LLVM 19.0.0git
AMDGPUUnifyDivergentExitNodes.cpp
Go to the documentation of this file.
1//===- AMDGPUUnifyDivergentExitNodes.cpp ----------------------------------===//
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 is a variant of the UnifyFunctionExitNodes pass. Rather than ensuring
10// there is at most one ret and one unreachable instruction, it ensures there is
11// at most one divergent exiting block.
12//
13// StructurizeCFG can't deal with multi-exit regions formed by branches to
14// multiple return nodes. It is not desirable to structurize regions with
15// uniform branches, so unifying those to the same return block as divergent
16// branches inhibits use of scalar branching. It still can't deal with the case
17// where one branch goes to return, and one unreachable. Replace unreachable in
18// this case with a return.
19//
20//===----------------------------------------------------------------------===//
21
23#include "AMDGPU.h"
24#include "SIDefines.h"
25#include "llvm/ADT/ArrayRef.h"
28#include "llvm/ADT/StringRef.h"
33#include "llvm/IR/BasicBlock.h"
34#include "llvm/IR/CFG.h"
35#include "llvm/IR/Constants.h"
36#include "llvm/IR/Dominators.h"
37#include "llvm/IR/Function.h"
38#include "llvm/IR/IRBuilder.h"
39#include "llvm/IR/InstrTypes.h"
41#include "llvm/IR/Intrinsics.h"
42#include "llvm/IR/IntrinsicsAMDGPU.h"
43#include "llvm/IR/Type.h"
45#include "llvm/Pass.h"
51
52using namespace llvm;
53
54#define DEBUG_TYPE "amdgpu-unify-divergent-exit-nodes"
55
56namespace {
57
58class AMDGPUUnifyDivergentExitNodesImpl {
59private:
60 const TargetTransformInfo *TTI = nullptr;
61
62public:
63 AMDGPUUnifyDivergentExitNodesImpl() = delete;
64 AMDGPUUnifyDivergentExitNodesImpl(const TargetTransformInfo *TTI)
65 : TTI(TTI) {}
66
67 // We can preserve non-critical-edgeness when we unify function exit nodes
68 BasicBlock *unifyReturnBlockSet(Function &F, DomTreeUpdater &DTU,
69 ArrayRef<BasicBlock *> ReturningBlocks,
71 bool run(Function &F, DominatorTree *DT, const PostDominatorTree &PDT,
72 const UniformityInfo &UA);
73};
74
75class AMDGPUUnifyDivergentExitNodes : public FunctionPass {
76public:
77 static char ID;
78 AMDGPUUnifyDivergentExitNodes() : FunctionPass(ID) {
81 }
82 void getAnalysisUsage(AnalysisUsage &AU) const override;
83 bool runOnFunction(Function &F) override;
84};
85} // end anonymous namespace
86
87char AMDGPUUnifyDivergentExitNodes::ID = 0;
88
89char &llvm::AMDGPUUnifyDivergentExitNodesID = AMDGPUUnifyDivergentExitNodes::ID;
90
91INITIALIZE_PASS_BEGIN(AMDGPUUnifyDivergentExitNodes, DEBUG_TYPE,
92 "Unify divergent function exit nodes", false, false)
96INITIALIZE_PASS_END(AMDGPUUnifyDivergentExitNodes, DEBUG_TYPE,
97 "Unify divergent function exit nodes", false, false)
98
99void AMDGPUUnifyDivergentExitNodes::getAnalysisUsage(AnalysisUsage &AU) const {
101 AU.addRequired<DominatorTreeWrapperPass>();
102
103 AU.addRequired<PostDominatorTreeWrapperPass>();
104
105 AU.addRequired<UniformityInfoWrapperPass>();
106
108 AU.addPreserved<DominatorTreeWrapperPass>();
109 // FIXME: preserve PostDominatorTreeWrapperPass
110 }
111
112 // We preserve the non-critical-edgeness property
113 AU.addPreservedID(BreakCriticalEdgesID);
114
115 FunctionPass::getAnalysisUsage(AU);
116
117 AU.addRequired<TargetTransformInfoWrapperPass>();
118}
119
120/// \returns true if \p BB is reachable through only uniform branches.
121/// XXX - Is there a more efficient way to find this?
122static bool isUniformlyReached(const UniformityInfo &UA, BasicBlock &BB) {
125
126 while (!Stack.empty()) {
127 BasicBlock *Top = Stack.pop_back_val();
128 if (!UA.isUniform(Top->getTerminator()))
129 return false;
130
131 for (BasicBlock *Pred : predecessors(Top)) {
132 if (Visited.insert(Pred).second)
133 Stack.push_back(Pred);
134 }
135 }
136
137 return true;
138}
139
140BasicBlock *AMDGPUUnifyDivergentExitNodesImpl::unifyReturnBlockSet(
141 Function &F, DomTreeUpdater &DTU, ArrayRef<BasicBlock *> ReturningBlocks,
142 StringRef Name) {
143 // Otherwise, we need to insert a new basic block into the function, add a PHI
144 // nodes (if the function returns values), and convert all of the return
145 // instructions into unconditional branches.
146 BasicBlock *NewRetBlock = BasicBlock::Create(F.getContext(), Name, &F);
147 IRBuilder<> B(NewRetBlock);
148
149 PHINode *PN = nullptr;
150 if (F.getReturnType()->isVoidTy()) {
151 B.CreateRetVoid();
152 } else {
153 // If the function doesn't return void... add a PHI node to the block...
154 PN = B.CreatePHI(F.getReturnType(), ReturningBlocks.size(),
155 "UnifiedRetVal");
156 B.CreateRet(PN);
157 }
158
159 // Loop over all of the blocks, replacing the return instruction with an
160 // unconditional branch.
161 std::vector<DominatorTree::UpdateType> Updates;
162 Updates.reserve(ReturningBlocks.size());
163 for (BasicBlock *BB : ReturningBlocks) {
164 // Add an incoming element to the PHI node for every return instruction that
165 // is merging into this new block...
166 if (PN)
167 PN->addIncoming(BB->getTerminator()->getOperand(0), BB);
168
169 // Remove and delete the return inst.
170 BB->getTerminator()->eraseFromParent();
171 BranchInst::Create(NewRetBlock, BB);
172 Updates.push_back({DominatorTree::Insert, BB, NewRetBlock});
173 }
174
176 DTU.applyUpdates(Updates);
177 Updates.clear();
178
179 for (BasicBlock *BB : ReturningBlocks) {
180 // Cleanup possible branch to unconditional branch to the return.
181 simplifyCFG(BB, *TTI, RequireAndPreserveDomTree ? &DTU : nullptr,
182 SimplifyCFGOptions().bonusInstThreshold(2));
183 }
184
185 return NewRetBlock;
186}
187
188bool AMDGPUUnifyDivergentExitNodesImpl::run(Function &F, DominatorTree *DT,
189 const PostDominatorTree &PDT,
190 const UniformityInfo &UA) {
191 assert(hasOnlySimpleTerminator(F) && "Unsupported block terminator.");
192
193 if (PDT.root_size() == 0 ||
194 (PDT.root_size() == 1 &&
195 !isa<BranchInst>(PDT.getRoot()->getTerminator())))
196 return false;
197
198 // Loop over all of the blocks in a function, tracking all of the blocks that
199 // return.
200 SmallVector<BasicBlock *, 4> ReturningBlocks;
201 SmallVector<BasicBlock *, 4> UnreachableBlocks;
202
203 // Dummy return block for infinite loop.
204 BasicBlock *DummyReturnBB = nullptr;
205
206 bool Changed = false;
207 std::vector<DominatorTree::UpdateType> Updates;
208
209 // TODO: For now we unify all exit blocks, even though they are uniformly
210 // reachable, if there are any exits not uniformly reached. This is to
211 // workaround the limitation of structurizer, which can not handle multiple
212 // function exits. After structurizer is able to handle multiple function
213 // exits, we should only unify UnreachableBlocks that are not uniformly
214 // reachable.
215 bool HasDivergentExitBlock = llvm::any_of(
216 PDT.roots(), [&](auto BB) { return !isUniformlyReached(UA, *BB); });
217
218 for (BasicBlock *BB : PDT.roots()) {
219 if (isa<ReturnInst>(BB->getTerminator())) {
220 if (HasDivergentExitBlock)
221 ReturningBlocks.push_back(BB);
222 } else if (isa<UnreachableInst>(BB->getTerminator())) {
223 if (HasDivergentExitBlock)
224 UnreachableBlocks.push_back(BB);
225 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
226
227 ConstantInt *BoolTrue = ConstantInt::getTrue(F.getContext());
228 if (DummyReturnBB == nullptr) {
229 DummyReturnBB = BasicBlock::Create(F.getContext(),
230 "DummyReturnBlock", &F);
231 Type *RetTy = F.getReturnType();
232 Value *RetVal = RetTy->isVoidTy() ? nullptr : PoisonValue::get(RetTy);
233 ReturnInst::Create(F.getContext(), RetVal, DummyReturnBB);
234 ReturningBlocks.push_back(DummyReturnBB);
235 }
236
237 if (BI->isUnconditional()) {
238 BasicBlock *LoopHeaderBB = BI->getSuccessor(0);
239 BI->eraseFromParent(); // Delete the unconditional branch.
240 // Add a new conditional branch with a dummy edge to the return block.
241 BranchInst::Create(LoopHeaderBB, DummyReturnBB, BoolTrue, BB);
242 Updates.push_back({DominatorTree::Insert, BB, DummyReturnBB});
243 } else { // Conditional branch.
245
246 // Create a new transition block to hold the conditional branch.
247 BasicBlock *TransitionBB = BB->splitBasicBlock(BI, "TransitionBlock");
248
249 Updates.reserve(Updates.size() + 2 * Successors.size() + 2);
250
251 // 'Successors' become successors of TransitionBB instead of BB,
252 // and TransitionBB becomes a single successor of BB.
253 Updates.push_back({DominatorTree::Insert, BB, TransitionBB});
254 for (BasicBlock *Successor : Successors) {
255 Updates.push_back({DominatorTree::Insert, TransitionBB, Successor});
256 Updates.push_back({DominatorTree::Delete, BB, Successor});
257 }
258
259 // Create a branch that will always branch to the transition block and
260 // references DummyReturnBB.
262 BranchInst::Create(TransitionBB, DummyReturnBB, BoolTrue, BB);
263 Updates.push_back({DominatorTree::Insert, BB, DummyReturnBB});
264 }
265 Changed = true;
266 }
267 }
268
269 if (!UnreachableBlocks.empty()) {
270 BasicBlock *UnreachableBlock = nullptr;
271
272 if (UnreachableBlocks.size() == 1) {
273 UnreachableBlock = UnreachableBlocks.front();
274 } else {
275 UnreachableBlock = BasicBlock::Create(F.getContext(),
276 "UnifiedUnreachableBlock", &F);
277 new UnreachableInst(F.getContext(), UnreachableBlock);
278
279 Updates.reserve(Updates.size() + UnreachableBlocks.size());
280 for (BasicBlock *BB : UnreachableBlocks) {
281 // Remove and delete the unreachable inst.
282 BB->getTerminator()->eraseFromParent();
283 BranchInst::Create(UnreachableBlock, BB);
284 Updates.push_back({DominatorTree::Insert, BB, UnreachableBlock});
285 }
286 Changed = true;
287 }
288
289 if (!ReturningBlocks.empty()) {
290 // Don't create a new unreachable inst if we have a return. The
291 // structurizer/annotator can't handle the multiple exits
292
293 Type *RetTy = F.getReturnType();
294 Value *RetVal = RetTy->isVoidTy() ? nullptr : PoisonValue::get(RetTy);
295 // Remove and delete the unreachable inst.
296 UnreachableBlock->getTerminator()->eraseFromParent();
297
298 Function *UnreachableIntrin =
299 Intrinsic::getDeclaration(F.getParent(), Intrinsic::amdgcn_unreachable);
300
301 // Insert a call to an intrinsic tracking that this is an unreachable
302 // point, in case we want to kill the active lanes or something later.
303 CallInst::Create(UnreachableIntrin, {}, "", UnreachableBlock);
304
305 // Don't create a scalar trap. We would only want to trap if this code was
306 // really reached, but a scalar trap would happen even if no lanes
307 // actually reached here.
308 ReturnInst::Create(F.getContext(), RetVal, UnreachableBlock);
309 ReturningBlocks.push_back(UnreachableBlock);
310 Changed = true;
311 }
312 }
313
314 // FIXME: add PDT here once simplifycfg is ready.
315 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
317 DTU.applyUpdates(Updates);
318 Updates.clear();
319
320 // Now handle return blocks.
321 if (ReturningBlocks.empty())
322 return Changed; // No blocks return
323
324 if (ReturningBlocks.size() == 1)
325 return Changed; // Already has a single return block
326
327 unifyReturnBlockSet(F, DTU, ReturningBlocks, "UnifiedReturnBlock");
328 return true;
329}
330
331bool AMDGPUUnifyDivergentExitNodes::runOnFunction(Function &F) {
332 DominatorTree *DT = nullptr;
334 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
335 const auto &PDT =
336 getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree();
337 const auto &UA = getAnalysis<UniformityInfoWrapperPass>().getUniformityInfo();
338 const auto *TranformInfo =
339 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
340 return AMDGPUUnifyDivergentExitNodesImpl(TranformInfo).run(F, DT, PDT, UA);
341}
342
346 DominatorTree *DT = nullptr;
349
350 const auto &PDT = AM.getResult<PostDominatorTreeAnalysis>(F);
351 const auto &UA = AM.getResult<UniformityInfoAnalysis>(F);
352 const auto *TransformInfo = &AM.getResult<TargetIRAnalysis>(F);
353 return AMDGPUUnifyDivergentExitNodesImpl(TransformInfo).run(F, DT, PDT, UA)
356}
static bool isUniformlyReached(const UniformityInfo &UA, BasicBlock &BB)
Unify divergent function exit nodes
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
This file contains the declarations for the subclasses of Constant, which represent the different fla...
return RetTy
Performs the initial survey of the specified function
std::string Name
This file provides various utilities for inspecting and working with the control flow graph in LLVM I...
#define F(x, y, z)
Definition: MD5.cpp:55
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:55
#define INITIALIZE_PASS_END(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:59
#define INITIALIZE_PASS_BEGIN(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:52
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
This file defines the SmallPtrSet class.
This file defines the SmallVector class.
This pass exposes codegen information to IR-level passes.
LLVM IR instance of the generic uniformity analysis.
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
A container for analyses that lazily runs them and caches their results.
Definition: PassManager.h:321
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:473
Represent the analysis usage information of a pass.
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...
Definition: ArrayRef.h:41
size_t size() const
size - Get the array size.
Definition: ArrayRef.h:165
LLVM Basic Block Representation.
Definition: BasicBlock.h:60
static BasicBlock * Create(LLVMContext &Context, const Twine &Name="", Function *Parent=nullptr, BasicBlock *InsertBefore=nullptr)
Creates a new BasicBlock.
Definition: BasicBlock.h:199
BasicBlock * splitBasicBlock(iterator I, const Twine &BBName="", bool Before=false)
Split the basic block into two basic blocks at the specified instruction.
Definition: BasicBlock.cpp:570
SymbolTableList< BasicBlock >::iterator eraseFromParent()
Unlink 'this' from the containing function and delete it.
Definition: BasicBlock.cpp:276
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.h:221
Conditional or Unconditional Branch instruction.
static BranchInst * Create(BasicBlock *IfTrue, BasicBlock::iterator InsertBefore)
static CallInst * Create(FunctionType *Ty, Value *F, const Twine &NameStr, BasicBlock::iterator InsertBefore)
This is the shared class of boolean and integer constants.
Definition: Constants.h:80
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:849
void applyUpdates(ArrayRef< DominatorTree::UpdateType > Updates)
Submit updates to all available trees.
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:279
iterator_range< root_iterator > roots()
size_t root_size() const
NodeT * getRoot() const
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:317
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree.
Definition: Dominators.h:162
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:311
virtual bool runOnFunction(Function &F)=0
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass.
bool isUniform(ConstValueRefT V) const
Whether V is uniform/non-divergent.
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:2667
InstListType::iterator eraseFromParent()
This method unlinks 'this' from the containing basic block and deletes it.
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
virtual void getAnalysisUsage(AnalysisUsage &) const
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: Pass.cpp:98
static PoisonValue * get(Type *T)
Static factory methods - Return an 'poison' object of the specified type.
Definition: Constants.cpp:1827
Analysis pass which computes a PostDominatorTree.
PostDominatorTree Class - Concrete subclass of DominatorTree that is used to compute the post-dominat...
A set of analyses that are preserved following a run of a transformation pass.
Definition: Analysis.h:109
static PreservedAnalyses none()
Convenience factory function for the empty preserved set.
Definition: Analysis.h:112
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: Analysis.h:115
static ReturnInst * Create(LLVMContext &C, Value *retVal, BasicBlock::iterator InsertBefore)
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:342
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements.
Definition: SmallPtrSet.h:427
bool empty() const
Definition: SmallVector.h:94
size_t size() const
Definition: SmallVector.h:91
void push_back(const T &Elt)
Definition: SmallVector.h:426
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
Definition: SmallVector.h:1209
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:50
Analysis pass providing the TargetTransformInfo.
Result run(const Function &F, FunctionAnalysisManager &)
Wrapper pass for TargetTransformInfo.
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:45
Analysis pass which computes UniformityInfo.
Legacy analysis pass which computes a CycleInfo.
This function has undefined behavior.
LLVM Value Representation.
Definition: Value.h:74
unsigned ID
LLVM IR allows to use arbitrary numbers as calling convention identifiers.
Definition: CallingConv.h:24
Function * getDeclaration(Module *M, ID id, ArrayRef< Type * > Tys=std::nullopt)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
Definition: Function.cpp:1469
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
bool hasOnlySimpleTerminator(const Function &F)
auto successors(const MachineBasicBlock *BB)
char & AMDGPUUnifyDivergentExitNodesID
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:1729
cl::opt< bool > RequireAndPreserveDomTree
This function is used to do simplification of a CFG.
bool simplifyCFG(BasicBlock *BB, const TargetTransformInfo &TTI, DomTreeUpdater *DTU=nullptr, const SimplifyCFGOptions &Options={}, ArrayRef< WeakVH > LoopHeaders={})
char & BreakCriticalEdgesID
auto predecessors(const MachineBasicBlock *BB)
void initializeAMDGPUUnifyDivergentExitNodesPass(PassRegistry &)