1 //===- MergedLoadStoreMotion.cpp - merge and hoist/sink load/stores -------===// 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 //! \file 10 //! This pass performs merges of loads and stores on both sides of a 11 // diamond (hammock). It hoists the loads and sinks the stores. 12 // 13 // The algorithm iteratively hoists two loads to the same address out of a 14 // diamond (hammock) and merges them into a single load in the header. Similar 15 // it sinks and merges two stores to the tail block (footer). The algorithm 16 // iterates over the instructions of one side of the diamond and attempts to 17 // find a matching load/store on the other side. New tail/footer block may be 18 // insterted if the tail/footer block has more predecessors (not only the two 19 // predecessors that are forming the diamond). It hoists / sinks when it thinks 20 // it safe to do so. This optimization helps with eg. hiding load latencies, 21 // triggering if-conversion, and reducing static code size. 22 // 23 // NOTE: This code no longer performs load hoisting, it is subsumed by GVNHoist. 24 // 25 //===----------------------------------------------------------------------===// 26 // 27 // 28 // Example: 29 // Diamond shaped code before merge: 30 // 31 // header: 32 // br %cond, label %if.then, label %if.else 33 // + + 34 // + + 35 // + + 36 // if.then: if.else: 37 // %lt = load %addr_l %le = load %addr_l 38 // <use %lt> <use %le> 39 // <...> <...> 40 // store %st, %addr_s store %se, %addr_s 41 // br label %if.end br label %if.end 42 // + + 43 // + + 44 // + + 45 // if.end ("footer"): 46 // <...> 47 // 48 // Diamond shaped code after merge: 49 // 50 // header: 51 // %l = load %addr_l 52 // br %cond, label %if.then, label %if.else 53 // + + 54 // + + 55 // + + 56 // if.then: if.else: 57 // <use %l> <use %l> 58 // <...> <...> 59 // br label %if.end br label %if.end 60 // + + 61 // + + 62 // + + 63 // if.end ("footer"): 64 // %s.sink = phi [%st, if.then], [%se, if.else] 65 // <...> 66 // store %s.sink, %addr_s 67 // <...> 68 // 69 // 70 //===----------------------- TODO -----------------------------------------===// 71 // 72 // 1) Generalize to regions other than diamonds 73 // 2) Be more aggressive merging memory operations 74 // Note that both changes require register pressure control 75 // 76 //===----------------------------------------------------------------------===// 77 78 #include "llvm/Transforms/Scalar/MergedLoadStoreMotion.h" 79 #include "llvm/ADT/Statistic.h" 80 #include "llvm/Analysis/AliasAnalysis.h" 81 #include "llvm/Analysis/CFG.h" 82 #include "llvm/Analysis/GlobalsModRef.h" 83 #include "llvm/Analysis/Loads.h" 84 #include "llvm/Analysis/ValueTracking.h" 85 #include "llvm/IR/Metadata.h" 86 #include "llvm/InitializePasses.h" 87 #include "llvm/Support/Debug.h" 88 #include "llvm/Support/raw_ostream.h" 89 #include "llvm/Transforms/Scalar.h" 90 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 91 92 using namespace llvm; 93 94 #define DEBUG_TYPE "mldst-motion" 95 96 namespace { 97 //===----------------------------------------------------------------------===// 98 // MergedLoadStoreMotion Pass 99 //===----------------------------------------------------------------------===// 100 class MergedLoadStoreMotion { 101 AliasAnalysis *AA = nullptr; 102 103 // The mergeLoad/Store algorithms could have Size0 * Size1 complexity, 104 // where Size0 and Size1 are the #instructions on the two sides of 105 // the diamond. The constant chosen here is arbitrary. Compiler Time 106 // Control is enforced by the check Size0 * Size1 < MagicCompileTimeControl. 107 const int MagicCompileTimeControl = 250; 108 109 const bool SplitFooterBB; 110 public: 111 MergedLoadStoreMotion(bool SplitFooterBB) : SplitFooterBB(SplitFooterBB) {} 112 bool run(Function &F, AliasAnalysis &AA); 113 114 private: 115 BasicBlock *getDiamondTail(BasicBlock *BB); 116 bool isDiamondHead(BasicBlock *BB); 117 // Routines for sinking stores 118 StoreInst *canSinkFromBlock(BasicBlock *BB, StoreInst *SI); 119 PHINode *getPHIOperand(BasicBlock *BB, StoreInst *S0, StoreInst *S1); 120 bool isStoreSinkBarrierInRange(const Instruction &Start, 121 const Instruction &End, MemoryLocation Loc); 122 bool canSinkStoresAndGEPs(StoreInst *S0, StoreInst *S1) const; 123 void sinkStoresAndGEPs(BasicBlock *BB, StoreInst *SinkCand, 124 StoreInst *ElseInst); 125 bool mergeStores(BasicBlock *BB); 126 }; 127 } // end anonymous namespace 128 129 /// 130 /// Return tail block of a diamond. 131 /// 132 BasicBlock *MergedLoadStoreMotion::getDiamondTail(BasicBlock *BB) { 133 assert(isDiamondHead(BB) && "Basic block is not head of a diamond"); 134 return BB->getTerminator()->getSuccessor(0)->getSingleSuccessor(); 135 } 136 137 /// 138 /// True when BB is the head of a diamond (hammock) 139 /// 140 bool MergedLoadStoreMotion::isDiamondHead(BasicBlock *BB) { 141 if (!BB) 142 return false; 143 auto *BI = dyn_cast<BranchInst>(BB->getTerminator()); 144 if (!BI || !BI->isConditional()) 145 return false; 146 147 BasicBlock *Succ0 = BI->getSuccessor(0); 148 BasicBlock *Succ1 = BI->getSuccessor(1); 149 150 if (!Succ0->getSinglePredecessor()) 151 return false; 152 if (!Succ1->getSinglePredecessor()) 153 return false; 154 155 BasicBlock *Succ0Succ = Succ0->getSingleSuccessor(); 156 BasicBlock *Succ1Succ = Succ1->getSingleSuccessor(); 157 // Ignore triangles. 158 if (!Succ0Succ || !Succ1Succ || Succ0Succ != Succ1Succ) 159 return false; 160 return true; 161 } 162 163 164 /// 165 /// True when instruction is a sink barrier for a store 166 /// located in Loc 167 /// 168 /// Whenever an instruction could possibly read or modify the 169 /// value being stored or protect against the store from 170 /// happening it is considered a sink barrier. 171 /// 172 bool MergedLoadStoreMotion::isStoreSinkBarrierInRange(const Instruction &Start, 173 const Instruction &End, 174 MemoryLocation Loc) { 175 for (const Instruction &Inst : 176 make_range(Start.getIterator(), End.getIterator())) 177 if (Inst.mayThrow()) 178 return true; 179 return AA->canInstructionRangeModRef(Start, End, Loc, ModRefInfo::ModRef); 180 } 181 182 /// 183 /// Check if \p BB contains a store to the same address as \p SI 184 /// 185 /// \return The store in \p when it is safe to sink. Otherwise return Null. 186 /// 187 StoreInst *MergedLoadStoreMotion::canSinkFromBlock(BasicBlock *BB1, 188 StoreInst *Store0) { 189 LLVM_DEBUG(dbgs() << "can Sink? : "; Store0->dump(); dbgs() << "\n"); 190 BasicBlock *BB0 = Store0->getParent(); 191 for (Instruction &Inst : reverse(*BB1)) { 192 auto *Store1 = dyn_cast<StoreInst>(&Inst); 193 if (!Store1) 194 continue; 195 196 MemoryLocation Loc0 = MemoryLocation::get(Store0); 197 MemoryLocation Loc1 = MemoryLocation::get(Store1); 198 if (AA->isMustAlias(Loc0, Loc1) && Store0->isSameOperationAs(Store1) && 199 !isStoreSinkBarrierInRange(*Store1->getNextNode(), BB1->back(), Loc1) && 200 !isStoreSinkBarrierInRange(*Store0->getNextNode(), BB0->back(), Loc0)) { 201 return Store1; 202 } 203 } 204 return nullptr; 205 } 206 207 /// 208 /// Create a PHI node in BB for the operands of S0 and S1 209 /// 210 PHINode *MergedLoadStoreMotion::getPHIOperand(BasicBlock *BB, StoreInst *S0, 211 StoreInst *S1) { 212 // Create a phi if the values mismatch. 213 Value *Opd1 = S0->getValueOperand(); 214 Value *Opd2 = S1->getValueOperand(); 215 if (Opd1 == Opd2) 216 return nullptr; 217 218 auto *NewPN = PHINode::Create(Opd1->getType(), 2, Opd2->getName() + ".sink", 219 &BB->front()); 220 NewPN->applyMergedLocation(S0->getDebugLoc(), S1->getDebugLoc()); 221 NewPN->addIncoming(Opd1, S0->getParent()); 222 NewPN->addIncoming(Opd2, S1->getParent()); 223 return NewPN; 224 } 225 226 /// 227 /// Check if 2 stores can be sunk together with corresponding GEPs 228 /// 229 bool MergedLoadStoreMotion::canSinkStoresAndGEPs(StoreInst *S0, 230 StoreInst *S1) const { 231 auto *A0 = dyn_cast<Instruction>(S0->getPointerOperand()); 232 auto *A1 = dyn_cast<Instruction>(S1->getPointerOperand()); 233 return A0 && A1 && A0->isIdenticalTo(A1) && A0->hasOneUse() && 234 (A0->getParent() == S0->getParent()) && A1->hasOneUse() && 235 (A1->getParent() == S1->getParent()) && isa<GetElementPtrInst>(A0); 236 } 237 238 /// 239 /// Merge two stores to same address and sink into \p BB 240 /// 241 /// Also sinks GEP instruction computing the store address 242 /// 243 void MergedLoadStoreMotion::sinkStoresAndGEPs(BasicBlock *BB, StoreInst *S0, 244 StoreInst *S1) { 245 // Only one definition? 246 auto *A0 = dyn_cast<Instruction>(S0->getPointerOperand()); 247 auto *A1 = dyn_cast<Instruction>(S1->getPointerOperand()); 248 LLVM_DEBUG(dbgs() << "Sink Instruction into BB \n"; BB->dump(); 249 dbgs() << "Instruction Left\n"; S0->dump(); dbgs() << "\n"; 250 dbgs() << "Instruction Right\n"; S1->dump(); dbgs() << "\n"); 251 // Hoist the instruction. 252 BasicBlock::iterator InsertPt = BB->getFirstInsertionPt(); 253 // Intersect optional metadata. 254 S0->andIRFlags(S1); 255 S0->dropUnknownNonDebugMetadata(); 256 257 // Create the new store to be inserted at the join point. 258 StoreInst *SNew = cast<StoreInst>(S0->clone()); 259 Instruction *ANew = A0->clone(); 260 SNew->insertBefore(&*InsertPt); 261 ANew->insertBefore(SNew); 262 263 assert(S0->getParent() == A0->getParent()); 264 assert(S1->getParent() == A1->getParent()); 265 266 // New PHI operand? Use it. 267 if (PHINode *NewPN = getPHIOperand(BB, S0, S1)) 268 SNew->setOperand(0, NewPN); 269 S0->eraseFromParent(); 270 S1->eraseFromParent(); 271 A0->replaceAllUsesWith(ANew); 272 A0->eraseFromParent(); 273 A1->replaceAllUsesWith(ANew); 274 A1->eraseFromParent(); 275 } 276 277 /// 278 /// True when two stores are equivalent and can sink into the footer 279 /// 280 /// Starting from a diamond head block, iterate over the instructions in one 281 /// successor block and try to match a store in the second successor. 282 /// 283 bool MergedLoadStoreMotion::mergeStores(BasicBlock *HeadBB) { 284 285 bool MergedStores = false; 286 BasicBlock *TailBB = getDiamondTail(HeadBB); 287 BasicBlock *SinkBB = TailBB; 288 assert(SinkBB && "Footer of a diamond cannot be empty"); 289 290 succ_iterator SI = succ_begin(HeadBB); 291 assert(SI != succ_end(HeadBB) && "Diamond head cannot have zero successors"); 292 BasicBlock *Pred0 = *SI; 293 ++SI; 294 assert(SI != succ_end(HeadBB) && "Diamond head cannot have single successor"); 295 BasicBlock *Pred1 = *SI; 296 // tail block of a diamond/hammock? 297 if (Pred0 == Pred1) 298 return false; // No. 299 // bail out early if we can not merge into the footer BB 300 if (!SplitFooterBB && TailBB->hasNPredecessorsOrMore(3)) 301 return false; 302 // #Instructions in Pred1 for Compile Time Control 303 auto InstsNoDbg = Pred1->instructionsWithoutDebug(); 304 int Size1 = std::distance(InstsNoDbg.begin(), InstsNoDbg.end()); 305 int NStores = 0; 306 307 for (BasicBlock::reverse_iterator RBI = Pred0->rbegin(), RBE = Pred0->rend(); 308 RBI != RBE;) { 309 310 Instruction *I = &*RBI; 311 ++RBI; 312 313 // Don't sink non-simple (atomic, volatile) stores. 314 auto *S0 = dyn_cast<StoreInst>(I); 315 if (!S0 || !S0->isSimple()) 316 continue; 317 318 ++NStores; 319 if (NStores * Size1 >= MagicCompileTimeControl) 320 break; 321 if (StoreInst *S1 = canSinkFromBlock(Pred1, S0)) { 322 if (!canSinkStoresAndGEPs(S0, S1)) 323 // Don't attempt to sink below stores that had to stick around 324 // But after removal of a store and some of its feeding 325 // instruction search again from the beginning since the iterator 326 // is likely stale at this point. 327 break; 328 329 if (SinkBB == TailBB && TailBB->hasNPredecessorsOrMore(3)) { 330 // We have more than 2 predecessors. Insert a new block 331 // postdominating 2 predecessors we're going to sink from. 332 SinkBB = SplitBlockPredecessors(TailBB, {Pred0, Pred1}, ".sink.split"); 333 if (!SinkBB) 334 break; 335 } 336 337 MergedStores = true; 338 sinkStoresAndGEPs(SinkBB, S0, S1); 339 RBI = Pred0->rbegin(); 340 RBE = Pred0->rend(); 341 LLVM_DEBUG(dbgs() << "Search again\n"; Instruction *I = &*RBI; I->dump()); 342 } 343 } 344 return MergedStores; 345 } 346 347 bool MergedLoadStoreMotion::run(Function &F, AliasAnalysis &AA) { 348 this->AA = &AA; 349 350 bool Changed = false; 351 LLVM_DEBUG(dbgs() << "Instruction Merger\n"); 352 353 // Merge unconditional branches, allowing PRE to catch more 354 // optimization opportunities. 355 // This loop doesn't care about newly inserted/split blocks 356 // since they never will be diamond heads. 357 for (BasicBlock &BB : make_early_inc_range(F)) 358 // Hoist equivalent loads and sink stores 359 // outside diamonds when possible 360 if (isDiamondHead(&BB)) 361 Changed |= mergeStores(&BB); 362 return Changed; 363 } 364 365 namespace { 366 class MergedLoadStoreMotionLegacyPass : public FunctionPass { 367 const bool SplitFooterBB; 368 public: 369 static char ID; // Pass identification, replacement for typeid 370 MergedLoadStoreMotionLegacyPass(bool SplitFooterBB = false) 371 : FunctionPass(ID), SplitFooterBB(SplitFooterBB) { 372 initializeMergedLoadStoreMotionLegacyPassPass( 373 *PassRegistry::getPassRegistry()); 374 } 375 376 /// 377 /// Run the transformation for each function 378 /// 379 bool runOnFunction(Function &F) override { 380 if (skipFunction(F)) 381 return false; 382 MergedLoadStoreMotion Impl(SplitFooterBB); 383 return Impl.run(F, getAnalysis<AAResultsWrapperPass>().getAAResults()); 384 } 385 386 private: 387 void getAnalysisUsage(AnalysisUsage &AU) const override { 388 if (!SplitFooterBB) 389 AU.setPreservesCFG(); 390 AU.addRequired<AAResultsWrapperPass>(); 391 AU.addPreserved<GlobalsAAWrapperPass>(); 392 } 393 }; 394 395 char MergedLoadStoreMotionLegacyPass::ID = 0; 396 } // anonymous namespace 397 398 /// 399 /// createMergedLoadStoreMotionPass - The public interface to this file. 400 /// 401 FunctionPass *llvm::createMergedLoadStoreMotionPass(bool SplitFooterBB) { 402 return new MergedLoadStoreMotionLegacyPass(SplitFooterBB); 403 } 404 405 INITIALIZE_PASS_BEGIN(MergedLoadStoreMotionLegacyPass, "mldst-motion", 406 "MergedLoadStoreMotion", false, false) 407 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 408 INITIALIZE_PASS_END(MergedLoadStoreMotionLegacyPass, "mldst-motion", 409 "MergedLoadStoreMotion", false, false) 410 411 PreservedAnalyses 412 MergedLoadStoreMotionPass::run(Function &F, FunctionAnalysisManager &AM) { 413 MergedLoadStoreMotion Impl(Options.SplitFooterBB); 414 auto &AA = AM.getResult<AAManager>(F); 415 if (!Impl.run(F, AA)) 416 return PreservedAnalyses::all(); 417 418 PreservedAnalyses PA; 419 if (!Options.SplitFooterBB) 420 PA.preserveSet<CFGAnalyses>(); 421 PA.preserve<GlobalsAA>(); 422 return PA; 423 } 424