1//===- CodeGenPrepare.cpp - Prepare a function for code generation --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass munges the code in the input function to better prepare it for 11// SelectionDAG-based code generation. This works around limitations in it's 12// basic-block-at-a-time approach. It should eventually be removed. 13// 14//===----------------------------------------------------------------------===// 15 16#include "llvm/CodeGen/Passes.h" 17#include "llvm/ADT/DenseMap.h" 18#include "llvm/ADT/SmallSet.h" 19#include "llvm/ADT/Statistic.h" 20#include "llvm/Analysis/InstructionSimplify.h" 21#include "llvm/Analysis/TargetLibraryInfo.h" 22#include "llvm/Analysis/TargetTransformInfo.h" 23#include "llvm/IR/CallSite.h" 24#include "llvm/IR/Constants.h" 25#include "llvm/IR/DataLayout.h" 26#include "llvm/IR/DerivedTypes.h" 27#include "llvm/IR/Dominators.h" 28#include "llvm/IR/Function.h" 29#include "llvm/IR/GetElementPtrTypeIterator.h" 30#include "llvm/IR/IRBuilder.h" 31#include "llvm/IR/InlineAsm.h" 32#include "llvm/IR/Instructions.h" 33#include "llvm/IR/IntrinsicInst.h" 34#include "llvm/IR/MDBuilder.h" 35#include "llvm/IR/PatternMatch.h" 36#include "llvm/IR/Statepoint.h" 37#include "llvm/IR/ValueHandle.h" 38#include "llvm/IR/ValueMap.h" 39#include "llvm/Pass.h" 40#include "llvm/Support/CommandLine.h" 41#include "llvm/Support/Debug.h" 42#include "llvm/Support/raw_ostream.h" 43#include "llvm/Target/TargetLowering.h" 44#include "llvm/Target/TargetSubtargetInfo.h" 45#include "llvm/Transforms/Utils/BasicBlockUtils.h" 46#include "llvm/Transforms/Utils/BuildLibCalls.h" 47#include "llvm/Transforms/Utils/BypassSlowDivision.h" 48#include "llvm/Transforms/Utils/Local.h" 49#include "llvm/Transforms/Utils/SimplifyLibCalls.h" 50using namespace llvm; 51using namespace llvm::PatternMatch; 52 53#define DEBUG_TYPE "codegenprepare" 54 55STATISTIC(NumBlocksElim, "Number of blocks eliminated"); 56STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated"); 57STATISTIC(NumGEPsElim, "Number of GEPs converted to casts"); 58STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of " 59 "sunken Cmps"); 60STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses " 61 "of sunken Casts"); 62STATISTIC(NumMemoryInsts, "Number of memory instructions whose address " 63 "computations were sunk"); 64STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads"); 65STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized"); 66STATISTIC(NumRetsDup, "Number of return instructions duplicated"); 67STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved"); 68STATISTIC(NumSelectsExpanded, "Number of selects turned into branches"); 69STATISTIC(NumAndCmpsMoved, "Number of and/cmp's pushed into branches"); 70STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed"); 71 72static cl::opt<bool> DisableBranchOpts( 73 "disable-cgp-branch-opts", cl::Hidden, cl::init(false), 74 cl::desc("Disable branch optimizations in CodeGenPrepare")); 75 76static cl::opt<bool> 77 DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false), 78 cl::desc("Disable GC optimizations in CodeGenPrepare")); 79 80static cl::opt<bool> DisableSelectToBranch( 81 "disable-cgp-select2branch", cl::Hidden, cl::init(false), 82 cl::desc("Disable select to branch conversion.")); 83 84static cl::opt<bool> AddrSinkUsingGEPs( 85 "addr-sink-using-gep", cl::Hidden, cl::init(false), 86 cl::desc("Address sinking in CGP using GEPs.")); 87 88static cl::opt<bool> EnableAndCmpSinking( 89 "enable-andcmp-sinking", cl::Hidden, cl::init(true), 90 cl::desc("Enable sinkinig and/cmp into branches.")); 91 92static cl::opt<bool> DisableStoreExtract( 93 "disable-cgp-store-extract", cl::Hidden, cl::init(false), 94 cl::desc("Disable store(extract) optimizations in CodeGenPrepare")); 95 96static cl::opt<bool> StressStoreExtract( 97 "stress-cgp-store-extract", cl::Hidden, cl::init(false), 98 cl::desc("Stress test store(extract) optimizations in CodeGenPrepare")); 99 100static cl::opt<bool> DisableExtLdPromotion( 101 "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false), 102 cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in " 103 "CodeGenPrepare")); 104 105static cl::opt<bool> StressExtLdPromotion( 106 "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false), 107 cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) " 108 "optimization in CodeGenPrepare")); 109 110namespace { 111typedef SmallPtrSet<Instruction *, 16> SetOfInstrs; 112struct TypeIsSExt { 113 Type *Ty; 114 bool IsSExt; 115 TypeIsSExt(Type *Ty, bool IsSExt) : Ty(Ty), IsSExt(IsSExt) {} 116}; 117typedef DenseMap<Instruction *, TypeIsSExt> InstrToOrigTy; 118class TypePromotionTransaction; 119 120 class CodeGenPrepare : public FunctionPass { 121 /// TLI - Keep a pointer of a TargetLowering to consult for determining 122 /// transformation profitability. 123 const TargetMachine *TM; 124 const TargetLowering *TLI; 125 const TargetTransformInfo *TTI; 126 const TargetLibraryInfo *TLInfo; 127 128 /// CurInstIterator - As we scan instructions optimizing them, this is the 129 /// next instruction to optimize. Xforms that can invalidate this should 130 /// update it. 131 BasicBlock::iterator CurInstIterator; 132 133 /// Keeps track of non-local addresses that have been sunk into a block. 134 /// This allows us to avoid inserting duplicate code for blocks with 135 /// multiple load/stores of the same address. 136 ValueMap<Value*, Value*> SunkAddrs; 137 138 /// Keeps track of all truncates inserted for the current function. 139 SetOfInstrs InsertedTruncsSet; 140 /// Keeps track of the type of the related instruction before their 141 /// promotion for the current function. 142 InstrToOrigTy PromotedInsts; 143 144 /// ModifiedDT - If CFG is modified in anyway. 145 bool ModifiedDT; 146 147 /// OptSize - True if optimizing for size. 148 bool OptSize; 149 150 public: 151 static char ID; // Pass identification, replacement for typeid 152 explicit CodeGenPrepare(const TargetMachine *TM = nullptr) 153 : FunctionPass(ID), TM(TM), TLI(nullptr), TTI(nullptr) { 154 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry()); 155 } 156 bool runOnFunction(Function &F) override; 157 158 const char *getPassName() const override { return "CodeGen Prepare"; } 159 160 void getAnalysisUsage(AnalysisUsage &AU) const override { 161 AU.addPreserved<DominatorTreeWrapperPass>(); 162 AU.addRequired<TargetLibraryInfoWrapperPass>(); 163 AU.addRequired<TargetTransformInfoWrapperPass>(); 164 } 165 166 private: 167 bool EliminateFallThrough(Function &F); 168 bool EliminateMostlyEmptyBlocks(Function &F); 169 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; 170 void EliminateMostlyEmptyBlock(BasicBlock *BB); 171 bool OptimizeBlock(BasicBlock &BB, bool& ModifiedDT); 172 bool OptimizeInst(Instruction *I, bool& ModifiedDT); 173 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy); 174 bool OptimizeInlineAsmInst(CallInst *CS); 175 bool OptimizeCallInst(CallInst *CI, bool& ModifiedDT); 176 bool MoveExtToFormExtLoad(Instruction *&I); 177 bool OptimizeExtUses(Instruction *I); 178 bool OptimizeSelectInst(SelectInst *SI); 179 bool OptimizeShuffleVectorInst(ShuffleVectorInst *SI); 180 bool OptimizeExtractElementInst(Instruction *Inst); 181 bool DupRetToEnableTailCallOpts(BasicBlock *BB); 182 bool PlaceDbgValues(Function &F); 183 bool sinkAndCmp(Function &F); 184 bool ExtLdPromotion(TypePromotionTransaction &TPT, LoadInst *&LI, 185 Instruction *&Inst, 186 const SmallVectorImpl<Instruction *> &Exts, 187 unsigned CreatedInstCost); 188 bool splitBranchCondition(Function &F); 189 bool simplifyOffsetableRelocate(Instruction &I); 190 }; 191} 192 193char CodeGenPrepare::ID = 0; 194INITIALIZE_TM_PASS(CodeGenPrepare, "codegenprepare", 195 "Optimize for code generation", false, false) 196 197FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) { 198 return new CodeGenPrepare(TM); 199} 200 201bool CodeGenPrepare::runOnFunction(Function &F) { 202 if (skipOptnoneFunction(F)) 203 return false; 204 205 bool EverMadeChange = false; 206 // Clear per function information. 207 InsertedTruncsSet.clear(); 208 PromotedInsts.clear(); 209 210 ModifiedDT = false; 211 if (TM) 212 TLI = TM->getSubtargetImpl(F)->getTargetLowering(); 213 TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 214 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 215 OptSize = F.hasFnAttribute(Attribute::OptimizeForSize); 216 217 /// This optimization identifies DIV instructions that can be 218 /// profitably bypassed and carried out with a shorter, faster divide. 219 if (!OptSize && TLI && TLI->isSlowDivBypassed()) { 220 const DenseMap<unsigned int, unsigned int> &BypassWidths = 221 TLI->getBypassSlowDivWidths(); 222 for (Function::iterator I = F.begin(); I != F.end(); I++) 223 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths); 224 } 225 226 // Eliminate blocks that contain only PHI nodes and an 227 // unconditional branch. 228 EverMadeChange |= EliminateMostlyEmptyBlocks(F); 229 230 // llvm.dbg.value is far away from the value then iSel may not be able 231 // handle it properly. iSel will drop llvm.dbg.value if it can not 232 // find a node corresponding to the value. 233 EverMadeChange |= PlaceDbgValues(F); 234 235 // If there is a mask, compare against zero, and branch that can be combined 236 // into a single target instruction, push the mask and compare into branch 237 // users. Do this before OptimizeBlock -> OptimizeInst -> 238 // OptimizeCmpExpression, which perturbs the pattern being searched for. 239 if (!DisableBranchOpts) { 240 EverMadeChange |= sinkAndCmp(F); 241 EverMadeChange |= splitBranchCondition(F); 242 } 243 244 bool MadeChange = true; 245 while (MadeChange) { 246 MadeChange = false; 247 for (Function::iterator I = F.begin(); I != F.end(); ) { 248 BasicBlock *BB = I++; 249 bool ModifiedDTOnIteration = false; 250 MadeChange |= OptimizeBlock(*BB, ModifiedDTOnIteration); 251 252 // Restart BB iteration if the dominator tree of the Function was changed 253 if (ModifiedDTOnIteration) 254 break; 255 } 256 EverMadeChange |= MadeChange; 257 } 258 259 SunkAddrs.clear(); 260 261 if (!DisableBranchOpts) { 262 MadeChange = false; 263 SmallPtrSet<BasicBlock*, 8> WorkList; 264 for (BasicBlock &BB : F) { 265 SmallVector<BasicBlock *, 2> Successors(succ_begin(&BB), succ_end(&BB)); 266 MadeChange |= ConstantFoldTerminator(&BB, true); 267 if (!MadeChange) continue; 268 269 for (SmallVectorImpl<BasicBlock*>::iterator 270 II = Successors.begin(), IE = Successors.end(); II != IE; ++II) 271 if (pred_begin(*II) == pred_end(*II)) 272 WorkList.insert(*II); 273 } 274 275 // Delete the dead blocks and any of their dead successors. 276 MadeChange |= !WorkList.empty(); 277 while (!WorkList.empty()) { 278 BasicBlock *BB = *WorkList.begin(); 279 WorkList.erase(BB); 280 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB)); 281 282 DeleteDeadBlock(BB); 283 284 for (SmallVectorImpl<BasicBlock*>::iterator 285 II = Successors.begin(), IE = Successors.end(); II != IE; ++II) 286 if (pred_begin(*II) == pred_end(*II)) 287 WorkList.insert(*II); 288 } 289 290 // Merge pairs of basic blocks with unconditional branches, connected by 291 // a single edge. 292 if (EverMadeChange || MadeChange) 293 MadeChange |= EliminateFallThrough(F); 294 295 EverMadeChange |= MadeChange; 296 } 297 298 if (!DisableGCOpts) { 299 SmallVector<Instruction *, 2> Statepoints; 300 for (BasicBlock &BB : F) 301 for (Instruction &I : BB) 302 if (isStatepoint(I)) 303 Statepoints.push_back(&I); 304 for (auto &I : Statepoints) 305 EverMadeChange |= simplifyOffsetableRelocate(*I); 306 } 307 308 return EverMadeChange; 309} 310 311/// EliminateFallThrough - Merge basic blocks which are connected 312/// by a single edge, where one of the basic blocks has a single successor 313/// pointing to the other basic block, which has a single predecessor. 314bool CodeGenPrepare::EliminateFallThrough(Function &F) { 315 bool Changed = false; 316 // Scan all of the blocks in the function, except for the entry block. 317 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) { 318 BasicBlock *BB = I++; 319 // If the destination block has a single pred, then this is a trivial 320 // edge, just collapse it. 321 BasicBlock *SinglePred = BB->getSinglePredecessor(); 322 323 // Don't merge if BB's address is taken. 324 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue; 325 326 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator()); 327 if (Term && !Term->isConditional()) { 328 Changed = true; 329 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n"); 330 // Remember if SinglePred was the entry block of the function. 331 // If so, we will need to move BB back to the entry position. 332 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 333 MergeBasicBlockIntoOnlyPred(BB, nullptr); 334 335 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 336 BB->moveBefore(&BB->getParent()->getEntryBlock()); 337 338 // We have erased a block. Update the iterator. 339 I = BB; 340 } 341 } 342 return Changed; 343} 344 345/// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes, 346/// debug info directives, and an unconditional branch. Passes before isel 347/// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for 348/// isel. Start by eliminating these blocks so we can split them the way we 349/// want them. 350bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) { 351 bool MadeChange = false; 352 // Note that this intentionally skips the entry block. 353 for (Function::iterator I = std::next(F.begin()), E = F.end(); I != E;) { 354 BasicBlock *BB = I++; 355 356 // If this block doesn't end with an uncond branch, ignore it. 357 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 358 if (!BI || !BI->isUnconditional()) 359 continue; 360 361 // If the instruction before the branch (skipping debug info) isn't a phi 362 // node, then other stuff is happening here. 363 BasicBlock::iterator BBI = BI; 364 if (BBI != BB->begin()) { 365 --BBI; 366 while (isa<DbgInfoIntrinsic>(BBI)) { 367 if (BBI == BB->begin()) 368 break; 369 --BBI; 370 } 371 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI)) 372 continue; 373 } 374 375 // Do not break infinite loops. 376 BasicBlock *DestBB = BI->getSuccessor(0); 377 if (DestBB == BB) 378 continue; 379 380 if (!CanMergeBlocks(BB, DestBB)) 381 continue; 382 383 EliminateMostlyEmptyBlock(BB); 384 MadeChange = true; 385 } 386 return MadeChange; 387} 388 389/// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a 390/// single uncond branch between them, and BB contains no other non-phi 391/// instructions. 392bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB, 393 const BasicBlock *DestBB) const { 394 // We only want to eliminate blocks whose phi nodes are used by phi nodes in 395 // the successor. If there are more complex condition (e.g. preheaders), 396 // don't mess around with them. 397 BasicBlock::const_iterator BBI = BB->begin(); 398 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 399 for (const User *U : PN->users()) { 400 const Instruction *UI = cast<Instruction>(U); 401 if (UI->getParent() != DestBB || !isa<PHINode>(UI)) 402 return false; 403 // If User is inside DestBB block and it is a PHINode then check 404 // incoming value. If incoming value is not from BB then this is 405 // a complex condition (e.g. preheaders) we want to avoid here. 406 if (UI->getParent() == DestBB) { 407 if (const PHINode *UPN = dyn_cast<PHINode>(UI)) 408 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { 409 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); 410 if (Insn && Insn->getParent() == BB && 411 Insn->getParent() != UPN->getIncomingBlock(I)) 412 return false; 413 } 414 } 415 } 416 } 417 418 // If BB and DestBB contain any common predecessors, then the phi nodes in BB 419 // and DestBB may have conflicting incoming values for the block. If so, we 420 // can't merge the block. 421 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); 422 if (!DestBBPN) return true; // no conflict. 423 424 // Collect the preds of BB. 425 SmallPtrSet<const BasicBlock*, 16> BBPreds; 426 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 427 // It is faster to get preds from a PHI than with pred_iterator. 428 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 429 BBPreds.insert(BBPN->getIncomingBlock(i)); 430 } else { 431 BBPreds.insert(pred_begin(BB), pred_end(BB)); 432 } 433 434 // Walk the preds of DestBB. 435 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { 436 BasicBlock *Pred = DestBBPN->getIncomingBlock(i); 437 if (BBPreds.count(Pred)) { // Common predecessor? 438 BBI = DestBB->begin(); 439 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 440 const Value *V1 = PN->getIncomingValueForBlock(Pred); 441 const Value *V2 = PN->getIncomingValueForBlock(BB); 442 443 // If V2 is a phi node in BB, look up what the mapped value will be. 444 if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) 445 if (V2PN->getParent() == BB) 446 V2 = V2PN->getIncomingValueForBlock(Pred); 447 448 // If there is a conflict, bail out. 449 if (V1 != V2) return false; 450 } 451 } 452 } 453 454 return true; 455} 456 457 458/// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and 459/// an unconditional branch in it. 460void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { 461 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 462 BasicBlock *DestBB = BI->getSuccessor(0); 463 464 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB); 465 466 // If the destination block has a single pred, then this is a trivial edge, 467 // just collapse it. 468 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) { 469 if (SinglePred != DestBB) { 470 // Remember if SinglePred was the entry block of the function. If so, we 471 // will need to move BB back to the entry position. 472 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 473 MergeBasicBlockIntoOnlyPred(DestBB, nullptr); 474 475 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 476 BB->moveBefore(&BB->getParent()->getEntryBlock()); 477 478 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"); 479 return; 480 } 481 } 482 483 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB 484 // to handle the new incoming edges it is about to have. 485 PHINode *PN; 486 for (BasicBlock::iterator BBI = DestBB->begin(); 487 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 488 // Remove the incoming value for BB, and remember it. 489 Value *InVal = PN->removeIncomingValue(BB, false); 490 491 // Two options: either the InVal is a phi node defined in BB or it is some 492 // value that dominates BB. 493 PHINode *InValPhi = dyn_cast<PHINode>(InVal); 494 if (InValPhi && InValPhi->getParent() == BB) { 495 // Add all of the input values of the input PHI as inputs of this phi. 496 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) 497 PN->addIncoming(InValPhi->getIncomingValue(i), 498 InValPhi->getIncomingBlock(i)); 499 } else { 500 // Otherwise, add one instance of the dominating value for each edge that 501 // we will be adding. 502 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 503 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 504 PN->addIncoming(InVal, BBPN->getIncomingBlock(i)); 505 } else { 506 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 507 PN->addIncoming(InVal, *PI); 508 } 509 } 510 } 511 512 // The PHIs are now updated, change everything that refers to BB to use 513 // DestBB and remove BB. 514 BB->replaceAllUsesWith(DestBB); 515 BB->eraseFromParent(); 516 ++NumBlocksElim; 517 518 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n"); 519} 520 521// Computes a map of base pointer relocation instructions to corresponding 522// derived pointer relocation instructions given a vector of all relocate calls 523static void computeBaseDerivedRelocateMap( 524 const SmallVectorImpl<User *> &AllRelocateCalls, 525 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> & 526 RelocateInstMap) { 527 // Collect information in two maps: one primarily for locating the base object 528 // while filling the second map; the second map is the final structure holding 529 // a mapping between Base and corresponding Derived relocate calls 530 DenseMap<std::pair<unsigned, unsigned>, IntrinsicInst *> RelocateIdxMap; 531 for (auto &U : AllRelocateCalls) { 532 GCRelocateOperands ThisRelocate(U); 533 IntrinsicInst *I = cast<IntrinsicInst>(U); 534 auto K = std::make_pair(ThisRelocate.basePtrIndex(), 535 ThisRelocate.derivedPtrIndex()); 536 RelocateIdxMap.insert(std::make_pair(K, I)); 537 } 538 for (auto &Item : RelocateIdxMap) { 539 std::pair<unsigned, unsigned> Key = Item.first; 540 if (Key.first == Key.second) 541 // Base relocation: nothing to insert 542 continue; 543 544 IntrinsicInst *I = Item.second; 545 auto BaseKey = std::make_pair(Key.first, Key.first); 546 547 // We're iterating over RelocateIdxMap so we cannot modify it. 548 auto MaybeBase = RelocateIdxMap.find(BaseKey); 549 if (MaybeBase == RelocateIdxMap.end()) 550 // TODO: We might want to insert a new base object relocate and gep off 551 // that, if there are enough derived object relocates. 552 continue; 553 554 RelocateInstMap[MaybeBase->second].push_back(I); 555 } 556} 557 558// Accepts a GEP and extracts the operands into a vector provided they're all 559// small integer constants 560static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP, 561 SmallVectorImpl<Value *> &OffsetV) { 562 for (unsigned i = 1; i < GEP->getNumOperands(); i++) { 563 // Only accept small constant integer operands 564 auto Op = dyn_cast<ConstantInt>(GEP->getOperand(i)); 565 if (!Op || Op->getZExtValue() > 20) 566 return false; 567 } 568 569 for (unsigned i = 1; i < GEP->getNumOperands(); i++) 570 OffsetV.push_back(GEP->getOperand(i)); 571 return true; 572} 573 574// Takes a RelocatedBase (base pointer relocation instruction) and Targets to 575// replace, computes a replacement, and affects it. 576static bool 577simplifyRelocatesOffABase(IntrinsicInst *RelocatedBase, 578 const SmallVectorImpl<IntrinsicInst *> &Targets) { 579 bool MadeChange = false; 580 for (auto &ToReplace : Targets) { 581 GCRelocateOperands MasterRelocate(RelocatedBase); 582 GCRelocateOperands ThisRelocate(ToReplace); 583 584 assert(ThisRelocate.basePtrIndex() == MasterRelocate.basePtrIndex() && 585 "Not relocating a derived object of the original base object"); 586 if (ThisRelocate.basePtrIndex() == ThisRelocate.derivedPtrIndex()) { 587 // A duplicate relocate call. TODO: coalesce duplicates. 588 continue; 589 } 590 591 Value *Base = ThisRelocate.basePtr(); 592 auto Derived = dyn_cast<GetElementPtrInst>(ThisRelocate.derivedPtr()); 593 if (!Derived || Derived->getPointerOperand() != Base) 594 continue; 595 596 SmallVector<Value *, 2> OffsetV; 597 if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV)) 598 continue; 599 600 // Create a Builder and replace the target callsite with a gep 601 IRBuilder<> Builder(ToReplace); 602 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc()); 603 Value *Replacement = Builder.CreateGEP( 604 Derived->getSourceElementType(), RelocatedBase, makeArrayRef(OffsetV)); 605 Instruction *ReplacementInst = cast<Instruction>(Replacement); 606 ReplacementInst->removeFromParent(); 607 ReplacementInst->insertAfter(RelocatedBase); 608 Replacement->takeName(ToReplace); 609 ToReplace->replaceAllUsesWith(Replacement); 610 ToReplace->eraseFromParent(); 611 612 MadeChange = true; 613 } 614 return MadeChange; 615} 616 617// Turns this: 618// 619// %base = ... 620// %ptr = gep %base + 15 621// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr) 622// %base' = relocate(%tok, i32 4, i32 4) 623// %ptr' = relocate(%tok, i32 4, i32 5) 624// %val = load %ptr' 625// 626// into this: 627// 628// %base = ... 629// %ptr = gep %base + 15 630// %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr) 631// %base' = gc.relocate(%tok, i32 4, i32 4) 632// %ptr' = gep %base' + 15 633// %val = load %ptr' 634bool CodeGenPrepare::simplifyOffsetableRelocate(Instruction &I) { 635 bool MadeChange = false; 636 SmallVector<User *, 2> AllRelocateCalls; 637 638 for (auto *U : I.users()) 639 if (isGCRelocate(dyn_cast<Instruction>(U))) 640 // Collect all the relocate calls associated with a statepoint 641 AllRelocateCalls.push_back(U); 642 643 // We need atleast one base pointer relocation + one derived pointer 644 // relocation to mangle 645 if (AllRelocateCalls.size() < 2) 646 return false; 647 648 // RelocateInstMap is a mapping from the base relocate instruction to the 649 // corresponding derived relocate instructions 650 DenseMap<IntrinsicInst *, SmallVector<IntrinsicInst *, 2>> RelocateInstMap; 651 computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap); 652 if (RelocateInstMap.empty()) 653 return false; 654 655 for (auto &Item : RelocateInstMap) 656 // Item.first is the RelocatedBase to offset against 657 // Item.second is the vector of Targets to replace 658 MadeChange = simplifyRelocatesOffABase(Item.first, Item.second); 659 return MadeChange; 660} 661 662/// SinkCast - Sink the specified cast instruction into its user blocks 663static bool SinkCast(CastInst *CI) { 664 BasicBlock *DefBB = CI->getParent(); 665 666 /// InsertedCasts - Only insert a cast in each block once. 667 DenseMap<BasicBlock*, CastInst*> InsertedCasts; 668 669 bool MadeChange = false; 670 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end(); 671 UI != E; ) { 672 Use &TheUse = UI.getUse(); 673 Instruction *User = cast<Instruction>(*UI); 674 675 // Figure out which BB this cast is used in. For PHI's this is the 676 // appropriate predecessor block. 677 BasicBlock *UserBB = User->getParent(); 678 if (PHINode *PN = dyn_cast<PHINode>(User)) { 679 UserBB = PN->getIncomingBlock(TheUse); 680 } 681 682 // Preincrement use iterator so we don't invalidate it. 683 ++UI; 684 685 // If this user is in the same block as the cast, don't change the cast. 686 if (UserBB == DefBB) continue; 687 688 // If we have already inserted a cast into this block, use it. 689 CastInst *&InsertedCast = InsertedCasts[UserBB]; 690 691 if (!InsertedCast) { 692 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); 693 InsertedCast = 694 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 695 InsertPt); 696 } 697 698 // Replace a use of the cast with a use of the new cast. 699 TheUse = InsertedCast; 700 MadeChange = true; 701 ++NumCastUses; 702 } 703 704 // If we removed all uses, nuke the cast. 705 if (CI->use_empty()) { 706 CI->eraseFromParent(); 707 MadeChange = true; 708 } 709 710 return MadeChange; 711} 712 713/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop 714/// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC), 715/// sink it into user blocks to reduce the number of virtual 716/// registers that must be created and coalesced. 717/// 718/// Return true if any changes are made. 719/// 720static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ 721 // If this is a noop copy, 722 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 723 EVT DstVT = TLI.getValueType(CI->getType()); 724 725 // This is an fp<->int conversion? 726 if (SrcVT.isInteger() != DstVT.isInteger()) 727 return false; 728 729 // If this is an extension, it will be a zero or sign extension, which 730 // isn't a noop. 731 if (SrcVT.bitsLT(DstVT)) return false; 732 733 // If these values will be promoted, find out what they will be promoted 734 // to. This helps us consider truncates on PPC as noop copies when they 735 // are. 736 if (TLI.getTypeAction(CI->getContext(), SrcVT) == 737 TargetLowering::TypePromoteInteger) 738 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT); 739 if (TLI.getTypeAction(CI->getContext(), DstVT) == 740 TargetLowering::TypePromoteInteger) 741 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT); 742 743 // If, after promotion, these are the same types, this is a noop copy. 744 if (SrcVT != DstVT) 745 return false; 746 747 return SinkCast(CI); 748} 749 750/// CombineUAddWithOverflow - try to combine CI into a call to the 751/// llvm.uadd.with.overflow intrinsic if possible. 752/// 753/// Return true if any changes were made. 754static bool CombineUAddWithOverflow(CmpInst *CI) { 755 Value *A, *B; 756 Instruction *AddI; 757 if (!match(CI, 758 m_UAddWithOverflow(m_Value(A), m_Value(B), m_Instruction(AddI)))) 759 return false; 760 761 Type *Ty = AddI->getType(); 762 if (!isa<IntegerType>(Ty)) 763 return false; 764 765 // We don't want to move around uses of condition values this late, so we we 766 // check if it is legal to create the call to the intrinsic in the basic 767 // block containing the icmp: 768 769 if (AddI->getParent() != CI->getParent() && !AddI->hasOneUse()) 770 return false; 771 772#ifndef NDEBUG 773 // Someday m_UAddWithOverflow may get smarter, but this is a safe assumption 774 // for now: 775 if (AddI->hasOneUse()) 776 assert(*AddI->user_begin() == CI && "expected!"); 777#endif 778 779 Module *M = CI->getParent()->getParent()->getParent(); 780 Value *F = Intrinsic::getDeclaration(M, Intrinsic::uadd_with_overflow, Ty); 781 782 auto *InsertPt = AddI->hasOneUse() ? CI : AddI; 783 784 auto *UAddWithOverflow = 785 CallInst::Create(F, {A, B}, "uadd.overflow", InsertPt); 786 auto *UAdd = ExtractValueInst::Create(UAddWithOverflow, 0, "uadd", InsertPt); 787 auto *Overflow = 788 ExtractValueInst::Create(UAddWithOverflow, 1, "overflow", InsertPt); 789 790 CI->replaceAllUsesWith(Overflow); 791 AddI->replaceAllUsesWith(UAdd); 792 CI->eraseFromParent(); 793 AddI->eraseFromParent(); 794 return true; 795} 796 797/// SinkCmpExpression - Sink the given CmpInst into user blocks to reduce 798/// the number of virtual registers that must be created and coalesced. This is 799/// a clear win except on targets with multiple condition code registers 800/// (PowerPC), where it might lose; some adjustment may be wanted there. 801/// 802/// Return true if any changes are made. 803static bool SinkCmpExpression(CmpInst *CI) { 804 BasicBlock *DefBB = CI->getParent(); 805 806 /// InsertedCmp - Only insert a cmp in each block once. 807 DenseMap<BasicBlock*, CmpInst*> InsertedCmps; 808 809 bool MadeChange = false; 810 for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end(); 811 UI != E; ) { 812 Use &TheUse = UI.getUse(); 813 Instruction *User = cast<Instruction>(*UI); 814 815 // Preincrement use iterator so we don't invalidate it. 816 ++UI; 817 818 // Don't bother for PHI nodes. 819 if (isa<PHINode>(User)) 820 continue; 821 822 // Figure out which BB this cmp is used in. 823 BasicBlock *UserBB = User->getParent(); 824 825 // If this user is in the same block as the cmp, don't change the cmp. 826 if (UserBB == DefBB) continue; 827 828 // If we have already inserted a cmp into this block, use it. 829 CmpInst *&InsertedCmp = InsertedCmps[UserBB]; 830 831 if (!InsertedCmp) { 832 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); 833 InsertedCmp = 834 CmpInst::Create(CI->getOpcode(), 835 CI->getPredicate(), CI->getOperand(0), 836 CI->getOperand(1), "", InsertPt); 837 } 838 839 // Replace a use of the cmp with a use of the new cmp. 840 TheUse = InsertedCmp; 841 MadeChange = true; 842 ++NumCmpUses; 843 } 844 845 // If we removed all uses, nuke the cmp. 846 if (CI->use_empty()) { 847 CI->eraseFromParent(); 848 MadeChange = true; 849 } 850 851 return MadeChange; 852} 853 854static bool OptimizeCmpExpression(CmpInst *CI) { 855 if (SinkCmpExpression(CI)) 856 return true; 857 858 if (CombineUAddWithOverflow(CI)) 859 return true; 860 861 return false; 862} 863 864/// isExtractBitsCandidateUse - Check if the candidates could 865/// be combined with shift instruction, which includes: 866/// 1. Truncate instruction 867/// 2. And instruction and the imm is a mask of the low bits: 868/// imm & (imm+1) == 0 869static bool isExtractBitsCandidateUse(Instruction *User) { 870 if (!isa<TruncInst>(User)) { 871 if (User->getOpcode() != Instruction::And || 872 !isa<ConstantInt>(User->getOperand(1))) 873 return false; 874 875 const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue(); 876 877 if ((Cimm & (Cimm + 1)).getBoolValue()) 878 return false; 879 } 880 return true; 881} 882 883/// SinkShiftAndTruncate - sink both shift and truncate instruction 884/// to the use of truncate's BB. 885static bool 886SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI, 887 DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts, 888 const TargetLowering &TLI) { 889 BasicBlock *UserBB = User->getParent(); 890 DenseMap<BasicBlock *, CastInst *> InsertedTruncs; 891 TruncInst *TruncI = dyn_cast<TruncInst>(User); 892 bool MadeChange = false; 893 894 for (Value::user_iterator TruncUI = TruncI->user_begin(), 895 TruncE = TruncI->user_end(); 896 TruncUI != TruncE;) { 897 898 Use &TruncTheUse = TruncUI.getUse(); 899 Instruction *TruncUser = cast<Instruction>(*TruncUI); 900 // Preincrement use iterator so we don't invalidate it. 901 902 ++TruncUI; 903 904 int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode()); 905 if (!ISDOpcode) 906 continue; 907 908 // If the use is actually a legal node, there will not be an 909 // implicit truncate. 910 // FIXME: always querying the result type is just an 911 // approximation; some nodes' legality is determined by the 912 // operand or other means. There's no good way to find out though. 913 if (TLI.isOperationLegalOrCustom( 914 ISDOpcode, TLI.getValueType(TruncUser->getType(), true))) 915 continue; 916 917 // Don't bother for PHI nodes. 918 if (isa<PHINode>(TruncUser)) 919 continue; 920 921 BasicBlock *TruncUserBB = TruncUser->getParent(); 922 923 if (UserBB == TruncUserBB) 924 continue; 925 926 BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB]; 927 CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB]; 928 929 if (!InsertedShift && !InsertedTrunc) { 930 BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt(); 931 // Sink the shift 932 if (ShiftI->getOpcode() == Instruction::AShr) 933 InsertedShift = 934 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt); 935 else 936 InsertedShift = 937 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt); 938 939 // Sink the trunc 940 BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt(); 941 TruncInsertPt++; 942 943 InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift, 944 TruncI->getType(), "", TruncInsertPt); 945 946 MadeChange = true; 947 948 TruncTheUse = InsertedTrunc; 949 } 950 } 951 return MadeChange; 952} 953 954/// OptimizeExtractBits - sink the shift *right* instruction into user blocks if 955/// the uses could potentially be combined with this shift instruction and 956/// generate BitExtract instruction. It will only be applied if the architecture 957/// supports BitExtract instruction. Here is an example: 958/// BB1: 959/// %x.extract.shift = lshr i64 %arg1, 32 960/// BB2: 961/// %x.extract.trunc = trunc i64 %x.extract.shift to i16 962/// ==> 963/// 964/// BB2: 965/// %x.extract.shift.1 = lshr i64 %arg1, 32 966/// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16 967/// 968/// CodeGen will recoginze the pattern in BB2 and generate BitExtract 969/// instruction. 970/// Return true if any changes are made. 971static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI, 972 const TargetLowering &TLI) { 973 BasicBlock *DefBB = ShiftI->getParent(); 974 975 /// Only insert instructions in each block once. 976 DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts; 977 978 bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(ShiftI->getType())); 979 980 bool MadeChange = false; 981 for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end(); 982 UI != E;) { 983 Use &TheUse = UI.getUse(); 984 Instruction *User = cast<Instruction>(*UI); 985 // Preincrement use iterator so we don't invalidate it. 986 ++UI; 987 988 // Don't bother for PHI nodes. 989 if (isa<PHINode>(User)) 990 continue; 991 992 if (!isExtractBitsCandidateUse(User)) 993 continue; 994 995 BasicBlock *UserBB = User->getParent(); 996 997 if (UserBB == DefBB) { 998 // If the shift and truncate instruction are in the same BB. The use of 999 // the truncate(TruncUse) may still introduce another truncate if not 1000 // legal. In this case, we would like to sink both shift and truncate 1001 // instruction to the BB of TruncUse. 1002 // for example: 1003 // BB1: 1004 // i64 shift.result = lshr i64 opnd, imm 1005 // trunc.result = trunc shift.result to i16 1006 // 1007 // BB2: 1008 // ----> We will have an implicit truncate here if the architecture does 1009 // not have i16 compare. 1010 // cmp i16 trunc.result, opnd2 1011 // 1012 if (isa<TruncInst>(User) && shiftIsLegal 1013 // If the type of the truncate is legal, no trucate will be 1014 // introduced in other basic blocks. 1015 && (!TLI.isTypeLegal(TLI.getValueType(User->getType())))) 1016 MadeChange = 1017 SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI); 1018 1019 continue; 1020 } 1021 // If we have already inserted a shift into this block, use it. 1022 BinaryOperator *&InsertedShift = InsertedShifts[UserBB]; 1023 1024 if (!InsertedShift) { 1025 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); 1026 1027 if (ShiftI->getOpcode() == Instruction::AShr) 1028 InsertedShift = 1029 BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, "", InsertPt); 1030 else 1031 InsertedShift = 1032 BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, "", InsertPt); 1033 1034 MadeChange = true; 1035 } 1036 1037 // Replace a use of the shift with a use of the new shift. 1038 TheUse = InsertedShift; 1039 } 1040 1041 // If we removed all uses, nuke the shift. 1042 if (ShiftI->use_empty()) 1043 ShiftI->eraseFromParent(); 1044 1045 return MadeChange; 1046} 1047 1048// ScalarizeMaskedLoad() translates masked load intrinsic, like 1049// <16 x i32 > @llvm.masked.load( <16 x i32>* %addr, i32 align, 1050// <16 x i1> %mask, <16 x i32> %passthru) 1051// to a chain of basic blocks, whith loading element one-by-one if 1052// the appropriate mask bit is set 1053// 1054// %1 = bitcast i8* %addr to i32* 1055// %2 = extractelement <16 x i1> %mask, i32 0 1056// %3 = icmp eq i1 %2, true 1057// br i1 %3, label %cond.load, label %else 1058// 1059//cond.load: ; preds = %0 1060// %4 = getelementptr i32* %1, i32 0 1061// %5 = load i32* %4 1062// %6 = insertelement <16 x i32> undef, i32 %5, i32 0 1063// br label %else 1064// 1065//else: ; preds = %0, %cond.load 1066// %res.phi.else = phi <16 x i32> [ %6, %cond.load ], [ undef, %0 ] 1067// %7 = extractelement <16 x i1> %mask, i32 1 1068// %8 = icmp eq i1 %7, true 1069// br i1 %8, label %cond.load1, label %else2 1070// 1071//cond.load1: ; preds = %else 1072// %9 = getelementptr i32* %1, i32 1 1073// %10 = load i32* %9 1074// %11 = insertelement <16 x i32> %res.phi.else, i32 %10, i32 1 1075// br label %else2 1076// 1077//else2: ; preds = %else, %cond.load1 1078// %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ] 1079// %12 = extractelement <16 x i1> %mask, i32 2 1080// %13 = icmp eq i1 %12, true 1081// br i1 %13, label %cond.load4, label %else5 1082// 1083static void ScalarizeMaskedLoad(CallInst *CI) { 1084 Value *Ptr = CI->getArgOperand(0); 1085 Value *Src0 = CI->getArgOperand(3); 1086 Value *Mask = CI->getArgOperand(2); 1087 VectorType *VecType = dyn_cast<VectorType>(CI->getType()); 1088 Type *EltTy = VecType->getElementType(); 1089 1090 assert(VecType && "Unexpected return type of masked load intrinsic"); 1091 1092 IRBuilder<> Builder(CI->getContext()); 1093 Instruction *InsertPt = CI; 1094 BasicBlock *IfBlock = CI->getParent(); 1095 BasicBlock *CondBlock = nullptr; 1096 BasicBlock *PrevIfBlock = CI->getParent(); 1097 Builder.SetInsertPoint(InsertPt); 1098 1099 Builder.SetCurrentDebugLocation(CI->getDebugLoc()); 1100 1101 // Bitcast %addr fron i8* to EltTy* 1102 Type *NewPtrType = 1103 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace()); 1104 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType); 1105 Value *UndefVal = UndefValue::get(VecType); 1106 1107 // The result vector 1108 Value *VResult = UndefVal; 1109 1110 PHINode *Phi = nullptr; 1111 Value *PrevPhi = UndefVal; 1112 1113 unsigned VectorWidth = VecType->getNumElements(); 1114 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 1115 1116 // Fill the "else" block, created in the previous iteration 1117 // 1118 // %res.phi.else3 = phi <16 x i32> [ %11, %cond.load1 ], [ %res.phi.else, %else ] 1119 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx 1120 // %to_load = icmp eq i1 %mask_1, true 1121 // br i1 %to_load, label %cond.load, label %else 1122 // 1123 if (Idx > 0) { 1124 Phi = Builder.CreatePHI(VecType, 2, "res.phi.else"); 1125 Phi->addIncoming(VResult, CondBlock); 1126 Phi->addIncoming(PrevPhi, PrevIfBlock); 1127 PrevPhi = Phi; 1128 VResult = Phi; 1129 } 1130 1131 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx)); 1132 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate, 1133 ConstantInt::get(Predicate->getType(), 1)); 1134 1135 // Create "cond" block 1136 // 1137 // %EltAddr = getelementptr i32* %1, i32 0 1138 // %Elt = load i32* %EltAddr 1139 // VResult = insertelement <16 x i32> VResult, i32 %Elt, i32 Idx 1140 // 1141 CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.load"); 1142 Builder.SetInsertPoint(InsertPt); 1143 1144 Value *Gep = 1145 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx)); 1146 LoadInst* Load = Builder.CreateLoad(Gep, false); 1147 VResult = Builder.CreateInsertElement(VResult, Load, Builder.getInt32(Idx)); 1148 1149 // Create "else" block, fill it in the next iteration 1150 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else"); 1151 Builder.SetInsertPoint(InsertPt); 1152 Instruction *OldBr = IfBlock->getTerminator(); 1153 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr); 1154 OldBr->eraseFromParent(); 1155 PrevIfBlock = IfBlock; 1156 IfBlock = NewIfBlock; 1157 } 1158 1159 Phi = Builder.CreatePHI(VecType, 2, "res.phi.select"); 1160 Phi->addIncoming(VResult, CondBlock); 1161 Phi->addIncoming(PrevPhi, PrevIfBlock); 1162 Value *NewI = Builder.CreateSelect(Mask, Phi, Src0); 1163 CI->replaceAllUsesWith(NewI); 1164 CI->eraseFromParent(); 1165} 1166 1167// ScalarizeMaskedStore() translates masked store intrinsic, like 1168// void @llvm.masked.store(<16 x i32> %src, <16 x i32>* %addr, i32 align, 1169// <16 x i1> %mask) 1170// to a chain of basic blocks, that stores element one-by-one if 1171// the appropriate mask bit is set 1172// 1173// %1 = bitcast i8* %addr to i32* 1174// %2 = extractelement <16 x i1> %mask, i32 0 1175// %3 = icmp eq i1 %2, true 1176// br i1 %3, label %cond.store, label %else 1177// 1178// cond.store: ; preds = %0 1179// %4 = extractelement <16 x i32> %val, i32 0 1180// %5 = getelementptr i32* %1, i32 0 1181// store i32 %4, i32* %5 1182// br label %else 1183// 1184// else: ; preds = %0, %cond.store 1185// %6 = extractelement <16 x i1> %mask, i32 1 1186// %7 = icmp eq i1 %6, true 1187// br i1 %7, label %cond.store1, label %else2 1188// 1189// cond.store1: ; preds = %else 1190// %8 = extractelement <16 x i32> %val, i32 1 1191// %9 = getelementptr i32* %1, i32 1 1192// store i32 %8, i32* %9 1193// br label %else2 1194// . . . 1195static void ScalarizeMaskedStore(CallInst *CI) { 1196 Value *Ptr = CI->getArgOperand(1); 1197 Value *Src = CI->getArgOperand(0); 1198 Value *Mask = CI->getArgOperand(3); 1199 1200 VectorType *VecType = dyn_cast<VectorType>(Src->getType()); 1201 Type *EltTy = VecType->getElementType(); 1202 1203 assert(VecType && "Unexpected data type in masked store intrinsic"); 1204 1205 IRBuilder<> Builder(CI->getContext()); 1206 Instruction *InsertPt = CI; 1207 BasicBlock *IfBlock = CI->getParent(); 1208 Builder.SetInsertPoint(InsertPt); 1209 Builder.SetCurrentDebugLocation(CI->getDebugLoc()); 1210 1211 // Bitcast %addr fron i8* to EltTy* 1212 Type *NewPtrType = 1213 EltTy->getPointerTo(cast<PointerType>(Ptr->getType())->getAddressSpace()); 1214 Value *FirstEltPtr = Builder.CreateBitCast(Ptr, NewPtrType); 1215 1216 unsigned VectorWidth = VecType->getNumElements(); 1217 for (unsigned Idx = 0; Idx < VectorWidth; ++Idx) { 1218 1219 // Fill the "else" block, created in the previous iteration 1220 // 1221 // %mask_1 = extractelement <16 x i1> %mask, i32 Idx 1222 // %to_store = icmp eq i1 %mask_1, true 1223 // br i1 %to_load, label %cond.store, label %else 1224 // 1225 Value *Predicate = Builder.CreateExtractElement(Mask, Builder.getInt32(Idx)); 1226 Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_EQ, Predicate, 1227 ConstantInt::get(Predicate->getType(), 1)); 1228 1229 // Create "cond" block 1230 // 1231 // %OneElt = extractelement <16 x i32> %Src, i32 Idx 1232 // %EltAddr = getelementptr i32* %1, i32 0 1233 // %store i32 %OneElt, i32* %EltAddr 1234 // 1235 BasicBlock *CondBlock = IfBlock->splitBasicBlock(InsertPt, "cond.store"); 1236 Builder.SetInsertPoint(InsertPt); 1237 1238 Value *OneElt = Builder.CreateExtractElement(Src, Builder.getInt32(Idx)); 1239 Value *Gep = 1240 Builder.CreateInBoundsGEP(EltTy, FirstEltPtr, Builder.getInt32(Idx)); 1241 Builder.CreateStore(OneElt, Gep); 1242 1243 // Create "else" block, fill it in the next iteration 1244 BasicBlock *NewIfBlock = CondBlock->splitBasicBlock(InsertPt, "else"); 1245 Builder.SetInsertPoint(InsertPt); 1246 Instruction *OldBr = IfBlock->getTerminator(); 1247 BranchInst::Create(CondBlock, NewIfBlock, Cmp, OldBr); 1248 OldBr->eraseFromParent(); 1249 IfBlock = NewIfBlock; 1250 } 1251 CI->eraseFromParent(); 1252} 1253 1254bool CodeGenPrepare::OptimizeCallInst(CallInst *CI, bool& ModifiedDT) { 1255 BasicBlock *BB = CI->getParent(); 1256 1257 // Lower inline assembly if we can. 1258 // If we found an inline asm expession, and if the target knows how to 1259 // lower it to normal LLVM code, do so now. 1260 if (TLI && isa<InlineAsm>(CI->getCalledValue())) { 1261 if (TLI->ExpandInlineAsm(CI)) { 1262 // Avoid invalidating the iterator. 1263 CurInstIterator = BB->begin(); 1264 // Avoid processing instructions out of order, which could cause 1265 // reuse before a value is defined. 1266 SunkAddrs.clear(); 1267 return true; 1268 } 1269 // Sink address computing for memory operands into the block. 1270 if (OptimizeInlineAsmInst(CI)) 1271 return true; 1272 } 1273 1274 const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr; 1275 1276 // Align the pointer arguments to this call if the target thinks it's a good 1277 // idea 1278 unsigned MinSize, PrefAlign; 1279 if (TLI && TD && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) { 1280 for (auto &Arg : CI->arg_operands()) { 1281 // We want to align both objects whose address is used directly and 1282 // objects whose address is used in casts and GEPs, though it only makes 1283 // sense for GEPs if the offset is a multiple of the desired alignment and 1284 // if size - offset meets the size threshold. 1285 if (!Arg->getType()->isPointerTy()) 1286 continue; 1287 APInt Offset(TD->getPointerSizeInBits( 1288 cast<PointerType>(Arg->getType())->getAddressSpace()), 0); 1289 Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*TD, Offset); 1290 uint64_t Offset2 = Offset.getLimitedValue(); 1291 if ((Offset2 & (PrefAlign-1)) != 0) 1292 continue; 1293 AllocaInst *AI; 1294 if ((AI = dyn_cast<AllocaInst>(Val)) && 1295 AI->getAlignment() < PrefAlign && 1296 TD->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2) 1297 AI->setAlignment(PrefAlign); 1298 // Global variables can only be aligned if they are defined in this 1299 // object (i.e. they are uniquely initialized in this object), and 1300 // over-aligning global variables that have an explicit section is 1301 // forbidden. 1302 GlobalVariable *GV; 1303 if ((GV = dyn_cast<GlobalVariable>(Val)) && 1304 GV->hasUniqueInitializer() && 1305 !GV->hasSection() && 1306 GV->getAlignment() < PrefAlign && 1307 TD->getTypeAllocSize( 1308 GV->getType()->getElementType()) >= MinSize + Offset2) 1309 GV->setAlignment(PrefAlign); 1310 } 1311 // If this is a memcpy (or similar) then we may be able to improve the 1312 // alignment 1313 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) { 1314 unsigned Align = getKnownAlignment(MI->getDest(), *TD); 1315 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) 1316 Align = std::min(Align, getKnownAlignment(MTI->getSource(), *TD)); 1317 if (Align > MI->getAlignment()) 1318 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align)); 1319 } 1320 } 1321 1322 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI); 1323 if (II) { 1324 switch (II->getIntrinsicID()) { 1325 default: break; 1326 case Intrinsic::objectsize: { 1327 // Lower all uses of llvm.objectsize.* 1328 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1); 1329 Type *ReturnTy = CI->getType(); 1330 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL); 1331 1332 // Substituting this can cause recursive simplifications, which can 1333 // invalidate our iterator. Use a WeakVH to hold onto it in case this 1334 // happens. 1335 WeakVH IterHandle(CurInstIterator); 1336 1337 replaceAndRecursivelySimplify(CI, RetVal, 1338 TLInfo, nullptr); 1339 1340 // If the iterator instruction was recursively deleted, start over at the 1341 // start of the block. 1342 if (IterHandle != CurInstIterator) { 1343 CurInstIterator = BB->begin(); 1344 SunkAddrs.clear(); 1345 } 1346 return true; 1347 } 1348 case Intrinsic::masked_load: { 1349 // Scalarize unsupported vector masked load 1350 if (!TTI->isLegalMaskedLoad(CI->getType(), 1)) { 1351 ScalarizeMaskedLoad(CI); 1352 ModifiedDT = true; 1353 return true; 1354 } 1355 return false; 1356 } 1357 case Intrinsic::masked_store: { 1358 if (!TTI->isLegalMaskedStore(CI->getArgOperand(0)->getType(), 1)) { 1359 ScalarizeMaskedStore(CI); 1360 ModifiedDT = true; 1361 return true; 1362 } 1363 return false; 1364 } 1365 } 1366 1367 if (TLI) { 1368 SmallVector<Value*, 2> PtrOps; 1369 Type *AccessTy; 1370 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy)) 1371 while (!PtrOps.empty()) 1372 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy)) 1373 return true; 1374 } 1375 } 1376 1377 // From here on out we're working with named functions. 1378 if (!CI->getCalledFunction()) return false; 1379 1380 // Lower all default uses of _chk calls. This is very similar 1381 // to what InstCombineCalls does, but here we are only lowering calls 1382 // to fortified library functions (e.g. __memcpy_chk) that have the default 1383 // "don't know" as the objectsize. Anything else should be left alone. 1384 FortifiedLibCallSimplifier Simplifier(TLInfo, true); 1385 if (Value *V = Simplifier.optimizeCall(CI)) { 1386 CI->replaceAllUsesWith(V); 1387 CI->eraseFromParent(); 1388 return true; 1389 } 1390 return false; 1391} 1392 1393/// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return 1394/// instructions to the predecessor to enable tail call optimizations. The 1395/// case it is currently looking for is: 1396/// @code 1397/// bb0: 1398/// %tmp0 = tail call i32 @f0() 1399/// br label %return 1400/// bb1: 1401/// %tmp1 = tail call i32 @f1() 1402/// br label %return 1403/// bb2: 1404/// %tmp2 = tail call i32 @f2() 1405/// br label %return 1406/// return: 1407/// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ] 1408/// ret i32 %retval 1409/// @endcode 1410/// 1411/// => 1412/// 1413/// @code 1414/// bb0: 1415/// %tmp0 = tail call i32 @f0() 1416/// ret i32 %tmp0 1417/// bb1: 1418/// %tmp1 = tail call i32 @f1() 1419/// ret i32 %tmp1 1420/// bb2: 1421/// %tmp2 = tail call i32 @f2() 1422/// ret i32 %tmp2 1423/// @endcode 1424bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) { 1425 if (!TLI) 1426 return false; 1427 1428 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()); 1429 if (!RI) 1430 return false; 1431 1432 PHINode *PN = nullptr; 1433 BitCastInst *BCI = nullptr; 1434 Value *V = RI->getReturnValue(); 1435 if (V) { 1436 BCI = dyn_cast<BitCastInst>(V); 1437 if (BCI) 1438 V = BCI->getOperand(0); 1439 1440 PN = dyn_cast<PHINode>(V); 1441 if (!PN) 1442 return false; 1443 } 1444 1445 if (PN && PN->getParent() != BB) 1446 return false; 1447 1448 // It's not safe to eliminate the sign / zero extension of the return value. 1449 // See llvm::isInTailCallPosition(). 1450 const Function *F = BB->getParent(); 1451 AttributeSet CallerAttrs = F->getAttributes(); 1452 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) || 1453 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt)) 1454 return false; 1455 1456 // Make sure there are no instructions between the PHI and return, or that the 1457 // return is the first instruction in the block. 1458 if (PN) { 1459 BasicBlock::iterator BI = BB->begin(); 1460 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI)); 1461 if (&*BI == BCI) 1462 // Also skip over the bitcast. 1463 ++BI; 1464 if (&*BI != RI) 1465 return false; 1466 } else { 1467 BasicBlock::iterator BI = BB->begin(); 1468 while (isa<DbgInfoIntrinsic>(BI)) ++BI; 1469 if (&*BI != RI) 1470 return false; 1471 } 1472 1473 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail 1474 /// call. 1475 SmallVector<CallInst*, 4> TailCalls; 1476 if (PN) { 1477 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) { 1478 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I)); 1479 // Make sure the phi value is indeed produced by the tail call. 1480 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) && 1481 TLI->mayBeEmittedAsTailCall(CI)) 1482 TailCalls.push_back(CI); 1483 } 1484 } else { 1485 SmallPtrSet<BasicBlock*, 4> VisitedBBs; 1486 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) { 1487 if (!VisitedBBs.insert(*PI).second) 1488 continue; 1489 1490 BasicBlock::InstListType &InstList = (*PI)->getInstList(); 1491 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin(); 1492 BasicBlock::InstListType::reverse_iterator RE = InstList.rend(); 1493 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI)); 1494 if (RI == RE) 1495 continue; 1496 1497 CallInst *CI = dyn_cast<CallInst>(&*RI); 1498 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI)) 1499 TailCalls.push_back(CI); 1500 } 1501 } 1502 1503 bool Changed = false; 1504 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) { 1505 CallInst *CI = TailCalls[i]; 1506 CallSite CS(CI); 1507 1508 // Conservatively require the attributes of the call to match those of the 1509 // return. Ignore noalias because it doesn't affect the call sequence. 1510 AttributeSet CalleeAttrs = CS.getAttributes(); 1511 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex). 1512 removeAttribute(Attribute::NoAlias) != 1513 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex). 1514 removeAttribute(Attribute::NoAlias)) 1515 continue; 1516 1517 // Make sure the call instruction is followed by an unconditional branch to 1518 // the return block. 1519 BasicBlock *CallBB = CI->getParent(); 1520 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator()); 1521 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB) 1522 continue; 1523 1524 // Duplicate the return into CallBB. 1525 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB); 1526 ModifiedDT = Changed = true; 1527 ++NumRetsDup; 1528 } 1529 1530 // If we eliminated all predecessors of the block, delete the block now. 1531 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB)) 1532 BB->eraseFromParent(); 1533 1534 return Changed; 1535} 1536 1537//===----------------------------------------------------------------------===// 1538// Memory Optimization 1539//===----------------------------------------------------------------------===// 1540 1541namespace { 1542 1543/// ExtAddrMode - This is an extended version of TargetLowering::AddrMode 1544/// which holds actual Value*'s for register values. 1545struct ExtAddrMode : public TargetLowering::AddrMode { 1546 Value *BaseReg; 1547 Value *ScaledReg; 1548 ExtAddrMode() : BaseReg(nullptr), ScaledReg(nullptr) {} 1549 void print(raw_ostream &OS) const; 1550 void dump() const; 1551 1552 bool operator==(const ExtAddrMode& O) const { 1553 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) && 1554 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) && 1555 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale); 1556 } 1557}; 1558 1559#ifndef NDEBUG 1560static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) { 1561 AM.print(OS); 1562 return OS; 1563} 1564#endif 1565 1566void ExtAddrMode::print(raw_ostream &OS) const { 1567 bool NeedPlus = false; 1568 OS << "["; 1569 if (BaseGV) { 1570 OS << (NeedPlus ? " + " : "") 1571 << "GV:"; 1572 BaseGV->printAsOperand(OS, /*PrintType=*/false); 1573 NeedPlus = true; 1574 } 1575 1576 if (BaseOffs) { 1577 OS << (NeedPlus ? " + " : "") 1578 << BaseOffs; 1579 NeedPlus = true; 1580 } 1581 1582 if (BaseReg) { 1583 OS << (NeedPlus ? " + " : "") 1584 << "Base:"; 1585 BaseReg->printAsOperand(OS, /*PrintType=*/false); 1586 NeedPlus = true; 1587 } 1588 if (Scale) { 1589 OS << (NeedPlus ? " + " : "") 1590 << Scale << "*"; 1591 ScaledReg->printAsOperand(OS, /*PrintType=*/false); 1592 } 1593 1594 OS << ']'; 1595} 1596 1597#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1598void ExtAddrMode::dump() const { 1599 print(dbgs()); 1600 dbgs() << '\n'; 1601} 1602#endif 1603 1604/// \brief This class provides transaction based operation on the IR. 1605/// Every change made through this class is recorded in the internal state and 1606/// can be undone (rollback) until commit is called. 1607class TypePromotionTransaction { 1608 1609 /// \brief This represents the common interface of the individual transaction. 1610 /// Each class implements the logic for doing one specific modification on 1611 /// the IR via the TypePromotionTransaction. 1612 class TypePromotionAction { 1613 protected: 1614 /// The Instruction modified. 1615 Instruction *Inst; 1616 1617 public: 1618 /// \brief Constructor of the action. 1619 /// The constructor performs the related action on the IR. 1620 TypePromotionAction(Instruction *Inst) : Inst(Inst) {} 1621 1622 virtual ~TypePromotionAction() {} 1623 1624 /// \brief Undo the modification done by this action. 1625 /// When this method is called, the IR must be in the same state as it was 1626 /// before this action was applied. 1627 /// \pre Undoing the action works if and only if the IR is in the exact same 1628 /// state as it was directly after this action was applied. 1629 virtual void undo() = 0; 1630 1631 /// \brief Advocate every change made by this action. 1632 /// When the results on the IR of the action are to be kept, it is important 1633 /// to call this function, otherwise hidden information may be kept forever. 1634 virtual void commit() { 1635 // Nothing to be done, this action is not doing anything. 1636 } 1637 }; 1638 1639 /// \brief Utility to remember the position of an instruction. 1640 class InsertionHandler { 1641 /// Position of an instruction. 1642 /// Either an instruction: 1643 /// - Is the first in a basic block: BB is used. 1644 /// - Has a previous instructon: PrevInst is used. 1645 union { 1646 Instruction *PrevInst; 1647 BasicBlock *BB; 1648 } Point; 1649 /// Remember whether or not the instruction had a previous instruction. 1650 bool HasPrevInstruction; 1651 1652 public: 1653 /// \brief Record the position of \p Inst. 1654 InsertionHandler(Instruction *Inst) { 1655 BasicBlock::iterator It = Inst; 1656 HasPrevInstruction = (It != (Inst->getParent()->begin())); 1657 if (HasPrevInstruction) 1658 Point.PrevInst = --It; 1659 else 1660 Point.BB = Inst->getParent(); 1661 } 1662 1663 /// \brief Insert \p Inst at the recorded position. 1664 void insert(Instruction *Inst) { 1665 if (HasPrevInstruction) { 1666 if (Inst->getParent()) 1667 Inst->removeFromParent(); 1668 Inst->insertAfter(Point.PrevInst); 1669 } else { 1670 Instruction *Position = Point.BB->getFirstInsertionPt(); 1671 if (Inst->getParent()) 1672 Inst->moveBefore(Position); 1673 else 1674 Inst->insertBefore(Position); 1675 } 1676 } 1677 }; 1678 1679 /// \brief Move an instruction before another. 1680 class InstructionMoveBefore : public TypePromotionAction { 1681 /// Original position of the instruction. 1682 InsertionHandler Position; 1683 1684 public: 1685 /// \brief Move \p Inst before \p Before. 1686 InstructionMoveBefore(Instruction *Inst, Instruction *Before) 1687 : TypePromotionAction(Inst), Position(Inst) { 1688 DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Before << "\n"); 1689 Inst->moveBefore(Before); 1690 } 1691 1692 /// \brief Move the instruction back to its original position. 1693 void undo() override { 1694 DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n"); 1695 Position.insert(Inst); 1696 } 1697 }; 1698 1699 /// \brief Set the operand of an instruction with a new value. 1700 class OperandSetter : public TypePromotionAction { 1701 /// Original operand of the instruction. 1702 Value *Origin; 1703 /// Index of the modified instruction. 1704 unsigned Idx; 1705 1706 public: 1707 /// \brief Set \p Idx operand of \p Inst with \p NewVal. 1708 OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal) 1709 : TypePromotionAction(Inst), Idx(Idx) { 1710 DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n" 1711 << "for:" << *Inst << "\n" 1712 << "with:" << *NewVal << "\n"); 1713 Origin = Inst->getOperand(Idx); 1714 Inst->setOperand(Idx, NewVal); 1715 } 1716 1717 /// \brief Restore the original value of the instruction. 1718 void undo() override { 1719 DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n" 1720 << "for: " << *Inst << "\n" 1721 << "with: " << *Origin << "\n"); 1722 Inst->setOperand(Idx, Origin); 1723 } 1724 }; 1725 1726 /// \brief Hide the operands of an instruction. 1727 /// Do as if this instruction was not using any of its operands. 1728 class OperandsHider : public TypePromotionAction { 1729 /// The list of original operands. 1730 SmallVector<Value *, 4> OriginalValues; 1731 1732 public: 1733 /// \brief Remove \p Inst from the uses of the operands of \p Inst. 1734 OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) { 1735 DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n"); 1736 unsigned NumOpnds = Inst->getNumOperands(); 1737 OriginalValues.reserve(NumOpnds); 1738 for (unsigned It = 0; It < NumOpnds; ++It) { 1739 // Save the current operand. 1740 Value *Val = Inst->getOperand(It); 1741 OriginalValues.push_back(Val); 1742 // Set a dummy one. 1743 // We could use OperandSetter here, but that would implied an overhead 1744 // that we are not willing to pay. 1745 Inst->setOperand(It, UndefValue::get(Val->getType())); 1746 } 1747 } 1748 1749 /// \brief Restore the original list of uses. 1750 void undo() override { 1751 DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n"); 1752 for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It) 1753 Inst->setOperand(It, OriginalValues[It]); 1754 } 1755 }; 1756 1757 /// \brief Build a truncate instruction. 1758 class TruncBuilder : public TypePromotionAction { 1759 Value *Val; 1760 public: 1761 /// \brief Build a truncate instruction of \p Opnd producing a \p Ty 1762 /// result. 1763 /// trunc Opnd to Ty. 1764 TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) { 1765 IRBuilder<> Builder(Opnd); 1766 Val = Builder.CreateTrunc(Opnd, Ty, "promoted"); 1767 DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n"); 1768 } 1769 1770 /// \brief Get the built value. 1771 Value *getBuiltValue() { return Val; } 1772 1773 /// \brief Remove the built instruction. 1774 void undo() override { 1775 DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n"); 1776 if (Instruction *IVal = dyn_cast<Instruction>(Val)) 1777 IVal->eraseFromParent(); 1778 } 1779 }; 1780 1781 /// \brief Build a sign extension instruction. 1782 class SExtBuilder : public TypePromotionAction { 1783 Value *Val; 1784 public: 1785 /// \brief Build a sign extension instruction of \p Opnd producing a \p Ty 1786 /// result. 1787 /// sext Opnd to Ty. 1788 SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty) 1789 : TypePromotionAction(InsertPt) { 1790 IRBuilder<> Builder(InsertPt); 1791 Val = Builder.CreateSExt(Opnd, Ty, "promoted"); 1792 DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n"); 1793 } 1794 1795 /// \brief Get the built value. 1796 Value *getBuiltValue() { return Val; } 1797 1798 /// \brief Remove the built instruction. 1799 void undo() override { 1800 DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n"); 1801 if (Instruction *IVal = dyn_cast<Instruction>(Val)) 1802 IVal->eraseFromParent(); 1803 } 1804 }; 1805 1806 /// \brief Build a zero extension instruction. 1807 class ZExtBuilder : public TypePromotionAction { 1808 Value *Val; 1809 public: 1810 /// \brief Build a zero extension instruction of \p Opnd producing a \p Ty 1811 /// result. 1812 /// zext Opnd to Ty. 1813 ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty) 1814 : TypePromotionAction(InsertPt) { 1815 IRBuilder<> Builder(InsertPt); 1816 Val = Builder.CreateZExt(Opnd, Ty, "promoted"); 1817 DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n"); 1818 } 1819 1820 /// \brief Get the built value. 1821 Value *getBuiltValue() { return Val; } 1822 1823 /// \brief Remove the built instruction. 1824 void undo() override { 1825 DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n"); 1826 if (Instruction *IVal = dyn_cast<Instruction>(Val)) 1827 IVal->eraseFromParent(); 1828 } 1829 }; 1830 1831 /// \brief Mutate an instruction to another type. 1832 class TypeMutator : public TypePromotionAction { 1833 /// Record the original type. 1834 Type *OrigTy; 1835 1836 public: 1837 /// \brief Mutate the type of \p Inst into \p NewTy. 1838 TypeMutator(Instruction *Inst, Type *NewTy) 1839 : TypePromotionAction(Inst), OrigTy(Inst->getType()) { 1840 DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTy 1841 << "\n"); 1842 Inst->mutateType(NewTy); 1843 } 1844 1845 /// \brief Mutate the instruction back to its original type. 1846 void undo() override { 1847 DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTy 1848 << "\n"); 1849 Inst->mutateType(OrigTy); 1850 } 1851 }; 1852 1853 /// \brief Replace the uses of an instruction by another instruction. 1854 class UsesReplacer : public TypePromotionAction { 1855 /// Helper structure to keep track of the replaced uses. 1856 struct InstructionAndIdx { 1857 /// The instruction using the instruction. 1858 Instruction *Inst; 1859 /// The index where this instruction is used for Inst. 1860 unsigned Idx; 1861 InstructionAndIdx(Instruction *Inst, unsigned Idx) 1862 : Inst(Inst), Idx(Idx) {} 1863 }; 1864 1865 /// Keep track of the original uses (pair Instruction, Index). 1866 SmallVector<InstructionAndIdx, 4> OriginalUses; 1867 typedef SmallVectorImpl<InstructionAndIdx>::iterator use_iterator; 1868 1869 public: 1870 /// \brief Replace all the use of \p Inst by \p New. 1871 UsesReplacer(Instruction *Inst, Value *New) : TypePromotionAction(Inst) { 1872 DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *New 1873 << "\n"); 1874 // Record the original uses. 1875 for (Use &U : Inst->uses()) { 1876 Instruction *UserI = cast<Instruction>(U.getUser()); 1877 OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo())); 1878 } 1879 // Now, we can replace the uses. 1880 Inst->replaceAllUsesWith(New); 1881 } 1882 1883 /// \brief Reassign the original uses of Inst to Inst. 1884 void undo() override { 1885 DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n"); 1886 for (use_iterator UseIt = OriginalUses.begin(), 1887 EndIt = OriginalUses.end(); 1888 UseIt != EndIt; ++UseIt) { 1889 UseIt->Inst->setOperand(UseIt->Idx, Inst); 1890 } 1891 } 1892 }; 1893 1894 /// \brief Remove an instruction from the IR. 1895 class InstructionRemover : public TypePromotionAction { 1896 /// Original position of the instruction. 1897 InsertionHandler Inserter; 1898 /// Helper structure to hide all the link to the instruction. In other 1899 /// words, this helps to do as if the instruction was removed. 1900 OperandsHider Hider; 1901 /// Keep track of the uses replaced, if any. 1902 UsesReplacer *Replacer; 1903 1904 public: 1905 /// \brief Remove all reference of \p Inst and optinally replace all its 1906 /// uses with New. 1907 /// \pre If !Inst->use_empty(), then New != nullptr 1908 InstructionRemover(Instruction *Inst, Value *New = nullptr) 1909 : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst), 1910 Replacer(nullptr) { 1911 if (New) 1912 Replacer = new UsesReplacer(Inst, New); 1913 DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n"); 1914 Inst->removeFromParent(); 1915 } 1916 1917 ~InstructionRemover() override { delete Replacer; } 1918 1919 /// \brief Really remove the instruction. 1920 void commit() override { delete Inst; } 1921 1922 /// \brief Resurrect the instruction and reassign it to the proper uses if 1923 /// new value was provided when build this action. 1924 void undo() override { 1925 DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n"); 1926 Inserter.insert(Inst); 1927 if (Replacer) 1928 Replacer->undo(); 1929 Hider.undo(); 1930 } 1931 }; 1932 1933public: 1934 /// Restoration point. 1935 /// The restoration point is a pointer to an action instead of an iterator 1936 /// because the iterator may be invalidated but not the pointer. 1937 typedef const TypePromotionAction *ConstRestorationPt; 1938 /// Advocate every changes made in that transaction. 1939 void commit(); 1940 /// Undo all the changes made after the given point. 1941 void rollback(ConstRestorationPt Point); 1942 /// Get the current restoration point. 1943 ConstRestorationPt getRestorationPoint() const; 1944 1945 /// \name API for IR modification with state keeping to support rollback. 1946 /// @{ 1947 /// Same as Instruction::setOperand. 1948 void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal); 1949 /// Same as Instruction::eraseFromParent. 1950 void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr); 1951 /// Same as Value::replaceAllUsesWith. 1952 void replaceAllUsesWith(Instruction *Inst, Value *New); 1953 /// Same as Value::mutateType. 1954 void mutateType(Instruction *Inst, Type *NewTy); 1955 /// Same as IRBuilder::createTrunc. 1956 Value *createTrunc(Instruction *Opnd, Type *Ty); 1957 /// Same as IRBuilder::createSExt. 1958 Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty); 1959 /// Same as IRBuilder::createZExt. 1960 Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty); 1961 /// Same as Instruction::moveBefore. 1962 void moveBefore(Instruction *Inst, Instruction *Before); 1963 /// @} 1964 1965private: 1966 /// The ordered list of actions made so far. 1967 SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions; 1968 typedef SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator CommitPt; 1969}; 1970 1971void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx, 1972 Value *NewVal) { 1973 Actions.push_back( 1974 make_unique<TypePromotionTransaction::OperandSetter>(Inst, Idx, NewVal)); 1975} 1976 1977void TypePromotionTransaction::eraseInstruction(Instruction *Inst, 1978 Value *NewVal) { 1979 Actions.push_back( 1980 make_unique<TypePromotionTransaction::InstructionRemover>(Inst, NewVal)); 1981} 1982 1983void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst, 1984 Value *New) { 1985 Actions.push_back(make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New)); 1986} 1987 1988void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) { 1989 Actions.push_back(make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy)); 1990} 1991 1992Value *TypePromotionTransaction::createTrunc(Instruction *Opnd, 1993 Type *Ty) { 1994 std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty)); 1995 Value *Val = Ptr->getBuiltValue(); 1996 Actions.push_back(std::move(Ptr)); 1997 return Val; 1998} 1999 2000Value *TypePromotionTransaction::createSExt(Instruction *Inst, 2001 Value *Opnd, Type *Ty) { 2002 std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty)); 2003 Value *Val = Ptr->getBuiltValue(); 2004 Actions.push_back(std::move(Ptr)); 2005 return Val; 2006} 2007 2008Value *TypePromotionTransaction::createZExt(Instruction *Inst, 2009 Value *Opnd, Type *Ty) { 2010 std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty)); 2011 Value *Val = Ptr->getBuiltValue(); 2012 Actions.push_back(std::move(Ptr)); 2013 return Val; 2014} 2015 2016void TypePromotionTransaction::moveBefore(Instruction *Inst, 2017 Instruction *Before) { 2018 Actions.push_back( 2019 make_unique<TypePromotionTransaction::InstructionMoveBefore>(Inst, Before)); 2020} 2021 2022TypePromotionTransaction::ConstRestorationPt 2023TypePromotionTransaction::getRestorationPoint() const { 2024 return !Actions.empty() ? Actions.back().get() : nullptr; 2025} 2026 2027void TypePromotionTransaction::commit() { 2028 for (CommitPt It = Actions.begin(), EndIt = Actions.end(); It != EndIt; 2029 ++It) 2030 (*It)->commit(); 2031 Actions.clear(); 2032} 2033 2034void TypePromotionTransaction::rollback( 2035 TypePromotionTransaction::ConstRestorationPt Point) { 2036 while (!Actions.empty() && Point != Actions.back().get()) { 2037 std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val(); 2038 Curr->undo(); 2039 } 2040} 2041 2042/// \brief A helper class for matching addressing modes. 2043/// 2044/// This encapsulates the logic for matching the target-legal addressing modes. 2045class AddressingModeMatcher { 2046 SmallVectorImpl<Instruction*> &AddrModeInsts; 2047 const TargetMachine &TM; 2048 const TargetLowering &TLI; 2049 2050 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and 2051 /// the memory instruction that we're computing this address for. 2052 Type *AccessTy; 2053 Instruction *MemoryInst; 2054 2055 /// AddrMode - This is the addressing mode that we're building up. This is 2056 /// part of the return value of this addressing mode matching stuff. 2057 ExtAddrMode &AddrMode; 2058 2059 /// The truncate instruction inserted by other CodeGenPrepare optimizations. 2060 const SetOfInstrs &InsertedTruncs; 2061 /// A map from the instructions to their type before promotion. 2062 InstrToOrigTy &PromotedInsts; 2063 /// The ongoing transaction where every action should be registered. 2064 TypePromotionTransaction &TPT; 2065 2066 /// IgnoreProfitability - This is set to true when we should not do 2067 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode 2068 /// always returns true. 2069 bool IgnoreProfitability; 2070 2071 AddressingModeMatcher(SmallVectorImpl<Instruction *> &AMI, 2072 const TargetMachine &TM, Type *AT, Instruction *MI, 2073 ExtAddrMode &AM, const SetOfInstrs &InsertedTruncs, 2074 InstrToOrigTy &PromotedInsts, 2075 TypePromotionTransaction &TPT) 2076 : AddrModeInsts(AMI), TM(TM), 2077 TLI(*TM.getSubtargetImpl(*MI->getParent()->getParent()) 2078 ->getTargetLowering()), 2079 AccessTy(AT), MemoryInst(MI), AddrMode(AM), 2080 InsertedTruncs(InsertedTruncs), PromotedInsts(PromotedInsts), TPT(TPT) { 2081 IgnoreProfitability = false; 2082 } 2083public: 2084 2085 /// Match - Find the maximal addressing mode that a load/store of V can fold, 2086 /// give an access type of AccessTy. This returns a list of involved 2087 /// instructions in AddrModeInsts. 2088 /// \p InsertedTruncs The truncate instruction inserted by other 2089 /// CodeGenPrepare 2090 /// optimizations. 2091 /// \p PromotedInsts maps the instructions to their type before promotion. 2092 /// \p The ongoing transaction where every action should be registered. 2093 static ExtAddrMode Match(Value *V, Type *AccessTy, 2094 Instruction *MemoryInst, 2095 SmallVectorImpl<Instruction*> &AddrModeInsts, 2096 const TargetMachine &TM, 2097 const SetOfInstrs &InsertedTruncs, 2098 InstrToOrigTy &PromotedInsts, 2099 TypePromotionTransaction &TPT) { 2100 ExtAddrMode Result; 2101 2102 bool Success = AddressingModeMatcher(AddrModeInsts, TM, AccessTy, 2103 MemoryInst, Result, InsertedTruncs, 2104 PromotedInsts, TPT).MatchAddr(V, 0); 2105 (void)Success; assert(Success && "Couldn't select *anything*?"); 2106 return Result; 2107 } 2108private: 2109 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth); 2110 bool MatchAddr(Value *V, unsigned Depth); 2111 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth, 2112 bool *MovedAway = nullptr); 2113 bool IsProfitableToFoldIntoAddressingMode(Instruction *I, 2114 ExtAddrMode &AMBefore, 2115 ExtAddrMode &AMAfter); 2116 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2); 2117 bool IsPromotionProfitable(unsigned NewCost, unsigned OldCost, 2118 Value *PromotedOperand) const; 2119}; 2120 2121/// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode. 2122/// Return true and update AddrMode if this addr mode is legal for the target, 2123/// false if not. 2124bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale, 2125 unsigned Depth) { 2126 // If Scale is 1, then this is the same as adding ScaleReg to the addressing 2127 // mode. Just process that directly. 2128 if (Scale == 1) 2129 return MatchAddr(ScaleReg, Depth); 2130 2131 // If the scale is 0, it takes nothing to add this. 2132 if (Scale == 0) 2133 return true; 2134 2135 // If we already have a scale of this value, we can add to it, otherwise, we 2136 // need an available scale field. 2137 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg) 2138 return false; 2139 2140 ExtAddrMode TestAddrMode = AddrMode; 2141 2142 // Add scale to turn X*4+X*3 -> X*7. This could also do things like 2143 // [A+B + A*7] -> [B+A*8]. 2144 TestAddrMode.Scale += Scale; 2145 TestAddrMode.ScaledReg = ScaleReg; 2146 2147 // If the new address isn't legal, bail out. 2148 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) 2149 return false; 2150 2151 // It was legal, so commit it. 2152 AddrMode = TestAddrMode; 2153 2154 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now 2155 // to see if ScaleReg is actually X+C. If so, we can turn this into adding 2156 // X*Scale + C*Scale to addr mode. 2157 ConstantInt *CI = nullptr; Value *AddLHS = nullptr; 2158 if (isa<Instruction>(ScaleReg) && // not a constant expr. 2159 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) { 2160 TestAddrMode.ScaledReg = AddLHS; 2161 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale; 2162 2163 // If this addressing mode is legal, commit it and remember that we folded 2164 // this instruction. 2165 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) { 2166 AddrModeInsts.push_back(cast<Instruction>(ScaleReg)); 2167 AddrMode = TestAddrMode; 2168 return true; 2169 } 2170 } 2171 2172 // Otherwise, not (x+c)*scale, just return what we have. 2173 return true; 2174} 2175 2176/// MightBeFoldableInst - This is a little filter, which returns true if an 2177/// addressing computation involving I might be folded into a load/store 2178/// accessing it. This doesn't need to be perfect, but needs to accept at least 2179/// the set of instructions that MatchOperationAddr can. 2180static bool MightBeFoldableInst(Instruction *I) { 2181 switch (I->getOpcode()) { 2182 case Instruction::BitCast: 2183 case Instruction::AddrSpaceCast: 2184 // Don't touch identity bitcasts. 2185 if (I->getType() == I->getOperand(0)->getType()) 2186 return false; 2187 return I->getType()->isPointerTy() || I->getType()->isIntegerTy(); 2188 case Instruction::PtrToInt: 2189 // PtrToInt is always a noop, as we know that the int type is pointer sized. 2190 return true; 2191 case Instruction::IntToPtr: 2192 // We know the input is intptr_t, so this is foldable. 2193 return true; 2194 case Instruction::Add: 2195 return true; 2196 case Instruction::Mul: 2197 case Instruction::Shl: 2198 // Can only handle X*C and X << C. 2199 return isa<ConstantInt>(I->getOperand(1)); 2200 case Instruction::GetElementPtr: 2201 return true; 2202 default: 2203 return false; 2204 } 2205} 2206 2207/// \brief Check whether or not \p Val is a legal instruction for \p TLI. 2208/// \note \p Val is assumed to be the product of some type promotion. 2209/// Therefore if \p Val has an undefined state in \p TLI, this is assumed 2210/// to be legal, as the non-promoted value would have had the same state. 2211static bool isPromotedInstructionLegal(const TargetLowering &TLI, Value *Val) { 2212 Instruction *PromotedInst = dyn_cast<Instruction>(Val); 2213 if (!PromotedInst) 2214 return false; 2215 int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode()); 2216 // If the ISDOpcode is undefined, it was undefined before the promotion. 2217 if (!ISDOpcode) 2218 return true; 2219 // Otherwise, check if the promoted instruction is legal or not. 2220 return TLI.isOperationLegalOrCustom( 2221 ISDOpcode, TLI.getValueType(PromotedInst->getType())); 2222} 2223 2224/// \brief Hepler class to perform type promotion. 2225class TypePromotionHelper { 2226 /// \brief Utility function to check whether or not a sign or zero extension 2227 /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by 2228 /// either using the operands of \p Inst or promoting \p Inst. 2229 /// The type of the extension is defined by \p IsSExt. 2230 /// In other words, check if: 2231 /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType. 2232 /// #1 Promotion applies: 2233 /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...). 2234 /// #2 Operand reuses: 2235 /// ext opnd1 to ConsideredExtType. 2236 /// \p PromotedInsts maps the instructions to their type before promotion. 2237 static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType, 2238 const InstrToOrigTy &PromotedInsts, bool IsSExt); 2239 2240 /// \brief Utility function to determine if \p OpIdx should be promoted when 2241 /// promoting \p Inst. 2242 static bool shouldExtOperand(const Instruction *Inst, int OpIdx) { 2243 if (isa<SelectInst>(Inst) && OpIdx == 0) 2244 return false; 2245 return true; 2246 } 2247 2248 /// \brief Utility function to promote the operand of \p Ext when this 2249 /// operand is a promotable trunc or sext or zext. 2250 /// \p PromotedInsts maps the instructions to their type before promotion. 2251 /// \p CreatedInstsCost[out] contains the cost of all instructions 2252 /// created to promote the operand of Ext. 2253 /// Newly added extensions are inserted in \p Exts. 2254 /// Newly added truncates are inserted in \p Truncs. 2255 /// Should never be called directly. 2256 /// \return The promoted value which is used instead of Ext. 2257 static Value *promoteOperandForTruncAndAnyExt( 2258 Instruction *Ext, TypePromotionTransaction &TPT, 2259 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, 2260 SmallVectorImpl<Instruction *> *Exts, 2261 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI); 2262 2263 /// \brief Utility function to promote the operand of \p Ext when this 2264 /// operand is promotable and is not a supported trunc or sext. 2265 /// \p PromotedInsts maps the instructions to their type before promotion. 2266 /// \p CreatedInstsCost[out] contains the cost of all the instructions 2267 /// created to promote the operand of Ext. 2268 /// Newly added extensions are inserted in \p Exts. 2269 /// Newly added truncates are inserted in \p Truncs. 2270 /// Should never be called directly. 2271 /// \return The promoted value which is used instead of Ext. 2272 static Value *promoteOperandForOther(Instruction *Ext, 2273 TypePromotionTransaction &TPT, 2274 InstrToOrigTy &PromotedInsts, 2275 unsigned &CreatedInstsCost, 2276 SmallVectorImpl<Instruction *> *Exts, 2277 SmallVectorImpl<Instruction *> *Truncs, 2278 const TargetLowering &TLI, bool IsSExt); 2279 2280 /// \see promoteOperandForOther. 2281 static Value *signExtendOperandForOther( 2282 Instruction *Ext, TypePromotionTransaction &TPT, 2283 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, 2284 SmallVectorImpl<Instruction *> *Exts, 2285 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) { 2286 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost, 2287 Exts, Truncs, TLI, true); 2288 } 2289 2290 /// \see promoteOperandForOther. 2291 static Value *zeroExtendOperandForOther( 2292 Instruction *Ext, TypePromotionTransaction &TPT, 2293 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, 2294 SmallVectorImpl<Instruction *> *Exts, 2295 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) { 2296 return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost, 2297 Exts, Truncs, TLI, false); 2298 } 2299 2300public: 2301 /// Type for the utility function that promotes the operand of Ext. 2302 typedef Value *(*Action)(Instruction *Ext, TypePromotionTransaction &TPT, 2303 InstrToOrigTy &PromotedInsts, 2304 unsigned &CreatedInstsCost, 2305 SmallVectorImpl<Instruction *> *Exts, 2306 SmallVectorImpl<Instruction *> *Truncs, 2307 const TargetLowering &TLI); 2308 /// \brief Given a sign/zero extend instruction \p Ext, return the approriate 2309 /// action to promote the operand of \p Ext instead of using Ext. 2310 /// \return NULL if no promotable action is possible with the current 2311 /// sign extension. 2312 /// \p InsertedTruncs keeps track of all the truncate instructions inserted by 2313 /// the others CodeGenPrepare optimizations. This information is important 2314 /// because we do not want to promote these instructions as CodeGenPrepare 2315 /// will reinsert them later. Thus creating an infinite loop: create/remove. 2316 /// \p PromotedInsts maps the instructions to their type before promotion. 2317 static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedTruncs, 2318 const TargetLowering &TLI, 2319 const InstrToOrigTy &PromotedInsts); 2320}; 2321 2322bool TypePromotionHelper::canGetThrough(const Instruction *Inst, 2323 Type *ConsideredExtType, 2324 const InstrToOrigTy &PromotedInsts, 2325 bool IsSExt) { 2326 // The promotion helper does not know how to deal with vector types yet. 2327 // To be able to fix that, we would need to fix the places where we 2328 // statically extend, e.g., constants and such. 2329 if (Inst->getType()->isVectorTy()) 2330 return false; 2331 2332 // We can always get through zext. 2333 if (isa<ZExtInst>(Inst)) 2334 return true; 2335 2336 // sext(sext) is ok too. 2337 if (IsSExt && isa<SExtInst>(Inst)) 2338 return true; 2339 2340 // We can get through binary operator, if it is legal. In other words, the 2341 // binary operator must have a nuw or nsw flag. 2342 const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst); 2343 if (BinOp && isa<OverflowingBinaryOperator>(BinOp) && 2344 ((!IsSExt && BinOp->hasNoUnsignedWrap()) || 2345 (IsSExt && BinOp->hasNoSignedWrap()))) 2346 return true; 2347 2348 // Check if we can do the following simplification. 2349 // ext(trunc(opnd)) --> ext(opnd) 2350 if (!isa<TruncInst>(Inst)) 2351 return false; 2352 2353 Value *OpndVal = Inst->getOperand(0); 2354 // Check if we can use this operand in the extension. 2355 // If the type is larger than the result type of the extension, 2356 // we cannot. 2357 if (!OpndVal->getType()->isIntegerTy() || 2358 OpndVal->getType()->getIntegerBitWidth() > 2359 ConsideredExtType->getIntegerBitWidth()) 2360 return false; 2361 2362 // If the operand of the truncate is not an instruction, we will not have 2363 // any information on the dropped bits. 2364 // (Actually we could for constant but it is not worth the extra logic). 2365 Instruction *Opnd = dyn_cast<Instruction>(OpndVal); 2366 if (!Opnd) 2367 return false; 2368 2369 // Check if the source of the type is narrow enough. 2370 // I.e., check that trunc just drops extended bits of the same kind of 2371 // the extension. 2372 // #1 get the type of the operand and check the kind of the extended bits. 2373 const Type *OpndType; 2374 InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd); 2375 if (It != PromotedInsts.end() && It->second.IsSExt == IsSExt) 2376 OpndType = It->second.Ty; 2377 else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd))) 2378 OpndType = Opnd->getOperand(0)->getType(); 2379 else 2380 return false; 2381 2382 // #2 check that the truncate just drop extended bits. 2383 if (Inst->getType()->getIntegerBitWidth() >= OpndType->getIntegerBitWidth()) 2384 return true; 2385 2386 return false; 2387} 2388 2389TypePromotionHelper::Action TypePromotionHelper::getAction( 2390 Instruction *Ext, const SetOfInstrs &InsertedTruncs, 2391 const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) { 2392 assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) && 2393 "Unexpected instruction type"); 2394 Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0)); 2395 Type *ExtTy = Ext->getType(); 2396 bool IsSExt = isa<SExtInst>(Ext); 2397 // If the operand of the extension is not an instruction, we cannot 2398 // get through. 2399 // If it, check we can get through. 2400 if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt)) 2401 return nullptr; 2402 2403 // Do not promote if the operand has been added by codegenprepare. 2404 // Otherwise, it means we are undoing an optimization that is likely to be 2405 // redone, thus causing potential infinite loop. 2406 if (isa<TruncInst>(ExtOpnd) && InsertedTruncs.count(ExtOpnd)) 2407 return nullptr; 2408 2409 // SExt or Trunc instructions. 2410 // Return the related handler. 2411 if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) || 2412 isa<ZExtInst>(ExtOpnd)) 2413 return promoteOperandForTruncAndAnyExt; 2414 2415 // Regular instruction. 2416 // Abort early if we will have to insert non-free instructions. 2417 if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType())) 2418 return nullptr; 2419 return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther; 2420} 2421 2422Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt( 2423 llvm::Instruction *SExt, TypePromotionTransaction &TPT, 2424 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, 2425 SmallVectorImpl<Instruction *> *Exts, 2426 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) { 2427 // By construction, the operand of SExt is an instruction. Otherwise we cannot 2428 // get through it and this method should not be called. 2429 Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0)); 2430 Value *ExtVal = SExt; 2431 bool HasMergedNonFreeExt = false; 2432 if (isa<ZExtInst>(SExtOpnd)) { 2433 // Replace s|zext(zext(opnd)) 2434 // => zext(opnd). 2435 HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd); 2436 Value *ZExt = 2437 TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType()); 2438 TPT.replaceAllUsesWith(SExt, ZExt); 2439 TPT.eraseInstruction(SExt); 2440 ExtVal = ZExt; 2441 } else { 2442 // Replace z|sext(trunc(opnd)) or sext(sext(opnd)) 2443 // => z|sext(opnd). 2444 TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0)); 2445 } 2446 CreatedInstsCost = 0; 2447 2448 // Remove dead code. 2449 if (SExtOpnd->use_empty()) 2450 TPT.eraseInstruction(SExtOpnd); 2451 2452 // Check if the extension is still needed. 2453 Instruction *ExtInst = dyn_cast<Instruction>(ExtVal); 2454 if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) { 2455 if (ExtInst) { 2456 if (Exts) 2457 Exts->push_back(ExtInst); 2458 CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt; 2459 } 2460 return ExtVal; 2461 } 2462 2463 // At this point we have: ext ty opnd to ty. 2464 // Reassign the uses of ExtInst to the opnd and remove ExtInst. 2465 Value *NextVal = ExtInst->getOperand(0); 2466 TPT.eraseInstruction(ExtInst, NextVal); 2467 return NextVal; 2468} 2469 2470Value *TypePromotionHelper::promoteOperandForOther( 2471 Instruction *Ext, TypePromotionTransaction &TPT, 2472 InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, 2473 SmallVectorImpl<Instruction *> *Exts, 2474 SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI, 2475 bool IsSExt) { 2476 // By construction, the operand of Ext is an instruction. Otherwise we cannot 2477 // get through it and this method should not be called. 2478 Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0)); 2479 CreatedInstsCost = 0; 2480 if (!ExtOpnd->hasOneUse()) { 2481 // ExtOpnd will be promoted. 2482 // All its uses, but Ext, will need to use a truncated value of the 2483 // promoted version. 2484 // Create the truncate now. 2485 Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType()); 2486 if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) { 2487 ITrunc->removeFromParent(); 2488 // Insert it just after the definition. 2489 ITrunc->insertAfter(ExtOpnd); 2490 if (Truncs) 2491 Truncs->push_back(ITrunc); 2492 } 2493 2494 TPT.replaceAllUsesWith(ExtOpnd, Trunc); 2495 // Restore the operand of Ext (which has been replace by the previous call 2496 // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext. 2497 TPT.setOperand(Ext, 0, ExtOpnd); 2498 } 2499 2500 // Get through the Instruction: 2501 // 1. Update its type. 2502 // 2. Replace the uses of Ext by Inst. 2503 // 3. Extend each operand that needs to be extended. 2504 2505 // Remember the original type of the instruction before promotion. 2506 // This is useful to know that the high bits are sign extended bits. 2507 PromotedInsts.insert(std::pair<Instruction *, TypeIsSExt>( 2508 ExtOpnd, TypeIsSExt(ExtOpnd->getType(), IsSExt))); 2509 // Step #1. 2510 TPT.mutateType(ExtOpnd, Ext->getType()); 2511 // Step #2. 2512 TPT.replaceAllUsesWith(Ext, ExtOpnd); 2513 // Step #3. 2514 Instruction *ExtForOpnd = Ext; 2515 2516 DEBUG(dbgs() << "Propagate Ext to operands\n"); 2517 for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx; 2518 ++OpIdx) { 2519 DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n'); 2520 if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() || 2521 !shouldExtOperand(ExtOpnd, OpIdx)) { 2522 DEBUG(dbgs() << "No need to propagate\n"); 2523 continue; 2524 } 2525 // Check if we can statically extend the operand. 2526 Value *Opnd = ExtOpnd->getOperand(OpIdx); 2527 if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) { 2528 DEBUG(dbgs() << "Statically extend\n"); 2529 unsigned BitWidth = Ext->getType()->getIntegerBitWidth(); 2530 APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth) 2531 : Cst->getValue().zext(BitWidth); 2532 TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal)); 2533 continue; 2534 } 2535 // UndefValue are typed, so we have to statically sign extend them. 2536 if (isa<UndefValue>(Opnd)) { 2537 DEBUG(dbgs() << "Statically extend\n"); 2538 TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType())); 2539 continue; 2540 } 2541 2542 // Otherwise we have to explicity sign extend the operand. 2543 // Check if Ext was reused to extend an operand. 2544 if (!ExtForOpnd) { 2545 // If yes, create a new one. 2546 DEBUG(dbgs() << "More operands to ext\n"); 2547 Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType()) 2548 : TPT.createZExt(Ext, Opnd, Ext->getType()); 2549 if (!isa<Instruction>(ValForExtOpnd)) { 2550 TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd); 2551 continue; 2552 } 2553 ExtForOpnd = cast<Instruction>(ValForExtOpnd); 2554 } 2555 if (Exts) 2556 Exts->push_back(ExtForOpnd); 2557 TPT.setOperand(ExtForOpnd, 0, Opnd); 2558 2559 // Move the sign extension before the insertion point. 2560 TPT.moveBefore(ExtForOpnd, ExtOpnd); 2561 TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd); 2562 CreatedInstsCost += !TLI.isExtFree(ExtForOpnd); 2563 // If more sext are required, new instructions will have to be created. 2564 ExtForOpnd = nullptr; 2565 } 2566 if (ExtForOpnd == Ext) { 2567 DEBUG(dbgs() << "Extension is useless now\n"); 2568 TPT.eraseInstruction(Ext); 2569 } 2570 return ExtOpnd; 2571} 2572 2573/// IsPromotionProfitable - Check whether or not promoting an instruction 2574/// to a wider type was profitable. 2575/// \p NewCost gives the cost of extension instructions created by the 2576/// promotion. 2577/// \p OldCost gives the cost of extension instructions before the promotion 2578/// plus the number of instructions that have been 2579/// matched in the addressing mode the promotion. 2580/// \p PromotedOperand is the value that has been promoted. 2581/// \return True if the promotion is profitable, false otherwise. 2582bool AddressingModeMatcher::IsPromotionProfitable( 2583 unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const { 2584 DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCost << '\n'); 2585 // The cost of the new extensions is greater than the cost of the 2586 // old extension plus what we folded. 2587 // This is not profitable. 2588 if (NewCost > OldCost) 2589 return false; 2590 if (NewCost < OldCost) 2591 return true; 2592 // The promotion is neutral but it may help folding the sign extension in 2593 // loads for instance. 2594 // Check that we did not create an illegal instruction. 2595 return isPromotedInstructionLegal(TLI, PromotedOperand); 2596} 2597 2598/// MatchOperationAddr - Given an instruction or constant expr, see if we can 2599/// fold the operation into the addressing mode. If so, update the addressing 2600/// mode and return true, otherwise return false without modifying AddrMode. 2601/// If \p MovedAway is not NULL, it contains the information of whether or 2602/// not AddrInst has to be folded into the addressing mode on success. 2603/// If \p MovedAway == true, \p AddrInst will not be part of the addressing 2604/// because it has been moved away. 2605/// Thus AddrInst must not be added in the matched instructions. 2606/// This state can happen when AddrInst is a sext, since it may be moved away. 2607/// Therefore, AddrInst may not be valid when MovedAway is true and it must 2608/// not be referenced anymore. 2609bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode, 2610 unsigned Depth, 2611 bool *MovedAway) { 2612 // Avoid exponential behavior on extremely deep expression trees. 2613 if (Depth >= 5) return false; 2614 2615 // By default, all matched instructions stay in place. 2616 if (MovedAway) 2617 *MovedAway = false; 2618 2619 switch (Opcode) { 2620 case Instruction::PtrToInt: 2621 // PtrToInt is always a noop, as we know that the int type is pointer sized. 2622 return MatchAddr(AddrInst->getOperand(0), Depth); 2623 case Instruction::IntToPtr: 2624 // This inttoptr is a no-op if the integer type is pointer sized. 2625 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) == 2626 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace())) 2627 return MatchAddr(AddrInst->getOperand(0), Depth); 2628 return false; 2629 case Instruction::BitCast: 2630 case Instruction::AddrSpaceCast: 2631 // BitCast is always a noop, and we can handle it as long as it is 2632 // int->int or pointer->pointer (we don't want int<->fp or something). 2633 if ((AddrInst->getOperand(0)->getType()->isPointerTy() || 2634 AddrInst->getOperand(0)->getType()->isIntegerTy()) && 2635 // Don't touch identity bitcasts. These were probably put here by LSR, 2636 // and we don't want to mess around with them. Assume it knows what it 2637 // is doing. 2638 AddrInst->getOperand(0)->getType() != AddrInst->getType()) 2639 return MatchAddr(AddrInst->getOperand(0), Depth); 2640 return false; 2641 case Instruction::Add: { 2642 // Check to see if we can merge in the RHS then the LHS. If so, we win. 2643 ExtAddrMode BackupAddrMode = AddrMode; 2644 unsigned OldSize = AddrModeInsts.size(); 2645 // Start a transaction at this point. 2646 // The LHS may match but not the RHS. 2647 // Therefore, we need a higher level restoration point to undo partially 2648 // matched operation. 2649 TypePromotionTransaction::ConstRestorationPt LastKnownGood = 2650 TPT.getRestorationPoint(); 2651 2652 if (MatchAddr(AddrInst->getOperand(1), Depth+1) && 2653 MatchAddr(AddrInst->getOperand(0), Depth+1)) 2654 return true; 2655 2656 // Restore the old addr mode info. 2657 AddrMode = BackupAddrMode; 2658 AddrModeInsts.resize(OldSize); 2659 TPT.rollback(LastKnownGood); 2660 2661 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS. 2662 if (MatchAddr(AddrInst->getOperand(0), Depth+1) && 2663 MatchAddr(AddrInst->getOperand(1), Depth+1)) 2664 return true; 2665 2666 // Otherwise we definitely can't merge the ADD in. 2667 AddrMode = BackupAddrMode; 2668 AddrModeInsts.resize(OldSize); 2669 TPT.rollback(LastKnownGood); 2670 break; 2671 } 2672 //case Instruction::Or: 2673 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD. 2674 //break; 2675 case Instruction::Mul: 2676 case Instruction::Shl: { 2677 // Can only handle X*C and X << C. 2678 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); 2679 if (!RHS) 2680 return false; 2681 int64_t Scale = RHS->getSExtValue(); 2682 if (Opcode == Instruction::Shl) 2683 Scale = 1LL << Scale; 2684 2685 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth); 2686 } 2687 case Instruction::GetElementPtr: { 2688 // Scan the GEP. We check it if it contains constant offsets and at most 2689 // one variable offset. 2690 int VariableOperand = -1; 2691 unsigned VariableScale = 0; 2692 2693 int64_t ConstantOffset = 0; 2694 const DataLayout *TD = TLI.getDataLayout(); 2695 gep_type_iterator GTI = gep_type_begin(AddrInst); 2696 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) { 2697 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 2698 const StructLayout *SL = TD->getStructLayout(STy); 2699 unsigned Idx = 2700 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue(); 2701 ConstantOffset += SL->getElementOffset(Idx); 2702 } else { 2703 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType()); 2704 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) { 2705 ConstantOffset += CI->getSExtValue()*TypeSize; 2706 } else if (TypeSize) { // Scales of zero don't do anything. 2707 // We only allow one variable index at the moment. 2708 if (VariableOperand != -1) 2709 return false; 2710 2711 // Remember the variable index. 2712 VariableOperand = i; 2713 VariableScale = TypeSize; 2714 } 2715 } 2716 } 2717 2718 // A common case is for the GEP to only do a constant offset. In this case, 2719 // just add it to the disp field and check validity. 2720 if (VariableOperand == -1) { 2721 AddrMode.BaseOffs += ConstantOffset; 2722 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){ 2723 // Check to see if we can fold the base pointer in too. 2724 if (MatchAddr(AddrInst->getOperand(0), Depth+1)) 2725 return true; 2726 } 2727 AddrMode.BaseOffs -= ConstantOffset; 2728 return false; 2729 } 2730 2731 // Save the valid addressing mode in case we can't match. 2732 ExtAddrMode BackupAddrMode = AddrMode; 2733 unsigned OldSize = AddrModeInsts.size(); 2734 2735 // See if the scale and offset amount is valid for this target. 2736 AddrMode.BaseOffs += ConstantOffset; 2737 2738 // Match the base operand of the GEP. 2739 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) { 2740 // If it couldn't be matched, just stuff the value in a register. 2741 if (AddrMode.HasBaseReg) { 2742 AddrMode = BackupAddrMode; 2743 AddrModeInsts.resize(OldSize); 2744 return false; 2745 } 2746 AddrMode.HasBaseReg = true; 2747 AddrMode.BaseReg = AddrInst->getOperand(0); 2748 } 2749 2750 // Match the remaining variable portion of the GEP. 2751 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale, 2752 Depth)) { 2753 // If it couldn't be matched, try stuffing the base into a register 2754 // instead of matching it, and retrying the match of the scale. 2755 AddrMode = BackupAddrMode; 2756 AddrModeInsts.resize(OldSize); 2757 if (AddrMode.HasBaseReg) 2758 return false; 2759 AddrMode.HasBaseReg = true; 2760 AddrMode.BaseReg = AddrInst->getOperand(0); 2761 AddrMode.BaseOffs += ConstantOffset; 2762 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), 2763 VariableScale, Depth)) { 2764 // If even that didn't work, bail. 2765 AddrMode = BackupAddrMode; 2766 AddrModeInsts.resize(OldSize); 2767 return false; 2768 } 2769 } 2770 2771 return true; 2772 } 2773 case Instruction::SExt: 2774 case Instruction::ZExt: { 2775 Instruction *Ext = dyn_cast<Instruction>(AddrInst); 2776 if (!Ext) 2777 return false; 2778 2779 // Try to move this ext out of the way of the addressing mode. 2780 // Ask for a method for doing so. 2781 TypePromotionHelper::Action TPH = 2782 TypePromotionHelper::getAction(Ext, InsertedTruncs, TLI, PromotedInsts); 2783 if (!TPH) 2784 return false; 2785 2786 TypePromotionTransaction::ConstRestorationPt LastKnownGood = 2787 TPT.getRestorationPoint(); 2788 unsigned CreatedInstsCost = 0; 2789 unsigned ExtCost = !TLI.isExtFree(Ext); 2790 Value *PromotedOperand = 2791 TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI); 2792 // SExt has been moved away. 2793 // Thus either it will be rematched later in the recursive calls or it is 2794 // gone. Anyway, we must not fold it into the addressing mode at this point. 2795 // E.g., 2796 // op = add opnd, 1 2797 // idx = ext op 2798 // addr = gep base, idx 2799 // is now: 2800 // promotedOpnd = ext opnd <- no match here 2801 // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls) 2802 // addr = gep base, op <- match 2803 if (MovedAway) 2804 *MovedAway = true; 2805 2806 assert(PromotedOperand && 2807 "TypePromotionHelper should have filtered out those cases"); 2808 2809 ExtAddrMode BackupAddrMode = AddrMode; 2810 unsigned OldSize = AddrModeInsts.size(); 2811 2812 if (!MatchAddr(PromotedOperand, Depth) || 2813 // The total of the new cost is equals to the cost of the created 2814 // instructions. 2815 // The total of the old cost is equals to the cost of the extension plus 2816 // what we have saved in the addressing mode. 2817 !IsPromotionProfitable(CreatedInstsCost, 2818 ExtCost + (AddrModeInsts.size() - OldSize), 2819 PromotedOperand)) { 2820 AddrMode = BackupAddrMode; 2821 AddrModeInsts.resize(OldSize); 2822 DEBUG(dbgs() << "Sign extension does not pay off: rollback\n"); 2823 TPT.rollback(LastKnownGood); 2824 return false; 2825 } 2826 return true; 2827 } 2828 } 2829 return false; 2830} 2831 2832/// MatchAddr - If we can, try to add the value of 'Addr' into the current 2833/// addressing mode. If Addr can't be added to AddrMode this returns false and 2834/// leaves AddrMode unmodified. This assumes that Addr is either a pointer type 2835/// or intptr_t for the target. 2836/// 2837bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) { 2838 // Start a transaction at this point that we will rollback if the matching 2839 // fails. 2840 TypePromotionTransaction::ConstRestorationPt LastKnownGood = 2841 TPT.getRestorationPoint(); 2842 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) { 2843 // Fold in immediates if legal for the target. 2844 AddrMode.BaseOffs += CI->getSExtValue(); 2845 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 2846 return true; 2847 AddrMode.BaseOffs -= CI->getSExtValue(); 2848 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) { 2849 // If this is a global variable, try to fold it into the addressing mode. 2850 if (!AddrMode.BaseGV) { 2851 AddrMode.BaseGV = GV; 2852 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 2853 return true; 2854 AddrMode.BaseGV = nullptr; 2855 } 2856 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) { 2857 ExtAddrMode BackupAddrMode = AddrMode; 2858 unsigned OldSize = AddrModeInsts.size(); 2859 2860 // Check to see if it is possible to fold this operation. 2861 bool MovedAway = false; 2862 if (MatchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) { 2863 // This instruction may have been move away. If so, there is nothing 2864 // to check here. 2865 if (MovedAway) 2866 return true; 2867 // Okay, it's possible to fold this. Check to see if it is actually 2868 // *profitable* to do so. We use a simple cost model to avoid increasing 2869 // register pressure too much. 2870 if (I->hasOneUse() || 2871 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) { 2872 AddrModeInsts.push_back(I); 2873 return true; 2874 } 2875 2876 // It isn't profitable to do this, roll back. 2877 //cerr << "NOT FOLDING: " << *I; 2878 AddrMode = BackupAddrMode; 2879 AddrModeInsts.resize(OldSize); 2880 TPT.rollback(LastKnownGood); 2881 } 2882 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) { 2883 if (MatchOperationAddr(CE, CE->getOpcode(), Depth)) 2884 return true; 2885 TPT.rollback(LastKnownGood); 2886 } else if (isa<ConstantPointerNull>(Addr)) { 2887 // Null pointer gets folded without affecting the addressing mode. 2888 return true; 2889 } 2890 2891 // Worse case, the target should support [reg] addressing modes. :) 2892 if (!AddrMode.HasBaseReg) { 2893 AddrMode.HasBaseReg = true; 2894 AddrMode.BaseReg = Addr; 2895 // Still check for legality in case the target supports [imm] but not [i+r]. 2896 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 2897 return true; 2898 AddrMode.HasBaseReg = false; 2899 AddrMode.BaseReg = nullptr; 2900 } 2901 2902 // If the base register is already taken, see if we can do [r+r]. 2903 if (AddrMode.Scale == 0) { 2904 AddrMode.Scale = 1; 2905 AddrMode.ScaledReg = Addr; 2906 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 2907 return true; 2908 AddrMode.Scale = 0; 2909 AddrMode.ScaledReg = nullptr; 2910 } 2911 // Couldn't match. 2912 TPT.rollback(LastKnownGood); 2913 return false; 2914} 2915 2916/// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified 2917/// inline asm call are due to memory operands. If so, return true, otherwise 2918/// return false. 2919static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal, 2920 const TargetMachine &TM) { 2921 const Function *F = CI->getParent()->getParent(); 2922 const TargetLowering *TLI = TM.getSubtargetImpl(*F)->getTargetLowering(); 2923 const TargetRegisterInfo *TRI = TM.getSubtargetImpl(*F)->getRegisterInfo(); 2924 TargetLowering::AsmOperandInfoVector TargetConstraints = 2925 TLI->ParseConstraints(TRI, ImmutableCallSite(CI)); 2926 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { 2927 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i]; 2928 2929 // Compute the constraint code and ConstraintType to use. 2930 TLI->ComputeConstraintToUse(OpInfo, SDValue()); 2931 2932 // If this asm operand is our Value*, and if it isn't an indirect memory 2933 // operand, we can't fold it! 2934 if (OpInfo.CallOperandVal == OpVal && 2935 (OpInfo.ConstraintType != TargetLowering::C_Memory || 2936 !OpInfo.isIndirect)) 2937 return false; 2938 } 2939 2940 return true; 2941} 2942 2943/// FindAllMemoryUses - Recursively walk all the uses of I until we find a 2944/// memory use. If we find an obviously non-foldable instruction, return true. 2945/// Add the ultimately found memory instructions to MemoryUses. 2946static bool FindAllMemoryUses( 2947 Instruction *I, 2948 SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses, 2949 SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetMachine &TM) { 2950 // If we already considered this instruction, we're done. 2951 if (!ConsideredInsts.insert(I).second) 2952 return false; 2953 2954 // If this is an obviously unfoldable instruction, bail out. 2955 if (!MightBeFoldableInst(I)) 2956 return true; 2957 2958 // Loop over all the uses, recursively processing them. 2959 for (Use &U : I->uses()) { 2960 Instruction *UserI = cast<Instruction>(U.getUser()); 2961 2962 if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) { 2963 MemoryUses.push_back(std::make_pair(LI, U.getOperandNo())); 2964 continue; 2965 } 2966 2967 if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) { 2968 unsigned opNo = U.getOperandNo(); 2969 if (opNo == 0) return true; // Storing addr, not into addr. 2970 MemoryUses.push_back(std::make_pair(SI, opNo)); 2971 continue; 2972 } 2973 2974 if (CallInst *CI = dyn_cast<CallInst>(UserI)) { 2975 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue()); 2976 if (!IA) return true; 2977 2978 // If this is a memory operand, we're cool, otherwise bail out. 2979 if (!IsOperandAMemoryOperand(CI, IA, I, TM)) 2980 return true; 2981 continue; 2982 } 2983 2984 if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TM)) 2985 return true; 2986 } 2987 2988 return false; 2989} 2990 2991/// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at 2992/// the use site that we're folding it into. If so, there is no cost to 2993/// include it in the addressing mode. KnownLive1 and KnownLive2 are two values 2994/// that we know are live at the instruction already. 2995bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1, 2996 Value *KnownLive2) { 2997 // If Val is either of the known-live values, we know it is live! 2998 if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2) 2999 return true; 3000 3001 // All values other than instructions and arguments (e.g. constants) are live. 3002 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true; 3003 3004 // If Val is a constant sized alloca in the entry block, it is live, this is 3005 // true because it is just a reference to the stack/frame pointer, which is 3006 // live for the whole function. 3007 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val)) 3008 if (AI->isStaticAlloca()) 3009 return true; 3010 3011 // Check to see if this value is already used in the memory instruction's 3012 // block. If so, it's already live into the block at the very least, so we 3013 // can reasonably fold it. 3014 return Val->isUsedInBasicBlock(MemoryInst->getParent()); 3015} 3016 3017/// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing 3018/// mode of the machine to fold the specified instruction into a load or store 3019/// that ultimately uses it. However, the specified instruction has multiple 3020/// uses. Given this, it may actually increase register pressure to fold it 3021/// into the load. For example, consider this code: 3022/// 3023/// X = ... 3024/// Y = X+1 3025/// use(Y) -> nonload/store 3026/// Z = Y+1 3027/// load Z 3028/// 3029/// In this case, Y has multiple uses, and can be folded into the load of Z 3030/// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to 3031/// be live at the use(Y) line. If we don't fold Y into load Z, we use one 3032/// fewer register. Since Y can't be folded into "use(Y)" we don't increase the 3033/// number of computations either. 3034/// 3035/// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If 3036/// X was live across 'load Z' for other reasons, we actually *would* want to 3037/// fold the addressing mode in the Z case. This would make Y die earlier. 3038bool AddressingModeMatcher:: 3039IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore, 3040 ExtAddrMode &AMAfter) { 3041 if (IgnoreProfitability) return true; 3042 3043 // AMBefore is the addressing mode before this instruction was folded into it, 3044 // and AMAfter is the addressing mode after the instruction was folded. Get 3045 // the set of registers referenced by AMAfter and subtract out those 3046 // referenced by AMBefore: this is the set of values which folding in this 3047 // address extends the lifetime of. 3048 // 3049 // Note that there are only two potential values being referenced here, 3050 // BaseReg and ScaleReg (global addresses are always available, as are any 3051 // folded immediates). 3052 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg; 3053 3054 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their 3055 // lifetime wasn't extended by adding this instruction. 3056 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg)) 3057 BaseReg = nullptr; 3058 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg)) 3059 ScaledReg = nullptr; 3060 3061 // If folding this instruction (and it's subexprs) didn't extend any live 3062 // ranges, we're ok with it. 3063 if (!BaseReg && !ScaledReg) 3064 return true; 3065 3066 // If all uses of this instruction are ultimately load/store/inlineasm's, 3067 // check to see if their addressing modes will include this instruction. If 3068 // so, we can fold it into all uses, so it doesn't matter if it has multiple 3069 // uses. 3070 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses; 3071 SmallPtrSet<Instruction*, 16> ConsideredInsts; 3072 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TM)) 3073 return false; // Has a non-memory, non-foldable use! 3074 3075 // Now that we know that all uses of this instruction are part of a chain of 3076 // computation involving only operations that could theoretically be folded 3077 // into a memory use, loop over each of these uses and see if they could 3078 // *actually* fold the instruction. 3079 SmallVector<Instruction*, 32> MatchedAddrModeInsts; 3080 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) { 3081 Instruction *User = MemoryUses[i].first; 3082 unsigned OpNo = MemoryUses[i].second; 3083 3084 // Get the access type of this use. If the use isn't a pointer, we don't 3085 // know what it accesses. 3086 Value *Address = User->getOperand(OpNo); 3087 if (!Address->getType()->isPointerTy()) 3088 return false; 3089 Type *AddressAccessTy = Address->getType()->getPointerElementType(); 3090 3091 // Do a match against the root of this address, ignoring profitability. This 3092 // will tell us if the addressing mode for the memory operation will 3093 // *actually* cover the shared instruction. 3094 ExtAddrMode Result; 3095 TypePromotionTransaction::ConstRestorationPt LastKnownGood = 3096 TPT.getRestorationPoint(); 3097 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TM, AddressAccessTy, 3098 MemoryInst, Result, InsertedTruncs, 3099 PromotedInsts, TPT); 3100 Matcher.IgnoreProfitability = true; 3101 bool Success = Matcher.MatchAddr(Address, 0); 3102 (void)Success; assert(Success && "Couldn't select *anything*?"); 3103 3104 // The match was to check the profitability, the changes made are not 3105 // part of the original matcher. Therefore, they should be dropped 3106 // otherwise the original matcher will not present the right state. 3107 TPT.rollback(LastKnownGood); 3108 3109 // If the match didn't cover I, then it won't be shared by it. 3110 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(), 3111 I) == MatchedAddrModeInsts.end()) 3112 return false; 3113 3114 MatchedAddrModeInsts.clear(); 3115 } 3116 3117 return true; 3118} 3119 3120} // end anonymous namespace 3121 3122/// IsNonLocalValue - Return true if the specified values are defined in a 3123/// different basic block than BB. 3124static bool IsNonLocalValue(Value *V, BasicBlock *BB) { 3125 if (Instruction *I = dyn_cast<Instruction>(V)) 3126 return I->getParent() != BB; 3127 return false; 3128} 3129 3130/// OptimizeMemoryInst - Load and Store Instructions often have 3131/// addressing modes that can do significant amounts of computation. As such, 3132/// instruction selection will try to get the load or store to do as much 3133/// computation as possible for the program. The problem is that isel can only 3134/// see within a single block. As such, we sink as much legal addressing mode 3135/// stuff into the block as possible. 3136/// 3137/// This method is used to optimize both load/store and inline asms with memory 3138/// operands. 3139bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr, 3140 Type *AccessTy) { 3141 Value *Repl = Addr; 3142 3143 // Try to collapse single-value PHI nodes. This is necessary to undo 3144 // unprofitable PRE transformations. 3145 SmallVector<Value*, 8> worklist; 3146 SmallPtrSet<Value*, 16> Visited; 3147 worklist.push_back(Addr); 3148 3149 // Use a worklist to iteratively look through PHI nodes, and ensure that 3150 // the addressing mode obtained from the non-PHI roots of the graph 3151 // are equivalent. 3152 Value *Consensus = nullptr; 3153 unsigned NumUsesConsensus = 0; 3154 bool IsNumUsesConsensusValid = false; 3155 SmallVector<Instruction*, 16> AddrModeInsts; 3156 ExtAddrMode AddrMode; 3157 TypePromotionTransaction TPT; 3158 TypePromotionTransaction::ConstRestorationPt LastKnownGood = 3159 TPT.getRestorationPoint(); 3160 while (!worklist.empty()) { 3161 Value *V = worklist.back(); 3162 worklist.pop_back(); 3163 3164 // Break use-def graph loops. 3165 if (!Visited.insert(V).second) { 3166 Consensus = nullptr; 3167 break; 3168 } 3169 3170 // For a PHI node, push all of its incoming values. 3171 if (PHINode *P = dyn_cast<PHINode>(V)) { 3172 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) 3173 worklist.push_back(P->getIncomingValue(i)); 3174 continue; 3175 } 3176 3177 // For non-PHIs, determine the addressing mode being computed. 3178 SmallVector<Instruction*, 16> NewAddrModeInsts; 3179 ExtAddrMode NewAddrMode = AddressingModeMatcher::Match( 3180 V, AccessTy, MemoryInst, NewAddrModeInsts, *TM, InsertedTruncsSet, 3181 PromotedInsts, TPT); 3182 3183 // This check is broken into two cases with very similar code to avoid using 3184 // getNumUses() as much as possible. Some values have a lot of uses, so 3185 // calling getNumUses() unconditionally caused a significant compile-time 3186 // regression. 3187 if (!Consensus) { 3188 Consensus = V; 3189 AddrMode = NewAddrMode; 3190 AddrModeInsts = NewAddrModeInsts; 3191 continue; 3192 } else if (NewAddrMode == AddrMode) { 3193 if (!IsNumUsesConsensusValid) { 3194 NumUsesConsensus = Consensus->getNumUses(); 3195 IsNumUsesConsensusValid = true; 3196 } 3197 3198 // Ensure that the obtained addressing mode is equivalent to that obtained 3199 // for all other roots of the PHI traversal. Also, when choosing one 3200 // such root as representative, select the one with the most uses in order 3201 // to keep the cost modeling heuristics in AddressingModeMatcher 3202 // applicable. 3203 unsigned NumUses = V->getNumUses(); 3204 if (NumUses > NumUsesConsensus) { 3205 Consensus = V; 3206 NumUsesConsensus = NumUses; 3207 AddrModeInsts = NewAddrModeInsts; 3208 } 3209 continue; 3210 } 3211 3212 Consensus = nullptr; 3213 break; 3214 } 3215 3216 // If the addressing mode couldn't be determined, or if multiple different 3217 // ones were determined, bail out now. 3218 if (!Consensus) { 3219 TPT.rollback(LastKnownGood); 3220 return false; 3221 } 3222 TPT.commit(); 3223 3224 // Check to see if any of the instructions supersumed by this addr mode are 3225 // non-local to I's BB. 3226 bool AnyNonLocal = false; 3227 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { 3228 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) { 3229 AnyNonLocal = true; 3230 break; 3231 } 3232 } 3233 3234 // If all the instructions matched are already in this BB, don't do anything. 3235 if (!AnyNonLocal) { 3236 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n"); 3237 return false; 3238 } 3239 3240 // Insert this computation right after this user. Since our caller is 3241 // scanning from the top of the BB to the bottom, reuse of the expr are 3242 // guaranteed to happen later. 3243 IRBuilder<> Builder(MemoryInst); 3244 3245 // Now that we determined the addressing expression we want to use and know 3246 // that we have to sink it into this block. Check to see if we have already 3247 // done this for some other load/store instr in this block. If so, reuse the 3248 // computation. 3249 Value *&SunkAddr = SunkAddrs[Addr]; 3250 if (SunkAddr) { 3251 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for " 3252 << *MemoryInst << "\n"); 3253 if (SunkAddr->getType() != Addr->getType()) 3254 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType()); 3255 } else if (AddrSinkUsingGEPs || 3256 (!AddrSinkUsingGEPs.getNumOccurrences() && TM && 3257 TM->getSubtargetImpl(*MemoryInst->getParent()->getParent()) 3258 ->useAA())) { 3259 // By default, we use the GEP-based method when AA is used later. This 3260 // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities. 3261 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " 3262 << *MemoryInst << "\n"); 3263 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType()); 3264 Value *ResultPtr = nullptr, *ResultIndex = nullptr; 3265 3266 // First, find the pointer. 3267 if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) { 3268 ResultPtr = AddrMode.BaseReg; 3269 AddrMode.BaseReg = nullptr; 3270 } 3271 3272 if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) { 3273 // We can't add more than one pointer together, nor can we scale a 3274 // pointer (both of which seem meaningless). 3275 if (ResultPtr || AddrMode.Scale != 1) 3276 return false; 3277 3278 ResultPtr = AddrMode.ScaledReg; 3279 AddrMode.Scale = 0; 3280 } 3281 3282 if (AddrMode.BaseGV) { 3283 if (ResultPtr) 3284 return false; 3285 3286 ResultPtr = AddrMode.BaseGV; 3287 } 3288 3289 // If the real base value actually came from an inttoptr, then the matcher 3290 // will look through it and provide only the integer value. In that case, 3291 // use it here. 3292 if (!ResultPtr && AddrMode.BaseReg) { 3293 ResultPtr = 3294 Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), "sunkaddr"); 3295 AddrMode.BaseReg = nullptr; 3296 } else if (!ResultPtr && AddrMode.Scale == 1) { 3297 ResultPtr = 3298 Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), "sunkaddr"); 3299 AddrMode.Scale = 0; 3300 } 3301 3302 if (!ResultPtr && 3303 !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) { 3304 SunkAddr = Constant::getNullValue(Addr->getType()); 3305 } else if (!ResultPtr) { 3306 return false; 3307 } else { 3308 Type *I8PtrTy = 3309 Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace()); 3310 Type *I8Ty = Builder.getInt8Ty(); 3311 3312 // Start with the base register. Do this first so that subsequent address 3313 // matching finds it last, which will prevent it from trying to match it 3314 // as the scaled value in case it happens to be a mul. That would be 3315 // problematic if we've sunk a different mul for the scale, because then 3316 // we'd end up sinking both muls. 3317 if (AddrMode.BaseReg) { 3318 Value *V = AddrMode.BaseReg; 3319 if (V->getType() != IntPtrTy) 3320 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr"); 3321 3322 ResultIndex = V; 3323 } 3324 3325 // Add the scale value. 3326 if (AddrMode.Scale) { 3327 Value *V = AddrMode.ScaledReg; 3328 if (V->getType() == IntPtrTy) { 3329 // done. 3330 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 3331 cast<IntegerType>(V->getType())->getBitWidth()) { 3332 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr"); 3333 } else { 3334 // It is only safe to sign extend the BaseReg if we know that the math 3335 // required to create it did not overflow before we extend it. Since 3336 // the original IR value was tossed in favor of a constant back when 3337 // the AddrMode was created we need to bail out gracefully if widths 3338 // do not match instead of extending it. 3339 Instruction *I = dyn_cast_or_null<Instruction>(ResultIndex); 3340 if (I && (ResultIndex != AddrMode.BaseReg)) 3341 I->eraseFromParent(); 3342 return false; 3343 } 3344 3345 if (AddrMode.Scale != 1) 3346 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale), 3347 "sunkaddr"); 3348 if (ResultIndex) 3349 ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr"); 3350 else 3351 ResultIndex = V; 3352 } 3353 3354 // Add in the Base Offset if present. 3355 if (AddrMode.BaseOffs) { 3356 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 3357 if (ResultIndex) { 3358 // We need to add this separately from the scale above to help with 3359 // SDAG consecutive load/store merging. 3360 if (ResultPtr->getType() != I8PtrTy) 3361 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy); 3362 ResultPtr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr"); 3363 } 3364 3365 ResultIndex = V; 3366 } 3367 3368 if (!ResultIndex) { 3369 SunkAddr = ResultPtr; 3370 } else { 3371 if (ResultPtr->getType() != I8PtrTy) 3372 ResultPtr = Builder.CreateBitCast(ResultPtr, I8PtrTy); 3373 SunkAddr = Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr"); 3374 } 3375 3376 if (SunkAddr->getType() != Addr->getType()) 3377 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType()); 3378 } 3379 } else { 3380 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " 3381 << *MemoryInst << "\n"); 3382 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType()); 3383 Value *Result = nullptr; 3384 3385 // Start with the base register. Do this first so that subsequent address 3386 // matching finds it last, which will prevent it from trying to match it 3387 // as the scaled value in case it happens to be a mul. That would be 3388 // problematic if we've sunk a different mul for the scale, because then 3389 // we'd end up sinking both muls. 3390 if (AddrMode.BaseReg) { 3391 Value *V = AddrMode.BaseReg; 3392 if (V->getType()->isPointerTy()) 3393 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr"); 3394 if (V->getType() != IntPtrTy) 3395 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr"); 3396 Result = V; 3397 } 3398 3399 // Add the scale value. 3400 if (AddrMode.Scale) { 3401 Value *V = AddrMode.ScaledReg; 3402 if (V->getType() == IntPtrTy) { 3403 // done. 3404 } else if (V->getType()->isPointerTy()) { 3405 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr"); 3406 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 3407 cast<IntegerType>(V->getType())->getBitWidth()) { 3408 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr"); 3409 } else { 3410 // It is only safe to sign extend the BaseReg if we know that the math 3411 // required to create it did not overflow before we extend it. Since 3412 // the original IR value was tossed in favor of a constant back when 3413 // the AddrMode was created we need to bail out gracefully if widths 3414 // do not match instead of extending it. 3415 Instruction *I = dyn_cast_or_null<Instruction>(Result); 3416 if (I && (Result != AddrMode.BaseReg)) 3417 I->eraseFromParent(); 3418 return false; 3419 } 3420 if (AddrMode.Scale != 1) 3421 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale), 3422 "sunkaddr"); 3423 if (Result) 3424 Result = Builder.CreateAdd(Result, V, "sunkaddr"); 3425 else 3426 Result = V; 3427 } 3428 3429 // Add in the BaseGV if present. 3430 if (AddrMode.BaseGV) { 3431 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr"); 3432 if (Result) 3433 Result = Builder.CreateAdd(Result, V, "sunkaddr"); 3434 else 3435 Result = V; 3436 } 3437 3438 // Add in the Base Offset if present. 3439 if (AddrMode.BaseOffs) { 3440 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 3441 if (Result) 3442 Result = Builder.CreateAdd(Result, V, "sunkaddr"); 3443 else 3444 Result = V; 3445 } 3446 3447 if (!Result) 3448 SunkAddr = Constant::getNullValue(Addr->getType()); 3449 else 3450 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr"); 3451 } 3452 3453 MemoryInst->replaceUsesOfWith(Repl, SunkAddr); 3454 3455 // If we have no uses, recursively delete the value and all dead instructions 3456 // using it. 3457 if (Repl->use_empty()) { 3458 // This can cause recursive deletion, which can invalidate our iterator. 3459 // Use a WeakVH to hold onto it in case this happens. 3460 WeakVH IterHandle(CurInstIterator); 3461 BasicBlock *BB = CurInstIterator->getParent(); 3462 3463 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo); 3464 3465 if (IterHandle != CurInstIterator) { 3466 // If the iterator instruction was recursively deleted, start over at the 3467 // start of the block. 3468 CurInstIterator = BB->begin(); 3469 SunkAddrs.clear(); 3470 } 3471 } 3472 ++NumMemoryInsts; 3473 return true; 3474} 3475 3476/// OptimizeInlineAsmInst - If there are any memory operands, use 3477/// OptimizeMemoryInst to sink their address computing into the block when 3478/// possible / profitable. 3479bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) { 3480 bool MadeChange = false; 3481 3482 const TargetRegisterInfo *TRI = 3483 TM->getSubtargetImpl(*CS->getParent()->getParent())->getRegisterInfo(); 3484 TargetLowering::AsmOperandInfoVector 3485 TargetConstraints = TLI->ParseConstraints(TRI, CS); 3486 unsigned ArgNo = 0; 3487 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { 3488 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i]; 3489 3490 // Compute the constraint code and ConstraintType to use. 3491 TLI->ComputeConstraintToUse(OpInfo, SDValue()); 3492 3493 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 3494 OpInfo.isIndirect) { 3495 Value *OpVal = CS->getArgOperand(ArgNo++); 3496 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType()); 3497 } else if (OpInfo.Type == InlineAsm::isInput) 3498 ArgNo++; 3499 } 3500 3501 return MadeChange; 3502} 3503 3504/// \brief Check if all the uses of \p Inst are equivalent (or free) zero or 3505/// sign extensions. 3506static bool hasSameExtUse(Instruction *Inst, const TargetLowering &TLI) { 3507 assert(!Inst->use_empty() && "Input must have at least one use"); 3508 const Instruction *FirstUser = cast<Instruction>(*Inst->user_begin()); 3509 bool IsSExt = isa<SExtInst>(FirstUser); 3510 Type *ExtTy = FirstUser->getType(); 3511 for (const User *U : Inst->users()) { 3512 const Instruction *UI = cast<Instruction>(U); 3513 if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI))) 3514 return false; 3515 Type *CurTy = UI->getType(); 3516 // Same input and output types: Same instruction after CSE. 3517 if (CurTy == ExtTy) 3518 continue; 3519 3520 // If IsSExt is true, we are in this situation: 3521 // a = Inst 3522 // b = sext ty1 a to ty2 3523 // c = sext ty1 a to ty3 3524 // Assuming ty2 is shorter than ty3, this could be turned into: 3525 // a = Inst 3526 // b = sext ty1 a to ty2 3527 // c = sext ty2 b to ty3 3528 // However, the last sext is not free. 3529 if (IsSExt) 3530 return false; 3531 3532 // This is a ZExt, maybe this is free to extend from one type to another. 3533 // In that case, we would not account for a different use. 3534 Type *NarrowTy; 3535 Type *LargeTy; 3536 if (ExtTy->getScalarType()->getIntegerBitWidth() > 3537 CurTy->getScalarType()->getIntegerBitWidth()) { 3538 NarrowTy = CurTy; 3539 LargeTy = ExtTy; 3540 } else { 3541 NarrowTy = ExtTy; 3542 LargeTy = CurTy; 3543 } 3544 3545 if (!TLI.isZExtFree(NarrowTy, LargeTy)) 3546 return false; 3547 } 3548 // All uses are the same or can be derived from one another for free. 3549 return true; 3550} 3551 3552/// \brief Try to form ExtLd by promoting \p Exts until they reach a 3553/// load instruction. 3554/// If an ext(load) can be formed, it is returned via \p LI for the load 3555/// and \p Inst for the extension. 3556/// Otherwise LI == nullptr and Inst == nullptr. 3557/// When some promotion happened, \p TPT contains the proper state to 3558/// revert them. 3559/// 3560/// \return true when promoting was necessary to expose the ext(load) 3561/// opportunity, false otherwise. 3562/// 3563/// Example: 3564/// \code 3565/// %ld = load i32* %addr 3566/// %add = add nuw i32 %ld, 4 3567/// %zext = zext i32 %add to i64 3568/// \endcode 3569/// => 3570/// \code 3571/// %ld = load i32* %addr 3572/// %zext = zext i32 %ld to i64 3573/// %add = add nuw i64 %zext, 4 3574/// \encode 3575/// Thanks to the promotion, we can match zext(load i32*) to i64. 3576bool CodeGenPrepare::ExtLdPromotion(TypePromotionTransaction &TPT, 3577 LoadInst *&LI, Instruction *&Inst, 3578 const SmallVectorImpl<Instruction *> &Exts, 3579 unsigned CreatedInstsCost = 0) { 3580 // Iterate over all the extensions to see if one form an ext(load). 3581 for (auto I : Exts) { 3582 // Check if we directly have ext(load). 3583 if ((LI = dyn_cast<LoadInst>(I->getOperand(0)))) { 3584 Inst = I; 3585 // No promotion happened here. 3586 return false; 3587 } 3588 // Check whether or not we want to do any promotion. 3589 if (!TLI || !TLI->enableExtLdPromotion() || DisableExtLdPromotion) 3590 continue; 3591 // Get the action to perform the promotion. 3592 TypePromotionHelper::Action TPH = TypePromotionHelper::getAction( 3593 I, InsertedTruncsSet, *TLI, PromotedInsts); 3594 // Check if we can promote. 3595 if (!TPH) 3596 continue; 3597 // Save the current state. 3598 TypePromotionTransaction::ConstRestorationPt LastKnownGood = 3599 TPT.getRestorationPoint(); 3600 SmallVector<Instruction *, 4> NewExts; 3601 unsigned NewCreatedInstsCost = 0; 3602 unsigned ExtCost = !TLI->isExtFree(I); 3603 // Promote. 3604 Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost, 3605 &NewExts, nullptr, *TLI); 3606 assert(PromotedVal && 3607 "TypePromotionHelper should have filtered out those cases"); 3608 3609 // We would be able to merge only one extension in a load. 3610 // Therefore, if we have more than 1 new extension we heuristically 3611 // cut this search path, because it means we degrade the code quality. 3612 // With exactly 2, the transformation is neutral, because we will merge 3613 // one extension but leave one. However, we optimistically keep going, 3614 // because the new extension may be removed too. 3615 long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost; 3616 TotalCreatedInstsCost -= ExtCost; 3617 if (!StressExtLdPromotion && 3618 (TotalCreatedInstsCost > 1 || 3619 !isPromotedInstructionLegal(*TLI, PromotedVal))) { 3620 // The promotion is not profitable, rollback to the previous state. 3621 TPT.rollback(LastKnownGood); 3622 continue; 3623 } 3624 // The promotion is profitable. 3625 // Check if it exposes an ext(load). 3626 (void)ExtLdPromotion(TPT, LI, Inst, NewExts, TotalCreatedInstsCost); 3627 if (LI && (StressExtLdPromotion || NewCreatedInstsCost <= ExtCost || 3628 // If we have created a new extension, i.e., now we have two 3629 // extensions. We must make sure one of them is merged with 3630 // the load, otherwise we may degrade the code quality. 3631 (LI->hasOneUse() || hasSameExtUse(LI, *TLI)))) 3632 // Promotion happened. 3633 return true; 3634 // If this does not help to expose an ext(load) then, rollback. 3635 TPT.rollback(LastKnownGood); 3636 } 3637 // None of the extension can form an ext(load). 3638 LI = nullptr; 3639 Inst = nullptr; 3640 return false; 3641} 3642 3643/// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same 3644/// basic block as the load, unless conditions are unfavorable. This allows 3645/// SelectionDAG to fold the extend into the load. 3646/// \p I[in/out] the extension may be modified during the process if some 3647/// promotions apply. 3648/// 3649bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *&I) { 3650 // Try to promote a chain of computation if it allows to form 3651 // an extended load. 3652 TypePromotionTransaction TPT; 3653 TypePromotionTransaction::ConstRestorationPt LastKnownGood = 3654 TPT.getRestorationPoint(); 3655 SmallVector<Instruction *, 1> Exts; 3656 Exts.push_back(I); 3657 // Look for a load being extended. 3658 LoadInst *LI = nullptr; 3659 Instruction *OldExt = I; 3660 bool HasPromoted = ExtLdPromotion(TPT, LI, I, Exts); 3661 if (!LI || !I) { 3662 assert(!HasPromoted && !LI && "If we did not match any load instruction " 3663 "the code must remain the same"); 3664 I = OldExt; 3665 return false; 3666 } 3667 3668 // If they're already in the same block, there's nothing to do. 3669 // Make the cheap checks first if we did not promote. 3670 // If we promoted, we need to check if it is indeed profitable. 3671 if (!HasPromoted && LI->getParent() == I->getParent()) 3672 return false; 3673 3674 EVT VT = TLI->getValueType(I->getType()); 3675 EVT LoadVT = TLI->getValueType(LI->getType()); 3676 3677 // If the load has other users and the truncate is not free, this probably 3678 // isn't worthwhile. 3679 if (!LI->hasOneUse() && TLI && 3680 (TLI->isTypeLegal(LoadVT) || !TLI->isTypeLegal(VT)) && 3681 !TLI->isTruncateFree(I->getType(), LI->getType())) { 3682 I = OldExt; 3683 TPT.rollback(LastKnownGood); 3684 return false; 3685 } 3686 3687 // Check whether the target supports casts folded into loads. 3688 unsigned LType; 3689 if (isa<ZExtInst>(I)) 3690 LType = ISD::ZEXTLOAD; 3691 else { 3692 assert(isa<SExtInst>(I) && "Unexpected ext type!"); 3693 LType = ISD::SEXTLOAD; 3694 } 3695 if (TLI && !TLI->isLoadExtLegal(LType, VT, LoadVT)) { 3696 I = OldExt; 3697 TPT.rollback(LastKnownGood); 3698 return false; 3699 } 3700 3701 // Move the extend into the same block as the load, so that SelectionDAG 3702 // can fold it. 3703 TPT.commit(); 3704 I->removeFromParent(); 3705 I->insertAfter(LI); 3706 ++NumExtsMoved; 3707 return true; 3708} 3709 3710bool CodeGenPrepare::OptimizeExtUses(Instruction *I) { 3711 BasicBlock *DefBB = I->getParent(); 3712 3713 // If the result of a {s|z}ext and its source are both live out, rewrite all 3714 // other uses of the source with result of extension. 3715 Value *Src = I->getOperand(0); 3716 if (Src->hasOneUse()) 3717 return false; 3718 3719 // Only do this xform if truncating is free. 3720 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType())) 3721 return false; 3722 3723 // Only safe to perform the optimization if the source is also defined in 3724 // this block. 3725 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) 3726 return false; 3727 3728 bool DefIsLiveOut = false; 3729 for (User *U : I->users()) { 3730 Instruction *UI = cast<Instruction>(U); 3731 3732 // Figure out which BB this ext is used in. 3733 BasicBlock *UserBB = UI->getParent(); 3734 if (UserBB == DefBB) continue; 3735 DefIsLiveOut = true; 3736 break; 3737 } 3738 if (!DefIsLiveOut) 3739 return false; 3740 3741 // Make sure none of the uses are PHI nodes. 3742 for (User *U : Src->users()) { 3743 Instruction *UI = cast<Instruction>(U); 3744 BasicBlock *UserBB = UI->getParent(); 3745 if (UserBB == DefBB) continue; 3746 // Be conservative. We don't want this xform to end up introducing 3747 // reloads just before load / store instructions. 3748 if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI)) 3749 return false; 3750 } 3751 3752 // InsertedTruncs - Only insert one trunc in each block once. 3753 DenseMap<BasicBlock*, Instruction*> InsertedTruncs; 3754 3755 bool MadeChange = false; 3756 for (Use &U : Src->uses()) { 3757 Instruction *User = cast<Instruction>(U.getUser()); 3758 3759 // Figure out which BB this ext is used in. 3760 BasicBlock *UserBB = User->getParent(); 3761 if (UserBB == DefBB) continue; 3762 3763 // Both src and def are live in this block. Rewrite the use. 3764 Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; 3765 3766 if (!InsertedTrunc) { 3767 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); 3768 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt); 3769 InsertedTruncsSet.insert(InsertedTrunc); 3770 } 3771 3772 // Replace a use of the {s|z}ext source with a use of the result. 3773 U = InsertedTrunc; 3774 ++NumExtUses; 3775 MadeChange = true; 3776 } 3777 3778 return MadeChange; 3779} 3780 3781/// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be 3782/// turned into an explicit branch. 3783static bool isFormingBranchFromSelectProfitable(SelectInst *SI) { 3784 // FIXME: This should use the same heuristics as IfConversion to determine 3785 // whether a select is better represented as a branch. This requires that 3786 // branch probability metadata is preserved for the select, which is not the 3787 // case currently. 3788 3789 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition()); 3790 3791 // If the branch is predicted right, an out of order CPU can avoid blocking on 3792 // the compare. Emit cmovs on compares with a memory operand as branches to 3793 // avoid stalls on the load from memory. If the compare has more than one use 3794 // there's probably another cmov or setcc around so it's not worth emitting a 3795 // branch. 3796 if (!Cmp) 3797 return false; 3798 3799 Value *CmpOp0 = Cmp->getOperand(0); 3800 Value *CmpOp1 = Cmp->getOperand(1); 3801 3802 // We check that the memory operand has one use to avoid uses of the loaded 3803 // value directly after the compare, making branches unprofitable. 3804 return Cmp->hasOneUse() && 3805 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) || 3806 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse())); 3807} 3808 3809 3810/// If we have a SelectInst that will likely profit from branch prediction, 3811/// turn it into a branch. 3812bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) { 3813 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1); 3814 3815 // Can we convert the 'select' to CF ? 3816 if (DisableSelectToBranch || OptSize || !TLI || VectorCond) 3817 return false; 3818 3819 TargetLowering::SelectSupportKind SelectKind; 3820 if (VectorCond) 3821 SelectKind = TargetLowering::VectorMaskSelect; 3822 else if (SI->getType()->isVectorTy()) 3823 SelectKind = TargetLowering::ScalarCondVectorVal; 3824 else 3825 SelectKind = TargetLowering::ScalarValSelect; 3826 3827 // Do we have efficient codegen support for this kind of 'selects' ? 3828 if (TLI->isSelectSupported(SelectKind)) { 3829 // We have efficient codegen support for the select instruction. 3830 // Check if it is profitable to keep this 'select'. 3831 if (!TLI->isPredictableSelectExpensive() || 3832 !isFormingBranchFromSelectProfitable(SI)) 3833 return false; 3834 } 3835 3836 ModifiedDT = true; 3837 3838 // First, we split the block containing the select into 2 blocks. 3839 BasicBlock *StartBlock = SI->getParent(); 3840 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI)); 3841 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end"); 3842 3843 // Create a new block serving as the landing pad for the branch. 3844 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid", 3845 NextBlock->getParent(), NextBlock); 3846 3847 // Move the unconditional branch from the block with the select in it into our 3848 // landing pad block. 3849 StartBlock->getTerminator()->eraseFromParent(); 3850 BranchInst::Create(NextBlock, SmallBlock); 3851 3852 // Insert the real conditional branch based on the original condition. 3853 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI); 3854 3855 // The select itself is replaced with a PHI Node. 3856 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin()); 3857 PN->takeName(SI); 3858 PN->addIncoming(SI->getTrueValue(), StartBlock); 3859 PN->addIncoming(SI->getFalseValue(), SmallBlock); 3860 SI->replaceAllUsesWith(PN); 3861 SI->eraseFromParent(); 3862 3863 // Instruct OptimizeBlock to skip to the next block. 3864 CurInstIterator = StartBlock->end(); 3865 ++NumSelectsExpanded; 3866 return true; 3867} 3868 3869static bool isBroadcastShuffle(ShuffleVectorInst *SVI) { 3870 SmallVector<int, 16> Mask(SVI->getShuffleMask()); 3871 int SplatElem = -1; 3872 for (unsigned i = 0; i < Mask.size(); ++i) { 3873 if (SplatElem != -1 && Mask[i] != -1 && Mask[i] != SplatElem) 3874 return false; 3875 SplatElem = Mask[i]; 3876 } 3877 3878 return true; 3879} 3880 3881/// Some targets have expensive vector shifts if the lanes aren't all the same 3882/// (e.g. x86 only introduced "vpsllvd" and friends with AVX2). In these cases 3883/// it's often worth sinking a shufflevector splat down to its use so that 3884/// codegen can spot all lanes are identical. 3885bool CodeGenPrepare::OptimizeShuffleVectorInst(ShuffleVectorInst *SVI) { 3886 BasicBlock *DefBB = SVI->getParent(); 3887 3888 // Only do this xform if variable vector shifts are particularly expensive. 3889 if (!TLI || !TLI->isVectorShiftByScalarCheap(SVI->getType())) 3890 return false; 3891 3892 // We only expect better codegen by sinking a shuffle if we can recognise a 3893 // constant splat. 3894 if (!isBroadcastShuffle(SVI)) 3895 return false; 3896 3897 // InsertedShuffles - Only insert a shuffle in each block once. 3898 DenseMap<BasicBlock*, Instruction*> InsertedShuffles; 3899 3900 bool MadeChange = false; 3901 for (User *U : SVI->users()) { 3902 Instruction *UI = cast<Instruction>(U); 3903 3904 // Figure out which BB this ext is used in. 3905 BasicBlock *UserBB = UI->getParent(); 3906 if (UserBB == DefBB) continue; 3907 3908 // For now only apply this when the splat is used by a shift instruction. 3909 if (!UI->isShift()) continue; 3910 3911 // Everything checks out, sink the shuffle if the user's block doesn't 3912 // already have a copy. 3913 Instruction *&InsertedShuffle = InsertedShuffles[UserBB]; 3914 3915 if (!InsertedShuffle) { 3916 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); 3917 InsertedShuffle = new ShuffleVectorInst(SVI->getOperand(0), 3918 SVI->getOperand(1), 3919 SVI->getOperand(2), "", InsertPt); 3920 } 3921 3922 UI->replaceUsesOfWith(SVI, InsertedShuffle); 3923 MadeChange = true; 3924 } 3925 3926 // If we removed all uses, nuke the shuffle. 3927 if (SVI->use_empty()) { 3928 SVI->eraseFromParent(); 3929 MadeChange = true; 3930 } 3931 3932 return MadeChange; 3933} 3934 3935namespace { 3936/// \brief Helper class to promote a scalar operation to a vector one. 3937/// This class is used to move downward extractelement transition. 3938/// E.g., 3939/// a = vector_op <2 x i32> 3940/// b = extractelement <2 x i32> a, i32 0 3941/// c = scalar_op b 3942/// store c 3943/// 3944/// => 3945/// a = vector_op <2 x i32> 3946/// c = vector_op a (equivalent to scalar_op on the related lane) 3947/// * d = extractelement <2 x i32> c, i32 0 3948/// * store d 3949/// Assuming both extractelement and store can be combine, we get rid of the 3950/// transition. 3951class VectorPromoteHelper { 3952 /// Used to perform some checks on the legality of vector operations. 3953 const TargetLowering &TLI; 3954 3955 /// Used to estimated the cost of the promoted chain. 3956 const TargetTransformInfo &TTI; 3957 3958 /// The transition being moved downwards. 3959 Instruction *Transition; 3960 /// The sequence of instructions to be promoted. 3961 SmallVector<Instruction *, 4> InstsToBePromoted; 3962 /// Cost of combining a store and an extract. 3963 unsigned StoreExtractCombineCost; 3964 /// Instruction that will be combined with the transition. 3965 Instruction *CombineInst; 3966 3967 /// \brief The instruction that represents the current end of the transition. 3968 /// Since we are faking the promotion until we reach the end of the chain 3969 /// of computation, we need a way to get the current end of the transition. 3970 Instruction *getEndOfTransition() const { 3971 if (InstsToBePromoted.empty()) 3972 return Transition; 3973 return InstsToBePromoted.back(); 3974 } 3975 3976 /// \brief Return the index of the original value in the transition. 3977 /// E.g., for "extractelement <2 x i32> c, i32 1" the original value, 3978 /// c, is at index 0. 3979 unsigned getTransitionOriginalValueIdx() const { 3980 assert(isa<ExtractElementInst>(Transition) && 3981 "Other kind of transitions are not supported yet"); 3982 return 0; 3983 } 3984 3985 /// \brief Return the index of the index in the transition. 3986 /// E.g., for "extractelement <2 x i32> c, i32 0" the index 3987 /// is at index 1. 3988 unsigned getTransitionIdx() const { 3989 assert(isa<ExtractElementInst>(Transition) && 3990 "Other kind of transitions are not supported yet"); 3991 return 1; 3992 } 3993 3994 /// \brief Get the type of the transition. 3995 /// This is the type of the original value. 3996 /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the 3997 /// transition is <2 x i32>. 3998 Type *getTransitionType() const { 3999 return Transition->getOperand(getTransitionOriginalValueIdx())->getType(); 4000 } 4001 4002 /// \brief Promote \p ToBePromoted by moving \p Def downward through. 4003 /// I.e., we have the following sequence: 4004 /// Def = Transition <ty1> a to <ty2> 4005 /// b = ToBePromoted <ty2> Def, ... 4006 /// => 4007 /// b = ToBePromoted <ty1> a, ... 4008 /// Def = Transition <ty1> ToBePromoted to <ty2> 4009 void promoteImpl(Instruction *ToBePromoted); 4010 4011 /// \brief Check whether or not it is profitable to promote all the 4012 /// instructions enqueued to be promoted. 4013 bool isProfitableToPromote() { 4014 Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx()); 4015 unsigned Index = isa<ConstantInt>(ValIdx) 4016 ? cast<ConstantInt>(ValIdx)->getZExtValue() 4017 : -1; 4018 Type *PromotedType = getTransitionType(); 4019 4020 StoreInst *ST = cast<StoreInst>(CombineInst); 4021 unsigned AS = ST->getPointerAddressSpace(); 4022 unsigned Align = ST->getAlignment(); 4023 // Check if this store is supported. 4024 if (!TLI.allowsMisalignedMemoryAccesses( 4025 TLI.getValueType(ST->getValueOperand()->getType()), AS, Align)) { 4026 // If this is not supported, there is no way we can combine 4027 // the extract with the store. 4028 return false; 4029 } 4030 4031 // The scalar chain of computation has to pay for the transition 4032 // scalar to vector. 4033 // The vector chain has to account for the combining cost. 4034 uint64_t ScalarCost = 4035 TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index); 4036 uint64_t VectorCost = StoreExtractCombineCost; 4037 for (const auto &Inst : InstsToBePromoted) { 4038 // Compute the cost. 4039 // By construction, all instructions being promoted are arithmetic ones. 4040 // Moreover, one argument is a constant that can be viewed as a splat 4041 // constant. 4042 Value *Arg0 = Inst->getOperand(0); 4043 bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) || 4044 isa<ConstantFP>(Arg0); 4045 TargetTransformInfo::OperandValueKind Arg0OVK = 4046 IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue 4047 : TargetTransformInfo::OK_AnyValue; 4048 TargetTransformInfo::OperandValueKind Arg1OVK = 4049 !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue 4050 : TargetTransformInfo::OK_AnyValue; 4051 ScalarCost += TTI.getArithmeticInstrCost( 4052 Inst->getOpcode(), Inst->getType(), Arg0OVK, Arg1OVK); 4053 VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType, 4054 Arg0OVK, Arg1OVK); 4055 } 4056 DEBUG(dbgs() << "Estimated cost of computation to be promoted:\nScalar: " 4057 << ScalarCost << "\nVector: " << VectorCost << '\n'); 4058 return ScalarCost > VectorCost; 4059 } 4060 4061 /// \brief Generate a constant vector with \p Val with the same 4062 /// number of elements as the transition. 4063 /// \p UseSplat defines whether or not \p Val should be replicated 4064 /// accross the whole vector. 4065 /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>, 4066 /// otherwise we generate a vector with as many undef as possible: 4067 /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only 4068 /// used at the index of the extract. 4069 Value *getConstantVector(Constant *Val, bool UseSplat) const { 4070 unsigned ExtractIdx = UINT_MAX; 4071 if (!UseSplat) { 4072 // If we cannot determine where the constant must be, we have to 4073 // use a splat constant. 4074 Value *ValExtractIdx = Transition->getOperand(getTransitionIdx()); 4075 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx)) 4076 ExtractIdx = CstVal->getSExtValue(); 4077 else 4078 UseSplat = true; 4079 } 4080 4081 unsigned End = getTransitionType()->getVectorNumElements(); 4082 if (UseSplat) 4083 return ConstantVector::getSplat(End, Val); 4084 4085 SmallVector<Constant *, 4> ConstVec; 4086 UndefValue *UndefVal = UndefValue::get(Val->getType()); 4087 for (unsigned Idx = 0; Idx != End; ++Idx) { 4088 if (Idx == ExtractIdx) 4089 ConstVec.push_back(Val); 4090 else 4091 ConstVec.push_back(UndefVal); 4092 } 4093 return ConstantVector::get(ConstVec); 4094 } 4095 4096 /// \brief Check if promoting to a vector type an operand at \p OperandIdx 4097 /// in \p Use can trigger undefined behavior. 4098 static bool canCauseUndefinedBehavior(const Instruction *Use, 4099 unsigned OperandIdx) { 4100 // This is not safe to introduce undef when the operand is on 4101 // the right hand side of a division-like instruction. 4102 if (OperandIdx != 1) 4103 return false; 4104 switch (Use->getOpcode()) { 4105 default: 4106 return false; 4107 case Instruction::SDiv: 4108 case Instruction::UDiv: 4109 case Instruction::SRem: 4110 case Instruction::URem: 4111 return true; 4112 case Instruction::FDiv: 4113 case Instruction::FRem: 4114 return !Use->hasNoNaNs(); 4115 } 4116 llvm_unreachable(nullptr); 4117 } 4118 4119public: 4120 VectorPromoteHelper(const TargetLowering &TLI, const TargetTransformInfo &TTI, 4121 Instruction *Transition, unsigned CombineCost) 4122 : TLI(TLI), TTI(TTI), Transition(Transition), 4123 StoreExtractCombineCost(CombineCost), CombineInst(nullptr) { 4124 assert(Transition && "Do not know how to promote null"); 4125 } 4126 4127 /// \brief Check if we can promote \p ToBePromoted to \p Type. 4128 bool canPromote(const Instruction *ToBePromoted) const { 4129 // We could support CastInst too. 4130 return isa<BinaryOperator>(ToBePromoted); 4131 } 4132 4133 /// \brief Check if it is profitable to promote \p ToBePromoted 4134 /// by moving downward the transition through. 4135 bool shouldPromote(const Instruction *ToBePromoted) const { 4136 // Promote only if all the operands can be statically expanded. 4137 // Indeed, we do not want to introduce any new kind of transitions. 4138 for (const Use &U : ToBePromoted->operands()) { 4139 const Value *Val = U.get(); 4140 if (Val == getEndOfTransition()) { 4141 // If the use is a division and the transition is on the rhs, 4142 // we cannot promote the operation, otherwise we may create a 4143 // division by zero. 4144 if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo())) 4145 return false; 4146 continue; 4147 } 4148 if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) && 4149 !isa<ConstantFP>(Val)) 4150 return false; 4151 } 4152 // Check that the resulting operation is legal. 4153 int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode()); 4154 if (!ISDOpcode) 4155 return false; 4156 return StressStoreExtract || 4157 TLI.isOperationLegalOrCustom( 4158 ISDOpcode, TLI.getValueType(getTransitionType(), true)); 4159 } 4160 4161 /// \brief Check whether or not \p Use can be combined 4162 /// with the transition. 4163 /// I.e., is it possible to do Use(Transition) => AnotherUse? 4164 bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); } 4165 4166 /// \brief Record \p ToBePromoted as part of the chain to be promoted. 4167 void enqueueForPromotion(Instruction *ToBePromoted) { 4168 InstsToBePromoted.push_back(ToBePromoted); 4169 } 4170 4171 /// \brief Set the instruction that will be combined with the transition. 4172 void recordCombineInstruction(Instruction *ToBeCombined) { 4173 assert(canCombine(ToBeCombined) && "Unsupported instruction to combine"); 4174 CombineInst = ToBeCombined; 4175 } 4176 4177 /// \brief Promote all the instructions enqueued for promotion if it is 4178 /// is profitable. 4179 /// \return True if the promotion happened, false otherwise. 4180 bool promote() { 4181 // Check if there is something to promote. 4182 // Right now, if we do not have anything to combine with, 4183 // we assume the promotion is not profitable. 4184 if (InstsToBePromoted.empty() || !CombineInst) 4185 return false; 4186 4187 // Check cost. 4188 if (!StressStoreExtract && !isProfitableToPromote()) 4189 return false; 4190 4191 // Promote. 4192 for (auto &ToBePromoted : InstsToBePromoted) 4193 promoteImpl(ToBePromoted); 4194 InstsToBePromoted.clear(); 4195 return true; 4196 } 4197}; 4198} // End of anonymous namespace. 4199 4200void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) { 4201 // At this point, we know that all the operands of ToBePromoted but Def 4202 // can be statically promoted. 4203 // For Def, we need to use its parameter in ToBePromoted: 4204 // b = ToBePromoted ty1 a 4205 // Def = Transition ty1 b to ty2 4206 // Move the transition down. 4207 // 1. Replace all uses of the promoted operation by the transition. 4208 // = ... b => = ... Def. 4209 assert(ToBePromoted->getType() == Transition->getType() && 4210 "The type of the result of the transition does not match " 4211 "the final type"); 4212 ToBePromoted->replaceAllUsesWith(Transition); 4213 // 2. Update the type of the uses. 4214 // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def. 4215 Type *TransitionTy = getTransitionType(); 4216 ToBePromoted->mutateType(TransitionTy); 4217 // 3. Update all the operands of the promoted operation with promoted 4218 // operands. 4219 // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a. 4220 for (Use &U : ToBePromoted->operands()) { 4221 Value *Val = U.get(); 4222 Value *NewVal = nullptr; 4223 if (Val == Transition) 4224 NewVal = Transition->getOperand(getTransitionOriginalValueIdx()); 4225 else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) || 4226 isa<ConstantFP>(Val)) { 4227 // Use a splat constant if it is not safe to use undef. 4228 NewVal = getConstantVector( 4229 cast<Constant>(Val), 4230 isa<UndefValue>(Val) || 4231 canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo())); 4232 } else 4233 llvm_unreachable("Did you modified shouldPromote and forgot to update " 4234 "this?"); 4235 ToBePromoted->setOperand(U.getOperandNo(), NewVal); 4236 } 4237 Transition->removeFromParent(); 4238 Transition->insertAfter(ToBePromoted); 4239 Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted); 4240} 4241 4242/// Some targets can do store(extractelement) with one instruction. 4243/// Try to push the extractelement towards the stores when the target 4244/// has this feature and this is profitable. 4245bool CodeGenPrepare::OptimizeExtractElementInst(Instruction *Inst) { 4246 unsigned CombineCost = UINT_MAX; 4247 if (DisableStoreExtract || !TLI || 4248 (!StressStoreExtract && 4249 !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(), 4250 Inst->getOperand(1), CombineCost))) 4251 return false; 4252 4253 // At this point we know that Inst is a vector to scalar transition. 4254 // Try to move it down the def-use chain, until: 4255 // - We can combine the transition with its single use 4256 // => we got rid of the transition. 4257 // - We escape the current basic block 4258 // => we would need to check that we are moving it at a cheaper place and 4259 // we do not do that for now. 4260 BasicBlock *Parent = Inst->getParent(); 4261 DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n'); 4262 VectorPromoteHelper VPH(*TLI, *TTI, Inst, CombineCost); 4263 // If the transition has more than one use, assume this is not going to be 4264 // beneficial. 4265 while (Inst->hasOneUse()) { 4266 Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin()); 4267 DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n'); 4268 4269 if (ToBePromoted->getParent() != Parent) { 4270 DEBUG(dbgs() << "Instruction to promote is in a different block (" 4271 << ToBePromoted->getParent()->getName() 4272 << ") than the transition (" << Parent->getName() << ").\n"); 4273 return false; 4274 } 4275 4276 if (VPH.canCombine(ToBePromoted)) { 4277 DEBUG(dbgs() << "Assume " << *Inst << '\n' 4278 << "will be combined with: " << *ToBePromoted << '\n'); 4279 VPH.recordCombineInstruction(ToBePromoted); 4280 bool Changed = VPH.promote(); 4281 NumStoreExtractExposed += Changed; 4282 return Changed; 4283 } 4284 4285 DEBUG(dbgs() << "Try promoting.\n"); 4286 if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted)) 4287 return false; 4288 4289 DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n"); 4290 4291 VPH.enqueueForPromotion(ToBePromoted); 4292 Inst = ToBePromoted; 4293 } 4294 return false; 4295} 4296 4297bool CodeGenPrepare::OptimizeInst(Instruction *I, bool& ModifiedDT) { 4298 if (PHINode *P = dyn_cast<PHINode>(I)) { 4299 // It is possible for very late stage optimizations (such as SimplifyCFG) 4300 // to introduce PHI nodes too late to be cleaned up. If we detect such a 4301 // trivial PHI, go ahead and zap it here. 4302 const DataLayout &DL = I->getModule()->getDataLayout(); 4303 if (Value *V = SimplifyInstruction(P, DL, TLInfo, nullptr)) { 4304 P->replaceAllUsesWith(V); 4305 P->eraseFromParent(); 4306 ++NumPHIsElim; 4307 return true; 4308 } 4309 return false; 4310 } 4311 4312 if (CastInst *CI = dyn_cast<CastInst>(I)) { 4313 // If the source of the cast is a constant, then this should have 4314 // already been constant folded. The only reason NOT to constant fold 4315 // it is if something (e.g. LSR) was careful to place the constant 4316 // evaluation in a block other than then one that uses it (e.g. to hoist 4317 // the address of globals out of a loop). If this is the case, we don't 4318 // want to forward-subst the cast. 4319 if (isa<Constant>(CI->getOperand(0))) 4320 return false; 4321 4322 if (TLI && OptimizeNoopCopyExpression(CI, *TLI)) 4323 return true; 4324 4325 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) { 4326 /// Sink a zext or sext into its user blocks if the target type doesn't 4327 /// fit in one register 4328 if (TLI && TLI->getTypeAction(CI->getContext(), 4329 TLI->getValueType(CI->getType())) == 4330 TargetLowering::TypeExpandInteger) { 4331 return SinkCast(CI); 4332 } else { 4333 bool MadeChange = MoveExtToFormExtLoad(I); 4334 return MadeChange | OptimizeExtUses(I); 4335 } 4336 } 4337 return false; 4338 } 4339 4340 if (CmpInst *CI = dyn_cast<CmpInst>(I)) 4341 if (!TLI || !TLI->hasMultipleConditionRegisters()) 4342 return OptimizeCmpExpression(CI); 4343 4344 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 4345 if (TLI) 4346 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType()); 4347 return false; 4348 } 4349 4350 if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 4351 if (TLI) 4352 return OptimizeMemoryInst(I, SI->getOperand(1), 4353 SI->getOperand(0)->getType()); 4354 return false; 4355 } 4356 4357 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I); 4358 4359 if (BinOp && (BinOp->getOpcode() == Instruction::AShr || 4360 BinOp->getOpcode() == Instruction::LShr)) { 4361 ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1)); 4362 if (TLI && CI && TLI->hasExtractBitsInsn()) 4363 return OptimizeExtractBits(BinOp, CI, *TLI); 4364 4365 return false; 4366 } 4367 4368 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 4369 if (GEPI->hasAllZeroIndices()) { 4370 /// The GEP operand must be a pointer, so must its result -> BitCast 4371 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 4372 GEPI->getName(), GEPI); 4373 GEPI->replaceAllUsesWith(NC); 4374 GEPI->eraseFromParent(); 4375 ++NumGEPsElim; 4376 OptimizeInst(NC, ModifiedDT); 4377 return true; 4378 } 4379 return false; 4380 } 4381 4382 if (CallInst *CI = dyn_cast<CallInst>(I)) 4383 return OptimizeCallInst(CI, ModifiedDT); 4384 4385 if (SelectInst *SI = dyn_cast<SelectInst>(I)) 4386 return OptimizeSelectInst(SI); 4387 4388 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) 4389 return OptimizeShuffleVectorInst(SVI); 4390 4391 if (isa<ExtractElementInst>(I)) 4392 return OptimizeExtractElementInst(I); 4393 4394 return false; 4395} 4396 4397// In this pass we look for GEP and cast instructions that are used 4398// across basic blocks and rewrite them to improve basic-block-at-a-time 4399// selection. 4400bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB, bool& ModifiedDT) { 4401 SunkAddrs.clear(); 4402 bool MadeChange = false; 4403 4404 CurInstIterator = BB.begin(); 4405 while (CurInstIterator != BB.end()) { 4406 MadeChange |= OptimizeInst(CurInstIterator++, ModifiedDT); 4407 if (ModifiedDT) 4408 return true; 4409 } 4410 MadeChange |= DupRetToEnableTailCallOpts(&BB); 4411 4412 return MadeChange; 4413} 4414 4415// llvm.dbg.value is far away from the value then iSel may not be able 4416// handle it properly. iSel will drop llvm.dbg.value if it can not 4417// find a node corresponding to the value. 4418bool CodeGenPrepare::PlaceDbgValues(Function &F) { 4419 bool MadeChange = false; 4420 for (BasicBlock &BB : F) { 4421 Instruction *PrevNonDbgInst = nullptr; 4422 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) { 4423 Instruction *Insn = BI++; 4424 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn); 4425 // Leave dbg.values that refer to an alloca alone. These 4426 // instrinsics describe the address of a variable (= the alloca) 4427 // being taken. They should not be moved next to the alloca 4428 // (and to the beginning of the scope), but rather stay close to 4429 // where said address is used. 4430 if (!DVI || (DVI->getValue() && isa<AllocaInst>(DVI->getValue()))) { 4431 PrevNonDbgInst = Insn; 4432 continue; 4433 } 4434 4435 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue()); 4436 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) { 4437 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI); 4438 DVI->removeFromParent(); 4439 if (isa<PHINode>(VI)) 4440 DVI->insertBefore(VI->getParent()->getFirstInsertionPt()); 4441 else 4442 DVI->insertAfter(VI); 4443 MadeChange = true; 4444 ++NumDbgValueMoved; 4445 } 4446 } 4447 } 4448 return MadeChange; 4449} 4450 4451// If there is a sequence that branches based on comparing a single bit 4452// against zero that can be combined into a single instruction, and the 4453// target supports folding these into a single instruction, sink the 4454// mask and compare into the branch uses. Do this before OptimizeBlock -> 4455// OptimizeInst -> OptimizeCmpExpression, which perturbs the pattern being 4456// searched for. 4457bool CodeGenPrepare::sinkAndCmp(Function &F) { 4458 if (!EnableAndCmpSinking) 4459 return false; 4460 if (!TLI || !TLI->isMaskAndBranchFoldingLegal()) 4461 return false; 4462 bool MadeChange = false; 4463 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) { 4464 BasicBlock *BB = I++; 4465 4466 // Does this BB end with the following? 4467 // %andVal = and %val, #single-bit-set 4468 // %icmpVal = icmp %andResult, 0 4469 // br i1 %cmpVal label %dest1, label %dest2" 4470 BranchInst *Brcc = dyn_cast<BranchInst>(BB->getTerminator()); 4471 if (!Brcc || !Brcc->isConditional()) 4472 continue; 4473 ICmpInst *Cmp = dyn_cast<ICmpInst>(Brcc->getOperand(0)); 4474 if (!Cmp || Cmp->getParent() != BB) 4475 continue; 4476 ConstantInt *Zero = dyn_cast<ConstantInt>(Cmp->getOperand(1)); 4477 if (!Zero || !Zero->isZero()) 4478 continue; 4479 Instruction *And = dyn_cast<Instruction>(Cmp->getOperand(0)); 4480 if (!And || And->getOpcode() != Instruction::And || And->getParent() != BB) 4481 continue; 4482 ConstantInt* Mask = dyn_cast<ConstantInt>(And->getOperand(1)); 4483 if (!Mask || !Mask->getUniqueInteger().isPowerOf2()) 4484 continue; 4485 DEBUG(dbgs() << "found and; icmp ?,0; brcc\n"); DEBUG(BB->dump()); 4486 4487 // Push the "and; icmp" for any users that are conditional branches. 4488 // Since there can only be one branch use per BB, we don't need to keep 4489 // track of which BBs we insert into. 4490 for (Value::use_iterator UI = Cmp->use_begin(), E = Cmp->use_end(); 4491 UI != E; ) { 4492 Use &TheUse = *UI; 4493 // Find brcc use. 4494 BranchInst *BrccUser = dyn_cast<BranchInst>(*UI); 4495 ++UI; 4496 if (!BrccUser || !BrccUser->isConditional()) 4497 continue; 4498 BasicBlock *UserBB = BrccUser->getParent(); 4499 if (UserBB == BB) continue; 4500 DEBUG(dbgs() << "found Brcc use\n"); 4501 4502 // Sink the "and; icmp" to use. 4503 MadeChange = true; 4504 BinaryOperator *NewAnd = 4505 BinaryOperator::CreateAnd(And->getOperand(0), And->getOperand(1), "", 4506 BrccUser); 4507 CmpInst *NewCmp = 4508 CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), NewAnd, Zero, 4509 "", BrccUser); 4510 TheUse = NewCmp; 4511 ++NumAndCmpsMoved; 4512 DEBUG(BrccUser->getParent()->dump()); 4513 } 4514 } 4515 return MadeChange; 4516} 4517 4518/// \brief Retrieve the probabilities of a conditional branch. Returns true on 4519/// success, or returns false if no or invalid metadata was found. 4520static bool extractBranchMetadata(BranchInst *BI, 4521 uint64_t &ProbTrue, uint64_t &ProbFalse) { 4522 assert(BI->isConditional() && 4523 "Looking for probabilities on unconditional branch?"); 4524 auto *ProfileData = BI->getMetadata(LLVMContext::MD_prof); 4525 if (!ProfileData || ProfileData->getNumOperands() != 3) 4526 return false; 4527 4528 const auto *CITrue = 4529 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(1)); 4530 const auto *CIFalse = 4531 mdconst::dyn_extract<ConstantInt>(ProfileData->getOperand(2)); 4532 if (!CITrue || !CIFalse) 4533 return false; 4534 4535 ProbTrue = CITrue->getValue().getZExtValue(); 4536 ProbFalse = CIFalse->getValue().getZExtValue(); 4537 4538 return true; 4539} 4540 4541/// \brief Scale down both weights to fit into uint32_t. 4542static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) { 4543 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse; 4544 uint32_t Scale = (NewMax / UINT32_MAX) + 1; 4545 NewTrue = NewTrue / Scale; 4546 NewFalse = NewFalse / Scale; 4547} 4548 4549/// \brief Some targets prefer to split a conditional branch like: 4550/// \code 4551/// %0 = icmp ne i32 %a, 0 4552/// %1 = icmp ne i32 %b, 0 4553/// %or.cond = or i1 %0, %1 4554/// br i1 %or.cond, label %TrueBB, label %FalseBB 4555/// \endcode 4556/// into multiple branch instructions like: 4557/// \code 4558/// bb1: 4559/// %0 = icmp ne i32 %a, 0 4560/// br i1 %0, label %TrueBB, label %bb2 4561/// bb2: 4562/// %1 = icmp ne i32 %b, 0 4563/// br i1 %1, label %TrueBB, label %FalseBB 4564/// \endcode 4565/// This usually allows instruction selection to do even further optimizations 4566/// and combine the compare with the branch instruction. Currently this is 4567/// applied for targets which have "cheap" jump instructions. 4568/// 4569/// FIXME: Remove the (equivalent?) implementation in SelectionDAG. 4570/// 4571bool CodeGenPrepare::splitBranchCondition(Function &F) { 4572 if (!TM || !TM->Options.EnableFastISel || !TLI || TLI->isJumpExpensive()) 4573 return false; 4574 4575 bool MadeChange = false; 4576 for (auto &BB : F) { 4577 // Does this BB end with the following? 4578 // %cond1 = icmp|fcmp|binary instruction ... 4579 // %cond2 = icmp|fcmp|binary instruction ... 4580 // %cond.or = or|and i1 %cond1, cond2 4581 // br i1 %cond.or label %dest1, label %dest2" 4582 BinaryOperator *LogicOp; 4583 BasicBlock *TBB, *FBB; 4584 if (!match(BB.getTerminator(), m_Br(m_OneUse(m_BinOp(LogicOp)), TBB, FBB))) 4585 continue; 4586 4587 unsigned Opc; 4588 Value *Cond1, *Cond2; 4589 if (match(LogicOp, m_And(m_OneUse(m_Value(Cond1)), 4590 m_OneUse(m_Value(Cond2))))) 4591 Opc = Instruction::And; 4592 else if (match(LogicOp, m_Or(m_OneUse(m_Value(Cond1)), 4593 m_OneUse(m_Value(Cond2))))) 4594 Opc = Instruction::Or; 4595 else 4596 continue; 4597 4598 if (!match(Cond1, m_CombineOr(m_Cmp(), m_BinOp())) || 4599 !match(Cond2, m_CombineOr(m_Cmp(), m_BinOp())) ) 4600 continue; 4601 4602 DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump()); 4603 4604 // Create a new BB. 4605 auto *InsertBefore = std::next(Function::iterator(BB)) 4606 .getNodePtrUnchecked(); 4607 auto TmpBB = BasicBlock::Create(BB.getContext(), 4608 BB.getName() + ".cond.split", 4609 BB.getParent(), InsertBefore); 4610 4611 // Update original basic block by using the first condition directly by the 4612 // branch instruction and removing the no longer needed and/or instruction. 4613 auto *Br1 = cast<BranchInst>(BB.getTerminator()); 4614 Br1->setCondition(Cond1); 4615 LogicOp->eraseFromParent(); 4616 4617 // Depending on the conditon we have to either replace the true or the false 4618 // successor of the original branch instruction. 4619 if (Opc == Instruction::And) 4620 Br1->setSuccessor(0, TmpBB); 4621 else 4622 Br1->setSuccessor(1, TmpBB); 4623 4624 // Fill in the new basic block. 4625 auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB); 4626 if (auto *I = dyn_cast<Instruction>(Cond2)) { 4627 I->removeFromParent(); 4628 I->insertBefore(Br2); 4629 } 4630 4631 // Update PHI nodes in both successors. The original BB needs to be 4632 // replaced in one succesor's PHI nodes, because the branch comes now from 4633 // the newly generated BB (NewBB). In the other successor we need to add one 4634 // incoming edge to the PHI nodes, because both branch instructions target 4635 // now the same successor. Depending on the original branch condition 4636 // (and/or) we have to swap the successors (TrueDest, FalseDest), so that 4637 // we perfrom the correct update for the PHI nodes. 4638 // This doesn't change the successor order of the just created branch 4639 // instruction (or any other instruction). 4640 if (Opc == Instruction::Or) 4641 std::swap(TBB, FBB); 4642 4643 // Replace the old BB with the new BB. 4644 for (auto &I : *TBB) { 4645 PHINode *PN = dyn_cast<PHINode>(&I); 4646 if (!PN) 4647 break; 4648 int i; 4649 while ((i = PN->getBasicBlockIndex(&BB)) >= 0) 4650 PN->setIncomingBlock(i, TmpBB); 4651 } 4652 4653 // Add another incoming edge form the new BB. 4654 for (auto &I : *FBB) { 4655 PHINode *PN = dyn_cast<PHINode>(&I); 4656 if (!PN) 4657 break; 4658 auto *Val = PN->getIncomingValueForBlock(&BB); 4659 PN->addIncoming(Val, TmpBB); 4660 } 4661 4662 // Update the branch weights (from SelectionDAGBuilder:: 4663 // FindMergedConditions). 4664 if (Opc == Instruction::Or) { 4665 // Codegen X | Y as: 4666 // BB1: 4667 // jmp_if_X TBB 4668 // jmp TmpBB 4669 // TmpBB: 4670 // jmp_if_Y TBB 4671 // jmp FBB 4672 // 4673 4674 // We have flexibility in setting Prob for BB1 and Prob for NewBB. 4675 // The requirement is that 4676 // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB) 4677 // = TrueProb for orignal BB. 4678 // Assuming the orignal weights are A and B, one choice is to set BB1's 4679 // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice 4680 // assumes that 4681 // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB. 4682 // Another choice is to assume TrueProb for BB1 equals to TrueProb for 4683 // TmpBB, but the math is more complicated. 4684 uint64_t TrueWeight, FalseWeight; 4685 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) { 4686 uint64_t NewTrueWeight = TrueWeight; 4687 uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight; 4688 scaleWeights(NewTrueWeight, NewFalseWeight); 4689 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext()) 4690 .createBranchWeights(TrueWeight, FalseWeight)); 4691 4692 NewTrueWeight = TrueWeight; 4693 NewFalseWeight = 2 * FalseWeight; 4694 scaleWeights(NewTrueWeight, NewFalseWeight); 4695 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext()) 4696 .createBranchWeights(TrueWeight, FalseWeight)); 4697 } 4698 } else { 4699 // Codegen X & Y as: 4700 // BB1: 4701 // jmp_if_X TmpBB 4702 // jmp FBB 4703 // TmpBB: 4704 // jmp_if_Y TBB 4705 // jmp FBB 4706 // 4707 // This requires creation of TmpBB after CurBB. 4708 4709 // We have flexibility in setting Prob for BB1 and Prob for TmpBB. 4710 // The requirement is that 4711 // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB) 4712 // = FalseProb for orignal BB. 4713 // Assuming the orignal weights are A and B, one choice is to set BB1's 4714 // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice 4715 // assumes that 4716 // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB. 4717 uint64_t TrueWeight, FalseWeight; 4718 if (extractBranchMetadata(Br1, TrueWeight, FalseWeight)) { 4719 uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight; 4720 uint64_t NewFalseWeight = FalseWeight; 4721 scaleWeights(NewTrueWeight, NewFalseWeight); 4722 Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext()) 4723 .createBranchWeights(TrueWeight, FalseWeight)); 4724 4725 NewTrueWeight = 2 * TrueWeight; 4726 NewFalseWeight = FalseWeight; 4727 scaleWeights(NewTrueWeight, NewFalseWeight); 4728 Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext()) 4729 .createBranchWeights(TrueWeight, FalseWeight)); 4730 } 4731 } 4732 4733 // Note: No point in getting fancy here, since the DT info is never 4734 // available to CodeGenPrepare. 4735 ModifiedDT = true; 4736 4737 MadeChange = true; 4738 4739 DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump(); 4740 TmpBB->dump()); 4741 } 4742 return MadeChange; 4743} 4744