loc Clang source location. |
1049 | /// \param SchedKind Schedule kind, specified by the 'dist_schedule' clause. |
1050 | /// \param Values Input arguments for the construct. |
1051 | /// |
1052 | virtual void emitDistributeStaticInit(CodeGenFunction &CGF, |
1053 | SourceLocation Loc, |
1054 | OpenMPDistScheduleClauseKind SchedKind, |
1055 | const StaticRTInput &Values); |
1056 | |
1057 | /// Call the appropriate runtime routine to notify that we finished |
1058 | /// iteration of the ordered loop with the dynamic scheduling. |
1059 | /// |
1060 | /// \param CGF Reference to current CodeGenFunction. |
1061 | /// \param Loc Clang source location. |
1062 | /// \param IVSize Size of the iteration variable in bits. |
1063 | /// \param IVSigned Sign of the iteration variable. |
1064 | /// |
1065 | virtual void emitForOrderedIterationEnd(CodeGenFunction &CGF, |
1066 | SourceLocation Loc, unsigned IVSize, |
1067 | bool IVSigned); |
1068 | |
1069 | /// Call the appropriate runtime routine to notify that we finished |
1070 | /// all the work with current loop. |
1071 | /// |
1072 | /// \param CGF Reference to current CodeGenFunction. |
1073 | /// \param Loc Clang source location. |
1074 | /// \param DKind Kind of the directive for which the static finish is emitted. |
1075 | /// |
1076 | virtual void emitForStaticFinish(CodeGenFunction &CGF, SourceLocation Loc, |
1077 | OpenMPDirectiveKind DKind); |
1078 | |
1079 | /// Call __kmpc_dispatch_next( |
1080 | /// ident_t *loc, kmp_int32 tid, kmp_int32 *p_lastiter, |
1081 | /// kmp_int[32|64] *p_lower, kmp_int[32|64] *p_upper, |
1082 | /// kmp_int[32|64] *p_stride); |
1083 | /// \param IVSize Size of the iteration variable in bits. |
1084 | /// \param IVSigned Sign of the iteration variable. |
1085 | /// \param IL Address of the output variable in which the flag of the |
1086 | /// last iteration is returned. |
1087 | /// \param LB Address of the output variable in which the lower iteration |
1088 | /// number is returned. |
1089 | /// \param UB Address of the output variable in which the upper iteration |
1090 | /// number is returned. |
1091 | /// \param ST Address of the output variable in which the stride value is |
1092 | /// returned. |
1093 | virtual llvm::Value *emitForNext(CodeGenFunction &CGF, SourceLocation Loc, |
1094 | unsigned IVSize, bool IVSigned, |
1095 | Address IL, Address LB, |
1096 | Address UB, Address ST); |
1097 | |
1098 | /// Emits call to void __kmpc_push_num_threads(ident_t *loc, kmp_int32 |
1099 | /// global_tid, kmp_int32 num_threads) to generate code for 'num_threads' |
1100 | /// clause. |
1101 | /// \param NumThreads An integer value of threads. |
1102 | virtual void emitNumThreadsClause(CodeGenFunction &CGF, |
1103 | llvm::Value *NumThreads, |
1104 | SourceLocation Loc); |
1105 | |
1106 | /// Emit call to void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 |
1107 | /// global_tid, int proc_bind) to generate code for 'proc_bind' clause. |
1108 | virtual void emitProcBindClause(CodeGenFunction &CGF, |
1109 | OpenMPProcBindClauseKind ProcBind, |
1110 | SourceLocation Loc); |
1111 | |
1112 | /// Returns address of the threadprivate variable for the current |
1113 | /// thread. |
1114 | /// \param VD Threadprivate variable. |
1115 | /// \param VDAddr Address of the global variable \a VD. |
1116 | /// \param Loc Location of the reference to threadprivate var. |
1117 | /// \return Address of the threadprivate variable for the current thread. |
1118 | virtual Address getAddrOfThreadPrivate(CodeGenFunction &CGF, |
1119 | const VarDecl *VD, |
1120 | Address VDAddr, |
1121 | SourceLocation Loc); |
1122 | |
1123 | /// Returns the address of the variable marked as declare target with link |
1124 | /// clause. |
1125 | virtual Address getAddrOfDeclareTargetLink(const VarDecl *VD); |
1126 | |
1127 | /// Emit a code for initialization of threadprivate variable. It emits |
1128 | /// a call to runtime library which adds initial value to the newly created |
1129 | /// threadprivate variable (if it is not constant) and registers destructor |
1130 | /// for the variable (if any). |
1131 | /// \param VD Threadprivate variable. |
1132 | /// \param VDAddr Address of the global variable \a VD. |
1133 | /// \param Loc Location of threadprivate declaration. |
1134 | /// \param PerformInit true if initialization expression is not constant. |
1135 | virtual llvm::Function * |
1136 | emitThreadPrivateVarDefinition(const VarDecl *VD, Address VDAddr, |
1137 | SourceLocation Loc, bool PerformInit, |
1138 | CodeGenFunction *CGF = nullptr); |
1139 | |
1140 | /// Emit a code for initialization of declare target variable. |
1141 | /// \param VD Declare target variable. |
1142 | /// \param Addr Address of the global variable \a VD. |
1143 | /// \param PerformInit true if initialization expression is not constant. |
1144 | virtual bool emitDeclareTargetVarDefinition(const VarDecl *VD, |
1145 | llvm::GlobalVariable *Addr, |
1146 | bool PerformInit); |
1147 | |
1148 | /// Creates artificial threadprivate variable with name \p Name and type \p |
1149 | /// VarType. |
1150 | /// \param VarType Type of the artificial threadprivate variable. |
1151 | /// \param Name Name of the artificial threadprivate variable. |
1152 | virtual Address getAddrOfArtificialThreadPrivate(CodeGenFunction &CGF, |
1153 | QualType VarType, |
1154 | StringRef Name); |
1155 | |
1156 | /// Emit flush of the variables specified in 'omp flush' directive. |
1157 | /// \param Vars List of variables to flush. |
1158 | virtual void emitFlush(CodeGenFunction &CGF, ArrayRef<const Expr *> Vars, |
1159 | SourceLocation Loc); |
1160 | |
1161 | /// Emit task region for the task directive. The task region is |
1162 | /// emitted in several steps: |
1163 | /// 1. Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 |
1164 | /// gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, |
1165 | /// kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the |
1166 | /// function: |
1167 | /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { |
1168 | /// TaskFunction(gtid, tt->part_id, tt->shareds); |
1169 | /// return 0; |
1170 | /// } |
1171 | /// 2. Copy a list of shared variables to field shareds of the resulting |
1172 | /// structure kmp_task_t returned by the previous call (if any). |
1173 | /// 3. Copy a pointer to destructions function to field destructions of the |
1174 | /// resulting structure kmp_task_t. |
1175 | /// 4. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, |
1176 | /// kmp_task_t *new_task), where new_task is a resulting structure from |
1177 | /// previous items. |
1178 | /// \param D Current task directive. |
1179 | /// \param TaskFunction An LLVM function with type void (*)(i32 /*gtid*/, i32 |
1180 | /// /*part_id*/, captured_struct */*__context*/); |
1181 | /// \param SharedsTy A type which contains references the shared variables. |
1182 | /// \param Shareds Context with the list of shared variables from the \p |
1183 | /// TaskFunction. |
1184 | /// \param IfCond Not a nullptr if 'if' clause was specified, nullptr |
1185 | /// otherwise. |
1186 | /// \param Data Additional data for task generation like tiednsee, final |
1187 | /// state, list of privates etc. |
1188 | virtual void emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc, |
1189 | const OMPExecutableDirective &D, |
1190 | llvm::Function *TaskFunction, QualType SharedsTy, |
1191 | Address Shareds, const Expr *IfCond, |
1192 | const OMPTaskDataTy &Data); |
1193 | |
1194 | /// Emit task region for the taskloop directive. The taskloop region is |
1195 | /// emitted in several steps: |
1196 | /// 1. Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 |
1197 | /// gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, |
1198 | /// kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the |
1199 | /// function: |
1200 | /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { |
1201 | /// TaskFunction(gtid, tt->part_id, tt->shareds); |
1202 | /// return 0; |
1203 | /// } |
1204 | /// 2. Copy a list of shared variables to field shareds of the resulting |
1205 | /// structure kmp_task_t returned by the previous call (if any). |
1206 | /// 3. Copy a pointer to destructions function to field destructions of the |
1207 | /// resulting structure kmp_task_t. |
1208 | /// 4. Emit a call to void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t |
1209 | /// *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int |
1210 | /// nogroup, int sched, kmp_uint64 grainsize, void *task_dup ), where new_task |
1211 | /// is a resulting structure from |
1212 | /// previous items. |
1213 | /// \param D Current task directive. |
1214 | /// \param TaskFunction An LLVM function with type void (*)(i32 /*gtid*/, i32 |
1215 | /// /*part_id*/, captured_struct */*__context*/); |
1216 | /// \param SharedsTy A type which contains references the shared variables. |
1217 | /// \param Shareds Context with the list of shared variables from the \p |
1218 | /// TaskFunction. |
1219 | /// \param IfCond Not a nullptr if 'if' clause was specified, nullptr |
1220 | /// otherwise. |
1221 | /// \param Data Additional data for task generation like tiednsee, final |
1222 | /// state, list of privates etc. |
1223 | virtual void emitTaskLoopCall(CodeGenFunction &CGF, SourceLocation Loc, |
1224 | const OMPLoopDirective &D, |
1225 | llvm::Function *TaskFunction, |
1226 | QualType SharedsTy, Address Shareds, |
1227 | const Expr *IfCond, const OMPTaskDataTy &Data); |
1228 | |
1229 | /// Emit code for the directive that does not require outlining. |
1230 | /// |
1231 | /// \param InnermostKind Kind of innermost directive (for simple directives it |
1232 | /// is a directive itself, for combined - its innermost directive). |
1233 | /// \param CodeGen Code generation sequence for the \a D directive. |
1234 | /// \param HasCancel true if region has inner cancel directive, false |
1235 | /// otherwise. |
1236 | virtual void emitInlinedDirective(CodeGenFunction &CGF, |
1237 | OpenMPDirectiveKind InnermostKind, |
1238 | const RegionCodeGenTy &CodeGen, |
1239 | bool HasCancel = false); |
1240 | |
1241 | /// Emits reduction function. |
1242 | /// \param ArgsType Array type containing pointers to reduction variables. |
1243 | /// \param Privates List of private copies for original reduction arguments. |
1244 | /// \param LHSExprs List of LHS in \a ReductionOps reduction operations. |
1245 | /// \param RHSExprs List of RHS in \a ReductionOps reduction operations. |
1246 | /// \param ReductionOps List of reduction operations in form 'LHS binop RHS' |
1247 | /// or 'operator binop(LHS, RHS)'. |
1248 | llvm::Function *emitReductionFunction(SourceLocation Loc, |
1249 | llvm::Type *ArgsType, |
1250 | ArrayRef<const Expr *> Privates, |
1251 | ArrayRef<const Expr *> LHSExprs, |
1252 | ArrayRef<const Expr *> RHSExprs, |
1253 | ArrayRef<const Expr *> ReductionOps); |
1254 | |
1255 | /// Emits single reduction combiner |
1256 | void emitSingleReductionCombiner(CodeGenFunction &CGF, |
1257 | const Expr *ReductionOp, |
1258 | const Expr *PrivateRef, |
1259 | const DeclRefExpr *LHS, |
1260 | const DeclRefExpr *RHS); |
1261 | |
1262 | struct ReductionOptionsTy { |
1263 | bool WithNowait; |
1264 | bool SimpleReduction; |
1265 | OpenMPDirectiveKind ReductionKind; |
1266 | }; |
1267 | /// Emit a code for reduction clause. Next code should be emitted for |
1268 | /// reduction: |
1269 | /// \code |
1270 | /// |
1271 | /// static kmp_critical_name lock = { 0 }; |
1272 | /// |
1273 | /// void reduce_func(void *lhs[<n>], void *rhs[<n>]) { |
1274 | /// ... |
1275 | /// *(Type<i>*)lhs[i] = RedOp<i>(*(Type<i>*)lhs[i], *(Type<i>*)rhs[i]); |
1276 | /// ... |
1277 | /// } |
1278 | /// |
1279 | /// ... |
1280 | /// void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; |
1281 | /// switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), |
1282 | /// RedList, reduce_func, &<lock>)) { |
1283 | /// case 1: |
1284 | /// ... |
1285 | /// <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]); |
1286 | /// ... |
1287 | /// __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); |
1288 | /// break; |
1289 | /// case 2: |
1290 | /// ... |
1291 | /// Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i])); |
1292 | /// ... |
1293 | /// break; |
1294 | /// default:; |
1295 | /// } |
1296 | /// \endcode |
1297 | /// |
1298 | /// \param Privates List of private copies for original reduction arguments. |
1299 | /// \param LHSExprs List of LHS in \a ReductionOps reduction operations. |
1300 | /// \param RHSExprs List of RHS in \a ReductionOps reduction operations. |
1301 | /// \param ReductionOps List of reduction operations in form 'LHS binop RHS' |
1302 | /// or 'operator binop(LHS, RHS)'. |
1303 | /// \param Options List of options for reduction codegen: |
1304 | /// WithNowait true if parent directive has also nowait clause, false |
1305 | /// otherwise. |
1306 | /// SimpleReduction Emit reduction operation only. Used for omp simd |
1307 | /// directive on the host. |
1308 | /// ReductionKind The kind of reduction to perform. |
1309 | virtual void emitReduction(CodeGenFunction &CGF, SourceLocation Loc, |
1310 | ArrayRef<const Expr *> Privates, |
1311 | ArrayRef<const Expr *> LHSExprs, |
1312 | ArrayRef<const Expr *> RHSExprs, |
1313 | ArrayRef<const Expr *> ReductionOps, |
1314 | ReductionOptionsTy Options); |
1315 | |
1316 | /// Emit a code for initialization of task reduction clause. Next code |
1317 | /// should be emitted for reduction: |
1318 | /// \code |
1319 | /// |
1320 | /// _task_red_item_t red_data[n]; |
1321 | /// ... |
1322 | /// red_data[i].shar = &origs[i]; |
1323 | /// red_data[i].size = sizeof(origs[i]); |
1324 | /// red_data[i].f_init = (void*)RedInit<i>; |
1325 | /// red_data[i].f_fini = (void*)RedDest<i>; |
1326 | /// red_data[i].f_comb = (void*)RedOp<i>; |
1327 | /// red_data[i].flags = <Flag_i>; |
1328 | /// ... |
1329 | /// void* tg1 = __kmpc_task_reduction_init(gtid, n, red_data); |
1330 | /// \endcode |
1331 | /// |
1332 | /// \param LHSExprs List of LHS in \a Data.ReductionOps reduction operations. |
1333 | /// \param RHSExprs List of RHS in \a Data.ReductionOps reduction operations. |
1334 | /// \param Data Additional data for task generation like tiedness, final |
1335 | /// state, list of privates, reductions etc. |
1336 | virtual llvm::Value *emitTaskReductionInit(CodeGenFunction &CGF, |
1337 | SourceLocation Loc, |
1338 | ArrayRef<const Expr *> LHSExprs, |
1339 | ArrayRef<const Expr *> RHSExprs, |
1340 | const OMPTaskDataTy &Data); |
1341 | |
1342 | /// Required to resolve existing problems in the runtime. Emits threadprivate |
1343 | /// variables to store the size of the VLAs/array sections for |
1344 | /// initializer/combiner/finalizer functions + emits threadprivate variable to |
1345 | /// store the pointer to the original reduction item for the custom |
1346 | /// initializer defined by declare reduction construct. |
1347 | /// \param RCG Allows to reuse an existing data for the reductions. |
1348 | /// \param N Reduction item for which fixups must be emitted. |
1349 | virtual void emitTaskReductionFixups(CodeGenFunction &CGF, SourceLocation Loc, |
1350 | ReductionCodeGen &RCG, unsigned N); |
1351 | |
1352 | /// Get the address of `void *` type of the privatue copy of the reduction |
1353 | /// item specified by the \p SharedLVal. |
1354 | /// \param ReductionsPtr Pointer to the reduction data returned by the |
1355 | /// emitTaskReductionInit function. |
1356 | /// \param SharedLVal Address of the original reduction item. |
1357 | virtual Address getTaskReductionItem(CodeGenFunction &CGF, SourceLocation Loc, |
1358 | llvm::Value *ReductionsPtr, |
1359 | LValue SharedLVal); |
1360 | |
1361 | /// Emit code for 'taskwait' directive. |
1362 | virtual void emitTaskwaitCall(CodeGenFunction &CGF, SourceLocation Loc); |
1363 | |
1364 | /// Emit code for 'cancellation point' construct. |
1365 | /// \param CancelRegion Region kind for which the cancellation point must be |
1366 | /// emitted. |
1367 | /// |
1368 | virtual void emitCancellationPointCall(CodeGenFunction &CGF, |
1369 | SourceLocation Loc, |
1370 | OpenMPDirectiveKind CancelRegion); |
1371 | |
1372 | /// Emit code for 'cancel' construct. |
1373 | /// \param IfCond Condition in the associated 'if' clause, if it was |
1374 | /// specified, nullptr otherwise. |
1375 | /// \param CancelRegion Region kind for which the cancel must be emitted. |
1376 | /// |
1377 | virtual void emitCancelCall(CodeGenFunction &CGF, SourceLocation Loc, |
1378 | const Expr *IfCond, |
1379 | OpenMPDirectiveKind CancelRegion); |
1380 | |
1381 | /// Emit outilined function for 'target' directive. |
1382 | /// \param D Directive to emit. |
1383 | /// \param ParentName Name of the function that encloses the target region. |
1384 | /// \param OutlinedFn Outlined function value to be defined by this call. |
1385 | /// \param OutlinedFnID Outlined function ID value to be defined by this call. |
1386 | /// \param IsOffloadEntry True if the outlined function is an offload entry. |
1387 | /// \param CodeGen Code generation sequence for the \a D directive. |
1388 | /// An outlined function may not be an entry if, e.g. the if clause always |
1389 | /// evaluates to false. |
1390 | virtual void emitTargetOutlinedFunction(const OMPExecutableDirective &D, |
1391 | StringRef ParentName, |
1392 | llvm::Function *&OutlinedFn, |
1393 | llvm::Constant *&OutlinedFnID, |
1394 | bool IsOffloadEntry, |
1395 | const RegionCodeGenTy &CodeGen); |
1396 | |
1397 | /// Emit code that pushes the trip count of loops associated with constructs |
1398 | /// 'target teams distribute' and 'teams distribute parallel for'. |
1399 | /// \param SizeEmitter Emits the int64 value for the number of iterations of |
1400 | /// the associated loop. |
1401 | virtual void emitTargetNumIterationsCall( |
1402 | CodeGenFunction &CGF, const OMPExecutableDirective &D, const Expr *Device, |
1403 | const llvm::function_ref<llvm::Value *( |
1404 | CodeGenFunction &CGF, const OMPLoopDirective &D)> &SizeEmitter); |
1405 | |
1406 | /// Emit the target offloading code associated with \a D. The emitted |
1407 | /// code attempts offloading the execution to the device, an the event of |
1408 | /// a failure it executes the host version outlined in \a OutlinedFn. |
1409 | /// \param D Directive to emit. |
1410 | /// \param OutlinedFn Host version of the code to be offloaded. |
1411 | /// \param OutlinedFnID ID of host version of the code to be offloaded. |
1412 | /// \param IfCond Expression evaluated in if clause associated with the target |
1413 | /// directive, or null if no if clause is used. |
1414 | /// \param Device Expression evaluated in device clause associated with the |
1415 | /// target directive, or null if no device clause is used. |
1416 | virtual void emitTargetCall(CodeGenFunction &CGF, |
1417 | const OMPExecutableDirective &D, |
1418 | llvm::Function *OutlinedFn, |
1419 | llvm::Value *OutlinedFnID, const Expr *IfCond, |
1420 | const Expr *Device); |
1421 | |
1422 | /// Emit the target regions enclosed in \a GD function definition or |
1423 | /// the function itself in case it is a valid device function. Returns true if |
1424 | /// \a GD was dealt with successfully. |
1425 | /// \param GD Function to scan. |
1426 | virtual bool emitTargetFunctions(GlobalDecl GD); |
1427 | |
1428 | /// Emit the global variable if it is a valid device global variable. |
1429 | /// Returns true if \a GD was dealt with successfully. |
1430 | /// \param GD Variable declaration to emit. |
1431 | virtual bool emitTargetGlobalVariable(GlobalDecl GD); |
1432 | |
1433 | /// Checks if the provided global decl \a GD is a declare target variable and |
1434 | /// registers it when emitting code for the host. |
1435 | virtual void registerTargetGlobalVariable(const VarDecl *VD, |
1436 | llvm::Constant *Addr); |
1437 | |
1438 | /// Registers provided target firstprivate variable as global on the |
1439 | /// target. |
1440 | llvm::Constant *registerTargetFirstprivateCopy(CodeGenFunction &CGF, |
1441 | const VarDecl *VD); |
1442 | |
1443 | /// Emit the global \a GD if it is meaningful for the target. Returns |
1444 | /// if it was emitted successfully. |
1445 | /// \param GD Global to scan. |
1446 | virtual bool emitTargetGlobal(GlobalDecl GD); |
1447 | |
1448 | /// Creates and returns a registration function for when at least one |
1449 | /// requires directives was used in the current module. |
1450 | llvm::Function *emitRequiresDirectiveRegFun(); |
1451 | |
1452 | /// Creates the offloading descriptor in the event any target region |
1453 | /// was emitted in the current module and return the function that registers |
1454 | /// it. |
1455 | virtual llvm::Function *emitRegistrationFunction(); |
1456 | |
1457 | /// Emits code for teams call of the \a OutlinedFn with |
1458 | /// variables captured in a record which address is stored in \a |
1459 | /// CapturedStruct. |
1460 | /// \param OutlinedFn Outlined function to be run by team masters. Type of |
1461 | /// this function is void(*)(kmp_int32 *, kmp_int32, struct context_vars*). |
1462 | /// \param CapturedVars A pointer to the record with the references to |
1463 | /// variables used in \a OutlinedFn function. |
1464 | /// |
1465 | virtual void emitTeamsCall(CodeGenFunction &CGF, |
1466 | const OMPExecutableDirective &D, |
1467 | SourceLocation Loc, llvm::Function *OutlinedFn, |
1468 | ArrayRef<llvm::Value *> CapturedVars); |
1469 | |
1470 | /// Emits call to void __kmpc_push_num_teams(ident_t *loc, kmp_int32 |
1471 | /// global_tid, kmp_int32 num_teams, kmp_int32 thread_limit) to generate code |
1472 | /// for num_teams clause. |
1473 | /// \param NumTeams An integer expression of teams. |
1474 | /// \param ThreadLimit An integer expression of threads. |
1475 | virtual void emitNumTeamsClause(CodeGenFunction &CGF, const Expr *NumTeams, |
1476 | const Expr *ThreadLimit, SourceLocation Loc); |
1477 | |
1478 | /// Struct that keeps all the relevant information that should be kept |
1479 | /// throughout a 'target data' region. |
1480 | class TargetDataInfo { |
1481 | /// Set to true if device pointer information have to be obtained. |
1482 | bool RequiresDevicePointerInfo = false; |
1483 | |
1484 | public: |
1485 | /// The array of base pointer passed to the runtime library. |
1486 | llvm::Value *BasePointersArray = nullptr; |
1487 | /// The array of section pointers passed to the runtime library. |
1488 | llvm::Value *PointersArray = nullptr; |
1489 | /// The array of sizes passed to the runtime library. |
1490 | llvm::Value *SizesArray = nullptr; |
1491 | /// The array of map types passed to the runtime library. |
1492 | llvm::Value *MapTypesArray = nullptr; |
1493 | /// The total number of pointers passed to the runtime library. |
1494 | unsigned NumberOfPtrs = 0u; |
1495 | /// Map between the a declaration of a capture and the corresponding base |
1496 | /// pointer address where the runtime returns the device pointers. |
1497 | llvm::DenseMap<const ValueDecl *, Address> CaptureDeviceAddrMap; |
1498 | |
1499 | explicit TargetDataInfo() {} |
1500 | explicit TargetDataInfo(bool RequiresDevicePointerInfo) |
1501 | : RequiresDevicePointerInfo(RequiresDevicePointerInfo) {} |
1502 | /// Clear information about the data arrays. |
1503 | void clearArrayInfo() { |
1504 | BasePointersArray = nullptr; |
1505 | PointersArray = nullptr; |
1506 | SizesArray = nullptr; |
1507 | MapTypesArray = nullptr; |
1508 | NumberOfPtrs = 0u; |
1509 | } |
1510 | /// Return true if the current target data information has valid arrays. |
1511 | bool isValid() { |
1512 | return BasePointersArray && PointersArray && SizesArray && |
1513 | MapTypesArray && NumberOfPtrs; |
1514 | } |
1515 | bool requiresDevicePointerInfo() { return RequiresDevicePointerInfo; } |
1516 | }; |
1517 | |
1518 | /// Emit the target data mapping code associated with \a D. |
1519 | /// \param D Directive to emit. |
1520 | /// \param IfCond Expression evaluated in if clause associated with the |
1521 | /// target directive, or null if no device clause is used. |
1522 | /// \param Device Expression evaluated in device clause associated with the |
1523 | /// target directive, or null if no device clause is used. |
1524 | /// \param Info A record used to store information that needs to be preserved |
1525 | /// until the region is closed. |
1526 | virtual void emitTargetDataCalls(CodeGenFunction &CGF, |
1527 | const OMPExecutableDirective &D, |
1528 | const Expr *IfCond, const Expr *Device, |
1529 | const RegionCodeGenTy &CodeGen, |
1530 | TargetDataInfo &Info); |
1531 | |
1532 | /// Emit the data mapping/movement code associated with the directive |
1533 | /// \a D that should be of the form 'target [{enter|exit} data | update]'. |
1534 | /// \param D Directive to emit. |
1535 | /// \param IfCond Expression evaluated in if clause associated with the target |
1536 | /// directive, or null if no if clause is used. |
1537 | /// \param Device Expression evaluated in device clause associated with the |
1538 | /// target directive, or null if no device clause is used. |
1539 | virtual void emitTargetDataStandAloneCall(CodeGenFunction &CGF, |
1540 | const OMPExecutableDirective &D, |
1541 | const Expr *IfCond, |
1542 | const Expr *Device); |
1543 | |
1544 | /// Marks function \a Fn with properly mangled versions of vector functions. |
1545 | /// \param FD Function marked as 'declare simd'. |
1546 | /// \param Fn LLVM function that must be marked with 'declare simd' |
1547 | /// attributes. |
1548 | virtual void emitDeclareSimdFunction(const FunctionDecl *FD, |
1549 | llvm::Function *Fn); |
1550 | |
1551 | /// Emit initialization for doacross loop nesting support. |
1552 | /// \param D Loop-based construct used in doacross nesting construct. |
1553 | virtual void emitDoacrossInit(CodeGenFunction &CGF, const OMPLoopDirective &D, |
1554 | ArrayRef<Expr *> NumIterations); |
1555 | |
1556 | /// Emit code for doacross ordered directive with 'depend' clause. |
1557 | /// \param C 'depend' clause with 'sink|source' dependency kind. |
1558 | virtual void emitDoacrossOrdered(CodeGenFunction &CGF, |
1559 | const OMPDependClause *C); |
1560 | |
1561 | /// Translates the native parameter of outlined function if this is required |
1562 | /// for target. |
1563 | /// \param FD Field decl from captured record for the parameter. |
1564 | /// \param NativeParam Parameter itself. |
1565 | virtual const VarDecl *translateParameter(const FieldDecl *FD, |
1566 | const VarDecl *NativeParam) const { |
1567 | return NativeParam; |
1568 | } |
1569 | |
1570 | /// Gets the address of the native argument basing on the address of the |
1571 | /// target-specific parameter. |
1572 | /// \param NativeParam Parameter itself. |
1573 | /// \param TargetParam Corresponding target-specific parameter. |
1574 | virtual Address getParameterAddress(CodeGenFunction &CGF, |
1575 | const VarDecl *NativeParam, |
1576 | const VarDecl *TargetParam) const; |
1577 | |
1578 | /// Choose default schedule type and chunk value for the |
1579 | /// dist_schedule clause. |
1580 | virtual void getDefaultDistScheduleAndChunk(CodeGenFunction &CGF, |
1581 | const OMPLoopDirective &S, OpenMPDistScheduleClauseKind &ScheduleKind, |
1582 | llvm::Value *&Chunk) const {} |
1583 | |
1584 | /// Choose default schedule type and chunk value for the |
1585 | /// schedule clause. |
1586 | virtual void getDefaultScheduleAndChunk(CodeGenFunction &CGF, |
1587 | const OMPLoopDirective &S, OpenMPScheduleClauseKind &ScheduleKind, |
1588 | const Expr *&ChunkExpr) const; |
1589 | |
1590 | /// Emits call of the outlined function with the provided arguments, |
1591 | /// translating these arguments to correct target-specific arguments. |
1592 | virtual void |
1593 | emitOutlinedFunctionCall(CodeGenFunction &CGF, SourceLocation Loc, |
1594 | llvm::FunctionCallee OutlinedFn, |
1595 | ArrayRef<llvm::Value *> Args = llvm::None) const; |
1596 | |
1597 | /// Emits OpenMP-specific function prolog. |
1598 | /// Required for device constructs. |
1599 | virtual void emitFunctionProlog(CodeGenFunction &CGF, const Decl *D); |
1600 | |
1601 | /// Gets the OpenMP-specific address of the local variable. |
1602 | virtual Address getAddressOfLocalVariable(CodeGenFunction &CGF, |
1603 | const VarDecl *VD); |
1604 | |
1605 | /// Marks the declaration as already emitted for the device code and returns |
1606 | /// true, if it was marked already, and false, otherwise. |
1607 | bool markAsGlobalTarget(GlobalDecl GD); |
1608 | |
1609 | /// Emit deferred declare target variables marked for deferred emission. |
1610 | void emitDeferredTargetDecls() const; |
1611 | |
1612 | /// Adjust some parameters for the target-based directives, like addresses of |
1613 | /// the variables captured by reference in lambdas. |
1614 | virtual void |
1615 | adjustTargetSpecificDataForLambdas(CodeGenFunction &CGF, |
1616 | const OMPExecutableDirective &D) const; |
1617 | |
1618 | /// Perform check on requires decl to ensure that target architecture |
1619 | /// supports unified addressing |
1620 | virtual void checkArchForUnifiedAddressing(const OMPRequiresDecl *D); |
1621 | |
1622 | /// Checks if the variable has associated OMPAllocateDeclAttr attribute with |
1623 | /// the predefined allocator and translates it into the corresponding address |
1624 | /// space. |
1625 | virtual bool hasAllocateAttributeForGlobalVar(const VarDecl *VD, LangAS &AS); |
1626 | }; |
1627 | |
1628 | /// Class supports emissionof SIMD-only code. |
1629 | class CGOpenMPSIMDRuntime final : public CGOpenMPRuntime { |
1630 | public: |
1631 | explicit CGOpenMPSIMDRuntime(CodeGenModule &CGM) : CGOpenMPRuntime(CGM) {} |
1632 | ~CGOpenMPSIMDRuntime() override {} |
1633 | |
1634 | /// Emits outlined function for the specified OpenMP parallel directive |
1635 | /// \a D. This outlined function has type void(*)(kmp_int32 *ThreadID, |
1636 | /// kmp_int32 BoundID, struct context_vars*). |
1637 | /// \param D OpenMP directive. |
1638 | /// \param ThreadIDVar Variable for thread id in the current OpenMP region. |
1639 | /// \param InnermostKind Kind of innermost directive (for simple directives it |
1640 | /// is a directive itself, for combined - its innermost directive). |
1641 | /// \param CodeGen Code generation sequence for the \a D directive. |
1642 | llvm::Function * |
1643 | emitParallelOutlinedFunction(const OMPExecutableDirective &D, |
1644 | const VarDecl *ThreadIDVar, |
1645 | OpenMPDirectiveKind InnermostKind, |
1646 | const RegionCodeGenTy &CodeGen) override; |
1647 | |
1648 | /// Emits outlined function for the specified OpenMP teams directive |
1649 | /// \a D. This outlined function has type void(*)(kmp_int32 *ThreadID, |
1650 | /// kmp_int32 BoundID, struct context_vars*). |
1651 | /// \param D OpenMP directive. |
1652 | /// \param ThreadIDVar Variable for thread id in the current OpenMP region. |
1653 | /// \param InnermostKind Kind of innermost directive (for simple directives it |
1654 | /// is a directive itself, for combined - its innermost directive). |
1655 | /// \param CodeGen Code generation sequence for the \a D directive. |
1656 | llvm::Function * |
1657 | emitTeamsOutlinedFunction(const OMPExecutableDirective &D, |
1658 | const VarDecl *ThreadIDVar, |
1659 | OpenMPDirectiveKind InnermostKind, |
1660 | const RegionCodeGenTy &CodeGen) override; |
1661 | |
1662 | /// Emits outlined function for the OpenMP task directive \a D. This |
1663 | /// outlined function has type void(*)(kmp_int32 ThreadID, struct task_t* |
1664 | /// TaskT). |
1665 | /// \param D OpenMP directive. |
1666 | /// \param ThreadIDVar Variable for thread id in the current OpenMP region. |
1667 | /// \param PartIDVar Variable for partition id in the current OpenMP untied |
1668 | /// task region. |
1669 | /// \param TaskTVar Variable for task_t argument. |
1670 | /// \param InnermostKind Kind of innermost directive (for simple directives it |
1671 | /// is a directive itself, for combined - its innermost directive). |
1672 | /// \param CodeGen Code generation sequence for the \a D directive. |
1673 | /// \param Tied true if task is generated for tied task, false otherwise. |
1674 | /// \param NumberOfParts Number of parts in untied task. Ignored for tied |
1675 | /// tasks. |
1676 | /// |
1677 | llvm::Function *emitTaskOutlinedFunction( |
1678 | const OMPExecutableDirective &D, const VarDecl *ThreadIDVar, |
1679 | const VarDecl *PartIDVar, const VarDecl *TaskTVar, |
1680 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen, |
1681 | bool Tied, unsigned &NumberOfParts) override; |
1682 | |
1683 | /// Emits code for parallel or serial call of the \a OutlinedFn with |
1684 | /// variables captured in a record which address is stored in \a |
1685 | /// CapturedStruct. |
1686 | /// \param OutlinedFn Outlined function to be run in parallel threads. Type of |
1687 | /// this function is void(*)(kmp_int32 *, kmp_int32, struct context_vars*). |
1688 | /// \param CapturedVars A pointer to the record with the references to |
1689 | /// variables used in \a OutlinedFn function. |
1690 | /// \param IfCond Condition in the associated 'if' clause, if it was |
1691 | /// specified, nullptr otherwise. |
1692 | /// |
1693 | void emitParallelCall(CodeGenFunction &CGF, SourceLocation Loc, |
1694 | llvm::Function *OutlinedFn, |
1695 | ArrayRef<llvm::Value *> CapturedVars, |
1696 | const Expr *IfCond) override; |
1697 | |
1698 | /// Emits a critical region. |
1699 | /// \param CriticalName Name of the critical region. |
1700 | /// \param CriticalOpGen Generator for the statement associated with the given |
1701 | /// critical region. |
1702 | /// \param Hint Value of the 'hint' clause (optional). |
1703 | void emitCriticalRegion(CodeGenFunction &CGF, StringRef CriticalName, |
1704 | const RegionCodeGenTy &CriticalOpGen, |
1705 | SourceLocation Loc, |
1706 | const Expr *Hint = nullptr) override; |
1707 | |
1708 | /// Emits a master region. |
1709 | /// \param MasterOpGen Generator for the statement associated with the given |
1710 | /// master region. |
1711 | void emitMasterRegion(CodeGenFunction &CGF, |
1712 | const RegionCodeGenTy &MasterOpGen, |
1713 | SourceLocation Loc) override; |
1714 | |
1715 | /// Emits code for a taskyield directive. |
1716 | void emitTaskyieldCall(CodeGenFunction &CGF, SourceLocation Loc) override; |
1717 | |
1718 | /// Emit a taskgroup region. |
1719 | /// \param TaskgroupOpGen Generator for the statement associated with the |
1720 | /// given taskgroup region. |
1721 | void emitTaskgroupRegion(CodeGenFunction &CGF, |
1722 | const RegionCodeGenTy &TaskgroupOpGen, |
1723 | SourceLocation Loc) override; |
1724 | |
1725 | /// Emits a single region. |
1726 | /// \param SingleOpGen Generator for the statement associated with the given |
1727 | /// single region. |
1728 | void emitSingleRegion(CodeGenFunction &CGF, |
1729 | const RegionCodeGenTy &SingleOpGen, SourceLocation Loc, |
1730 | ArrayRef<const Expr *> CopyprivateVars, |
1731 | ArrayRef<const Expr *> DestExprs, |
1732 | ArrayRef<const Expr *> SrcExprs, |
1733 | ArrayRef<const Expr *> AssignmentOps) override; |
1734 | |
1735 | /// Emit an ordered region. |
1736 | /// \param OrderedOpGen Generator for the statement associated with the given |
1737 | /// ordered region. |
1738 | void emitOrderedRegion(CodeGenFunction &CGF, |
1739 | const RegionCodeGenTy &OrderedOpGen, |
1740 | SourceLocation Loc, bool IsThreads) override; |
1741 | |
1742 | /// Emit an implicit/explicit barrier for OpenMP threads. |
1743 | /// \param Kind Directive for which this implicit barrier call must be |
1744 | /// generated. Must be OMPD_barrier for explicit barrier generation. |
1745 | /// \param EmitChecks true if need to emit checks for cancellation barriers. |
1746 | /// \param ForceSimpleCall true simple barrier call must be emitted, false if |
1747 | /// runtime class decides which one to emit (simple or with cancellation |
1748 | /// checks). |
1749 | /// |
1750 | void emitBarrierCall(CodeGenFunction &CGF, SourceLocation Loc, |
1751 | OpenMPDirectiveKind Kind, bool EmitChecks = true, |
1752 | bool ForceSimpleCall = false) override; |
1753 | |
1754 | /// This is used for non static scheduled types and when the ordered |
1755 | /// clause is present on the loop construct. |
1756 | /// Depending on the loop schedule, it is necessary to call some runtime |
1757 | /// routine before start of the OpenMP loop to get the loop upper / lower |
1758 | /// bounds \a LB and \a UB and stride \a ST. |
1759 | /// |
1760 | /// \param CGF Reference to current CodeGenFunction. |
1761 | /// \param Loc Clang source location. |
1762 | /// \param ScheduleKind Schedule kind, specified by the 'schedule' clause. |
1763 | /// \param IVSize Size of the iteration variable in bits. |
1764 | /// \param IVSigned Sign of the iteration variable. |
1765 | /// \param Ordered true if loop is ordered, false otherwise. |
1766 | /// \param DispatchValues struct containing llvm values for lower bound, upper |
1767 | /// bound, and chunk expression. |
1768 | /// For the default (nullptr) value, the chunk 1 will be used. |
1769 | /// |
1770 | void emitForDispatchInit(CodeGenFunction &CGF, SourceLocation Loc, |
1771 | const OpenMPScheduleTy &ScheduleKind, |
1772 | unsigned IVSize, bool IVSigned, bool Ordered, |
1773 | const DispatchRTInput &DispatchValues) override; |
1774 | |
1775 | /// Call the appropriate runtime routine to initialize it before start |
1776 | /// of loop. |
1777 | /// |
1778 | /// This is used only in case of static schedule, when the user did not |
1779 | /// specify a ordered clause on the loop construct. |
1780 | /// Depending on the loop schedule, it is necessary to call some runtime |
1781 | /// routine before start of the OpenMP loop to get the loop upper / lower |
1782 | /// bounds LB and UB and stride ST. |
1783 | /// |
1784 | /// \param CGF Reference to current CodeGenFunction. |
1785 | /// \param Loc Clang source location. |
1786 | /// \param DKind Kind of the directive. |
1787 | /// \param ScheduleKind Schedule kind, specified by the 'schedule' clause. |
1788 | /// \param Values Input arguments for the construct. |
1789 | /// |
1790 | void emitForStaticInit(CodeGenFunction &CGF, SourceLocation Loc, |
1791 | OpenMPDirectiveKind DKind, |
1792 | const OpenMPScheduleTy &ScheduleKind, |
1793 | const StaticRTInput &Values) override; |
1794 | |
1795 | /// |
1796 | /// \param CGF Reference to current CodeGenFunction. |
1797 | /// \param Loc Clang source location. |
1798 | /// \param SchedKind Schedule kind, specified by the 'dist_schedule' clause. |
1799 | /// \param Values Input arguments for the construct. |
1800 | /// |
1801 | void emitDistributeStaticInit(CodeGenFunction &CGF, SourceLocation Loc, |
1802 | OpenMPDistScheduleClauseKind SchedKind, |
1803 | const StaticRTInput &Values) override; |
1804 | |
1805 | /// Call the appropriate runtime routine to notify that we finished |
1806 | /// iteration of the ordered loop with the dynamic scheduling. |
1807 | /// |
1808 | /// \param CGF Reference to current CodeGenFunction. |
1809 | /// \param Loc Clang source location. |
1810 | /// \param IVSize Size of the iteration variable in bits. |
1811 | /// \param IVSigned Sign of the iteration variable. |
1812 | /// |
1813 | void emitForOrderedIterationEnd(CodeGenFunction &CGF, SourceLocation Loc, |
1814 | unsigned IVSize, bool IVSigned) override; |
1815 | |
1816 | /// Call the appropriate runtime routine to notify that we finished |
1817 | /// all the work with current loop. |
1818 | /// |
1819 | /// \param CGF Reference to current CodeGenFunction. |
1820 | /// \param Loc Clang source location. |
1821 | /// \param DKind Kind of the directive for which the static finish is emitted. |
1822 | /// |
1823 | void emitForStaticFinish(CodeGenFunction &CGF, SourceLocation Loc, |
1824 | OpenMPDirectiveKind DKind) override; |
1825 | |
1826 | /// Call __kmpc_dispatch_next( |
1827 | /// ident_t *loc, kmp_int32 tid, kmp_int32 *p_lastiter, |
1828 | /// kmp_int[32|64] *p_lower, kmp_int[32|64] *p_upper, |
1829 | /// kmp_int[32|64] *p_stride); |
1830 | /// \param IVSize Size of the iteration variable in bits. |
1831 | /// \param IVSigned Sign of the iteration variable. |
1832 | /// \param IL Address of the output variable in which the flag of the |
1833 | /// last iteration is returned. |
1834 | /// \param LB Address of the output variable in which the lower iteration |
1835 | /// number is returned. |
1836 | /// \param UB Address of the output variable in which the upper iteration |
1837 | /// number is returned. |
1838 | /// \param ST Address of the output variable in which the stride value is |
1839 | /// returned. |
1840 | llvm::Value *emitForNext(CodeGenFunction &CGF, SourceLocation Loc, |
1841 | unsigned IVSize, bool IVSigned, Address IL, |
1842 | Address LB, Address UB, Address ST) override; |
1843 | |
1844 | /// Emits call to void __kmpc_push_num_threads(ident_t *loc, kmp_int32 |
1845 | /// global_tid, kmp_int32 num_threads) to generate code for 'num_threads' |
1846 | /// clause. |
1847 | /// \param NumThreads An integer value of threads. |
1848 | void emitNumThreadsClause(CodeGenFunction &CGF, llvm::Value *NumThreads, |
1849 | SourceLocation Loc) override; |
1850 | |
1851 | /// Emit call to void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 |
1852 | /// global_tid, int proc_bind) to generate code for 'proc_bind' clause. |
1853 | void emitProcBindClause(CodeGenFunction &CGF, |
1854 | OpenMPProcBindClauseKind ProcBind, |
1855 | SourceLocation Loc) override; |
1856 | |
1857 | /// Returns address of the threadprivate variable for the current |
1858 | /// thread. |
1859 | /// \param VD Threadprivate variable. |
1860 | /// \param VDAddr Address of the global variable \a VD. |
1861 | /// \param Loc Location of the reference to threadprivate var. |
1862 | /// \return Address of the threadprivate variable for the current thread. |
1863 | Address getAddrOfThreadPrivate(CodeGenFunction &CGF, const VarDecl *VD, |
1864 | Address VDAddr, SourceLocation Loc) override; |
1865 | |
1866 | /// Emit a code for initialization of threadprivate variable. It emits |
1867 | /// a call to runtime library which adds initial value to the newly created |
1868 | /// threadprivate variable (if it is not constant) and registers destructor |
1869 | /// for the variable (if any). |
1870 | /// \param VD Threadprivate variable. |
1871 | /// \param VDAddr Address of the global variable \a VD. |
1872 | /// \param Loc Location of threadprivate declaration. |
1873 | /// \param PerformInit true if initialization expression is not constant. |
1874 | llvm::Function * |
1875 | emitThreadPrivateVarDefinition(const VarDecl *VD, Address VDAddr, |
1876 | SourceLocation Loc, bool PerformInit, |
1877 | CodeGenFunction *CGF = nullptr) override; |
1878 | |
1879 | /// Creates artificial threadprivate variable with name \p Name and type \p |
1880 | /// VarType. |
1881 | /// \param VarType Type of the artificial threadprivate variable. |
1882 | /// \param Name Name of the artificial threadprivate variable. |
1883 | Address getAddrOfArtificialThreadPrivate(CodeGenFunction &CGF, |
1884 | QualType VarType, |
1885 | StringRef Name) override; |
1886 | |
1887 | /// Emit flush of the variables specified in 'omp flush' directive. |
1888 | /// \param Vars List of variables to flush. |
1889 | void emitFlush(CodeGenFunction &CGF, ArrayRef<const Expr *> Vars, |
1890 | SourceLocation Loc) override; |
1891 | |
1892 | /// Emit task region for the task directive. The task region is |
1893 | /// emitted in several steps: |
1894 | /// 1. Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 |
1895 | /// gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, |
1896 | /// kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the |
1897 | /// function: |
1898 | /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { |
1899 | /// TaskFunction(gtid, tt->part_id, tt->shareds); |
1900 | /// return 0; |
1901 | /// } |
1902 | /// 2. Copy a list of shared variables to field shareds of the resulting |
1903 | /// structure kmp_task_t returned by the previous call (if any). |
1904 | /// 3. Copy a pointer to destructions function to field destructions of the |
1905 | /// resulting structure kmp_task_t. |
1906 | /// 4. Emit a call to kmp_int32 __kmpc_omp_task(ident_t *, kmp_int32 gtid, |
1907 | /// kmp_task_t *new_task), where new_task is a resulting structure from |
1908 | /// previous items. |
1909 | /// \param D Current task directive. |
1910 | /// \param TaskFunction An LLVM function with type void (*)(i32 /*gtid*/, i32 |
1911 | /// /*part_id*/, captured_struct */*__context*/); |
1912 | /// \param SharedsTy A type which contains references the shared variables. |
1913 | /// \param Shareds Context with the list of shared variables from the \p |
1914 | /// TaskFunction. |
1915 | /// \param IfCond Not a nullptr if 'if' clause was specified, nullptr |
1916 | /// otherwise. |
1917 | /// \param Data Additional data for task generation like tiednsee, final |
1918 | /// state, list of privates etc. |
1919 | void emitTaskCall(CodeGenFunction &CGF, SourceLocation Loc, |
1920 | const OMPExecutableDirective &D, |
1921 | llvm::Function *TaskFunction, QualType SharedsTy, |
1922 | Address Shareds, const Expr *IfCond, |
1923 | const OMPTaskDataTy &Data) override; |
1924 | |
1925 | /// Emit task region for the taskloop directive. The taskloop region is |
1926 | /// emitted in several steps: |
1927 | /// 1. Emit a call to kmp_task_t *__kmpc_omp_task_alloc(ident_t *, kmp_int32 |
1928 | /// gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, size_t sizeof_shareds, |
1929 | /// kmp_routine_entry_t *task_entry). Here task_entry is a pointer to the |
1930 | /// function: |
1931 | /// kmp_int32 .omp_task_entry.(kmp_int32 gtid, kmp_task_t *tt) { |
1932 | /// TaskFunction(gtid, tt->part_id, tt->shareds); |
1933 | /// return 0; |
1934 | /// } |
1935 | /// 2. Copy a list of shared variables to field shareds of the resulting |
1936 | /// structure kmp_task_t returned by the previous call (if any). |
1937 | /// 3. Copy a pointer to destructions function to field destructions of the |
1938 | /// resulting structure kmp_task_t. |
1939 | /// 4. Emit a call to void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t |
1940 | /// *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int |
1941 | /// nogroup, int sched, kmp_uint64 grainsize, void *task_dup ), where new_task |
1942 | /// is a resulting structure from |
1943 | /// previous items. |
1944 | /// \param D Current task directive. |
1945 | /// \param TaskFunction An LLVM function with type void (*)(i32 /*gtid*/, i32 |
1946 | /// /*part_id*/, captured_struct */*__context*/); |
1947 | /// \param SharedsTy A type which contains references the shared variables. |
1948 | /// \param Shareds Context with the list of shared variables from the \p |
1949 | /// TaskFunction. |
1950 | /// \param IfCond Not a nullptr if 'if' clause was specified, nullptr |
1951 | /// otherwise. |
1952 | /// \param Data Additional data for task generation like tiednsee, final |
1953 | /// state, list of privates etc. |
1954 | void emitTaskLoopCall(CodeGenFunction &CGF, SourceLocation Loc, |
1955 | const OMPLoopDirective &D, llvm::Function *TaskFunction, |
1956 | QualType SharedsTy, Address Shareds, const Expr *IfCond, |
1957 | const OMPTaskDataTy &Data) override; |
1958 | |
1959 | /// Emit a code for reduction clause. Next code should be emitted for |
1960 | /// reduction: |
1961 | /// \code |
1962 | /// |
1963 | /// static kmp_critical_name lock = { 0 }; |
1964 | /// |
1965 | /// void reduce_func(void *lhs[<n>], void *rhs[<n>]) { |
1966 | /// ... |
1967 | /// *(Type<i>*)lhs[i] = RedOp<i>(*(Type<i>*)lhs[i], *(Type<i>*)rhs[i]); |
1968 | /// ... |
1969 | /// } |
1970 | /// |
1971 | /// ... |
1972 | /// void *RedList[<n>] = {&<RHSExprs>[0], ..., &<RHSExprs>[<n>-1]}; |
1973 | /// switch (__kmpc_reduce{_nowait}(<loc>, <gtid>, <n>, sizeof(RedList), |
1974 | /// RedList, reduce_func, &<lock>)) { |
1975 | /// case 1: |
1976 | /// ... |
1977 | /// <LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i]); |
1978 | /// ... |
1979 | /// __kmpc_end_reduce{_nowait}(<loc>, <gtid>, &<lock>); |
1980 | /// break; |
1981 | /// case 2: |
1982 | /// ... |
1983 | /// Atomic(<LHSExprs>[i] = RedOp<i>(*<LHSExprs>[i], *<RHSExprs>[i])); |
1984 | /// ... |
1985 | /// break; |
1986 | /// default:; |
1987 | /// } |
1988 | /// \endcode |
1989 | /// |
1990 | /// \param Privates List of private copies for original reduction arguments. |
1991 | /// \param LHSExprs List of LHS in \a ReductionOps reduction operations. |
1992 | /// \param RHSExprs List of RHS in \a ReductionOps reduction operations. |
1993 | /// \param ReductionOps List of reduction operations in form 'LHS binop RHS' |
1994 | /// or 'operator binop(LHS, RHS)'. |
1995 | /// \param Options List of options for reduction codegen: |
1996 | /// WithNowait true if parent directive has also nowait clause, false |
1997 | /// otherwise. |
1998 | /// SimpleReduction Emit reduction operation only. Used for omp simd |
1999 | /// directive on the host. |
2000 | /// ReductionKind The kind of reduction to perform. |
2001 | void emitReduction(CodeGenFunction &CGF, SourceLocation Loc, |
2002 | ArrayRef<const Expr *> Privates, |
2003 | ArrayRef<const Expr *> LHSExprs, |
2004 | ArrayRef<const Expr *> RHSExprs, |
2005 | ArrayRef<const Expr *> ReductionOps, |
2006 | ReductionOptionsTy Options) override; |
2007 | |
2008 | /// Emit a code for initialization of task reduction clause. Next code |
2009 | /// should be emitted for reduction: |
2010 | /// \code |
2011 | /// |
2012 | /// _task_red_item_t red_data[n]; |
2013 | /// ... |
2014 | /// red_data[i].shar = &origs[i]; |
2015 | /// red_data[i].size = sizeof(origs[i]); |
2016 | /// red_data[i].f_init = (void*)RedInit<i>; |
2017 | /// red_data[i].f_fini = (void*)RedDest<i>; |
2018 | /// red_data[i].f_comb = (void*)RedOp<i>; |
2019 | /// red_data[i].flags = <Flag_i>; |
2020 | /// ... |
2021 | /// void* tg1 = __kmpc_task_reduction_init(gtid, n, red_data); |
2022 | /// \endcode |
2023 | /// |
2024 | /// \param LHSExprs List of LHS in \a Data.ReductionOps reduction operations. |
2025 | /// \param RHSExprs List of RHS in \a Data.ReductionOps reduction operations. |
2026 | /// \param Data Additional data for task generation like tiedness, final |
2027 | /// state, list of privates, reductions etc. |
2028 | llvm::Value *emitTaskReductionInit(CodeGenFunction &CGF, SourceLocation Loc, |
2029 | ArrayRef<const Expr *> LHSExprs, |
2030 | ArrayRef<const Expr *> RHSExprs, |
2031 | const OMPTaskDataTy &Data) override; |
2032 | |
2033 | /// Required to resolve existing problems in the runtime. Emits threadprivate |
2034 | /// variables to store the size of the VLAs/array sections for |
2035 | /// initializer/combiner/finalizer functions + emits threadprivate variable to |
2036 | /// store the pointer to the original reduction item for the custom |
2037 | /// initializer defined by declare reduction construct. |
2038 | /// \param RCG Allows to reuse an existing data for the reductions. |
2039 | /// \param N Reduction item for which fixups must be emitted. |
2040 | void emitTaskReductionFixups(CodeGenFunction &CGF, SourceLocation Loc, |
2041 | ReductionCodeGen &RCG, unsigned N) override; |
2042 | |
2043 | /// Get the address of `void *` type of the privatue copy of the reduction |
2044 | /// item specified by the \p SharedLVal. |
2045 | /// \param ReductionsPtr Pointer to the reduction data returned by the |
2046 | /// emitTaskReductionInit function. |
2047 | /// \param SharedLVal Address of the original reduction item. |
2048 | Address getTaskReductionItem(CodeGenFunction &CGF, SourceLocation Loc, |
2049 | llvm::Value *ReductionsPtr, |
2050 | LValue SharedLVal) override; |
2051 | |
2052 | /// Emit code for 'taskwait' directive. |
2053 | void emitTaskwaitCall(CodeGenFunction &CGF, SourceLocation Loc) override; |
2054 | |
2055 | /// Emit code for 'cancellation point' construct. |
2056 | /// \param CancelRegion Region kind for which the cancellation point must be |
2057 | /// emitted. |
2058 | /// |
2059 | void emitCancellationPointCall(CodeGenFunction &CGF, SourceLocation Loc, |
2060 | OpenMPDirectiveKind CancelRegion) override; |
2061 | |
2062 | /// Emit code for 'cancel' construct. |
2063 | /// \param IfCond Condition in the associated 'if' clause, if it was |
2064 | /// specified, nullptr otherwise. |
2065 | /// \param CancelRegion Region kind for which the cancel must be emitted. |
2066 | /// |
2067 | void emitCancelCall(CodeGenFunction &CGF, SourceLocation Loc, |
2068 | const Expr *IfCond, |
2069 | OpenMPDirectiveKind CancelRegion) override; |
2070 | |
2071 | /// Emit outilined function for 'target' directive. |
2072 | /// \param D Directive to emit. |
2073 | /// \param ParentName Name of the function that encloses the target region. |
2074 | /// \param OutlinedFn Outlined function value to be defined by this call. |
2075 | /// \param OutlinedFnID Outlined function ID value to be defined by this call. |
2076 | /// \param IsOffloadEntry True if the outlined function is an offload entry. |
2077 | /// \param CodeGen Code generation sequence for the \a D directive. |
2078 | /// An outlined function may not be an entry if, e.g. the if clause always |
2079 | /// evaluates to false. |
2080 | void emitTargetOutlinedFunction(const OMPExecutableDirective &D, |
2081 | StringRef ParentName, |
2082 | llvm::Function *&OutlinedFn, |
2083 | llvm::Constant *&OutlinedFnID, |
2084 | bool IsOffloadEntry, |
2085 | const RegionCodeGenTy &CodeGen) override; |
2086 | |
2087 | /// Emit the target offloading code associated with \a D. The emitted |
2088 | /// code attempts offloading the execution to the device, an the event of |
2089 | /// a failure it executes the host version outlined in \a OutlinedFn. |
2090 | /// \param D Directive to emit. |
2091 | /// \param OutlinedFn Host version of the code to be offloaded. |
2092 | /// \param OutlinedFnID ID of host version of the code to be offloaded. |
2093 | /// \param IfCond Expression evaluated in if clause associated with the target |
2094 | /// directive, or null if no if clause is used. |
2095 | /// \param Device Expression evaluated in device clause associated with the |
2096 | /// target directive, or null if no device clause is used. |
2097 | void emitTargetCall(CodeGenFunction &CGF, const OMPExecutableDirective &D, |
2098 | llvm::Function *OutlinedFn, llvm::Value *OutlinedFnID, |
2099 | const Expr *IfCond, const Expr *Device) override; |
2100 | |
2101 | /// Emit the target regions enclosed in \a GD function definition or |
2102 | /// the function itself in case it is a valid device function. Returns true if |
2103 | /// \a GD was dealt with successfully. |
2104 | /// \param GD Function to scan. |
2105 | bool emitTargetFunctions(GlobalDecl GD) override; |
2106 | |
2107 | /// Emit the global variable if it is a valid device global variable. |
2108 | /// Returns true if \a GD was dealt with successfully. |
2109 | /// \param GD Variable declaration to emit. |
2110 | bool emitTargetGlobalVariable(GlobalDecl GD) override; |
2111 | |
2112 | /// Emit the global \a GD if it is meaningful for the target. Returns |
2113 | /// if it was emitted successfully. |
2114 | /// \param GD Global to scan. |
2115 | bool emitTargetGlobal(GlobalDecl GD) override; |
2116 | |
2117 | /// Creates the offloading descriptor in the event any target region |
2118 | /// was emitted in the current module and return the function that registers |
2119 | /// it. |
2120 | llvm::Function *emitRegistrationFunction() override; |
2121 | |
2122 | /// Emits code for teams call of the \a OutlinedFn with |
2123 | /// variables captured in a record which address is stored in \a |
2124 | /// CapturedStruct. |
2125 | /// \param OutlinedFn Outlined function to be run by team masters. Type of |
2126 | /// this function is void(*)(kmp_int32 *, kmp_int32, struct context_vars*). |
2127 | /// \param CapturedVars A pointer to the record with the references to |
2128 | /// variables used in \a OutlinedFn function. |
2129 | /// |
2130 | void emitTeamsCall(CodeGenFunction &CGF, const OMPExecutableDirective &D, |
2131 | SourceLocation Loc, llvm::Function *OutlinedFn, |
2132 | ArrayRef<llvm::Value *> CapturedVars) override; |
2133 | |
2134 | /// Emits call to void __kmpc_push_num_teams(ident_t *loc, kmp_int32 |
2135 | /// global_tid, kmp_int32 num_teams, kmp_int32 thread_limit) to generate code |
2136 | /// for num_teams clause. |
2137 | /// \param NumTeams An integer expression of teams. |
2138 | /// \param ThreadLimit An integer expression of threads. |
2139 | void emitNumTeamsClause(CodeGenFunction &CGF, const Expr *NumTeams, |
2140 | const Expr *ThreadLimit, SourceLocation Loc) override; |
2141 | |
2142 | /// Emit the target data mapping code associated with \a D. |
2143 | /// \param D Directive to emit. |
2144 | /// \param IfCond Expression evaluated in if clause associated with the |
2145 | /// target directive, or null if no device clause is used. |
2146 | /// \param Device Expression evaluated in device clause associated with the |
2147 | /// target directive, or null if no device clause is used. |
2148 | /// \param Info A record used to store information that needs to be preserved |
2149 | /// until the region is closed. |
2150 | void emitTargetDataCalls(CodeGenFunction &CGF, |
2151 | const OMPExecutableDirective &D, const Expr *IfCond, |
2152 | const Expr *Device, const RegionCodeGenTy &CodeGen, |
2153 | TargetDataInfo &Info) override; |
2154 | |
2155 | /// Emit the data mapping/movement code associated with the directive |
2156 | /// \a D that should be of the form 'target [{enter|exit} data | update]'. |
2157 | /// \param D Directive to emit. |
2158 | /// \param IfCond Expression evaluated in if clause associated with the target |
2159 | /// directive, or null if no if clause is used. |
2160 | /// \param Device Expression evaluated in device clause associated with the |
2161 | /// target directive, or null if no device clause is used. |
2162 | void emitTargetDataStandAloneCall(CodeGenFunction &CGF, |
2163 | const OMPExecutableDirective &D, |
2164 | const Expr *IfCond, |
2165 | const Expr *Device) override; |
2166 | |
2167 | /// Emit initialization for doacross loop nesting support. |
2168 | /// \param D Loop-based construct used in doacross nesting construct. |
2169 | void emitDoacrossInit(CodeGenFunction &CGF, const OMPLoopDirective &D, |
2170 | ArrayRef<Expr *> NumIterations) override; |
2171 | |
2172 | /// Emit code for doacross ordered directive with 'depend' clause. |
2173 | /// \param C 'depend' clause with 'sink|source' dependency kind. |
2174 | void emitDoacrossOrdered(CodeGenFunction &CGF, |
2175 | const OMPDependClause *C) override; |
2176 | |
2177 | /// Translates the native parameter of outlined function if this is required |
2178 | /// for target. |
2179 | /// \param FD Field decl from captured record for the parameter. |
2180 | /// \param NativeParam Parameter itself. |
2181 | const VarDecl *translateParameter(const FieldDecl *FD, |
2182 | const VarDecl *NativeParam) const override; |
2183 | |
2184 | /// Gets the address of the native argument basing on the address of the |
2185 | /// target-specific parameter. |
2186 | /// \param NativeParam Parameter itself. |
2187 | /// \param TargetParam Corresponding target-specific parameter. |
2188 | Address getParameterAddress(CodeGenFunction &CGF, const VarDecl *NativeParam, |
2189 | const VarDecl *TargetParam) const override; |
2190 | |
2191 | /// Gets the OpenMP-specific address of the local variable. |
2192 | Address getAddressOfLocalVariable(CodeGenFunction &CGF, |
2193 | const VarDecl *VD) override { |
2194 | return Address::invalid(); |
2195 | } |
2196 | }; |
2197 | |
2198 | } // namespace CodeGen |
2199 | } // namespace clang |
2200 | |
2201 | #endif |