LLVM 24.0.0git
SPIRVInstructionSelector.cpp
Go to the documentation of this file.
1//===- SPIRVInstructionSelector.cpp ------------------------------*- 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 file implements the targeting of the InstructionSelector class for
10// SPIRV.
11// TODO: This should be generated by TableGen.
12//
13//===----------------------------------------------------------------------===//
14
15#include "MCTargetDesc/SPIRVBaseInfo.h"
16#include "MCTargetDesc/SPIRVMCTargetDesc.h"
17#include "SPIRV.h"
18#include "SPIRVGlobalRegistry.h"
19#include "SPIRVInstrInfo.h"
20#include "SPIRVRegisterInfo.h"
21#include "SPIRVTargetMachine.h"
22#include "SPIRVTypeInst.h"
23#include "SPIRVUtils.h"
24#include "llvm/ADT/APFloat.h"
25#include "llvm/ADT/STLExtras.h"
26#include "llvm/ADT/STLFunctionalExtras.h"
27#include "llvm/ADT/SmallSet.h"
28#include "llvm/ADT/StringExtras.h"
29#include "llvm/CodeGen/GlobalISel/GIMatchTableExecutorImpl.h"
30#include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
31#include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
32#include "llvm/CodeGen/MachineInstrBuilder.h"
33#include "llvm/CodeGen/MachineRegisterInfo.h"
34#include "llvm/CodeGen/Register.h"
35#include "llvm/CodeGen/TargetOpcodes.h"
36#include "llvm/IR/DiagnosticInfo.h"
37#include "llvm/IR/IntrinsicsSPIRV.h"
38#include "llvm/Support/Debug.h"
39#include "llvm/Support/ErrorHandling.h"
40#include <functional>
41#include <optional>
42
43#define DEBUG_TYPE "spirv-isel"
44
45using namespace llvm;
46namespace CL = SPIRV::OpenCLExtInst;
47namespace GL = SPIRV::GLSLExtInst;
48
49using ExtInstList =
50 std::vector<std::pair<SPIRV::InstructionSet::InstructionSet, uint32_t>>;
51
52namespace {
53
54struct ImageOperands {
55 std::optional<Register> Bias;
56 std::optional<Register> Offset;
57 std::optional<Register> MinLod;
58 std::optional<Register> GradX;
59 std::optional<Register> GradY;
60 std::optional<Register> Lod;
61 std::optional<Register> Compare;
62};
63
64llvm::SPIRV::SelectionControl::SelectionControl
65getSelectionOperandForImm(int Imm) {
66 if (Imm == 2)
67 return SPIRV::SelectionControl::Flatten;
68 if (Imm == 1)
69 return SPIRV::SelectionControl::DontFlatten;
70 if (Imm == 0)
71 return SPIRV::SelectionControl::None;
72 llvm_unreachable("Invalid immediate");
73}
74
75#define GET_GLOBALISEL_PREDICATE_BITSET
76#include "SPIRVGenGlobalISel.inc"
77#undef GET_GLOBALISEL_PREDICATE_BITSET
78
79class SPIRVInstructionSelector : public InstructionSelector {
80 const SPIRVSubtarget &STI;
81 const SPIRVInstrInfo &TII;
82 const SPIRVRegisterInfo &TRI;
83 const RegisterBankInfo &RBI;
84 SPIRVGlobalRegistry &GR;
85 MachineRegisterInfo *MRI;
86 MachineFunction *HasVRegsReset = nullptr;
87
88 /// We need to keep track of the number we give to anonymous global values to
89 /// generate the same name every time when this is needed.
90 mutable DenseMap<const GlobalValue *, unsigned> UnnamedGlobalIDs;
91 SmallPtrSet<MachineInstr *, 8> DeadMIs;
92
93public:
94 SPIRVInstructionSelector(const SPIRVTargetMachine &TM,
95 const SPIRVSubtarget &ST,
96 const RegisterBankInfo &RBI);
97 void setupMF(MachineFunction &MF, GISelValueTracking *VT,
98 CodeGenCoverage *CoverageInfo, ProfileSummaryInfo *PSI,
99 BlockFrequencyInfo *BFI) override;
100 // Common selection code. Instruction-specific selection occurs in spvSelect.
101 bool select(MachineInstr &I) override;
102 static const char *getName() { return DEBUG_TYPE; }
103
104#define GET_GLOBALISEL_PREDICATES_DECL
105#include "SPIRVGenGlobalISel.inc"
106#undef GET_GLOBALISEL_PREDICATES_DECL
107
108#define GET_GLOBALISEL_TEMPORARIES_DECL
109#include "SPIRVGenGlobalISel.inc"
110#undef GET_GLOBALISEL_TEMPORARIES_DECL
111
112private:
113 void resetVRegsType(MachineFunction &MF);
114 void removeDeadInstruction(MachineInstr &MI) const;
115 void removeOpNamesForDeadMI(MachineInstr &MI) const;
116
117 // tblgen-erated 'select' implementation, used as the initial selector for
118 // the patterns that don't require complex C++.
119 bool selectImpl(MachineInstr &I, CodeGenCoverage &CoverageInfo) const;
120
121 // All instruction-specific selection that didn't happen in "select()".
122 // Is basically a large Switch/Case delegating to all other select method.
123 bool spvSelect(Register ResVReg, SPIRVTypeInst ResType,
124 MachineInstr &I) const;
125
126 bool selectFirstBitHigh(Register ResVReg, SPIRVTypeInst ResType,
127 MachineInstr &I, bool IsSigned) const;
128
129 bool selectFirstBitLow(Register ResVReg, SPIRVTypeInst ResType,
130 MachineInstr &I) const;
131
132 bool selectFirstBitSet16(Register ResVReg, SPIRVTypeInst ResType,
133 MachineInstr &I, unsigned ExtendOpcode,
134 unsigned BitSetOpcode) const;
135
136 bool selectFirstBitSet32(Register ResVReg, SPIRVTypeInst ResType,
137 MachineInstr &I, Register SrcReg,
138 unsigned BitSetOpcode) const;
139
140 bool selectFirstBitSet64(Register ResVReg, SPIRVTypeInst ResType,
141 MachineInstr &I, Register SrcReg,
142 unsigned BitSetOpcode, bool SwapPrimarySide) const;
143
144 bool selectGlobalValue(Register ResVReg, MachineInstr &I,
145 const MachineInstr *Init = nullptr) const;
146
147 bool selectOpWithSrcs(Register ResVReg, SPIRVTypeInst ResType,
148 MachineInstr &I, ArrayRef<Register> SrcRegs,
149 unsigned Opcode) const;
150
151 bool selectUnOp(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
152 unsigned Opcode) const;
153
154 bool selectBitcast(Register ResVReg, SPIRVTypeInst ResType,
155 MachineInstr &I) const;
156
157 bool selectLoad(Register ResVReg, SPIRVTypeInst ResType,
158 MachineInstr &I) const;
159 bool selectAtomicLoad(Register ResVReg, SPIRVTypeInst ResType,
160 MachineInstr &I) const;
161 bool selectStore(MachineInstr &I) const;
162 bool selectAtomicStore(MachineInstr &I) const;
163
164 bool selectStackSave(Register ResVReg, SPIRVTypeInst ResType,
165 MachineInstr &I) const;
166 bool selectStackRestore(MachineInstr &I) const;
167
168 bool selectMemOperation(Register ResVReg, MachineInstr &I) const;
169 Register getOrCreateMemSetGlobal(MachineInstr &I) const;
170 bool selectCopyMemory(MachineInstr &I, Register SrcReg) const;
171 bool selectCopyMemorySized(MachineInstr &I, Register SrcReg) const;
172
173 bool selectAtomicRMW(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
174 unsigned NewOpcode, unsigned NegateOpcode = 0) const;
175
176 // Creates an integer-typed register with bitwidth equal to pointer size.
177 Register createPtrSizedIntReg(MachineIRBuilder &MIRBuilder) const;
178 // Emit an OpConvertPtrToU that converts the pointer value in \p PtrVal into
179 // an integer of equal bitwidth, returning the register holding the result.
180 Register convertPtrToInt(Register PtrVal, MachineIRBuilder &MIRBuilder) const;
181 // Emit an OpBitcast that reinterprets the pointer \p Ptr as a pointer to an
182 // integer of pointer size in storage class \p SC, returning the result.
183 Register castPtrToPtrToInt(Register Ptr, SPIRV::StorageClass::StorageClass SC,
184 MachineIRBuilder &MIRBuilder) const;
185 // Handle atomic loads, stores and exchanges of pointer types by casting
186 // to/from integer types as needed.
187 bool selectAtomicPtrValue(
188 Register ResVReg, SPIRVTypeInst ResType, MachineIRBuilder &MIRBuilder,
189 function_ref<Register(SPIRVTypeInst IntType)> EmitAtomic) const;
190
191 bool selectAtomicCmpXchg(Register ResVReg, SPIRVTypeInst ResType,
192 MachineInstr &I) const;
193
194 bool selectFence(MachineInstr &I) const;
195
196 bool selectAddrSpaceCast(Register ResVReg, SPIRVTypeInst ResType,
197 MachineInstr &I) const;
198
199 bool selectPtrMask(Register ResVReg, SPIRVTypeInst ResType,
200 MachineInstr &I) const;
201
202 bool selectAnyOrAll(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
203 unsigned OpType) const;
204
205 bool selectAll(Register ResVReg, SPIRVTypeInst ResType,
206 MachineInstr &I) const;
207
208 bool selectAny(Register ResVReg, SPIRVTypeInst ResType,
209 MachineInstr &I) const;
210
211 bool selectBitreverse(Register ResVReg, SPIRVTypeInst ResType,
212 MachineInstr &I) const;
213
214 bool selectBitreverseViaI32(Register ResVReg, SPIRVTypeInst ResType,
215 MachineInstr &I, Register Op) const;
216
217 bool selectBitreverse64(Register ResVReg, SPIRVTypeInst ResType,
218 MachineInstr &I, Register SrcReg) const;
219
220 bool selectBitreverseNative(Register ResVReg, SPIRVTypeInst ResType,
221 MachineInstr &I, Register Op) const;
222
223 bool selectBuildVector(Register ResVReg, SPIRVTypeInst ResType,
224 MachineInstr &I) const;
225 bool selectSplatVector(Register ResVReg, SPIRVTypeInst ResType,
226 MachineInstr &I) const;
227 bool selectConcatVectors(Register ResVReg, SPIRVTypeInst ResType,
228 MachineInstr &I) const;
229
230 bool selectCmp(Register ResVReg, SPIRVTypeInst ResType,
231 unsigned comparisonOpcode, MachineInstr &I) const;
232 bool selectDiscard(Register ResVReg, SPIRVTypeInst ResType,
233 MachineInstr &I) const;
234
235 bool selectICmp(Register ResVReg, SPIRVTypeInst ResType,
236 MachineInstr &I) const;
237 bool selectFCmp(Register ResVReg, SPIRVTypeInst ResType,
238 MachineInstr &I) const;
239
240 bool selectSign(Register ResVReg, SPIRVTypeInst ResType,
241 MachineInstr &I) const;
242
243 bool selectFloatDot(Register ResVReg, SPIRVTypeInst ResType,
244 MachineInstr &I) const;
245
246 bool selectOverflowArith(Register ResVReg, SPIRVTypeInst ResType,
247 MachineInstr &I, unsigned Opcode) const;
248 bool selectDebugTrap(Register ResVReg, SPIRVTypeInst ResType,
249 MachineInstr &I) const;
250 bool selectPrefetch(MachineInstr &I) const;
251
252 bool selectIntegerDot(Register ResVReg, SPIRVTypeInst ResType,
253 MachineInstr &I, bool Signed) const;
254
255 bool selectIntegerDotExpansion(Register ResVReg, SPIRVTypeInst ResType,
256 MachineInstr &I) const;
257
258 bool selectOpIsInf(Register ResVReg, SPIRVTypeInst ResType,
259 MachineInstr &I) const;
260
261 bool selectOpIsNan(Register ResVReg, SPIRVTypeInst ResType,
262 MachineInstr &I) const;
263
264 bool selectOpIsFinite(Register ResVReg, SPIRVTypeInst ResType,
265 MachineInstr &I) const;
266
267 bool selectOpIsNormal(Register ResVReg, SPIRVTypeInst ResType,
268 MachineInstr &I) const;
269
270 bool selectPopCount(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
271 unsigned Opcode) const;
272
273 bool selectPopCount16(Register ResVReg, SPIRVTypeInst ResType,
274 MachineInstr &I, unsigned ExtOpcode,
275 unsigned Opcode) const;
276
277 bool selectPopCount32(Register ResVReg, SPIRVTypeInst ResType,
278 MachineInstr &I, Register SrcReg,
279 unsigned Opcode) const;
280
281 bool selectPopCount64(Register ResVReg, SPIRVTypeInst ResType,
282 MachineInstr &I, Register SrcReg,
283 unsigned Opcode) const;
284
285 template <bool Signed>
286 bool selectDot4AddPacked(Register ResVReg, SPIRVTypeInst ResType,
287 MachineInstr &I) const;
288 template <bool Signed>
289 bool selectDot4AddPackedExpansion(Register ResVReg, SPIRVTypeInst ResType,
290 MachineInstr &I) const;
291
292 bool selectWavePrefixBitCount(Register ResVReg, SPIRVTypeInst ResType,
293 MachineInstr &I) const;
294
295 template <typename PickOpcodeFn>
296 bool selectWaveReduce(Register ResVReg, SPIRVTypeInst ResType,
297 MachineInstr &I, bool IsUnsigned,
298 PickOpcodeFn &&PickOpcode) const;
299
300 bool selectWaveReduceOp(Register ResVReg, SPIRVTypeInst ResType,
301 MachineInstr &I, unsigned Opcode) const;
302
303 bool selectWaveReduceMax(Register ResVReg, SPIRVTypeInst ResType,
304 MachineInstr &I, bool IsUnsigned) const;
305
306 bool selectWaveReduceMin(Register ResVReg, SPIRVTypeInst ResType,
307 MachineInstr &I, bool IsUnsigned) const;
308
309 bool selectWaveReduceSum(Register ResVReg, SPIRVTypeInst ResType,
310 MachineInstr &I) const;
311
312 bool selectWaveReduceProduct(Register ResVReg, const SPIRVTypeInst ResType,
313 MachineInstr &I) const;
314
315 template <typename PickOpcodeFn>
316 bool selectWaveExclusiveScan(Register ResVReg, SPIRVTypeInst ResType,
317 MachineInstr &I, bool IsUnsigned,
318 PickOpcodeFn &&PickOpcode) const;
319
320 bool selectWaveExclusiveScanSum(Register ResVReg, SPIRVTypeInst ResType,
321 MachineInstr &I) const;
322
323 bool selectWaveExclusiveScanProduct(Register ResVReg, SPIRVTypeInst ResType,
324 MachineInstr &I) const;
325
326 bool selectQuadSwap(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
327 unsigned Direction) const;
328
329 bool selectConst(Register ResVReg, SPIRVTypeInst ResType,
330 MachineInstr &I) const;
331
332 bool selectSelect(Register ResVReg, SPIRVTypeInst ResType,
333 MachineInstr &I) const;
334 bool selectBoolToInt(Register ResVReg, SPIRVTypeInst ResType,
335 Register BooleanVReg, MachineInstr &InsertAt,
336 bool IsSigned) const;
337 bool selectIToF(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
338 bool IsSigned, unsigned Opcode) const;
339 bool selectExt(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
340 bool IsSigned) const;
341
342 bool selectTrunc(Register ResVReg, SPIRVTypeInst ResType,
343 MachineInstr &I) const;
344
345 bool selectSUCmp(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
346 bool IsSigned) const;
347
348 bool selectIntToBool(Register IntReg, Register ResVReg, MachineInstr &I,
349 SPIRVTypeInst intTy, SPIRVTypeInst boolTy) const;
350
351 bool selectOpUndef(Register ResVReg, SPIRVTypeInst ResType,
352 MachineInstr &I) const;
353 bool selectFreeze(Register ResVReg, SPIRVTypeInst ResType,
354 MachineInstr &I) const;
355 bool selectIntrinsic(Register ResVReg, SPIRVTypeInst ResType,
356 MachineInstr &I) const;
357 bool selectExtractVal(Register ResVReg, SPIRVTypeInst ResType,
358 MachineInstr &I) const;
359 bool selectInsertVal(Register ResVReg, SPIRVTypeInst ResType,
360 MachineInstr &I) const;
361 bool selectExtractElt(Register ResVReg, SPIRVTypeInst ResType,
362 MachineInstr &I) const;
363 bool selectInsertElt(Register ResVReg, SPIRVTypeInst ResType,
364 MachineInstr &I) const;
365 bool selectGEP(Register ResVReg, SPIRVTypeInst ResType,
366 MachineInstr &I) const;
367
368 bool selectMaskedGather(Register ResVReg, SPIRVTypeInst ResType,
369 MachineInstr &I) const;
370 bool selectMaskedScatter(MachineInstr &I) const;
371
372 bool diagnoseUnsupported(const MachineInstr &I, const Twine &Msg) const;
373
374 bool selectAbort(MachineInstr &I) const;
375 bool selectTrap(MachineInstr &I) const;
376 bool selectFrameIndex(Register ResVReg, SPIRVTypeInst ResType,
377 MachineInstr &I) const;
378 bool selectAllocaArray(Register ResVReg, SPIRVTypeInst ResType,
379 MachineInstr &I) const;
380
381 bool selectBranch(MachineInstr &I) const;
382 bool selectBranchCond(MachineInstr &I) const;
383
384 bool selectPhi(Register ResVReg, MachineInstr &I) const;
385
386 bool selectExtInst(Register ResVReg, SPIRVTypeInst RestType, MachineInstr &I,
387 GL::GLSLExtInst GLInst, bool setMIFlags = true,
388 bool useMISrc = true,
389 ArrayRef<Register> SrcRegs = {}) const;
390 bool selectExtInst(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
391 CL::OpenCLExtInst CLInst, bool setMIFlags = true,
392 bool useMISrc = true,
393 ArrayRef<Register> SrcRegs = {}) const;
394 bool selectExtInst(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
395 CL::OpenCLExtInst CLInst, GL::GLSLExtInst GLInst,
396 bool setMIFlags = true, bool useMISrc = true,
397 ArrayRef<Register> SrcRegs = {}) const;
398 bool selectExtInst(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
399 const ExtInstList &ExtInsts, bool setMIFlags = true,
400 bool useMISrc = true,
401 ArrayRef<Register> SrcRegs = {}) const;
402
403 bool selectLog10(Register ResVReg, SPIRVTypeInst ResType,
404 MachineInstr &I) const;
405
406 bool selectFpowi(Register ResVReg, SPIRVTypeInst ResType,
407 MachineInstr &I) const;
408
409 bool selectSaturate(Register ResVReg, SPIRVTypeInst ResType,
410 MachineInstr &I) const;
411
412 bool selectWaveOpInst(Register ResVReg, SPIRVTypeInst ResType,
413 MachineInstr &I, unsigned Opcode) const;
414
415 bool selectBarrierInst(MachineInstr &I, unsigned Scope, unsigned MemSem,
416 bool WithGroupSync) const;
417
418 bool selectWaveActiveCountBits(Register ResVReg, SPIRVTypeInst ResType,
419 MachineInstr &I) const;
420
421 bool selectWaveActiveAllEqual(Register ResVReg, SPIRVTypeInst ResType,
422 MachineInstr &I) const;
423
424 bool selectUnmergeValues(MachineInstr &I) const;
425
426 bool selectHandleFromBinding(Register &ResVReg, SPIRVTypeInst ResType,
427 MachineInstr &I) const;
428
429 bool selectCounterHandleFromBinding(Register &ResVReg, SPIRVTypeInst ResType,
430 MachineInstr &I) const;
431
432 bool selectReadImageIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
433 MachineInstr &I) const;
434 bool selectGetDimensionsIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
435 MachineInstr &I) const;
436 bool selectGetDimensionsLevelsIntrinsic(Register &ResVReg,
437 SPIRVTypeInst ResType,
438 MachineInstr &I) const;
439 bool selectGetDimensionsMSIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
440 MachineInstr &I) const;
441 bool
442 selectImageQuerySize(Register ImageReg, Register &ResVReg, MachineInstr &I,
443 std::optional<Register> LodReg = std::nullopt) const;
444 bool selectSampleBasicIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
445 MachineInstr &I) const;
446 bool selectCalculateLodIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
447 MachineInstr &I) const;
448 bool selectSampleBiasIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
449 MachineInstr &I) const;
450 bool selectSampleGradIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
451 MachineInstr &I) const;
452 bool selectSampleLevelIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
453 MachineInstr &I) const;
454 bool selectLoadLevelIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
455 MachineInstr &I) const;
456 bool selectSampleCmpIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
457 MachineInstr &I) const;
458 bool selectSampleCmpLevelZeroIntrinsic(Register &ResVReg,
459 SPIRVTypeInst ResType,
460 MachineInstr &I) const;
461 bool selectGatherIntrinsic(Register &ResVReg, SPIRVTypeInst ResType,
462 MachineInstr &I) const;
463 bool selectImageWriteIntrinsic(MachineInstr &I) const;
464 bool selectResourceGetPointer(Register &ResVReg, SPIRVTypeInst ResType,
465 MachineInstr &I) const;
466 bool selectPushConstantGetPointer(Register &ResVReg, SPIRVTypeInst ResType,
467 MachineInstr &I) const;
468 bool selectResourceNonUniformIndex(Register &ResVReg, SPIRVTypeInst ResType,
469 MachineInstr &I) const;
470 bool selectModf(Register ResVReg, SPIRVTypeInst ResType,
471 MachineInstr &I) const;
472 bool selectUpdateCounter(Register &ResVReg, SPIRVTypeInst ResType,
473 MachineInstr &I) const;
474 bool selectFrexp(Register ResVReg, SPIRVTypeInst ResType,
475 MachineInstr &I) const;
476
477 bool selectCopySign(Register ResVReg, SPIRVTypeInst ResType,
478 MachineInstr &I) const;
479
480 bool selectLdexp(Register ResVReg, SPIRVTypeInst ResType,
481 MachineInstr &I) const;
482 bool selectSincos(Register ResVReg, SPIRVTypeInst ResType,
483 MachineInstr &I) const;
484 bool selectExp10(Register ResVReg, SPIRVTypeInst ResType,
485 MachineInstr &I) const;
486 bool selectDerivativeInst(Register ResVReg, SPIRVTypeInst ResType,
487 MachineInstr &I, const unsigned DPdOpCode) const;
488 // Utilities
489 Register buildI32Constant(uint32_t Val, MachineInstr &I,
490 SPIRVTypeInst ResType = nullptr) const;
491 Register buildI32ConstantInEntryBlock(uint32_t Val, MachineInstr &I,
492 SPIRVTypeInst ResType = nullptr) const;
493
494 Register buildZerosVal(SPIRVTypeInst ResType, MachineInstr &I) const;
495 bool isScalarOrVectorIntConstantZero(Register Reg) const;
496 Register buildZerosValF(SPIRVTypeInst ResType, MachineInstr &I) const;
497 Register buildOnesVal(bool AllOnes, SPIRVTypeInst ResType,
498 MachineInstr &I) const;
499 Register buildOnesValF(SPIRVTypeInst ResType, MachineInstr &I) const;
500
501 bool wrapIntoSpecConstantOp(MachineInstr &I,
502 SmallVector<Register> &CompositeArgs) const;
503
504 Register getUcharPtrTypeReg(MachineInstr &I,
505 SPIRV::StorageClass::StorageClass SC) const;
506 MachineInstrBuilder buildSpecConstantOp(MachineInstr &I, Register Dest,
507 Register Src, Register DestType,
508 uint32_t Opcode) const;
509 MachineInstrBuilder buildConstGenericPtr(MachineInstr &I, Register SrcPtr,
510 SPIRVTypeInst SrcPtrTy) const;
511 Register buildPointerToResource(SPIRVTypeInst ResType,
512 SPIRV::StorageClass::StorageClass SC,
513 uint32_t Set, uint32_t Binding,
514 uint32_t ArraySize, Register IndexReg,
515 StringRef Name,
516 MachineIRBuilder MIRBuilder) const;
517 SPIRVTypeInst widenTypeToVec4(SPIRVTypeInst Type, MachineInstr &I) const;
518 bool extractSubvector(Register &ResVReg, SPIRVTypeInst ResType,
519 Register &ReadReg, MachineInstr &InsertionPoint) const;
520 bool generateImageReadOrFetch(Register &ResVReg, SPIRVTypeInst ResType,
521 Register ImageReg, Register IdxReg,
522 DebugLoc Loc, MachineInstr &Pos,
523 const ImageOperands *ImOps = nullptr) const;
524 bool generateSampleImage(Register ResVReg, SPIRVTypeInst ResType,
525 Register ImageReg, Register SamplerReg,
526 Register CoordinateReg, const ImageOperands &ImOps,
527 DebugLoc Loc, MachineInstr &I) const;
528 bool BuildCOPY(Register DestReg, Register SrcReg, MachineInstr &I) const;
529 bool loadVec3BuiltinInputID(SPIRV::BuiltIn::BuiltIn BuiltInValue,
530 Register ResVReg, SPIRVTypeInst ResType,
531 MachineInstr &I) const;
532 bool loadBuiltinInputID(SPIRV::BuiltIn::BuiltIn BuiltInValue,
533 Register ResVReg, SPIRVTypeInst ResType,
534 MachineInstr &I) const;
535 bool loadHandleBeforePosition(Register &HandleReg, SPIRVTypeInst ResType,
536 GIntrinsic &HandleDef, MachineInstr &Pos) const;
537 void decorateUsesAsNonUniform(Register &NonUniformReg) const;
538 bool errorIfInstrOutsideShader(MachineInstr &I) const;
539
540 bool
541 handle64BitOverflow(Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
542 Register SrcReg, unsigned int Opcode,
543 std::function<bool(Register, SPIRVTypeInst,
544 MachineInstr &, Register, unsigned)>
545 CallbackFunction) const;
546};
547
548bool sampledTypeIsSignedInteger(const llvm::Type *HandleType) {
549 const TargetExtType *TET = cast<TargetExtType>(HandleType);
550 if (TET->getTargetExtName() == "spirv.Image") {
551 return false;
552 }
553 assert(TET->getTargetExtName() == "spirv.SignedImage");
554 return TET->getTypeParameter(0)->isIntegerTy();
555}
556} // end anonymous namespace
557
558#define GET_GLOBALISEL_IMPL
559#include "SPIRVGenGlobalISel.inc"
560#undef GET_GLOBALISEL_IMPL
561
562SPIRVInstructionSelector::SPIRVInstructionSelector(const SPIRVTargetMachine &TM,
563 const SPIRVSubtarget &ST,
564 const RegisterBankInfo &RBI)
565 : InstructionSelector(), STI(ST), TII(*ST.getInstrInfo()),
566 TRI(*ST.getRegisterInfo()), RBI(RBI), GR(*ST.getSPIRVGlobalRegistry()),
567 MRI(nullptr),
568#define GET_GLOBALISEL_PREDICATES_INIT
569#include "SPIRVGenGlobalISel.inc"
570#undef GET_GLOBALISEL_PREDICATES_INIT
571#define GET_GLOBALISEL_TEMPORARIES_INIT
572#include "SPIRVGenGlobalISel.inc"
573#undef GET_GLOBALISEL_TEMPORARIES_INIT
574{
575}
576
577void SPIRVInstructionSelector::setupMF(MachineFunction &MF,
578 GISelValueTracking *VT,
579 CodeGenCoverage *CoverageInfo,
580 ProfileSummaryInfo *PSI,
581 BlockFrequencyInfo *BFI) {
582 MRI = &MF.getRegInfo();
583 GR.setCurrentFunc(MF);
584 InstructionSelector::setupMF(MF, VT, CoverageInfo, PSI, BFI);
585}
586
587// Ensure that register classes correspond to pattern matching rules.
588void SPIRVInstructionSelector::resetVRegsType(MachineFunction &MF) {
589 if (HasVRegsReset == &MF)
590 return;
591 HasVRegsReset = &MF;
592
593 MachineRegisterInfo &MRI = MF.getRegInfo();
594 for (unsigned I = 0, E = MRI.getNumVirtRegs(); I != E; ++I) {
595 Register Reg = Register::index2VirtReg(I);
596 LLT RegType = MRI.getType(Reg);
597 if (RegType.isScalar())
598 MRI.setType(Reg, LLT::scalar(64));
599 else if (RegType.isPointer())
600 MRI.setType(Reg, LLT::pointer(0, 64));
601 else if (RegType.isVector())
602 MRI.setType(Reg, LLT::fixed_vector(2, LLT::scalar(64)));
603 }
604 for (const auto &MBB : MF) {
605 for (const auto &MI : MBB) {
606 if (isPreISelGenericOpcode(MI.getOpcode()))
607 GR.erase(&MI);
608 if (MI.getOpcode() != SPIRV::ASSIGN_TYPE)
609 continue;
610
611 Register DstReg = MI.getOperand(0).getReg();
612 LLT DstType = MRI.getType(DstReg);
613 Register SrcReg = MI.getOperand(1).getReg();
614 LLT SrcType = MRI.getType(SrcReg);
615 if (DstType != SrcType)
616 MRI.setType(DstReg, MRI.getType(SrcReg));
617
618 const TargetRegisterClass *DstRC = MRI.getRegClassOrNull(DstReg);
619 const TargetRegisterClass *SrcRC = MRI.getRegClassOrNull(SrcReg);
620 if (DstRC != SrcRC && SrcRC)
621 MRI.setRegClass(DstReg, SrcRC);
622 }
623 }
624}
625
626// Return true if the MachineInstr represents a constant register
627static bool isConstReg(MachineRegisterInfo *MRI, MachineInstr *OpDef) {
628
629 SmallVector<MachineInstr *> Stack = {OpDef};
630 SmallPtrSet<MachineInstr *, 4> Visited;
631
632 while (!Stack.empty()) {
633 MachineInstr *MI = Stack.pop_back_val();
634 MI = passCopy(MI, MRI);
635 if (!Visited.insert(MI).second)
636 continue;
637 switch (MI->getOpcode()) {
638 case TargetOpcode::G_INTRINSIC:
639 case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
640 case TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS: {
641 GIntrinsic *GIntr = cast<GIntrinsic>(MI);
642 unsigned IntrID = GIntr->getIntrinsicID();
643 if (IntrID != Intrinsic::spv_const_composite &&
644 IntrID != Intrinsic::spv_undef && IntrID != Intrinsic::spv_poison)
645 return false;
646 continue;
647 }
648 case TargetOpcode::G_BUILD_VECTOR:
649 case TargetOpcode::G_SPLAT_VECTOR:
650 for (unsigned i = OpDef->getNumExplicitDefs();
651 i < OpDef->getNumOperands(); i++) {
652 if (!OpDef->getOperand(i).isReg())
653 continue;
654 MachineInstr *OpNestedDef =
655 MRI->getVRegDef(OpDef->getOperand(i).getReg());
656 Stack.push_back(OpNestedDef);
657 }
658 continue;
659 case TargetOpcode::G_CONSTANT:
660 case TargetOpcode::G_FCONSTANT:
661 case TargetOpcode::G_IMPLICIT_DEF:
662 case SPIRV::OpConstantTrue:
663 case SPIRV::OpConstantFalse:
664 case SPIRV::OpConstantI:
665 case SPIRV::OpConstantF:
666 case SPIRV::OpConstantComposite:
667 case SPIRV::OpConstantCompositeContinuedINTEL:
668 case SPIRV::OpConstantSampler:
669 case SPIRV::OpConstantNull:
670 case SPIRV::OpUndef:
671 case SPIRV::OpPoisonKHR:
672 case SPIRV::OpConstantFunctionPointerINTEL:
673 continue;
674 default:
675 return false;
676 }
677 }
678 return true;
679}
680
681// Return true if the virtual register represents a constant
682static bool isConstReg(MachineRegisterInfo *MRI, Register OpReg) {
683 if (MachineInstr *OpDef = MRI->getVRegDef(OpReg))
684 return isConstReg(MRI, OpDef);
685 return false;
686}
687
688// TODO(168736): We should make this either a flag in tabelgen
689// or reduce our dependence on the global registry, so we can remove this
690// function. It can easily be missed when new intrinsics are added.
691
692// Most SPIR-V intrinsics are considered to have side-effects in their tablegen
693// definition because they are referenced in the global registry. This is a list
694// of intrinsics that have no side effects other than their references in the
695// global registry.
696static bool intrinsicHasSideEffects(Intrinsic::ID ID) {
697 switch (ID) {
698 // This is not an exhaustive list and may need to be updated.
699 case Intrinsic::spv_all:
700 case Intrinsic::spv_alloca:
701 case Intrinsic::spv_any:
702 case Intrinsic::spv_bitcast:
703 case Intrinsic::spv_const_composite:
704 case Intrinsic::spv_degrees:
705 case Intrinsic::spv_distance:
706 case Intrinsic::spv_extractelt:
707 case Intrinsic::spv_extractv:
708 case Intrinsic::spv_faceforward:
709 case Intrinsic::spv_fdot:
710 case Intrinsic::spv_firstbitlow:
711 case Intrinsic::spv_firstbitshigh:
712 case Intrinsic::spv_firstbituhigh:
713 case Intrinsic::spv_frac:
714 case Intrinsic::spv_gep:
715 case Intrinsic::spv_global_offset:
716 case Intrinsic::spv_global_size:
717 case Intrinsic::spv_group_id:
718 case Intrinsic::spv_insertelt:
719 case Intrinsic::spv_insertv:
720 case Intrinsic::spv_isinf:
721 case Intrinsic::spv_isnan:
722 case Intrinsic::spv_isfinite:
723 case Intrinsic::spv_isnormal:
724 case Intrinsic::spv_lerp:
725 case Intrinsic::spv_length:
726 case Intrinsic::spv_normalize:
727 case Intrinsic::spv_num_subgroups:
728 case Intrinsic::spv_num_workgroups:
729 case Intrinsic::spv_ptrcast:
730 case Intrinsic::spv_radians:
731 case Intrinsic::spv_reflect:
732 case Intrinsic::spv_refract:
733 case Intrinsic::spv_resource_getbasepointer:
734 case Intrinsic::spv_resource_getpointer:
735 case Intrinsic::spv_resource_handlefrombinding:
736 case Intrinsic::spv_resource_handlefromimplicitbinding:
737 case Intrinsic::spv_resource_nonuniformindex:
738 case Intrinsic::spv_resource_sample:
739 case Intrinsic::spv_rsqrt:
740 case Intrinsic::spv_saturate:
741 case Intrinsic::spv_sdot:
742 case Intrinsic::spv_sign:
743 case Intrinsic::spv_smoothstep:
744 case Intrinsic::spv_subgroup_id:
745 case Intrinsic::spv_subgroup_local_invocation_id:
746 case Intrinsic::spv_subgroup_max_size:
747 case Intrinsic::spv_subgroup_size:
748 case Intrinsic::spv_thread_id:
749 case Intrinsic::spv_thread_id_in_group:
750 case Intrinsic::spv_udot:
751 case Intrinsic::spv_undef:
752 case Intrinsic::spv_value_md:
753 case Intrinsic::spv_workgroup_size:
754 return false;
755 default:
756 return true;
757 }
758}
759
760// TODO(168736): We should make this either a flag in tabelgen
761// or reduce our dependence on the global registry, so we can remove this
762// function. It can easily be missed when new intrinsics are added.
763static bool isOpcodeWithNoSideEffects(unsigned Opcode) {
764 switch (Opcode) {
765 case SPIRV::OpTypeVoid:
766 case SPIRV::OpTypeBool:
767 case SPIRV::OpTypeInt:
768 case SPIRV::OpTypeFloat:
769 case SPIRV::OpTypeVector:
770 case SPIRV::OpTypeVectorIdEXT:
771 case SPIRV::OpTypeMatrix:
772 case SPIRV::OpTypeImage:
773 case SPIRV::OpTypeSampler:
774 case SPIRV::OpTypeSampledImage:
775 case SPIRV::OpTypeArray:
776 case SPIRV::OpTypeRuntimeArray:
777 case SPIRV::OpTypeStruct:
778 case SPIRV::OpTypeOpaque:
779 case SPIRV::OpTypePointer:
780 case SPIRV::OpTypeFunction:
781 case SPIRV::OpTypeEvent:
782 case SPIRV::OpTypeDeviceEvent:
783 case SPIRV::OpTypeReserveId:
784 case SPIRV::OpTypeQueue:
785 case SPIRV::OpTypePipe:
786 case SPIRV::OpTypeForwardPointer:
787 case SPIRV::OpTypePipeStorage:
788 case SPIRV::OpTypeNamedBarrier:
789 case SPIRV::OpTypeAccelerationStructureNV:
790 case SPIRV::OpTypeCooperativeMatrixNV:
791 case SPIRV::OpTypeCooperativeMatrixKHR:
792 return true;
793 default:
794 return false;
795 }
796}
797
798bool isDead(const MachineInstr &MI, const MachineRegisterInfo &MRI) {
799 // If there are no definitions, then assume there is some other
800 // side-effect that makes this instruction live.
801 if (MI.getNumDefs() == 0)
802 return false;
803
804 for (const auto &MO : MI.all_defs()) {
805 Register Reg = MO.getReg();
806 if (Reg.isPhysical()) {
807 LLVM_DEBUG(dbgs() << "Not dead: def of physical register " << Reg);
808 return false;
809 }
810 for (const auto &UseMI : MRI.use_nodbg_instructions(Reg)) {
811 if (UseMI.getOpcode() != SPIRV::OpName) {
812 LLVM_DEBUG(dbgs() << "Not dead: def " << MO << " has use in " << UseMI);
813 return false;
814 }
815 }
816 }
817
818 if (MI.getOpcode() == TargetOpcode::LOCAL_ESCAPE || MI.isFakeUse() ||
819 MI.isLifetimeMarker()) {
820 LLVM_DEBUG(
821 dbgs()
822 << "Not dead: Opcode is LOCAL_ESCAPE, fake use, or lifetime marker.\n");
823 return false;
824 }
825 if (MI.isPHI()) {
826 LLVM_DEBUG(dbgs() << "Dead: Phi instruction with no uses.\n");
827 return true;
828 }
829
830 // It is possible that the only side effect is that the instruction is
831 // referenced in the global registry. If that is the only side effect, the
832 // intrinsic is dead.
833 if (MI.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS ||
834 MI.getOpcode() == TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS) {
835 const auto &Intr = cast<GIntrinsic>(MI);
836 if (!intrinsicHasSideEffects(Intr.getIntrinsicID())) {
837 LLVM_DEBUG(dbgs() << "Dead: Intrinsic with no real side effects.\n");
838 return true;
839 }
840 }
841
842 if (MI.mayStore() || MI.isCall() ||
843 (MI.mayLoad() && MI.hasOrderedMemoryRef()) || MI.isPosition() ||
844 MI.isDebugInstr() || MI.isTerminator() || MI.isJumpTableDebugInfo()) {
845 LLVM_DEBUG(dbgs() << "Not dead: instruction has side effects.\n");
846 return false;
847 }
848
849 if (isPreISelGenericOpcode(MI.getOpcode())) {
850 // TODO: Is there a generic way to check if the opcode has side effects?
851 LLVM_DEBUG(dbgs() << "Dead: Generic opcode with no uses.\n");
852 return true;
853 }
854
855 if (isOpcodeWithNoSideEffects(MI.getOpcode())) {
856 LLVM_DEBUG(dbgs() << "Dead: known opcode with no side effects\n");
857 return true;
858 }
859
860 return false;
861}
862
863void SPIRVInstructionSelector::removeOpNamesForDeadMI(MachineInstr &MI) const {
864 // Delete the OpName that uses the result if there is one.
865 for (const auto &MO : MI.all_defs()) {
866 Register Reg = MO.getReg();
867 if (Reg.isPhysical())
868 continue;
869 SmallVector<MachineInstr *, 4> UselessOpNames;
870 for (MachineInstr &UseMI : MRI->use_nodbg_instructions(Reg)) {
871 assert(UseMI.getOpcode() == SPIRV::OpName &&
872 "There is still a use of the dead function.");
873 UselessOpNames.push_back(&UseMI);
874 }
875 for (MachineInstr *OpNameMI : UselessOpNames) {
876 GR.invalidateMachineInstr(OpNameMI);
877 OpNameMI->eraseFromParent();
878 }
879 }
880}
881
882void SPIRVInstructionSelector::removeDeadInstruction(MachineInstr &MI) const {
883 salvageDebugInfo(*MRI, MI);
884 GR.invalidateMachineInstr(&MI);
885 removeOpNamesForDeadMI(MI);
886 MI.eraseFromParent();
887}
888
889bool SPIRVInstructionSelector::select(MachineInstr &I) {
890 resetVRegsType(*I.getParent()->getParent());
891
892 assert(I.getParent() && "Instruction should be in a basic block!");
893 assert(I.getParent()->getParent() && "Instruction should be in a function!");
894
895 LLVM_DEBUG(dbgs() << "Checking if instruction is dead: " << I;);
896 if (isDead(I, *MRI)) {
897 LLVM_DEBUG(dbgs() << "Instruction is dead.\n");
898 removeDeadInstruction(I);
899 return true;
900 }
901
902 Register Opcode = I.getOpcode();
903 // If it's not a GMIR instruction, we've selected it already.
904 if (!isPreISelGenericOpcode(Opcode)) {
905 if (Opcode == SPIRV::ASSIGN_TYPE) { // These pseudos aren't needed any more.
906 Register DstReg = I.getOperand(0).getReg();
907 Register SrcReg = I.getOperand(1).getReg();
908 auto *Def = MRI->getVRegDef(SrcReg);
909 if (isTypeFoldingSupported(Def->getOpcode()) &&
910 Def->getOpcode() != TargetOpcode::G_CONSTANT &&
911 Def->getOpcode() != TargetOpcode::G_FCONSTANT) {
912 if (Def->getOpcode() == TargetOpcode::G_SELECT) {
913 Register SelectDstReg = Def->getOperand(0).getReg();
914 bool SuccessToSelectSelect [[maybe_unused]] = selectSelect(
915 SelectDstReg, GR.getSPIRVTypeForVReg(SelectDstReg), *Def);
916 assert(SuccessToSelectSelect);
917 GR.invalidateMachineInstr(Def);
918 Def->eraseFromParent();
919 MRI->replaceRegWith(DstReg, SelectDstReg);
920 GR.invalidateMachineInstr(&I);
921 I.eraseFromParent();
922 return true;
923 }
924
925 bool Res = selectImpl(I, *CoverageInfo);
926 LLVM_DEBUG({
927 if (!Res && Def->getOpcode() != TargetOpcode::G_CONSTANT) {
928 dbgs() << "Unexpected pattern in ASSIGN_TYPE.\nInstruction: ";
929 I.print(dbgs());
930 }
931 });
932 assert(Res || Def->getOpcode() == TargetOpcode::G_CONSTANT);
933 if (Res) {
934 if (!isTriviallyDead(*Def, *MRI) && isDead(*Def, *MRI))
935 DeadMIs.insert(Def);
936 return Res;
937 }
938 }
939 MRI->setRegClass(SrcReg, MRI->getRegClass(DstReg));
940 MRI->replaceRegWith(SrcReg, DstReg);
941 GR.invalidateMachineInstr(&I);
942 I.eraseFromParent();
943 return true;
944 } else if (I.getNumDefs() == 1) {
945 // Make all vregs 64 bits (for SPIR-V IDs).
946 MRI->setType(I.getOperand(0).getReg(), LLT::scalar(64));
947 }
948 constrainSelectedInstRegOperands(I, TII, TRI, RBI);
949 return true;
950 }
951
952 if (DeadMIs.contains(&I)) {
953 // if the instruction has been already made dead by folding it away
954 // erase it
955 LLVM_DEBUG(dbgs() << "Instruction is folded and dead.\n");
956 removeDeadInstruction(I);
957 DeadMIs.erase(&I);
958 return true;
959 }
960
961 if (I.getNumOperands() != I.getNumExplicitOperands()) {
962 LLVM_DEBUG(errs() << "Generic instr has unexpected implicit operands\n");
963 return false;
964 }
965
966 // Common code for getting return reg+type, and removing selected instr
967 // from parent occurs here. Instr-specific selection happens in spvSelect().
968 bool HasDefs = I.getNumDefs() > 0;
969 Register ResVReg = HasDefs ? I.getOperand(0).getReg() : Register(0);
970 SPIRVTypeInst ResType = HasDefs ? GR.getSPIRVTypeForVReg(ResVReg) : nullptr;
971 assert(!HasDefs || ResType || I.getOpcode() == TargetOpcode::G_GLOBAL_VALUE ||
972 I.getOpcode() == TargetOpcode::G_IMPLICIT_DEF);
973 if (spvSelect(ResVReg, ResType, I)) {
974 if (HasDefs) // Make all vregs 64 bits (for SPIR-V IDs).
975 for (unsigned i = 0; i < I.getNumDefs(); ++i)
976 MRI->setType(I.getOperand(i).getReg(), LLT::scalar(64));
977 GR.invalidateMachineInstr(&I);
978 I.eraseFromParent();
979 return true;
980 }
981 return false;
982}
983
984static bool mayApplyGenericSelection(unsigned Opcode) {
985 switch (Opcode) {
986 case TargetOpcode::G_CONSTANT:
987 case TargetOpcode::G_FCONSTANT:
988 return false;
989 }
990 return isTypeFoldingSupported(Opcode);
991}
992
993bool SPIRVInstructionSelector::BuildCOPY(Register DestReg, Register SrcReg,
994 MachineInstr &I) const {
995 const TargetRegisterClass *DstRC = MRI->getRegClassOrNull(DestReg);
996 const TargetRegisterClass *SrcRC = MRI->getRegClassOrNull(SrcReg);
997 if (DstRC != SrcRC && SrcRC)
998 MRI->setRegClass(DestReg, SrcRC);
999 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(TargetOpcode::COPY))
1000 .addDef(DestReg)
1001 .addUse(SrcReg)
1002 .constrainAllUses(TII, TRI, RBI);
1003 return true;
1004}
1005
1006bool SPIRVInstructionSelector::spvSelect(Register ResVReg,
1007 SPIRVTypeInst ResType,
1008 MachineInstr &I) const {
1009 const unsigned Opcode = I.getOpcode();
1010 if (mayApplyGenericSelection(Opcode))
1011 return selectImpl(I, *CoverageInfo);
1012 switch (Opcode) {
1013 case TargetOpcode::G_CONSTANT:
1014 case TargetOpcode::G_FCONSTANT:
1015 return selectConst(ResVReg, ResType, I);
1016 case TargetOpcode::G_GLOBAL_VALUE:
1017 return selectGlobalValue(ResVReg, I);
1018 case TargetOpcode::G_IMPLICIT_DEF:
1019 return selectOpUndef(ResVReg, ResType, I);
1020 case TargetOpcode::G_FREEZE:
1021 return selectFreeze(ResVReg, ResType, I);
1022
1023 case TargetOpcode::G_INTRINSIC:
1024 case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
1025 case TargetOpcode::G_INTRINSIC_CONVERGENT:
1026 case TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS:
1027 return selectIntrinsic(ResVReg, ResType, I);
1028 case TargetOpcode::G_BITREVERSE:
1029 return selectBitreverse(ResVReg, ResType, I);
1030
1031 case TargetOpcode::G_BUILD_VECTOR:
1032 return selectBuildVector(ResVReg, ResType, I);
1033 case TargetOpcode::G_SPLAT_VECTOR:
1034 return selectSplatVector(ResVReg, ResType, I);
1035 case TargetOpcode::G_CONCAT_VECTORS:
1036 return selectConcatVectors(ResVReg, ResType, I);
1037
1038 case TargetOpcode::G_SHUFFLE_VECTOR: {
1039 MachineBasicBlock &BB = *I.getParent();
1040 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorShuffle))
1041 .addDef(ResVReg)
1042 .addUse(GR.getSPIRVTypeID(ResType))
1043 .addUse(I.getOperand(1).getReg())
1044 .addUse(I.getOperand(2).getReg());
1045 for (auto V : I.getOperand(3).getShuffleMask())
1046 MIB.addImm(V);
1047 MIB.constrainAllUses(TII, TRI, RBI);
1048 return true;
1049 }
1050 case TargetOpcode::G_MEMMOVE:
1051 case TargetOpcode::G_MEMCPY:
1052 case TargetOpcode::G_MEMCPY_INLINE:
1053 case TargetOpcode::G_MEMSET:
1054 case TargetOpcode::G_MEMSET_INLINE:
1055 return selectMemOperation(ResVReg, I);
1056
1057 case TargetOpcode::G_ICMP:
1058 return selectICmp(ResVReg, ResType, I);
1059 case TargetOpcode::G_FCMP:
1060 return selectFCmp(ResVReg, ResType, I);
1061
1062 case TargetOpcode::G_FRAME_INDEX:
1063 return selectFrameIndex(ResVReg, ResType, I);
1064
1065 case TargetOpcode::G_LOAD:
1066 return selectLoad(ResVReg, ResType, I);
1067 case TargetOpcode::G_STORE:
1068 return selectStore(I);
1069
1070 case TargetOpcode::G_BR:
1071 return selectBranch(I);
1072 case TargetOpcode::G_BRCOND:
1073 return selectBranchCond(I);
1074
1075 case TargetOpcode::G_PHI:
1076 return selectPhi(ResVReg, I);
1077
1078 case TargetOpcode::G_FPTOSI:
1079 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertFToS);
1080 case TargetOpcode::G_FPTOUI:
1081 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertFToU);
1082
1083 case TargetOpcode::G_FPTOSI_SAT:
1084 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertFToS);
1085 case TargetOpcode::G_FPTOUI_SAT:
1086 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertFToU);
1087
1088 case TargetOpcode::G_SITOFP:
1089 return selectIToF(ResVReg, ResType, I, true, SPIRV::OpConvertSToF);
1090 case TargetOpcode::G_UITOFP:
1091 return selectIToF(ResVReg, ResType, I, false, SPIRV::OpConvertUToF);
1092
1093 case TargetOpcode::G_CTPOP:
1094 return selectPopCount(ResVReg, ResType, I, SPIRV::OpBitCount);
1095 case TargetOpcode::G_SMIN:
1096 return selectExtInst(ResVReg, ResType, I, CL::s_min, GL::SMin);
1097 case TargetOpcode::G_UMIN:
1098 return selectExtInst(ResVReg, ResType, I, CL::u_min, GL::UMin);
1099
1100 case TargetOpcode::G_SMAX:
1101 return selectExtInst(ResVReg, ResType, I, CL::s_max, GL::SMax);
1102 case TargetOpcode::G_UMAX:
1103 return selectExtInst(ResVReg, ResType, I, CL::u_max, GL::UMax);
1104
1105 case TargetOpcode::G_SCMP:
1106 return selectSUCmp(ResVReg, ResType, I, true);
1107 case TargetOpcode::G_UCMP:
1108 return selectSUCmp(ResVReg, ResType, I, false);
1109 case TargetOpcode::G_LROUND:
1110 case TargetOpcode::G_LLROUND: {
1111 Register regForLround =
1112 MRI->createVirtualRegister(MRI->getRegClass(ResVReg), "lround");
1113 MRI->setRegClass(regForLround, &SPIRV::iIDRegClass);
1114 GR.assignSPIRVTypeToVReg(GR.getSPIRVTypeForVReg(I.getOperand(1).getReg()),
1115 regForLround, *(I.getParent()->getParent()));
1116 selectExtInst(regForLround, GR.getSPIRVTypeForVReg(regForLround), I,
1117 CL::round, GL::Round, /* setMIFlags */ false);
1118 MachineBasicBlock &BB = *I.getParent();
1119 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConvertFToS))
1120 .addDef(ResVReg)
1121 .addUse(GR.getSPIRVTypeID(ResType))
1122 .addUse(regForLround);
1123 MIB.constrainAllUses(TII, TRI, RBI);
1124 return true;
1125 }
1126 case TargetOpcode::G_STRICT_FMA:
1127 case TargetOpcode::G_FMA: {
1128 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_fma)) {
1129 MachineBasicBlock &BB = *I.getParent();
1130 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpFmaKHR))
1131 .addDef(ResVReg)
1132 .addUse(GR.getSPIRVTypeID(ResType))
1133 .addUse(I.getOperand(1).getReg())
1134 .addUse(I.getOperand(2).getReg())
1135 .addUse(I.getOperand(3).getReg())
1136 .setMIFlags(I.getFlags());
1137 MIB.constrainAllUses(TII, TRI, RBI);
1138 return true;
1139 }
1140 return selectExtInst(ResVReg, ResType, I, CL::fma, GL::Fma);
1141 }
1142
1143 case TargetOpcode::G_FLDEXP:
1144 case TargetOpcode::G_STRICT_FLDEXP:
1145 return selectLdexp(ResVReg, ResType, I);
1146
1147 case TargetOpcode::G_FPOW:
1148 return selectExtInst(ResVReg, ResType, I, CL::pow, GL::Pow);
1149 case TargetOpcode::G_FPOWI:
1150 return selectFpowi(ResVReg, ResType, I);
1151
1152 case TargetOpcode::G_FEXP:
1153 return selectExtInst(ResVReg, ResType, I, CL::exp, GL::Exp);
1154 case TargetOpcode::G_FEXP2:
1155 return selectExtInst(ResVReg, ResType, I, CL::exp2, GL::Exp2);
1156 case TargetOpcode::G_FEXP10:
1157 return selectExp10(ResVReg, ResType, I);
1158
1159 case TargetOpcode::G_FMODF:
1160 return selectModf(ResVReg, ResType, I);
1161 case TargetOpcode::G_FSINCOS:
1162 return selectSincos(ResVReg, ResType, I);
1163
1164 case TargetOpcode::G_FLOG:
1165 return selectExtInst(ResVReg, ResType, I, CL::log, GL::Log);
1166 case TargetOpcode::G_FLOG2:
1167 return selectExtInst(ResVReg, ResType, I, CL::log2, GL::Log2);
1168 case TargetOpcode::G_FLOG10:
1169 return selectLog10(ResVReg, ResType, I);
1170
1171 case TargetOpcode::G_FABS:
1172 return selectExtInst(ResVReg, ResType, I, CL::fabs, GL::FAbs);
1173 case TargetOpcode::G_ABS:
1174 return selectExtInst(ResVReg, ResType, I, CL::s_abs, GL::SAbs);
1175
1176 case TargetOpcode::G_FMINNUM:
1177 case TargetOpcode::G_FMINIMUM:
1178 return selectExtInst(ResVReg, ResType, I, CL::fmin, GL::NMin);
1179 case TargetOpcode::G_FMAXNUM:
1180 case TargetOpcode::G_FMAXIMUM:
1181 return selectExtInst(ResVReg, ResType, I, CL::fmax, GL::NMax);
1182
1183 case TargetOpcode::G_FCOPYSIGN:
1184 return selectCopySign(ResVReg, ResType, I);
1185
1186 case TargetOpcode::G_FCEIL:
1187 return selectExtInst(ResVReg, ResType, I, CL::ceil, GL::Ceil);
1188 case TargetOpcode::G_FFLOOR:
1189 return selectExtInst(ResVReg, ResType, I, CL::floor, GL::Floor);
1190
1191 case TargetOpcode::G_FCOS:
1192 return selectExtInst(ResVReg, ResType, I, CL::cos, GL::Cos);
1193 case TargetOpcode::G_FSIN:
1194 return selectExtInst(ResVReg, ResType, I, CL::sin, GL::Sin);
1195 case TargetOpcode::G_FTAN:
1196 return selectExtInst(ResVReg, ResType, I, CL::tan, GL::Tan);
1197 case TargetOpcode::G_FACOS:
1198 return selectExtInst(ResVReg, ResType, I, CL::acos, GL::Acos);
1199 case TargetOpcode::G_FASIN:
1200 return selectExtInst(ResVReg, ResType, I, CL::asin, GL::Asin);
1201 case TargetOpcode::G_FATAN:
1202 return selectExtInst(ResVReg, ResType, I, CL::atan, GL::Atan);
1203 case TargetOpcode::G_FATAN2:
1204 return selectExtInst(ResVReg, ResType, I, CL::atan2, GL::Atan2);
1205 case TargetOpcode::G_FCOSH:
1206 return selectExtInst(ResVReg, ResType, I, CL::cosh, GL::Cosh);
1207 case TargetOpcode::G_FSINH:
1208 return selectExtInst(ResVReg, ResType, I, CL::sinh, GL::Sinh);
1209 case TargetOpcode::G_FTANH:
1210 return selectExtInst(ResVReg, ResType, I, CL::tanh, GL::Tanh);
1211
1212 case TargetOpcode::G_STRICT_FSQRT:
1213 case TargetOpcode::G_FSQRT:
1214 return selectExtInst(ResVReg, ResType, I, CL::sqrt, GL::Sqrt);
1215
1216 case TargetOpcode::G_CTTZ:
1217 case TargetOpcode::G_CTTZ_ZERO_POISON:
1218 return selectExtInst(ResVReg, ResType, I, CL::ctz);
1219 case TargetOpcode::G_CTLZ:
1220 case TargetOpcode::G_CTLZ_ZERO_POISON:
1221 return selectExtInst(ResVReg, ResType, I, CL::clz);
1222
1223 case TargetOpcode::G_INTRINSIC_ROUND:
1224 return selectExtInst(ResVReg, ResType, I, CL::round, GL::Round);
1225 case TargetOpcode::G_INTRINSIC_ROUNDEVEN:
1226 return selectExtInst(ResVReg, ResType, I, CL::rint, GL::RoundEven);
1227 case TargetOpcode::G_INTRINSIC_TRUNC:
1228 return selectExtInst(ResVReg, ResType, I, CL::trunc, GL::Trunc);
1229 case TargetOpcode::G_FRINT:
1230 case TargetOpcode::G_FNEARBYINT:
1231 return selectExtInst(ResVReg, ResType, I, CL::rint, GL::RoundEven);
1232
1233 case TargetOpcode::G_SMULH:
1234 return selectExtInst(ResVReg, ResType, I, CL::s_mul_hi);
1235 case TargetOpcode::G_UMULH:
1236 return selectExtInst(ResVReg, ResType, I, CL::u_mul_hi);
1237
1238 case TargetOpcode::G_SADDSAT:
1239 return selectExtInst(ResVReg, ResType, I, CL::s_add_sat);
1240 case TargetOpcode::G_UADDSAT:
1241 return selectExtInst(ResVReg, ResType, I, CL::u_add_sat);
1242 case TargetOpcode::G_SSUBSAT:
1243 return selectExtInst(ResVReg, ResType, I, CL::s_sub_sat);
1244 case TargetOpcode::G_USUBSAT:
1245 return selectExtInst(ResVReg, ResType, I, CL::u_sub_sat);
1246
1247 case TargetOpcode::G_FFREXP:
1248 return selectFrexp(ResVReg, ResType, I);
1249
1250 case TargetOpcode::G_UADDO:
1251 return selectOverflowArith(ResVReg, ResType, I,
1252 isVectorType(ResType) ? SPIRV::OpIAddCarryV
1253 : SPIRV::OpIAddCarryS);
1254 case TargetOpcode::G_USUBO:
1255 return selectOverflowArith(ResVReg, ResType, I,
1256 isVectorType(ResType) ? SPIRV::OpISubBorrowV
1257 : SPIRV::OpISubBorrowS);
1258 case TargetOpcode::G_UMULO:
1259 return selectOverflowArith(ResVReg, ResType, I, SPIRV::OpUMulExtended);
1260 case TargetOpcode::G_SMULO:
1261 return selectOverflowArith(ResVReg, ResType, I, SPIRV::OpSMulExtended);
1262
1263 case TargetOpcode::G_SEXT:
1264 return selectExt(ResVReg, ResType, I, true);
1265 case TargetOpcode::G_ANYEXT:
1266 case TargetOpcode::G_ZEXT:
1267 return selectExt(ResVReg, ResType, I, false);
1268 case TargetOpcode::G_TRUNC:
1269 return selectTrunc(ResVReg, ResType, I);
1270 case TargetOpcode::G_FPTRUNC:
1271 case TargetOpcode::G_FPEXT:
1272 return selectUnOp(ResVReg, ResType, I, SPIRV::OpFConvert);
1273
1274 case TargetOpcode::G_PTRTOINT:
1275 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertPtrToU);
1276 case TargetOpcode::G_INTTOPTR:
1277 return selectUnOp(ResVReg, ResType, I, SPIRV::OpConvertUToPtr);
1278 case TargetOpcode::G_BITCAST:
1279 return selectBitcast(ResVReg, ResType, I);
1280 case TargetOpcode::G_ADDRSPACE_CAST:
1281 return selectAddrSpaceCast(ResVReg, ResType, I);
1282 case TargetOpcode::G_PTRMASK:
1283 return selectPtrMask(ResVReg, ResType, I);
1284 case TargetOpcode::G_PTR_ADD: {
1285 // Currently, we get G_PTR_ADD only applied to global variables.
1286 assert(I.getOperand(1).isReg() && I.getOperand(2).isReg());
1287 Register GV = I.getOperand(1).getReg();
1288 MachineRegisterInfo::def_instr_iterator II = MRI->def_instr_begin(GV);
1289 (void)II;
1290 assert(((*II).getOpcode() == TargetOpcode::G_GLOBAL_VALUE ||
1291 (*II).getOpcode() == TargetOpcode::COPY ||
1292 (*II).getOpcode() == SPIRV::OpVariable ||
1293 (*II).getOpcode() == SPIRV::OpUntypedVariableKHR) &&
1294 getImm(I.getOperand(2), MRI));
1295 // It may be the initialization of a global variable.
1296 bool IsGVInit = false;
1297 for (MachineRegisterInfo::use_instr_iterator
1298 UseIt = MRI->use_instr_begin(I.getOperand(0).getReg()),
1299 UseEnd = MRI->use_instr_end();
1300 UseIt != UseEnd; UseIt = std::next(UseIt)) {
1301 if ((*UseIt).getOpcode() == TargetOpcode::G_GLOBAL_VALUE ||
1302 (*UseIt).getOpcode() == SPIRV::OpSpecConstantOp ||
1303 (*UseIt).getOpcode() == SPIRV::OpVariable ||
1304 (*UseIt).getOpcode() == SPIRV::OpUntypedVariableKHR) {
1305 IsGVInit = true;
1306 break;
1307 }
1308 }
1309 MachineBasicBlock &BB = *I.getParent();
1310 // An untyped result needs OpUntypedInBoundsPtrAccessChainKHR, which spells
1311 // out the Base Type. The offset of G_PTR_ADD is a byte count, so the Base
1312 // Type is i8 and the offset is the Element index as is. The opcode is
1313 // picked by the result type alone because an untyped access chain accepts
1314 // a typed base while a typed access chain rejects an untyped result.
1315 SPIRVTypeInst GVType = GR.getSPIRVTypeForVReg(GV);
1316 const bool UseUntypedPointers =
1317 ResType->getOpcode() == SPIRV::OpTypeUntypedPointerKHR;
1318 if (UseUntypedPointers) {
1319 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSpecConstantOp))
1320 .addDef(ResVReg)
1321 .addUse(GR.getSPIRVTypeID(ResType))
1322 .addImm(static_cast<uint32_t>(
1323 SPIRV::Opcode::UntypedInBoundsPtrAccessChainKHR))
1324 .addUse(GR.getSPIRVTypeID(GR.getOrCreateSPIRVIntegerType(8, I, TII)))
1325 .addUse(GV)
1326 .addUse(I.getOperand(2).getReg())
1327 .constrainAllUses(TII, TRI, RBI);
1328 return true;
1329 }
1330 if (!IsGVInit) {
1331 SPIRVTypeInst GVPointeeType = GR.getPointeeType(GVType);
1332 SPIRVTypeInst ResPointeeType = GR.getPointeeType(ResType);
1333 if (GVPointeeType && ResPointeeType && GVPointeeType != ResPointeeType) {
1334 // Build a new virtual register that is associated with the required
1335 // data type.
1336 Register NewVReg = MRI->createGenericVirtualRegister(MRI->getType(GV));
1337 MRI->setRegClass(NewVReg, MRI->getRegClass(GV));
1338 // Having a correctly typed base we are ready to build the actually
1339 // required GEP. It may not be a constant though, because all Operands
1340 // of OpSpecConstantOp is to originate from other const instructions,
1341 // and only the AccessChain named opcodes accept a global OpVariable
1342 // instruction. We can't use an AccessChain opcode because of the type
1343 // mismatch between result and base types.
1344 if (!GR.isBitcastCompatible(ResType, GVType))
1345 return diagnoseUnsupported(
1346 I, "incompatible result and operand types in a bitcast");
1347 Register ResTypeReg = GR.getSPIRVTypeID(ResType);
1348 MachineInstrBuilder MIB =
1349 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpBitcast))
1350 .addDef(NewVReg)
1351 .addUse(ResTypeReg)
1352 .addUse(GV);
1353 MIB.constrainAllUses(TII, TRI, RBI);
1354 BuildMI(BB, I, I.getDebugLoc(),
1355 TII.get(STI.isLogicalSPIRV() ? SPIRV::OpInBoundsAccessChain
1356 : SPIRV::OpInBoundsPtrAccessChain))
1357 .addDef(ResVReg)
1358 .addUse(ResTypeReg)
1359 .addUse(NewVReg)
1360 .addUse(I.getOperand(2).getReg())
1361 .constrainAllUses(TII, TRI, RBI);
1362 } else {
1363 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSpecConstantOp))
1364 .addDef(ResVReg)
1365 .addUse(GR.getSPIRVTypeID(ResType))
1366 .addImm(
1367 static_cast<uint32_t>(SPIRV::Opcode::InBoundsPtrAccessChain))
1368 .addUse(GV)
1369 .addUse(I.getOperand(2).getReg())
1370 .constrainAllUses(TII, TRI, RBI);
1371 }
1372 return true;
1373 }
1374 // It's possible to translate G_PTR_ADD to OpSpecConstantOp: either to
1375 // initialize a global variable with a constant expression (e.g., the test
1376 // case opencl/basic/progvar_prog_scope_init.ll), or for another use case
1377 Register Idx = buildZerosVal(GR.getOrCreateSPIRVIntegerType(32, I, TII), I);
1378 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSpecConstantOp))
1379 .addDef(ResVReg)
1380 .addUse(GR.getSPIRVTypeID(ResType))
1381 .addImm(static_cast<uint32_t>(
1382 SPIRV::Opcode::InBoundsPtrAccessChain))
1383 .addUse(GV)
1384 .addUse(Idx)
1385 .addUse(I.getOperand(2).getReg());
1386 MIB.constrainAllUses(TII, TRI, RBI);
1387 return true;
1388 }
1389
1390 case TargetOpcode::G_ATOMICRMW_OR:
1391 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicOr);
1392 case TargetOpcode::G_ATOMICRMW_ADD:
1393 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicIAdd);
1394 case TargetOpcode::G_ATOMICRMW_AND:
1395 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicAnd);
1396 case TargetOpcode::G_ATOMICRMW_MAX:
1397 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicSMax);
1398 case TargetOpcode::G_ATOMICRMW_MIN:
1399 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicSMin);
1400 case TargetOpcode::G_ATOMICRMW_SUB:
1401 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicISub);
1402 case TargetOpcode::G_ATOMICRMW_XOR:
1403 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicXor);
1404 case TargetOpcode::G_ATOMICRMW_UMAX:
1405 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicUMax);
1406 case TargetOpcode::G_ATOMICRMW_UMIN:
1407 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicUMin);
1408 case TargetOpcode::G_ATOMICRMW_XCHG:
1409 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicExchange);
1410
1411 case TargetOpcode::G_ATOMICRMW_FADD:
1412 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicFAddEXT);
1413 case TargetOpcode::G_ATOMICRMW_FSUB:
1414 // Translate G_ATOMICRMW_FSUB to OpAtomicFAddEXT with negative value operand
1415 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicFAddEXT,
1416 isVectorType(ResType) ? SPIRV::OpFNegateV
1417 : SPIRV::OpFNegate);
1418 case TargetOpcode::G_ATOMICRMW_FMIN:
1419 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicFMinEXT);
1420 case TargetOpcode::G_ATOMICRMW_FMAX:
1421 return selectAtomicRMW(ResVReg, ResType, I, SPIRV::OpAtomicFMaxEXT);
1422
1423 case TargetOpcode::G_FENCE:
1424 return selectFence(I);
1425
1426 case TargetOpcode::G_STACKSAVE:
1427 return selectStackSave(ResVReg, ResType, I);
1428 case TargetOpcode::G_STACKRESTORE:
1429 return selectStackRestore(I);
1430
1431 case TargetOpcode::G_UNMERGE_VALUES:
1432 return selectUnmergeValues(I);
1433
1434 case TargetOpcode::G_TRAP:
1435 case TargetOpcode::G_UBSANTRAP:
1436 return selectTrap(I);
1437
1438 // Discard gen opcodes for intrinsics which we do not expect to actually
1439 // represent code after lowering or intrinsics which are not implemented but
1440 // should not crash when found in a customer's LLVM IR input.
1441 case TargetOpcode::DBG_LABEL:
1442 return true;
1443 case TargetOpcode::G_DEBUGTRAP:
1444 return selectDebugTrap(ResVReg, ResType, I);
1445 case TargetOpcode::G_PREFETCH:
1446 return selectPrefetch(I);
1447
1448 default:
1449 return false;
1450 }
1451}
1452
1453bool SPIRVInstructionSelector::selectDebugTrap(Register ResVReg,
1454 SPIRVTypeInst ResType,
1455 MachineInstr &I) const {
1456 unsigned Opcode = SPIRV::OpNop;
1457 MachineBasicBlock &BB = *I.getParent();
1458 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
1459 .constrainAllUses(TII, TRI, RBI);
1460 return true;
1461}
1462
1463bool SPIRVInstructionSelector::selectPrefetch(MachineInstr &I) const {
1464 // llvm.prefetch is missing info about how many bytes to prefetch that
1465 // OpUntypedPrefetchKHR instruction requires, so just drop the instruction
1466 // in case untyped pointers are enabled since OCL prefetch cannot take
1467 // untyped pointers.
1468 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_untyped_pointers))
1469 return true;
1470
1471 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std)) {
1472 MachineIRBuilder MIRBuilder(I);
1473 const unsigned PointerSize = GR.getPointerSize();
1474 const SPIRVTypeInst PointerSizeType =
1475 GR.getOrCreateSPIRVIntegerType(PointerSize, MIRBuilder);
1476 bool ZeroAsNull = !STI.isShader();
1477
1478 // Assume that a single element needs prefetch.
1479 Register ConstIntOne =
1480 GR.getOrCreateConstInt(1, I, PointerSizeType, TII, ZeroAsNull);
1481
1482 Register AddrVal = I.getOperand(0).getReg();
1483 Register ExtReg = MRI->createVirtualRegister(
1484 GR.getRegClass(GR.getOpTypeVoid(MIRBuilder)));
1485 return selectExtInst(ExtReg, GR.getOpTypeVoid(MIRBuilder), I, CL::prefetch,
1486 /*setMIFlags=*/false, /*useMSSrc=*/false,
1487 {AddrVal, ConstIntOne});
1488 }
1489 return true;
1490}
1491
1492bool SPIRVInstructionSelector::selectExtInst(Register ResVReg,
1493 SPIRVTypeInst ResType,
1494 MachineInstr &I,
1495 GL::GLSLExtInst GLInst,
1496 bool setMIFlags, bool useMISrc,
1497 ArrayRef<Register> SrcRegs) const {
1498 if (!STI.canUseExtInstSet(
1499 SPIRV::InstructionSet::InstructionSet::GLSL_std_450))
1500 return diagnoseUnsupported(
1501 I,
1502 "this instruction is only supported with the GLSL extended instruction "
1503 "set.");
1504 return selectExtInst(ResVReg, ResType, I,
1505 {{SPIRV::InstructionSet::GLSL_std_450, GLInst}},
1506 setMIFlags, useMISrc, SrcRegs);
1507}
1508
1509bool SPIRVInstructionSelector::selectExtInst(Register ResVReg,
1510 SPIRVTypeInst ResType,
1511 MachineInstr &I,
1512 CL::OpenCLExtInst CLInst,
1513 bool setMIFlags, bool useMISrc,
1514 ArrayRef<Register> SrcRegs) const {
1515 return selectExtInst(ResVReg, ResType, I,
1516 {{SPIRV::InstructionSet::OpenCL_std, CLInst}},
1517 setMIFlags, useMISrc, SrcRegs);
1518}
1519
1520bool SPIRVInstructionSelector::selectExtInst(
1521 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
1522 CL::OpenCLExtInst CLInst, GL::GLSLExtInst GLInst, bool setMIFlags,
1523 bool useMISrc, ArrayRef<Register> SrcRegs) const {
1524 ExtInstList ExtInsts = {{SPIRV::InstructionSet::OpenCL_std, CLInst},
1525 {SPIRV::InstructionSet::GLSL_std_450, GLInst}};
1526 return selectExtInst(ResVReg, ResType, I, ExtInsts, setMIFlags, useMISrc,
1527 SrcRegs);
1528}
1529
1530bool SPIRVInstructionSelector::selectExtInst(Register ResVReg,
1531 SPIRVTypeInst ResType,
1532 MachineInstr &I,
1533 const ExtInstList &Insts,
1534 bool setMIFlags, bool useMISrc,
1535 ArrayRef<Register> SrcRegs) const {
1536
1537 for (const auto &[InstructionSet, Opcode] : Insts) {
1538 if (!STI.canUseExtInstSet(InstructionSet))
1539 continue;
1540 MachineBasicBlock &BB = *I.getParent();
1541 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
1542 .addDef(ResVReg)
1543 .addUse(GR.getSPIRVTypeID(ResType))
1544 .addImm(static_cast<uint32_t>(InstructionSet))
1545 .addImm(Opcode);
1546 if (setMIFlags)
1547 MIB.setMIFlags(I.getFlags());
1548 if (useMISrc) {
1549 const unsigned NumOps = I.getNumOperands();
1550 unsigned Index = 1;
1551 if (Index < NumOps &&
1552 I.getOperand(Index).getType() ==
1553 MachineOperand::MachineOperandType::MO_IntrinsicID)
1554 Index = 2;
1555 for (; Index < NumOps; ++Index)
1556 MIB.add(I.getOperand(Index));
1557 } else {
1558 for (Register SReg : SrcRegs) {
1559 MIB.addUse(SReg);
1560 }
1561 }
1562 MIB.constrainAllUses(TII, TRI, RBI);
1563 return true;
1564 }
1565 return false;
1566}
1567
1568bool SPIRVInstructionSelector::selectCopySign(Register ResVReg,
1569 SPIRVTypeInst ResType,
1570 MachineInstr &I) const {
1571 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std))
1572 return selectExtInst(ResVReg, ResType, I, CL::copysign);
1573
1574 // There is no copysign instruction in the GLSL Extended Instruction set, so
1575 // it is implemented with bit manipulation:
1576 // bitcast((bitcast(magnitude) & ~signBit) | (bitcast(sign) & signBit))
1577 Register MagnitudeReg = I.getOperand(1).getReg();
1578 Register SignReg = I.getOperand(2).getReg();
1579
1580 const unsigned BitWidth = GR.getScalarOrVectorBitWidth(ResType);
1581 const unsigned ComponentCount = GR.getScalarOrVectorComponentCount(ResType);
1582 SPIRVTypeInst IntType = GR.getOrCreateSPIRVIntegerType(BitWidth, I, TII);
1583 const APInt SignMaskVal = APInt::getSignMask(BitWidth);
1584
1585 Register SignMask, NotSignMask;
1586 unsigned AndOpcode, OrOpcode;
1587 if (ComponentCount > 1) {
1588 IntType = GR.getOrCreateSPIRVVectorType(IntType, ComponentCount, I, TII);
1589 SignMask = GR.getOrCreateConstVector(SignMaskVal, I, IntType, TII);
1590 NotSignMask = GR.getOrCreateConstVector(~SignMaskVal, I, IntType, TII);
1591 AndOpcode = SPIRV::OpBitwiseAndV;
1592 OrOpcode = SPIRV::OpBitwiseOrV;
1593 } else {
1594 SignMask = GR.getOrCreateConstInt(SignMaskVal, I, IntType, TII);
1595 NotSignMask = GR.getOrCreateConstInt(~SignMaskVal, I, IntType, TII);
1596 AndOpcode = SPIRV::OpBitwiseAndS;
1597 OrOpcode = SPIRV::OpBitwiseOrS;
1598 }
1599
1600 auto EmitBitOp = [&](Register &ResReg, ArrayRef<Register> SrcRegs,
1601 unsigned Opcode) {
1602 ResReg = MRI->createVirtualRegister(GR.getRegClass(IntType));
1603 return selectOpWithSrcs(ResReg, IntType, I, SrcRegs, Opcode);
1604 };
1605
1606 Register MagnitudeInt, SignInt, MagnitudeBits, SignBits, CombinedInt;
1607 if (!EmitBitOp(MagnitudeInt, {MagnitudeReg}, SPIRV::OpBitcast) ||
1608 !EmitBitOp(SignInt, {SignReg}, SPIRV::OpBitcast) ||
1609 !EmitBitOp(MagnitudeBits, {MagnitudeInt, NotSignMask}, AndOpcode) ||
1610 !EmitBitOp(SignBits, {SignInt, SignMask}, AndOpcode) ||
1611 !EmitBitOp(CombinedInt, {MagnitudeBits, SignBits}, OrOpcode))
1612 return false;
1613
1614 return selectOpWithSrcs(ResVReg, ResType, I, {CombinedInt}, SPIRV::OpBitcast);
1615}
1616
1617bool SPIRVInstructionSelector::selectFrexp(Register ResVReg,
1618 SPIRVTypeInst ResType,
1619 MachineInstr &I) const {
1620 ExtInstList ExtInsts = {{SPIRV::InstructionSet::OpenCL_std, CL::frexp},
1621 {SPIRV::InstructionSet::GLSL_std_450, GL::Frexp}};
1622 for (const auto &Ex : ExtInsts) {
1623 SPIRV::InstructionSet::InstructionSet Set = Ex.first;
1624 uint32_t Opcode = Ex.second;
1625 if (!STI.canUseExtInstSet(Set))
1626 continue;
1627
1628 MachineIRBuilder MIRBuilder(I);
1629 SPIRVTypeInst PointeeTy = GR.getSPIRVTypeForVReg(I.getOperand(1).getReg());
1630 const SPIRVTypeInst PointerType = GR.getOrCreateSPIRVPointerType(
1631 PointeeTy, MIRBuilder, SPIRV::StorageClass::Function);
1632 Register PointerVReg =
1633 createVirtualRegister(PointerType, &GR, MRI, MRI->getMF());
1634
1635 auto It = getOpVariableMBBIt(*I.getMF());
1636 // An untyped pointer result type is only legal on OpUntypedVariableKHR,
1637 // but not on OpVariable.
1638 const bool IsUntyped =
1639 PointerType->getOpcode() == SPIRV::OpTypeUntypedPointerKHR;
1640 auto VarMIB =
1641 BuildMI(*It->getParent(), It, It->getDebugLoc(),
1642 TII.get(IsUntyped ? SPIRV::OpUntypedVariableKHR
1643 : SPIRV::OpVariable))
1644 .addDef(PointerVReg)
1645 .addUse(GR.getSPIRVTypeID(PointerType))
1646 .addImm(static_cast<uint32_t>(SPIRV::StorageClass::Function));
1647 if (IsUntyped)
1648 VarMIB.addUse(GR.getSPIRVTypeID(PointeeTy)); // Data Type
1649 VarMIB.constrainAllUses(TII, TRI, RBI);
1650
1651 SPIRVTypeInst MantissaTy = GR.getSPIRVTypeForVReg(I.getOperand(2).getReg());
1652 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
1653 .addDef(ResVReg)
1654 .addUse(GR.getSPIRVTypeID(MantissaTy))
1655 .addImm(static_cast<uint32_t>(Ex.first))
1656 .addImm(Opcode)
1657 .add(I.getOperand(2))
1658 .addUse(PointerVReg)
1659 .constrainAllUses(TII, TRI, RBI);
1660
1661 Register ExpResReg = I.getOperand(1).getReg();
1662 if (!MRI->use_nodbg_empty(ExpResReg))
1663 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpLoad))
1664 .addDef(ExpResReg)
1665 .addUse(GR.getSPIRVTypeID(PointeeTy))
1666 .addUse(PointerVReg)
1667 .constrainAllUses(TII, TRI, RBI);
1668 return true;
1669 }
1670 return false;
1671}
1672
1673bool SPIRVInstructionSelector::selectLdexp(Register ResVReg,
1674 SPIRVTypeInst ResType,
1675 MachineInstr &I) const {
1676 Register XReg = I.getOperand(1).getReg();
1677 Register ExpReg = I.getOperand(2).getReg();
1678
1679 // Both OpenCL.std ldexp and GLSL.std.450 Ldexp require the exponent to have
1680 // the same number of components as the result, but G_FLDEXP's exponent is
1681 // scalar even for vector operands. Splat it when the result is a vector.
1682 SPIRVTypeInst ExpType = GR.getSPIRVTypeForVReg(ExpReg);
1683 if (ResType->getOpcode() == SPIRV::OpTypeVector &&
1684 ExpType->getOpcode() != SPIRV::OpTypeVector) {
1685 unsigned NumElts = ResType->getOperand(2).getImm();
1686 SPIRVTypeInst ExpVecType =
1687 GR.getOrCreateSPIRVVectorType(ExpType, NumElts, I, TII);
1688 Register SplatReg = MRI->createVirtualRegister(GR.getRegClass(ExpVecType));
1689 GR.assignSPIRVTypeToVReg(ExpVecType, SplatReg, MRI->getMF());
1690 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
1691 TII.get(SPIRV::OpCompositeConstruct))
1692 .addDef(SplatReg)
1693 .addUse(GR.getSPIRVTypeID(ExpVecType));
1694 for (unsigned J = 0; J < NumElts; ++J)
1695 MIB.addUse(ExpReg);
1696 MIB.constrainAllUses(TII, TRI, RBI);
1697 ExpReg = SplatReg;
1698 }
1699
1700 return selectExtInst(ResVReg, ResType, I, CL::ldexp, GL::Ldexp,
1701 /*setMIFlags=*/true, /*useMISrc=*/false, {XReg, ExpReg});
1702}
1703
1704bool SPIRVInstructionSelector::selectSincos(Register ResVReg,
1705 SPIRVTypeInst ResType,
1706 MachineInstr &I) const {
1707 Register CosResVReg = I.getOperand(1).getReg();
1708 unsigned SrcIdx = I.getNumExplicitDefs();
1709 Register ResTypeReg = GR.getSPIRVTypeID(ResType);
1710
1711 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std)) {
1712 // OpenCL.std sincos(x, cosval*) -> returns sin(x), writes cos(x) to ptr.
1713 MachineIRBuilder MIRBuilder(I);
1714 const SPIRVTypeInst PointerType = GR.getOrCreateSPIRVPointerType(
1715 ResType, MIRBuilder, SPIRV::StorageClass::Function);
1716 Register PointerVReg =
1717 createVirtualRegister(PointerType, &GR, MRI, MRI->getMF());
1718
1719 auto It = getOpVariableMBBIt(*I.getMF());
1720 // An untyped pointer result type is only legal on OpUntypedVariableKHR,
1721 // but not on OpVariable.
1722 const bool IsUntyped =
1723 PointerType->getOpcode() == SPIRV::OpTypeUntypedPointerKHR;
1724 auto VarMIB =
1725 BuildMI(*It->getParent(), It, It->getDebugLoc(),
1726 TII.get(IsUntyped ? SPIRV::OpUntypedVariableKHR
1727 : SPIRV::OpVariable))
1728 .addDef(PointerVReg)
1729 .addUse(GR.getSPIRVTypeID(PointerType))
1730 .addImm(static_cast<uint32_t>(SPIRV::StorageClass::Function));
1731 if (IsUntyped)
1732 VarMIB.addUse(GR.getSPIRVTypeID(ResType)); // Data Type
1733 VarMIB.constrainAllUses(TII, TRI, RBI);
1734 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
1735 .addDef(ResVReg)
1736 .addUse(ResTypeReg)
1737 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::OpenCL_std))
1738 .addImm(CL::sincos)
1739 .add(I.getOperand(SrcIdx))
1740 .addUse(PointerVReg)
1741 .constrainAllUses(TII, TRI, RBI);
1742 if (!MRI->use_nodbg_empty(CosResVReg))
1743 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpLoad))
1744 .addDef(CosResVReg)
1745 .addUse(ResTypeReg)
1746 .addUse(PointerVReg)
1747 .constrainAllUses(TII, TRI, RBI);
1748 return true;
1749 } else if (STI.canUseExtInstSet(SPIRV::InstructionSet::GLSL_std_450)) {
1750 // GLSL.std.450 has no combined sincos; emit separate Sin and Cos.
1751 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
1752 .addDef(ResVReg)
1753 .addUse(ResTypeReg)
1754 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
1755 .addImm(GL::Sin)
1756 .add(I.getOperand(SrcIdx))
1757 .constrainAllUses(TII, TRI, RBI);
1758 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
1759 .addDef(CosResVReg)
1760 .addUse(ResTypeReg)
1761 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
1762 .addImm(GL::Cos)
1763 .add(I.getOperand(SrcIdx))
1764 .constrainAllUses(TII, TRI, RBI);
1765 return true;
1766 }
1767 return false;
1768}
1769
1770bool SPIRVInstructionSelector::selectOpWithSrcs(Register ResVReg,
1771 SPIRVTypeInst ResType,
1772 MachineInstr &I,
1773 ArrayRef<Register> Srcs,
1774 unsigned Opcode) const {
1775 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
1776 .addDef(ResVReg)
1777 .addUse(GR.getSPIRVTypeID(ResType));
1778 for (Register SReg : Srcs) {
1779 MIB.addUse(SReg);
1780 }
1781 MIB.constrainAllUses(TII, TRI, RBI);
1782 return true;
1783}
1784
1785bool SPIRVInstructionSelector::selectPopCount16(Register ResVReg,
1786 SPIRVTypeInst ResType,
1787 MachineInstr &I,
1788 unsigned ExtOpcode,
1789 unsigned Opcode) const {
1790 MachineIRBuilder MIRBuilder(I);
1791
1792 Register OpReg = I.getOperand(1).getReg();
1793 SPIRVTypeInst SrcType = GR.getSPIRVTypeForVReg(OpReg);
1794 unsigned ComponentCount = GR.getScalarOrVectorComponentCount(SrcType);
1795 bool IsScalar = !isVectorType(SrcType);
1796 SPIRVTypeInst I32Type = GR.getOrCreateSPIRVIntegerType(32, MIRBuilder);
1797
1798 SPIRVTypeInst ExtType =
1799 IsScalar ? I32Type
1800 : GR.getOrCreateSPIRVVectorType(I32Type, ComponentCount,
1801 MIRBuilder, /*IsSigned=*/false);
1802
1803 Register ExtReg = MRI->createVirtualRegister(GR.getRegClass(ExtType));
1804 // Always use OpUConvert to always use a 0 extend
1805 if (!selectOpWithSrcs(ExtReg, ExtType, I, {OpReg}, SPIRV::OpUConvert))
1806 return false;
1807
1808 Register PopCountReg = MRI->createVirtualRegister(GR.getRegClass(ExtType));
1809 if (!selectPopCount32(PopCountReg, ExtType, I, ExtReg, Opcode))
1810 return false;
1811
1812 return selectOpWithSrcs(ResVReg, ResType, I, {PopCountReg}, ExtOpcode);
1813}
1814
1815bool SPIRVInstructionSelector::selectPopCount32(Register ResVReg,
1816 SPIRVTypeInst ResType,
1817 MachineInstr &I,
1818 Register SrcReg,
1819 unsigned Opcode) const {
1820 return selectOpWithSrcs(ResVReg, ResType, I, {SrcReg}, Opcode);
1821}
1822
1823bool SPIRVInstructionSelector::selectPopCount64(Register ResVReg,
1824 SPIRVTypeInst ResType,
1825 MachineInstr &I,
1826 Register SrcReg,
1827 unsigned Opcode) const {
1828 MachineIRBuilder MIRBuilder(I);
1829
1830 SPIRVTypeInst SrcType = GR.getSPIRVTypeForVReg(SrcReg);
1831 unsigned ComponentCount = GR.getScalarOrVectorComponentCount(SrcType);
1832 bool IsScalar = !isVectorType(SrcType);
1833 SPIRVTypeInst I32Type = GR.getOrCreateSPIRVIntegerType(32, MIRBuilder);
1834
1835 // We need to work with a type matching the shape of the input but using 32
1836 // bit int instead of 64.
1837 SPIRVTypeInst WorkingType =
1838 IsScalar ? I32Type
1839 : GR.getOrCreateSPIRVVectorType(I32Type, ComponentCount,
1840 MIRBuilder, /*IsSigned=*/false);
1841
1842 // Truncate and count the low bits.
1843 Register Trunc = MRI->createVirtualRegister(GR.getRegClass(WorkingType));
1844 if (!selectOpWithSrcs(Trunc, WorkingType, I, {SrcReg}, SPIRV::OpUConvert))
1845 return false;
1846
1847 Register LowCount = MRI->createVirtualRegister(GR.getRegClass(WorkingType));
1848 if (!selectOpWithSrcs(LowCount, WorkingType, I, {Trunc}, SPIRV::OpBitCount))
1849 return false;
1850
1851 // Shift the high bits over and count them too.
1852 Register ShiftAmount = IsScalar
1853 ? GR.getOrCreateConstInt(32, I, SrcType, TII)
1854 : GR.getOrCreateConstVector(32, I, SrcType, TII);
1855 unsigned ShiftOp =
1856 IsScalar ? SPIRV::OpShiftRightLogicalS : SPIRV::OpShiftRightLogicalV;
1857 Register Shift = MRI->createVirtualRegister(GR.getRegClass(SrcType));
1858 if (!selectOpWithSrcs(Shift, SrcType, I, {SrcReg, ShiftAmount}, ShiftOp))
1859 return false;
1860
1861 Trunc = MRI->createVirtualRegister(GR.getRegClass(WorkingType));
1862 if (!selectOpWithSrcs(Trunc, WorkingType, I, {Shift}, SPIRV::OpUConvert))
1863 return false;
1864
1865 Register HighCount = MRI->createVirtualRegister(GR.getRegClass(WorkingType));
1866 if (!selectOpWithSrcs(HighCount, WorkingType, I, {Trunc}, SPIRV::OpBitCount))
1867 return false;
1868
1869 // Add them up and zext or sext back to 64 bit values.
1870 Register Sum = MRI->createVirtualRegister(GR.getRegClass(WorkingType));
1871 if (!selectOpWithSrcs(Sum, WorkingType, I, {HighCount, LowCount},
1872 IsScalar ? SPIRV::OpIAddS : SPIRV::OpIAddV))
1873 return false;
1874
1875 bool IsSigned = GR.isScalarOrVectorSigned(ResType);
1876 unsigned ConvOp = IsSigned ? SPIRV::OpSConvert : SPIRV::OpUConvert;
1877 return selectOpWithSrcs(ResVReg, ResType, I, {Sum}, ConvOp);
1878}
1879
1880bool SPIRVInstructionSelector::selectPopCount(Register ResVReg,
1881 SPIRVTypeInst ResType,
1882 MachineInstr &I,
1883 unsigned Opcode) const {
1884 // Vulkan restricts OpBitCount to 32-bit integers or vectors of 32-bit
1885 // integers unless VK_KHR_maintenance9 is enabled. Until VK_KHR_maintenance9
1886 // is core we will not generate OpBitCount with any other types when
1887 // targeting Vulkan.
1888 if (!STI.getTargetTriple().isVulkanOS())
1889 return selectUnOp(ResVReg, ResType, I, Opcode);
1890
1891 Register OpReg = I.getOperand(1).getReg();
1892 SPIRVTypeInst OpType = GR.getSPIRVTypeForVReg(OpReg);
1893 unsigned ExtOpcode = GR.isScalarOrVectorSigned(ResType) ? SPIRV::OpSConvert
1894 : SPIRV::OpUConvert;
1895 switch (GR.getScalarOrVectorBitWidth(OpType)) {
1896 case 8:
1897 case 16:
1898 return selectPopCount16(ResVReg, ResType, I, ExtOpcode, Opcode);
1899 case 32:
1900 return selectPopCount32(ResVReg, ResType, I, OpReg, Opcode);
1901 case 64:
1902 return selectPopCount64(ResVReg, ResType, I, OpReg, Opcode);
1903 default:
1904 return diagnoseUnsupported(I, "unsupported operand bit width for popcount");
1905 }
1906}
1907
1908bool SPIRVInstructionSelector::selectUnOp(Register ResVReg,
1909 SPIRVTypeInst ResType,
1910 MachineInstr &I,
1911 unsigned Opcode) const {
1912 if (STI.isPhysicalSPIRV() && I.getOperand(1).isReg()) {
1913 Register SrcReg = I.getOperand(1).getReg();
1914 bool IsGV = false;
1915 for (MachineRegisterInfo::def_instr_iterator DefIt =
1916 MRI->def_instr_begin(SrcReg);
1917 DefIt != MRI->def_instr_end(); DefIt = std::next(DefIt)) {
1918 unsigned DefOpCode = DefIt->getOpcode();
1919 if (DefOpCode == SPIRV::ASSIGN_TYPE || DefOpCode == TargetOpcode::COPY) {
1920 // We need special handling to look through the type assignment or the
1921 // COPY pseudo-op and see if this is a constant or a global.
1922 if (auto *VRD = getVRegDef(*MRI, DefIt->getOperand(1).getReg()))
1923 DefOpCode = VRD->getOpcode();
1924 }
1925 if (DefOpCode == TargetOpcode::G_GLOBAL_VALUE ||
1926 DefOpCode == TargetOpcode::G_CONSTANT ||
1927 DefOpCode == SPIRV::OpVariable ||
1928 DefOpCode == SPIRV::OpUntypedVariableKHR ||
1929 DefOpCode == SPIRV::OpConstantI) {
1930 IsGV = true;
1931 break;
1932 }
1933 }
1934 if (IsGV) {
1935 uint32_t SpecOpcode = 0;
1936 switch (Opcode) {
1937 case SPIRV::OpConvertPtrToU:
1938 SpecOpcode = static_cast<uint32_t>(SPIRV::Opcode::ConvertPtrToU);
1939 break;
1940 case SPIRV::OpConvertUToPtr:
1941 SpecOpcode = static_cast<uint32_t>(SPIRV::Opcode::ConvertUToPtr);
1942 break;
1943 }
1944 if (SpecOpcode) {
1945 BuildMI(*I.getParent(), I, I.getDebugLoc(),
1946 TII.get(SPIRV::OpSpecConstantOp))
1947 .addDef(ResVReg)
1948 .addUse(GR.getSPIRVTypeID(ResType))
1949 .addImm(SpecOpcode)
1950 .addUse(SrcReg)
1951 .constrainAllUses(TII, TRI, RBI);
1952 return true;
1953 }
1954 }
1955 }
1956 return selectOpWithSrcs(ResVReg, ResType, I, {I.getOperand(1).getReg()},
1957 Opcode);
1958}
1959
1960bool SPIRVInstructionSelector::selectBitcast(Register ResVReg,
1961 SPIRVTypeInst ResType,
1962 MachineInstr &I) const {
1963 Register OpReg = I.getOperand(1).getReg();
1964 SPIRVTypeInst OpType =
1965 OpReg.isValid() ? GR.getSPIRVTypeForVReg(OpReg) : nullptr;
1966 if (!GR.isBitcastCompatible(ResType, OpType))
1967 return diagnoseUnsupported(
1968 I, "incompatible result and operand types in a bitcast");
1969 return selectUnOp(ResVReg, ResType, I, SPIRV::OpBitcast);
1970}
1971
1972static void addMemoryOperands(MachineMemOperand *MemOp,
1973 MachineInstrBuilder &MIB,
1974 MachineIRBuilder &MIRBuilder,
1975 SPIRVGlobalRegistry &GR) {
1976 const SPIRVSubtarget *ST =
1977 static_cast<const SPIRVSubtarget *>(&MIRBuilder.getMF().getSubtarget());
1978 uint32_t SpvMemOp = static_cast<uint32_t>(SPIRV::MemoryOperand::None);
1979 if (MemOp->isVolatile())
1980 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Volatile);
1981 if (MemOp->isNonTemporal())
1982 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Nontemporal);
1983 // Aligned memory operand requires the Kernel capability.
1984 if (!ST->isShader() && MemOp->getAlign().value())
1985 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Aligned);
1986
1987 [[maybe_unused]] MachineInstr *AliasList = nullptr;
1988 [[maybe_unused]] MachineInstr *NoAliasList = nullptr;
1989 if (ST->canUseExtension(SPIRV::Extension::SPV_INTEL_memory_access_aliasing)) {
1990 if (auto *MD = MemOp->getAAInfo().Scope) {
1991 AliasList = GR.getOrAddMemAliasingINTELInst(MIRBuilder, MD);
1992 if (AliasList)
1993 SpvMemOp |=
1994 static_cast<uint32_t>(SPIRV::MemoryOperand::AliasScopeINTELMask);
1995 }
1996 if (auto *MD = MemOp->getAAInfo().NoAlias) {
1997 NoAliasList = GR.getOrAddMemAliasingINTELInst(MIRBuilder, MD);
1998 if (NoAliasList)
1999 SpvMemOp |=
2000 static_cast<uint32_t>(SPIRV::MemoryOperand::NoAliasINTELMask);
2001 }
2002 }
2003
2004 if (SpvMemOp != static_cast<uint32_t>(SPIRV::MemoryOperand::None)) {
2005 MIB.addImm(SpvMemOp);
2006 if (SpvMemOp & static_cast<uint32_t>(SPIRV::MemoryOperand::Aligned))
2007 MIB.addImm(MemOp->getAlign().value());
2008 if (AliasList)
2009 MIB.addUse(AliasList->getOperand(0).getReg());
2010 if (NoAliasList)
2011 MIB.addUse(NoAliasList->getOperand(0).getReg());
2012 }
2013}
2014
2015static void addMemoryOperands(uint64_t Flags, MachineInstrBuilder &MIB) {
2016 uint32_t SpvMemOp = static_cast<uint32_t>(SPIRV::MemoryOperand::None);
2017 if (Flags & MachineMemOperand::Flags::MOVolatile)
2018 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Volatile);
2019 if (Flags & MachineMemOperand::Flags::MONonTemporal)
2020 SpvMemOp |= static_cast<uint32_t>(SPIRV::MemoryOperand::Nontemporal);
2021
2022 if (SpvMemOp != static_cast<uint32_t>(SPIRV::MemoryOperand::None))
2023 MIB.addImm(SpvMemOp);
2024}
2025
2026bool SPIRVInstructionSelector::selectLoad(Register ResVReg,
2027 SPIRVTypeInst ResType,
2028 MachineInstr &I) const {
2029 unsigned OpOffset = isa<GIntrinsic>(I) ? 1 : 0;
2030 Register Ptr = I.getOperand(1 + OpOffset).getReg();
2031
2032 auto *PtrDef = getVRegDef(*MRI, Ptr);
2033 auto *IntPtrDef = dyn_cast<GIntrinsic>(PtrDef);
2034 if (IntPtrDef &&
2035 (IntPtrDef->getIntrinsicID() == Intrinsic::spv_resource_getbasepointer ||
2036 IntPtrDef->getIntrinsicID() == Intrinsic::spv_resource_getpointer)) {
2037
2038 Register HandleReg = IntPtrDef->getOperand(2).getReg();
2039 SPIRVTypeInst HandleType = GR.getSPIRVTypeForVReg(HandleReg);
2040 if (HandleType->getOpcode() == SPIRV::OpTypeImage) {
2041 Register NewHandleReg =
2042 MRI->createVirtualRegister(MRI->getRegClass(HandleReg));
2043 auto *HandleDef = cast<GIntrinsic>(getVRegDef(*MRI, HandleReg));
2044 if (!loadHandleBeforePosition(NewHandleReg, HandleType, *HandleDef, I)) {
2045 return false;
2046 }
2047
2048 Register IdxReg = IntPtrDef->getOperand(3).getReg();
2049 return generateImageReadOrFetch(ResVReg, ResType, NewHandleReg, IdxReg,
2050 I.getDebugLoc(), I);
2051 }
2052 }
2053
2054 MachineIRBuilder MIRBuilder(I);
2055
2056 if (I.getNumMemOperands()) {
2057 const MachineMemOperand *MemOp = *I.memoperands_begin();
2058 if (MemOp->isAtomic())
2059 return selectAtomicLoad(ResVReg, ResType, I);
2060 }
2061
2062 auto MIB = MIRBuilder.buildInstr(SPIRV::OpLoad)
2063 .addDef(ResVReg)
2064 .addUse(GR.getSPIRVTypeID(ResType))
2065 .addUse(Ptr);
2066 if (!I.getNumMemOperands()) {
2067 assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS ||
2068 I.getOpcode() ==
2069 TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS);
2070 addMemoryOperands(I.getOperand(2 + OpOffset).getImm(), MIB);
2071 } else {
2072 addMemoryOperands(*I.memoperands_begin(), MIB, MIRBuilder, GR);
2073 }
2074 MIB.constrainAllUses(TII, TRI, RBI);
2075 return true;
2076}
2077
2078Register SPIRVInstructionSelector::createPtrSizedIntReg(
2079 MachineIRBuilder &MIRBuilder) const {
2080 SPIRVTypeInst IntType =
2081 GR.getOrCreateSPIRVIntegerType(GR.getPointerSize(), MIRBuilder);
2082 Register Reg =
2083 MRI->createGenericVirtualRegister(LLT::scalar(GR.getPointerSize()));
2084 MRI->setRegClass(Reg, GR.getRegClass(IntType));
2085 GR.assignSPIRVTypeToVReg(IntType, Reg, MIRBuilder.getMF());
2086 return Reg;
2087}
2088
2089Register
2090SPIRVInstructionSelector::convertPtrToInt(Register PtrVal,
2091 MachineIRBuilder &MIRBuilder) const {
2092 SPIRVTypeInst IntType =
2093 GR.getOrCreateSPIRVIntegerType(GR.getPointerSize(), MIRBuilder);
2094 Register IntReg = createPtrSizedIntReg(MIRBuilder);
2095 MIRBuilder.buildInstr(SPIRV::OpConvertPtrToU)
2096 .addDef(IntReg)
2097 .addUse(GR.getSPIRVTypeID(IntType)) // Result type
2098 .addUse(PtrVal) // Pointer operand
2099 .constrainAllUses(TII, TRI, RBI);
2100 return IntReg;
2101}
2102
2103Register SPIRVInstructionSelector::castPtrToPtrToInt(
2104 Register Ptr, SPIRV::StorageClass::StorageClass SC,
2105 MachineIRBuilder &MIRBuilder) const {
2106 SPIRVTypeInst IntType =
2107 GR.getOrCreateSPIRVIntegerType(GR.getPointerSize(), MIRBuilder);
2108 SPIRVTypeInst PtrType =
2109 GR.getOrCreateSPIRVPointerType(IntType, MIRBuilder, SC);
2110 Register CastedPtr =
2111 MRI->createGenericVirtualRegister(LLT::scalar(GR.getPointerSize()));
2112 MRI->setRegClass(CastedPtr, GR.getRegClass(PtrType));
2113 GR.assignSPIRVTypeToVReg(PtrType, CastedPtr, MIRBuilder.getMF());
2114 MIRBuilder.buildInstr(SPIRV::OpBitcast)
2115 .addDef(CastedPtr)
2116 .addUse(GR.getSPIRVTypeID(PtrType))
2117 .addUse(Ptr)
2118 .constrainAllUses(TII, TRI, RBI);
2119 return CastedPtr;
2120}
2121
2122bool SPIRVInstructionSelector::selectAtomicPtrValue(
2123 Register ResVReg, SPIRVTypeInst ResType, MachineIRBuilder &MIRBuilder,
2124 function_ref<Register(SPIRVTypeInst IntType)> EmitAtomic) const {
2125 // Pointer-typed atomics are lowered by bitcasting the Ptr operand to a
2126 // pointer to an integer of the same size as the pointer, so that the actual
2127 // atomic instruction operates on integers as required by the spec. Value
2128 // operands and results are converted with OpConvertPtrToU/OpConvertUToPtr.
2129 unsigned PtrSize = GR.getPointerSize();
2130 SPIRVTypeInst IntType = GR.getOrCreateSPIRVIntegerType(PtrSize, MIRBuilder);
2131
2132 Register IntResult = EmitAtomic(IntType);
2133 if (IntResult.isValid())
2134 MIRBuilder.buildInstr(SPIRV::OpConvertUToPtr)
2135 .addDef(ResVReg)
2136 .addUse(GR.getSPIRVTypeID(ResType))
2137 .addUse(IntResult)
2138 .constrainAllUses(TII, TRI, RBI);
2139 return true;
2140}
2141
2142bool SPIRVInstructionSelector::selectAtomicLoad(Register ResVReg,
2143 SPIRVTypeInst ResType,
2144 MachineInstr &I) const {
2145 LLVMContext &Context = I.getMF()->getFunction().getContext();
2146
2147 unsigned OpOffset = isa<GIntrinsic>(I) ? 1 : 0;
2148 Register Ptr = I.getOperand(1 + OpOffset).getReg();
2149
2150 if (!ResType.isTypeIntOrFloat() && !ResType.isTypePtr())
2151 return diagnoseUnsupported(
2152 I, "Lowering to SPIR-V of atomic load is only "
2153 "allowed for integer, floating point or pointer types");
2154
2155 assert(I.getNumMemOperands());
2156 const MachineMemOperand &MemOp = **I.memoperands_begin();
2157 assert(MemOp.isAtomic());
2158
2159 uint32_t Scope = static_cast<uint32_t>(
2160 getMemScope(STI.getTargetTriple(), Context, MemOp.getSyncScopeID()));
2161 Register ScopeReg = buildI32Constant(Scope, I);
2162
2163 AtomicOrdering AO = MemOp.getSuccessOrdering();
2164 uint32_t StorageClass = static_cast<uint32_t>(getMemSemanticsForStorageClass(
2165 addressSpaceToStorageClass(MemOp.getAddrSpace(), STI)));
2166 uint32_t MemSem = static_cast<uint32_t>(getMemSemantics(AO));
2167 if (MemOp.isVolatile() && STI.getTargetTriple().isVulkanOS())
2168 MemSem |= static_cast<uint32_t>(SPIRV::MemorySemantics::Volatile);
2169 uint32_t Sem = getMemSemanticsWithStorageClass(STI.getTargetTriple(), MemSem,
2170 StorageClass);
2171 Register MemSemReg = buildI32Constant(Sem, I);
2172
2173 MachineIRBuilder MIRBuilder(I);
2174
2175 if (ResType.isTypePtr()) {
2176 if (!STI.isPhysicalSPIRV())
2177 return diagnoseUnsupported(
2178 I, "Lowering to SPIR-V of atomic load is only "
2179 "allowed for pointer types for physical addressing model");
2180 // If data to load is a pointer type we bitcast the Ptr parameter to pointer
2181 // to an integer type of the same size as the pointer size and then generate
2182 // OpAtomicLoad the return value of that OpAtomicLoad is an integer that is
2183 // converted back to a pointer type using OpConvertUToPtr.
2184 SPIRV::StorageClass::StorageClass SC =
2185 addressSpaceToStorageClass(MemOp.getAddrSpace(), STI);
2186 return selectAtomicPtrValue(
2187 ResVReg, ResType, MIRBuilder, [&](SPIRVTypeInst IntType) {
2188 Register CastedPtr = castPtrToPtrToInt(Ptr, SC, MIRBuilder);
2189 Register IntResult = createPtrSizedIntReg(MIRBuilder);
2190 MIRBuilder.buildInstr(SPIRV::OpAtomicLoad)
2191 .addDef(IntResult)
2192 .addUse(GR.getSPIRVTypeID(IntType))
2193 .addUse(CastedPtr)
2194 .addUse(ScopeReg)
2195 .addUse(MemSemReg)
2196 .constrainAllUses(TII, TRI, RBI);
2197 return IntResult;
2198 });
2199 }
2200 auto AtomicLoad = MIRBuilder.buildInstr(SPIRV::OpAtomicLoad)
2201 .addDef(ResVReg)
2202 .addUse(GR.getSPIRVTypeID(ResType))
2203 .addUse(Ptr)
2204 .addUse(ScopeReg)
2205 .addUse(MemSemReg);
2206 AtomicLoad.constrainAllUses(TII, TRI, RBI);
2207
2208 return true;
2209}
2210
2211bool SPIRVInstructionSelector::selectStore(MachineInstr &I) const {
2212 unsigned OpOffset = isa<GIntrinsic>(I) ? 1 : 0;
2213 Register StoreVal = I.getOperand(0 + OpOffset).getReg();
2214 Register Ptr = I.getOperand(1 + OpOffset).getReg();
2215
2216 auto *PtrDef = getVRegDef(*MRI, Ptr);
2217 auto *IntPtrDef = dyn_cast<GIntrinsic>(PtrDef);
2218 if (IntPtrDef &&
2219 (IntPtrDef->getIntrinsicID() == Intrinsic::spv_resource_getbasepointer ||
2220 IntPtrDef->getIntrinsicID() == Intrinsic::spv_resource_getpointer)) {
2221
2222 Register HandleReg = IntPtrDef->getOperand(2).getReg();
2223 Register NewHandleReg =
2224 MRI->createVirtualRegister(MRI->getRegClass(HandleReg));
2225 auto *HandleDef = cast<GIntrinsic>(getVRegDef(*MRI, HandleReg));
2226 SPIRVTypeInst HandleType = GR.getSPIRVTypeForVReg(HandleReg);
2227 if (!loadHandleBeforePosition(NewHandleReg, HandleType, *HandleDef, I)) {
2228 return false;
2229 }
2230
2231 Register IdxReg = IntPtrDef->getOperand(3).getReg();
2232 if (HandleType->getOpcode() == SPIRV::OpTypeImage) {
2233 SPIRVTypeInst SampledType =
2234 GR.getSPIRVTypeForVReg(HandleType->getOperand(1).getReg());
2235 SPIRVTypeInst StoreValCompType =
2236 GR.getScalarOrVectorComponentType(GR.getSPIRVTypeForVReg(StoreVal));
2237 if (StoreValCompType && StoreValCompType != SampledType) {
2238 // A wide element (e.g. int64_t2) is emulated with a narrower packed
2239 // image. This bitcasts the value to match the format.
2240 SPIRVTypeInst PackedType = widenTypeToVec4(SampledType, I);
2241 Register PackedReg =
2242 MRI->createVirtualRegister(GR.getRegClass(PackedType));
2243 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpBitcast))
2244 .addDef(PackedReg)
2245 .addUse(GR.getSPIRVTypeID(PackedType))
2246 .addUse(StoreVal)
2247 .constrainAllUses(TII, TRI, RBI);
2248 StoreVal = PackedReg;
2249 }
2250
2251 auto BMI = BuildMI(*I.getParent(), I, I.getDebugLoc(),
2252 TII.get(SPIRV::OpImageWrite))
2253 .addUse(NewHandleReg)
2254 .addUse(IdxReg)
2255 .addUse(StoreVal);
2256
2257 const llvm::Type *LLVMHandleType = GR.getTypeForSPIRVType(HandleType);
2258 if (sampledTypeIsSignedInteger(LLVMHandleType))
2259 BMI.addImm(0x1000); // SignExtend
2260
2261 BMI.constrainAllUses(TII, TRI, RBI);
2262 return true;
2263 }
2264 }
2265
2266 SPIRVTypeInst PointeeTy = GR.getPointeeType(GR.getSPIRVTypeForVReg(Ptr));
2267 SPIRVTypeInst StoreTy = GR.getSPIRVTypeForVReg(StoreVal);
2268 if (PointeeTy && PointeeTy->getOpcode() == SPIRV::OpTypeVectorIdEXT &&
2269 StoreTy->getOpcode() != SPIRV::OpTypeVectorIdEXT &&
2270 GR.getScalarOrVectorComponentCount(PointeeTy) == 1) {
2271 MachineInstr *StoreValDef = getVRegDef(*MRI, StoreVal);
2272 Register Reg = StoreValDef->getOperand(0).getReg();
2273 if (auto FC = GFConstant::getConstant(Reg, *MRI))
2274 StoreVal =
2275 GR.getOrCreateConstVector(FC->getScalarValue(), I, PointeeTy, TII);
2276 else if (auto IC = GIConstant::getConstant(Reg, *MRI))
2277 StoreVal =
2278 GR.getOrCreateConstVector(IC->getScalarValue(), I, PointeeTy, TII);
2279 else
2280 llvm_unreachable("Unexpected <1 x T> Store type!");
2281 }
2282
2283 if (I.getNumMemOperands()) {
2284 const MachineMemOperand *MemOp = *I.memoperands_begin();
2285 if (MemOp->isAtomic())
2286 return selectAtomicStore(I);
2287 }
2288
2289 // Stores into a read-only storage class produce invalid SPIR-V. Reject such
2290 // input with a diagnostic rather than silently emitting an OpStore that
2291 // validation rejects.
2292 SPIRV::StorageClass::StorageClass PtrSC = GR.getPointerStorageClass(Ptr);
2293 if (PtrSC == SPIRV::StorageClass::UniformConstant ||
2294 PtrSC == SPIRV::StorageClass::Input ||
2295 PtrSC == SPIRV::StorageClass::PushConstant)
2296 return diagnoseUnsupported(
2297 I, "store into a read-only SPIR-V storage class is not allowed");
2298
2299 MachineIRBuilder MIRBuilder(I);
2300 auto MIB = MIRBuilder.buildInstr(SPIRV::OpStore).addUse(Ptr).addUse(StoreVal);
2301 if (!I.getNumMemOperands()) {
2302 assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS ||
2303 I.getOpcode() ==
2304 TargetOpcode::G_INTRINSIC_CONVERGENT_W_SIDE_EFFECTS);
2305 addMemoryOperands(I.getOperand(2 + OpOffset).getImm(), MIB);
2306 } else {
2307 addMemoryOperands(*I.memoperands_begin(), MIB, MIRBuilder, GR);
2308 }
2309 MIB.constrainAllUses(TII, TRI, RBI);
2310 return true;
2311}
2312
2313bool SPIRVInstructionSelector::selectAtomicStore(MachineInstr &I) const {
2314 LLVMContext &Context = I.getMF()->getFunction().getContext();
2315
2316 unsigned OpOffset = isa<GIntrinsic>(I) ? 1 : 0;
2317 Register StoreVal = I.getOperand(0 + OpOffset).getReg();
2318 Register Ptr = I.getOperand(1 + OpOffset).getReg();
2319
2320 SPIRVTypeInst PtrType = GR.getSPIRVTypeForVReg(Ptr);
2321 SPIRVTypeInst PointeeType = GR.getPointeeType(PtrType);
2322 // For an untyped pointer the data type is the stored value's type.
2323 if (!PointeeType && PtrType &&
2324 PtrType->getOpcode() == SPIRV::OpTypeUntypedPointerKHR)
2325 PointeeType = GR.getSPIRVTypeForVReg(StoreVal);
2326 if (!PointeeType)
2327 return diagnoseUnsupported(I,
2328 "Lowering to SPIR-V of atomic store is only "
2329 "allowed for integer or floating point types");
2330
2331 assert(I.getNumMemOperands());
2332 const MachineMemOperand &MemOp = **I.memoperands_begin();
2333 assert(MemOp.isAtomic());
2334
2335 uint32_t Scope = static_cast<uint32_t>(
2336 getMemScope(STI.getTargetTriple(), Context, MemOp.getSyncScopeID()));
2337 Register ScopeReg = buildI32Constant(Scope, I);
2338
2339 AtomicOrdering AO = MemOp.getSuccessOrdering();
2340 uint32_t StorageClass = static_cast<uint32_t>(getMemSemanticsForStorageClass(
2341 addressSpaceToStorageClass(MemOp.getAddrSpace(), STI)));
2342 uint32_t MemSem = static_cast<uint32_t>(getMemSemantics(AO));
2343 if (MemOp.isVolatile() && STI.getTargetTriple().isVulkanOS())
2344 MemSem |= static_cast<uint32_t>(SPIRV::MemorySemantics::Volatile);
2345 uint32_t Sem = getMemSemanticsWithStorageClass(STI.getTargetTriple(), MemSem,
2346 StorageClass);
2347 Register MemSemReg = buildI32Constant(Sem, I);
2348 MachineIRBuilder MIRBuilder(I);
2349
2350 if (PointeeType.isTypePtr()) {
2351 if (!STI.isPhysicalSPIRV())
2352 return diagnoseUnsupported(
2353 I, "Lowering to SPIR-V of atomic store is only "
2354 "allowed for pointer types for physical addressing model");
2355 // If data to store is a pointer type we cast it to an integer type of the
2356 // same size as the pointer size using OpConvertPtrToU, bitcast Ptr
2357 // parameter to pointer to integer type and then generate OpAtomicStore
2358 // with casted values as required by spec.
2359 SPIRV::StorageClass::StorageClass SC =
2360 addressSpaceToStorageClass(MemOp.getAddrSpace(), STI);
2361 return selectAtomicPtrValue(
2362 Register(), SPIRVTypeInst(), MIRBuilder, [&](SPIRVTypeInst IntType) {
2363 Register ValueAsInt = convertPtrToInt(StoreVal, MIRBuilder);
2364 Register CastedPtr = castPtrToPtrToInt(Ptr, SC, MIRBuilder);
2365 MIRBuilder.buildInstr(SPIRV::OpAtomicStore)
2366 .addUse(CastedPtr)
2367 .addUse(ScopeReg)
2368 .addUse(MemSemReg)
2369 .addUse(ValueAsInt)
2370 .constrainAllUses(TII, TRI, RBI);
2371 // Stores produce no result, so no OpConvertUToPtr is needed.
2372 return Register();
2373 });
2374 }
2375
2376 if (!PointeeType.isTypeIntOrFloat())
2377 return diagnoseUnsupported(I,
2378 "Lowering to SPIR-V of atomic store is only "
2379 "allowed for integer or floating point types");
2380
2381 auto AtomicStore = MIRBuilder.buildInstr(SPIRV::OpAtomicStore)
2382 .addUse(Ptr)
2383 .addUse(ScopeReg)
2384 .addUse(MemSemReg)
2385 .addUse(StoreVal);
2386 AtomicStore.constrainAllUses(TII, TRI, RBI);
2387
2388 return true;
2389}
2390
2391bool SPIRVInstructionSelector::selectMaskedGather(Register ResVReg,
2392 SPIRVTypeInst ResType,
2393 MachineInstr &I) const {
2394 assert(I.getNumExplicitDefs() == 1 && "Expected single def for gather");
2395 // Operand indices:
2396 // 0: result (def)
2397 // 1: intrinsic ID
2398 // 2: vector of pointers
2399 // 3: alignment (i32 immediate)
2400 // 4: mask (vector of i1)
2401 // 5: passthru/fill value
2402 const Register PtrsReg = I.getOperand(2).getReg();
2403 const uint32_t Alignment = I.getOperand(3).getImm();
2404 const Register MaskReg = I.getOperand(4).getReg();
2405 const Register PassthruReg = I.getOperand(5).getReg();
2406 const Register AlignmentReg = buildI32Constant(Alignment, I);
2407
2408 MachineBasicBlock &BB = *I.getParent();
2409 auto MIB =
2410 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpMaskedGatherINTEL))
2411 .addDef(ResVReg)
2412 .addUse(GR.getSPIRVTypeID(ResType))
2413 .addUse(PtrsReg)
2414 .addUse(AlignmentReg)
2415 .addUse(MaskReg)
2416 .addUse(PassthruReg);
2417 MIB.constrainAllUses(TII, TRI, RBI);
2418 return true;
2419}
2420
2421bool SPIRVInstructionSelector::selectMaskedScatter(MachineInstr &I) const {
2422 assert(I.getNumExplicitDefs() == 0 && "Expected no defs for scatter");
2423 // Operand indices (no explicit defs):
2424 // 0: intrinsic ID
2425 // 1: value vector
2426 // 2: vector of pointers
2427 // 3: alignment (i32 immediate)
2428 // 4: mask (vector of i1)
2429 const Register ValuesReg = I.getOperand(1).getReg();
2430 const Register PtrsReg = I.getOperand(2).getReg();
2431 const uint32_t Alignment = I.getOperand(3).getImm();
2432 const Register MaskReg = I.getOperand(4).getReg();
2433 const Register AlignmentReg = buildI32Constant(Alignment, I);
2434 MachineBasicBlock &BB = *I.getParent();
2435
2436 auto MIB =
2437 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpMaskedScatterINTEL))
2438 .addUse(PtrsReg)
2439 .addUse(AlignmentReg)
2440 .addUse(MaskReg)
2441 .addUse(ValuesReg);
2442 MIB.constrainAllUses(TII, TRI, RBI);
2443 return true;
2444}
2445
2446bool SPIRVInstructionSelector::diagnoseUnsupported(const MachineInstr &I,
2447 const Twine &Msg) const {
2448 const Function &F = I.getMF()->getFunction();
2449 F.getContext().diagnose(
2450 DiagnosticInfoUnsupported(F, Msg, I.getDebugLoc(), DS_Error));
2451 return false;
2452}
2453
2454bool SPIRVInstructionSelector::selectStackSave(Register ResVReg,
2455 SPIRVTypeInst ResType,
2456 MachineInstr &I) const {
2457 if (!STI.canUseExtension(SPIRV::Extension::SPV_INTEL_variable_length_array))
2458 return diagnoseUnsupported(
2459 I, "llvm.stacksave intrinsic: this instruction requires the following "
2460 "SPIR-V extension: SPV_INTEL_variable_length_array");
2461 MachineBasicBlock &BB = *I.getParent();
2462 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSaveMemoryINTEL))
2463 .addDef(ResVReg)
2464 .addUse(GR.getSPIRVTypeID(ResType))
2465 .constrainAllUses(TII, TRI, RBI);
2466 return true;
2467}
2468
2469bool SPIRVInstructionSelector::selectStackRestore(MachineInstr &I) const {
2470 if (!STI.canUseExtension(SPIRV::Extension::SPV_INTEL_variable_length_array))
2471 return diagnoseUnsupported(
2472 I,
2473 "llvm.stackrestore intrinsic: this instruction requires the following "
2474 "SPIR-V extension: SPV_INTEL_variable_length_array");
2475 if (!I.getOperand(0).isReg())
2476 return false;
2477 MachineBasicBlock &BB = *I.getParent();
2478 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpRestoreMemoryINTEL))
2479 .addUse(I.getOperand(0).getReg())
2480 .constrainAllUses(TII, TRI, RBI);
2481 return true;
2482}
2483
2484Register
2485SPIRVInstructionSelector::getOrCreateMemSetGlobal(MachineInstr &I) const {
2486 MachineIRBuilder MIRBuilder(I);
2487 assert(I.getOperand(1).isReg() && I.getOperand(2).isReg());
2488
2489 // TODO: check if we have such GV, add init, use buildGlobalVariable.
2490 unsigned Num = getIConstVal(I.getOperand(2).getReg(), MRI);
2491 Function &CurFunction = GR.CurMF->getFunction();
2492 Type *LLVMArrTy =
2493 ArrayType::get(IntegerType::get(CurFunction.getContext(), 8), Num);
2494 GlobalVariable *GV = new GlobalVariable(*CurFunction.getParent(), LLVMArrTy,
2495 true, GlobalValue::InternalLinkage,
2496 Constant::getNullValue(LLVMArrTy));
2497
2498 Type *ValTy = Type::getInt8Ty(I.getMF()->getFunction().getContext());
2499 Type *ArrTy = ArrayType::get(ValTy, Num);
2500 SPIRVTypeInst VarTy = GR.getOrCreateSPIRVPointerType(
2501 ArrTy, MIRBuilder, SPIRV::StorageClass::UniformConstant);
2502
2503 SPIRVTypeInst SpvArrTy = GR.getOrCreateSPIRVType(
2504 ArrTy, MIRBuilder, SPIRV::AccessQualifier::None, false);
2505
2506 unsigned Val = getIConstVal(I.getOperand(1).getReg(), MRI);
2507 Register Const = GR.getOrCreateConstIntArray(Val, Num, I, SpvArrTy, TII);
2508
2509 Register VarReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
2510 // With SPV_KHR_untyped_pointers enabled, getOrCreateSPIRVPointerType returns
2511 // an untyped pointer type. An untyped pointer result type is only legal on
2512 // OpUntypedVariableKHR, not on OpVariable.
2513 // Pick the matching opcode/operands so the synthesized constant global is
2514 // valid SPIR-V.
2515 const bool IsUntyped = VarTy->getOpcode() == SPIRV::OpTypeUntypedPointerKHR;
2516 auto MIBVar = BuildMI(*I.getParent(), I, I.getDebugLoc(),
2517 TII.get(IsUntyped ? SPIRV::OpUntypedVariableKHR
2518 : SPIRV::OpVariable))
2519 .addDef(VarReg)
2520 .addUse(GR.getSPIRVTypeID(VarTy))
2521 .addImm(SPIRV::StorageClass::UniformConstant);
2522 if (IsUntyped)
2523 MIBVar.addUse(GR.getSPIRVTypeID(SpvArrTy)); // Data Type
2524 MIBVar.addUse(Const); // Initializer
2525 MIBVar.constrainAllUses(TII, TRI, RBI);
2526
2527 GR.add(GV, MIBVar);
2528 GR.addGlobalObject(GV, GR.CurMF, VarReg);
2529
2530 buildOpDecorate(VarReg, I, TII, SPIRV::Decoration::Constant, {});
2531 return VarReg;
2532}
2533
2534bool SPIRVInstructionSelector::selectCopyMemory(MachineInstr &I,
2535 Register SrcReg) const {
2536 MachineBasicBlock &BB = *I.getParent();
2537 Register DstReg = I.getOperand(0).getReg();
2538 SPIRVTypeInst DstTy = GR.getSPIRVTypeForVReg(DstReg);
2539 SPIRVTypeInst SrcTy = GR.getSPIRVTypeForVReg(SrcReg);
2540 if (GR.getPointeeType(DstTy) != GR.getPointeeType(SrcTy))
2541 return diagnoseUnsupported(
2542 I, "OpCopyMemory requires operands to have the same type");
2543 uint64_t CopySize = getIConstVal(I.getOperand(2).getReg(), MRI);
2544 SPIRVTypeInst PointeeTy = GR.getPointeeType(DstTy);
2545 const Type *LLVMPointeeTy = GR.getTypeForSPIRVType(PointeeTy);
2546 if (!LLVMPointeeTy)
2547 return diagnoseUnsupported(
2548 I, "Unable to determine pointee type size for OpCopyMemory");
2549 const DataLayout &DL = I.getMF()->getFunction().getDataLayout();
2550 if (CopySize != DL.getTypeStoreSize(const_cast<Type *>(LLVMPointeeTy)))
2551 return diagnoseUnsupported(
2552 I, "OpCopyMemory requires the size to match the pointee type size");
2553 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCopyMemory))
2554 .addUse(DstReg)
2555 .addUse(SrcReg);
2556 if (I.getNumMemOperands()) {
2557 MachineIRBuilder MIRBuilder(I);
2558 addMemoryOperands(*I.memoperands_begin(), MIB, MIRBuilder, GR);
2559 }
2560 MIB.constrainAllUses(TII, TRI, RBI);
2561 return true;
2562}
2563
2564bool SPIRVInstructionSelector::selectCopyMemorySized(MachineInstr &I,
2565 Register SrcReg) const {
2566 MachineBasicBlock &BB = *I.getParent();
2567 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCopyMemorySized))
2568 .addUse(I.getOperand(0).getReg())
2569 .addUse(SrcReg)
2570 .addUse(I.getOperand(2).getReg());
2571 if (I.getNumMemOperands()) {
2572 MachineIRBuilder MIRBuilder(I);
2573 addMemoryOperands(*I.memoperands_begin(), MIB, MIRBuilder, GR);
2574 }
2575 MIB.constrainAllUses(TII, TRI, RBI);
2576 return true;
2577}
2578
2579bool SPIRVInstructionSelector::selectMemOperation(Register ResVReg,
2580 MachineInstr &I) const {
2581 // Zero-sized memcpy/memmove/memset are no-ops.
2582 Register SizeReg = I.getOperand(2).getReg();
2583 if (MachineInstr *SizeDef = getDefInstrMaybeConstant(SizeReg, MRI);
2584 SizeDef && SizeDef->getOpcode() == TargetOpcode::G_CONSTANT &&
2585 getIConstVal(SizeReg, MRI) == 0)
2586 return true;
2587
2588 Register SrcReg = I.getOperand(1).getReg();
2589 if (I.getOpcode() == TargetOpcode::G_MEMSET ||
2590 I.getOpcode() == TargetOpcode::G_MEMSET_INLINE) {
2591 Register VarReg = getOrCreateMemSetGlobal(I);
2592 if (!VarReg.isValid())
2593 return false;
2594 Type *ValTy = Type::getInt8Ty(I.getMF()->getFunction().getContext());
2595 SPIRVTypeInst SourceTy = GR.getOrCreateSPIRVPointerType(
2596 ValTy, I, SPIRV::StorageClass::UniformConstant);
2597 SrcReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
2598 if (!selectOpWithSrcs(SrcReg, SourceTy, I, {VarReg}, SPIRV::OpBitcast))
2599 return false;
2600 }
2601 if (STI.isLogicalSPIRV()) {
2602 if (!selectCopyMemory(I, SrcReg))
2603 return false;
2604 } else {
2605 if (!selectCopyMemorySized(I, SrcReg))
2606 return false;
2607 }
2608 if (ResVReg.isValid() && ResVReg != I.getOperand(0).getReg())
2609 if (!BuildCOPY(ResVReg, I.getOperand(0).getReg(), I))
2610 return false;
2611 return true;
2612}
2613
2614bool SPIRVInstructionSelector::selectAtomicRMW(Register ResVReg,
2615 SPIRVTypeInst ResType,
2616 MachineInstr &I,
2617 unsigned NewOpcode,
2618 unsigned NegateOpcode) const {
2619 assert(I.hasOneMemOperand());
2620 const MachineMemOperand *MemOp = *I.memoperands_begin();
2621 uint32_t Scope = static_cast<uint32_t>(
2622 getMemScope(STI.getTargetTriple(), GR.CurMF->getFunction().getContext(),
2623 MemOp->getSyncScopeID()));
2624 Register ScopeReg = buildI32Constant(Scope, I);
2625
2626 Register Ptr = I.getOperand(1).getReg();
2627 uint32_t ScSem = static_cast<uint32_t>(
2628 getMemSemanticsForStorageClass(GR.getPointerStorageClass(Ptr)));
2629 AtomicOrdering AO = MemOp->getSuccessOrdering();
2630 uint32_t OrderSem = static_cast<uint32_t>(getMemSemantics(AO));
2631 Register MemSemReg = buildI32Constant(
2632 getMemSemanticsWithStorageClass(STI.getTargetTriple(), OrderSem, ScSem),
2633 I);
2634
2635 Register ValueReg = I.getOperand(2).getReg();
2636 if (NegateOpcode != 0) {
2637 // Translation with negative value operand is requested
2638 Register TmpReg = createVirtualRegister(ResType, &GR, MRI, MRI->getMF());
2639 if (!selectOpWithSrcs(TmpReg, ResType, I, {ValueReg}, NegateOpcode))
2640 return false;
2641 ValueReg = TmpReg;
2642 }
2643
2644 if (ResType.isTypePtr()) {
2645 if (NewOpcode != SPIRV::OpAtomicExchange)
2646 return diagnoseUnsupported(
2647 I, "Lowering to SPIR-V of this atomic operation is not "
2648 "allowed for pointer types");
2649 if (!STI.isPhysicalSPIRV())
2650 return diagnoseUnsupported(
2651 I, "Lowering to SPIR-V of atomic exchange is only "
2652 "allowed for pointer types for physical addressing model");
2653 // If the exchanged value is a pointer type we convert the value operand to
2654 // an integer type of the same size as the pointer size using
2655 // OpConvertPtrToU, bitcast the Ptr parameter to pointer to integer type and
2656 // then generate OpAtomicExchange on integers. The integer result is
2657 // converted back to a pointer type using OpConvertUToPtr, similar to atomic
2658 // load and store.
2659 MachineIRBuilder MIRBuilder(I);
2660 SPIRV::StorageClass::StorageClass SC = GR.getPointerStorageClass(Ptr);
2661 return selectAtomicPtrValue(
2662 ResVReg, ResType, MIRBuilder, [&](SPIRVTypeInst IntType) {
2663 Register ValueAsInt = convertPtrToInt(ValueReg, MIRBuilder);
2664 Register CastedPtr = castPtrToPtrToInt(Ptr, SC, MIRBuilder);
2665 Register ExchangeResReg = createPtrSizedIntReg(MIRBuilder);
2666 MIRBuilder.buildInstr(SPIRV::OpAtomicExchange)
2667 .addDef(ExchangeResReg)
2668 .addUse(GR.getSPIRVTypeID(IntType))
2669 .addUse(CastedPtr)
2670 .addUse(ScopeReg)
2671 .addUse(MemSemReg)
2672 .addUse(ValueAsInt)
2673 .constrainAllUses(TII, TRI, RBI);
2674 return ExchangeResReg;
2675 });
2676 }
2677
2678 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(NewOpcode))
2679 .addDef(ResVReg)
2680 .addUse(GR.getSPIRVTypeID(ResType))
2681 .addUse(Ptr)
2682 .addUse(ScopeReg)
2683 .addUse(MemSemReg)
2684 .addUse(ValueReg)
2685 .constrainAllUses(TII, TRI, RBI);
2686 return true;
2687}
2688
2689bool SPIRVInstructionSelector::selectUnmergeValues(MachineInstr &I) const {
2690 unsigned ArgI = I.getNumOperands() - 1;
2691 Register SrcReg =
2692 I.getOperand(ArgI).isReg() ? I.getOperand(ArgI).getReg() : Register(0);
2693 SPIRVTypeInst SrcType =
2694 SrcReg.isValid() ? GR.getSPIRVTypeForVReg(SrcReg) : nullptr;
2695 if (!SrcType || !isVectorType(SrcType))
2696 report_fatal_error(
2697 "cannot select G_UNMERGE_VALUES with a non-vector argument");
2698
2699 SPIRVTypeInst ScalarType = GR.getScalarOrVectorComponentType(SrcType);
2700 MachineBasicBlock &BB = *I.getParent();
2701 unsigned CurrentIndex = 0;
2702 for (unsigned i = 0; i < I.getNumDefs(); ++i) {
2703 Register ResVReg = I.getOperand(i).getReg();
2704 SPIRVTypeInst ResType = GR.getSPIRVTypeForVReg(ResVReg);
2705 if (!ResType) {
2706 LLT ResLLT = MRI->getType(ResVReg);
2707 assert(ResLLT.isValid());
2708 if (ResLLT.isVector()) {
2709 ResType = GR.getOrCreateSPIRVVectorType(
2710 ScalarType, ResLLT.getNumElements(), I, TII);
2711 } else {
2712 ResType = ScalarType;
2713 }
2714 MRI->setRegClass(ResVReg, GR.getRegClass(ResType));
2715 GR.assignSPIRVTypeToVReg(ResType, ResVReg, *GR.CurMF);
2716 }
2717
2718 if (isVectorType(ResType)) {
2719 Register UndefReg = GR.getOrCreateUndef(I, SrcType, TII);
2720 auto MIB =
2721 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorShuffle))
2722 .addDef(ResVReg)
2723 .addUse(GR.getSPIRVTypeID(ResType))
2724 .addUse(SrcReg)
2725 .addUse(UndefReg);
2726 unsigned NumElements = GR.getScalarOrVectorComponentCount(ResType);
2727 for (unsigned j = 0; j < NumElements; ++j) {
2728 MIB.addImm(CurrentIndex + j);
2729 }
2730 CurrentIndex += NumElements;
2731 MIB.constrainAllUses(TII, TRI, RBI);
2732 } else {
2733 auto MIB =
2734 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
2735 .addDef(ResVReg)
2736 .addUse(GR.getSPIRVTypeID(ResType))
2737 .addUse(SrcReg)
2738 .addImm(CurrentIndex);
2739 CurrentIndex++;
2740 MIB.constrainAllUses(TII, TRI, RBI);
2741 }
2742 }
2743 return true;
2744}
2745
2746bool SPIRVInstructionSelector::selectFence(MachineInstr &I) const {
2747 AtomicOrdering AO = AtomicOrdering(I.getOperand(0).getImm());
2748 uint32_t ScSem = STI.isShader()
2749 ? SPIRV::MemorySemantics::UniformMemory |
2750 SPIRV::MemorySemantics::WorkgroupMemory |
2751 SPIRV::MemorySemantics::ImageMemory
2752 : SPIRV::MemorySemantics::WorkgroupMemory |
2753 SPIRV::MemorySemantics::CrossWorkgroupMemory |
2754 SPIRV::MemorySemantics::ImageMemory;
2755 uint32_t MemSem = getMemSemanticsWithStorageClass(
2756 STI.getTargetTriple(), static_cast<uint32_t>(getMemSemantics(AO)), ScSem);
2757 Register MemSemReg = buildI32ConstantInEntryBlock(MemSem, I);
2758 SyncScope::ID Ord = SyncScope::ID(I.getOperand(1).getImm());
2759 uint32_t Scope = static_cast<uint32_t>(getMemScope(
2760 STI.getTargetTriple(), GR.CurMF->getFunction().getContext(), Ord));
2761 Register ScopeReg = buildI32ConstantInEntryBlock(Scope, I);
2762 MachineBasicBlock &BB = *I.getParent();
2763 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpMemoryBarrier))
2764 .addUse(ScopeReg)
2765 .addUse(MemSemReg)
2766 .constrainAllUses(TII, TRI, RBI);
2767 return true;
2768}
2769
2770bool SPIRVInstructionSelector::selectOverflowArith(Register ResVReg,
2771 SPIRVTypeInst ResType,
2772 MachineInstr &I,
2773 unsigned Opcode) const {
2774 Type *ResTy = nullptr;
2775 StringRef ResName;
2776 if (!GR.findValueAttrs(&I, ResTy, ResName))
2777 return diagnoseUnsupported(
2778 I,
2779 "Not enough info to select the arithmetic with overflow instruction");
2780 if (!ResTy || !ResTy->isStructTy())
2781 return diagnoseUnsupported(I,
2782 "Expect struct type result for the arithmetic "
2783 "with overflow instruction");
2784 // "Result Type must be from OpTypeStruct. The struct must have two members,
2785 // and the two members must be the same type."
2786 Type *ResElemTy = cast<StructType>(ResTy)->getElementType(0);
2787 ResTy = StructType::get(ResElemTy, ResElemTy);
2788 // Build SPIR-V types and constant(s) if needed.
2789 MachineIRBuilder MIRBuilder(I);
2790 SPIRVTypeInst StructType = GR.getOrCreateSPIRVType(
2791 ResTy, MIRBuilder, SPIRV::AccessQualifier::ReadWrite, false);
2792 assert(I.getNumDefs() > 1 && "Not enought operands");
2793 SPIRVTypeInst BoolType = GR.getOrCreateSPIRVBoolType(I, TII);
2794 unsigned N = GR.getScalarOrVectorComponentCount(ResType);
2795 if (N > 1 || (isVectorType(ResType) &&
2796 STI.canUseExtension(SPIRV::Extension::SPV_EXT_long_vector)))
2797 BoolType = GR.getOrCreateSPIRVVectorType(BoolType, N, I, TII);
2798 Register BoolTypeReg = GR.getSPIRVTypeID(BoolType);
2799 Register ZeroReg = buildZerosVal(ResType, I);
2800 // A new virtual register to store the result struct.
2801 Register StructVReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
2802 MRI->setRegClass(StructVReg, &SPIRV::IDRegClass);
2803 // Build the result name if needed.
2804 if (ResName.size() > 0)
2805 buildOpName(StructVReg, ResName, MIRBuilder);
2806 // Build the arithmetic with overflow instruction.
2807 MachineBasicBlock &BB = *I.getParent();
2808 auto MIB =
2809 BuildMI(BB, MIRBuilder.getInsertPt(), I.getDebugLoc(), TII.get(Opcode))
2810 .addDef(StructVReg)
2811 .addUse(GR.getSPIRVTypeID(StructType));
2812 for (unsigned i = I.getNumDefs(); i < I.getNumOperands(); ++i)
2813 MIB.addUse(I.getOperand(i).getReg());
2814 MIB.constrainAllUses(TII, TRI, RBI);
2815 // Build instructions to extract fields of the instruction's result.
2816 // A new virtual register to store the higher part of the result struct.
2817 Register HigherVReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
2818 MRI->setRegClass(HigherVReg, &SPIRV::iIDRegClass);
2819 for (unsigned i = 0; i < I.getNumDefs(); ++i) {
2820 auto MIB =
2821 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
2822 .addDef(i == 1 ? HigherVReg : I.getOperand(i).getReg())
2823 .addUse(GR.getSPIRVTypeID(ResType))
2824 .addUse(StructVReg)
2825 .addImm(i);
2826 MIB.constrainAllUses(TII, TRI, RBI);
2827 }
2828 // Build boolean value from the higher part.
2829 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpINotEqual))
2830 .addDef(I.getOperand(1).getReg())
2831 .addUse(BoolTypeReg)
2832 .addUse(HigherVReg)
2833 .addUse(ZeroReg)
2834 .constrainAllUses(TII, TRI, RBI);
2835 return true;
2836}
2837
2838bool SPIRVInstructionSelector::selectAtomicCmpXchg(Register ResVReg,
2839 SPIRVTypeInst ResType,
2840 MachineInstr &I) const {
2841 assert(isa<GIntrinsic>(I) &&
2842 "selectAtomicCmpXchg only handles the spv_cmpxchg intrinsic");
2843 Register Ptr = I.getOperand(2).getReg();
2844 Register ScopeReg = I.getOperand(5).getReg();
2845 Register MemSemEqReg = I.getOperand(6).getReg();
2846 Register MemSemNeqReg = I.getOperand(7).getReg();
2847 Register Cmp = I.getOperand(3).getReg();
2848 Register Val = I.getOperand(4).getReg();
2849 SPIRVTypeInst SpvValTy = GR.getSPIRVTypeForVReg(Val);
2850 Register ACmpRes = createVirtualRegister(SpvValTy, &GR, MRI, *I.getMF());
2851 const DebugLoc &DL = I.getDebugLoc();
2852 BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpAtomicCompareExchange))
2853 .addDef(ACmpRes)
2854 .addUse(GR.getSPIRVTypeID(SpvValTy))
2855 .addUse(Ptr)
2856 .addUse(ScopeReg)
2857 .addUse(MemSemEqReg)
2858 .addUse(MemSemNeqReg)
2859 .addUse(Val)
2860 .addUse(Cmp)
2861 .constrainAllUses(TII, TRI, RBI);
2862 SPIRVTypeInst BoolTy = GR.getOrCreateSPIRVBoolType(I, TII);
2863 Register CmpSuccReg = createVirtualRegister(BoolTy, &GR, MRI, *I.getMF());
2864 BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpIEqual))
2865 .addDef(CmpSuccReg)
2866 .addUse(GR.getSPIRVTypeID(BoolTy))
2867 .addUse(ACmpRes)
2868 .addUse(Cmp)
2869 .constrainAllUses(TII, TRI, RBI);
2870 Register TmpReg = createVirtualRegister(ResType, &GR, MRI, *I.getMF());
2871 BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpCompositeInsert))
2872 .addDef(TmpReg)
2873 .addUse(GR.getSPIRVTypeID(ResType))
2874 .addUse(ACmpRes)
2875 .addUse(GR.getOrCreateUndef(I, ResType, TII))
2876 .addImm(0)
2877 .constrainAllUses(TII, TRI, RBI);
2878 BuildMI(*I.getParent(), I, DL, TII.get(SPIRV::OpCompositeInsert))
2879 .addDef(ResVReg)
2880 .addUse(GR.getSPIRVTypeID(ResType))
2881 .addUse(CmpSuccReg)
2882 .addUse(TmpReg)
2883 .addImm(1)
2884 .constrainAllUses(TII, TRI, RBI);
2885 return true;
2886}
2887
2888static bool isUSMStorageClass(SPIRV::StorageClass::StorageClass SC) {
2889 switch (SC) {
2890 case SPIRV::StorageClass::DeviceOnlyINTEL:
2891 case SPIRV::StorageClass::HostOnlyINTEL:
2892 return true;
2893 default:
2894 return false;
2895 }
2896}
2897
2898// Returns true ResVReg is referred only from global vars and OpName's.
2899static bool isASCastInGVar(MachineRegisterInfo *MRI, Register ResVReg) {
2900 bool IsGRef = false;
2901 bool IsAllowedRefs =
2902 llvm::all_of(MRI->use_instructions(ResVReg), [&IsGRef](auto const &It) {
2903 unsigned Opcode = It.getOpcode();
2904 if (Opcode == SPIRV::OpConstantComposite ||
2905 Opcode == SPIRV::OpSpecConstantComposite ||
2906 Opcode == SPIRV::OpVariable ||
2907 Opcode == SPIRV::OpUntypedVariableKHR ||
2908 isSpvIntrinsic(It, Intrinsic::spv_init_global))
2909 return IsGRef = true;
2910 return Opcode == SPIRV::OpName;
2911 });
2912 return IsAllowedRefs && IsGRef;
2913}
2914
2915Register SPIRVInstructionSelector::getUcharPtrTypeReg(
2916 MachineInstr &I, SPIRV::StorageClass::StorageClass SC) const {
2917 return GR.getSPIRVTypeID(GR.getOrCreateSPIRVPointerType(
2918 Type::getInt8Ty(I.getMF()->getFunction().getContext()), I, SC));
2919}
2920
2921MachineInstrBuilder
2922SPIRVInstructionSelector::buildSpecConstantOp(MachineInstr &I, Register Dest,
2923 Register Src, Register DestType,
2924 uint32_t Opcode) const {
2925 return BuildMI(*I.getParent(), I, I.getDebugLoc(),
2926 TII.get(SPIRV::OpSpecConstantOp))
2927 .addDef(Dest)
2928 .addUse(DestType)
2929 .addImm(Opcode)
2930 .addUse(Src);
2931}
2932
2933MachineInstrBuilder
2934SPIRVInstructionSelector::buildConstGenericPtr(MachineInstr &I, Register SrcPtr,
2935 SPIRVTypeInst SrcPtrTy) const {
2936 SPIRVTypeInst GenericPtrTy =
2937 GR.changePointerStorageClass(SrcPtrTy, SPIRV::StorageClass::Generic, I);
2938 Register Tmp = MRI->createVirtualRegister(&SPIRV::pIDRegClass);
2939 MRI->setType(Tmp, LLT::pointer(storageClassToAddressSpace(
2940 SPIRV::StorageClass::Generic),
2941 GR.getPointerSize()));
2942 MachineFunction *MF = I.getParent()->getParent();
2943 GR.assignSPIRVTypeToVReg(GenericPtrTy, Tmp, *MF);
2944 MachineInstrBuilder MIB = buildSpecConstantOp(
2945 I, Tmp, SrcPtr, GR.getSPIRVTypeID(GenericPtrTy),
2946 static_cast<uint32_t>(SPIRV::Opcode::PtrCastToGeneric));
2947 GR.add(MIB.getInstr(), MIB);
2948 return MIB;
2949}
2950
2951// In SPIR-V address space casting can only happen to and from the Generic
2952// storage class. We can also only cast Workgroup, CrossWorkgroup, or Function
2953// pointers to and from Generic pointers. As such, we can convert e.g. from
2954// Workgroup to Function by going via a Generic pointer as an intermediary. All
2955// other combinations can only be done by a bitcast, and are probably not safe.
2956bool SPIRVInstructionSelector::selectAddrSpaceCast(Register ResVReg,
2957 SPIRVTypeInst ResType,
2958 MachineInstr &I) const {
2959 MachineBasicBlock &BB = *I.getParent();
2960 const DebugLoc &DL = I.getDebugLoc();
2961
2962 Register SrcPtr = I.getOperand(1).getReg();
2963 SPIRVTypeInst SrcPtrTy = GR.getSPIRVTypeForVReg(SrcPtr);
2964
2965 // don't generate a cast for a null that may be represented by OpTypeInt
2966 if (!SrcPtrTy || !SrcPtrTy.isPointer() || !ResType || !ResType.isPointer())
2967 return BuildCOPY(ResVReg, SrcPtr, I);
2968
2969 SPIRV::StorageClass::StorageClass SrcSC = GR.getPointerStorageClass(SrcPtrTy);
2970 SPIRV::StorageClass::StorageClass DstSC = GR.getPointerStorageClass(ResType);
2971
2972 if (isASCastInGVar(MRI, ResVReg)) {
2973 // AddrSpaceCast uses within OpVariable and OpConstantComposite instructions
2974 // are expressed by OpSpecConstantOp with an Opcode.
2975 // TODO: maybe insert a check whether the Kernel capability was declared and
2976 // so PtrCastToGeneric/GenericCastToPtr are available.
2977 unsigned SpecOpcode = [&]() -> unsigned {
2978 if (SrcSC == SPIRV::StorageClass::CodeSectionINTEL)
2979 return static_cast<uint32_t>(SPIRV::Opcode::Bitcast);
2980 if (DstSC == SPIRV::StorageClass::Generic && isGenericCastablePtr(SrcSC))
2981 return static_cast<uint32_t>(SPIRV::Opcode::PtrCastToGeneric);
2982 if (SrcSC == SPIRV::StorageClass::Generic && isGenericCastablePtr(DstSC))
2983 return static_cast<uint32_t>(SPIRV::Opcode::GenericCastToPtr);
2984 return 0u;
2985 }();
2986 // TODO: OpConstantComposite expects i8*, so we are forced to forget a
2987 // correct value of ResType and use general i8* instead. Maybe this should
2988 // be addressed in the emit-intrinsic step to infer a correct
2989 // OpConstantComposite type.
2990 if (SpecOpcode) {
2991 buildSpecConstantOp(I, ResVReg, SrcPtr, getUcharPtrTypeReg(I, DstSC),
2992 SpecOpcode)
2993 .constrainAllUses(TII, TRI, RBI);
2994 } else if (isGenericCastablePtr(SrcSC) && isGenericCastablePtr(DstSC)) {
2995 MachineInstrBuilder MIB = buildConstGenericPtr(I, SrcPtr, SrcPtrTy);
2996 MIB.constrainAllUses(TII, TRI, RBI);
2997 buildSpecConstantOp(
2998 I, ResVReg, MIB->getOperand(0).getReg(), getUcharPtrTypeReg(I, DstSC),
2999 static_cast<uint32_t>(SPIRV::Opcode::GenericCastToPtr))
3000 .constrainAllUses(TII, TRI, RBI);
3001 }
3002 return true;
3003 }
3004
3005 // don't generate a cast between identical storage classes
3006 if (SrcSC == DstSC)
3007 return BuildCOPY(ResVReg, SrcPtr, I);
3008
3009 if ((SrcSC == SPIRV::StorageClass::Function &&
3010 DstSC == SPIRV::StorageClass::Private) ||
3011 (DstSC == SPIRV::StorageClass::Function &&
3012 SrcSC == SPIRV::StorageClass::Private))
3013 return BuildCOPY(ResVReg, SrcPtr, I);
3014
3015 // Casting from an eligible pointer to Generic.
3016 if (DstSC == SPIRV::StorageClass::Generic && isGenericCastablePtr(SrcSC))
3017 return selectUnOp(ResVReg, ResType, I, SPIRV::OpPtrCastToGeneric);
3018 // Casting from Generic to an eligible pointer.
3019 if (SrcSC == SPIRV::StorageClass::Generic && isGenericCastablePtr(DstSC))
3020 return selectUnOp(ResVReg, ResType, I, SPIRV::OpGenericCastToPtr);
3021 // Casting between 2 eligible pointers using Generic as an intermediary.
3022 if (isGenericCastablePtr(SrcSC) && isGenericCastablePtr(DstSC)) {
3023 SPIRVTypeInst GenericPtrTy =
3024 GR.changePointerStorageClass(SrcPtrTy, SPIRV::StorageClass::Generic, I);
3025 Register Tmp = createVirtualRegister(GenericPtrTy, &GR, MRI, MRI->getMF());
3026 BuildMI(BB, I, DL, TII.get(SPIRV::OpPtrCastToGeneric))
3027 .addDef(Tmp)
3028 .addUse(GR.getSPIRVTypeID(GenericPtrTy))
3029 .addUse(SrcPtr)
3030 .constrainAllUses(TII, TRI, RBI);
3031 BuildMI(BB, I, DL, TII.get(SPIRV::OpGenericCastToPtr))
3032 .addDef(ResVReg)
3033 .addUse(GR.getSPIRVTypeID(ResType))
3034 .addUse(Tmp)
3035 .constrainAllUses(TII, TRI, RBI);
3036 return true;
3037 }
3038
3039 // Check if instructions from the SPV_INTEL_usm_storage_classes extension may
3040 // be applied
3041 if (isUSMStorageClass(SrcSC) && DstSC == SPIRV::StorageClass::CrossWorkgroup)
3042 return selectUnOp(ResVReg, ResType, I,
3043 SPIRV::OpPtrCastToCrossWorkgroupINTEL);
3044 if (SrcSC == SPIRV::StorageClass::CrossWorkgroup && isUSMStorageClass(DstSC))
3045 return selectUnOp(ResVReg, ResType, I,
3046 SPIRV::OpCrossWorkgroupCastToPtrINTEL);
3047 if (isUSMStorageClass(SrcSC) && DstSC == SPIRV::StorageClass::Generic)
3048 return selectUnOp(ResVReg, ResType, I, SPIRV::OpPtrCastToGeneric);
3049 if (SrcSC == SPIRV::StorageClass::Generic && isUSMStorageClass(DstSC))
3050 return selectUnOp(ResVReg, ResType, I, SPIRV::OpGenericCastToPtr);
3051
3052 // Bitcast for pointers requires that the address spaces must match
3053 return false;
3054}
3055
3056// G_PTRMASK - Apply a bitmask to a pointer value.
3057// Result = Ptr & Mask
3058// We need to convert the pointer to an integer, perform the AND operation,
3059// and convert back to a pointer.
3060bool SPIRVInstructionSelector::selectPtrMask(Register ResVReg,
3061 SPIRVTypeInst ResType,
3062 MachineInstr &I) const {
3063 if (STI.isLogicalSPIRV())
3064 return diagnoseUnsupported(
3065 I, "G_PTRMASK is not supported with logical SPIR-V");
3066 MachineBasicBlock &BB = *I.getParent();
3067 MachineFunction &MF = *BB.getParent();
3068 const DebugLoc &DL = I.getDebugLoc();
3069
3070 Register PtrReg = I.getOperand(1).getReg();
3071 Register MaskReg = I.getOperand(2).getReg();
3072
3073 SPIRVTypeInst MaskType = GR.getSPIRVTypeForVReg(MaskReg);
3074
3075 // Convert pointer to integer.
3076 Register PtrAsInt = MRI->createVirtualRegister(GR.getRegClass(MaskType));
3077 GR.assignSPIRVTypeToVReg(MaskType, PtrAsInt, MF);
3078
3079 BuildMI(BB, I, DL, TII.get(SPIRV::OpConvertPtrToU))
3080 .addDef(PtrAsInt)
3081 .addUse(GR.getSPIRVTypeID(MaskType))
3082 .addUse(PtrReg)
3083 .constrainAllUses(TII, TRI, RBI);
3084
3085 // Perform bitwise AND.
3086 Register MaskedInt = MRI->createVirtualRegister(GR.getRegClass(MaskType));
3087 GR.assignSPIRVTypeToVReg(MaskType, MaskedInt, MF);
3088
3089 unsigned AndOpcode = GR.getScalarOrVectorComponentCount(MaskType) > 1
3090 ? SPIRV::OpBitwiseAndV
3091 : SPIRV::OpBitwiseAndS;
3092
3093 BuildMI(BB, I, DL, TII.get(AndOpcode))
3094 .addDef(MaskedInt)
3095 .addUse(GR.getSPIRVTypeID(MaskType))
3096 .addUse(PtrAsInt)
3097 .addUse(MaskReg)
3098 .constrainAllUses(TII, TRI, RBI);
3099
3100 // Convert integer back to pointer.
3101 BuildMI(BB, I, DL, TII.get(SPIRV::OpConvertUToPtr))
3102 .addDef(ResVReg)
3103 .addUse(GR.getSPIRVTypeID(ResType))
3104 .addUse(MaskedInt)
3105 .constrainAllUses(TII, TRI, RBI);
3106
3107 return true;
3108}
3109
3110static unsigned getFCmpOpcode(unsigned PredNum) {
3111 auto Pred = static_cast<CmpInst::Predicate>(PredNum);
3112 switch (Pred) {
3113 case CmpInst::FCMP_OEQ:
3114 return SPIRV::OpFOrdEqual;
3115 case CmpInst::FCMP_OGE:
3116 return SPIRV::OpFOrdGreaterThanEqual;
3117 case CmpInst::FCMP_OGT:
3118 return SPIRV::OpFOrdGreaterThan;
3119 case CmpInst::FCMP_OLE:
3120 return SPIRV::OpFOrdLessThanEqual;
3121 case CmpInst::FCMP_OLT:
3122 return SPIRV::OpFOrdLessThan;
3123 case CmpInst::FCMP_ONE:
3124 return SPIRV::OpFOrdNotEqual;
3125 case CmpInst::FCMP_ORD:
3126 return SPIRV::OpOrdered;
3127 case CmpInst::FCMP_UEQ:
3128 return SPIRV::OpFUnordEqual;
3129 case CmpInst::FCMP_UGE:
3130 return SPIRV::OpFUnordGreaterThanEqual;
3131 case CmpInst::FCMP_UGT:
3132 return SPIRV::OpFUnordGreaterThan;
3133 case CmpInst::FCMP_ULE:
3134 return SPIRV::OpFUnordLessThanEqual;
3135 case CmpInst::FCMP_ULT:
3136 return SPIRV::OpFUnordLessThan;
3137 case CmpInst::FCMP_UNE:
3138 return SPIRV::OpFUnordNotEqual;
3139 case CmpInst::FCMP_UNO:
3140 return SPIRV::OpUnordered;
3141 default:
3142 llvm_unreachable("Unknown predicate type for FCmp");
3143 }
3144}
3145
3146static unsigned getICmpOpcode(unsigned PredNum) {
3147 auto Pred = static_cast<CmpInst::Predicate>(PredNum);
3148 switch (Pred) {
3149 case CmpInst::ICMP_EQ:
3150 return SPIRV::OpIEqual;
3151 case CmpInst::ICMP_NE:
3152 return SPIRV::OpINotEqual;
3153 case CmpInst::ICMP_SGE:
3154 return SPIRV::OpSGreaterThanEqual;
3155 case CmpInst::ICMP_SGT:
3156 return SPIRV::OpSGreaterThan;
3157 case CmpInst::ICMP_SLE:
3158 return SPIRV::OpSLessThanEqual;
3159 case CmpInst::ICMP_SLT:
3160 return SPIRV::OpSLessThan;
3161 case CmpInst::ICMP_UGE:
3162 return SPIRV::OpUGreaterThanEqual;
3163 case CmpInst::ICMP_UGT:
3164 return SPIRV::OpUGreaterThan;
3165 case CmpInst::ICMP_ULE:
3166 return SPIRV::OpULessThanEqual;
3167 case CmpInst::ICMP_ULT:
3168 return SPIRV::OpULessThan;
3169 default:
3170 llvm_unreachable("Unknown predicate type for ICmp");
3171 }
3172}
3173
3174static unsigned getPtrCmpOpcode(unsigned Pred) {
3175 switch (static_cast<CmpInst::Predicate>(Pred)) {
3176 case CmpInst::ICMP_EQ:
3177 return SPIRV::OpPtrEqual;
3178 case CmpInst::ICMP_NE:
3179 return SPIRV::OpPtrNotEqual;
3180 default:
3181 llvm_unreachable("Unknown predicate type for pointer comparison");
3182 }
3183}
3184
3185// Return the logical operation, or abort if none exists.
3186static unsigned getBoolCmpOpcode(unsigned PredNum) {
3187 auto Pred = static_cast<CmpInst::Predicate>(PredNum);
3188 switch (Pred) {
3189 case CmpInst::ICMP_EQ:
3190 return SPIRV::OpLogicalEqual;
3191 case CmpInst::ICMP_NE:
3192 return SPIRV::OpLogicalNotEqual;
3193 default:
3194 llvm_unreachable("Unknown predicate type for Bool comparison");
3195 }
3196}
3197
3198static APFloat getZeroFP(const Type *LLVMFloatTy) {
3199 if (!LLVMFloatTy)
3200 return APFloat::getZero(APFloat::IEEEsingle());
3201 switch (LLVMFloatTy->getScalarType()->getTypeID()) {
3202 case Type::HalfTyID:
3203 return APFloat::getZero(APFloat::IEEEhalf());
3204 case Type::BFloatTyID:
3205 return APFloat::getZero(APFloat::BFloat());
3206 default:
3207 case Type::FloatTyID:
3208 return APFloat::getZero(APFloat::IEEEsingle());
3209 case Type::DoubleTyID:
3210 return APFloat::getZero(APFloat::IEEEdouble());
3211 }
3212}
3213
3214static APFloat getOneFP(const Type *LLVMFloatTy) {
3215 if (!LLVMFloatTy)
3216 return APFloat::getOne(APFloat::IEEEsingle());
3217 switch (LLVMFloatTy->getScalarType()->getTypeID()) {
3218 case Type::HalfTyID:
3219 return APFloat::getOne(APFloat::IEEEhalf());
3220 case Type::BFloatTyID:
3221 return APFloat::getOne(APFloat::BFloat());
3222 default:
3223 case Type::FloatTyID:
3224 return APFloat::getOne(APFloat::IEEEsingle());
3225 case Type::DoubleTyID:
3226 return APFloat::getOne(APFloat::IEEEdouble());
3227 }
3228}
3229
3230bool SPIRVInstructionSelector::selectAnyOrAll(Register ResVReg,
3231 SPIRVTypeInst ResType,
3232 MachineInstr &I,
3233 unsigned OpAnyOrAll) const {
3234 assert(I.getNumOperands() == 3);
3235 assert(I.getOperand(2).isReg());
3236 MachineBasicBlock &BB = *I.getParent();
3237 Register InputRegister = I.getOperand(2).getReg();
3238 SPIRVTypeInst InputType = GR.getSPIRVTypeForVReg(InputRegister);
3239
3240 assert(InputType && "VReg has no type assigned");
3241
3242 bool IsBoolTy = GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeBool);
3243 if (IsBoolTy && !isVectorType(InputType)) {
3244 assert(ResVReg == I.getOperand(0).getReg());
3245 return BuildCOPY(ResVReg, InputRegister, I);
3246 }
3247
3248 bool IsFloatTy = GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeFloat);
3249 unsigned SpirvNotEqualId =
3250 IsFloatTy ? SPIRV::OpFOrdNotEqual : SPIRV::OpINotEqual;
3251 SPIRVTypeInst SpvBoolScalarTy = GR.getOrCreateSPIRVBoolType(I, TII);
3252 SPIRVTypeInst SpvBoolTy = SpvBoolScalarTy;
3253 Register NotEqualReg = ResVReg;
3254
3255 if (isVectorType(InputType)) {
3256 NotEqualReg =
3257 IsBoolTy ? InputRegister
3258 : createVirtualRegister(SpvBoolTy, &GR, MRI, MRI->getMF());
3259 const unsigned NumElts = GR.getScalarOrVectorComponentCount(InputType);
3260 SpvBoolTy = GR.getOrCreateSPIRVVectorType(SpvBoolTy, NumElts, I, TII);
3261 }
3262
3263 if (!IsBoolTy) {
3264 Register ConstZeroReg =
3265 IsFloatTy ? buildZerosValF(InputType, I) : buildZerosVal(InputType, I);
3266
3267 BuildMI(BB, I, I.getDebugLoc(), TII.get(SpirvNotEqualId))
3268 .addDef(NotEqualReg)
3269 .addUse(GR.getSPIRVTypeID(SpvBoolTy))
3270 .addUse(InputRegister)
3271 .addUse(ConstZeroReg)
3272 .constrainAllUses(TII, TRI, RBI);
3273 }
3274
3275 if (isVectorType(InputType))
3276 BuildMI(BB, I, I.getDebugLoc(), TII.get(OpAnyOrAll))
3277 .addDef(ResVReg)
3278 .addUse(GR.getSPIRVTypeID(SpvBoolScalarTy))
3279 .addUse(NotEqualReg)
3280 .constrainAllUses(TII, TRI, RBI);
3281 return true;
3282}
3283
3284bool SPIRVInstructionSelector::selectAll(Register ResVReg,
3285 SPIRVTypeInst ResType,
3286 MachineInstr &I) const {
3287 return selectAnyOrAll(ResVReg, ResType, I, SPIRV::OpAll);
3288}
3289
3290bool SPIRVInstructionSelector::selectAny(Register ResVReg,
3291 SPIRVTypeInst ResType,
3292 MachineInstr &I) const {
3293 return selectAnyOrAll(ResVReg, ResType, I, SPIRV::OpAny);
3294}
3295
3296// Select the OpDot instruction for the given float dot
3297bool SPIRVInstructionSelector::selectFloatDot(Register ResVReg,
3298 SPIRVTypeInst ResType,
3299 MachineInstr &I) const {
3300 assert(I.getNumOperands() == 4);
3301 assert(I.getOperand(2).isReg());
3302 assert(I.getOperand(3).isReg());
3303
3304 [[maybe_unused]] SPIRVTypeInst VecType =
3305 GR.getSPIRVTypeForVReg(I.getOperand(2).getReg());
3306
3307 assert(((VecType->getOpcode() == SPIRV::OpTypeVector &&
3308 GR.getScalarOrVectorComponentCount(VecType) > 1) ||
3309 VecType->getOpcode() == SPIRV::OpTypeVectorIdEXT) &&
3310 "dot product requires either a vector of at least 2 components or"
3311 " the SPV_EXT_long vector extension.");
3312
3313 [[maybe_unused]] SPIRVTypeInst EltType =
3314 GR.getScalarOrVectorComponentType(VecType);
3315
3316 assert(EltType->getOpcode() == SPIRV::OpTypeFloat);
3317
3318 MachineBasicBlock &BB = *I.getParent();
3319 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpDot))
3320 .addDef(ResVReg)
3321 .addUse(GR.getSPIRVTypeID(ResType))
3322 .addUse(I.getOperand(2).getReg())
3323 .addUse(I.getOperand(3).getReg())
3324 .constrainAllUses(TII, TRI, RBI);
3325 return true;
3326}
3327
3328bool SPIRVInstructionSelector::selectIntegerDot(Register ResVReg,
3329 SPIRVTypeInst ResType,
3330 MachineInstr &I,
3331 bool Signed) const {
3332 assert(I.getNumOperands() == 4);
3333 assert(I.getOperand(2).isReg());
3334 assert(I.getOperand(3).isReg());
3335 MachineBasicBlock &BB = *I.getParent();
3336
3337 auto DotOp = Signed ? SPIRV::OpSDot : SPIRV::OpUDot;
3338 BuildMI(BB, I, I.getDebugLoc(), TII.get(DotOp))
3339 .addDef(ResVReg)
3340 .addUse(GR.getSPIRVTypeID(ResType))
3341 .addUse(I.getOperand(2).getReg())
3342 .addUse(I.getOperand(3).getReg())
3343 .constrainAllUses(TII, TRI, RBI);
3344 return true;
3345}
3346
3347// Since pre-1.6 SPIRV has no integer dot implementation,
3348// expand by piecewise multiplying and adding the results
3349bool SPIRVInstructionSelector::selectIntegerDotExpansion(
3350 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
3351 assert(I.getNumOperands() == 4);
3352 assert(I.getOperand(2).isReg());
3353 assert(I.getOperand(3).isReg());
3354 MachineBasicBlock &BB = *I.getParent();
3355
3356 // Multiply the vectors, then sum the results
3357 Register Vec0 = I.getOperand(2).getReg();
3358 Register Vec1 = I.getOperand(3).getReg();
3359 Register TmpVec = MRI->createVirtualRegister(GR.getRegClass(ResType));
3360 SPIRVTypeInst VecType = GR.getSPIRVTypeForVReg(Vec0);
3361
3362 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIMulV))
3363 .addDef(TmpVec)
3364 .addUse(GR.getSPIRVTypeID(VecType))
3365 .addUse(Vec0)
3366 .addUse(Vec1)
3367 .constrainAllUses(TII, TRI, RBI);
3368
3369 assert(((VecType->getOpcode() == SPIRV::OpTypeVector &&
3370 GR.getScalarOrVectorComponentCount(VecType) > 1) ||
3371 VecType->getOpcode() == SPIRV::OpTypeVectorIdEXT) &&
3372 "dot product requires either a vector of at least 2 components "
3373 "or the SPV_EXT_long_vector extension.");
3374
3375 Register Res = MRI->createVirtualRegister(GR.getRegClass(ResType));
3376 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
3377 .addDef(Res)
3378 .addUse(GR.getSPIRVTypeID(ResType))
3379 .addUse(TmpVec)
3380 .addImm(0)
3381 .constrainAllUses(TII, TRI, RBI);
3382
3383 for (unsigned i = 1; i < GR.getScalarOrVectorComponentCount(VecType); i++) {
3384 Register Elt = MRI->createVirtualRegister(GR.getRegClass(ResType));
3385
3386 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
3387 .addDef(Elt)
3388 .addUse(GR.getSPIRVTypeID(ResType))
3389 .addUse(TmpVec)
3390 .addImm(i)
3391 .constrainAllUses(TII, TRI, RBI);
3392
3393 Register Sum = i < GR.getScalarOrVectorComponentCount(VecType) - 1
3394 ? MRI->createVirtualRegister(GR.getRegClass(ResType))
3395 : ResVReg;
3396
3397 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIAddS))
3398 .addDef(Sum)
3399 .addUse(GR.getSPIRVTypeID(ResType))
3400 .addUse(Res)
3401 .addUse(Elt)
3402 .constrainAllUses(TII, TRI, RBI);
3403 Res = Sum;
3404 }
3405
3406 return true;
3407}
3408
3409bool SPIRVInstructionSelector::selectOpIsInf(Register ResVReg,
3410 SPIRVTypeInst ResType,
3411 MachineInstr &I) const {
3412 MachineBasicBlock &BB = *I.getParent();
3413 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIsInf))
3414 .addDef(ResVReg)
3415 .addUse(GR.getSPIRVTypeID(ResType))
3416 .addUse(I.getOperand(2).getReg())
3417 .constrainAllUses(TII, TRI, RBI);
3418 return true;
3419}
3420
3421bool SPIRVInstructionSelector::selectOpIsNan(Register ResVReg,
3422 SPIRVTypeInst ResType,
3423 MachineInstr &I) const {
3424 MachineBasicBlock &BB = *I.getParent();
3425 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIsNan))
3426 .addDef(ResVReg)
3427 .addUse(GR.getSPIRVTypeID(ResType))
3428 .addUse(I.getOperand(2).getReg())
3429 .constrainAllUses(TII, TRI, RBI);
3430 return true;
3431}
3432
3433bool SPIRVInstructionSelector::selectOpIsFinite(Register ResVReg,
3434 SPIRVTypeInst ResType,
3435 MachineInstr &I) const {
3436 MachineBasicBlock &BB = *I.getParent();
3437 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIsFinite))
3438 .addDef(ResVReg)
3439 .addUse(GR.getSPIRVTypeID(ResType))
3440 .addUse(I.getOperand(2).getReg())
3441 .constrainAllUses(TII, TRI, RBI);
3442 return true;
3443}
3444
3445bool SPIRVInstructionSelector::selectOpIsNormal(Register ResVReg,
3446 SPIRVTypeInst ResType,
3447 MachineInstr &I) const {
3448 MachineBasicBlock &BB = *I.getParent();
3449 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIsNormal))
3450 .addDef(ResVReg)
3451 .addUse(GR.getSPIRVTypeID(ResType))
3452 .addUse(I.getOperand(2).getReg())
3453 .constrainAllUses(TII, TRI, RBI);
3454 return true;
3455}
3456
3457template <bool Signed>
3458bool SPIRVInstructionSelector::selectDot4AddPacked(Register ResVReg,
3459 SPIRVTypeInst ResType,
3460 MachineInstr &I) const {
3461 assert(I.getNumOperands() == 5);
3462 assert(I.getOperand(2).isReg());
3463 assert(I.getOperand(3).isReg());
3464 assert(I.getOperand(4).isReg());
3465 MachineBasicBlock &BB = *I.getParent();
3466
3467 Register Acc = I.getOperand(2).getReg();
3468 Register X = I.getOperand(3).getReg();
3469 Register Y = I.getOperand(4).getReg();
3470
3471 auto DotOp = Signed ? SPIRV::OpSDot : SPIRV::OpUDot;
3472 Register Dot = MRI->createVirtualRegister(GR.getRegClass(ResType));
3473 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(DotOp))
3474 .addDef(Dot)
3475 .addUse(GR.getSPIRVTypeID(ResType))
3476 .addUse(X)
3477 .addUse(Y);
3478 MIB.addImm(SPIRV::BuiltIn::PackedVectorFormat4x8Bit);
3479 MIB.constrainAllUses(TII, TRI, RBI);
3480
3481 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIAddS))
3482 .addDef(ResVReg)
3483 .addUse(GR.getSPIRVTypeID(ResType))
3484 .addUse(Dot)
3485 .addUse(Acc)
3486 .constrainAllUses(TII, TRI, RBI);
3487 return true;
3488}
3489
3490// Since pre-1.6 SPIRV has no DotProductInput4x8BitPacked implementation,
3491// extract the elements of the packed inputs, multiply them and add the result
3492// to the accumulator.
3493template <bool Signed>
3494bool SPIRVInstructionSelector::selectDot4AddPackedExpansion(
3495 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
3496 assert(I.getNumOperands() == 5);
3497 assert(I.getOperand(2).isReg());
3498 assert(I.getOperand(3).isReg());
3499 assert(I.getOperand(4).isReg());
3500 MachineBasicBlock &BB = *I.getParent();
3501
3502 Register Acc = I.getOperand(2).getReg();
3503 Register X = I.getOperand(3).getReg();
3504 Register Y = I.getOperand(4).getReg();
3505
3506 SPIRVTypeInst EltType = GR.getOrCreateSPIRVIntegerType(8, I, TII);
3507 auto ExtractOp =
3508 Signed ? SPIRV::OpBitFieldSExtract : SPIRV::OpBitFieldUExtract;
3509
3510 bool ZeroAsNull = !STI.isShader();
3511 // Extract the i8 element, multiply and add it to the accumulator
3512 for (unsigned i = 0; i < 4; i++) {
3513 // A[i]
3514 Register AElt = MRI->createVirtualRegister(&SPIRV::IDRegClass);
3515 BuildMI(BB, I, I.getDebugLoc(), TII.get(ExtractOp))
3516 .addDef(AElt)
3517 .addUse(GR.getSPIRVTypeID(ResType))
3518 .addUse(X)
3519 .addUse(GR.getOrCreateConstInt(i * 8, I, EltType, TII, ZeroAsNull))
3520 .addUse(GR.getOrCreateConstInt(8, I, EltType, TII, ZeroAsNull))
3521 .constrainAllUses(TII, TRI, RBI);
3522
3523 // B[i]
3524 Register BElt = MRI->createVirtualRegister(&SPIRV::IDRegClass);
3525 BuildMI(BB, I, I.getDebugLoc(), TII.get(ExtractOp))
3526 .addDef(BElt)
3527 .addUse(GR.getSPIRVTypeID(ResType))
3528 .addUse(Y)
3529 .addUse(GR.getOrCreateConstInt(i * 8, I, EltType, TII, ZeroAsNull))
3530 .addUse(GR.getOrCreateConstInt(8, I, EltType, TII, ZeroAsNull))
3531 .constrainAllUses(TII, TRI, RBI);
3532
3533 // A[i] * B[i]
3534 Register Mul = MRI->createVirtualRegister(&SPIRV::IDRegClass);
3535 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIMulS))
3536 .addDef(Mul)
3537 .addUse(GR.getSPIRVTypeID(ResType))
3538 .addUse(AElt)
3539 .addUse(BElt)
3540 .constrainAllUses(TII, TRI, RBI);
3541
3542 // Discard 24 highest-bits so that stored i32 register is i8 equivalent
3543 Register MaskMul = MRI->createVirtualRegister(&SPIRV::IDRegClass);
3544 BuildMI(BB, I, I.getDebugLoc(), TII.get(ExtractOp))
3545 .addDef(MaskMul)
3546 .addUse(GR.getSPIRVTypeID(ResType))
3547 .addUse(Mul)
3548 .addUse(GR.getOrCreateConstInt(0, I, EltType, TII, ZeroAsNull))
3549 .addUse(GR.getOrCreateConstInt(8, I, EltType, TII, ZeroAsNull))
3550 .constrainAllUses(TII, TRI, RBI);
3551
3552 // Acc = Acc + A[i] * B[i]
3553 Register Sum =
3554 i < 3 ? MRI->createVirtualRegister(&SPIRV::IDRegClass) : ResVReg;
3555 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpIAddS))
3556 .addDef(Sum)
3557 .addUse(GR.getSPIRVTypeID(ResType))
3558 .addUse(Acc)
3559 .addUse(MaskMul)
3560 .constrainAllUses(TII, TRI, RBI);
3561
3562 Acc = Sum;
3563 }
3564
3565 return true;
3566}
3567
3568/// Transform saturate(x) to clamp(x, 0.0f, 1.0f) as SPIRV
3569/// does not have a saturate builtin.
3570bool SPIRVInstructionSelector::selectSaturate(Register ResVReg,
3571 SPIRVTypeInst ResType,
3572 MachineInstr &I) const {
3573 assert(I.getNumOperands() == 3);
3574 assert(I.getOperand(2).isReg());
3575 MachineBasicBlock &BB = *I.getParent();
3576 Register VZero = buildZerosValF(ResType, I);
3577 Register VOne = buildOnesValF(ResType, I);
3578
3579 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
3580 .addDef(ResVReg)
3581 .addUse(GR.getSPIRVTypeID(ResType))
3582 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
3583 .addImm(GL::FClamp)
3584 .addUse(I.getOperand(2).getReg())
3585 .addUse(VZero)
3586 .addUse(VOne)
3587 .constrainAllUses(TII, TRI, RBI);
3588 return true;
3589}
3590
3591bool SPIRVInstructionSelector::selectSign(Register ResVReg,
3592 SPIRVTypeInst ResType,
3593 MachineInstr &I) const {
3594 assert(I.getNumOperands() == 3);
3595 assert(I.getOperand(2).isReg());
3596 MachineBasicBlock &BB = *I.getParent();
3597 Register InputRegister = I.getOperand(2).getReg();
3598 SPIRVTypeInst InputType = GR.getSPIRVTypeForVReg(InputRegister);
3599 auto &DL = I.getDebugLoc();
3600
3601 if (!InputType)
3602 return diagnoseUnsupported(I, "Input Type could not be determined.");
3603
3604 bool IsFloatTy = GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeFloat);
3605
3606 unsigned SignBitWidth = GR.getScalarOrVectorBitWidth(InputType);
3607 unsigned ResBitWidth = GR.getScalarOrVectorBitWidth(ResType);
3608
3609 bool NeedsConversion = IsFloatTy || SignBitWidth != ResBitWidth;
3610
3611 auto SignOpcode = IsFloatTy ? GL::FSign : GL::SSign;
3612 Register SignReg = NeedsConversion
3613 ? MRI->createVirtualRegister(&SPIRV::IDRegClass)
3614 : ResVReg;
3615
3616 BuildMI(BB, I, DL, TII.get(SPIRV::OpExtInst))
3617 .addDef(SignReg)
3618 .addUse(GR.getSPIRVTypeID(InputType))
3619 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
3620 .addImm(SignOpcode)
3621 .addUse(InputRegister)
3622 .constrainAllUses(TII, TRI, RBI);
3623
3624 if (NeedsConversion) {
3625 auto ConvertOpcode = IsFloatTy ? SPIRV::OpConvertFToS : SPIRV::OpSConvert;
3626 BuildMI(*I.getParent(), I, DL, TII.get(ConvertOpcode))
3627 .addDef(ResVReg)
3628 .addUse(GR.getSPIRVTypeID(ResType))
3629 .addUse(SignReg)
3630 .constrainAllUses(TII, TRI, RBI);
3631 }
3632
3633 return true;
3634}
3635
3636bool SPIRVInstructionSelector::selectWaveOpInst(Register ResVReg,
3637 SPIRVTypeInst ResType,
3638 MachineInstr &I,
3639 unsigned Opcode) const {
3640 MachineBasicBlock &BB = *I.getParent();
3641 SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3642
3643 auto BMI = BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
3644 .addDef(ResVReg)
3645 .addUse(GR.getSPIRVTypeID(ResType))
3646 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I,
3647 IntTy, TII, !STI.isShader()));
3648
3649 for (unsigned J = 2; J < I.getNumOperands(); J++) {
3650 BMI.addUse(I.getOperand(J).getReg());
3651 }
3652
3653 BMI.constrainAllUses(TII, TRI, RBI);
3654 return true;
3655}
3656
3657bool SPIRVInstructionSelector::selectBarrierInst(MachineInstr &I,
3658 unsigned Scope,
3659 unsigned MemSem,
3660 bool WithGroupSync) const {
3661 auto BarrierType =
3662 WithGroupSync ? SPIRV::OpControlBarrier : SPIRV::OpMemoryBarrier;
3663
3664 MemSem |= SPIRV::MemorySemantics::AcquireRelease;
3665
3666 assert(((Scope != SPIRV::Scope::Workgroup) ||
3667 ((MemSem & SPIRV::MemorySemantics::WorkgroupMemory) > 0)) &&
3668 "Workgroup Scope must set WorkGroupMemory semantic "
3669 "in Barrier instruction");
3670
3671 assert(((Scope != SPIRV::Scope::Device) ||
3672 ((MemSem & SPIRV::MemorySemantics::UniformMemory) > 0 &&
3673 (MemSem & SPIRV::MemorySemantics::ImageMemory) > 0)) &&
3674 "Device Scope must set UniformMemory and ImageMemory semantic "
3675 "in Barrier instruction");
3676
3677 MachineBasicBlock &BB = *I.getParent();
3678 auto MI = BuildMI(BB, I, I.getDebugLoc(), TII.get(BarrierType));
3679
3680 // OpControlBarrier needs to also set Execution Scope
3681 if (WithGroupSync) {
3682 Register ExecReg = buildI32Constant(SPIRV::Scope::Workgroup, I);
3683 MI.addUse(ExecReg);
3684 }
3685
3686 Register ScopeReg = buildI32Constant(Scope, I);
3687 Register MemSemReg = buildI32Constant(MemSem, I);
3688
3689 MI.addUse(ScopeReg).addUse(MemSemReg).constrainAllUses(TII, TRI, RBI);
3690 return true;
3691}
3692
3693bool SPIRVInstructionSelector::selectWaveActiveCountBits(
3694 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
3695
3696 SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3697 SPIRVTypeInst BallotType = GR.getOrCreateSPIRVVectorType(IntTy, 4, I, TII);
3698 Register BallotReg = MRI->createVirtualRegister(GR.getRegClass(BallotType));
3699 if (!selectWaveOpInst(BallotReg, BallotType, I,
3700 SPIRV::OpGroupNonUniformBallot))
3701 return false;
3702
3703 MachineBasicBlock &BB = *I.getParent();
3704 BuildMI(BB, I, I.getDebugLoc(),
3705 TII.get(SPIRV::OpGroupNonUniformBallotBitCount))
3706 .addDef(ResVReg)
3707 .addUse(GR.getSPIRVTypeID(ResType))
3708 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy, TII,
3709 !STI.isShader()))
3710 .addImm(SPIRV::GroupOperation::Reduce)
3711 .addUse(BallotReg)
3712 .constrainAllUses(TII, TRI, RBI);
3713
3714 return true;
3715}
3716
3717bool SPIRVInstructionSelector::selectWaveActiveAllEqual(Register ResVReg,
3718 SPIRVTypeInst ResType,
3719 MachineInstr &I) const {
3720 MachineBasicBlock &BB = *I.getParent();
3721 const DebugLoc &DL = I.getDebugLoc();
3722
3723 // Input to the intrinsic
3724 Register InputReg = I.getOperand(2).getReg();
3725 SPIRVTypeInst InputType = GR.getSPIRVTypeForVReg(InputReg);
3726
3727 // Determine if input is vector
3728 unsigned NumElems = GR.getScalarOrVectorComponentCount(InputType);
3729 bool IsVector = NumElems > 1 ||
3730 (InputType->getOpcode() == SPIRV::OpTypeVectorIdEXT &&
3731 STI.canUseExtension(SPIRV::Extension::SPV_EXT_long_vector));
3732
3733 // Determine element types
3734 SPIRVTypeInst ElemInputType = GR.getScalarOrVectorComponentType(InputType);
3735 SPIRVTypeInst ElemBoolType = GR.getScalarOrVectorComponentType(ResType);
3736
3737 // Subgroup scope constant
3738 SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3739 Register ScopeConst = GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy,
3740 TII, !STI.isShader());
3741
3742 // Scalar case
3743 if (!IsVector) {
3744 return selectWaveOpInst(ResVReg, ElemBoolType, I,
3745 SPIRV::OpGroupNonUniformAllEqual);
3746 }
3747
3748 // Vector case
3749 SmallVector<Register, 4> ElementResults;
3750 ElementResults.reserve(NumElems);
3751
3752 for (unsigned Idx = 0; Idx < NumElems; ++Idx) {
3753 // Extract element
3754 Register ElemInput = InputReg;
3755 Register Extracted =
3756 MRI->createVirtualRegister(GR.getRegClass(ElemInputType));
3757
3758 BuildMI(BB, I, DL, TII.get(SPIRV::OpCompositeExtract))
3759 .addDef(Extracted)
3760 .addUse(GR.getSPIRVTypeID(ElemInputType))
3761 .addUse(InputReg)
3762 .addImm(Idx)
3763 .constrainAllUses(TII, TRI, RBI);
3764
3765 ElemInput = Extracted;
3766
3767 // Emit per-element AllEqual
3768 Register ElemResult =
3769 MRI->createVirtualRegister(GR.getRegClass(ElemBoolType));
3770
3771 BuildMI(BB, I, DL, TII.get(SPIRV::OpGroupNonUniformAllEqual))
3772 .addDef(ElemResult)
3773 .addUse(GR.getSPIRVTypeID(ElemBoolType))
3774 .addUse(ScopeConst)
3775 .addUse(ElemInput)
3776 .constrainAllUses(TII, TRI, RBI);
3777
3778 ElementResults.push_back(ElemResult);
3779 }
3780
3781 // Reconstruct vector<bool>
3782 auto MIB = BuildMI(BB, I, DL, TII.get(SPIRV::OpCompositeConstruct))
3783 .addDef(ResVReg)
3784 .addUse(GR.getSPIRVTypeID(ResType));
3785 for (Register R : ElementResults)
3786 MIB.addUse(R);
3787
3788 MIB.constrainAllUses(TII, TRI, RBI);
3789
3790 return true;
3791}
3792
3793bool SPIRVInstructionSelector::selectWavePrefixBitCount(Register ResVReg,
3794 SPIRVTypeInst ResType,
3795 MachineInstr &I) const {
3796
3797 assert(I.getNumOperands() == 3);
3798
3799 auto Op = I.getOperand(2);
3800 assert(Op.isReg());
3801
3802 MachineBasicBlock &BB = *I.getParent();
3803 DebugLoc DL = I.getDebugLoc();
3804
3805 Register InputRegister = Op.getReg();
3806 SPIRVTypeInst InputType = GR.getSPIRVTypeForVReg(InputRegister);
3807
3808 if (!InputType)
3809 return diagnoseUnsupported(I, "Input Type could not be determined.");
3810
3811 if (InputType->getOpcode() != SPIRV::OpTypeBool)
3812 return diagnoseUnsupported(I, "WavePrefixBitCount requires boolean input");
3813
3814 // Types
3815 SPIRVTypeInst Int32Ty = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3816
3817 // Ballot result type: vector<uint32>
3818 // Match DXC: %v4uint for Subgroup size
3819 SPIRVTypeInst BallotTy = GR.getOrCreateSPIRVVectorType(Int32Ty, 4, I, TII);
3820
3821 // Create a vreg for the ballot result
3822 Register BallotVReg = MRI->createVirtualRegister(&SPIRV::IDRegClass);
3823
3824 // 1. OpGroupNonUniformBallot
3825 BuildMI(BB, I, DL, TII.get(SPIRV::OpGroupNonUniformBallot))
3826 .addDef(BallotVReg)
3827 .addUse(GR.getSPIRVTypeID(BallotTy))
3828 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, Int32Ty, TII))
3829 .addUse(InputRegister)
3830 .constrainAllUses(TII, TRI, RBI);
3831
3832 // 2. OpGroupNonUniformBallotBitCount
3833 BuildMI(BB, I, DL, TII.get(SPIRV::OpGroupNonUniformBallotBitCount))
3834 .addDef(ResVReg)
3835 .addUse(GR.getSPIRVTypeID(ResType))
3836 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, Int32Ty, TII))
3837 .addImm(SPIRV::GroupOperation::ExclusiveScan)
3838 .addUse(BallotVReg)
3839 .constrainAllUses(TII, TRI, RBI);
3840
3841 return true;
3842}
3843
3844bool SPIRVInstructionSelector::selectWaveReduceMax(Register ResVReg,
3845 SPIRVTypeInst ResType,
3846 MachineInstr &I,
3847 bool IsUnsigned) const {
3848 return selectWaveReduce(
3849 ResVReg, ResType, I, IsUnsigned,
3850 [&](Register InputRegister, bool IsUnsigned) {
3851 const bool IsFloatTy =
3852 GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeFloat);
3853 const auto IntOp = IsUnsigned ? SPIRV::OpGroupNonUniformUMax
3854 : SPIRV::OpGroupNonUniformSMax;
3855 return IsFloatTy ? SPIRV::OpGroupNonUniformFMax : IntOp;
3856 });
3857}
3858
3859bool SPIRVInstructionSelector::selectWaveReduceMin(Register ResVReg,
3860 SPIRVTypeInst ResType,
3861 MachineInstr &I,
3862 bool IsUnsigned) const {
3863 return selectWaveReduce(
3864 ResVReg, ResType, I, IsUnsigned,
3865 [&](Register InputRegister, bool IsUnsigned) {
3866 const bool IsFloatTy =
3867 GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeFloat);
3868 const auto IntOp = IsUnsigned ? SPIRV::OpGroupNonUniformUMin
3869 : SPIRV::OpGroupNonUniformSMin;
3870 return IsFloatTy ? SPIRV::OpGroupNonUniformFMin : IntOp;
3871 });
3872}
3873
3874bool SPIRVInstructionSelector::selectWaveReduceSum(Register ResVReg,
3875 SPIRVTypeInst ResType,
3876 MachineInstr &I) const {
3877 return selectWaveReduce(ResVReg, ResType, I, /*IsUnsigned*/ false,
3878 [&](Register InputRegister, bool IsUnsigned) {
3879 bool IsFloatTy = GR.isScalarOrVectorOfType(
3880 InputRegister, SPIRV::OpTypeFloat);
3881 return IsFloatTy ? SPIRV::OpGroupNonUniformFAdd
3882 : SPIRV::OpGroupNonUniformIAdd;
3883 });
3884}
3885
3886bool SPIRVInstructionSelector::selectWaveReduceProduct(Register ResVReg,
3887 SPIRVTypeInst ResType,
3888 MachineInstr &I) const {
3889 return selectWaveReduce(ResVReg, ResType, I, /*IsUnsigned*/ false,
3890 [&](Register InputRegister, bool IsUnsigned) {
3891 bool IsFloatTy = GR.isScalarOrVectorOfType(
3892 InputRegister, SPIRV::OpTypeFloat);
3893 return IsFloatTy ? SPIRV::OpGroupNonUniformFMul
3894 : SPIRV::OpGroupNonUniformIMul;
3895 });
3896}
3897
3898template <typename PickOpcodeFn>
3899bool SPIRVInstructionSelector::selectWaveReduce(
3900 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I, bool IsUnsigned,
3901 PickOpcodeFn &&PickOpcode) const {
3902 assert(I.getNumOperands() == 3);
3903 assert(I.getOperand(2).isReg());
3904 MachineBasicBlock &BB = *I.getParent();
3905 Register InputRegister = I.getOperand(2).getReg();
3906 SPIRVTypeInst InputType = GR.getSPIRVTypeForVReg(InputRegister);
3907
3908 if (!InputType)
3909 return diagnoseUnsupported(I, "Input Type could not be determined.");
3910
3911 SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3912 const unsigned Opcode = PickOpcode(InputRegister, IsUnsigned);
3913 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
3914 .addDef(ResVReg)
3915 .addUse(GR.getSPIRVTypeID(ResType))
3916 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy, TII,
3917 !STI.isShader()))
3918 .addImm(SPIRV::GroupOperation::Reduce)
3919 .addUse(I.getOperand(2).getReg())
3920 .constrainAllUses(TII, TRI, RBI);
3921 return true;
3922}
3923
3924bool SPIRVInstructionSelector::selectWaveReduceOp(Register ResVReg,
3925 SPIRVTypeInst ResType,
3926 MachineInstr &I,
3927 unsigned Opcode) const {
3928 return selectWaveReduce(
3929 ResVReg, ResType, I, false,
3930 [&](Register InputRegister, bool IsUnsigned) { return Opcode; });
3931}
3932
3933bool SPIRVInstructionSelector::selectWaveExclusiveScanSum(
3934 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
3935 return selectWaveExclusiveScan(ResVReg, ResType, I, /*IsUnsigned*/ false,
3936 [&](Register InputRegister, bool IsUnsigned) {
3937 bool IsFloatTy = GR.isScalarOrVectorOfType(
3938 InputRegister, SPIRV::OpTypeFloat);
3939 return IsFloatTy
3940 ? SPIRV::OpGroupNonUniformFAdd
3941 : SPIRV::OpGroupNonUniformIAdd;
3942 });
3943}
3944
3945bool SPIRVInstructionSelector::selectWaveExclusiveScanProduct(
3946 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
3947 return selectWaveExclusiveScan(ResVReg, ResType, I, /*IsUnsigned*/ false,
3948 [&](Register InputRegister, bool IsUnsigned) {
3949 bool IsFloatTy = GR.isScalarOrVectorOfType(
3950 InputRegister, SPIRV::OpTypeFloat);
3951 return IsFloatTy
3952 ? SPIRV::OpGroupNonUniformFMul
3953 : SPIRV::OpGroupNonUniformIMul;
3954 });
3955}
3956
3957template <typename PickOpcodeFn>
3958bool SPIRVInstructionSelector::selectWaveExclusiveScan(
3959 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I, bool IsUnsigned,
3960 PickOpcodeFn &&PickOpcode) const {
3961 assert(I.getNumOperands() == 3);
3962 assert(I.getOperand(2).isReg());
3963 MachineBasicBlock &BB = *I.getParent();
3964 Register InputRegister = I.getOperand(2).getReg();
3965 SPIRVTypeInst InputType = GR.getSPIRVTypeForVReg(InputRegister);
3966
3967 if (!InputType)
3968 return diagnoseUnsupported(I, "Input Type could not be determined.");
3969
3970 SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3971 const unsigned Opcode = PickOpcode(InputRegister, IsUnsigned);
3972 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
3973 .addDef(ResVReg)
3974 .addUse(GR.getSPIRVTypeID(ResType))
3975 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy, TII,
3976 !STI.isShader()))
3977 .addImm(SPIRV::GroupOperation::ExclusiveScan)
3978 .addUse(I.getOperand(2).getReg())
3979 .constrainAllUses(TII, TRI, RBI);
3980 return true;
3981}
3982
3983bool SPIRVInstructionSelector::selectQuadSwap(Register ResVReg,
3984 SPIRVTypeInst ResType,
3985 MachineInstr &I,
3986 unsigned Direction) const {
3987 assert(I.getNumOperands() == 3);
3988 assert(I.getOperand(2).isReg());
3989 MachineBasicBlock &BB = *I.getParent();
3990 Register InputRegister = I.getOperand(2).getReg();
3991
3992 SPIRVTypeInst IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
3993 bool ZeroAsNull = !STI.isShader();
3994 Register DirectionReg =
3995 GR.getOrCreateConstInt(Direction, I, IntTy, TII, ZeroAsNull);
3996 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpGroupNonUniformQuadSwap))
3997 .addDef(ResVReg)
3998 .addUse(GR.getSPIRVTypeID(ResType))
3999 .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy, TII,
4000 ZeroAsNull))
4001 .addUse(InputRegister)
4002 .addUse(DirectionReg)
4003 .constrainAllUses(TII, TRI, RBI);
4004 return true;
4005}
4006
4007bool SPIRVInstructionSelector::selectBitreverseViaI32(Register ResVReg,
4008 SPIRVTypeInst ResType,
4009 MachineInstr &I,
4010 Register Op) const {
4011 SPIRVTypeInst Int32Type = GR.getOrCreateSPIRVIntegerType(32, I, TII);
4012 const unsigned BitWidth = GR.getScalarOrVectorBitWidth(ResType);
4013 Register ShiftConst =
4014 GR.getOrCreateConstInt(32 - BitWidth, I, Int32Type, TII);
4015 unsigned ShiftOp = SPIRV::OpShiftRightLogicalS;
4016
4017 const unsigned N = GR.getScalarOrVectorComponentCount(ResType);
4018 const unsigned ExtendOpcode = GR.isScalarOrVectorSigned(ResType)
4019 ? SPIRV::OpSConvert
4020 : SPIRV::OpUConvert;
4021
4022 if (N > 1 || (ResType->getOpcode() == SPIRV::OpTypeVectorIdEXT &&
4023 STI.canUseExtension(SPIRV::Extension::SPV_EXT_long_vector))) {
4024 Int32Type = GR.getOrCreateSPIRVVectorType(Int32Type, N, I, TII);
4025 ShiftOp = SPIRV::OpShiftRightLogicalV;
4026
4027 // Vector shifts require a composite constant
4028 const Register CompositeReg =
4029 MRI->createVirtualRegister(GR.getRegClass(Int32Type));
4030 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
4031 TII.get(SPIRV::OpConstantComposite))
4032 .addDef(CompositeReg)
4033 .addUse(GR.getSPIRVTypeID(Int32Type));
4034 for (unsigned It = 0; It < N; ++It)
4035 MIB.addUse(ShiftConst);
4036 MIB.constrainAllUses(TII, TRI, RBI);
4037
4038 ShiftConst = CompositeReg;
4039 }
4040
4041 // Converts the input to i32 (or vector of i32)
4042 Register ExtReg = MRI->createVirtualRegister(GR.getRegClass(Int32Type));
4043 if (!selectOpWithSrcs(ExtReg, Int32Type, I, {Op}, ExtendOpcode))
4044 return false;
4045
4046 // Perform bitreverse on the i32 value
4047 Register BitrevReg = MRI->createVirtualRegister(GR.getRegClass(Int32Type));
4048 if (!selectBitreverseNative(BitrevReg, Int32Type, I, ExtReg))
4049 return false;
4050
4051 // Shift the bit-reversed value to get the final result.
4052 Register ShiftReg = MRI->createVirtualRegister(GR.getRegClass(Int32Type));
4053 if (!selectOpWithSrcs(ShiftReg, Int32Type, I, {BitrevReg, ShiftConst},
4054 ShiftOp))
4055 return false;
4056
4057 // Finally, convert the result back.
4058 return selectOpWithSrcs(ResVReg, ResType, I, {ShiftReg}, ExtendOpcode);
4059}
4060
4061bool SPIRVInstructionSelector::handle64BitOverflow(
4062 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I, Register SrcReg,
4063 unsigned int Opcode,
4064 std::function<bool(Register, SPIRVTypeInst, MachineInstr &, Register,
4065 unsigned)>
4066 CallbackFunction) const {
4067
4068 SPIRVTypeInst BaseType = GR.retrieveScalarOrVectorIntType(ResType);
4069 assert(BaseType->getOpcode() == SPIRV::OpTypeInt &&
4070 "handle64BitOverflow should only be used for integer types");
4071 unsigned ComponentCount = GR.getScalarOrVectorComponentCount(ResType);
4072 assert(ComponentCount < 5 && "Vec 5+ will generate invalid SPIR-V ops");
4073
4074 MachineIRBuilder MIRBuilder(I);
4075 SPIRVTypeInst I64Type = GR.getOrCreateSPIRVIntegerType(64, MIRBuilder);
4076 SPIRVTypeInst I64x2Type =
4077 GR.getOrCreateSPIRVVectorType(I64Type, 2, MIRBuilder, false);
4078 SPIRVTypeInst Vec2ResType =
4079 GR.getOrCreateSPIRVVectorType(BaseType, 2, MIRBuilder, false);
4080
4081 std::vector<Register> PartialRegs;
4082
4083 unsigned CurrentComponent = 0;
4084 for (; CurrentComponent + 1 < ComponentCount; CurrentComponent += 2) {
4085 Register PopCountResult =
4086 MRI->createVirtualRegister(GR.getRegClass(I64x2Type));
4087
4088 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
4089 TII.get(SPIRV::OpVectorShuffle))
4090 .addDef(PopCountResult)
4091 .addUse(GR.getSPIRVTypeID(I64x2Type))
4092 .addUse(SrcReg)
4093 .addUse(SrcReg)
4094 .addImm(CurrentComponent)
4095 .addImm(CurrentComponent + 1);
4096
4097 MIB.constrainAllUses(TII, TRI, RBI);
4098
4099 Register SubVecReg =
4100 MRI->createVirtualRegister(GR.getRegClass(Vec2ResType));
4101
4102 if (!CallbackFunction(SubVecReg, Vec2ResType, I, PopCountResult, Opcode))
4103 return false;
4104
4105 PartialRegs.push_back(SubVecReg);
4106 }
4107 // On odd component counts we need to handle one more component
4108 if (CurrentComponent != ComponentCount) {
4109 bool ZeroAsNull = !STI.isShader();
4110 Register FinalElemReg = MRI->createVirtualRegister(GR.getRegClass(I64Type));
4111 Register ConstIntLastIdx = GR.getOrCreateConstInt(
4112 ComponentCount - 1, I, BaseType, TII, ZeroAsNull);
4113
4114 if (!selectOpWithSrcs(FinalElemReg, I64Type, I, {SrcReg, ConstIntLastIdx},
4115 SPIRV::OpVectorExtractDynamic))
4116 return false;
4117
4118 Register FinalElemResReg =
4119 MRI->createVirtualRegister(GR.getRegClass(BaseType));
4120
4121 if (!CallbackFunction(FinalElemResReg, BaseType, I, FinalElemReg, Opcode))
4122 return false;
4123
4124 PartialRegs.push_back(FinalElemResReg);
4125 }
4126 // Join all the resulting registers back into the return type in order
4127 // (ie i32x2, i32x2, i32x1 -> i32x5)
4128 return selectOpWithSrcs(ResVReg, ResType, I, PartialRegs,
4129 SPIRV::OpCompositeConstruct);
4130}
4131
4132bool SPIRVInstructionSelector::selectBitreverse64(Register ResVReg,
4133 SPIRVTypeInst ResType,
4134 MachineInstr &I,
4135 Register SrcReg) const {
4136 unsigned ComponentCount = GR.getScalarOrVectorComponentCount(ResType);
4137 if (ComponentCount > 2)
4138 return handle64BitOverflow(
4139 ResVReg, ResType, I, SrcReg, SPIRV::OpBitReverse,
4140 [this](Register R, SPIRVTypeInst T, MachineInstr &I, Register S,
4141 unsigned O) { return this->selectBitreverse64(R, T, I, S); });
4142
4143 MachineIRBuilder MIRBuilder(I);
4144
4145 SPIRVTypeInst I32Type = GR.getOrCreateSPIRVIntegerType(32, MIRBuilder);
4146 SPIRVTypeInst VecI32Type = GR.getOrCreateSPIRVVectorType(
4147 I32Type, 2 * ComponentCount, MIRBuilder, /*IsSigned=*/false);
4148
4149 // Converts 64 bit into and array of 32 bit, containing 2 elements.
4150 Register Vec32 = MRI->createVirtualRegister(GR.getRegClass(VecI32Type));
4151 if (!selectOpWithSrcs(Vec32, VecI32Type, I, {SrcReg}, SPIRV::OpBitcast))
4152 return false;
4153
4154 // Apply bitreverse on each 32 bit lane
4155 Register Reverse32 = MRI->createVirtualRegister(GR.getRegClass(VecI32Type));
4156 if (!selectBitreverseNative(Reverse32, VecI32Type, I, Vec32))
4157 return false;
4158
4159 // Reversing a 64-bit value = reverse each 32-bit half AND swap them,
4160 // so the old High word becomes lane 0 (low) and old Low becomes lane 1
4161 // (high).
4162 Register SwappedVec = MRI->createVirtualRegister(GR.getRegClass(VecI32Type));
4163 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
4164 TII.get(SPIRV::OpVectorShuffle))
4165 .addDef(SwappedVec)
4166 .addUse(GR.getSPIRVTypeID(VecI32Type))
4167 .addUse(Reverse32)
4168 .addUse(Reverse32);
4169 for (unsigned J = 0; J < ComponentCount; ++J) {
4170 MIB.addImm(2 * J + 1);
4171 MIB.addImm(2 * J);
4172 }
4173 MIB.constrainAllUses(TII, TRI, RBI);
4174
4175 // Groups 32 bit vector back to 64 bit scalar.
4176 return selectOpWithSrcs(ResVReg, ResType, I, {SwappedVec}, SPIRV::OpBitcast);
4177}
4178
4179bool SPIRVInstructionSelector::selectBitreverseNative(Register ResVReg,
4180 SPIRVTypeInst ResType,
4181 MachineInstr &I,
4182 Register Op) const {
4183 MachineBasicBlock &BB = *I.getParent();
4184 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpBitReverse))
4185 .addDef(ResVReg)
4186 .addUse(GR.getSPIRVTypeID(ResType))
4187 .addUse(Op)
4188 .constrainAllUses(TII, TRI, RBI);
4189 return true;
4190}
4191
4192bool SPIRVInstructionSelector::selectBitreverse(Register ResVReg,
4193 SPIRVTypeInst ResType,
4194 MachineInstr &I) const {
4195 Register OpReg = I.getOperand(1).getReg();
4196
4197 // TODO: Fix shader behavior in case of VK_KHR_maintenance9 extension is set
4198 if (STI.isShader()) {
4199 SPIRVTypeInst OpType = GR.getSPIRVTypeForVReg(OpReg);
4200 switch (GR.getScalarOrVectorBitWidth(OpType)) {
4201 case 8:
4202 case 16:
4203 case 24:
4204 return selectBitreverseViaI32(ResVReg, ResType, I, OpReg);
4205 case 32:
4206 return selectBitreverseNative(ResVReg, ResType, I, OpReg);
4207 case 64:
4208 return selectBitreverse64(ResVReg, ResType, I, OpReg);
4209 }
4210 return SPIRVInstructionSelector::diagnoseUnsupported(
4211 I, "G_BITREVERSE only support 16,32,64 bits.");
4212 }
4213
4214 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_bit_instructions))
4215 return selectBitreverseNative(ResVReg, ResType, I, OpReg);
4216
4217 // Expansion bitreverse using bit manipulation operations
4218 // Algo: https://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
4219 const unsigned BitWidth = GR.getScalarOrVectorBitWidth(ResType);
4220 // TODO: add support for any bit width and bitwidth more than 64.
4221 if (BitWidth > 64 || !isPowerOf2_32(BitWidth))
4222 return false;
4223
4224 const unsigned N = GR.getScalarOrVectorComponentCount(ResType);
4225
4226 unsigned AndOp = SPIRV::OpBitwiseAndS;
4227 unsigned OrOp = SPIRV::OpBitwiseOrS;
4228 unsigned ShlOp = SPIRV::OpShiftLeftLogicalS;
4229 unsigned ShrOp = SPIRV::OpShiftRightLogicalS;
4230 if (N > 1 || (ResType->getOpcode() == SPIRV::OpTypeVectorIdEXT &&
4231 STI.canUseExtension(SPIRV::Extension::SPV_EXT_long_vector))) {
4232 AndOp = SPIRV::OpBitwiseAndV;
4233 OrOp = SPIRV::OpBitwiseOrV;
4234 ShlOp = SPIRV::OpShiftLeftLogicalV;
4235 ShrOp = SPIRV::OpShiftRightLogicalV;
4236 }
4237
4238 // Helper, one swap per step: ((input>>shift)&mask)|((input&mask)<<shift),
4239 // RPN: input shift >> mask & input mask & shift << |
4240 auto SwapBits = [&](const Register Input, const uint64_t Mask,
4241 const unsigned Shift) -> Register {
4242 auto CreateConst = [&](const uint64_t Value) -> Register {
4243 if (N == 1 &&
4244 (ResType->getOpcode() != SPIRV::OpTypeVectorIdEXT ||
4245 !STI.canUseExtension(SPIRV::Extension::SPV_EXT_long_vector)))
4246 return GR.getOrCreateConstInt(
4247 Value, I, GR.retrieveScalarOrVectorIntType(ResType), TII);
4248 return GR.getOrCreateConstVector(Value, I, ResType, TII);
4249 };
4250
4251 Register MaskReg = CreateConst(Mask);
4252 Register ShiftReg = CreateConst(Shift);
4253 Register T1 = MRI->createVirtualRegister(GR.getRegClass(ResType));
4254 Register T2 = MRI->createVirtualRegister(GR.getRegClass(ResType));
4255 Register T3 = MRI->createVirtualRegister(GR.getRegClass(ResType));
4256 Register T4 = MRI->createVirtualRegister(GR.getRegClass(ResType));
4257 Register Result = MRI->createVirtualRegister(GR.getRegClass(ResType));
4258
4259 if (!selectOpWithSrcs(T1, ResType, I, {Input, ShiftReg}, ShrOp) ||
4260 !selectOpWithSrcs(T2, ResType, I, {T1, MaskReg}, AndOp) ||
4261 !selectOpWithSrcs(T3, ResType, I, {Input, MaskReg}, AndOp) ||
4262 !selectOpWithSrcs(T4, ResType, I, {T3, ShiftReg}, ShlOp) ||
4263 !selectOpWithSrcs(Result, ResType, I, {T2, T4}, OrOp))
4264 return Register();
4265
4266 return Result;
4267 };
4268
4269 unsigned Shift = BitWidth;
4270 Register Result = OpReg;
4271 uint64_t Mask = ~0ull;
4272 while ((Shift >>= 1) > 0) {
4273 Mask ^= (Mask << Shift);
4274 Result = SwapBits(Result, Mask, Shift);
4275 if (!Result.isValid())
4276 return false;
4277 }
4278
4279 return BuildCOPY(ResVReg, Result, I);
4280}
4281
4282bool SPIRVInstructionSelector::selectFreeze(Register ResVReg,
4283 SPIRVTypeInst ResType,
4284 MachineInstr &I) const {
4285 assert(I.getOperand(0).isReg() && I.getOperand(1).isReg() &&
4286 "G_FREEZE must define and use a register");
4287 Register OpReg = I.getOperand(1).getReg();
4288
4289 // With SPV_KHR_poison_freeze, lower `freeze` to OpFreezeKHR.
4290 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_poison_freeze)) {
4291 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpFreezeKHR))
4292 .addDef(ResVReg)
4293 .addUse(GR.getSPIRVTypeID(ResType))
4294 .addUse(OpReg)
4295 .constrainAllUses(TII, TRI, RBI);
4296 return true;
4297 }
4298
4299 // There is no way to implement `freeze` correctly without support on SPIR-V
4300 // standard side, but we may at least address a simple (static) case when
4301 // undef/poison value presence is obvious. The main benefit of even
4302 // incomplete `freeze` support is preventing of translation from crashing due
4303 // to lack of support on legalization and instruction selection steps.
4304 if (MachineInstr *Def = MRI->getVRegDef(OpReg)) {
4305 if (Def->getOpcode() == TargetOpcode::COPY)
4306 Def = MRI->getVRegDef(Def->getOperand(1).getReg());
4307 Register Reg;
4308 switch (Def->getOpcode()) {
4309 case SPIRV::ASSIGN_TYPE:
4310 if (MachineInstr *AssignToDef =
4311 MRI->getVRegDef(Def->getOperand(1).getReg())) {
4312 if (AssignToDef->getOpcode() == TargetOpcode::G_IMPLICIT_DEF)
4313 Reg = Def->getOperand(2).getReg();
4314 }
4315 break;
4316 case SPIRV::OpUndef:
4317 Reg = Def->getOperand(1).getReg();
4318 break;
4319 }
4320 unsigned DestOpCode;
4321 if (Reg.isValid()) {
4322 DestOpCode = SPIRV::OpConstantNull;
4323 LLVM_DEBUG(dbgs() << "SPV_KHR_poison_freeze is not enabled. freeze of a "
4324 "static undef/poison lowered to OpConstantNull\n");
4325 } else {
4326 DestOpCode = TargetOpcode::COPY;
4327 Reg = OpReg;
4328 LLVM_DEBUG(dbgs() << "SPV_KHR_poison_freeze is not enabled. freeze "
4329 "skipped, lowered as a copy of the operand\n");
4330 }
4331 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(DestOpCode))
4332 .addDef(I.getOperand(0).getReg())
4333 .addUse(Reg)
4334 .constrainAllUses(TII, TRI, RBI);
4335 return true;
4336 }
4337 return false;
4338}
4339
4340bool SPIRVInstructionSelector::selectBuildVector(Register ResVReg,
4341 SPIRVTypeInst ResType,
4342 MachineInstr &I) const {
4343 unsigned N = 0;
4344 if (isVectorType(ResType))
4345 N = GR.getScalarOrVectorComponentCount(ResType);
4346 else if (ResType->getOpcode() == SPIRV::OpTypeArray)
4347 N = getArrayComponentCount(MRI, ResType);
4348 else
4349 report_fatal_error("Cannot select G_BUILD_VECTOR with a non-vector result");
4350 if (I.getNumExplicitOperands() - I.getNumExplicitDefs() != N)
4351 report_fatal_error("G_BUILD_VECTOR and the result type are inconsistent");
4352
4353 // check if we may construct a constant vector
4354 bool IsConst = true;
4355 for (unsigned i = I.getNumExplicitDefs();
4356 i < I.getNumExplicitOperands() && IsConst; ++i)
4357 if (!isConstReg(MRI, I.getOperand(i).getReg()))
4358 IsConst = false;
4359
4360 if (!IsConst && (N < 2 && ResType->getOpcode() != SPIRV::OpTypeVectorIdEXT))
4361 return diagnoseUnsupported(
4362 I, "There must be at least two constituent operands in a vector");
4363
4364 MRI->setRegClass(ResVReg, GR.getRegClass(ResType));
4365
4366 bool IsNullVector = IsConst && !STI.isShader();
4367 for (unsigned i = I.getNumExplicitDefs();
4368 i < I.getNumExplicitOperands() && IsNullVector; ++i) {
4369 MachineInstr *Def = getDef(I.getOperand(i), MRI);
4370 IsNullVector = Def && isNullOrNullSplat(*Def, *MRI);
4371 }
4372
4373 if (IsNullVector) {
4374 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull))
4375 .addDef(ResVReg)
4376 .addUse(GR.getSPIRVTypeID(ResType))
4377 .constrainAllUses(TII, TRI, RBI);
4378 return true;
4379 }
4380
4381 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
4382 TII.get(IsConst ? SPIRV::OpConstantComposite
4383 : SPIRV::OpCompositeConstruct))
4384 .addDef(ResVReg)
4385 .addUse(GR.getSPIRVTypeID(ResType));
4386 for (unsigned i = I.getNumExplicitDefs(); i < I.getNumExplicitOperands(); ++i)
4387 MIB.addUse(I.getOperand(i).getReg());
4388 MIB.constrainAllUses(TII, TRI, RBI);
4389 return true;
4390}
4391
4392bool SPIRVInstructionSelector::selectSplatVector(Register ResVReg,
4393 SPIRVTypeInst ResType,
4394 MachineInstr &I) const {
4395 unsigned N = 0;
4396 if (isVectorType(ResType))
4397 N = GR.getScalarOrVectorComponentCount(ResType);
4398 else if (ResType->getOpcode() == SPIRV::OpTypeArray)
4399 N = getArrayComponentCount(MRI, ResType);
4400 else
4401 report_fatal_error("Cannot select G_SPLAT_VECTOR with a non-vector result");
4402
4403 unsigned OpIdx = I.getNumExplicitDefs();
4404 if (!I.getOperand(OpIdx).isReg())
4405 report_fatal_error("Unexpected argument in G_SPLAT_VECTOR");
4406
4407 // check if we may construct a constant vector
4408 Register OpReg = I.getOperand(OpIdx).getReg();
4409 bool IsConst = isConstReg(MRI, OpReg);
4410
4411 if (!IsConst && (N < 2 && ResType->getOpcode() != SPIRV::OpTypeVectorIdEXT))
4412 return diagnoseUnsupported(
4413 I, "There must be at least two constituent operands in a vector");
4414
4415 MRI->setRegClass(ResVReg, GR.getRegClass(ResType));
4416 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
4417 TII.get(IsConst ? SPIRV::OpConstantComposite
4418 : SPIRV::OpCompositeConstruct))
4419 .addDef(ResVReg)
4420 .addUse(GR.getSPIRVTypeID(ResType));
4421 for (unsigned i = 0; i < N; ++i)
4422 MIB.addUse(OpReg);
4423 MIB.constrainAllUses(TII, TRI, RBI);
4424 return true;
4425}
4426
4427bool SPIRVInstructionSelector::selectConcatVectors(Register ResVReg,
4428 SPIRVTypeInst ResType,
4429 MachineInstr &I) const {
4430 // Implement G_CONCAT_VECTORS using OpCompositeConstruct, which allows vector
4431 // constituents that share the result's component type to be
4432 // concatenated in operand order.
4433 if (!isVectorType(ResType))
4434 report_fatal_error(
4435 "Cannot select G_CONCAT_VECTORS with a non-vector result");
4436
4437 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
4438 TII.get(SPIRV::OpCompositeConstruct))
4439 .addDef(ResVReg)
4440 .addUse(GR.getSPIRVTypeID(ResType));
4441 for (unsigned OpIdx = I.getNumExplicitDefs();
4442 OpIdx < I.getNumExplicitOperands(); ++OpIdx)
4443 MIB.addUse(I.getOperand(OpIdx).getReg());
4444 MIB.constrainAllUses(TII, TRI, RBI);
4445 return true;
4446}
4447
4448bool SPIRVInstructionSelector::selectDiscard(Register ResVReg,
4449 SPIRVTypeInst ResType,
4450 MachineInstr &I) const {
4451
4452 unsigned Opcode;
4453 MachineBasicBlock &BB = *I.getParent();
4454
4455 if (STI.canUseExtension(
4456 SPIRV::Extension::SPV_EXT_demote_to_helper_invocation) ||
4457 STI.isAtLeastSPIRVVer(llvm::VersionTuple(1, 6))) {
4458 Opcode = SPIRV::OpDemoteToHelperInvocation;
4459 } else {
4460 Opcode = SPIRV::OpKill;
4461 // OpKill must be the last operation of any basic block.
4462 for (MachineInstr &ToErase : make_early_inc_range(
4463 make_range(std::next(I.getIterator()), BB.instr_end()))) {
4464 GR.invalidateMachineInstr(&ToErase);
4465 ToErase.eraseFromParent();
4466 }
4467 }
4468
4469 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
4470 .constrainAllUses(TII, TRI, RBI);
4471 return true;
4472}
4473
4474bool SPIRVInstructionSelector::selectCmp(Register ResVReg,
4475 SPIRVTypeInst ResType, unsigned CmpOpc,
4476 MachineInstr &I) const {
4477 Register Cmp0 = I.getOperand(2).getReg();
4478 Register Cmp1 = I.getOperand(3).getReg();
4479 assert(GR.getSPIRVTypeForVReg(Cmp0)->getOpcode() ==
4480 GR.getSPIRVTypeForVReg(Cmp1)->getOpcode() &&
4481 "CMP operands should have the same type");
4482 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(CmpOpc))
4483 .addDef(ResVReg)
4484 .addUse(GR.getSPIRVTypeID(ResType))
4485 .addUse(Cmp0)
4486 .addUse(Cmp1)
4487 .setMIFlags(I.getFlags())
4488 .constrainAllUses(TII, TRI, RBI);
4489 return true;
4490}
4491
4492bool SPIRVInstructionSelector::selectICmp(Register ResVReg,
4493 SPIRVTypeInst ResType,
4494 MachineInstr &I) const {
4495 auto Pred = I.getOperand(1).getPredicate();
4496 unsigned CmpOpc;
4497
4498 Register CmpOperand = I.getOperand(2).getReg();
4499 SPIRVTypeInst CmpOperandType = GR.getSPIRVTypeForVReg(CmpOperand);
4500 bool IsPtrCmp = CmpOperandType && CmpOperandType.isPointer();
4501 if (IsPtrCmp) {
4502 CmpOpc = getPtrCmpOpcode(Pred);
4503 // OpPtrEqual/OpPtrNotEqual require both operands to share an identical
4504 // pointer type. If they are not OpBitcast is inserted.
4505 Register Op1 = I.getOperand(3).getReg();
4506 SPIRVTypeInst Ty0 = GR.getSPIRVTypeForVReg(CmpOperand);
4507 if (Ty0 != GR.getSPIRVTypeForVReg(Op1)) {
4508 Register NewOp1 = createVirtualRegister(Ty0, &GR, MRI, MRI->getMF());
4509 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpBitcast))
4510 .addDef(NewOp1)
4511 .addUse(GR.getSPIRVTypeID(Ty0))
4512 .addUse(Op1)
4513 .constrainAllUses(TII, TRI, RBI);
4514 I.getOperand(3).setReg(NewOp1);
4515 }
4516 } else if (GR.isScalarOrVectorOfType(CmpOperand, SPIRV::OpTypeBool))
4517 CmpOpc = getBoolCmpOpcode(Pred);
4518 else
4519 CmpOpc = getICmpOpcode(Pred);
4520 return selectCmp(ResVReg, ResType, CmpOpc, I);
4521}
4522
4523Register
4524SPIRVInstructionSelector::buildI32Constant(uint32_t Val, MachineInstr &I,
4525 SPIRVTypeInst ResType) const {
4526 Type *LLVMTy = IntegerType::get(GR.CurMF->getFunction().getContext(), 32);
4527 SPIRVTypeInst SpvI32Ty =
4528 ResType ? ResType : GR.getOrCreateSPIRVIntegerType(32, I, TII);
4529 // Find a constant in DT or build a new one.
4530 auto ConstInt = ConstantInt::get(LLVMTy, Val);
4531 Register NewReg = GR.find(ConstInt, GR.CurMF);
4532 if (!NewReg.isValid()) {
4533 NewReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
4534 MachineBasicBlock &BB = *I.getParent();
4535 MachineInstr *MI =
4536 Val == 0
4537 ? BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull))
4538 .addDef(NewReg)
4539 .addUse(GR.getSPIRVTypeID(SpvI32Ty))
4540 : BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantI))
4541 .addDef(NewReg)
4542 .addUse(GR.getSPIRVTypeID(SpvI32Ty))
4543 .addImm(APInt(32, Val).getZExtValue());
4544 constrainSelectedInstRegOperands(*MI, TII, TRI, RBI);
4545 GR.add(ConstInt, MI);
4546 }
4547 return NewReg;
4548}
4549
4550// Like buildI32Constant, but always inserts the constant definition in the
4551// entry block so it dominates all uses regardless of block ordering.
4552Register SPIRVInstructionSelector::buildI32ConstantInEntryBlock(
4553 uint32_t Val, MachineInstr &I, SPIRVTypeInst ResType) const {
4554 Type *LLVMTy = IntegerType::get(GR.CurMF->getFunction().getContext(), 32);
4555 SPIRVTypeInst SpvI32Ty =
4556 ResType ? ResType : GR.getOrCreateSPIRVIntegerType(32, I, TII);
4557 auto *ConstInt = ConstantInt::get(LLVMTy, Val);
4558 Register NewReg = GR.find(ConstInt, GR.CurMF);
4559 if (!NewReg.isValid()) {
4560 NewReg = MRI->createGenericVirtualRegister(LLT::scalar(64));
4561 auto InsertIt = getOpVariableMBBIt(*I.getMF());
4562 MachineBasicBlock &EntryBB = *InsertIt->getParent();
4563 MachineInstr *MI = nullptr;
4564 Register TypeReg = GR.getSPIRVTypeID(SpvI32Ty);
4565 DebugLoc DbgLoc = I.getDebugLoc();
4566 if (Val == 0) {
4567 MI = BuildMI(EntryBB, InsertIt, DbgLoc, TII.get(SPIRV::OpConstantNull))
4568 .addDef(NewReg)
4569 .addUse(TypeReg);
4570 } else {
4571 uint64_t ImmVal = APInt(32, Val).getZExtValue();
4572 MI = BuildMI(EntryBB, InsertIt, DbgLoc, TII.get(SPIRV::OpConstantI))
4573 .addDef(NewReg)
4574 .addUse(TypeReg)
4575 .addImm(ImmVal);
4576 }
4577 constrainSelectedInstRegOperands(*MI, TII, TRI, RBI);
4578 GR.add(ConstInt, MI);
4579 }
4580 return NewReg;
4581}
4582
4583bool SPIRVInstructionSelector::selectFCmp(Register ResVReg,
4584 SPIRVTypeInst ResType,
4585 MachineInstr &I) const {
4586 unsigned CmpOp = getFCmpOpcode(I.getOperand(1).getPredicate());
4587 return selectCmp(ResVReg, ResType, CmpOp, I);
4588}
4589
4590bool SPIRVInstructionSelector::selectExp10(Register ResVReg,
4591 SPIRVTypeInst ResType,
4592 MachineInstr &I) const {
4593 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std)) {
4594 return selectExtInst(ResVReg, ResType, I, CL::exp10);
4595 }
4596
4597 if (STI.canUseExtInstSet(SPIRV::InstructionSet::GLSL_std_450)) {
4598 /// There is no exp10 in GLSL. Use exp10(x) = exp2(x * log2(10)) instead
4599 /// log2(10) ~= 3.3219280948874l
4600
4601 if (!isVectorType(ResType) && !ResType.isAnyTypeFloat())
4602 return false;
4603
4604 MachineIRBuilder MIRBuilder(I);
4605
4606 SPIRVTypeInst SpirvScalarType = GR.getScalarOrVectorComponentType(ResType);
4607
4608 // Match the literal precision to the scalar type so the OpConstant
4609 // literal does not contain non-zero high-order bits that would fail
4610 // SPIR-V validation when the type is narrower than 32 bits (e.g. half).
4611 APFloat ConstVal(3.3219280948873623);
4612 bool LosesInfo;
4613 ConstVal.convert(
4614 getZeroFP(GR.getTypeForSPIRVType(SpirvScalarType)).getSemantics(),
4615 APFloat::rmNearestTiesToEven, &LosesInfo);
4616 Register ConstReg =
4617 GR.buildConstantFP(ConstVal, MIRBuilder, SpirvScalarType);
4618 Register ArgReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
4619 auto Opcode =
4620 isVectorType(ResType) ? SPIRV::OpVectorTimesScalar : SPIRV::OpFMulS;
4621
4622 if (!selectOpWithSrcs(ArgReg, ResType, I,
4623 {I.getOperand(1).getReg(), ConstReg}, Opcode))
4624 return false;
4625 if (!selectExtInst(ResVReg, ResType, I,
4626 {{SPIRV::InstructionSet::GLSL_std_450, GL::Exp2}}, false,
4627 false, {ArgReg}))
4628 return false;
4629
4630 return true;
4631 }
4632
4633 return false;
4634}
4635
4636Register SPIRVInstructionSelector::buildZerosVal(SPIRVTypeInst ResType,
4637 MachineInstr &I) const {
4638 // OpenCL uses nulls for Zero. In HLSL we don't use null constants.
4639 bool ZeroAsNull = !STI.isShader();
4640 if (isVectorType(ResType))
4641 return GR.getOrCreateConstVector(0UL, I, ResType, TII, ZeroAsNull);
4642 return GR.getOrCreateConstInt(0, I, ResType, TII, ZeroAsNull);
4643}
4644
4645bool SPIRVInstructionSelector::isScalarOrVectorIntConstantZero(
4646 Register Reg) const {
4647 SPIRVTypeInst Type = GR.getSPIRVTypeForVReg(Reg);
4648 if (!Type)
4649 return false;
4650 SPIRVTypeInst CompType = GR.getScalarOrVectorComponentType(Type);
4651 if (!CompType || CompType->getOpcode() != SPIRV::OpTypeInt)
4652 return false;
4653
4654 auto IsZero = [this](Register Reg) {
4655 MachineInstr *Def = getDefInstrMaybeConstant(Reg, MRI);
4656 if (!Def)
4657 return false;
4658
4659 if (Def->getOpcode() == SPIRV::OpConstantNull)
4660 return true;
4661
4662 if (Def->getOpcode() == TargetOpcode::G_CONSTANT ||
4663 Def->getOpcode() == SPIRV::OpConstantI)
4664 return getIConstVal(Reg, MRI) == 0;
4665
4666 return false;
4667 };
4668
4669 if (IsZero(Reg))
4670 return true;
4671
4672 MachineInstr *Def = MRI->getVRegDef(Reg);
4673 if (!Def)
4674 return false;
4675
4676 if (Def->getOpcode() == TargetOpcode::G_BUILD_VECTOR ||
4677 (Def->getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS &&
4678 cast<GIntrinsic>(Def)->getIntrinsicID() ==
4679 Intrinsic::spv_const_composite)) {
4680 unsigned StartOp = Def->getOpcode() == TargetOpcode::G_BUILD_VECTOR ? 1 : 2;
4681 for (unsigned i = StartOp; i < Def->getNumOperands(); ++i) {
4682 if (!IsZero(Def->getOperand(i).getReg()))
4683 return false;
4684 }
4685 return true;
4686 }
4687
4688 return false;
4689}
4690
4691Register SPIRVInstructionSelector::buildZerosValF(SPIRVTypeInst ResType,
4692 MachineInstr &I) const {
4693 // OpenCL uses nulls for Zero. In HLSL we don't use null constants.
4694 bool ZeroAsNull = !STI.isShader();
4695 APFloat VZero = getZeroFP(GR.getTypeForSPIRVType(ResType));
4696 if (isVectorType(ResType))
4697 return GR.getOrCreateConstVector(VZero, I, ResType, TII, ZeroAsNull);
4698 return GR.getOrCreateConstFP(VZero, I, ResType, TII, ZeroAsNull);
4699}
4700
4701Register SPIRVInstructionSelector::buildOnesValF(SPIRVTypeInst ResType,
4702 MachineInstr &I) const {
4703 // OpenCL uses nulls for Zero. In HLSL we don't use null constants.
4704 bool ZeroAsNull = !STI.isShader();
4705 APFloat VOne = getOneFP(GR.getTypeForSPIRVType(ResType));
4706 if (isVectorType(ResType))
4707 return GR.getOrCreateConstVector(VOne, I, ResType, TII, ZeroAsNull);
4708 return GR.getOrCreateConstFP(VOne, I, ResType, TII, ZeroAsNull);
4709}
4710
4711Register SPIRVInstructionSelector::buildOnesVal(bool AllOnes,
4712 SPIRVTypeInst ResType,
4713 MachineInstr &I) const {
4714 unsigned BitWidth = GR.getScalarOrVectorBitWidth(ResType);
4715 APInt One =
4716 AllOnes ? APInt::getAllOnes(BitWidth) : APInt::getOneBitSet(BitWidth, 0);
4717 if (isVectorType(ResType))
4718 return GR.getOrCreateConstVector(One, I, ResType, TII);
4719 return GR.getOrCreateConstInt(One, I, ResType, TII);
4720}
4721
4722bool SPIRVInstructionSelector::selectSelect(Register ResVReg,
4723 SPIRVTypeInst ResType,
4724 MachineInstr &I) const {
4725 Register SelectFirstArg = I.getOperand(2).getReg();
4726 Register SelectSecondArg = I.getOperand(3).getReg();
4727 assert(ResType == GR.getSPIRVTypeForVReg(SelectFirstArg) &&
4728 ResType == GR.getSPIRVTypeForVReg(SelectSecondArg));
4729
4730 bool IsFloatTy =
4731 GR.isScalarOrVectorOfType(SelectFirstArg, SPIRV::OpTypeFloat);
4732 bool IsPtrTy =
4733 GR.isScalarOrVectorOfType(SelectFirstArg, SPIRV::OpTypePointer);
4734
4735 bool IsScalarBool =
4736 GR.isScalarOfType(I.getOperand(1).getReg(), SPIRV::OpTypeBool);
4737 unsigned Opcode;
4738 if (isVectorType(GR.getSPIRVTypeForVReg(SelectFirstArg))) {
4739 if (IsFloatTy) {
4740 Opcode = IsScalarBool ? SPIRV::OpSelectVFSCond : SPIRV::OpSelectVFVCond;
4741 } else if (IsPtrTy) {
4742 Opcode = IsScalarBool ? SPIRV::OpSelectVPSCond : SPIRV::OpSelectVPVCond;
4743 } else {
4744 Opcode = IsScalarBool ? SPIRV::OpSelectVISCond : SPIRV::OpSelectVIVCond;
4745 }
4746 } else {
4747 assert(IsScalarBool && "OpSelect with a scalar result requires a scalar "
4748 "boolean condition");
4749 if (IsFloatTy) {
4750 Opcode = SPIRV::OpSelectSFSCond;
4751 } else if (IsPtrTy) {
4752 Opcode = SPIRV::OpSelectSPSCond;
4753 } else {
4754 Opcode = SPIRV::OpSelectSISCond;
4755 }
4756 }
4757 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
4758 .addDef(ResVReg)
4759 .addUse(GR.getSPIRVTypeID(ResType))
4760 .addUse(I.getOperand(1).getReg())
4761 .addUse(SelectFirstArg)
4762 .addUse(SelectSecondArg)
4763 .constrainAllUses(TII, TRI, RBI);
4764 return true;
4765}
4766
4767// This function is used to extend a bool or a vector of bools into an integer
4768// or vector of integers.
4769bool SPIRVInstructionSelector::selectBoolToInt(Register ResVReg,
4770 SPIRVTypeInst ResType,
4771 Register BooleanVReg,
4772 MachineInstr &InsertAt,
4773 bool IsSigned) const {
4774 // To extend a bool, we need to use OpSelect between constants.
4775 Register ZeroReg = buildZerosVal(ResType, InsertAt);
4776 Register OneReg = buildOnesVal(IsSigned, ResType, InsertAt);
4777 bool IsScalarBool = GR.isScalarOfType(BooleanVReg, SPIRV::OpTypeBool);
4778 unsigned Opcode =
4779 IsScalarBool ? SPIRV::OpSelectSISCond : SPIRV::OpSelectVIVCond;
4780 BuildMI(*InsertAt.getParent(), InsertAt, InsertAt.getDebugLoc(),
4781 TII.get(Opcode))
4782 .addDef(ResVReg)
4783 .addUse(GR.getSPIRVTypeID(ResType))
4784 .addUse(BooleanVReg)
4785 .addUse(OneReg)
4786 .addUse(ZeroReg)
4787 .constrainAllUses(TII, TRI, RBI);
4788 return true;
4789}
4790
4791bool SPIRVInstructionSelector::selectIToF(Register ResVReg,
4792 SPIRVTypeInst ResType,
4793 MachineInstr &I, bool IsSigned,
4794 unsigned Opcode) const {
4795 Register SrcReg = I.getOperand(1).getReg();
4796 // We can convert bool value directly to float type without OpConvert*ToF,
4797 // however the translator generates OpSelect+OpConvert*ToF, so we do the same.
4798 if (GR.isScalarOrVectorOfType(I.getOperand(1).getReg(), SPIRV::OpTypeBool)) {
4799 unsigned BitWidth = GR.getScalarOrVectorBitWidth(ResType);
4800 SPIRVTypeInst TmpType = GR.getOrCreateSPIRVIntegerType(BitWidth, I, TII);
4801 if (isVectorType(ResType)) {
4802 const unsigned NumElts = GR.getScalarOrVectorComponentCount(ResType);
4803 TmpType = GR.getOrCreateSPIRVVectorType(TmpType, NumElts, I, TII);
4804 }
4805 SrcReg = createVirtualRegister(TmpType, &GR, MRI, MRI->getMF());
4806 selectBoolToInt(SrcReg, TmpType, I.getOperand(1).getReg(), I, IsSigned);
4807 }
4808 return selectOpWithSrcs(ResVReg, ResType, I, {SrcReg}, Opcode);
4809}
4810
4811bool SPIRVInstructionSelector::selectExt(Register ResVReg,
4812 SPIRVTypeInst ResType, MachineInstr &I,
4813 bool IsSigned) const {
4814 Register SrcReg = I.getOperand(1).getReg();
4815 if (GR.isScalarOrVectorOfType(SrcReg, SPIRV::OpTypeBool))
4816 return selectBoolToInt(ResVReg, ResType, I.getOperand(1).getReg(), I,
4817 IsSigned);
4818
4819 SPIRVTypeInst SrcType = GR.getSPIRVTypeForVReg(SrcReg);
4820 if (ResType == SrcType)
4821 return BuildCOPY(ResVReg, SrcReg, I);
4822
4823 unsigned Opcode = IsSigned ? SPIRV::OpSConvert : SPIRV::OpUConvert;
4824 return selectUnOp(ResVReg, ResType, I, Opcode);
4825}
4826
4827bool SPIRVInstructionSelector::selectSUCmp(Register ResVReg,
4828 SPIRVTypeInst ResType,
4829 MachineInstr &I,
4830 bool IsSigned) const {
4831 MachineIRBuilder MIRBuilder(I);
4832 MachineRegisterInfo *MRI = MIRBuilder.getMRI();
4833 MachineBasicBlock &BB = *I.getParent();
4834 // Ensure we have bool.
4835 SPIRVTypeInst BoolType = GR.getOrCreateSPIRVBoolType(I, TII);
4836 unsigned N = GR.getScalarOrVectorComponentCount(ResType);
4837 if (N > 1 || (ResType->getOpcode() == SPIRV::OpTypeVectorIdEXT &&
4838 STI.canUseExtension(SPIRV::Extension::SPV_EXT_long_vector)))
4839 BoolType = GR.getOrCreateSPIRVVectorType(BoolType, N, I, TII);
4840 Register BoolTypeReg = GR.getSPIRVTypeID(BoolType);
4841 // Build less-than-equal and less-than.
4842 Register IsLessEqReg =
4843 createVirtualRegister(BoolType, &GR, MRI, MIRBuilder.getMF());
4844 BuildMI(BB, I, I.getDebugLoc(),
4845 TII.get(IsSigned ? SPIRV::OpSLessThanEqual : SPIRV::OpULessThanEqual))
4846 .addDef(IsLessEqReg)
4847 .addUse(BoolTypeReg)
4848 .addUse(I.getOperand(1).getReg())
4849 .addUse(I.getOperand(2).getReg())
4850 .constrainAllUses(TII, TRI, RBI);
4851 Register IsLessReg =
4852 createVirtualRegister(BoolType, &GR, MRI, MIRBuilder.getMF());
4853 BuildMI(BB, I, I.getDebugLoc(),
4854 TII.get(IsSigned ? SPIRV::OpSLessThan : SPIRV::OpULessThan))
4855 .addDef(IsLessReg)
4856 .addUse(BoolTypeReg)
4857 .addUse(I.getOperand(1).getReg())
4858 .addUse(I.getOperand(2).getReg())
4859 .constrainAllUses(TII, TRI, RBI);
4860 // Build selects.
4861 Register ResTypeReg = GR.getSPIRVTypeID(ResType);
4862 Register NegOneOrZeroReg =
4863 MRI->createVirtualRegister(GR.getRegClass(ResType));
4864 MRI->setType(NegOneOrZeroReg, LLT::scalar(64));
4865 GR.assignSPIRVTypeToVReg(ResType, NegOneOrZeroReg, MIRBuilder.getMF());
4866 unsigned SelectOpcode =
4867 (N > 1 || (ResType->getOpcode() == SPIRV::OpTypeVectorIdEXT &&
4868 STI.canUseExtension(SPIRV::Extension::SPV_EXT_long_vector)))
4869 ? SPIRV::OpSelectVIVCond
4870 : SPIRV::OpSelectSISCond;
4871 BuildMI(BB, I, I.getDebugLoc(), TII.get(SelectOpcode))
4872 .addDef(NegOneOrZeroReg)
4873 .addUse(ResTypeReg)
4874 .addUse(IsLessReg)
4875 .addUse(buildOnesVal(true, ResType, I)) // -1
4876 .addUse(buildZerosVal(ResType, I))
4877 .constrainAllUses(TII, TRI, RBI);
4878 BuildMI(BB, I, I.getDebugLoc(), TII.get(SelectOpcode))
4879 .addDef(ResVReg)
4880 .addUse(ResTypeReg)
4881 .addUse(IsLessEqReg)
4882 .addUse(NegOneOrZeroReg) // -1 or 0
4883 .addUse(buildOnesVal(false, ResType, I))
4884 .constrainAllUses(TII, TRI, RBI);
4885 return true;
4886}
4887
4888bool SPIRVInstructionSelector::selectIntToBool(Register IntReg,
4889 Register ResVReg,
4890 MachineInstr &I,
4891 SPIRVTypeInst IntTy,
4892 SPIRVTypeInst BoolTy) const {
4893 // To truncate to a bool, we use OpBitwiseAnd 1 and OpINotEqual to zero.
4894 Register BitIntReg = createVirtualRegister(IntTy, &GR, MRI, MRI->getMF());
4895 unsigned Opcode =
4896 isVectorType(IntTy) ? SPIRV::OpBitwiseAndV : SPIRV::OpBitwiseAndS;
4897 Register Zero = buildZerosVal(IntTy, I);
4898 Register One = buildOnesVal(false, IntTy, I);
4899 MachineBasicBlock &BB = *I.getParent();
4900 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
4901 .addDef(BitIntReg)
4902 .addUse(GR.getSPIRVTypeID(IntTy))
4903 .addUse(IntReg)
4904 .addUse(One)
4905 .constrainAllUses(TII, TRI, RBI);
4906 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpINotEqual))
4907 .addDef(ResVReg)
4908 .addUse(GR.getSPIRVTypeID(BoolTy))
4909 .addUse(BitIntReg)
4910 .addUse(Zero)
4911 .constrainAllUses(TII, TRI, RBI);
4912 return true;
4913}
4914
4915bool SPIRVInstructionSelector::selectTrunc(Register ResVReg,
4916 SPIRVTypeInst ResType,
4917 MachineInstr &I) const {
4918 Register IntReg = I.getOperand(1).getReg();
4919 const SPIRVTypeInst ArgType = GR.getSPIRVTypeForVReg(IntReg);
4920 if (GR.isScalarOrVectorOfType(ResVReg, SPIRV::OpTypeBool))
4921 return selectIntToBool(IntReg, ResVReg, I, ArgType, ResType);
4922 if (ArgType == ResType)
4923 return BuildCOPY(ResVReg, IntReg, I);
4924 bool IsSigned = GR.isScalarOrVectorSigned(ResType);
4925 unsigned Opcode = IsSigned ? SPIRV::OpSConvert : SPIRV::OpUConvert;
4926 return selectUnOp(ResVReg, ResType, I, Opcode);
4927}
4928
4929bool SPIRVInstructionSelector::selectConst(Register ResVReg,
4930 SPIRVTypeInst ResType,
4931 MachineInstr &I) const {
4932 unsigned Opcode = I.getOpcode();
4933 unsigned TpOpcode = ResType->getOpcode();
4934 Register Reg;
4935 if (ResType.isPointer() || TpOpcode == SPIRV::OpTypeEvent) {
4936 assert(Opcode == TargetOpcode::G_CONSTANT &&
4937 I.getOperand(1).getCImm()->isZero());
4938 MachineBasicBlock &DepMBB = I.getMF()->front();
4939 MachineIRBuilder MIRBuilder(DepMBB, DepMBB.getFirstNonPHI());
4940 Reg = GR.getOrCreateConstNullPtr(MIRBuilder, ResType);
4941 } else if (TpOpcode == SPIRV::OpTypeVectorIdEXT) {
4942 // We ended up here coming from a splat on a <1 x T> type, which
4943 // IRTranslator translated into a scalar, so we have to restore the
4944 // vectorness.
4945 assert(GR.getScalarOrVectorComponentCount(ResType) == 1 &&
4946 "Expected <1 x T> Vector!");
4947 if (Opcode == TargetOpcode::G_FCONSTANT)
4948 Reg = GR.getOrCreateConstVector(I.getOperand(1).getFPImm()->getValue(), I,
4949 ResType, TII);
4950 else // We handle vector of pointer here as well.
4951 Reg = GR.getOrCreateConstVector(I.getOperand(1).getCImm()->getValue(), I,
4952 ResType, TII);
4953 } else if (Opcode == TargetOpcode::G_FCONSTANT) {
4954 Reg = GR.getOrCreateConstFP(I.getOperand(1).getFPImm()->getValue(), I,
4955 ResType, TII, !STI.isShader());
4956
4957 } else {
4958 Reg = GR.getOrCreateConstInt(I.getOperand(1).getCImm()->getValue(), I,
4959 ResType, TII, !STI.isShader());
4960 }
4961 return Reg == ResVReg ? true : BuildCOPY(ResVReg, Reg, I);
4962}
4963
4964bool SPIRVInstructionSelector::selectOpUndef(Register ResVReg,
4965 SPIRVTypeInst ResType,
4966 MachineInstr &I) const {
4967 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpUndef))
4968 .addDef(ResVReg)
4969 .addUse(GR.getSPIRVTypeID(ResType))
4970 .constrainAllUses(TII, TRI, RBI);
4971 return true;
4972}
4973
4974bool SPIRVInstructionSelector::selectInsertVal(Register ResVReg,
4975 SPIRVTypeInst ResType,
4976 MachineInstr &I) const {
4977 MachineBasicBlock &BB = *I.getParent();
4978 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeInsert))
4979 .addDef(ResVReg)
4980 .addUse(GR.getSPIRVTypeID(ResType))
4981 // object to insert
4982 .addUse(I.getOperand(3).getReg())
4983 // composite to insert into
4984 .addUse(I.getOperand(2).getReg());
4985 for (unsigned i = 4; i < I.getNumOperands(); i++)
4986 MIB.addImm(foldImm(I.getOperand(i), MRI));
4987 MIB.constrainAllUses(TII, TRI, RBI);
4988 return true;
4989}
4990
4991bool SPIRVInstructionSelector::selectExtractVal(Register ResVReg,
4992 SPIRVTypeInst ResType,
4993 MachineInstr &I) const {
4994 Type *MaybeResTy = nullptr;
4995 StringRef ResName;
4996 if (GR.findValueAttrs(&I, MaybeResTy, ResName) &&
4997 MaybeResTy != GR.getTypeForSPIRVType(ResType)) {
4998 assert((!MaybeResTy || MaybeResTy->isAggregateType()) &&
4999 "Expected aggregate type for extractv instruction");
5000 ResType = GR.getOrCreateSPIRVType(MaybeResTy, I,
5001 SPIRV::AccessQualifier::ReadWrite, false);
5002 GR.assignSPIRVTypeToVReg(ResType, ResVReg, *I.getMF());
5003 }
5004 MachineBasicBlock &BB = *I.getParent();
5005 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
5006 .addDef(ResVReg)
5007 .addUse(GR.getSPIRVTypeID(ResType))
5008 .addUse(I.getOperand(2).getReg());
5009 for (unsigned i = 3; i < I.getNumOperands(); i++)
5010 MIB.addImm(foldImm(I.getOperand(i), MRI));
5011 MIB.constrainAllUses(TII, TRI, RBI);
5012 return true;
5013}
5014
5015bool SPIRVInstructionSelector::selectInsertElt(Register ResVReg,
5016 SPIRVTypeInst ResType,
5017 MachineInstr &I) const {
5018 if (getImm(I.getOperand(4), MRI))
5019 return selectInsertVal(ResVReg, ResType, I);
5020 MachineBasicBlock &BB = *I.getParent();
5021 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorInsertDynamic))
5022 .addDef(ResVReg)
5023 .addUse(GR.getSPIRVTypeID(ResType))
5024 .addUse(I.getOperand(2).getReg())
5025 .addUse(I.getOperand(3).getReg())
5026 .addUse(I.getOperand(4).getReg())
5027 .constrainAllUses(TII, TRI, RBI);
5028 return true;
5029}
5030
5031bool SPIRVInstructionSelector::selectExtractElt(Register ResVReg,
5032 SPIRVTypeInst ResType,
5033 MachineInstr &I) const {
5034 if (getImm(I.getOperand(3), MRI))
5035 return selectExtractVal(ResVReg, ResType, I);
5036 MachineBasicBlock &BB = *I.getParent();
5037 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVectorExtractDynamic))
5038 .addDef(ResVReg)
5039 .addUse(GR.getSPIRVTypeID(ResType))
5040 .addUse(I.getOperand(2).getReg())
5041 .addUse(I.getOperand(3).getReg())
5042 .constrainAllUses(TII, TRI, RBI);
5043 return true;
5044}
5045
5046bool SPIRVInstructionSelector::selectGEP(Register ResVReg,
5047 SPIRVTypeInst ResType,
5048 MachineInstr &I) const {
5049 const bool IsGEPInBounds = I.getOperand(2).getImm();
5050 // Pointers to opaque types stay typed even with the extension on, so emit the
5051 // untyped variant only when the result is actually an untyped pointer.
5052 const bool UseUntypedPointers =
5053 ResType->getOpcode() == SPIRV::OpTypeUntypedPointerKHR;
5054
5055 // Determine the opcode based on pointer type and bounds checking.
5056 // When using untyped pointers, use OpUntyped*AccessChainKHR variants.
5057 unsigned Opcode;
5058 if (UseUntypedPointers) {
5059 if (STI.isLogicalSPIRV()) {
5060 Opcode = IsGEPInBounds ? SPIRV::OpUntypedInBoundsAccessChainKHR
5061 : SPIRV::OpUntypedAccessChainKHR;
5062 } else {
5063 Opcode = IsGEPInBounds ? SPIRV::OpUntypedInBoundsPtrAccessChainKHR
5064 : SPIRV::OpUntypedPtrAccessChainKHR;
5065 }
5066 } else {
5067 // OpAccessChain could be used for OpenCL, but the SPIRV-LLVM Translator
5068 // only relies on PtrAccessChain, so we'll try not to deviate. For Vulkan
5069 // however, we have to use Op[InBounds]AccessChain.
5070 // FIXME: fix llvm-spirv.
5071 if (STI.isLogicalSPIRV()) {
5072 Opcode =
5073 IsGEPInBounds ? SPIRV::OpInBoundsAccessChain : SPIRV::OpAccessChain;
5074 } else {
5075 Opcode = IsGEPInBounds ? SPIRV::OpInBoundsPtrAccessChain
5076 : SPIRV::OpPtrAccessChain;
5077 }
5078 }
5079
5080 Register BaseReg = I.getOperand(3).getReg();
5081 auto Res = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
5082 .addDef(ResVReg)
5083 .addUse(GR.getSPIRVTypeID(ResType));
5084
5085 // For untyped access chains, we need to add the base type operand.
5086 if (UseUntypedPointers) {
5087 // Get the element type from the base pointer register.
5088 // For untyped pointers, this was stored when processing
5089 // spv_assign_ptr_type.
5090 SPIRVTypeInst BaseType = GR.getUntypedPtrElementType(BaseReg);
5091 if (!BaseType) {
5092 // Otherwise try the pointee type for mixed typed-pointer usage.
5093 SPIRVTypeInst BasePtrType = GR.getSPIRVTypeForVReg(BaseReg);
5094 BaseType = BasePtrType ? GR.getPointeeType(BasePtrType) : nullptr;
5095 }
5096 if (!BaseType) {
5097 // The base may be a not-yet-selected global. Read its value type from
5098 // the defining G_GLOBAL_VALUE, following copies.
5099 Register DefReg = BaseReg;
5100 MachineInstr *Def = MRI->getVRegDef(DefReg);
5101 while (Def && Def->getOpcode() == TargetOpcode::COPY &&
5102 Def->getOperand(1).isReg())
5103 Def = MRI->getVRegDef(Def->getOperand(1).getReg());
5104 if (Def && Def->getOpcode() == TargetOpcode::G_GLOBAL_VALUE)
5105 if (const auto *GVar =
5106 dyn_cast<GlobalVariable>(Def->getOperand(1).getGlobal()))
5107 BaseType = GR.getOrCreateSPIRVType(GVar->getValueType(), I,
5108 SPIRV::AccessQualifier::ReadWrite,
5109 /*EmitIR=*/false);
5110 }
5111 if (!BaseType)
5112 return diagnoseUnsupported(
5113 I, "could not deduce the base type of an untyped access chain");
5114 Res.addUse(GR.getSPIRVTypeID(BaseType));
5115 }
5116
5117 // Object to get a pointer to.
5118 Res.addUse(BaseReg);
5119
5120 const bool IsAccessChainOpcode =
5121 (Opcode == SPIRV::OpAccessChain ||
5122 Opcode == SPIRV::OpInBoundsAccessChain ||
5123 Opcode == SPIRV::OpUntypedAccessChainKHR ||
5124 Opcode == SPIRV::OpUntypedInBoundsAccessChainKHR);
5125
5126 assert((!IsAccessChainOpcode || (getImm(I.getOperand(4), MRI) &&
5127 foldImm(I.getOperand(4), MRI) == 0)) &&
5128 "Cannot translate GEP to OpAccessChain.");
5129
5130 // Adding indices.
5131 const unsigned StartingIndex = IsAccessChainOpcode ? 5 : 4;
5132 for (unsigned i = StartingIndex; i < I.getNumExplicitOperands(); ++i)
5133 Res.addUse(I.getOperand(i).getReg());
5134 Res.constrainAllUses(TII, TRI, RBI);
5135
5136 // IRTranslator doesn't like <1 x T> vectors, and treats them as scalars. This
5137 // creates an information loss issue, and also broken code where a
5138 // OpCompositeExtract is applied to the scalar return of an OpPtrAccessChain,
5139 // when it was meant to be applied to the OpCompositeInsert created <1 x T>.
5140 // TODO: should we also re-create the index <1 x T>?
5141 if (MRI->hasOneUse(ResVReg)) {
5142 MachineInstr &Extract = *MRI->use_instr_begin(ResVReg);
5143 if (Extract.getOpcode() == SPIRV::OpCompositeExtract) {
5144 SPIRVTypeInst V = GR.getOrCreateSPIRVVectorType(ResType, 1, Extract, TII);
5145 // We cannot use GlobalRegistry::getOrCreateUndef directly here because it
5146 // tries to use UndefVal::get, which does not work for TypedPointerType,
5147 // which we can get here if we're dealing with <1 x T*>.
5148 Register Tmp = createVirtualRegister(V, &GR, MRI, *I.getMF());
5149 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpUndef))
5150 .addDef(Tmp)
5151 .addUse(GR.getSPIRVTypeID(V))
5152 .constrainAllUses(TII, TRI, RBI);
5153 Register InsertReg = createVirtualRegister(V, &GR, MRI, *I.getMF());
5154 BuildMI(*I.getParent(), Extract, Extract.getDebugLoc(),
5155 TII.get(SPIRV::OpCompositeInsert))
5156 .addDef(InsertReg)
5157 .addUse(GR.getSPIRVTypeID(V))
5158 .addUse(ResVReg)
5159 .addUse(Tmp)
5160 .addImm(0)
5161 .constrainAllUses(TII, TRI, RBI);
5162 Extract.substituteRegister(ResVReg, InsertReg, 0, TRI);
5163 }
5164 }
5165 return true;
5166}
5167
5168// Maybe wrap a value into OpSpecConstantOp
5169bool SPIRVInstructionSelector::wrapIntoSpecConstantOp(
5170 MachineInstr &I, SmallVector<Register> &CompositeArgs) const {
5171 unsigned Lim = I.getNumExplicitOperands();
5172 for (unsigned i = I.getNumExplicitDefs() + 1; i < Lim; ++i) {
5173 Register OpReg = I.getOperand(i).getReg();
5174 MachineInstr *OpDefine = MRI->getVRegDef(OpReg);
5175 SPIRVTypeInst OpType = GR.getSPIRVTypeForVReg(OpReg);
5176 if (!OpDefine || !OpType || isConstReg(MRI, OpDefine) ||
5177 OpDefine->getOpcode() == TargetOpcode::G_ADDRSPACE_CAST ||
5178 OpDefine->getOpcode() == TargetOpcode::G_INTTOPTR ||
5179 GR.isAggregateType(OpType)) {
5180 // The case of G_ADDRSPACE_CAST inside spv_const_composite() is processed
5181 // by selectAddrSpaceCast(), and G_INTTOPTR is processed by selectUnOp()
5182 CompositeArgs.push_back(OpReg);
5183 continue;
5184 }
5185 MachineFunction *MF = I.getMF();
5186 Register WrapReg = GR.find(OpDefine, MF);
5187 if (WrapReg.isValid()) {
5188 CompositeArgs.push_back(WrapReg);
5189 continue;
5190 }
5191 SPIRVTypeInst WrapType = OpType;
5192 if (OpType->getOpcode() == SPIRV::OpTypePointer &&
5193 GR.getPointerStorageClass(OpType) ==
5194 SPIRV::StorageClass::CodeSectionINTEL) {
5195 WrapType = GR.changePointerStorageClass(OpType,
5196 SPIRV::StorageClass::Function, I);
5197 }
5198 WrapReg = MRI->createVirtualRegister(GR.getRegClass(WrapType));
5199 CompositeArgs.push_back(WrapReg);
5200 // Decorate the wrapper register and generate a new instruction
5201 MRI->setType(WrapReg, LLT::pointer(0, 64));
5202 GR.assignSPIRVTypeToVReg(WrapType, WrapReg, *MF);
5203 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
5204 TII.get(SPIRV::OpSpecConstantOp))
5205 .addDef(WrapReg)
5206 .addUse(GR.getSPIRVTypeID(WrapType))
5207 .addImm(static_cast<uint32_t>(SPIRV::Opcode::Bitcast))
5208 .addUse(OpReg);
5209 GR.add(OpDefine, MIB);
5210 MIB.constrainAllUses(TII, TRI, RBI);
5211 }
5212 return true;
5213}
5214
5215bool SPIRVInstructionSelector::selectDerivativeInst(
5216 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
5217 const unsigned DPdOpCode) const {
5218 // TODO: This should check specifically for Fragment Execution Model, but STI
5219 // doesn't provide that information yet. See #167562
5220 if (!errorIfInstrOutsideShader(I))
5221 return false;
5222
5223 // If the arg/result types are half then we need to wrap the instr in
5224 // conversions to float
5225 // This case occurs because a half arg/result is legal in HLSL but not spirv.
5226 Register SrcReg = I.getOperand(2).getReg();
5227 SPIRVTypeInst SrcType = GR.getSPIRVTypeForVReg(SrcReg);
5228 unsigned BitWidth = std::min(GR.getScalarOrVectorBitWidth(SrcType),
5229 GR.getScalarOrVectorBitWidth(ResType));
5230 if (BitWidth == 32)
5231 return BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(DPdOpCode))
5232 .addDef(ResVReg)
5233 .addUse(GR.getSPIRVTypeID(ResType))
5234 .addUse(I.getOperand(2).getReg());
5235
5236 MachineIRBuilder MIRBuilder(I);
5237 unsigned componentCount = GR.getScalarOrVectorComponentCount(SrcType);
5238 SPIRVTypeInst F32ConvertTy = GR.getOrCreateSPIRVFloatType(32, I, TII);
5239 if (componentCount != 1)
5240 F32ConvertTy = GR.getOrCreateSPIRVVectorType(F32ConvertTy, componentCount,
5241 MIRBuilder, false);
5242
5243 const TargetRegisterClass *RegClass = GR.getRegClass(SrcType);
5244 Register ConvertToVReg = MRI->createVirtualRegister(RegClass);
5245 Register DpdOpVReg = MRI->createVirtualRegister(RegClass);
5246
5247 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpFConvert))
5248 .addDef(ConvertToVReg)
5249 .addUse(GR.getSPIRVTypeID(F32ConvertTy))
5250 .addUse(SrcReg)
5251 .constrainAllUses(TII, TRI, RBI);
5252 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(DPdOpCode))
5253 .addDef(DpdOpVReg)
5254 .addUse(GR.getSPIRVTypeID(F32ConvertTy))
5255 .addUse(ConvertToVReg)
5256 .constrainAllUses(TII, TRI, RBI);
5257 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpFConvert))
5258 .addDef(ResVReg)
5259 .addUse(GR.getSPIRVTypeID(ResType))
5260 .addUse(DpdOpVReg)
5261 .constrainAllUses(TII, TRI, RBI);
5262 return true;
5263}
5264
5265bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
5266 SPIRVTypeInst ResType,
5267 MachineInstr &I) const {
5268 MachineBasicBlock &BB = *I.getParent();
5269 Intrinsic::ID IID = cast<GIntrinsic>(I).getIntrinsicID();
5270 switch (IID) {
5271 case Intrinsic::spv_load:
5272 return selectLoad(ResVReg, ResType, I);
5273 case Intrinsic::spv_atomic_load:
5274 return selectAtomicLoad(ResVReg, ResType, I);
5275 case Intrinsic::spv_store:
5276 return selectStore(I);
5277 case Intrinsic::spv_atomic_store:
5278 return selectAtomicStore(I);
5279 case Intrinsic::spv_extractv:
5280 return selectExtractVal(ResVReg, ResType, I);
5281 case Intrinsic::spv_insertv:
5282 return selectInsertVal(ResVReg, ResType, I);
5283 case Intrinsic::spv_extractelt:
5284 return selectExtractElt(ResVReg, ResType, I);
5285 case Intrinsic::spv_insertelt:
5286 return selectInsertElt(ResVReg, ResType, I);
5287 case Intrinsic::spv_gep:
5288 return selectGEP(ResVReg, ResType, I);
5289 case Intrinsic::spv_bitcast: {
5290 Register OpReg = I.getOperand(2).getReg();
5291 SPIRVTypeInst OpType =
5292 OpReg.isValid() ? GR.getSPIRVTypeForVReg(OpReg) : nullptr;
5293 if (!GR.isBitcastCompatible(ResType, OpType))
5294 report_fatal_error("incompatible result and operand types in a bitcast");
5295 return selectOpWithSrcs(ResVReg, ResType, I, {OpReg}, SPIRV::OpBitcast);
5296 }
5297 case Intrinsic::spv_unref_global:
5298 case Intrinsic::spv_init_global: {
5299 MachineInstr *MI = MRI->getVRegDef(I.getOperand(1).getReg());
5300 MachineInstr *Init = I.getNumExplicitOperands() > 2
5301 ? MRI->getVRegDef(I.getOperand(2).getReg())
5302 : nullptr;
5303 assert(MI);
5304 Register GVarVReg = MI->getOperand(0).getReg();
5305 if (!selectGlobalValue(GVarVReg, *MI, Init))
5306 return false;
5307 // We violate SSA form by inserting OpVariable and still having a gMIR
5308 // instruction %vreg = G_GLOBAL_VALUE @gvar. We need to fix this by erasing
5309 // the duplicated definition.
5310 if (MI->getOpcode() == TargetOpcode::G_GLOBAL_VALUE) {
5311 GR.invalidateMachineInstr(MI);
5312 MI->eraseFromParent();
5313 }
5314 return true;
5315 }
5316 case Intrinsic::spv_undef: {
5317 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpUndef))
5318 .addDef(ResVReg)
5319 .addUse(GR.getSPIRVTypeID(ResType));
5320 MIB.constrainAllUses(TII, TRI, RBI);
5321 return true;
5322 }
5323 case Intrinsic::spv_poison:
5324 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpPoisonKHR))
5325 .addDef(ResVReg)
5326 .addUse(GR.getSPIRVTypeID(ResType))
5327 .constrainAllUses(TII, TRI, RBI);
5328 return true;
5329 case Intrinsic::spv_freeze:
5330 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpFreezeKHR))
5331 .addDef(ResVReg)
5332 .addUse(GR.getSPIRVTypeID(ResType))
5333 .addUse(I.getOperand(2).getReg())
5334 .constrainAllUses(TII, TRI, RBI);
5335 return true;
5336 case Intrinsic::spv_named_boolean_spec_constant: {
5337 auto Opcode = I.getOperand(3).getImm() ? SPIRV::OpSpecConstantTrue
5338 : SPIRV::OpSpecConstantFalse;
5339
5340 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
5341 .addDef(I.getOperand(0).getReg())
5342 .addUse(GR.getSPIRVTypeID(ResType));
5343 MIB.constrainAllUses(TII, TRI, RBI);
5344 unsigned SpecId = I.getOperand(2).getImm();
5345 buildOpDecorate(I.getOperand(0).getReg(), *++MIB->getIterator(), TII,
5346 SPIRV::Decoration::SpecId, {SpecId});
5347
5348 return true;
5349 }
5350 case Intrinsic::spv_const_composite: {
5351 // If no values are attached, the composite is null constant.
5352 bool IsNull = I.getNumExplicitDefs() + 1 == I.getNumExplicitOperands();
5353 SmallVector<Register> CompositeArgs;
5354 MRI->setRegClass(ResVReg, GR.getRegClass(ResType));
5355
5356 // skip type MD node we already used when generated assign.type for this
5357 if (!IsNull) {
5358 if (!wrapIntoSpecConstantOp(I, CompositeArgs))
5359 return false;
5360 std::function<bool(Register)> HasSpecConstOperand =
5361 [&](Register Reg) -> bool {
5362 MachineInstr *Def = MRI->getVRegDef(Reg);
5363 if (!Def)
5364 return false;
5365 if (!isConstReg(MRI, Def))
5366 return true;
5367 // Recurse into not-yet-selected spv_const_composite intrinsics
5368 // to detect transitive spec constant operands.
5369 if (isSpvIntrinsic(*Def, Intrinsic::spv_const_composite)) {
5370 for (unsigned J = Def->getNumExplicitDefs() + 1;
5371 J < Def->getNumExplicitOperands(); ++J) {
5372 if (Def->getOperand(J).isReg() &&
5373 HasSpecConstOperand(Def->getOperand(J).getReg()))
5374 return true;
5375 }
5376 }
5377 return false;
5378 };
5379 bool HasSpecConst = llvm::any_of(CompositeArgs, HasSpecConstOperand);
5380 unsigned CompositeOpc = HasSpecConst ? SPIRV::OpSpecConstantComposite
5381 : SPIRV::OpConstantComposite;
5382 unsigned ContinuedOpc = HasSpecConst
5383 ? SPIRV::OpSpecConstantCompositeContinuedINTEL
5384 : SPIRV::OpConstantCompositeContinuedINTEL;
5385 MachineIRBuilder MIR(I);
5386 SmallVector<MachineInstr *, 4> Instructions = createContinuedInstructions(
5387 MIR, CompositeOpc, 3, ContinuedOpc, CompositeArgs, ResVReg,
5388 GR.getSPIRVTypeID(ResType));
5389 for (auto *Instr : Instructions) {
5390 Instr->setDebugLoc(I.getDebugLoc());
5391 constrainSelectedInstRegOperands(*Instr, TII, TRI, RBI);
5392 }
5393 return true;
5394 } else {
5395 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull))
5396 .addDef(ResVReg)
5397 .addUse(GR.getSPIRVTypeID(ResType));
5398 MIB.constrainAllUses(TII, TRI, RBI);
5399 return true;
5400 }
5401 }
5402 case Intrinsic::spv_assign_name: {
5403 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpName));
5404 MIB.addUse(I.getOperand(I.getNumExplicitDefs() + 1).getReg());
5405 for (unsigned i = I.getNumExplicitDefs() + 2;
5406 i < I.getNumExplicitOperands(); ++i) {
5407 MIB.addImm(I.getOperand(i).getImm());
5408 }
5409 MIB.constrainAllUses(TII, TRI, RBI);
5410 return true;
5411 }
5412 case Intrinsic::spv_switch: {
5413 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSwitch));
5414 for (unsigned i = 1; i < I.getNumExplicitOperands(); ++i) {
5415 if (I.getOperand(i).isReg())
5416 MIB.addReg(I.getOperand(i).getReg());
5417 else if (I.getOperand(i).isCImm())
5418 addNumImm(I.getOperand(i).getCImm()->getValue(), MIB);
5419 else if (I.getOperand(i).isMBB())
5420 MIB.addMBB(I.getOperand(i).getMBB());
5421 else
5422 llvm_unreachable("Unexpected OpSwitch operand");
5423 }
5424 MIB.constrainAllUses(TII, TRI, RBI);
5425 return true;
5426 }
5427 case Intrinsic::spv_loop_merge: {
5428 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpLoopMerge));
5429 for (unsigned i = 1; i < I.getNumExplicitOperands(); ++i) {
5430 if (I.getOperand(i).isMBB())
5431 MIB.addMBB(I.getOperand(i).getMBB());
5432 else
5433 MIB.addImm(foldImm(I.getOperand(i), MRI));
5434 }
5435 MIB.constrainAllUses(TII, TRI, RBI);
5436 return true;
5437 }
5438 case Intrinsic::spv_loop_control_intel: {
5439 auto MIB =
5440 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpLoopControlINTEL));
5441 for (unsigned J = 1; J < I.getNumExplicitOperands(); ++J)
5442 MIB.addImm(foldImm(I.getOperand(J), MRI));
5443 MIB.constrainAllUses(TII, TRI, RBI);
5444 return true;
5445 }
5446 case Intrinsic::spv_selection_merge: {
5447 auto MIB =
5448 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSelectionMerge));
5449 assert(I.getOperand(1).isMBB() &&
5450 "operand 1 to spv_selection_merge must be a basic block");
5451 MIB.addMBB(I.getOperand(1).getMBB());
5452 MIB.addImm(getSelectionOperandForImm(I.getOperand(2).getImm()));
5453 MIB.constrainAllUses(TII, TRI, RBI);
5454 return true;
5455 }
5456 case Intrinsic::spv_cmpxchg:
5457 return selectAtomicCmpXchg(ResVReg, ResType, I);
5458 case Intrinsic::spv_unreachable:
5459 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpUnreachable))
5460 .constrainAllUses(TII, TRI, RBI);
5461 return true;
5462 case Intrinsic::spv_abort:
5463 return selectAbort(I);
5464 case Intrinsic::spv_alloca:
5465 return selectFrameIndex(ResVReg, ResType, I);
5466 case Intrinsic::spv_alloca_array:
5467 return selectAllocaArray(ResVReg, ResType, I);
5468 case Intrinsic::spv_assume:
5469 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_expect_assume)) {
5470 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpAssumeTrueKHR))
5471 .addUse(I.getOperand(1).getReg())
5472 .constrainAllUses(TII, TRI, RBI);
5473 return true;
5474 }
5475 break;
5476 case Intrinsic::spv_expect:
5477 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_expect_assume)) {
5478 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExpectKHR))
5479 .addDef(ResVReg)
5480 .addUse(GR.getSPIRVTypeID(ResType))
5481 .addUse(I.getOperand(2).getReg())
5482 .addUse(I.getOperand(3).getReg())
5483 .constrainAllUses(TII, TRI, RBI);
5484 return true;
5485 }
5486 break;
5487 case Intrinsic::arithmetic_fence:
5488 if (STI.canUseExtension(SPIRV::Extension::SPV_EXT_arithmetic_fence)) {
5489 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpArithmeticFenceEXT))
5490 .addDef(ResVReg)
5491 .addUse(GR.getSPIRVTypeID(ResType))
5492 .addUse(I.getOperand(2).getReg())
5493 .constrainAllUses(TII, TRI, RBI);
5494 return true;
5495 } else
5496 return BuildCOPY(ResVReg, I.getOperand(2).getReg(), I);
5497 break;
5498 case Intrinsic::spv_thread_id:
5499 // The HLSL SV_DispatchThreadID semantic is lowered to llvm.spv.thread.id
5500 // intrinsic in LLVM IR for SPIR-V backend.
5501 //
5502 // In SPIR-V backend, llvm.spv.thread.id is now correctly translated to a
5503 // `GlobalInvocationId` builtin variable
5504 return loadVec3BuiltinInputID(SPIRV::BuiltIn::GlobalInvocationId, ResVReg,
5505 ResType, I);
5506 case Intrinsic::spv_thread_id_in_group:
5507 // The HLSL SV_GroupThreadId semantic is lowered to
5508 // llvm.spv.thread.id.in.group intrinsic in LLVM IR for SPIR-V backend.
5509 //
5510 // In SPIR-V backend, llvm.spv.thread.id.in.group is now correctly
5511 // translated to a `LocalInvocationId` builtin variable
5512 return loadVec3BuiltinInputID(SPIRV::BuiltIn::LocalInvocationId, ResVReg,
5513 ResType, I);
5514 case Intrinsic::spv_group_id:
5515 // The HLSL SV_GroupId semantic is lowered to
5516 // llvm.spv.group.id intrinsic in LLVM IR for SPIR-V backend.
5517 //
5518 // In SPIR-V backend, llvm.spv.group.id is now translated to a `WorkgroupId`
5519 // builtin variable
5520 return loadVec3BuiltinInputID(SPIRV::BuiltIn::WorkgroupId, ResVReg, ResType,
5521 I);
5522 case Intrinsic::spv_flattened_thread_id_in_group:
5523 // The HLSL SV_GroupIndex semantic is lowered to
5524 // llvm.spv.flattened.thread.id.in.group() intrinsic in LLVM IR for SPIR-V
5525 // backend.
5526 //
5527 // In SPIR-V backend, llvm.spv.flattened.thread.id.in.group is translated to
5528 // a `LocalInvocationIndex` builtin variable
5529 return loadBuiltinInputID(SPIRV::BuiltIn::LocalInvocationIndex, ResVReg,
5530 ResType, I);
5531 case Intrinsic::spv_workgroup_size:
5532 return loadVec3BuiltinInputID(SPIRV::BuiltIn::WorkgroupSize, ResVReg,
5533 ResType, I);
5534 case Intrinsic::spv_global_size:
5535 return loadVec3BuiltinInputID(SPIRV::BuiltIn::GlobalSize, ResVReg, ResType,
5536 I);
5537 case Intrinsic::spv_global_offset:
5538 return loadVec3BuiltinInputID(SPIRV::BuiltIn::GlobalOffset, ResVReg,
5539 ResType, I);
5540 case Intrinsic::spv_num_workgroups:
5541 return loadVec3BuiltinInputID(SPIRV::BuiltIn::NumWorkgroups, ResVReg,
5542 ResType, I);
5543 case Intrinsic::spv_subgroup_size:
5544 return loadBuiltinInputID(SPIRV::BuiltIn::SubgroupSize, ResVReg, ResType,
5545 I);
5546 case Intrinsic::spv_num_subgroups:
5547 return loadBuiltinInputID(SPIRV::BuiltIn::NumSubgroups, ResVReg, ResType,
5548 I);
5549 case Intrinsic::spv_subgroup_id:
5550 return loadBuiltinInputID(SPIRV::BuiltIn::SubgroupId, ResVReg, ResType, I);
5551 case Intrinsic::spv_subgroup_local_invocation_id:
5552 return loadBuiltinInputID(SPIRV::BuiltIn::SubgroupLocalInvocationId,
5553 ResVReg, ResType, I);
5554 case Intrinsic::spv_subgroup_max_size:
5555 return loadBuiltinInputID(SPIRV::BuiltIn::SubgroupMaxSize, ResVReg, ResType,
5556 I);
5557 case Intrinsic::spv_fdot:
5558 return selectFloatDot(ResVReg, ResType, I);
5559 case Intrinsic::spv_udot:
5560 case Intrinsic::spv_sdot:
5561 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_integer_dot_product) ||
5562 STI.isAtLeastSPIRVVer(VersionTuple(1, 6)))
5563 return selectIntegerDot(ResVReg, ResType, I,
5564 /*Signed=*/IID == Intrinsic::spv_sdot);
5565 return selectIntegerDotExpansion(ResVReg, ResType, I);
5566 case Intrinsic::spv_dot4add_i8packed:
5567 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_integer_dot_product) ||
5568 STI.isAtLeastSPIRVVer(VersionTuple(1, 6)))
5569 return selectDot4AddPacked<true>(ResVReg, ResType, I);
5570 return selectDot4AddPackedExpansion<true>(ResVReg, ResType, I);
5571 case Intrinsic::spv_dot4add_u8packed:
5572 if (STI.canUseExtension(SPIRV::Extension::SPV_KHR_integer_dot_product) ||
5573 STI.isAtLeastSPIRVVer(VersionTuple(1, 6)))
5574 return selectDot4AddPacked<false>(ResVReg, ResType, I);
5575 return selectDot4AddPackedExpansion<false>(ResVReg, ResType, I);
5576 case Intrinsic::spv_all:
5577 return selectAll(ResVReg, ResType, I);
5578 case Intrinsic::spv_any:
5579 return selectAny(ResVReg, ResType, I);
5580 case Intrinsic::spv_distance:
5581 return selectExtInst(ResVReg, ResType, I, CL::distance, GL::Distance);
5582 case Intrinsic::spv_lerp:
5583 return selectExtInst(ResVReg, ResType, I, CL::mix, GL::FMix);
5584 case Intrinsic::spv_length:
5585 return selectExtInst(ResVReg, ResType, I, CL::length, GL::Length);
5586 case Intrinsic::spv_degrees:
5587 return selectExtInst(ResVReg, ResType, I, CL::degrees, GL::Degrees);
5588 case Intrinsic::spv_faceforward:
5589 return selectExtInst(ResVReg, ResType, I, GL::FaceForward);
5590 case Intrinsic::spv_frac:
5591 return selectExtInst(ResVReg, ResType, I, CL::fract, GL::Fract);
5592 case Intrinsic::spv_isinf:
5593 return selectOpIsInf(ResVReg, ResType, I);
5594 case Intrinsic::spv_isnan:
5595 return selectOpIsNan(ResVReg, ResType, I);
5596 case Intrinsic::spv_isfinite:
5597 return selectOpIsFinite(ResVReg, ResType, I);
5598 case Intrinsic::spv_isnormal:
5599 return selectOpIsNormal(ResVReg, ResType, I);
5600 case Intrinsic::spv_normalize:
5601 return selectExtInst(ResVReg, ResType, I, CL::normalize, GL::Normalize);
5602 case Intrinsic::spv_refract:
5603 return selectExtInst(ResVReg, ResType, I, GL::Refract);
5604 case Intrinsic::spv_reflect:
5605 return selectExtInst(ResVReg, ResType, I, GL::Reflect);
5606 case Intrinsic::spv_rsqrt:
5607 return selectExtInst(ResVReg, ResType, I, CL::rsqrt, GL::InverseSqrt);
5608 case Intrinsic::spv_sign:
5609 return selectSign(ResVReg, ResType, I);
5610 case Intrinsic::spv_smoothstep:
5611 return selectExtInst(ResVReg, ResType, I, CL::smoothstep, GL::SmoothStep);
5612 case Intrinsic::spv_firstbituhigh: // There is no CL equivalent of FindUMsb
5613 return selectFirstBitHigh(ResVReg, ResType, I, /*IsSigned=*/false);
5614 case Intrinsic::spv_firstbitshigh: // There is no CL equivalent of FindSMsb
5615 return selectFirstBitHigh(ResVReg, ResType, I, /*IsSigned=*/true);
5616 case Intrinsic::spv_firstbitlow: // There is no CL equivlent of FindILsb
5617 return selectFirstBitLow(ResVReg, ResType, I);
5618 case Intrinsic::spv_all_memory_barrier:
5619 return selectBarrierInst(I, SPIRV::Scope::Device,
5620 SPIRV::MemorySemantics::UniformMemory |
5621 SPIRV::MemorySemantics::ImageMemory |
5622 SPIRV::MemorySemantics::WorkgroupMemory,
5623 /*WithGroupSync*/ false);
5624 case Intrinsic::spv_all_memory_barrier_with_group_sync:
5625 return selectBarrierInst(I, SPIRV::Scope::Device,
5626 SPIRV::MemorySemantics::UniformMemory |
5627 SPIRV::MemorySemantics::ImageMemory |
5628 SPIRV::MemorySemantics::WorkgroupMemory,
5629 /*WithGroupSync*/ true);
5630 case Intrinsic::spv_device_memory_barrier:
5631 return selectBarrierInst(I, SPIRV::Scope::Device,
5632 SPIRV::MemorySemantics::UniformMemory |
5633 SPIRV::MemorySemantics::ImageMemory,
5634 /*WithGroupSync*/ false);
5635 case Intrinsic::spv_device_memory_barrier_with_group_sync:
5636 return selectBarrierInst(I, SPIRV::Scope::Device,
5637 SPIRV::MemorySemantics::UniformMemory |
5638 SPIRV::MemorySemantics::ImageMemory,
5639 /*WithGroupSync*/ true);
5640 case Intrinsic::spv_group_memory_barrier:
5641 return selectBarrierInst(I, SPIRV::Scope::Workgroup,
5642 SPIRV::MemorySemantics::WorkgroupMemory,
5643 /*WithGroupSync*/ false);
5644 case Intrinsic::spv_group_memory_barrier_with_group_sync:
5645 return selectBarrierInst(I, SPIRV::Scope::Workgroup,
5646 SPIRV::MemorySemantics::WorkgroupMemory,
5647 /*WithGroupSync*/ true);
5648 case Intrinsic::spv_generic_cast_to_ptr_explicit: {
5649 Register PtrReg = I.getOperand(I.getNumExplicitDefs() + 1).getReg();
5650 SPIRV::StorageClass::StorageClass ResSC =
5651 GR.getPointerStorageClass(ResType);
5652 if (!isGenericCastablePtr(ResSC))
5653 return diagnoseUnsupported(I, "The target storage class is not castable "
5654 "from the Generic storage class");
5655 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpGenericCastToPtrExplicit))
5656 .addDef(ResVReg)
5657 .addUse(GR.getSPIRVTypeID(ResType))
5658 .addUse(PtrReg)
5659 .addImm(ResSC)
5660 .constrainAllUses(TII, TRI, RBI);
5661 return true;
5662 }
5663 case Intrinsic::spv_lifetime_start:
5664 case Intrinsic::spv_lifetime_end: {
5665 unsigned Op = IID == Intrinsic::spv_lifetime_start ? SPIRV::OpLifetimeStart
5666 : SPIRV::OpLifetimeStop;
5667 int64_t Size = I.getOperand(I.getNumExplicitDefs() + 1).getImm();
5668 Register PtrReg = I.getOperand(I.getNumExplicitDefs() + 2).getReg();
5669 if (Size == -1)
5670 Size = 0;
5671 BuildMI(BB, I, I.getDebugLoc(), TII.get(Op))
5672 .addUse(PtrReg)
5673 .addImm(Size)
5674 .constrainAllUses(TII, TRI, RBI);
5675 return true;
5676 }
5677 case Intrinsic::spv_saturate:
5678 return selectSaturate(ResVReg, ResType, I);
5679 case Intrinsic::spv_nclamp:
5680 return selectExtInst(ResVReg, ResType, I, CL::fclamp, GL::NClamp);
5681 case Intrinsic::spv_uclamp:
5682 return selectExtInst(ResVReg, ResType, I, CL::u_clamp, GL::UClamp);
5683 case Intrinsic::spv_sclamp:
5684 return selectExtInst(ResVReg, ResType, I, CL::s_clamp, GL::SClamp);
5685 case Intrinsic::spv_subgroup_prefix_bit_count:
5686 return selectWavePrefixBitCount(ResVReg, ResType, I);
5687 case Intrinsic::spv_wave_active_countbits:
5688 return selectWaveActiveCountBits(ResVReg, ResType, I);
5689 case Intrinsic::spv_wave_all_equal:
5690 return selectWaveActiveAllEqual(ResVReg, ResType, I);
5691 case Intrinsic::spv_wave_all:
5692 return selectWaveOpInst(ResVReg, ResType, I, SPIRV::OpGroupNonUniformAll);
5693 case Intrinsic::spv_wave_any:
5694 return selectWaveOpInst(ResVReg, ResType, I, SPIRV::OpGroupNonUniformAny);
5695 case Intrinsic::spv_subgroup_ballot:
5696 return selectWaveOpInst(ResVReg, ResType, I,
5697 SPIRV::OpGroupNonUniformBallot);
5698 case Intrinsic::spv_wave_is_first_lane:
5699 return selectWaveOpInst(ResVReg, ResType, I, SPIRV::OpGroupNonUniformElect);
5700 case Intrinsic::spv_wave_reduce_or:
5701 return selectWaveReduceOp(ResVReg, ResType, I,
5702 SPIRV::OpGroupNonUniformBitwiseOr);
5703 case Intrinsic::spv_wave_reduce_xor:
5704 return selectWaveReduceOp(ResVReg, ResType, I,
5705 SPIRV::OpGroupNonUniformBitwiseXor);
5706 case Intrinsic::spv_wave_reduce_and:
5707 return selectWaveReduceOp(ResVReg, ResType, I,
5708 SPIRV::OpGroupNonUniformBitwiseAnd);
5709 case Intrinsic::spv_wave_reduce_umax:
5710 return selectWaveReduceMax(ResVReg, ResType, I, /*IsUnsigned*/ true);
5711 case Intrinsic::spv_wave_reduce_max:
5712 return selectWaveReduceMax(ResVReg, ResType, I, /*IsUnsigned*/ false);
5713 case Intrinsic::spv_wave_reduce_umin:
5714 return selectWaveReduceMin(ResVReg, ResType, I, /*IsUnsigned*/ true);
5715 case Intrinsic::spv_wave_reduce_min:
5716 return selectWaveReduceMin(ResVReg, ResType, I, /*IsUnsigned*/ false);
5717 case Intrinsic::spv_wave_reduce_sum:
5718 return selectWaveReduceSum(ResVReg, ResType, I);
5719 case Intrinsic::spv_wave_product:
5720 return selectWaveReduceProduct(ResVReg, ResType, I);
5721 case Intrinsic::spv_wave_readlane:
5722 return selectWaveOpInst(ResVReg, ResType, I,
5723 SPIRV::OpGroupNonUniformShuffle);
5724 case Intrinsic::spv_wave_readlane_first:
5725 return selectWaveOpInst(ResVReg, ResType, I,
5726 SPIRV::OpGroupNonUniformBroadcastFirst);
5727 case Intrinsic::spv_wave_prefix_sum:
5728 return selectWaveExclusiveScanSum(ResVReg, ResType, I);
5729 case Intrinsic::spv_wave_prefix_product:
5730 return selectWaveExclusiveScanProduct(ResVReg, ResType, I);
5731 case Intrinsic::spv_quad_read_across_x: {
5732 return selectQuadSwap(ResVReg, ResType, I, /*Direction*/ 0);
5733 }
5734 case Intrinsic::spv_quad_read_across_y: {
5735 return selectQuadSwap(ResVReg, ResType, I, /*Direction*/ 1);
5736 }
5737 case Intrinsic::spv_quad_read_across_diagonal: {
5738 return selectQuadSwap(ResVReg, ResType, I, /*Direction*/ 2);
5739 }
5740 case Intrinsic::spv_radians:
5741 return selectExtInst(ResVReg, ResType, I, CL::radians, GL::Radians);
5742 // Discard intrinsics which we do not expect to actually represent code after
5743 // lowering or intrinsics which are not implemented but should not crash when
5744 // found in a customer's LLVM IR input.
5745 case Intrinsic::instrprof_increment:
5746 case Intrinsic::instrprof_increment_step:
5747 case Intrinsic::instrprof_value_profile:
5748 break;
5749 // Discard internal intrinsics.
5750 case Intrinsic::spv_value_md:
5751 break;
5752 case Intrinsic::spv_resource_handlefrombinding: {
5753 return selectHandleFromBinding(ResVReg, ResType, I);
5754 }
5755 case Intrinsic::spv_resource_counterhandlefrombinding:
5756 return selectCounterHandleFromBinding(ResVReg, ResType, I);
5757 case Intrinsic::spv_resource_updatecounter:
5758 return selectUpdateCounter(ResVReg, ResType, I);
5759 case Intrinsic::spv_resource_store_typedbuffer: {
5760 return selectImageWriteIntrinsic(I);
5761 }
5762 case Intrinsic::spv_resource_load_typedbuffer: {
5763 return selectReadImageIntrinsic(ResVReg, ResType, I);
5764 }
5765 case Intrinsic::spv_resource_load_level: {
5766 return selectLoadLevelIntrinsic(ResVReg, ResType, I);
5767 }
5768 case Intrinsic::spv_resource_getdimensions_x:
5769 case Intrinsic::spv_resource_getdimensions_xy:
5770 case Intrinsic::spv_resource_getdimensions_xyz: {
5771 return selectGetDimensionsIntrinsic(ResVReg, ResType, I);
5772 }
5773 case Intrinsic::spv_resource_getdimensions_levels_x:
5774 case Intrinsic::spv_resource_getdimensions_levels_xy:
5775 case Intrinsic::spv_resource_getdimensions_levels_xyz: {
5776 return selectGetDimensionsLevelsIntrinsic(ResVReg, ResType, I);
5777 }
5778 case Intrinsic::spv_resource_getdimensions_ms_xy:
5779 case Intrinsic::spv_resource_getdimensions_ms_xyz: {
5780 return selectGetDimensionsMSIntrinsic(ResVReg, ResType, I);
5781 }
5782 case Intrinsic::spv_resource_calculate_lod:
5783 case Intrinsic::spv_resource_calculate_lod_unclamped:
5784 return selectCalculateLodIntrinsic(ResVReg, ResType, I);
5785 case Intrinsic::spv_resource_sample:
5786 case Intrinsic::spv_resource_sample_clamp:
5787 return selectSampleBasicIntrinsic(ResVReg, ResType, I);
5788 case Intrinsic::spv_resource_samplebias:
5789 case Intrinsic::spv_resource_samplebias_clamp:
5790 return selectSampleBiasIntrinsic(ResVReg, ResType, I);
5791 case Intrinsic::spv_resource_samplegrad:
5792 case Intrinsic::spv_resource_samplegrad_clamp:
5793 return selectSampleGradIntrinsic(ResVReg, ResType, I);
5794 case Intrinsic::spv_resource_samplelevel:
5795 return selectSampleLevelIntrinsic(ResVReg, ResType, I);
5796 case Intrinsic::spv_resource_samplecmp:
5797 case Intrinsic::spv_resource_samplecmp_clamp:
5798 return selectSampleCmpIntrinsic(ResVReg, ResType, I);
5799 case Intrinsic::spv_resource_samplecmplevelzero:
5800 return selectSampleCmpLevelZeroIntrinsic(ResVReg, ResType, I);
5801 case Intrinsic::spv_resource_gather:
5802 case Intrinsic::spv_resource_gather_cmp:
5803 return selectGatherIntrinsic(ResVReg, ResType, I);
5804 case Intrinsic::spv_resource_getbasepointer:
5805 case Intrinsic::spv_resource_getpointer: {
5806 return selectResourceGetPointer(ResVReg, ResType, I);
5807 }
5808 case Intrinsic::spv_pushconstant_getpointer: {
5809 return selectPushConstantGetPointer(ResVReg, ResType, I);
5810 }
5811 case Intrinsic::spv_discard: {
5812 return selectDiscard(ResVReg, ResType, I);
5813 }
5814 case Intrinsic::spv_resource_nonuniformindex: {
5815 return selectResourceNonUniformIndex(ResVReg, ResType, I);
5816 }
5817 case Intrinsic::spv_unpackhalf2x16: {
5818 return selectExtInst(ResVReg, ResType, I, GL::UnpackHalf2x16);
5819 }
5820 case Intrinsic::spv_packhalf2x16: {
5821 return selectExtInst(ResVReg, ResType, I, GL::PackHalf2x16);
5822 }
5823 case Intrinsic::spv_ddx:
5824 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpDPdx);
5825 case Intrinsic::spv_ddy:
5826 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpDPdy);
5827 case Intrinsic::spv_ddx_coarse:
5828 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpDPdxCoarse);
5829 case Intrinsic::spv_ddy_coarse:
5830 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpDPdyCoarse);
5831 case Intrinsic::spv_ddx_fine:
5832 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpDPdxFine);
5833 case Intrinsic::spv_ddy_fine:
5834 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpDPdyFine);
5835 case Intrinsic::spv_fwidth:
5836 return selectDerivativeInst(ResVReg, ResType, I, SPIRV::OpFwidth);
5837 case Intrinsic::spv_masked_gather:
5838 if (STI.canUseExtension(SPIRV::Extension::SPV_INTEL_masked_gather_scatter))
5839 return selectMaskedGather(ResVReg, ResType, I);
5840 return diagnoseUnsupported(
5841 I, "llvm.masked.gather requires SPV_INTEL_masked_gather_scatter");
5842 case Intrinsic::spv_masked_scatter:
5843 if (STI.canUseExtension(SPIRV::Extension::SPV_INTEL_masked_gather_scatter))
5844 return selectMaskedScatter(I);
5845 return diagnoseUnsupported(
5846 I, "llvm.masked.scatter requires SPV_INTEL_masked_gather_scatter");
5847 case Intrinsic::returnaddress:
5848 case Intrinsic::frameaddress: {
5849 // SPIR-V does not have a stack or return address. Lower to null.
5850 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpConstantNull))
5851 .addDef(ResVReg)
5852 .addUse(GR.getSPIRVTypeID(ResType));
5853 MIB.constrainAllUses(TII, TRI, RBI);
5854 return true;
5855 }
5856 default:
5857 return diagnoseUnsupported(I, "intrinsic selection not implemented.");
5858 }
5859 return true;
5860}
5861
5862bool SPIRVInstructionSelector::selectHandleFromBinding(Register &ResVReg,
5863 SPIRVTypeInst ResType,
5864 MachineInstr &I) const {
5865 // The images need to be loaded in the same basic block as their use. We defer
5866 // loading the image to the intrinsic that uses it.
5867 if (ResType->getOpcode() == SPIRV::OpTypeImage)
5868 return true;
5869
5870 return loadHandleBeforePosition(ResVReg, GR.getSPIRVTypeForVReg(ResVReg),
5871 *cast<GIntrinsic>(&I), I);
5872}
5873
5874bool SPIRVInstructionSelector::selectCounterHandleFromBinding(
5875 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
5876 auto &Intr = cast<GIntrinsic>(I);
5877 assert(Intr.getIntrinsicID() ==
5878 Intrinsic::spv_resource_counterhandlefrombinding);
5879
5880 // Extract information from the intrinsic call.
5881 Register MainHandleReg = Intr.getOperand(2).getReg();
5882 auto *MainHandleDef = cast<GIntrinsic>(getVRegDef(*MRI, MainHandleReg));
5883 assert(MainHandleDef->getIntrinsicID() ==
5884 Intrinsic::spv_resource_handlefrombinding);
5885
5886 uint32_t Set = getIConstVal(Intr.getOperand(3).getReg(), MRI);
5887 uint32_t Binding = getIConstVal(Intr.getOperand(4).getReg(), MRI);
5888 uint32_t ArraySize = getIConstVal(MainHandleDef->getOperand(4).getReg(), MRI);
5889 Register IndexReg = MainHandleDef->getOperand(5).getReg();
5890 std::string CounterName =
5891 getStringValueFromReg(MainHandleDef->getOperand(6).getReg(), *MRI) +
5892 ".counter";
5893
5894 // Create the counter variable.
5895 MachineIRBuilder MIRBuilder(I);
5896 Register CounterVarReg =
5897 buildPointerToResource(SPIRVTypeInst(GR.getPointeeType(ResType)),
5898 GR.getPointerStorageClass(ResType), Set, Binding,
5899 ArraySize, IndexReg, CounterName, MIRBuilder);
5900
5901 return BuildCOPY(ResVReg, CounterVarReg, I);
5902}
5903
5904bool SPIRVInstructionSelector::selectUpdateCounter(Register &ResVReg,
5905 SPIRVTypeInst ResType,
5906 MachineInstr &I) const {
5907 auto &Intr = cast<GIntrinsic>(I);
5908 assert(Intr.getIntrinsicID() == Intrinsic::spv_resource_updatecounter);
5909
5910 Register CounterHandleReg = Intr.getOperand(2).getReg();
5911 Register IncrReg = Intr.getOperand(3).getReg();
5912
5913 // The counter handle is a pointer to the counter variable (which is a struct
5914 // containing an i32). We need to get a pointer to that i32 member to do the
5915 // atomic operation.
5916#ifndef NDEBUG
5917 SPIRVTypeInst CounterVarType = GR.getSPIRVTypeForVReg(CounterHandleReg);
5918 SPIRVTypeInst CounterVarPointeeType = GR.getPointeeType(CounterVarType);
5919 assert(CounterVarPointeeType &&
5920 CounterVarPointeeType->getOpcode() == SPIRV::OpTypeStruct &&
5921 "Counter variable must be a struct");
5922 assert(GR.getPointerStorageClass(CounterVarType) ==
5923 SPIRV::StorageClass::StorageBuffer &&
5924 "Counter variable must be in the storage buffer storage class");
5925 assert(CounterVarPointeeType->getNumOperands() == 2 &&
5926 "Counter variable must have exactly 1 member in the struct");
5927 const SPIRVTypeInst MemberType =
5928 GR.getSPIRVTypeForVReg(CounterVarPointeeType->getOperand(1).getReg());
5929 assert(MemberType->getOpcode() == SPIRV::OpTypeInt &&
5930 "Counter variable struct must have a single i32 member");
5931#endif
5932
5933 // The struct has a single i32 member.
5934 MachineIRBuilder MIRBuilder(I);
5935 const Type *LLVMIntType =
5936 Type::getInt32Ty(I.getMF()->getFunction().getContext());
5937
5938 SPIRVTypeInst IntPtrType = GR.getOrCreateSPIRVPointerType(
5939 LLVMIntType, MIRBuilder, SPIRV::StorageClass::StorageBuffer);
5940
5941 Register Zero = buildI32Constant(0, I);
5942
5943 Register PtrToCounter =
5944 MRI->createVirtualRegister(GR.getRegClass(IntPtrType));
5945 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpAccessChain))
5946 .addDef(PtrToCounter)
5947 .addUse(GR.getSPIRVTypeID(IntPtrType))
5948 .addUse(CounterHandleReg)
5949 .addUse(Zero)
5950 .constrainAllUses(TII, TRI, RBI);
5951
5952 // For UAV/SSBO counters, the scope is Device. The counter variable is not
5953 // used as a flag. So the memory semantics can be None.
5954 Register Scope = buildI32Constant(SPIRV::Scope::Device, I);
5955 Register Semantics = buildI32Constant(SPIRV::MemorySemantics::None, I);
5956
5957 int64_t IncrVal = getIConstValSext(IncrReg, MRI);
5958 Register Incr = buildI32Constant(static_cast<uint32_t>(IncrVal), I);
5959
5960 Register AtomicRes = MRI->createVirtualRegister(GR.getRegClass(ResType));
5961 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpAtomicIAdd))
5962 .addDef(AtomicRes)
5963 .addUse(GR.getSPIRVTypeID(ResType))
5964 .addUse(PtrToCounter)
5965 .addUse(Scope)
5966 .addUse(Semantics)
5967 .addUse(Incr)
5968 .constrainAllUses(TII, TRI, RBI);
5969 if (IncrVal >= 0) {
5970 return BuildCOPY(ResVReg, AtomicRes, I);
5971 }
5972
5973 // In HLSL, IncrementCounter returns the value *before* the increment, while
5974 // DecrementCounter returns the value *after* the decrement. Both are lowered
5975 // to the same atomic intrinsic which returns the value *before* the
5976 // operation. So for decrements (negative IncrVal), we must subtract the
5977 // increment value from the result to get the post-decrement value.
5978 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpIAddS))
5979 .addDef(ResVReg)
5980 .addUse(GR.getSPIRVTypeID(ResType))
5981 .addUse(AtomicRes)
5982 .addUse(Incr)
5983 .constrainAllUses(TII, TRI, RBI);
5984 return true;
5985}
5986bool SPIRVInstructionSelector::selectReadImageIntrinsic(Register &ResVReg,
5987 SPIRVTypeInst ResType,
5988 MachineInstr &I) const {
5989
5990 // If the load of the image is in a different basic block, then
5991 // this will generate invalid code. A proper solution is to move
5992 // the OpLoad from selectHandleFromBinding here. However, to do
5993 // that we will need to change the return type of the intrinsic.
5994 // We will do that when we can, but for now trying to move forward with other
5995 // issues.
5996 Register ImageReg = I.getOperand(2).getReg();
5997 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
5998 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
5999 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
6000 *ImageDef, I)) {
6001 return false;
6002 }
6003
6004 Register IdxReg = I.getOperand(3).getReg();
6005 DebugLoc Loc = I.getDebugLoc();
6006 MachineInstr &Pos = I;
6007
6008 return generateImageReadOrFetch(ResVReg, ResType, NewImageReg, IdxReg, Loc,
6009 Pos);
6010}
6011
6012bool SPIRVInstructionSelector::generateSampleImage(
6013 Register ResVReg, SPIRVTypeInst ResType, Register ImageReg,
6014 Register SamplerReg, Register CoordinateReg, const ImageOperands &ImOps,
6015 DebugLoc Loc, MachineInstr &Pos) const {
6016 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
6017 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
6018 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
6019 *ImageDef, Pos)) {
6020 return false;
6021 }
6022
6023 auto *SamplerDef = cast<GIntrinsic>(getVRegDef(*MRI, SamplerReg));
6024 Register NewSamplerReg =
6025 MRI->createVirtualRegister(MRI->getRegClass(SamplerReg));
6026 if (!loadHandleBeforePosition(NewSamplerReg,
6027 GR.getSPIRVTypeForVReg(SamplerReg), *SamplerDef,
6028 Pos)) {
6029 return false;
6030 }
6031
6032 MachineIRBuilder MIRBuilder(Pos);
6033 SPIRVTypeInst SampledImageType = GR.getOrCreateOpTypeSampledImage(
6034 GR.getSPIRVTypeForVReg(ImageReg), MIRBuilder);
6035 Register SampledImageReg =
6036 MRI->createVirtualRegister(GR.getRegClass(SampledImageType));
6037
6038 BuildMI(*Pos.getParent(), Pos, Loc, TII.get(SPIRV::OpSampledImage))
6039 .addDef(SampledImageReg)
6040 .addUse(GR.getSPIRVTypeID(SampledImageType))
6041 .addUse(NewImageReg)
6042 .addUse(NewSamplerReg)
6043 .constrainAllUses(TII, TRI, RBI);
6044
6045 bool IsExplicitLod = ImOps.GradX.has_value() || ImOps.GradY.has_value() ||
6046 ImOps.Lod.has_value();
6047 unsigned Opcode = IsExplicitLod ? SPIRV::OpImageSampleExplicitLod
6048 : SPIRV::OpImageSampleImplicitLod;
6049 if (ImOps.Compare)
6050 Opcode = IsExplicitLod ? SPIRV::OpImageSampleDrefExplicitLod
6051 : SPIRV::OpImageSampleDrefImplicitLod;
6052
6053 auto MIB = BuildMI(*Pos.getParent(), Pos, Loc, TII.get(Opcode))
6054 .addDef(ResVReg)
6055 .addUse(GR.getSPIRVTypeID(ResType))
6056 .addUse(SampledImageReg)
6057 .addUse(CoordinateReg);
6058
6059 if (ImOps.Compare)
6060 MIB.addUse(*ImOps.Compare);
6061
6062 uint32_t ImageOperands = 0;
6063 if (ImOps.Bias)
6064 ImageOperands |= SPIRV::ImageOperand::Bias;
6065 if (ImOps.Lod)
6066 ImageOperands |= SPIRV::ImageOperand::Lod;
6067 if (ImOps.GradX && ImOps.GradY)
6068 ImageOperands |= SPIRV::ImageOperand::Grad;
6069 if (ImOps.Offset && !isScalarOrVectorIntConstantZero(*ImOps.Offset)) {
6070 if (isConstReg(MRI, *ImOps.Offset))
6071 ImageOperands |= SPIRV::ImageOperand::ConstOffset;
6072 else {
6073 Pos.emitGenericError(
6074 "Non-constant offsets are not supported in sample instructions.");
6075 return false;
6076 }
6077 }
6078 if (ImOps.MinLod)
6079 ImageOperands |= SPIRV::ImageOperand::MinLod;
6080
6081 if (ImageOperands != 0) {
6082 MIB.addImm(ImageOperands);
6083 if (ImageOperands & SPIRV::ImageOperand::Bias)
6084 MIB.addUse(*ImOps.Bias);
6085 if (ImageOperands & SPIRV::ImageOperand::Lod)
6086 MIB.addUse(*ImOps.Lod);
6087 if (ImageOperands & SPIRV::ImageOperand::Grad) {
6088 MIB.addUse(*ImOps.GradX);
6089 MIB.addUse(*ImOps.GradY);
6090 }
6091 if (ImageOperands &
6092 (SPIRV::ImageOperand::ConstOffset | SPIRV::ImageOperand::Offset))
6093 MIB.addUse(*ImOps.Offset);
6094 if (ImageOperands & SPIRV::ImageOperand::MinLod)
6095 MIB.addUse(*ImOps.MinLod);
6096 }
6097
6098 MIB.constrainAllUses(TII, TRI, RBI);
6099 return true;
6100}
6101
6102bool SPIRVInstructionSelector::selectImageQuerySize(
6103 Register ImageReg, Register &ResVReg, MachineInstr &I,
6104 std::optional<Register> LodReg) const {
6105 unsigned Opcode =
6106 LodReg ? SPIRV::OpImageQuerySizeLod : SPIRV::OpImageQuerySize;
6107 SPIRVTypeInst ImageType = GR.getSPIRVTypeForVReg(ImageReg);
6108 assert(ImageType && ImageType->getOpcode() == SPIRV::OpTypeImage &&
6109 "ImageReg is not an image type.");
6110
6111 auto Dim = static_cast<SPIRV::Dim::Dim>(ImageType->getOperand(2).getImm());
6112 bool IsArray = ImageType->getOperand(4).getImm() != 0;
6113 unsigned NumComponents = 0;
6114 switch (Dim) {
6115 case SPIRV::Dim::DIM_1D:
6116 case SPIRV::Dim::DIM_Buffer:
6117 NumComponents = IsArray ? 2 : 1;
6118 break;
6119 case SPIRV::Dim::DIM_2D:
6120 case SPIRV::Dim::DIM_Cube:
6121 case SPIRV::Dim::DIM_Rect:
6122 NumComponents = IsArray ? 3 : 2;
6123 break;
6124 case SPIRV::Dim::DIM_3D:
6125 NumComponents = 3;
6126 break;
6127 default:
6128 I.emitGenericError("Unsupported image dimension for OpImageQuerySize.");
6129 return false;
6130 }
6131
6132 SPIRVTypeInst I32Ty = GR.getOrCreateSPIRVIntegerType(32, I, TII);
6133 SPIRVTypeInst ResType =
6134 NumComponents == 1
6135 ? I32Ty
6136 : GR.getOrCreateSPIRVVectorType(I32Ty, NumComponents, I, TII);
6137
6138 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
6139 .addDef(ResVReg)
6140 .addUse(GR.getSPIRVTypeID(ResType))
6141 .addUse(ImageReg);
6142 if (LodReg)
6143 MIB.addUse(*LodReg);
6144 MIB.constrainAllUses(TII, TRI, RBI);
6145 return true;
6146}
6147
6148bool SPIRVInstructionSelector::selectGetDimensionsIntrinsic(
6149 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
6150 Register ImageReg = I.getOperand(2).getReg();
6151 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
6152 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
6153 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
6154 *ImageDef, I)) {
6155 return false;
6156 }
6157 return selectImageQuerySize(NewImageReg, ResVReg, I);
6158}
6159
6160bool SPIRVInstructionSelector::selectGetDimensionsLevelsIntrinsic(
6161 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
6162 Register ImageReg = I.getOperand(2).getReg();
6163 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
6164 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
6165 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
6166 *ImageDef, I)) {
6167 return false;
6168 }
6169
6170 Register SizeReg = MRI->createVirtualRegister(&SPIRV::iIDRegClass);
6171 Register LodReg = I.getOperand(3).getReg();
6172
6173 assert(GR.getSPIRVTypeForVReg(NewImageReg)->getOperand(6).getImm() == 1 &&
6174 "OpImageQuerySizeLod and OpImageQueryLevels require a sampled image");
6175
6176 if (!selectImageQuerySize(NewImageReg, SizeReg, I, LodReg)) {
6177 return false;
6178 }
6179
6180 SPIRVTypeInst I32Ty = GR.getOrCreateSPIRVIntegerType(32, I, TII);
6181 Register LevelsReg = MRI->createVirtualRegister(&SPIRV::iIDRegClass);
6182 BuildMI(*I.getParent(), I, I.getDebugLoc(),
6183 TII.get(SPIRV::OpImageQueryLevels))
6184 .addDef(LevelsReg)
6185 .addUse(GR.getSPIRVTypeID(I32Ty))
6186 .addUse(NewImageReg)
6187 .constrainAllUses(TII, TRI, RBI);
6188
6189 BuildMI(*I.getParent(), I, I.getDebugLoc(),
6190 TII.get(SPIRV::OpCompositeConstruct))
6191 .addDef(ResVReg)
6192 .addUse(GR.getSPIRVTypeID(ResType))
6193 .addUse(SizeReg)
6194 .addUse(LevelsReg)
6195 .constrainAllUses(TII, TRI, RBI);
6196
6197 return true;
6198}
6199
6200bool SPIRVInstructionSelector::selectGetDimensionsMSIntrinsic(
6201 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
6202 Register ImageReg = I.getOperand(2).getReg();
6203 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
6204 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
6205 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
6206 *ImageDef, I)) {
6207 return false;
6208 }
6209
6210 Register SizeReg = MRI->createVirtualRegister(&SPIRV::iIDRegClass);
6211
6212 assert(GR.getSPIRVTypeForVReg(NewImageReg)->getOperand(5).getImm() == 1 &&
6213 "OpImageQuerySamples requires a multisampled image");
6214
6215 if (!selectImageQuerySize(NewImageReg, SizeReg, I)) {
6216 return false;
6217 }
6218
6219 Register SamplesReg = MRI->createVirtualRegister(&SPIRV::iIDRegClass);
6220
6221 SPIRVTypeInst I32Ty = GR.getOrCreateSPIRVIntegerType(32, I, TII);
6222 BuildMI(*I.getParent(), I, I.getDebugLoc(),
6223 TII.get(SPIRV::OpImageQuerySamples))
6224 .addDef(SamplesReg)
6225 .addUse(GR.getSPIRVTypeID(I32Ty))
6226 .addUse(NewImageReg)
6227 .constrainAllUses(TII, TRI, RBI);
6228
6229 BuildMI(*I.getParent(), I, I.getDebugLoc(),
6230 TII.get(SPIRV::OpCompositeConstruct))
6231 .addDef(ResVReg)
6232 .addUse(GR.getSPIRVTypeID(ResType))
6233 .addUse(SizeReg)
6234 .addUse(SamplesReg)
6235 .constrainAllUses(TII, TRI, RBI);
6236
6237 return true;
6238}
6239
6240bool SPIRVInstructionSelector::selectCalculateLodIntrinsic(
6241 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
6242 Register ImageReg = I.getOperand(2).getReg();
6243 Register SamplerReg = I.getOperand(3).getReg();
6244 Register CoordinateReg = I.getOperand(4).getReg();
6245
6246 auto *ImageDef = dyn_cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
6247 if (!ImageDef)
6248 return false;
6249 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
6250 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
6251 *ImageDef, I)) {
6252 return false;
6253 }
6254
6255 auto *SamplerDef = dyn_cast<GIntrinsic>(getVRegDef(*MRI, SamplerReg));
6256 if (!SamplerDef)
6257 return false;
6258 Register NewSamplerReg =
6259 MRI->createVirtualRegister(MRI->getRegClass(SamplerReg));
6260 if (!loadHandleBeforePosition(
6261 NewSamplerReg, GR.getSPIRVTypeForVReg(SamplerReg), *SamplerDef, I)) {
6262 return false;
6263 }
6264
6265 MachineIRBuilder MIRBuilder(I);
6266 SPIRVTypeInst SampledImageType = GR.getOrCreateOpTypeSampledImage(
6267 GR.getSPIRVTypeForVReg(ImageReg), MIRBuilder);
6268 Register SampledImageReg =
6269 MRI->createVirtualRegister(GR.getRegClass(SampledImageType));
6270
6271 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpSampledImage))
6272 .addDef(SampledImageReg)
6273 .addUse(GR.getSPIRVTypeID(SampledImageType))
6274 .addUse(NewImageReg)
6275 .addUse(NewSamplerReg)
6276 .constrainAllUses(TII, TRI, RBI);
6277
6278 SPIRVTypeInst Vec2Ty = GR.getOrCreateSPIRVVectorType(ResType, 2, I, TII);
6279 Register QueryResultReg = MRI->createVirtualRegister(GR.getRegClass(Vec2Ty));
6280
6281 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpImageQueryLod))
6282 .addDef(QueryResultReg)
6283 .addUse(GR.getSPIRVTypeID(Vec2Ty))
6284 .addUse(SampledImageReg)
6285 .addUse(CoordinateReg)
6286 .constrainAllUses(TII, TRI, RBI);
6287
6288 unsigned ExtractedIndex =
6289 cast<GIntrinsic>(I).getIntrinsicID() ==
6290 Intrinsic::spv_resource_calculate_lod_unclamped
6291 ? 1
6292 : 0;
6293
6294 MachineInstrBuilder MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
6295 TII.get(SPIRV::OpCompositeExtract))
6296 .addDef(ResVReg)
6297 .addUse(GR.getSPIRVTypeID(ResType))
6298 .addUse(QueryResultReg)
6299 .addImm(ExtractedIndex);
6300
6301 MIB.constrainAllUses(TII, TRI, RBI);
6302 return true;
6303}
6304
6305bool SPIRVInstructionSelector::selectSampleBasicIntrinsic(
6306 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
6307 Register ImageReg = I.getOperand(2).getReg();
6308 Register SamplerReg = I.getOperand(3).getReg();
6309 Register CoordinateReg = I.getOperand(4).getReg();
6310 ImageOperands ImOps;
6311 if (I.getNumOperands() > 5)
6312 ImOps.Offset = I.getOperand(5).getReg();
6313 if (I.getNumOperands() > 6)
6314 ImOps.MinLod = I.getOperand(6).getReg();
6315 return generateSampleImage(ResVReg, ResType, ImageReg, SamplerReg,
6316 CoordinateReg, ImOps, I.getDebugLoc(), I);
6317}
6318
6319bool SPIRVInstructionSelector::selectSampleBiasIntrinsic(
6320 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
6321 Register ImageReg = I.getOperand(2).getReg();
6322 Register SamplerReg = I.getOperand(3).getReg();
6323 Register CoordinateReg = I.getOperand(4).getReg();
6324 ImageOperands ImOps;
6325 ImOps.Bias = I.getOperand(5).getReg();
6326 if (I.getNumOperands() > 6)
6327 ImOps.Offset = I.getOperand(6).getReg();
6328 if (I.getNumOperands() > 7)
6329 ImOps.MinLod = I.getOperand(7).getReg();
6330 return generateSampleImage(ResVReg, ResType, ImageReg, SamplerReg,
6331 CoordinateReg, ImOps, I.getDebugLoc(), I);
6332}
6333
6334bool SPIRVInstructionSelector::selectSampleGradIntrinsic(
6335 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
6336 Register ImageReg = I.getOperand(2).getReg();
6337 Register SamplerReg = I.getOperand(3).getReg();
6338 Register CoordinateReg = I.getOperand(4).getReg();
6339 ImageOperands ImOps;
6340 ImOps.GradX = I.getOperand(5).getReg();
6341 ImOps.GradY = I.getOperand(6).getReg();
6342 if (I.getNumOperands() > 7)
6343 ImOps.Offset = I.getOperand(7).getReg();
6344 if (I.getNumOperands() > 8)
6345 ImOps.MinLod = I.getOperand(8).getReg();
6346 return generateSampleImage(ResVReg, ResType, ImageReg, SamplerReg,
6347 CoordinateReg, ImOps, I.getDebugLoc(), I);
6348}
6349
6350bool SPIRVInstructionSelector::selectSampleLevelIntrinsic(
6351 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
6352 Register ImageReg = I.getOperand(2).getReg();
6353 Register SamplerReg = I.getOperand(3).getReg();
6354 Register CoordinateReg = I.getOperand(4).getReg();
6355 ImageOperands ImOps;
6356 ImOps.Lod = I.getOperand(5).getReg();
6357 if (I.getNumOperands() > 6)
6358 ImOps.Offset = I.getOperand(6).getReg();
6359 return generateSampleImage(ResVReg, ResType, ImageReg, SamplerReg,
6360 CoordinateReg, ImOps, I.getDebugLoc(), I);
6361}
6362
6363bool SPIRVInstructionSelector::selectSampleCmpIntrinsic(Register &ResVReg,
6364 SPIRVTypeInst ResType,
6365 MachineInstr &I) const {
6366 Register ImageReg = I.getOperand(2).getReg();
6367 Register SamplerReg = I.getOperand(3).getReg();
6368 Register CoordinateReg = I.getOperand(4).getReg();
6369 ImageOperands ImOps;
6370 ImOps.Compare = I.getOperand(5).getReg();
6371 if (I.getNumOperands() > 6)
6372 ImOps.Offset = I.getOperand(6).getReg();
6373 if (I.getNumOperands() > 7)
6374 ImOps.MinLod = I.getOperand(7).getReg();
6375 return generateSampleImage(ResVReg, ResType, ImageReg, SamplerReg,
6376 CoordinateReg, ImOps, I.getDebugLoc(), I);
6377}
6378
6379bool SPIRVInstructionSelector::selectLoadLevelIntrinsic(Register &ResVReg,
6380 SPIRVTypeInst ResType,
6381 MachineInstr &I) const {
6382 Register ImageReg = I.getOperand(2).getReg();
6383 Register CoordinateReg = I.getOperand(3).getReg();
6384 Register LodReg = I.getOperand(4).getReg();
6385
6386 ImageOperands ImOps;
6387 ImOps.Lod = LodReg;
6388 if (I.getNumOperands() > 5)
6389 ImOps.Offset = I.getOperand(5).getReg();
6390
6391 auto *ImageDef = dyn_cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
6392 if (!ImageDef)
6393 return false;
6394
6395 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
6396 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
6397 *ImageDef, I)) {
6398 return false;
6399 }
6400
6401 return generateImageReadOrFetch(ResVReg, ResType, NewImageReg, CoordinateReg,
6402 I.getDebugLoc(), I, &ImOps);
6403}
6404
6405bool SPIRVInstructionSelector::selectSampleCmpLevelZeroIntrinsic(
6406 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
6407 Register ImageReg = I.getOperand(2).getReg();
6408 Register SamplerReg = I.getOperand(3).getReg();
6409 Register CoordinateReg = I.getOperand(4).getReg();
6410 ImageOperands ImOps;
6411 ImOps.Compare = I.getOperand(5).getReg();
6412 if (I.getNumOperands() > 6)
6413 ImOps.Offset = I.getOperand(6).getReg();
6414 SPIRVTypeInst FloatTy = GR.getOrCreateSPIRVFloatType(32, I, TII);
6415 ImOps.Lod = GR.getOrCreateConstFP(APFloat(0.0f), I, FloatTy, TII);
6416 return generateSampleImage(ResVReg, ResType, ImageReg, SamplerReg,
6417 CoordinateReg, ImOps, I.getDebugLoc(), I);
6418}
6419
6420bool SPIRVInstructionSelector::selectGatherIntrinsic(Register &ResVReg,
6421 SPIRVTypeInst ResType,
6422 MachineInstr &I) const {
6423 Register ImageReg = I.getOperand(2).getReg();
6424 Register SamplerReg = I.getOperand(3).getReg();
6425 Register CoordinateReg = I.getOperand(4).getReg();
6426 SPIRVTypeInst ImageType = GR.getSPIRVTypeForVReg(ImageReg);
6427 assert(ImageType && ImageType->getOpcode() == SPIRV::OpTypeImage &&
6428 "ImageReg is not an image type.");
6429
6430 Register ComponentOrCompareReg;
6431 Register OffsetReg;
6432
6433 ComponentOrCompareReg = I.getOperand(5).getReg();
6434 OffsetReg = I.getOperand(6).getReg();
6435 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
6436 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
6437 if (!loadHandleBeforePosition(NewImageReg, ImageType, *ImageDef, I)) {
6438 return false;
6439 }
6440
6441 auto Dim = static_cast<SPIRV::Dim::Dim>(ImageType->getOperand(2).getImm());
6442 if (Dim != SPIRV::Dim::DIM_2D && Dim != SPIRV::Dim::DIM_Cube &&
6443 Dim != SPIRV::Dim::DIM_Rect) {
6444 I.emitGenericError(
6445 "Gather operations are only supported for 2D, Cube, and Rect images.");
6446 return false;
6447 }
6448
6449 auto *SamplerDef = cast<GIntrinsic>(getVRegDef(*MRI, SamplerReg));
6450 Register NewSamplerReg =
6451 MRI->createVirtualRegister(MRI->getRegClass(SamplerReg));
6452 if (!loadHandleBeforePosition(
6453 NewSamplerReg, GR.getSPIRVTypeForVReg(SamplerReg), *SamplerDef, I)) {
6454 return false;
6455 }
6456
6457 MachineIRBuilder MIRBuilder(I);
6458 SPIRVTypeInst SampledImageType =
6459 GR.getOrCreateOpTypeSampledImage(ImageType, MIRBuilder);
6460 Register SampledImageReg =
6461 MRI->createVirtualRegister(GR.getRegClass(SampledImageType));
6462
6463 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpSampledImage))
6464 .addDef(SampledImageReg)
6465 .addUse(GR.getSPIRVTypeID(SampledImageType))
6466 .addUse(NewImageReg)
6467 .addUse(NewSamplerReg)
6468 .constrainAllUses(TII, TRI, RBI);
6469
6470 auto IntrId = cast<GIntrinsic>(I).getIntrinsicID();
6471 bool IsGatherCmp = IntrId == Intrinsic::spv_resource_gather_cmp;
6472 unsigned Opcode =
6473 IsGatherCmp ? SPIRV::OpImageDrefGather : SPIRV::OpImageGather;
6474
6475 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode))
6476 .addDef(ResVReg)
6477 .addUse(GR.getSPIRVTypeID(ResType))
6478 .addUse(SampledImageReg)
6479 .addUse(CoordinateReg)
6480 .addUse(ComponentOrCompareReg);
6481
6482 uint32_t ImageOperands = 0;
6483 if (OffsetReg && !isScalarOrVectorIntConstantZero(OffsetReg)) {
6484 if (Dim == SPIRV::Dim::DIM_Cube) {
6485 I.emitGenericError(
6486 "Gather operations with offset are not supported for Cube images.");
6487 return false;
6488 }
6489 if (isConstReg(MRI, OffsetReg))
6490 ImageOperands |= SPIRV::ImageOperand::ConstOffset;
6491 else {
6492 ImageOperands |= SPIRV::ImageOperand::Offset;
6493 }
6494 }
6495
6496 if (ImageOperands != 0) {
6497 MIB.addImm(ImageOperands);
6498 if (ImageOperands &
6499 (SPIRV::ImageOperand::ConstOffset | SPIRV::ImageOperand::Offset))
6500 MIB.addUse(OffsetReg);
6501 }
6502
6503 MIB.constrainAllUses(TII, TRI, RBI);
6504 return true;
6505}
6506
6507bool SPIRVInstructionSelector::generateImageReadOrFetch(
6508 Register &ResVReg, SPIRVTypeInst ResType, Register ImageReg,
6509 Register IdxReg, DebugLoc Loc, MachineInstr &Pos,
6510 const ImageOperands *ImOps) const {
6511 SPIRVTypeInst ImageType = GR.getSPIRVTypeForVReg(ImageReg);
6512 assert(ImageType && ImageType->getOpcode() == SPIRV::OpTypeImage &&
6513 "ImageReg is not an image type.");
6514
6515 bool IsSignedInteger =
6516 sampledTypeIsSignedInteger(GR.getTypeForSPIRVType(ImageType));
6517 // Check if the "sampled" operand of the image type is 1.
6518 // https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#OpImageFetch
6519 auto SampledOp = ImageType->getOperand(6);
6520 bool IsFetch = (SampledOp.getImm() == 1);
6521
6522 auto AddOperands = [&](MachineInstrBuilder &MIB) {
6523 uint32_t ImageOperandsMask = 0;
6524 if (IsSignedInteger)
6525 ImageOperandsMask |= 0x1000; // SignExtend
6526
6527 if (IsFetch && ImOps) {
6528 if (ImOps->Lod)
6529 ImageOperandsMask |= SPIRV::ImageOperand::Lod;
6530 if (ImOps->Offset && !isScalarOrVectorIntConstantZero(*ImOps->Offset)) {
6531 if (isConstReg(MRI, *ImOps->Offset))
6532 ImageOperandsMask |= SPIRV::ImageOperand::ConstOffset;
6533 else
6534 ImageOperandsMask |= SPIRV::ImageOperand::Offset;
6535 }
6536 }
6537
6538 if (ImageOperandsMask != 0) {
6539 MIB.addImm(ImageOperandsMask);
6540 if (IsFetch && ImOps) {
6541 if (ImOps->Lod)
6542 MIB.addUse(*ImOps->Lod);
6543 if (ImOps->Offset &&
6544 (ImageOperandsMask &
6545 (SPIRV::ImageOperand::Offset | SPIRV::ImageOperand::ConstOffset)))
6546 MIB.addUse(*ImOps->Offset);
6547 }
6548 }
6549 };
6550
6551 uint64_t ResultSize = GR.getScalarOrVectorComponentCount(ResType);
6552
6553 // A wide element (e.g. int64_t2) is emulated with a narrower packed image, so
6554 // its sampled type is different from the result.
6555 SPIRVTypeInst SampledType =
6556 GR.getSPIRVTypeForVReg(ImageType->getOperand(1).getReg());
6557 bool IsPacked = SampledType != GR.getScalarOrVectorComponentType(ResType);
6558 SPIRVTypeInst ReadType =
6559 widenTypeToVec4(IsPacked ? SampledType : ResType, Pos);
6560 bool ReadTypeMatchesResult = ReadType == ResType;
6561 // Read directly into the result, or into a temporary to bitcast/extract.
6562 Register ReadReg = ReadTypeMatchesResult
6563 ? ResVReg
6564 : MRI->createVirtualRegister(GR.getRegClass(ReadType));
6565
6566 auto BMI =
6567 BuildMI(*Pos.getParent(), Pos, Loc,
6568 TII.get(IsFetch ? SPIRV::OpImageFetch : SPIRV::OpImageRead))
6569 .addDef(ReadReg)
6570 .addUse(GR.getSPIRVTypeID(ReadType))
6571 .addUse(ImageReg)
6572 .addUse(IdxReg);
6573 AddOperands(BMI);
6574 BMI.constrainAllUses(TII, TRI, RBI);
6575
6576 if (ReadTypeMatchesResult)
6577 return true;
6578
6579 if (IsPacked) {
6580 // Reinterpret the packed vector as the wide result type.
6581 BuildMI(*Pos.getParent(), Pos, Loc, TII.get(SPIRV::OpBitcast))
6582 .addDef(ResVReg)
6583 .addUse(GR.getSPIRVTypeID(ResType))
6584 .addUse(ReadReg)
6585 .constrainAllUses(TII, TRI, RBI);
6586 return true;
6587 }
6588
6589 if (ResultSize == 1) {
6590 BuildMI(*Pos.getParent(), Pos, Loc, TII.get(SPIRV::OpCompositeExtract))
6591 .addDef(ResVReg)
6592 .addUse(GR.getSPIRVTypeID(ResType))
6593 .addUse(ReadReg)
6594 .addImm(0)
6595 .constrainAllUses(TII, TRI, RBI);
6596 return true;
6597 }
6598 return extractSubvector(ResVReg, ResType, ReadReg, Pos);
6599}
6600
6601bool SPIRVInstructionSelector::selectResourceGetPointer(Register &ResVReg,
6602 SPIRVTypeInst ResType,
6603 MachineInstr &I) const {
6604 Register ResourcePtr = I.getOperand(2).getReg();
6605 SPIRVTypeInst RegType = GR.getSPIRVTypeForVReg(ResourcePtr, I.getMF());
6606 if (RegType->getOpcode() == SPIRV::OpTypeImage) {
6607 // For texel buffers, the index into the image is part of the OpImageRead or
6608 // OpImageWrite instructions. So we will do nothing in this case. This
6609 // intrinsic will be combined with the load or store when selecting the load
6610 // or store.
6611 return true;
6612 }
6613
6614 assert(ResType->getOpcode() == SPIRV::OpTypePointer);
6615 MachineIRBuilder MIRBuilder(I);
6616
6617 Register ZeroReg =
6618 buildZerosVal(GR.getOrCreateSPIRVIntegerType(32, I, TII), I);
6619 auto MIB =
6620 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpAccessChain))
6621 .addDef(ResVReg)
6622 .addUse(GR.getSPIRVTypeID(ResType))
6623 .addUse(ResourcePtr)
6624 .addUse(ZeroReg);
6625
6626 if (I.getNumExplicitOperands() > 3) {
6627 Register IndexReg = I.getOperand(3).getReg();
6628 MIB.addUse(IndexReg);
6629 }
6630 MIB.constrainAllUses(TII, TRI, RBI);
6631 return true;
6632}
6633
6634bool SPIRVInstructionSelector::selectPushConstantGetPointer(
6635 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
6636 MRI->replaceRegWith(ResVReg, I.getOperand(2).getReg());
6637 return true;
6638}
6639
6640bool SPIRVInstructionSelector::selectResourceNonUniformIndex(
6641 Register &ResVReg, SPIRVTypeInst ResType, MachineInstr &I) const {
6642 Register ObjReg = I.getOperand(2).getReg();
6643 if (!BuildCOPY(ResVReg, ObjReg, I))
6644 return false;
6645
6646 buildOpDecorate(ResVReg, I, TII, SPIRV::Decoration::NonUniformEXT, {});
6647 // Check for the registers that use the index marked as non-uniform
6648 // and recursively mark them as non-uniform.
6649 // Per the spec, it's necessary that the final argument used for
6650 // load/store/sample/atomic must be decorated, so we need to propagate the
6651 // decoration through access chains and copies.
6652 // https://docs.vulkan.org/samples/latest/samples/extensions/descriptor_indexing/README.html#_when_to_use_non_uniform_indexing_qualifier
6653 decorateUsesAsNonUniform(ResVReg);
6654 return true;
6655}
6656
6657void SPIRVInstructionSelector::decorateUsesAsNonUniform(
6658 Register &NonUniformReg) const {
6659 llvm::SmallVector<std::pair<Register, MachineInstr *>> WorkList = {
6660 {NonUniformReg, nullptr}};
6661 llvm::SmallSet<Register, 8> Visited;
6662 while (WorkList.size() > 0) {
6663 auto [CurrentReg, DefMI] = WorkList.pop_back_val();
6664
6665 if (!Visited.insert(CurrentReg).second)
6666 continue;
6667
6668 bool IsDecorated = false;
6669 for (MachineInstr &Use : MRI->use_instructions(CurrentReg)) {
6670 if (Use.getOpcode() == SPIRV::OpDecorate &&
6671 Use.getOperand(1).getImm() == SPIRV::Decoration::NonUniformEXT) {
6672 IsDecorated = true;
6673 continue;
6674 }
6675 // Check if the instruction has the result register and add it to the
6676 // worklist.
6677 if (Use.getOperand(0).isReg() && Use.getOperand(0).isDef()) {
6678 Register ResultReg = Use.getOperand(0).getReg();
6679 if (ResultReg == CurrentReg)
6680 continue;
6681 WorkList.push_back({ResultReg, &Use});
6682 }
6683 }
6684
6685 if (!IsDecorated) {
6686 MachineBasicBlock &MBB = *DefMI->getParent();
6687 MachineInstr &InsertPt =
6688 DefMI->isPHI() ? *MBB.getFirstNonPHI() : *DefMI->getNextNode();
6689 buildOpDecorate(CurrentReg, InsertPt, TII,
6690 SPIRV::Decoration::NonUniformEXT, {});
6691 }
6692 }
6693}
6694
6695bool SPIRVInstructionSelector::extractSubvector(
6696 Register &ResVReg, SPIRVTypeInst ResType, Register &ReadReg,
6697 MachineInstr &InsertionPoint) const {
6698 SPIRVTypeInst InputType = GR.getResultType(ReadReg);
6699 [[maybe_unused]] uint64_t InputSize =
6700 GR.getScalarOrVectorComponentCount(InputType);
6701 uint64_t ResultSize = GR.getScalarOrVectorComponentCount(ResType);
6702 [[maybe_unused]] bool IsLongVectorEXT =
6703 STI.canUseExtension(SPIRV::Extension::SPV_EXT_long_vector);
6704 assert((InputSize > 1 || IsLongVectorEXT) && "The input must be a vector.");
6705 assert((ResultSize > 1 || IsLongVectorEXT) && "The result must be a vector.");
6706 assert(ResultSize < InputSize &&
6707 "Cannot extract more element than there are in the input.");
6708 SmallVector<Register> ComponentRegisters;
6709 SPIRVTypeInst ScalarType = GR.getScalarOrVectorComponentType(ResType);
6710 const TargetRegisterClass *ScalarRegClass = GR.getRegClass(ScalarType);
6711 for (uint64_t I = 0; I < ResultSize; I++) {
6712 Register ComponentReg = MRI->createVirtualRegister(ScalarRegClass);
6713 BuildMI(*InsertionPoint.getParent(), InsertionPoint,
6714 InsertionPoint.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
6715 .addDef(ComponentReg)
6716 .addUse(ScalarType->getOperand(0).getReg())
6717 .addUse(ReadReg)
6718 .addImm(I)
6719 .constrainAllUses(TII, TRI, RBI);
6720 ComponentRegisters.emplace_back(ComponentReg);
6721 }
6722
6723 MachineInstrBuilder MIB = BuildMI(*InsertionPoint.getParent(), InsertionPoint,
6724 InsertionPoint.getDebugLoc(),
6725 TII.get(SPIRV::OpCompositeConstruct))
6726 .addDef(ResVReg)
6727 .addUse(GR.getSPIRVTypeID(ResType));
6728
6729 for (Register ComponentReg : ComponentRegisters)
6730 MIB.addUse(ComponentReg);
6731 MIB.constrainAllUses(TII, TRI, RBI);
6732 return true;
6733}
6734
6735bool SPIRVInstructionSelector::selectImageWriteIntrinsic(
6736 MachineInstr &I) const {
6737 // If the load of the image is in a different basic block, then
6738 // this will generate invalid code. A proper solution is to move
6739 // the OpLoad from selectHandleFromBinding here. However, to do
6740 // that we will need to change the return type of the intrinsic.
6741 // We will do that when we can, but for now trying to move forward with other
6742 // issues.
6743 Register ImageReg = I.getOperand(1).getReg();
6744 auto *ImageDef = cast<GIntrinsic>(getVRegDef(*MRI, ImageReg));
6745 Register NewImageReg = MRI->createVirtualRegister(MRI->getRegClass(ImageReg));
6746 if (!loadHandleBeforePosition(NewImageReg, GR.getSPIRVTypeForVReg(ImageReg),
6747 *ImageDef, I)) {
6748 return false;
6749 }
6750
6751 Register CoordinateReg = I.getOperand(2).getReg();
6752 Register DataReg = I.getOperand(3).getReg();
6753 assert(isVectorType(GR.getResultType(DataReg)));
6754 assert(GR.getScalarOrVectorComponentCount(GR.getResultType(DataReg)) == 4);
6755 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpImageWrite))
6756 .addUse(NewImageReg)
6757 .addUse(CoordinateReg)
6758 .addUse(DataReg)
6759 .constrainAllUses(TII, TRI, RBI);
6760 return true;
6761}
6762
6763Register SPIRVInstructionSelector::buildPointerToResource(
6764 SPIRVTypeInst SpirvResType, SPIRV::StorageClass::StorageClass SC,
6765 uint32_t Set, uint32_t Binding, uint32_t ArraySize, Register IndexReg,
6766 StringRef Name, MachineIRBuilder MIRBuilder) const {
6767 const Type *ResType = GR.getTypeForSPIRVType(SpirvResType);
6768 if (ArraySize == 1) {
6769 SPIRVTypeInst PtrType =
6770 GR.getOrCreateSPIRVPointerType(ResType, MIRBuilder, SC);
6771 assert(GR.getPointeeType(PtrType) == SpirvResType &&
6772 "SpirvResType did not have an explicit layout.");
6773 return GR.getOrCreateGlobalVariableWithBinding(PtrType, Set, Binding, Name,
6774 MIRBuilder);
6775 }
6776
6777 const Type *VarType = ArrayType::get(const_cast<Type *>(ResType), ArraySize);
6778 SPIRVTypeInst VarPointerType =
6779 GR.getOrCreateSPIRVPointerType(VarType, MIRBuilder, SC);
6780 Register VarReg = GR.getOrCreateGlobalVariableWithBinding(
6781 VarPointerType, Set, Binding, Name, MIRBuilder);
6782
6783 SPIRVTypeInst ResPointerType =
6784 GR.getOrCreateSPIRVPointerType(ResType, MIRBuilder, SC);
6785 Register AcReg = MRI->createVirtualRegister(GR.getRegClass(ResPointerType));
6786
6787 MIRBuilder.buildInstr(SPIRV::OpAccessChain)
6788 .addDef(AcReg)
6789 .addUse(GR.getSPIRVTypeID(ResPointerType))
6790 .addUse(VarReg)
6791 .addUse(IndexReg);
6792
6793 return AcReg;
6794}
6795
6796bool SPIRVInstructionSelector::selectFirstBitSet16(
6797 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I,
6798 unsigned ExtendOpcode, unsigned BitSetOpcode) const {
6799 Register ExtReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
6800 if (!selectOpWithSrcs(ExtReg, ResType, I, {I.getOperand(2).getReg()},
6801 ExtendOpcode))
6802 return false;
6803
6804 return selectFirstBitSet32(ResVReg, ResType, I, ExtReg, BitSetOpcode);
6805}
6806
6807bool SPIRVInstructionSelector::selectFirstBitSet32(
6808 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I, Register SrcReg,
6809 unsigned BitSetOpcode) const {
6810 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
6811 .addDef(ResVReg)
6812 .addUse(GR.getSPIRVTypeID(ResType))
6813 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
6814 .addImm(BitSetOpcode)
6815 .addUse(SrcReg)
6816 .constrainAllUses(TII, TRI, RBI);
6817 return true;
6818}
6819
6820bool SPIRVInstructionSelector::selectFirstBitSet64(
6821 Register ResVReg, SPIRVTypeInst ResType, MachineInstr &I, Register SrcReg,
6822 unsigned BitSetOpcode, bool SwapPrimarySide) const {
6823 unsigned ComponentCount = GR.getScalarOrVectorComponentCount(ResType);
6824 SPIRVTypeInst BaseType = GR.retrieveScalarOrVectorIntType(ResType);
6825 bool ZeroAsNull = !STI.isShader();
6826 Register ConstIntZero =
6827 GR.getOrCreateConstInt(0, I, BaseType, TII, ZeroAsNull);
6828 Register ConstIntOne =
6829 GR.getOrCreateConstInt(1, I, BaseType, TII, ZeroAsNull);
6830
6831 // SPIRV doesn't support vectors with more than 4 components. Since the
6832 // algoritm below converts i64 -> i32x2 and i64x4 -> i32x8 it can only
6833 // operate on vectors with 2 or less components. When largers vectors are
6834 // seen. Split them, recurse, then recombine them.
6835 // TODO: handle the case where SPV_EXT_long_vector is enabled.
6836 if (ComponentCount > 2) {
6837 auto Func = [this, SwapPrimarySide](Register ResVReg, SPIRVTypeInst ResType,
6838 MachineInstr &I, Register SrcReg,
6839 unsigned Opcode) -> bool {
6840 return this->selectFirstBitSet64(ResVReg, ResType, I, SrcReg, Opcode,
6841 SwapPrimarySide);
6842 };
6843
6844 return handle64BitOverflow(ResVReg, ResType, I, SrcReg, BitSetOpcode, Func);
6845 }
6846
6847 // 1. Split int64 into 2 pieces using a bitcast
6848 MachineIRBuilder MIRBuilder(I);
6849 SPIRVTypeInst PostCastType = GR.getOrCreateSPIRVVectorType(
6850 BaseType, 2 * ComponentCount, MIRBuilder, false);
6851 Register BitcastReg =
6852 MRI->createVirtualRegister(GR.getRegClass(PostCastType));
6853
6854 if (!selectOpWithSrcs(BitcastReg, PostCastType, I, {SrcReg},
6855 SPIRV::OpBitcast))
6856 return false;
6857
6858 // 2. Find the first set bit from the primary side for all the pieces in #1
6859 Register FBSReg = MRI->createVirtualRegister(GR.getRegClass(PostCastType));
6860 if (!selectFirstBitSet32(FBSReg, PostCastType, I, BitcastReg, BitSetOpcode))
6861 return false;
6862
6863 // 3. Split result vector into high bits and low bits
6864 Register HighReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
6865 Register LowReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
6866
6867 bool IsScalarRes = !isVectorType(ResType);
6868 if (IsScalarRes) {
6869 // if scalar do a vector extract
6870 if (!selectOpWithSrcs(HighReg, ResType, I, {FBSReg, ConstIntOne},
6871 SPIRV::OpVectorExtractDynamic))
6872 return false;
6873 if (!selectOpWithSrcs(LowReg, ResType, I, {FBSReg, ConstIntZero},
6874 SPIRV::OpVectorExtractDynamic))
6875 return false;
6876 } else {
6877 // if vector do a shufflevector
6878 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
6879 TII.get(SPIRV::OpVectorShuffle))
6880 .addDef(HighReg)
6881 .addUse(GR.getSPIRVTypeID(ResType))
6882 .addUse(FBSReg)
6883 // Per the spec, repeat the vector if only one vec is needed
6884 .addUse(FBSReg);
6885
6886 // high bits are stored in even natural indexes. Extract them from FBSReg
6887 for (unsigned J = 1; J < ComponentCount * 2; J += 2) {
6888 MIB.addImm(J);
6889 }
6890
6891 MIB.constrainAllUses(TII, TRI, RBI);
6892
6893 MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(),
6894 TII.get(SPIRV::OpVectorShuffle))
6895 .addDef(LowReg)
6896 .addUse(GR.getSPIRVTypeID(ResType))
6897 .addUse(FBSReg)
6898 // Per the spec, repeat the vector if only one vec is needed
6899 .addUse(FBSReg);
6900
6901 // low bits are stored in odd natural indices. Extract them from FBSReg
6902 for (unsigned J = 0; J < ComponentCount * 2; J += 2) {
6903 MIB.addImm(J);
6904 }
6905 MIB.constrainAllUses(TII, TRI, RBI);
6906 }
6907
6908 // 4. Check the result. When primary bits == -1 use secondary, otherwise use
6909 // primary
6910 SPIRVTypeInst BoolType = GR.getOrCreateSPIRVBoolType(I, TII);
6911 Register NegOneReg;
6912 Register Reg0;
6913 Register Reg32;
6914 unsigned SelectOp;
6915 unsigned AddOp;
6916
6917 if (IsScalarRes) {
6918 NegOneReg =
6919 GR.getOrCreateConstInt((unsigned)-1, I, ResType, TII, ZeroAsNull);
6920 Reg0 = GR.getOrCreateConstInt(0, I, ResType, TII, ZeroAsNull);
6921 Reg32 = GR.getOrCreateConstInt(32, I, ResType, TII, ZeroAsNull);
6922 SelectOp = SPIRV::OpSelectSISCond;
6923 AddOp = SPIRV::OpIAddS;
6924 } else {
6925 BoolType = GR.getOrCreateSPIRVVectorType(BoolType, ComponentCount,
6926 MIRBuilder, false);
6927 NegOneReg =
6928 GR.getOrCreateConstVector((unsigned)-1, I, ResType, TII, ZeroAsNull);
6929 Reg0 = GR.getOrCreateConstVector(0, I, ResType, TII, ZeroAsNull);
6930 Reg32 = GR.getOrCreateConstVector(32, I, ResType, TII, ZeroAsNull);
6931 SelectOp = SPIRV::OpSelectVIVCond;
6932 AddOp = SPIRV::OpIAddV;
6933 }
6934
6935 Register PrimaryReg = HighReg;
6936 Register SecondaryReg = LowReg;
6937 Register RegPrimaryOffset = Reg32;
6938 Register RegSecondaryOffset = Reg0;
6939
6940 // By default the emitted opcodes check for the set bit from the MSB side.
6941 // Setting SwapPrimarySide checks the set bit from the LSB side
6942 if (SwapPrimarySide) {
6943 PrimaryReg = LowReg;
6944 SecondaryReg = HighReg;
6945 RegPrimaryOffset = Reg0;
6946 RegSecondaryOffset = Reg32;
6947 }
6948
6949 Register RegSecondaryHasVal =
6950 MRI->createVirtualRegister(GR.getRegClass(BoolType));
6951 if (!selectOpWithSrcs(RegSecondaryHasVal, BoolType, I,
6952 {SecondaryReg, NegOneReg}, SPIRV::OpINotEqual))
6953 return false;
6954
6955 Register RegPrimaryHasVal =
6956 MRI->createVirtualRegister(GR.getRegClass(BoolType));
6957 if (!selectOpWithSrcs(RegPrimaryHasVal, BoolType, I, {PrimaryReg, NegOneReg},
6958 SPIRV::OpINotEqual))
6959 return false;
6960
6961 // Pass 1: seed with secondary (lower-priority fallback)
6962 // ReturnBits = secondaryHasVal ? SecondaryBits : -1
6963 // Add = secondaryHasVal ? SecondaryOffset : 0
6964 Register RegReturnBits = MRI->createVirtualRegister(GR.getRegClass(ResType));
6965 if (!selectOpWithSrcs(RegReturnBits, ResType, I,
6966 {RegSecondaryHasVal, SecondaryReg, NegOneReg},
6967 SelectOp))
6968 return false;
6969
6970 Register RegAdd;
6971 if (SwapPrimarySide) {
6972 RegAdd = MRI->createVirtualRegister(GR.getRegClass(ResType));
6973 if (!selectOpWithSrcs(RegAdd, ResType, I,
6974 {RegSecondaryHasVal, RegSecondaryOffset, Reg0},
6975 SelectOp))
6976 return false;
6977 } else {
6978 RegAdd = Reg0;
6979 }
6980
6981 // Pass 2: override with primary (higher priority) if it has a valid result
6982 // ReturnBits2 = primaryHasVal ? PrimaryBits : ReturnBits
6983 // Add2 = primaryHasVal ? PrimaryOffset : Add
6984 Register RegReturnBits2 = MRI->createVirtualRegister(GR.getRegClass(ResType));
6985 if (!selectOpWithSrcs(RegReturnBits2, ResType, I,
6986 {RegPrimaryHasVal, PrimaryReg, RegReturnBits},
6987 SelectOp))
6988 return false;
6989
6990 Register RegAdd2 = MRI->createVirtualRegister(GR.getRegClass(ResType));
6991 if (!selectOpWithSrcs(RegAdd2, ResType, I,
6992 {RegPrimaryHasVal, RegPrimaryOffset, RegAdd}, SelectOp))
6993 return false;
6994
6995 return selectOpWithSrcs(ResVReg, ResType, I, {RegReturnBits2, RegAdd2},
6996 AddOp);
6997}
6998
6999bool SPIRVInstructionSelector::selectFirstBitHigh(Register ResVReg,
7000 SPIRVTypeInst ResType,
7001 MachineInstr &I,
7002 bool IsSigned) const {
7003 // FindUMsb and FindSMsb intrinsics only support 32 bit integers
7004 Register OpReg = I.getOperand(2).getReg();
7005 SPIRVTypeInst OpType = GR.getSPIRVTypeForVReg(OpReg);
7006 // zero or sign extend
7007 unsigned ExtendOpcode = IsSigned ? SPIRV::OpSConvert : SPIRV::OpUConvert;
7008 unsigned BitSetOpcode = IsSigned ? GL::FindSMsb : GL::FindUMsb;
7009
7010 switch (GR.getScalarOrVectorBitWidth(OpType)) {
7011 case 16:
7012 return selectFirstBitSet16(ResVReg, ResType, I, ExtendOpcode, BitSetOpcode);
7013 case 32:
7014 return selectFirstBitSet32(ResVReg, ResType, I, OpReg, BitSetOpcode);
7015 case 64:
7016 return selectFirstBitSet64(ResVReg, ResType, I, OpReg, BitSetOpcode,
7017 /*SwapPrimarySide=*/false);
7018 default:
7019 return diagnoseUnsupported(
7020 I,
7021 "spv_firstbituhigh and spv_firstbitshigh only support 16,32,64 bits.");
7022 }
7023}
7024
7025bool SPIRVInstructionSelector::selectFirstBitLow(Register ResVReg,
7026 SPIRVTypeInst ResType,
7027 MachineInstr &I) const {
7028 // FindILsb intrinsic only supports 32 bit integers
7029 Register OpReg = I.getOperand(2).getReg();
7030 SPIRVTypeInst OpType = GR.getSPIRVTypeForVReg(OpReg);
7031 // OpUConvert treats the operand bits as an unsigned i16 and zero extends it
7032 // to an unsigned i32. As this leaves all the least significant bits unchanged
7033 // so the first set bit from the LSB side doesn't change.
7034 unsigned ExtendOpcode = SPIRV::OpUConvert;
7035 unsigned BitSetOpcode = GL::FindILsb;
7036
7037 switch (GR.getScalarOrVectorBitWidth(OpType)) {
7038 case 16:
7039 return selectFirstBitSet16(ResVReg, ResType, I, ExtendOpcode, BitSetOpcode);
7040 case 32:
7041 return selectFirstBitSet32(ResVReg, ResType, I, OpReg, BitSetOpcode);
7042 case 64:
7043 return selectFirstBitSet64(ResVReg, ResType, I, OpReg, BitSetOpcode,
7044 /*SwapPrimarySide=*/true);
7045 default:
7046 return diagnoseUnsupported(I,
7047 "spv_firstbitlow only supports 16,32,64 bits.");
7048 }
7049}
7050
7051bool SPIRVInstructionSelector::selectAllocaArray(Register ResVReg,
7052 SPIRVTypeInst ResType,
7053 MachineInstr &I) const {
7054 // there was an allocation size parameter to the allocation instruction
7055 // that is not 1
7056 MachineBasicBlock &BB = *I.getParent();
7057 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpVariableLengthArrayINTEL))
7058 .addDef(ResVReg)
7059 .addUse(GR.getSPIRVTypeID(ResType))
7060 .addUse(I.getOperand(2).getReg())
7061 .constrainAllUses(TII, TRI, RBI);
7062 if (!STI.isShader()) {
7063 unsigned Alignment = I.getOperand(3).getImm();
7064 buildOpDecorate(ResVReg, I, TII, SPIRV::Decoration::Alignment, {Alignment});
7065 }
7066 return true;
7067}
7068
7069// Returns true iff `Ty` is a concrete SPIR-V type per the SPV_KHR_abort
7070// definition: a numerical scalar (int/float), a (physical) pointer, a vector,
7071// matrix or any aggregate (array/struct) recursively containing only such
7072// types. OpTypeBool, OpTypeVoid, opaque handles and similar abstract
7073// non-concrete types are rejected.
7074static bool isConcreteSPIRVType(SPIRVTypeInst Ty,
7075 const SPIRVGlobalRegistry &GR) {
7076 SmallVector<SPIRVTypeInst, 4> Worklist{Ty};
7077 while (!Worklist.empty()) {
7078 SPIRVTypeInst T = Worklist.pop_back_val();
7079 switch (T->getOpcode()) {
7080 case SPIRV::OpTypeInt:
7081 case SPIRV::OpTypeFloat:
7082 case SPIRV::OpTypePointer:
7083 break;
7084 case SPIRV::OpTypeVector:
7085 case SPIRV::OpTypeVectorIdEXT:
7086 case SPIRV::OpTypeMatrix:
7087 case SPIRV::OpTypeArray: {
7088 Register OperandReg = T->getOperand(1).getReg();
7089 SPIRVTypeInst ElementT = GR.getSPIRVTypeForVReg(OperandReg);
7090 Worklist.push_back(ElementT);
7091 } break;
7092 case SPIRV::OpTypeStruct:
7093 for (unsigned Idx = 1, E = T->getNumOperands(); Idx < E; ++Idx) {
7094 Register OperandReg = T->getOperand(Idx).getReg();
7095 SPIRVTypeInst ElementT = GR.getSPIRVTypeForVReg(OperandReg);
7096 Worklist.push_back(ElementT);
7097 }
7098 break;
7099 default:
7100 return false;
7101 }
7102 }
7103 return true;
7104}
7105
7106static bool containsStorageBufferPointer(SPIRVTypeInst Ty,
7107 const SPIRVGlobalRegistry &GR,
7108 SmallSet<Register, 8> &Visited) {
7109 Register TypeReg = Ty->getOperand(0).getReg();
7110 if (!Visited.insert(TypeReg).second)
7111 return false;
7112
7113 switch (Ty->getOpcode()) {
7114 case SPIRV::OpTypePointer:
7115 if (Ty->getOperand(1).getImm() == SPIRV::StorageClass::StorageBuffer)
7116 return true;
7117 return containsStorageBufferPointer(
7118 GR.getSPIRVTypeForVReg(Ty->getOperand(2).getReg()), GR, Visited);
7119 case SPIRV::OpTypeArray:
7120 case SPIRV::OpTypeRuntimeArray:
7121 return containsStorageBufferPointer(
7122 GR.getSPIRVTypeForVReg(Ty->getOperand(1).getReg()), GR, Visited);
7123 case SPIRV::OpTypeStruct:
7124 for (unsigned I = 1; I < Ty->getNumOperands(); ++I)
7125 if (containsStorageBufferPointer(
7126 GR.getSPIRVTypeForVReg(Ty->getOperand(I).getReg()), GR, Visited))
7127 return true;
7128 return false;
7129 default:
7130 return false;
7131 }
7132}
7133
7134static bool containsStorageBufferPointer(SPIRVTypeInst Ty,
7135 const SPIRVGlobalRegistry &GR) {
7136 SmallSet<Register, 8> Visited;
7137 return containsStorageBufferPointer(Ty, GR, Visited);
7138}
7139
7140bool SPIRVInstructionSelector::selectAbort(MachineInstr &I) const {
7141 assert(I.getNumExplicitOperands() == 2);
7142
7143 Register MsgReg = I.getOperand(1).getReg();
7144 SPIRVTypeInst MsgType = GR.getSPIRVTypeForVReg(MsgReg);
7145 assert(MsgType && "Message argument of llvm.spv.abort has no SPIR-V type");
7146
7147 if (!isConcreteSPIRVType(MsgType, GR))
7148 return diagnoseUnsupported(
7149 I,
7150 "llvm.spv.abort message type must be a concrete SPIR-V type (numerical "
7151 "scalar, pointer, vector, matrix, or aggregate of such types)");
7152
7153 MachineBasicBlock &BB = *I.getParent();
7154 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpAbortKHR))
7155 .addUse(GR.getSPIRVTypeID(MsgType))
7156 .addUse(MsgReg)
7157 .constrainAllUses(TII, TRI, RBI);
7158 return true;
7159}
7160
7161bool SPIRVInstructionSelector::selectTrap(MachineInstr &I) const {
7162 // When the SPV_KHR_abort extension is disabled, drop the G_TRAP and
7163 // G_UBSANTRAP silently.
7164 if (!STI.canUseExtension(SPIRV::Extension::SPV_KHR_abort))
7165 return true;
7166
7167 // Use the 32-bit integer constant for the abort "message" argument:
7168 // - G_UBSANTRAP operand is zero-extended to 32 bits.
7169 // - "All ones" constant is used for G_TRAP.
7170 uint32_t MsgVal = ~0u;
7171 if (I.getOpcode() == TargetOpcode::G_UBSANTRAP)
7172 MsgVal = static_cast<uint32_t>(I.getOperand(0).getImm());
7173
7174 SPIRVTypeInst MsgType = GR.getOrCreateSPIRVIntegerType(32, I, TII);
7175 Register MsgReg = buildI32ConstantInEntryBlock(MsgVal, I, MsgType);
7176
7177 MachineBasicBlock &BB = *I.getParent();
7178 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpAbortKHR))
7179 .addUse(GR.getSPIRVTypeID(MsgType))
7180 .addUse(MsgReg)
7181 .constrainAllUses(TII, TRI, RBI);
7182 return true;
7183}
7184
7185bool SPIRVInstructionSelector::selectFrameIndex(Register ResVReg,
7186 SPIRVTypeInst ResType,
7187 MachineInstr &I) const {
7188 // Change order of instructions if needed: all OpVariable instructions in a
7189 // function must be the first instructions in the first block
7190 auto It = getOpVariableMBBIt(*I.getMF());
7191
7192 // Pointers to opaque types stay typed even with the extension on, so emit the
7193 // untyped variant only when the result is actually an untyped pointer.
7194 bool UseUntypedPointers =
7195 ResType->getOpcode() == SPIRV::OpTypeUntypedPointerKHR;
7196 unsigned Opcode =
7197 UseUntypedPointers ? SPIRV::OpUntypedVariableKHR : SPIRV::OpVariable;
7198
7199 if (!UseUntypedPointers && containsStorageBufferPointer(ResType, GR)) {
7200 MachineIRBuilder MIRBuilder(I);
7201 if (!STI.isAtLeastSPIRVVer(VersionTuple(1, 3)))
7202 MIRBuilder.buildInstr(SPIRV::OpExtension)
7203 .addImm(SPIRV::Extension::SPV_KHR_variable_pointers);
7204 MIRBuilder.buildInstr(SPIRV::OpCapability)
7205 .addImm(SPIRV::Capability::VariablePointersStorageBuffer);
7206 }
7207
7208 auto MIB = BuildMI(*It->getParent(), It, It->getDebugLoc(), TII.get(Opcode))
7209 .addDef(ResVReg)
7210 .addUse(GR.getSPIRVTypeID(ResType))
7211 .addImm(static_cast<uint32_t>(SPIRV::StorageClass::Function));
7212
7213 // OpUntypedVariableKHR takes an extra Data Type operand right after the
7214 // storage class.
7215 if (UseUntypedPointers) {
7216 // Get the element type that was stored when processing spv_assign_ptr_type.
7217 SPIRVTypeInst DataType = GR.getUntypedPtrElementType(ResVReg);
7218 if (!DataType)
7219 return diagnoseUnsupported(
7220 I, "could not deduce the data type of an untyped variable");
7221 MIB.addUse(GR.getSPIRVTypeID(DataType));
7222 }
7223 MIB.constrainAllUses(TII, TRI, RBI);
7224
7225 if (!STI.isShader()) {
7226 unsigned Alignment = I.getOperand(2).getImm();
7227 buildOpDecorate(ResVReg, *It, TII, SPIRV::Decoration::Alignment,
7228 {Alignment});
7229 }
7230 return true;
7231}
7232
7233bool SPIRVInstructionSelector::selectBranch(MachineInstr &I) const {
7234 // InstructionSelector walks backwards through the instructions. We can use
7235 // both a G_BR and a G_BRCOND to create an OpBranchConditional. We hit G_BR
7236 // first, so can generate an OpBranchConditional here. If there is no
7237 // G_BRCOND, we just use OpBranch for a regular unconditional branch.
7238 const MachineInstr *PrevI = I.getPrevNode();
7239 MachineBasicBlock &MBB = *I.getParent();
7240 if (PrevI != nullptr && PrevI->getOpcode() == TargetOpcode::G_BRCOND) {
7241 BuildMI(MBB, I, I.getDebugLoc(), TII.get(SPIRV::OpBranchConditional))
7242 .addUse(PrevI->getOperand(0).getReg())
7243 .addMBB(PrevI->getOperand(1).getMBB())
7244 .addMBB(I.getOperand(0).getMBB())
7245 .constrainAllUses(TII, TRI, RBI);
7246 return true;
7247 }
7248 BuildMI(MBB, I, I.getDebugLoc(), TII.get(SPIRV::OpBranch))
7249 .addMBB(I.getOperand(0).getMBB())
7250 .constrainAllUses(TII, TRI, RBI);
7251 return true;
7252}
7253
7254bool SPIRVInstructionSelector::selectBranchCond(MachineInstr &I) const {
7255 // InstructionSelector walks backwards through the instructions. For an
7256 // explicit conditional branch with no fallthrough, we use both a G_BR and a
7257 // G_BRCOND to create an OpBranchConditional. We should hit G_BR first, and
7258 // generate the OpBranchConditional in selectBranch above.
7259 //
7260 // If an OpBranchConditional has been generated, we simply return, as the work
7261 // is alread done. If there is no OpBranchConditional, LLVM must be relying on
7262 // implicit fallthrough to the next basic block, so we need to create an
7263 // OpBranchConditional with an explicit "false" argument pointing to the next
7264 // basic block that LLVM would fall through to.
7265 const MachineInstr *NextI = I.getNextNode();
7266 // Check if this has already been successfully selected.
7267 if (NextI != nullptr && NextI->getOpcode() == SPIRV::OpBranchConditional)
7268 return true;
7269 // Must be relying on implicit block fallthrough, so generate an
7270 // OpBranchConditional with the "next" basic block as the "false" target.
7271 MachineBasicBlock &MBB = *I.getParent();
7272 unsigned NextMBBNum = MBB.getNextNode()->getNumber();
7273 MachineBasicBlock *NextMBB = I.getMF()->getBlockNumbered(NextMBBNum);
7274 BuildMI(MBB, I, I.getDebugLoc(), TII.get(SPIRV::OpBranchConditional))
7275 .addUse(I.getOperand(0).getReg())
7276 .addMBB(I.getOperand(1).getMBB())
7277 .addMBB(NextMBB)
7278 .constrainAllUses(TII, TRI, RBI);
7279 return true;
7280}
7281
7282bool SPIRVInstructionSelector::selectPhi(Register ResVReg,
7283 MachineInstr &I) const {
7284 auto MIB =
7285 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(TargetOpcode::PHI))
7286 .addDef(ResVReg);
7287 const unsigned NumOps = I.getNumOperands();
7288 for (unsigned i = 1; i < NumOps; i += 2) {
7289 MIB.addUse(I.getOperand(i + 0).getReg());
7290 MIB.addMBB(I.getOperand(i + 1).getMBB());
7291 }
7292 MIB.constrainAllUses(TII, TRI, RBI);
7293 return true;
7294}
7295
7296bool SPIRVInstructionSelector::selectGlobalValue(
7297 Register ResVReg, MachineInstr &I, const MachineInstr *Init) const {
7298 // FIXME: don't use MachineIRBuilder here, replace it with BuildMI.
7299 MachineIRBuilder MIRBuilder(I);
7300 const GlobalValue *GV = I.getOperand(1).getGlobal();
7301 Type *GVType = toTypedPointer(GR.getDeducedGlobalValueType(GV));
7302
7303 std::string GlobalIdent;
7304 if (!GV->hasName()) {
7305 unsigned &ID = UnnamedGlobalIDs[GV];
7306 if (ID == 0)
7307 ID = UnnamedGlobalIDs.size();
7308 GlobalIdent = "__unnamed_" + Twine(ID).str();
7309 } else {
7310 GlobalIdent = GV->getName();
7311 }
7312
7313 // Behaviour of functions as operands depends on availability of the
7314 // corresponding extension (SPV_INTEL_function_pointers):
7315 // - If there is an extension to operate with functions as operands:
7316 // We create a proper constant operand and evaluate a correct type for a
7317 // function pointer.
7318 // - Without the required extension:
7319 // We have functions as operands in tests with blocks of instruction e.g. in
7320 // transcoding/global_block.ll. These operands are not used and should be
7321 // substituted by zero constants. Their type is expected to be always
7322 // OpTypePointer Function %uchar.
7323 if (isa<Function>(GV)) {
7324 const Constant *ConstVal = GV;
7325 MachineBasicBlock &BB = *I.getParent();
7326 Register NewReg = GR.find(ConstVal, GR.CurMF);
7327 if (!NewReg.isValid()) {
7328 const Function *GVFun =
7329 STI.canUseExtension(SPIRV::Extension::SPV_INTEL_function_pointers)
7330 ? dyn_cast<Function>(GV)
7331 : nullptr;
7332 SPIRVTypeInst ResType = GR.getOrCreateSPIRVPointerType(
7333 GVType, I,
7334 GVFun ? SPIRV::StorageClass::CodeSectionINTEL
7335 : addressSpaceToStorageClass(GV->getAddressSpace(), STI));
7336 if (GVFun) {
7337 // References to a function via function pointers generate virtual
7338 // registers without a definition. We will resolve it later, during
7339 // module analysis stage.
7340 Register ResTypeReg = GR.getSPIRVTypeID(ResType);
7341 MachineRegisterInfo *MRI = MIRBuilder.getMRI();
7342 Register FuncVReg =
7343 MRI->createGenericVirtualRegister(GR.getRegType(ResType));
7344 MRI->setRegClass(FuncVReg, &SPIRV::pIDRegClass);
7345 GR.assignSPIRVTypeToVReg(ResType, FuncVReg, *GR.CurMF);
7346 MachineInstrBuilder MIB1 =
7347 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpUndef))
7348 .addDef(FuncVReg)
7349 .addUse(ResTypeReg);
7350 MachineInstrBuilder MIB2 =
7351 BuildMI(BB, I, I.getDebugLoc(),
7352 TII.get(SPIRV::OpConstantFunctionPointerINTEL))
7353 .addDef(ResVReg)
7354 .addUse(ResTypeReg)
7355 .addUse(FuncVReg);
7356 GR.add(ConstVal, MIB2);
7357 // mapping the function pointer to the used Function
7358 GR.recordFunctionPointer(&MIB2.getInstr()->getOperand(2), GVFun);
7359 GR.assignSPIRVTypeToVReg(ResType, ResVReg, *GR.CurMF);
7360 MIB1.constrainAllUses(TII, TRI, RBI);
7361 MIB2.constrainAllUses(TII, TRI, RBI);
7362 return true;
7363 }
7364 MachineInstrBuilder MIB3 =
7365 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpUndef))
7366 .addDef(ResVReg)
7367 .addUse(GR.getSPIRVTypeID(ResType));
7368 GR.add(ConstVal, MIB3);
7369 GR.recordFunctionPointer(&MIB3.getInstr()->getOperand(0),
7370 cast<Function>(GV));
7371 MIB3.constrainAllUses(TII, TRI, RBI);
7372 return true;
7373 }
7374 assert(NewReg != ResVReg);
7375 return BuildCOPY(ResVReg, NewReg, I);
7376 }
7377 auto GlobalVar = cast<GlobalVariable>(GV);
7378 assert(GlobalVar->getName() != "llvm.global.annotations");
7379
7380 // Skip empty declaration for GVs with initializers till we get the decl with
7381 // passed initializer.
7382 if (hasInitializer(GlobalVar) && !Init)
7383 return true;
7384
7385 const std::optional<SPIRV::LinkageType::LinkageType> LnkType =
7386 getSpirvLinkageTypeFor(STI, *GV);
7387
7388 if (LnkType && *LnkType == SPIRV::LinkageType::Import)
7389 Init = nullptr;
7390
7391 const unsigned AddrSpace = GV->getAddressSpace();
7392 SPIRV::StorageClass::StorageClass StorageClass =
7393 addressSpaceToStorageClass(AddrSpace, STI);
7394 SPIRVTypeInst ResType =
7395 GR.getOrCreateSPIRVPointerType(GVType, I, StorageClass);
7396 Register Reg = GR.buildGlobalVariable(
7397 ResVReg, ResType, GlobalIdent, GV, StorageClass, Init,
7398 GlobalVar->isConstant(), LnkType, MIRBuilder, true);
7399 // TODO: For AMDGCN, we pipe externally_initialized through via
7400 // HostAccessINTEL, with ReadWrite (3) access, which is we then handle during
7401 // reverse translation. We should remove this once SPIR-V gains the ability to
7402 // express the concept.
7403 if (GlobalVar->isExternallyInitialized() &&
7404 STI.getTargetTriple().getVendor() == Triple::AMD) {
7405 constexpr unsigned ReadWriteINTEL = 3u;
7406 buildOpDecorate(Reg, MIRBuilder, SPIRV::Decoration::HostAccessINTEL,
7407 {ReadWriteINTEL});
7408 MachineInstrBuilder MIB(*MF, --MIRBuilder.getInsertPt());
7409 addStringImm(GV->getName(), MIB);
7410 }
7411 return Reg.isValid();
7412}
7413
7414bool SPIRVInstructionSelector::selectLog10(Register ResVReg,
7415 SPIRVTypeInst ResType,
7416 MachineInstr &I) const {
7417 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std)) {
7418 return selectExtInst(ResVReg, ResType, I, CL::log10);
7419 }
7420
7421 // There is no log10 instruction in the GLSL Extended Instruction set, so it
7422 // is implemented as:
7423 // log10(x) = log2(x) * (1 / log2(10))
7424 // = log2(x) * 0.30103
7425
7426 MachineIRBuilder MIRBuilder(I);
7427 MachineBasicBlock &BB = *I.getParent();
7428
7429 // Build log2(x).
7430 Register VarReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
7431 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
7432 .addDef(VarReg)
7433 .addUse(GR.getSPIRVTypeID(ResType))
7434 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::GLSL_std_450))
7435 .addImm(GL::Log2)
7436 .add(I.getOperand(1))
7437 .constrainAllUses(TII, TRI, RBI);
7438
7439 // Build 0.30103.
7440 assert(isVectorType(ResType) || ResType.isAnyTypeFloat());
7441 // TODO: Add matrix implementation once supported by the HLSL frontend.
7442 SPIRVTypeInst SpirvScalarType = GR.getScalarOrVectorComponentType(ResType);
7443 // The literal must match the precision of the scalar type, otherwise the
7444 // OpConstant will contain non-zero high-order bits and fail SPIR-V
7445 // validation when the type is narrower than 32 bits (e.g. half).
7446 APFloat ScaleVal(0.30103);
7447 bool LosesInfo;
7448 ScaleVal.convert(
7449 getZeroFP(GR.getTypeForSPIRVType(SpirvScalarType)).getSemantics(),
7450 APFloat::rmNearestTiesToEven, &LosesInfo);
7451 Register ScaleReg = GR.buildConstantFP(ScaleVal, MIRBuilder, SpirvScalarType);
7452
7453 // Multiply log2(x) by 0.30103 to get log10(x) result.
7454 auto Opcode =
7455 isVectorType(ResType) ? SPIRV::OpVectorTimesScalar : SPIRV::OpFMulS;
7456 BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
7457 .addDef(ResVReg)
7458 .addUse(GR.getSPIRVTypeID(ResType))
7459 .addUse(VarReg)
7460 .addUse(ScaleReg)
7461 .constrainAllUses(TII, TRI, RBI);
7462 return true;
7463}
7464
7465bool SPIRVInstructionSelector::selectFpowi(Register ResVReg,
7466 SPIRVTypeInst ResType,
7467 MachineInstr &I) const {
7468 // On OpenCL targets, pown(gentype x, intn n) maps directly.
7469 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std))
7470 return selectExtInst(ResVReg, ResType, I, CL::pown);
7471
7472 // On GLSL (Vulkan) targets, there is no integer-exponent power instruction.
7473 // Lower as: Pow(base, OpConvertSToF(exp)).
7474 if (STI.canUseExtInstSet(SPIRV::InstructionSet::GLSL_std_450)) {
7475 Register BaseReg = I.getOperand(1).getReg();
7476 Register ExpReg = I.getOperand(2).getReg();
7477 Register FloatExpReg = MRI->createVirtualRegister(GR.getRegClass(ResType));
7478 if (!selectOpWithSrcs(FloatExpReg, ResType, I, {ExpReg},
7479 SPIRV::OpConvertSToF))
7480 return false;
7481 return selectExtInst(ResVReg, ResType, I, GL::Pow,
7482 /*setMIFlags=*/true, /*useMISrc=*/false,
7483 {BaseReg, FloatExpReg});
7484 }
7485 return false;
7486}
7487
7488bool SPIRVInstructionSelector::selectModf(Register ResVReg,
7489 SPIRVTypeInst ResType,
7490 MachineInstr &I) const {
7491 // llvm.modf has a single arg --the number to be decomposed-- and returns a
7492 // struct { restype, restype }, while OpenCLLIB::modf has two args --the
7493 // number to be decomposed and a pointer--, returns the fractional part and
7494 // the integral part is stored in the pointer argument. Therefore, we can't
7495 // use directly the OpenCLLIB::modf intrinsic. However, we can do some
7496 // scaffolding to make it work. The idea is to create an alloca instruction
7497 // to get a ptr, pass this ptr to OpenCL::modf, and then load the value
7498 // from this ptr to place it in the struct. llvm.modf returns the fractional
7499 // part as the first element of the result, and the integral part as the
7500 // second element of the result.
7501
7502 // At this point, the return type is not a struct anymore, but rather two
7503 // independent elements of SPIRVResType. We can get each independent element
7504 // from I.getDefs() or I.getOperands().
7505 if (STI.canUseExtInstSet(SPIRV::InstructionSet::OpenCL_std)) {
7506 MachineIRBuilder MIRBuilder(I);
7507 SPIRVTypeInst FloatType =
7508 GR.getSPIRVTypeForVReg(I.getOperand(I.getNumExplicitDefs()).getReg());
7509 // Get pointer type for alloca variable.
7510 const SPIRVTypeInst PtrType = GR.getOrCreateSPIRVPointerType(
7511 FloatType, MIRBuilder, SPIRV::StorageClass::Function);
7512 // Create new register for the pointer type of alloca variable.
7513 Register PtrTyReg =
7514 MIRBuilder.getMRI()->createVirtualRegister(&SPIRV::iIDRegClass);
7515 MIRBuilder.getMRI()->setType(
7516 PtrTyReg,
7517 LLT::pointer(storageClassToAddressSpace(SPIRV::StorageClass::Function),
7518 GR.getPointerSize()));
7519
7520 // Assign SPIR-V type of the pointer type of the alloca variable to the
7521 // new register.
7522 GR.assignSPIRVTypeToVReg(PtrType, PtrTyReg, MIRBuilder.getMF());
7523 MachineBasicBlock::iterator VarPos = getOpVariableMBBIt(*I.getMF());
7524 MachineBasicBlock &EntryBB = I.getMF()->front();
7525 const bool IsUntyped =
7526 PtrType->getOpcode() == SPIRV::OpTypeUntypedPointerKHR;
7527 auto AllocaMIB =
7528 BuildMI(EntryBB, VarPos, I.getDebugLoc(),
7529 TII.get(IsUntyped ? SPIRV::OpUntypedVariableKHR
7530 : SPIRV::OpVariable))
7531 .addDef(PtrTyReg)
7532 .addUse(GR.getSPIRVTypeID(PtrType))
7533 .addImm(static_cast<uint32_t>(SPIRV::StorageClass::Function));
7534 if (IsUntyped)
7535 AllocaMIB.addUse(GR.getSPIRVTypeID(ResType)); // Data Type
7536 Register Variable = AllocaMIB->getOperand(0).getReg();
7537
7538 MachineBasicBlock &BB = *I.getParent();
7539 // Create the OpenCLLIB::modf instruction.
7540 auto MIB =
7541 BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpExtInst))
7542 .addDef(ResVReg)
7543 .addUse(GR.getSPIRVTypeID(FloatType))
7544 .addImm(static_cast<uint32_t>(SPIRV::InstructionSet::OpenCL_std))
7545 .addImm(CL::modf)
7546 .setMIFlags(I.getFlags())
7547 .add(I.getOperand(I.getNumExplicitDefs())) // Floating point value.
7548 .addUse(Variable); // Pointer to integral part.
7549 // Assign the integral part stored in the ptr to the second element of the
7550 // result.
7551 Register IntegralPartReg = I.getOperand(1).getReg();
7552 if (IntegralPartReg.isValid() && !MRI->use_nodbg_empty(IntegralPartReg)) {
7553 // Load the value from the pointer to integral part.
7554 auto LoadMIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpLoad))
7555 .addDef(IntegralPartReg)
7556 .addUse(GR.getSPIRVTypeID(FloatType))
7557 .addUse(Variable);
7558 LoadMIB.constrainAllUses(TII, TRI, RBI);
7559 }
7560
7561 MIB.constrainAllUses(TII, TRI, RBI);
7562 return true;
7563 } else if (STI.canUseExtInstSet(SPIRV::InstructionSet::GLSL_std_450)) {
7564 assert(false && "GLSL::Modf is deprecated.");
7565 // FIXME: GL::Modf is deprecated, use Modfstruct instead.
7566 return false;
7567 }
7568 return false;
7569}
7570
7571// Generate the instructions to load 3-element vector builtin input
7572// IDs/Indices.
7573// Like: GlobalInvocationId, LocalInvocationId, etc....
7574
7575bool SPIRVInstructionSelector::loadVec3BuiltinInputID(
7576 SPIRV::BuiltIn::BuiltIn BuiltInValue, Register ResVReg,
7577 SPIRVTypeInst ResType, MachineInstr &I) const {
7578 MachineIRBuilder MIRBuilder(I);
7579 const SPIRVTypeInst Vec3Ty =
7580 GR.getOrCreateSPIRVVectorType(ResType, 3, MIRBuilder, false);
7581 const SPIRVTypeInst PtrType = GR.getOrCreateSPIRVPointerType(
7582 Vec3Ty, MIRBuilder, SPIRV::StorageClass::Input);
7583
7584 // Create new register for the input ID builtin variable.
7585 Register NewRegister =
7586 MIRBuilder.getMRI()->createVirtualRegister(&SPIRV::iIDRegClass);
7587 MIRBuilder.getMRI()->setType(NewRegister, LLT::pointer(0, 64));
7588 GR.assignSPIRVTypeToVReg(PtrType, NewRegister, MIRBuilder.getMF());
7589
7590 // Build global variable with the necessary decorations for the input ID
7591 // builtin variable.
7592 Register Variable = GR.buildGlobalVariable(
7593 NewRegister, PtrType, getLinkStringForBuiltIn(BuiltInValue), nullptr,
7594 SPIRV::StorageClass::Input, nullptr, true, std::nullopt, MIRBuilder,
7595 false);
7596
7597 // Create new register for loading value.
7598 MachineRegisterInfo *MRI = MIRBuilder.getMRI();
7599 Register LoadedRegister = MRI->createVirtualRegister(&SPIRV::iIDRegClass);
7600 MIRBuilder.getMRI()->setType(LoadedRegister, LLT::pointer(0, 64));
7601 GR.assignSPIRVTypeToVReg(Vec3Ty, LoadedRegister, MIRBuilder.getMF());
7602
7603 // Load v3uint value from the global variable.
7604 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpLoad))
7605 .addDef(LoadedRegister)
7606 .addUse(GR.getSPIRVTypeID(Vec3Ty))
7607 .addUse(Variable);
7608
7609 // Get the input ID index. Expecting operand is a constant immediate value,
7610 // wrapped in a type assignment.
7611 assert(I.getOperand(2).isReg());
7612 const uint32_t ThreadId = foldImm(I.getOperand(2), MRI);
7613
7614 // Extract the input ID from the loaded vector value.
7615 MachineBasicBlock &BB = *I.getParent();
7616 auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpCompositeExtract))
7617 .addDef(ResVReg)
7618 .addUse(GR.getSPIRVTypeID(ResType))
7619 .addUse(LoadedRegister)
7620 .addImm(ThreadId);
7621 MIB.constrainAllUses(TII, TRI, RBI);
7622 return true;
7623}
7624
7625// Generate the instructions to load 32-bit integer builtin input IDs/Indices.
7626// Like LocalInvocationIndex
7627bool SPIRVInstructionSelector::loadBuiltinInputID(
7628 SPIRV::BuiltIn::BuiltIn BuiltInValue, Register ResVReg,
7629 SPIRVTypeInst ResType, MachineInstr &I) const {
7630 MachineIRBuilder MIRBuilder(I);
7631 const SPIRVTypeInst PtrType = GR.getOrCreateSPIRVPointerType(
7632 ResType, MIRBuilder, SPIRV::StorageClass::Input);
7633
7634 // Create new register for the input ID builtin variable.
7635 Register NewRegister =
7636 MIRBuilder.getMRI()->createVirtualRegister(GR.getRegClass(PtrType));
7637 MIRBuilder.getMRI()->setType(
7638 NewRegister,
7639 LLT::pointer(storageClassToAddressSpace(SPIRV::StorageClass::Input),
7640 GR.getPointerSize()));
7641 GR.assignSPIRVTypeToVReg(PtrType, NewRegister, MIRBuilder.getMF());
7642
7643 // Build global variable with the necessary decorations for the input ID
7644 // builtin variable.
7645 Register Variable = GR.buildGlobalVariable(
7646 NewRegister, PtrType, getLinkStringForBuiltIn(BuiltInValue), nullptr,
7647 SPIRV::StorageClass::Input, nullptr, true, std::nullopt, MIRBuilder,
7648 false);
7649
7650 // Load uint value from the global variable.
7651 auto MIB = BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(SPIRV::OpLoad))
7652 .addDef(ResVReg)
7653 .addUse(GR.getSPIRVTypeID(ResType))
7654 .addUse(Variable);
7655
7656 MIB.constrainAllUses(TII, TRI, RBI);
7657 return true;
7658}
7659
7660SPIRVTypeInst SPIRVInstructionSelector::widenTypeToVec4(SPIRVTypeInst Type,
7661 MachineInstr &I) const {
7662 MachineIRBuilder MIRBuilder(I);
7663 if (!isVectorType(Type))
7664 return GR.getOrCreateSPIRVVectorType(Type, 4, MIRBuilder, false);
7665
7666 if (GR.getScalarOrVectorComponentCount(Type) == 4)
7667 return Type;
7668
7669 SPIRVTypeInst ScalarType = GR.getScalarOrVectorComponentType(Type);
7670 return GR.getOrCreateSPIRVVectorType(ScalarType, 4, MIRBuilder, false);
7671}
7672
7673bool SPIRVInstructionSelector::loadHandleBeforePosition(
7674 Register &HandleReg, SPIRVTypeInst ResType, GIntrinsic &HandleDef,
7675 MachineInstr &Pos) const {
7676
7677 assert(HandleDef.getIntrinsicID() ==
7678 Intrinsic::spv_resource_handlefrombinding);
7679 uint32_t Set = foldImm(HandleDef.getOperand(2), MRI);
7680 uint32_t Binding = foldImm(HandleDef.getOperand(3), MRI);
7681 uint32_t ArraySize = foldImm(HandleDef.getOperand(4), MRI);
7682 Register IndexReg = HandleDef.getOperand(5).getReg();
7683 std::string Name =
7684 getStringValueFromReg(HandleDef.getOperand(6).getReg(), *MRI);
7685
7686 bool IsStructuredBuffer = ResType->getOpcode() == SPIRV::OpTypePointer;
7687 MachineIRBuilder MIRBuilder(HandleDef);
7688 SPIRVTypeInst VarType = ResType;
7689 SPIRV::StorageClass::StorageClass SC = SPIRV::StorageClass::UniformConstant;
7690
7691 if (IsStructuredBuffer) {
7692 VarType = GR.getPointeeType(ResType);
7693 SC = GR.getPointerStorageClass(ResType);
7694 }
7695
7696 // ArraySize 0 means an unbounded array and we need to set to required
7697 // capability.
7698 if (ArraySize == 0)
7699 MIRBuilder.buildInstr(SPIRV::OpCapability)
7700 .addImm(SPIRV::Capability::RuntimeDescriptorArrayEXT);
7701
7702 Register VarReg =
7703 buildPointerToResource(SPIRVTypeInst(VarType), SC, Set, Binding,
7704 ArraySize, IndexReg, Name, MIRBuilder);
7705
7706 // The handle for the buffer is the pointer to the resource. For an image, the
7707 // handle is the image object. So images get an extra load.
7708 uint32_t LoadOpcode =
7709 IsStructuredBuffer ? SPIRV::OpCopyObject : SPIRV::OpLoad;
7710 GR.assignSPIRVTypeToVReg(ResType, HandleReg, *Pos.getMF());
7711 BuildMI(*Pos.getParent(), Pos, HandleDef.getDebugLoc(), TII.get(LoadOpcode))
7712 .addDef(HandleReg)
7713 .addUse(GR.getSPIRVTypeID(ResType))
7714 .addUse(VarReg)
7715 .constrainAllUses(TII, TRI, RBI);
7716 return true;
7717}
7718
7719bool SPIRVInstructionSelector::errorIfInstrOutsideShader(
7720 MachineInstr &I) const {
7721 if (!STI.isShader())
7722 return diagnoseUnsupported(
7723 I, "this instruction is only supported in shaders.");
7724 return true;
7725}
7726
7727namespace llvm {
7728InstructionSelector *
7729createSPIRVInstructionSelector(const SPIRVTargetMachine &TM,
7730 const SPIRVSubtarget &Subtarget,
7731 const RegisterBankInfo &RBI) {
7732 return new SPIRVInstructionSelector(TM, Subtarget, RBI);
7733}
7734} // namespace llvm
UseMI
MachineInstrBuilder & UseMI
Definition AArch64ExpandPseudoInsts.cpp:125
DefMI
MachineInstrBuilder MachineInstrBuilder & DefMI
Definition AArch64ExpandPseudoInsts.cpp:126
GET_GLOBALISEL_PREDICATES_INIT
#define GET_GLOBALISEL_PREDICATES_INIT
GET_GLOBALISEL_TEMPORARIES_INIT
#define GET_GLOBALISEL_TEMPORARIES_INIT
assert
assert(UImm &&(UImm !=~static_cast< T >(0)) &&"Invalid immediate!")
Imm
unsigned Imm
Definition AMDGPUDisassembler.cpp:539
uint64_t
unsigned uint64_t
Definition AMDGPUDisassembler.cpp:540
APFloat.h
This file declares a class to represent arbitrary precision floating point values and provide a varie...
selectUnmergeValues
static bool selectUnmergeValues(MachineInstrBuilder &MIB, const ARMBaseInstrInfo &TII, MachineRegisterInfo &MRI, const TargetRegisterInfo &TRI, const RegisterBankInfo &RBI)
Definition ARMInstructionSelector.cpp:264
MBB
MachineBasicBlock & MBB
Definition ARMSLSHardening.cpp:71
DL
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
Definition ARMSLSHardening.cpp:73
SwapBits
static uint8_t SwapBits(uint8_t Val)
Definition BPFMCCodeEmitter.cpp:122
true
basic Basic Alias true
Definition BasicAliasAnalysis.cpp:2167
X
#define X(NUM, ENUM, NAME)
Definition ELF.h:857
E
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
Binding
DXIL Resource Implicit Binding
Definition DXILResourceImplicitBinding.cpp:185
DEBUG_TYPE
#define DEBUG_TYPE
Definition GenericCycleImpl.h:34
GenericMachineInstrs.h
Declares convenience wrapper classes for interpreting MachineInstr instances as specific generic oper...
TII
const HexagonInstrInfo * TII
Definition HexagonCopyToCombine.cpp:118
MI
IRTranslator LLVM IR MI
Definition IRTranslator.cpp:858
LLVMIntType
LLVMTypeRef LLVMIntType(unsigned NumBits)
Definition Core.cpp:742
NumOps
const size_t AbstractManglingParser< Derived, Alloc >::NumOps
Definition ItaniumDemangle.h:3474
Direction
Loop::LoopBounds::Direction Direction
Definition LoopInfo.cpp:253
F
#define F(x, y, z)
Definition MD5.cpp:54
I
#define I(x, y, z)
Definition MD5.cpp:57
Reg
Register Reg
Definition MachineSink.cpp:2127
TRI
Register const TargetRegisterInfo * TRI
Definition MachineSink.cpp:2128
Register
Promote Memory to Register
Definition Mem2Reg.cpp:110
T
#define T
Definition Mips16ISelLowering.cpp:282
T1
#define T1
Definition Mips16ISelLowering.cpp:281
II
uint64_t IntrinsicInst * II
Definition NVVMIntrRange.cpp:46
getName
static StringRef getName(Value *V)
Definition ProvenanceAnalysisEvaluator.cpp:20
getFCmpOpcode
static unsigned getFCmpOpcode(CmpInst::Predicate Pred, unsigned Size)
Definition RISCVInstructionSelector.cpp:2064
isConcreteSPIRVType
static bool isConcreteSPIRVType(SPIRVTypeInst Ty, const SPIRVGlobalRegistry &GR)
Definition SPIRVInstructionSelector.cpp:7074
getOneFP
static APFloat getOneFP(const Type *LLVMFloatTy)
Definition SPIRVInstructionSelector.cpp:3214
isUSMStorageClass
static bool isUSMStorageClass(SPIRV::StorageClass::StorageClass SC)
Definition SPIRVInstructionSelector.cpp:2888
isASCastInGVar
static bool isASCastInGVar(MachineRegisterInfo *MRI, Register ResVReg)
Definition SPIRVInstructionSelector.cpp:2899
mayApplyGenericSelection
static bool mayApplyGenericSelection(unsigned Opcode)
Definition SPIRVInstructionSelector.cpp:984
getZeroFP
static APFloat getZeroFP(const Type *LLVMFloatTy)
Definition SPIRVInstructionSelector.cpp:3198
ExtInstList
std::vector< std::pair< SPIRV::InstructionSet::InstructionSet, uint32_t > > ExtInstList
Definition SPIRVInstructionSelector.cpp:49
intrinsicHasSideEffects
static bool intrinsicHasSideEffects(Intrinsic::ID ID)
Definition SPIRVInstructionSelector.cpp:696
getBoolCmpOpcode
static unsigned getBoolCmpOpcode(unsigned PredNum)
Definition SPIRVInstructionSelector.cpp:3186
getICmpOpcode
static unsigned getICmpOpcode(unsigned PredNum)
Definition SPIRVInstructionSelector.cpp:3146
isOpcodeWithNoSideEffects
static bool isOpcodeWithNoSideEffects(unsigned Opcode)
Definition SPIRVInstructionSelector.cpp:763
addMemoryOperands
static void addMemoryOperands(MachineMemOperand *MemOp, MachineInstrBuilder &MIB, MachineIRBuilder &MIRBuilder, SPIRVGlobalRegistry &GR)
Definition SPIRVInstructionSelector.cpp:1972
containsStorageBufferPointer
static bool containsStorageBufferPointer(SPIRVTypeInst Ty, const SPIRVGlobalRegistry &GR, SmallSet< Register, 8 > &Visited)
Definition SPIRVInstructionSelector.cpp:7106
isConstReg
static bool isConstReg(MachineRegisterInfo *MRI, MachineInstr *OpDef)
Definition SPIRVInstructionSelector.cpp:627
getPtrCmpOpcode
static unsigned getPtrCmpOpcode(unsigned Pred)
Definition SPIRVInstructionSelector.cpp:3174
isDead
bool isDead(const MachineInstr &MI, const MachineRegisterInfo &MRI)
Definition SPIRVInstructionSelector.cpp:798
convertPtrToInt
static Register convertPtrToInt(Register Reg, LLT ConvTy, SPIRVTypeInst SpvType, LegalizerHelper &Helper, MachineRegisterInfo &MRI, SPIRVGlobalRegistry *GR)
Definition SPIRVLegalizerInfo.cpp:682
Msg
const char * Msg
Definition SPIRVModuleAnalysis.cpp:59
STLExtras.h
This file contains some templates that are useful if you are working with the STL at all.
BaseType
BaseType
A given derived pointer can have multiple base pointers through phi/selects.
Definition SafepointIRVerifier.cpp:314
SmallSet.h
This file defines the SmallSet class.
StringExtras.h
This file contains some functions that are useful when dealing with strings.
LLVM_DEBUG
#define LLVM_DEBUG(...)
Definition Debug.h:119
Y
static TableGen::Emitter::Opt Y("gen-skeleton-entry", EmitSkeleton, "Generate example skeleton entry")
CallbackFunction
static ManagedStatic< cl::opt< FnT >, OptCreatorT > CallbackFunction
Definition TableGenBackend.cpp:53
Mul
BinaryOperator * Mul
Definition X86PartialReduction.cpp:77
llvm::APFloatBase::IEEEsingle
static const fltSemantics & IEEEsingle()
Definition APFloat.h:304
llvm::APFloatBase::BFloat
static const fltSemantics & BFloat()
Definition APFloat.h:303
llvm::APFloatBase::IEEEdouble
static const fltSemantics & IEEEdouble()
Definition APFloat.h:305
llvm::APFloatBase::IEEEhalf
static const fltSemantics & IEEEhalf()
Definition APFloat.h:302
llvm::APFloat::getSemantics
const fltSemantics & getSemantics() const
Definition APFloat.h:1591
llvm::APFloat::getOne
static APFloat getOne(const fltSemantics &Sem, bool Negative=false)
Factory for Positive and Negative One.
Definition APFloat.h:1192
llvm::APFloat::getZero
static APFloat getZero(const fltSemantics &Sem, bool Negative=false)
Factory for Positive and Negative Zero.
Definition APFloat.h:1183
llvm::APInt::getAllOnes
static APInt getAllOnes(unsigned numBits)
Return an APInt of a specified width with all bits set.
Definition APInt.h:230
llvm::APInt::getSignMask
static APInt getSignMask(unsigned BitWidth)
Get the SignMask for a specific bit width.
Definition APInt.h:225
llvm::ArrayRef
Represent a constant reference to an array (0 or more elements consecutively in memory),...
Definition ArrayRef.h:40
llvm::BlockFrequencyInfo
BlockFrequencyInfo pass uses BlockFrequencyInfoImpl implementation to estimate IR basic block frequen...
Definition BlockFrequencyInfo.h:38
llvm::CmpInst::Predicate
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition InstrTypes.h:740
llvm::CmpInst::FCMP_OEQ
@ FCMP_OEQ
0 0 0 1 True if ordered and equal
Definition InstrTypes.h:743
llvm::CmpInst::ICMP_SLT
@ ICMP_SLT
signed less than
Definition InstrTypes.h:769
llvm::CmpInst::ICMP_SLE
@ ICMP_SLE
signed less or equal
Definition InstrTypes.h:770
llvm::CmpInst::FCMP_OLT
@ FCMP_OLT
0 1 0 0 True if ordered and less than
Definition InstrTypes.h:746
llvm::CmpInst::FCMP_ULE
@ FCMP_ULE
1 1 0 1 True if unordered, less than, or equal
Definition InstrTypes.h:755
llvm::CmpInst::FCMP_OGT
@ FCMP_OGT
0 0 1 0 True if ordered and greater than
Definition InstrTypes.h:744
llvm::CmpInst::FCMP_OGE
@ FCMP_OGE
0 0 1 1 True if ordered and greater than or equal
Definition InstrTypes.h:745
llvm::CmpInst::ICMP_UGE
@ ICMP_UGE
unsigned greater or equal
Definition InstrTypes.h:764
llvm::CmpInst::ICMP_UGT
@ ICMP_UGT
unsigned greater than
Definition InstrTypes.h:763
llvm::CmpInst::ICMP_SGT
@ ICMP_SGT
signed greater than
Definition InstrTypes.h:767
llvm::CmpInst::FCMP_ULT
@ FCMP_ULT
1 1 0 0 True if unordered or less than
Definition InstrTypes.h:754
llvm::CmpInst::FCMP_ONE
@ FCMP_ONE
0 1 1 0 True if ordered and operands are unequal
Definition InstrTypes.h:748
llvm::CmpInst::FCMP_UEQ
@ FCMP_UEQ
1 0 0 1 True if unordered or equal
Definition InstrTypes.h:751
llvm::CmpInst::ICMP_ULT
@ ICMP_ULT
unsigned less than
Definition InstrTypes.h:765
llvm::CmpInst::FCMP_UGT
@ FCMP_UGT
1 0 1 0 True if unordered or greater than
Definition InstrTypes.h:752
llvm::CmpInst::FCMP_OLE
@ FCMP_OLE
0 1 0 1 True if ordered and less than or equal
Definition InstrTypes.h:747
llvm::CmpInst::FCMP_ORD
@ FCMP_ORD
0 1 1 1 True if ordered (no nans)
Definition InstrTypes.h:749
llvm::CmpInst::ICMP_NE
@ ICMP_NE
not equal
Definition InstrTypes.h:762
llvm::CmpInst::ICMP_SGE
@ ICMP_SGE
signed greater or equal
Definition InstrTypes.h:768
llvm::CmpInst::FCMP_UNE
@ FCMP_UNE
1 1 1 0 True if unordered or not equal
Definition InstrTypes.h:756
llvm::CmpInst::ICMP_ULE
@ ICMP_ULE
unsigned less or equal
Definition InstrTypes.h:766
llvm::CmpInst::FCMP_UGE
@ FCMP_UGE
1 0 1 1 True if unordered, greater than, or equal
Definition InstrTypes.h:753
llvm::CmpInst::FCMP_UNO
@ FCMP_UNO
1 0 0 0 True if unordered: isnan(X) | isnan(Y)
Definition InstrTypes.h:750
llvm::Constant::getNullValue
static LLVM_ABI Constant * getNullValue(Type *Ty)
Constructor to create a '0' constant of arbitrary type.
Definition Constants.cpp:395
llvm::DenseMapBase::size
unsigned size() const
Definition DenseMap.h:733
llvm::Function::getContext
LLVMContext & getContext() const
getContext - Return a reference to the LLVMContext associated with this function.
Definition Function.cpp:356
llvm::GFConstant::getConstant
static LLVM_ABI std::optional< GFConstant > getConstant(Register Const, const MachineRegisterInfo &MRI)
Definition Utils.cpp:2038
llvm::GIConstant::getConstant
static LLVM_ABI std::optional< GIConstant > getConstant(Register Const, const MachineRegisterInfo &MRI)
Definition Utils.cpp:1998
llvm::GIntrinsic
Represents a call to an intrinsic.
Definition GenericMachineInstrs.h:564
llvm::GIntrinsic::getIntrinsicID
Intrinsic::ID getIntrinsicID() const
Definition GenericMachineInstrs.h:566
llvm::GlobalValue::getAddressSpace
unsigned getAddressSpace() const
Definition GlobalValue.h:211
llvm::GlobalValue::getParent
Module * getParent()
Get the module that this global value is contained inside of...
Definition GlobalValue.h:728
llvm::GlobalValue::InternalLinkage
@ InternalLinkage
Rename collisions when linking (static functions).
Definition GlobalValue.h:60
llvm::IntegerType::get
static LLVM_ABI IntegerType * get(LLVMContext &C, unsigned NumBits)
This static method is the primary way of constructing an IntegerType.
Definition Type.cpp:338
llvm::LLT::scalar
static constexpr LLT scalar(unsigned SizeInBits)
Get a low-level scalar or aggregate "bag of bits".
Definition LowLevelType.h:89
llvm::LLT::isValid
constexpr bool isValid() const
Definition LowLevelType.h:263
llvm::LLT::getNumElements
constexpr uint16_t getNumElements() const
Returns the number of elements in a vector LLT.
Definition LowLevelType.h:351
llvm::LLT::isVector
constexpr bool isVector() const
Definition LowLevelType.h:290
llvm::LLT::pointer
static constexpr LLT pointer(unsigned AddressSpace, unsigned SizeInBits)
Get a low-level pointer in the given address space.
Definition LowLevelType.h:116
llvm::LLT::fixed_vector
static constexpr LLT fixed_vector(unsigned NumElements, unsigned ScalarSizeInBits)
Get a low-level fixed-width vector of some number of elements and element width.
Definition LowLevelType.h:204
llvm::MachineBasicBlock::getNumber
int getNumber() const
MachineBasicBlocks are uniquely numbered at the function level, unless they're not in a MachineFuncti...
Definition MachineBasicBlock.h:1271
llvm::MachineBasicBlock::getFirstNonPHI
LLVM_ABI iterator getFirstNonPHI()
Returns a pointer to the first instruction in this block that is not a PHINode instruction.
Definition MachineBasicBlock.cpp:204
llvm::MachineBasicBlock::getParent
const MachineFunction * getParent() const
Return the MachineFunction containing this basic block.
Definition MachineBasicBlock.h:323
llvm::MachineBasicBlock::iterator
MachineInstrBundleIterator< MachineInstr > iterator
Definition MachineBasicBlock.h:341
llvm::MachineFunction::getRegInfo
MachineRegisterInfo & getRegInfo()
getRegInfo - Return information about the registers currently in use.
Definition MachineFunction.h:788
llvm::MachineFunction::getFunction
Function & getFunction()
Return the LLVM function that this machine code represents.
Definition MachineFunction.h:739
llvm::MachineIRBuilder
Helper class to build MachineInstr.
Definition MachineIRBuilder.h:238
llvm::MachineIRBuilder::getInsertPt
MachineBasicBlock::iterator getInsertPt()
Current insertion point for new instructions.
Definition MachineIRBuilder.h:336
llvm::MachineIRBuilder::buildInstr
MachineInstrBuilder buildInstr(unsigned Opcode)
Build and insert <empty> = Opcode <empty>.
Definition MachineIRBuilder.h:424
llvm::MachineIRBuilder::getMF
MachineFunction & getMF()
Getter for the function we currently build.
Definition MachineIRBuilder.h:290
llvm::MachineIRBuilder::getMRI
MachineRegisterInfo * getMRI()
Getter for MRI.
Definition MachineIRBuilder.h:312
llvm::MachineInstrBuilder::constrainAllUses
void constrainAllUses(const TargetInstrInfo &TII, const TargetRegisterInfo &TRI, const RegisterBankInfo &RBI) const
Definition MachineInstrBuilder.h:441
llvm::MachineInstrBuilder::addUse
const MachineInstrBuilder & addUse(Register RegNo, RegState Flags={}, unsigned SubReg=0) const
Add a virtual register use operand.
Definition MachineInstrBuilder.h:225
llvm::MachineInstrBuilder::addReg
const MachineInstrBuilder & addReg(Register RegNo, RegState Flags={}, unsigned SubReg=0) const
Add a new virtual register operand.
Definition MachineInstrBuilder.h:199
llvm::MachineInstrBuilder::addImm
const MachineInstrBuilder & addImm(int64_t Val) const
Add a new immediate operand.
Definition MachineInstrBuilder.h:233
llvm::MachineInstrBuilder::add
const MachineInstrBuilder & add(const MachineOperand &MO) const
Definition MachineInstrBuilder.h:326
llvm::MachineInstrBuilder::addMBB
const MachineInstrBuilder & addMBB(MachineBasicBlock *MBB, unsigned TargetFlags=0) const
Definition MachineInstrBuilder.h:248
llvm::MachineInstrBuilder::addDef
const MachineInstrBuilder & addDef(Register RegNo, RegState Flags={}, unsigned SubReg=0) const
Add a virtual register definition operand.
Definition MachineInstrBuilder.h:218
llvm::MachineInstrBuilder::setMIFlags
const MachineInstrBuilder & setMIFlags(unsigned Flags) const
Definition MachineInstrBuilder.h:379
llvm::MachineInstrBuilder::getInstr
MachineInstr * getInstr() const
If conversion operators fail, use this method to get the MachineInstr explicitly.
Definition MachineInstrBuilder.h:191
llvm::MachineInstr
Representation of each machine instruction.
Definition MachineInstr.h:74
llvm::MachineInstr::getOpcode
unsigned getOpcode() const
Returns the opcode of this MachineInstr.
Definition MachineInstr.h:609
llvm::MachineInstr::getParent
const MachineBasicBlock * getParent() const
Definition MachineInstr.h:381
llvm::MachineInstr::getNumOperands
unsigned getNumOperands() const
Retuns the total number of operands.
Definition MachineInstr.h:612
llvm::MachineInstr::getNumExplicitOperands
LLVM_ABI unsigned getNumExplicitOperands() const
Returns the number of non-implicit operands.
Definition MachineInstr.cpp:857
llvm::MachineInstr::getNumExplicitDefs
LLVM_ABI unsigned getNumExplicitDefs() const
Returns the number of non-implicit definitions.
Definition MachineInstr.cpp:875
llvm::MachineInstr::substituteRegister
LLVM_ABI void substituteRegister(Register FromReg, Register ToReg, unsigned SubIdx, const TargetRegisterInfo &RegInfo)
Replace all occurrences of FromReg with ToReg:SubIdx, properly composing subreg indices where necessa...
Definition MachineInstr.cpp:1350
llvm::MachineInstr::emitGenericError
LLVM_ABI void emitGenericError(const Twine &ErrMsg) const
Definition MachineInstr.cpp:2392
llvm::MachineInstr::getMF
LLVM_ABI const MachineFunction * getMF() const
Return the function that contains the basic block that this instruction belongs to.
Definition MachineInstr.cpp:792
llvm::MachineInstr::getDebugLoc
const DebugLoc & getDebugLoc() const
Returns the debug location id of this MachineInstr.
Definition MachineInstr.h:533
llvm::MachineInstr::getOperand
const MachineOperand & getOperand(unsigned i) const
Definition MachineInstr.h:617
llvm::MachineMemOperand
A description of a memory reference used in the backend.
Definition MachineMemOperand.h:141
llvm::MachineMemOperand::MOVolatile
@ MOVolatile
The memory access is volatile.
Definition MachineMemOperand.h:152
llvm::MachineMemOperand::MONonTemporal
@ MONonTemporal
The memory access is non-temporal.
Definition MachineMemOperand.h:154
llvm::MachineOperand::getImm
int64_t getImm() const
Definition MachineOperand.h:563
llvm::MachineOperand::isReg
bool isReg() const
isReg - Tests if this is a MO_Register operand.
Definition MachineOperand.h:334
llvm::MachineOperand::getMBB
MachineBasicBlock * getMBB() const
Definition MachineOperand.h:578
llvm::MachineOperand::getReg
Register getReg() const
getReg - Returns the register number.
Definition MachineOperand.h:375
llvm::MachineRegisterInfo
MachineRegisterInfo - Keep track of information for virtual and physical registers,...
Definition MachineRegisterInfo.h:55
llvm::MachineRegisterInfo::use_instr_iterator
defusechain_instr_iterator< true, false, false, true > use_instr_iterator
use_instr_iterator/use_instr_begin/use_instr_end - Walk all uses of the specified register,...
Definition MachineRegisterInfo.h:498
llvm::MachineRegisterInfo::getRegClass
const TargetRegisterClass * getRegClass(Register Reg) const
Return the register class of the specified virtual register.
Definition MachineRegisterInfo.h:684
llvm::MachineRegisterInfo::getVRegDef
LLVM_ABI LLVM_READONLY MachineInstr * getVRegDef(Register Reg) const
getVRegDef - Return the machine instr that defines the specified virtual register or null if none is ...
Definition MachineRegisterInfo.cpp:406
llvm::MachineRegisterInfo::use_instr_begin
use_instr_iterator use_instr_begin(Register RegNo) const
Definition MachineRegisterInfo.h:500
llvm::MachineRegisterInfo::use_nodbg_empty
bool use_nodbg_empty(Register RegNo) const
use_nodbg_empty - Return true if there are no non-Debug instructions using the specified register.
Definition MachineRegisterInfo.h:598
llvm::MachineRegisterInfo::def_instr_end
static def_instr_iterator def_instr_end()
Definition MachineRegisterInfo.h:424
llvm::MachineRegisterInfo::def_instr_iterator
defusechain_instr_iterator< false, true, false, true > def_instr_iterator
def_instr_iterator/def_instr_begin/def_instr_end - Walk all defs of the specified register,...
Definition MachineRegisterInfo.h:419
llvm::MachineRegisterInfo::createVirtualRegister
LLVM_ABI Register createVirtualRegister(const TargetRegisterClass *RegClass, StringRef Name="")
createVirtualRegister - Create and return a new virtual register in the function with the specified r...
Definition MachineRegisterInfo.cpp:156
llvm::MachineRegisterInfo::def_instr_begin
def_instr_iterator def_instr_begin(Register RegNo) const
Definition MachineRegisterInfo.h:421
llvm::MachineRegisterInfo::getType
LLT getType(Register Reg) const
Get the low-level type of Reg or LLT{} if Reg is not a generic (target independent) virtual register.
Definition MachineRegisterInfo.h:807
llvm::MachineRegisterInfo::hasOneUse
bool hasOneUse(Register RegNo) const
hasOneUse - Return true if there is exactly one instruction using the specified register.
Definition MachineRegisterInfo.h:533
llvm::MachineRegisterInfo::use_instr_end
static use_instr_iterator use_instr_end()
Definition MachineRegisterInfo.h:503
llvm::MachineRegisterInfo::use_nodbg_instructions
iterator_range< use_instr_nodbg_iterator > use_nodbg_instructions(Register Reg) const
Definition MachineRegisterInfo.h:575
llvm::MachineRegisterInfo::setType
LLVM_ABI void setType(Register VReg, LLT Ty)
Set the low-level type of VReg to Ty.
Definition MachineRegisterInfo.cpp:187
llvm::MachineRegisterInfo::getMF
const MachineFunction & getMF() const
Definition MachineRegisterInfo.h:209
llvm::MachineRegisterInfo::setRegClass
LLVM_ABI void setRegClass(Register Reg, const TargetRegisterClass *RC)
setRegClass - Set the register class of the specified virtual register.
Definition MachineRegisterInfo.cpp:60
llvm::MachineRegisterInfo::createGenericVirtualRegister
LLVM_ABI Register createGenericVirtualRegister(LLT Ty, StringRef Name="")
Create and return a new generic virtual register with low-level type Ty.
Definition MachineRegisterInfo.cpp:193
llvm::MachineRegisterInfo::getRegClassOrNull
const TargetRegisterClass * getRegClassOrNull(Register Reg) const
Return the register class of Reg, or null if Reg has not been assigned a register class yet.
Definition MachineRegisterInfo.h:701
llvm::MachineRegisterInfo::use_instructions
iterator_range< use_instr_iterator > use_instructions(Register Reg) const
Definition MachineRegisterInfo.h:508
llvm::MachineRegisterInfo::getNumVirtRegs
unsigned getNumVirtRegs() const
getNumVirtRegs - Return the number of virtual registers created.
Definition MachineRegisterInfo.h:832
llvm::MachineRegisterInfo::replaceRegWith
LLVM_ABI void replaceRegWith(Register FromReg, Register ToReg)
replaceRegWith - Replace all instances of FromReg with ToReg in the machine function.
Definition MachineRegisterInfo.cpp:388
llvm::ProfileSummaryInfo
Analysis providing profile information.
Definition ProfileSummaryInfo.h:42
llvm::RegisterBankInfo
Holds all the information related to register banks.
Definition RegisterBankInfo.h:41
llvm::Register
Wrapper class representing virtual and physical registers.
Definition Register.h:20
llvm::Register::isValid
constexpr bool isValid() const
Definition Register.h:112
llvm::Register::isPhysical
constexpr bool isPhysical() const
Return true if the specified register number is in the physical register namespace.
Definition Register.h:83
llvm::SPIRVGlobalRegistry::isScalarOrVectorSigned
bool isScalarOrVectorSigned(SPIRVTypeInst Type) const
Definition SPIRVGlobalRegistry.cpp:1547
llvm::SPIRVGlobalRegistry::getOrCreateOpTypeSampledImage
SPIRVTypeInst getOrCreateOpTypeSampledImage(SPIRVTypeInst ImageType, MachineIRBuilder &MIRBuilder)
Definition SPIRVGlobalRegistry.cpp:1791
llvm::SPIRVGlobalRegistry::assignSPIRVTypeToVReg
void assignSPIRVTypeToVReg(SPIRVTypeInst Type, Register VReg, const MachineFunction &MF)
Definition SPIRVGlobalRegistry.cpp:119
llvm::SPIRVGlobalRegistry::getOrCreateSPIRVPointerType
SPIRVTypeInst getOrCreateSPIRVPointerType(const Type *BaseType, MachineIRBuilder &MIRBuilder, SPIRV::StorageClass::StorageClass SC, bool ForceTyped=false)
Definition SPIRVGlobalRegistry.cpp:2064
llvm::SPIRVGlobalRegistry::getRegClass
const TargetRegisterClass * getRegClass(SPIRVTypeInst SpvType) const
Definition SPIRVGlobalRegistry.cpp:2197
llvm::SPIRVGlobalRegistry::getOrAddMemAliasingINTELInst
MachineInstr * getOrAddMemAliasingINTELInst(MachineIRBuilder &MIRBuilder, const MDNode *AliasingListMD)
Definition SPIRVGlobalRegistry.cpp:2264
llvm::SPIRVGlobalRegistry::isAggregateType
bool isAggregateType(SPIRVTypeInst Type) const
Definition SPIRVGlobalRegistry.h:381
llvm::SPIRVGlobalRegistry::getScalarOrVectorBitWidth
unsigned getScalarOrVectorBitWidth(SPIRVTypeInst Type) const
Definition SPIRVGlobalRegistry.cpp:1498
llvm::SPIRVGlobalRegistry::getOrCreateSPIRVIntegerType
SPIRVTypeInst getOrCreateSPIRVIntegerType(unsigned BitWidth, MachineIRBuilder &MIRBuilder)
Definition SPIRVGlobalRegistry.cpp:1921
llvm::SPIRVGlobalRegistry::getOrCreateSPIRVVectorType
SPIRVTypeInst getOrCreateSPIRVVectorType(SPIRVTypeInst BaseType, unsigned NumElements, MachineIRBuilder &MIRBuilder, bool EmitIR)
Definition SPIRVGlobalRegistry.cpp:2020
llvm::SPIRVGlobalRegistry::buildGlobalVariable
Register buildGlobalVariable(Register Reg, SPIRVTypeInst BaseType, StringRef Name, const GlobalValue *GV, SPIRV::StorageClass::StorageClass Storage, const MachineInstr *Init, bool IsConst, const std::optional< SPIRV::LinkageType::LinkageType > &LinkageType, MachineIRBuilder &MIRBuilder, bool IsInstSelector)
Definition SPIRVGlobalRegistry.cpp:816
llvm::SPIRVGlobalRegistry::getResultType
SPIRVTypeInst getResultType(Register VReg, MachineFunction *MF=nullptr)
Definition SPIRVGlobalRegistry.cpp:1378
llvm::SPIRVGlobalRegistry::getScalarOrVectorComponentCount
unsigned getScalarOrVectorComponentCount(Register VReg) const
Definition SPIRVGlobalRegistry.cpp:1470
llvm::SPIRVGlobalRegistry::getTypeForSPIRVType
const Type * getTypeForSPIRVType(SPIRVTypeInst Ty) const
Definition SPIRVGlobalRegistry.h:341
llvm::SPIRVGlobalRegistry::isBitcastCompatible
bool isBitcastCompatible(SPIRVTypeInst Type1, SPIRVTypeInst Type2) const
Definition SPIRVGlobalRegistry.cpp:1558
llvm::SPIRVGlobalRegistry::getOrCreateConstFP
Register getOrCreateConstFP(APFloat Val, MachineInstr &I, SPIRVTypeInst SpvType, const SPIRVInstrInfo &TII, bool ZeroAsNull=true)
Definition SPIRVGlobalRegistry.cpp:369
llvm::SPIRVGlobalRegistry::getRegType
LLT getRegType(SPIRVTypeInst SpvType) const
Definition SPIRVGlobalRegistry.cpp:2224
llvm::SPIRVGlobalRegistry::getOpTypeVoid
SPIRVTypeInst getOpTypeVoid(MachineIRBuilder &MIRBuilder)
Definition SPIRVGlobalRegistry.cpp:214
llvm::SPIRVGlobalRegistry::invalidateMachineInstr
void invalidateMachineInstr(MachineInstr *MI)
Definition SPIRVGlobalRegistry.cpp:222
llvm::SPIRVGlobalRegistry::getOrCreateSPIRVBoolType
SPIRVTypeInst getOrCreateSPIRVBoolType(MachineIRBuilder &MIRBuilder, bool EmitIR)
Definition SPIRVGlobalRegistry.cpp:1995
llvm::SPIRVGlobalRegistry::isScalarOfType
bool isScalarOfType(Register VReg, unsigned TypeOpcode) const
Definition SPIRVGlobalRegistry.cpp:1436
llvm::SPIRVGlobalRegistry::getSPIRVTypeID
Register getSPIRVTypeID(SPIRVTypeInst SpirvType) const
Definition SPIRVGlobalRegistry.cpp:1199
llvm::SPIRVGlobalRegistry::getOrCreateConstInt
Register getOrCreateConstInt(uint64_t Val, MachineInstr &I, SPIRVTypeInst SpvType, const SPIRVInstrInfo &TII, bool ZeroAsNull=true)
Definition SPIRVGlobalRegistry.cpp:419
llvm::SPIRVGlobalRegistry::getOrCreateConstIntArray
Register getOrCreateConstIntArray(uint64_t Val, size_t Num, MachineInstr &I, SPIRVTypeInst SpvType, const SPIRVInstrInfo &TII)
Definition SPIRVGlobalRegistry.cpp:682
llvm::SPIRVGlobalRegistry::findValueAttrs
bool findValueAttrs(const MachineInstr *Key, Type *&Ty, StringRef &Name)
Definition SPIRVGlobalRegistry.h:190
llvm::SPIRVGlobalRegistry::retrieveScalarOrVectorIntType
SPIRVTypeInst retrieveScalarOrVectorIntType(SPIRVTypeInst Type) const
Definition SPIRVGlobalRegistry.cpp:1541
llvm::SPIRVGlobalRegistry::getOrCreateGlobalVariableWithBinding
Register getOrCreateGlobalVariableWithBinding(SPIRVTypeInst VarType, uint32_t Set, uint32_t Binding, StringRef Name, MachineIRBuilder &MIRBuilder)
Definition SPIRVGlobalRegistry.cpp:959
llvm::SPIRVGlobalRegistry::changePointerStorageClass
SPIRVTypeInst changePointerStorageClass(SPIRVTypeInst PtrType, SPIRV::StorageClass::StorageClass SC, MachineInstr &I)
Definition SPIRVGlobalRegistry.cpp:2086
llvm::SPIRVGlobalRegistry::getOrCreateConstVector
Register getOrCreateConstVector(uint64_t Val, MachineInstr &I, SPIRVTypeInst SpvType, const SPIRVInstrInfo &TII, bool ZeroAsNull=true)
Definition SPIRVGlobalRegistry.cpp:619
llvm::SPIRVGlobalRegistry::buildConstantFP
Register buildConstantFP(APFloat Val, MachineIRBuilder &MIRBuilder, SPIRVTypeInst SpvType=nullptr)
Definition SPIRVGlobalRegistry.cpp:527
llvm::SPIRVGlobalRegistry::addGlobalObject
void addGlobalObject(const Value *V, const MachineFunction *MF, Register R)
Definition SPIRVGlobalRegistry.h:136
llvm::SPIRVGlobalRegistry::getScalarOrVectorComponentType
SPIRVTypeInst getScalarOrVectorComponentType(SPIRVTypeInst Type) const
Definition SPIRVGlobalRegistry.cpp:1486
llvm::SPIRVGlobalRegistry::recordFunctionPointer
void recordFunctionPointer(const MachineOperand *MO, const Function *F)
Definition SPIRVGlobalRegistry.h:283
llvm::SPIRVGlobalRegistry::getOrCreateSPIRVFloatType
SPIRVTypeInst getOrCreateSPIRVFloatType(unsigned BitWidth, MachineInstr &I, const SPIRVInstrInfo &TII)
Definition SPIRVGlobalRegistry.cpp:1974
llvm::SPIRVGlobalRegistry::getPointeeType
SPIRVTypeInst getPointeeType(SPIRVTypeInst PtrType)
Definition SPIRVGlobalRegistry.cpp:1552
llvm::SPIRVGlobalRegistry::getOrCreateSPIRVType
SPIRVTypeInst getOrCreateSPIRVType(const Type *Type, MachineInstr &I, SPIRV::AccessQualifier::AccessQualifier AQ, bool EmitIR)
Definition SPIRVGlobalRegistry.h:327
llvm::SPIRVGlobalRegistry::isScalarOrVectorOfType
bool isScalarOrVectorOfType(Register VReg, unsigned TypeOpcode) const
Definition SPIRVGlobalRegistry.cpp:1443
llvm::SPIRVGlobalRegistry::setCurrentFunc
MachineFunction * setCurrentFunc(MachineFunction &MF)
Definition SPIRVGlobalRegistry.h:374
llvm::SPIRVGlobalRegistry::getOrCreateConstNullPtr
Register getOrCreateConstNullPtr(MachineIRBuilder &MIRBuilder, SPIRVTypeInst SpvType)
Definition SPIRVGlobalRegistry.cpp:765
llvm::SPIRVGlobalRegistry::getSPIRVTypeForVReg
SPIRVTypeInst getSPIRVTypeForVReg(Register VReg, const MachineFunction *MF=nullptr) const
Definition SPIRVGlobalRegistry.cpp:1367
llvm::SPIRVGlobalRegistry::getDeducedGlobalValueType
Type * getDeducedGlobalValueType(const GlobalValue *Global)
Definition SPIRVGlobalRegistry.h:236
llvm::SPIRVGlobalRegistry::getOrCreateUndef
Register getOrCreateUndef(MachineInstr &I, SPIRVTypeInst SpvType, const SPIRVInstrInfo &TII)
Definition SPIRVGlobalRegistry.cpp:2166
llvm::SPIRVGlobalRegistry::getPointerStorageClass
SPIRV::StorageClass::StorageClass getPointerStorageClass(Register VReg) const
Definition SPIRVGlobalRegistry.cpp:1577
llvm::SPIRVGlobalRegistry::getUntypedPtrElementType
SPIRVTypeInst getUntypedPtrElementType(Register Reg) const
Definition SPIRVGlobalRegistry.h:231
llvm::SPIRVIRMapping::erase
bool erase(const MachineInstr *MI)
Definition SPIRVIRMapping.h:217
llvm::SPIRVIRMapping::add
bool add(SPIRV::IRHandle Handle, const MachineInstr *MI)
Definition SPIRVIRMapping.h:195
llvm::SPIRVIRMapping::find
Register find(SPIRV::IRHandle Handle, const MachineFunction *MF)
Definition SPIRVIRMapping.h:241
llvm::SPIRVSubtarget::isPhysicalSPIRV
bool isPhysicalSPIRV() const
Definition SPIRVSubtarget.h:92
llvm::SPIRVSubtarget::isAtLeastSPIRVVer
bool isAtLeastSPIRVVer(VersionTuple VerToCompareTo) const
Definition SPIRVSubtarget.cpp:149
llvm::SPIRVSubtarget::canUseExtInstSet
bool canUseExtInstSet(SPIRV::InstructionSet::InstructionSet E) const
Definition SPIRVSubtarget.cpp:136
llvm::SPIRVSubtarget::isLogicalSPIRV
bool isLogicalSPIRV() const
Definition SPIRVSubtarget.h:89
llvm::SPIRVSubtarget::canUseExtension
bool canUseExtension(SPIRV::Extension::Extension E) const
Definition SPIRVSubtarget.cpp:132
llvm::SPIRVTypeInst::isTypeIntOrFloat
bool isTypeIntOrFloat() const
Definition SPIRVTypeInst.h:57
llvm::SPIRVTypeInst::isAnyTypeFloat
bool isAnyTypeFloat() const
Definition SPIRVTypeInst.cpp:36
llvm::SmallPtrSetImpl::erase
bool erase(PtrType Ptr)
Remove pointer from the set.
Definition SmallPtrSet.h:388
llvm::SmallPtrSetImpl::insert
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition SmallPtrSet.h:373
llvm::SmallPtrSetImpl::contains
bool contains(ConstPtrType Ptr) const
Definition SmallPtrSet.h:453
llvm::SmallPtrSet
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements.
Definition SmallPtrSet.h:519
llvm::SmallSet
SmallSet - This maintains a set of unique values, optimizing for the case when the set is small (less...
Definition SmallSet.h:134
llvm::SmallSet::insert
std::pair< const_iterator, bool > insert(const T &V)
insert - Insert an element into the set if it isn't already there.
Definition SmallSet.h:184
llvm::SmallVectorImpl::emplace_back
reference emplace_back(ArgTypes &&... Args)
Definition SmallVector.h:993
llvm::SmallVectorImpl::reserve
void reserve(size_type N)
Definition SmallVector.h:698
llvm::SmallVectorTemplateBase::push_back
void push_back(const T &Elt)
Definition SmallVector.h:423
llvm::SmallVector
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
Definition SmallVector.h:1252
llvm::StringRef::size
constexpr size_t size() const
Get the string size.
Definition StringRef.h:144
llvm::StructType::get
static LLVM_ABI StructType * get(LLVMContext &Context, ArrayRef< Type * > Elements, bool isPacked=false)
This static method is the primary way to create a literal StructType.
Definition Type.cpp:467
llvm::Twine
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition Twine.h:82
llvm::Type
The instances of the Type class are immutable: once they are created, they are never changed.
Definition Type.h:46
llvm::Type::HalfTyID
@ HalfTyID
16-bit floating point type
Definition Type.h:57
llvm::Type::FloatTyID
@ FloatTyID
32-bit floating point type
Definition Type.h:59
llvm::Type::BFloatTyID
@ BFloatTyID
16-bit floating point type (7-bit significand)
Definition Type.h:58
llvm::Type::DoubleTyID
@ DoubleTyID
64-bit floating point type
Definition Type.h:60
llvm::Type::getScalarType
Type * getScalarType() const
If this is a vector type, return the element type, otherwise return 'this'.
Definition Type.h:363
llvm::Type::isStructTy
bool isStructTy() const
True if this is an instance of StructType.
Definition Type.h:271
llvm::Type::isAggregateType
bool isAggregateType() const
Return true if the type is an aggregate type.
Definition Type.h:314
llvm::Type::getTypeID
TypeID getTypeID() const
Return the type id for the type.
Definition Type.h:138
llvm::User::getOperand
Value * getOperand(unsigned i) const
Definition User.h:207
llvm::Value::hasName
bool hasName() const
Definition Value.h:263
llvm::Value::getName
LLVM_ABI StringRef getName() const
Return a constant reference to the value's name.
Definition Value.cpp:319
llvm::function_ref
An efficient, type-erasing, non-owning reference to a callable.
Definition STLFunctionalExtras.h:37
llvm::ilist_node_impl::getIterator
self_iterator getIterator()
Definition ilist_node.h:123
llvm::ilist_node_with_parent::getNextNode
NodeTy * getNextNode()
Get the next node, or nullptr for the list tail.
Definition ilist_node.h:348
uint32_t
llvm_unreachable
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
Definition ErrorHandling.h:164
R600_InstFlag::FC
@ FC
Definition R600Defines.h:32
llvm::AMDGPU::HSAMD::Kernel::Arg::Key::IsConst
constexpr char IsConst[]
Key for Kernel::Arg::Metadata::mIsConst.
Definition AMDGPUMetadata.h:197
llvm::BitmaskEnumDetail::Mask
constexpr std::underlying_type_t< E > Mask()
Get a bitmask with 1s in all places up to the high-order bit of E's largest value.
Definition BitmaskEnum.h:126
llvm::dwarf_linker::DebugSectionKind::DebugLoc
@ DebugLoc
Definition DWARFLinkerBase.h:35
llvm::ms_demangle::QualifierMangleMode::Result
@ Result
Definition MicrosoftDemangle.h:132
llvm::rdf::Def
NodeAddr< DefNode * > Def
Definition RDFGraph.h:384
llvm::rdf::Instr
NodeAddr< InstrNode * > Instr
Definition RDFGraph.h:389
llvm::rdf::Use
NodeAddr< UseNode * > Use
Definition RDFGraph.h:385
llvm::rdf::Func
NodeAddr< FuncNode * > Func
Definition RDFGraph.h:393
llvm::sframe::BaseReg
BaseReg
Stack frame base register. Bit 0 of FREInfo.Info.
Definition SFrame.h:77
llvm::vputils::getOpcode
unsigned getOpcode(const VPValue *V)
Return the instruction opcode for the recipe defining V or 0 for unsupported recipes and VPValues not...
Definition VPlanUtils.cpp:398
llvm
This is an optimization pass for GlobalISel generic memory operations.
Definition DefaultTargetInfo.h:21
llvm::Offset
@ Offset
Definition DWP.cpp:577
llvm::all_of
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly.
Definition STLExtras.h:1755
llvm::addStringImm
void addStringImm(StringRef Str, MCInst &Inst)
Definition SPIRVUtils.cpp:163
llvm::getOpVariableMBBIt
MachineBasicBlock::iterator getOpVariableMBBIt(MachineFunction &MF)
Definition SPIRVUtils.cpp:337
llvm::getIConstValSext
int64_t getIConstValSext(Register ConstReg, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:520
llvm::BuildMI
MachineInstrBuilder BuildMI(MachineFunction &MF, const MIMetadata &MIMD, const MCInstrDesc &MCID)
Builder interface. Specify how to create the initial instruction itself.
Definition MachineInstrBuilder.h:449
llvm::isTypeFoldingSupported
bool isTypeFoldingSupported(unsigned Opcode)
Definition SPIRVUtils.cpp:1159
llvm::getMemSemanticsWithStorageClass
uint32_t getMemSemanticsWithStorageClass(const Triple &TT, uint32_t OrderSem, uint32_t StorageClassSem)
Definition SPIRVUtils.cpp:456
llvm::dyn_cast
decltype(auto) dyn_cast(const From &Val)
dyn_cast<X> - Return the argument parameter cast to the specified type.
Definition Casting.h:643
llvm::getDef
MachineInstr * getDef(const MachineOperand &MO, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:1171
llvm::addNumImm
void addNumImm(const APInt &Imm, MachineInstrBuilder &MIB)
Definition SPIRVUtils.cpp:193
llvm::salvageDebugInfo
LLVM_ABI void salvageDebugInfo(const MachineRegisterInfo &MRI, MachineInstr &MI)
Assuming the instruction MI is going to be deleted, attempt to salvage debug users of MI by writing t...
Definition Utils.cpp:1676
llvm::constrainSelectedInstRegOperands
LLVM_ABI void constrainSelectedInstRegOperands(MachineInstr &I, const TargetInstrInfo &TII, const TargetRegisterInfo &TRI, const RegisterBankInfo &RBI)
Mutate the newly-selected instruction I to constrain its (possibly generic) virtual register operands...
Definition Utils.cpp:159
llvm::isPreISelGenericOpcode
bool isPreISelGenericOpcode(unsigned Opcode)
Check whether the given Opcode is a generic opcode that is not supposed to appear after ISel.
Definition TargetOpcodes.h:30
llvm::make_range
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
Definition iterator_range.h:67
llvm::createVirtualRegister
Register createVirtualRegister(SPIRVTypeInst SpvType, SPIRVGlobalRegistry *GR, MachineRegisterInfo *MRI, const MachineFunction &MF)
Definition SPIRVUtils.cpp:980
llvm::getArrayComponentCount
unsigned getArrayComponentCount(const MachineRegisterInfo *MRI, const MachineInstr *ResType)
Definition SPIRVUtils.cpp:1196
llvm::make_early_inc_range
iterator_range< early_inc_iterator_impl< detail::IterOfRange< RangeT > > > make_early_inc_range(RangeT &&Range)
Make a range that does early increment to allow mutation of the underlying range without disrupting i...
Definition STLExtras.h:649
llvm::buildOpDecorate
void buildOpDecorate(Register Reg, MachineIRBuilder &MIRBuilder, SPIRV::Decoration::Decoration Dec, ArrayRef< uint32_t > DecArgs, StringRef StrImm)
Definition SPIRVUtils.cpp:249
llvm::isNullOrNullSplat
LLVM_ABI bool isNullOrNullSplat(const MachineInstr &MI, const MachineRegisterInfo &MRI, bool AllowUndefs=false)
Return true if the value is a constant 0 integer or a splatted vector of a constant 0 integer (with n...
Definition Utils.cpp:1539
llvm::getMemScope
SPIRV::Scope::Scope getMemScope(const Triple &TT, LLVMContext &Ctx, SyncScope::ID Id)
Definition SPIRVUtils.cpp:464
llvm::getIConstVal
uint64_t getIConstVal(Register ConstReg, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:514
llvm::createContinuedInstructions
SmallVector< MachineInstr *, 4 > createContinuedInstructions(MachineIRBuilder &MIRBuilder, unsigned Opcode, unsigned MinWC, unsigned ContinuedOpcode, ArrayRef< Register > Args, Register ReturnRegister, Register TypeID)
Definition SPIRVUtils.cpp:1050
llvm::getMemSemanticsForStorageClass
SPIRV::MemorySemantics::MemorySemantics getMemSemanticsForStorageClass(SPIRV::StorageClass::StorageClass SC)
Definition SPIRVUtils.cpp:416
llvm::storageClassToAddressSpace
constexpr unsigned storageClassToAddressSpace(SPIRV::StorageClass::StorageClass SC)
Definition SPIRVUtils.h:245
llvm::isVectorType
bool isVectorType(SPIRVTypeInst SPVTy)
Definition SPIRVUtils.cpp:1007
llvm::Value
RelativeUniformCounterPtr ValuesPtrExpr VTableAddr Value
Definition InstrProf.h:143
llvm::buildOpName
void buildOpName(Register Target, StringRef Name, MachineIRBuilder &MIRBuilder)
Definition SPIRVUtils.cpp:222
llvm::any_of
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly.
Definition STLExtras.h:1762
llvm::getImm
MachineInstr * getImm(const MachineOperand &MO, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:1177
llvm::isPowerOf2_32
constexpr bool isPowerOf2_32(uint32_t Value)
Return true if the argument is a power of two > 0.
Definition MathExtras.h:280
llvm::toTypedPointer
Type * toTypedPointer(Type *Ty)
Definition SPIRVUtils.h:479
llvm::dbgs
LLVM_ABI raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition Debug.cpp:209
llvm::report_fatal_error
LLVM_ABI void report_fatal_error(Error Err, bool gen_crash_diag=true)
Definition Error.cpp:163
llvm::isGenericCastablePtr
constexpr bool isGenericCastablePtr(SPIRV::StorageClass::StorageClass SC)
Definition SPIRVUtils.h:229
llvm::SmallVector
class LLVM_GSL_OWNER SmallVector
Forward declaration of SmallVector so that calculateSmallVectorDefaultInlinedElements can reference s...
Definition SmallVector.h:1178
llvm::passCopy
MachineInstr * passCopy(MachineInstr *Def, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:1164
llvm::isa
bool isa(const From &Val)
isa<X> - Return true if the parameter to the template is an instance of one of the template type argu...
Definition Casting.h:547
llvm::getSpirvLinkageTypeFor
std::optional< SPIRV::LinkageType::LinkageType > getSpirvLinkageTypeFor(const SPIRVSubtarget &ST, const GlobalValue &GV)
Definition SPIRVUtils.cpp:1270
llvm::errs
LLVM_ABI raw_fd_ostream & errs()
This returns a reference to a raw_ostream for standard error.
Definition raw_ostream.cpp:890
llvm::addressSpaceToStorageClass
SPIRV::StorageClass::StorageClass addressSpaceToStorageClass(unsigned AddrSpace, const SPIRVSubtarget &STI)
Definition SPIRVUtils.cpp:376
llvm::AtomicOrdering
AtomicOrdering
Atomic ordering for LLVM's memory model.
Definition AtomicOrdering.h:56
llvm::createSPIRVInstructionSelector
InstructionSelector * createSPIRVInstructionSelector(const SPIRVTargetMachine &TM, const SPIRVSubtarget &Subtarget, const RegisterBankInfo &RBI)
Definition SPIRVInstructionSelector.cpp:7729
llvm::getStringValueFromReg
std::string getStringValueFromReg(Register Reg, MachineRegisterInfo &MRI)
Definition SPIRVUtils.cpp:183
llvm::foldImm
int64_t foldImm(const MachineOperand &MO, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:1186
llvm::Op
DWARFExpression::Operation Op
Definition DWARFExpressionPrinter.cpp:24
llvm::ArrayRef
ArrayRef(const T &OneElt) -> ArrayRef< T >
llvm::getDefInstrMaybeConstant
MachineInstr * getDefInstrMaybeConstant(Register &ConstReg, const MachineRegisterInfo *MRI)
Definition SPIRVUtils.cpp:491
llvm::BitWidth
constexpr unsigned BitWidth
Definition BitmaskEnum.h:219
llvm::cast
decltype(auto) cast(const From &Val)
cast<X> - Return the argument parameter cast to the specified type.
Definition Casting.h:559
llvm::hasInitializer
bool hasInitializer(const GlobalVariable *GV)
Definition SPIRVUtils.h:364
llvm::isSpvIntrinsic
bool isSpvIntrinsic(const MachineInstr &MI, Intrinsic::ID IntrinsicID)
Definition SPIRVUtils.cpp:526
llvm::getVRegDef
MachineInstr * getVRegDef(MachineRegisterInfo &MRI, Register Reg)
Definition SPIRVUtils.cpp:931
llvm::getMemSemantics
SPIRV::MemorySemantics::MemorySemantics getMemSemantics(AtomicOrdering Ord)
Definition SPIRVUtils.cpp:438
llvm::getLinkStringForBuiltIn
std::string getLinkStringForBuiltIn(SPIRV::BuiltIn::BuiltIn BuiltInValue)
Definition SPIRVBaseInfo.cpp:204
llvm::isTriviallyDead
LLVM_ABI bool isTriviallyDead(const MachineInstr &MI, const MachineRegisterInfo &MRI)
Check whether an instruction MI is dead: it only defines dead virtual registers, and doesn't have oth...
Definition Utils.cpp:224
llvm::TargetRegisterClass
MCRegisterClass TargetRegisterClass
Definition FastISel.h:58
N
#define N
llvm::MemOp::isVolatile
bool isVolatile() const
Definition TargetLowering.h:187
Generated on for LLVM by doxygen 1.14.0