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 UnifyDivergentExitNodes 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 
22 #include "AMDGPU.h"
23 #include "llvm/ADT/ArrayRef.h"
24 #include "llvm/ADT/SmallPtrSet.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/StringRef.h"
27 #include "llvm/Analysis/LegacyDivergenceAnalysis.h"
28 #include "llvm/Analysis/PostDominators.h"
29 #include "llvm/Analysis/TargetTransformInfo.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/CFG.h"
32 #include "llvm/IR/Constants.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/InstrTypes.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/Intrinsics.h"
37 #include "llvm/IR/IRBuilder.h"
38 #include "llvm/IR/Type.h"
39 #include "llvm/InitializePasses.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/Casting.h"
42 #include "llvm/Transforms/Scalar.h"
43 #include "llvm/Transforms/Utils.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 
46 using namespace llvm;
47 
48 #define DEBUG_TYPE "amdgpu-unify-divergent-exit-nodes"
49 
50 namespace {
51 
52 class AMDGPUUnifyDivergentExitNodes : public FunctionPass {
53 public:
54   static char ID; // Pass identification, replacement for typeid
55 
56   AMDGPUUnifyDivergentExitNodes() : FunctionPass(ID) {
57     initializeAMDGPUUnifyDivergentExitNodesPass(*PassRegistry::getPassRegistry());
58   }
59 
60   // We can preserve non-critical-edgeness when we unify function exit nodes
61   void getAnalysisUsage(AnalysisUsage &AU) const override;
62   bool runOnFunction(Function &F) override;
63 };
64 
65 } // end anonymous namespace
66 
67 char AMDGPUUnifyDivergentExitNodes::ID = 0;
68 
69 char &llvm::AMDGPUUnifyDivergentExitNodesID = AMDGPUUnifyDivergentExitNodes::ID;
70 
71 INITIALIZE_PASS_BEGIN(AMDGPUUnifyDivergentExitNodes, DEBUG_TYPE,
72                      "Unify divergent function exit nodes", false, false)
73 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
74 INITIALIZE_PASS_DEPENDENCY(LegacyDivergenceAnalysis)
75 INITIALIZE_PASS_END(AMDGPUUnifyDivergentExitNodes, DEBUG_TYPE,
76                     "Unify divergent function exit nodes", false, false)
77 
78 void AMDGPUUnifyDivergentExitNodes::getAnalysisUsage(AnalysisUsage &AU) const{
79   // TODO: Preserve dominator tree.
80   AU.addRequired<PostDominatorTreeWrapperPass>();
81 
82   AU.addRequired<LegacyDivergenceAnalysis>();
83 
84   // No divergent values are changed, only blocks and branch edges.
85   AU.addPreserved<LegacyDivergenceAnalysis>();
86 
87   // We preserve the non-critical-edgeness property
88   AU.addPreservedID(BreakCriticalEdgesID);
89 
90   // This is a cluster of orthogonal Transforms
91   AU.addPreservedID(LowerSwitchID);
92   FunctionPass::getAnalysisUsage(AU);
93 
94   AU.addRequired<TargetTransformInfoWrapperPass>();
95 }
96 
97 /// \returns true if \p BB is reachable through only uniform branches.
98 /// XXX - Is there a more efficient way to find this?
99 static bool isUniformlyReached(const LegacyDivergenceAnalysis &DA,
100                                BasicBlock &BB) {
101   SmallVector<BasicBlock *, 8> Stack;
102   SmallPtrSet<BasicBlock *, 8> Visited;
103 
104   for (BasicBlock *Pred : predecessors(&BB))
105     Stack.push_back(Pred);
106 
107   while (!Stack.empty()) {
108     BasicBlock *Top = Stack.pop_back_val();
109     if (!DA.isUniform(Top->getTerminator()))
110       return false;
111 
112     for (BasicBlock *Pred : predecessors(Top)) {
113       if (Visited.insert(Pred).second)
114         Stack.push_back(Pred);
115     }
116   }
117 
118   return true;
119 }
120 
121 static void removeDoneExport(Function &F) {
122   ConstantInt *BoolFalse = ConstantInt::getFalse(F.getContext());
123   for (BasicBlock &BB : F) {
124     for (Instruction &I : BB) {
125       if (IntrinsicInst *Intrin = llvm::dyn_cast<IntrinsicInst>(&I)) {
126         if (Intrin->getIntrinsicID() == Intrinsic::amdgcn_exp) {
127           Intrin->setArgOperand(6, BoolFalse); // done
128         } else if (Intrin->getIntrinsicID() == Intrinsic::amdgcn_exp_compr) {
129           Intrin->setArgOperand(4, BoolFalse); // done
130         }
131       }
132     }
133   }
134 }
135 
136 static BasicBlock *unifyReturnBlockSet(Function &F,
137                                        ArrayRef<BasicBlock *> ReturningBlocks,
138                                        bool InsertExport,
139                                        const TargetTransformInfo &TTI,
140                                        StringRef Name) {
141   // Otherwise, we need to insert a new basic block into the function, add a PHI
142   // nodes (if the function returns values), and convert all of the return
143   // instructions into unconditional branches.
144   BasicBlock *NewRetBlock = BasicBlock::Create(F.getContext(), Name, &F);
145   IRBuilder<> B(NewRetBlock);
146 
147   if (InsertExport) {
148     // Ensure that there's only one "done" export in the shader by removing the
149     // "done" bit set on the original final export. More than one "done" export
150     // can lead to undefined behavior.
151     removeDoneExport(F);
152 
153     Value *Undef = UndefValue::get(B.getFloatTy());
154     B.CreateIntrinsic(Intrinsic::amdgcn_exp, { B.getFloatTy() },
155                       {
156                         B.getInt32(9), // target, SQ_EXP_NULL
157                         B.getInt32(0), // enabled channels
158                         Undef, Undef, Undef, Undef, // values
159                         B.getTrue(), // done
160                         B.getTrue(), // valid mask
161                       });
162   }
163 
164   PHINode *PN = nullptr;
165   if (F.getReturnType()->isVoidTy()) {
166     B.CreateRetVoid();
167   } else {
168     // If the function doesn't return void... add a PHI node to the block...
169     PN = B.CreatePHI(F.getReturnType(), ReturningBlocks.size(),
170                      "UnifiedRetVal");
171     assert(!InsertExport);
172     B.CreateRet(PN);
173   }
174 
175   // Loop over all of the blocks, replacing the return instruction with an
176   // unconditional branch.
177   for (BasicBlock *BB : ReturningBlocks) {
178     // Add an incoming element to the PHI node for every return instruction that
179     // is merging into this new block...
180     if (PN)
181       PN->addIncoming(BB->getTerminator()->getOperand(0), BB);
182 
183     // Remove and delete the return inst.
184     BB->getTerminator()->eraseFromParent();
185     BranchInst::Create(NewRetBlock, BB);
186   }
187 
188   for (BasicBlock *BB : ReturningBlocks) {
189     // Cleanup possible branch to unconditional branch to the return.
190     simplifyCFG(BB, TTI, {2});
191   }
192 
193   return NewRetBlock;
194 }
195 
196 bool AMDGPUUnifyDivergentExitNodes::runOnFunction(Function &F) {
197   auto &PDT = getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree();
198 
199   // If there's only one exit, we don't need to do anything, unless this is a
200   // pixel shader and that exit is an infinite loop, since we still have to
201   // insert an export in that case.
202   if (PDT.root_size() <= 1 && F.getCallingConv() != CallingConv::AMDGPU_PS)
203     return false;
204 
205   LegacyDivergenceAnalysis &DA = getAnalysis<LegacyDivergenceAnalysis>();
206 
207   // Loop over all of the blocks in a function, tracking all of the blocks that
208   // return.
209   SmallVector<BasicBlock *, 4> ReturningBlocks;
210   SmallVector<BasicBlock *, 4> UniformlyReachedRetBlocks;
211   SmallVector<BasicBlock *, 4> UnreachableBlocks;
212 
213   // Dummy return block for infinite loop.
214   BasicBlock *DummyReturnBB = nullptr;
215 
216   bool InsertExport = false;
217 
218   bool Changed = false;
219   for (BasicBlock *BB : PDT.roots()) {
220     if (isa<ReturnInst>(BB->getTerminator())) {
221       if (!isUniformlyReached(DA, *BB))
222         ReturningBlocks.push_back(BB);
223       else
224         UniformlyReachedRetBlocks.push_back(BB);
225     } else if (isa<UnreachableInst>(BB->getTerminator())) {
226       if (!isUniformlyReached(DA, *BB))
227         UnreachableBlocks.push_back(BB);
228     } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
229 
230       ConstantInt *BoolTrue = ConstantInt::getTrue(F.getContext());
231       if (DummyReturnBB == nullptr) {
232         DummyReturnBB = BasicBlock::Create(F.getContext(),
233                                            "DummyReturnBlock", &F);
234         Type *RetTy = F.getReturnType();
235         Value *RetVal = RetTy->isVoidTy() ? nullptr : UndefValue::get(RetTy);
236 
237         // For pixel shaders, the producer guarantees that an export is
238         // executed before each return instruction. However, if there is an
239         // infinite loop and we insert a return ourselves, we need to uphold
240         // that guarantee by inserting a null export. This can happen e.g. in
241         // an infinite loop with kill instructions, which is supposed to
242         // terminate. However, we don't need to do this if there is a non-void
243         // return value, since then there is an epilog afterwards which will
244         // still export.
245         //
246         // Note: In the case where only some threads enter the infinite loop,
247         // this can result in the null export happening redundantly after the
248         // original exports. However, The last "real" export happens after all
249         // the threads that didn't enter an infinite loop converged, which
250         // means that the only extra threads to execute the null export are
251         // threads that entered the infinite loop, and they only could've
252         // exited through being killed which sets their exec bit to 0.
253         // Therefore, unless there's an actual infinite loop, which can have
254         // invalid results, or there's a kill after the last export, which we
255         // assume the frontend won't do, this export will have the same exec
256         // mask as the last "real" export, and therefore the valid mask will be
257         // overwritten with the same value and will still be correct. Also,
258         // even though this forces an extra unnecessary export wait, we assume
259         // that this happens rare enough in practice to that we don't have to
260         // worry about performance.
261         if (F.getCallingConv() == CallingConv::AMDGPU_PS &&
262             RetTy->isVoidTy()) {
263           InsertExport = true;
264         }
265 
266         ReturnInst::Create(F.getContext(), RetVal, DummyReturnBB);
267         ReturningBlocks.push_back(DummyReturnBB);
268       }
269 
270       if (BI->isUnconditional()) {
271         BasicBlock *LoopHeaderBB = BI->getSuccessor(0);
272         BI->eraseFromParent(); // Delete the unconditional branch.
273         // Add a new conditional branch with a dummy edge to the return block.
274         BranchInst::Create(LoopHeaderBB, DummyReturnBB, BoolTrue, BB);
275       } else { // Conditional branch.
276         // Create a new transition block to hold the conditional branch.
277         BasicBlock *TransitionBB = BB->splitBasicBlock(BI, "TransitionBlock");
278 
279         // Create a branch that will always branch to the transition block and
280         // references DummyReturnBB.
281         BB->getTerminator()->eraseFromParent();
282         BranchInst::Create(TransitionBB, DummyReturnBB, BoolTrue, BB);
283       }
284       Changed = true;
285     }
286   }
287 
288   if (!UnreachableBlocks.empty()) {
289     BasicBlock *UnreachableBlock = nullptr;
290 
291     if (UnreachableBlocks.size() == 1) {
292       UnreachableBlock = UnreachableBlocks.front();
293     } else {
294       UnreachableBlock = BasicBlock::Create(F.getContext(),
295                                             "UnifiedUnreachableBlock", &F);
296       new UnreachableInst(F.getContext(), UnreachableBlock);
297 
298       for (BasicBlock *BB : UnreachableBlocks) {
299         // Remove and delete the unreachable inst.
300         BB->getTerminator()->eraseFromParent();
301         BranchInst::Create(UnreachableBlock, BB);
302       }
303       Changed = true;
304     }
305 
306     if (!ReturningBlocks.empty()) {
307       // Don't create a new unreachable inst if we have a return. The
308       // structurizer/annotator can't handle the multiple exits
309 
310       Type *RetTy = F.getReturnType();
311       Value *RetVal = RetTy->isVoidTy() ? nullptr : UndefValue::get(RetTy);
312       // Remove and delete the unreachable inst.
313       UnreachableBlock->getTerminator()->eraseFromParent();
314 
315       Function *UnreachableIntrin =
316         Intrinsic::getDeclaration(F.getParent(), Intrinsic::amdgcn_unreachable);
317 
318       // Insert a call to an intrinsic tracking that this is an unreachable
319       // point, in case we want to kill the active lanes or something later.
320       CallInst::Create(UnreachableIntrin, {}, "", UnreachableBlock);
321 
322       // Don't create a scalar trap. We would only want to trap if this code was
323       // really reached, but a scalar trap would happen even if no lanes
324       // actually reached here.
325       ReturnInst::Create(F.getContext(), RetVal, UnreachableBlock);
326       ReturningBlocks.push_back(UnreachableBlock);
327       Changed = true;
328     }
329   }
330 
331   // Now handle return blocks.
332   if (ReturningBlocks.empty())
333     return Changed; // No blocks return
334 
335   if (ReturningBlocks.size() == 1 && !InsertExport)
336     return Changed; // Already has a single return block
337 
338   const TargetTransformInfo &TTI
339     = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
340 
341   // Unify returning blocks. If we are going to insert the export it is also
342   // necessary to include blocks that are uniformly reached, because in addition
343   // to inserting the export the "done" bits on existing exports will be cleared
344   // and we do not want to end up with the normal export in a non-unified,
345   // uniformly reached block with the "done" bit cleared.
346   auto BlocksToUnify = std::move(ReturningBlocks);
347   if (InsertExport) {
348     BlocksToUnify.insert(BlocksToUnify.end(), UniformlyReachedRetBlocks.begin(),
349                          UniformlyReachedRetBlocks.end());
350   }
351 
352   unifyReturnBlockSet(F, BlocksToUnify, InsertExport, TTI,
353                       "UnifiedReturnBlock");
354   return true;
355 }
356