1 //===-- Sink.cpp - Code Sinking -------------------------------------------===// 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 pass moves instructions into successor blocks, when possible, so that 10 // they aren't executed on paths where their results aren't needed. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Transforms/Scalar/Sink.h" 15 #include "llvm/ADT/Statistic.h" 16 #include "llvm/Analysis/AliasAnalysis.h" 17 #include "llvm/Analysis/LoopInfo.h" 18 #include "llvm/Analysis/ValueTracking.h" 19 #include "llvm/IR/CFG.h" 20 #include "llvm/IR/DataLayout.h" 21 #include "llvm/IR/Dominators.h" 22 #include "llvm/IR/IntrinsicInst.h" 23 #include "llvm/IR/Module.h" 24 #include "llvm/InitializePasses.h" 25 #include "llvm/Support/Debug.h" 26 #include "llvm/Support/raw_ostream.h" 27 #include "llvm/Transforms/Scalar.h" 28 using namespace llvm; 29 30 #define DEBUG_TYPE "sink" 31 32 STATISTIC(NumSunk, "Number of instructions sunk"); 33 STATISTIC(NumSinkIter, "Number of sinking iterations"); 34 35 static bool isSafeToMove(Instruction *Inst, AliasAnalysis &AA, 36 SmallPtrSetImpl<Instruction *> &Stores) { 37 38 if (Inst->mayWriteToMemory()) { 39 Stores.insert(Inst); 40 return false; 41 } 42 43 if (LoadInst *L = dyn_cast<LoadInst>(Inst)) { 44 MemoryLocation Loc = MemoryLocation::get(L); 45 for (Instruction *S : Stores) 46 if (isModSet(AA.getModRefInfo(S, Loc))) 47 return false; 48 } 49 50 if (Inst->isTerminator() || isa<PHINode>(Inst) || Inst->isEHPad() || 51 Inst->mayThrow()) 52 return false; 53 54 if (auto *Call = dyn_cast<CallBase>(Inst)) { 55 // Convergent operations cannot be made control-dependent on additional 56 // values. 57 if (Call->isConvergent()) 58 return false; 59 60 for (Instruction *S : Stores) 61 if (isModSet(AA.getModRefInfo(S, Call))) 62 return false; 63 } 64 65 return true; 66 } 67 68 /// IsAcceptableTarget - Return true if it is possible to sink the instruction 69 /// in the specified basic block. 70 static bool IsAcceptableTarget(Instruction *Inst, BasicBlock *SuccToSinkTo, 71 DominatorTree &DT, LoopInfo &LI) { 72 assert(Inst && "Instruction to be sunk is null"); 73 assert(SuccToSinkTo && "Candidate sink target is null"); 74 75 // It's never legal to sink an instruction into a block which terminates in an 76 // EH-pad. 77 if (SuccToSinkTo->getTerminator()->isExceptionalTerminator()) 78 return false; 79 80 // If the block has multiple predecessors, this would introduce computation 81 // on different code paths. We could split the critical edge, but for now we 82 // just punt. 83 // FIXME: Split critical edges if not backedges. 84 if (SuccToSinkTo->getUniquePredecessor() != Inst->getParent()) { 85 // We cannot sink a load across a critical edge - there may be stores in 86 // other code paths. 87 if (Inst->mayReadFromMemory()) 88 return false; 89 90 // We don't want to sink across a critical edge if we don't dominate the 91 // successor. We could be introducing calculations to new code paths. 92 if (!DT.dominates(Inst->getParent(), SuccToSinkTo)) 93 return false; 94 95 // Don't sink instructions into a loop. 96 Loop *succ = LI.getLoopFor(SuccToSinkTo); 97 Loop *cur = LI.getLoopFor(Inst->getParent()); 98 if (succ != nullptr && succ != cur) 99 return false; 100 } 101 102 return true; 103 } 104 105 /// SinkInstruction - Determine whether it is safe to sink the specified machine 106 /// instruction out of its current block into a successor. 107 static bool SinkInstruction(Instruction *Inst, 108 SmallPtrSetImpl<Instruction *> &Stores, 109 DominatorTree &DT, LoopInfo &LI, AAResults &AA) { 110 111 // Don't sink static alloca instructions. CodeGen assumes allocas outside the 112 // entry block are dynamically sized stack objects. 113 if (AllocaInst *AI = dyn_cast<AllocaInst>(Inst)) 114 if (AI->isStaticAlloca()) 115 return false; 116 117 // Check if it's safe to move the instruction. 118 if (!isSafeToMove(Inst, AA, Stores)) 119 return false; 120 121 // FIXME: This should include support for sinking instructions within the 122 // block they are currently in to shorten the live ranges. We often get 123 // instructions sunk into the top of a large block, but it would be better to 124 // also sink them down before their first use in the block. This xform has to 125 // be careful not to *increase* register pressure though, e.g. sinking 126 // "x = y + z" down if it kills y and z would increase the live ranges of y 127 // and z and only shrink the live range of x. 128 129 // SuccToSinkTo - This is the successor to sink this instruction to, once we 130 // decide. 131 BasicBlock *SuccToSinkTo = nullptr; 132 133 // Find the nearest common dominator of all users as the candidate. 134 BasicBlock *BB = Inst->getParent(); 135 for (Use &U : Inst->uses()) { 136 Instruction *UseInst = cast<Instruction>(U.getUser()); 137 BasicBlock *UseBlock = UseInst->getParent(); 138 // Don't worry about dead users. 139 if (!DT.isReachableFromEntry(UseBlock)) 140 continue; 141 if (PHINode *PN = dyn_cast<PHINode>(UseInst)) { 142 // PHI nodes use the operand in the predecessor block, not the block with 143 // the PHI. 144 unsigned Num = PHINode::getIncomingValueNumForOperand(U.getOperandNo()); 145 UseBlock = PN->getIncomingBlock(Num); 146 } 147 if (SuccToSinkTo) 148 SuccToSinkTo = DT.findNearestCommonDominator(SuccToSinkTo, UseBlock); 149 else 150 SuccToSinkTo = UseBlock; 151 // The current basic block needs to dominate the candidate. 152 if (!DT.dominates(BB, SuccToSinkTo)) 153 return false; 154 } 155 156 if (SuccToSinkTo) { 157 // The nearest common dominator may be in a parent loop of BB, which may not 158 // be beneficial. Find an ancestor. 159 while (SuccToSinkTo != BB && 160 !IsAcceptableTarget(Inst, SuccToSinkTo, DT, LI)) 161 SuccToSinkTo = DT.getNode(SuccToSinkTo)->getIDom()->getBlock(); 162 if (SuccToSinkTo == BB) 163 SuccToSinkTo = nullptr; 164 } 165 166 // If we couldn't find a block to sink to, ignore this instruction. 167 if (!SuccToSinkTo) 168 return false; 169 170 LLVM_DEBUG(dbgs() << "Sink" << *Inst << " ("; 171 Inst->getParent()->printAsOperand(dbgs(), false); dbgs() << " -> "; 172 SuccToSinkTo->printAsOperand(dbgs(), false); dbgs() << ")\n"); 173 174 // Move the instruction. 175 Inst->moveBefore(&*SuccToSinkTo->getFirstInsertionPt()); 176 return true; 177 } 178 179 static bool ProcessBlock(BasicBlock &BB, DominatorTree &DT, LoopInfo &LI, 180 AAResults &AA) { 181 // Can't sink anything out of a block that has less than two successors. 182 if (BB.getTerminator()->getNumSuccessors() <= 1) return false; 183 184 // Don't bother sinking code out of unreachable blocks. In addition to being 185 // unprofitable, it can also lead to infinite looping, because in an 186 // unreachable loop there may be nowhere to stop. 187 if (!DT.isReachableFromEntry(&BB)) return false; 188 189 bool MadeChange = false; 190 191 // Walk the basic block bottom-up. Remember if we saw a store. 192 BasicBlock::iterator I = BB.end(); 193 --I; 194 bool ProcessedBegin = false; 195 SmallPtrSet<Instruction *, 8> Stores; 196 do { 197 Instruction *Inst = &*I; // The instruction to sink. 198 199 // Predecrement I (if it's not begin) so that it isn't invalidated by 200 // sinking. 201 ProcessedBegin = I == BB.begin(); 202 if (!ProcessedBegin) 203 --I; 204 205 if (isa<DbgInfoIntrinsic>(Inst)) 206 continue; 207 208 if (SinkInstruction(Inst, Stores, DT, LI, AA)) { 209 ++NumSunk; 210 MadeChange = true; 211 } 212 213 // If we just processed the first instruction in the block, we're done. 214 } while (!ProcessedBegin); 215 216 return MadeChange; 217 } 218 219 static bool iterativelySinkInstructions(Function &F, DominatorTree &DT, 220 LoopInfo &LI, AAResults &AA) { 221 bool MadeChange, EverMadeChange = false; 222 223 do { 224 MadeChange = false; 225 LLVM_DEBUG(dbgs() << "Sinking iteration " << NumSinkIter << "\n"); 226 // Process all basic blocks. 227 for (BasicBlock &I : F) 228 MadeChange |= ProcessBlock(I, DT, LI, AA); 229 EverMadeChange |= MadeChange; 230 NumSinkIter++; 231 } while (MadeChange); 232 233 return EverMadeChange; 234 } 235 236 PreservedAnalyses SinkingPass::run(Function &F, FunctionAnalysisManager &AM) { 237 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 238 auto &LI = AM.getResult<LoopAnalysis>(F); 239 auto &AA = AM.getResult<AAManager>(F); 240 241 if (!iterativelySinkInstructions(F, DT, LI, AA)) 242 return PreservedAnalyses::all(); 243 244 PreservedAnalyses PA; 245 PA.preserveSet<CFGAnalyses>(); 246 return PA; 247 } 248 249 namespace { 250 class SinkingLegacyPass : public FunctionPass { 251 public: 252 static char ID; // Pass identification 253 SinkingLegacyPass() : FunctionPass(ID) { 254 initializeSinkingLegacyPassPass(*PassRegistry::getPassRegistry()); 255 } 256 257 bool runOnFunction(Function &F) override { 258 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 259 auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 260 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); 261 262 return iterativelySinkInstructions(F, DT, LI, AA); 263 } 264 265 void getAnalysisUsage(AnalysisUsage &AU) const override { 266 AU.setPreservesCFG(); 267 FunctionPass::getAnalysisUsage(AU); 268 AU.addRequired<AAResultsWrapperPass>(); 269 AU.addRequired<DominatorTreeWrapperPass>(); 270 AU.addRequired<LoopInfoWrapperPass>(); 271 AU.addPreserved<DominatorTreeWrapperPass>(); 272 AU.addPreserved<LoopInfoWrapperPass>(); 273 } 274 }; 275 } // end anonymous namespace 276 277 char SinkingLegacyPass::ID = 0; 278 INITIALIZE_PASS_BEGIN(SinkingLegacyPass, "sink", "Code sinking", false, false) 279 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 280 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 281 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 282 INITIALIZE_PASS_END(SinkingLegacyPass, "sink", "Code sinking", false, false) 283 284 FunctionPass *llvm::createSinkingPass() { return new SinkingLegacyPass(); } 285