CodeGenPrepare.cpp revision 84d1b40d448663050f12fb4dee052db907ac4748
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#define DEBUG_TYPE "codegenprepare" 17#include "llvm/Transforms/Scalar.h" 18#include "llvm/Constants.h" 19#include "llvm/DerivedTypes.h" 20#include "llvm/Function.h" 21#include "llvm/InlineAsm.h" 22#include "llvm/Instructions.h" 23#include "llvm/Pass.h" 24#include "llvm/Target/TargetAsmInfo.h" 25#include "llvm/Target/TargetData.h" 26#include "llvm/Target/TargetLowering.h" 27#include "llvm/Target/TargetMachine.h" 28#include "llvm/Transforms/Utils/BasicBlockUtils.h" 29#include "llvm/Transforms/Utils/Local.h" 30#include "llvm/ADT/DenseMap.h" 31#include "llvm/ADT/SmallSet.h" 32#include "llvm/Support/CallSite.h" 33#include "llvm/Support/Compiler.h" 34#include "llvm/Support/Debug.h" 35#include "llvm/Support/GetElementPtrTypeIterator.h" 36#include "llvm/Support/PatternMatch.h" 37using namespace llvm; 38using namespace llvm::PatternMatch; 39 40namespace { 41 class VISIBILITY_HIDDEN CodeGenPrepare : public FunctionPass { 42 /// TLI - Keep a pointer of a TargetLowering to consult for determining 43 /// transformation profitability. 44 const TargetLowering *TLI; 45 public: 46 static char ID; // Pass identification, replacement for typeid 47 explicit CodeGenPrepare(const TargetLowering *tli = 0) 48 : FunctionPass(&ID), TLI(tli) {} 49 bool runOnFunction(Function &F); 50 51 private: 52 bool EliminateMostlyEmptyBlocks(Function &F); 53 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; 54 void EliminateMostlyEmptyBlock(BasicBlock *BB); 55 bool OptimizeBlock(BasicBlock &BB); 56 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy, 57 DenseMap<Value*,Value*> &SunkAddrs); 58 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS, 59 DenseMap<Value*,Value*> &SunkAddrs); 60 bool OptimizeExtUses(Instruction *I); 61 }; 62} 63 64char CodeGenPrepare::ID = 0; 65static RegisterPass<CodeGenPrepare> X("codegenprepare", 66 "Optimize for code generation"); 67 68FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) { 69 return new CodeGenPrepare(TLI); 70} 71 72 73bool CodeGenPrepare::runOnFunction(Function &F) { 74 bool EverMadeChange = false; 75 76 // First pass, eliminate blocks that contain only PHI nodes and an 77 // unconditional branch. 78 EverMadeChange |= EliminateMostlyEmptyBlocks(F); 79 80 bool MadeChange = true; 81 while (MadeChange) { 82 MadeChange = false; 83 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 84 MadeChange |= OptimizeBlock(*BB); 85 EverMadeChange |= MadeChange; 86 } 87 return EverMadeChange; 88} 89 90/// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes 91/// and an unconditional branch. Passes before isel (e.g. LSR/loopsimplify) 92/// often split edges in ways that are non-optimal for isel. Start by 93/// eliminating these blocks so we can split them the way we want them. 94bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) { 95 bool MadeChange = false; 96 // Note that this intentionally skips the entry block. 97 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) { 98 BasicBlock *BB = I++; 99 100 // If this block doesn't end with an uncond branch, ignore it. 101 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 102 if (!BI || !BI->isUnconditional()) 103 continue; 104 105 // If the instruction before the branch isn't a phi node, then other stuff 106 // is happening here. 107 BasicBlock::iterator BBI = BI; 108 if (BBI != BB->begin()) { 109 --BBI; 110 if (!isa<PHINode>(BBI)) continue; 111 } 112 113 // Do not break infinite loops. 114 BasicBlock *DestBB = BI->getSuccessor(0); 115 if (DestBB == BB) 116 continue; 117 118 if (!CanMergeBlocks(BB, DestBB)) 119 continue; 120 121 EliminateMostlyEmptyBlock(BB); 122 MadeChange = true; 123 } 124 return MadeChange; 125} 126 127/// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a 128/// single uncond branch between them, and BB contains no other non-phi 129/// instructions. 130bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB, 131 const BasicBlock *DestBB) const { 132 // We only want to eliminate blocks whose phi nodes are used by phi nodes in 133 // the successor. If there are more complex condition (e.g. preheaders), 134 // don't mess around with them. 135 BasicBlock::const_iterator BBI = BB->begin(); 136 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 137 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end(); 138 UI != E; ++UI) { 139 const Instruction *User = cast<Instruction>(*UI); 140 if (User->getParent() != DestBB || !isa<PHINode>(User)) 141 return false; 142 // If User is inside DestBB block and it is a PHINode then check 143 // incoming value. If incoming value is not from BB then this is 144 // a complex condition (e.g. preheaders) we want to avoid here. 145 if (User->getParent() == DestBB) { 146 if (const PHINode *UPN = dyn_cast<PHINode>(User)) 147 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { 148 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); 149 if (Insn && Insn->getParent() == BB && 150 Insn->getParent() != UPN->getIncomingBlock(I)) 151 return false; 152 } 153 } 154 } 155 } 156 157 // If BB and DestBB contain any common predecessors, then the phi nodes in BB 158 // and DestBB may have conflicting incoming values for the block. If so, we 159 // can't merge the block. 160 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); 161 if (!DestBBPN) return true; // no conflict. 162 163 // Collect the preds of BB. 164 SmallPtrSet<const BasicBlock*, 16> BBPreds; 165 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 166 // It is faster to get preds from a PHI than with pred_iterator. 167 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 168 BBPreds.insert(BBPN->getIncomingBlock(i)); 169 } else { 170 BBPreds.insert(pred_begin(BB), pred_end(BB)); 171 } 172 173 // Walk the preds of DestBB. 174 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { 175 BasicBlock *Pred = DestBBPN->getIncomingBlock(i); 176 if (BBPreds.count(Pred)) { // Common predecessor? 177 BBI = DestBB->begin(); 178 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 179 const Value *V1 = PN->getIncomingValueForBlock(Pred); 180 const Value *V2 = PN->getIncomingValueForBlock(BB); 181 182 // If V2 is a phi node in BB, look up what the mapped value will be. 183 if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) 184 if (V2PN->getParent() == BB) 185 V2 = V2PN->getIncomingValueForBlock(Pred); 186 187 // If there is a conflict, bail out. 188 if (V1 != V2) return false; 189 } 190 } 191 } 192 193 return true; 194} 195 196 197/// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and 198/// an unconditional branch in it. 199void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { 200 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 201 BasicBlock *DestBB = BI->getSuccessor(0); 202 203 DOUT << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB; 204 205 // If the destination block has a single pred, then this is a trivial edge, 206 // just collapse it. 207 if (DestBB->getSinglePredecessor()) { 208 // If DestBB has single-entry PHI nodes, fold them. 209 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { 210 Value *NewVal = PN->getIncomingValue(0); 211 // Replace self referencing PHI with undef, it must be dead. 212 if (NewVal == PN) NewVal = UndefValue::get(PN->getType()); 213 PN->replaceAllUsesWith(NewVal); 214 PN->eraseFromParent(); 215 } 216 217 // Splice all the PHI nodes from BB over to DestBB. 218 DestBB->getInstList().splice(DestBB->begin(), BB->getInstList(), 219 BB->begin(), BI); 220 221 // Anything that branched to BB now branches to DestBB. 222 BB->replaceAllUsesWith(DestBB); 223 224 // Nuke BB. 225 BB->eraseFromParent(); 226 227 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 228 return; 229 } 230 231 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB 232 // to handle the new incoming edges it is about to have. 233 PHINode *PN; 234 for (BasicBlock::iterator BBI = DestBB->begin(); 235 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 236 // Remove the incoming value for BB, and remember it. 237 Value *InVal = PN->removeIncomingValue(BB, false); 238 239 // Two options: either the InVal is a phi node defined in BB or it is some 240 // value that dominates BB. 241 PHINode *InValPhi = dyn_cast<PHINode>(InVal); 242 if (InValPhi && InValPhi->getParent() == BB) { 243 // Add all of the input values of the input PHI as inputs of this phi. 244 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) 245 PN->addIncoming(InValPhi->getIncomingValue(i), 246 InValPhi->getIncomingBlock(i)); 247 } else { 248 // Otherwise, add one instance of the dominating value for each edge that 249 // we will be adding. 250 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 251 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 252 PN->addIncoming(InVal, BBPN->getIncomingBlock(i)); 253 } else { 254 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 255 PN->addIncoming(InVal, *PI); 256 } 257 } 258 } 259 260 // The PHIs are now updated, change everything that refers to BB to use 261 // DestBB and remove BB. 262 BB->replaceAllUsesWith(DestBB); 263 BB->eraseFromParent(); 264 265 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 266} 267 268 269/// SplitEdgeNicely - Split the critical edge from TI to its specified 270/// successor if it will improve codegen. We only do this if the successor has 271/// phi nodes (otherwise critical edges are ok). If there is already another 272/// predecessor of the succ that is empty (and thus has no phi nodes), use it 273/// instead of introducing a new block. 274static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, Pass *P) { 275 BasicBlock *TIBB = TI->getParent(); 276 BasicBlock *Dest = TI->getSuccessor(SuccNum); 277 assert(isa<PHINode>(Dest->begin()) && 278 "This should only be called if Dest has a PHI!"); 279 280 // As a hack, never split backedges of loops. Even though the copy for any 281 // PHIs inserted on the backedge would be dead for exits from the loop, we 282 // assume that the cost of *splitting* the backedge would be too high. 283 if (Dest == TIBB) 284 return; 285 286 /// TIPHIValues - This array is lazily computed to determine the values of 287 /// PHIs in Dest that TI would provide. 288 SmallVector<Value*, 32> TIPHIValues; 289 290 // Check to see if Dest has any blocks that can be used as a split edge for 291 // this terminator. 292 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { 293 BasicBlock *Pred = *PI; 294 // To be usable, the pred has to end with an uncond branch to the dest. 295 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); 296 if (!PredBr || !PredBr->isUnconditional() || 297 // Must be empty other than the branch. 298 &Pred->front() != PredBr || 299 // Cannot be the entry block; its label does not get emitted. 300 Pred == &(Dest->getParent()->getEntryBlock())) 301 continue; 302 303 // Finally, since we know that Dest has phi nodes in it, we have to make 304 // sure that jumping to Pred will have the same affect as going to Dest in 305 // terms of PHI values. 306 PHINode *PN; 307 unsigned PHINo = 0; 308 bool FoundMatch = true; 309 for (BasicBlock::iterator I = Dest->begin(); 310 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { 311 if (PHINo == TIPHIValues.size()) 312 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); 313 314 // If the PHI entry doesn't work, we can't use this pred. 315 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { 316 FoundMatch = false; 317 break; 318 } 319 } 320 321 // If we found a workable predecessor, change TI to branch to Succ. 322 if (FoundMatch) { 323 Dest->removePredecessor(TIBB); 324 TI->setSuccessor(SuccNum, Pred); 325 return; 326 } 327 } 328 329 SplitCriticalEdge(TI, SuccNum, P, true); 330} 331 332/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop 333/// copy (e.g. it's casting from one pointer type to another, int->uint, or 334/// int->sbyte on PPC), sink it into user blocks to reduce the number of virtual 335/// registers that must be created and coalesced. 336/// 337/// Return true if any changes are made. 338/// 339static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ 340 // If this is a noop copy, 341 MVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 342 MVT DstVT = TLI.getValueType(CI->getType()); 343 344 // This is an fp<->int conversion? 345 if (SrcVT.isInteger() != DstVT.isInteger()) 346 return false; 347 348 // If this is an extension, it will be a zero or sign extension, which 349 // isn't a noop. 350 if (SrcVT.bitsLT(DstVT)) return false; 351 352 // If these values will be promoted, find out what they will be promoted 353 // to. This helps us consider truncates on PPC as noop copies when they 354 // are. 355 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote) 356 SrcVT = TLI.getTypeToTransformTo(SrcVT); 357 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote) 358 DstVT = TLI.getTypeToTransformTo(DstVT); 359 360 // If, after promotion, these are the same types, this is a noop copy. 361 if (SrcVT != DstVT) 362 return false; 363 364 BasicBlock *DefBB = CI->getParent(); 365 366 /// InsertedCasts - Only insert a cast in each block once. 367 DenseMap<BasicBlock*, CastInst*> InsertedCasts; 368 369 bool MadeChange = false; 370 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 371 UI != E; ) { 372 Use &TheUse = UI.getUse(); 373 Instruction *User = cast<Instruction>(*UI); 374 375 // Figure out which BB this cast is used in. For PHI's this is the 376 // appropriate predecessor block. 377 BasicBlock *UserBB = User->getParent(); 378 if (PHINode *PN = dyn_cast<PHINode>(User)) { 379 unsigned OpVal = UI.getOperandNo()/2; 380 UserBB = PN->getIncomingBlock(OpVal); 381 } 382 383 // Preincrement use iterator so we don't invalidate it. 384 ++UI; 385 386 // If this user is in the same block as the cast, don't change the cast. 387 if (UserBB == DefBB) continue; 388 389 // If we have already inserted a cast into this block, use it. 390 CastInst *&InsertedCast = InsertedCasts[UserBB]; 391 392 if (!InsertedCast) { 393 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 394 395 InsertedCast = 396 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 397 InsertPt); 398 MadeChange = true; 399 } 400 401 // Replace a use of the cast with a use of the new cast. 402 TheUse = InsertedCast; 403 } 404 405 // If we removed all uses, nuke the cast. 406 if (CI->use_empty()) { 407 CI->eraseFromParent(); 408 MadeChange = true; 409 } 410 411 return MadeChange; 412} 413 414/// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce 415/// the number of virtual registers that must be created and coalesced. This is 416/// a clear win except on targets with multiple condition code registers 417/// (PowerPC), where it might lose; some adjustment may be wanted there. 418/// 419/// Return true if any changes are made. 420static bool OptimizeCmpExpression(CmpInst *CI) { 421 BasicBlock *DefBB = CI->getParent(); 422 423 /// InsertedCmp - Only insert a cmp in each block once. 424 DenseMap<BasicBlock*, CmpInst*> InsertedCmps; 425 426 bool MadeChange = false; 427 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 428 UI != E; ) { 429 Use &TheUse = UI.getUse(); 430 Instruction *User = cast<Instruction>(*UI); 431 432 // Preincrement use iterator so we don't invalidate it. 433 ++UI; 434 435 // Don't bother for PHI nodes. 436 if (isa<PHINode>(User)) 437 continue; 438 439 // Figure out which BB this cmp is used in. 440 BasicBlock *UserBB = User->getParent(); 441 442 // If this user is in the same block as the cmp, don't change the cmp. 443 if (UserBB == DefBB) continue; 444 445 // If we have already inserted a cmp into this block, use it. 446 CmpInst *&InsertedCmp = InsertedCmps[UserBB]; 447 448 if (!InsertedCmp) { 449 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 450 451 InsertedCmp = 452 CmpInst::Create(CI->getOpcode(), CI->getPredicate(), CI->getOperand(0), 453 CI->getOperand(1), "", InsertPt); 454 MadeChange = true; 455 } 456 457 // Replace a use of the cmp with a use of the new cmp. 458 TheUse = InsertedCmp; 459 } 460 461 // If we removed all uses, nuke the cmp. 462 if (CI->use_empty()) 463 CI->eraseFromParent(); 464 465 return MadeChange; 466} 467 468/// EraseDeadInstructions - Erase any dead instructions, recursively. 469static void EraseDeadInstructions(Value *V) { 470 Instruction *I = dyn_cast<Instruction>(V); 471 if (!I || !I->use_empty()) return; 472 473 SmallPtrSet<Instruction*, 16> Insts; 474 Insts.insert(I); 475 476 while (!Insts.empty()) { 477 I = *Insts.begin(); 478 Insts.erase(I); 479 if (isInstructionTriviallyDead(I)) { 480 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 481 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i))) 482 Insts.insert(U); 483 I->eraseFromParent(); 484 } 485 } 486} 487 488//===----------------------------------------------------------------------===// 489// Addressing Mode Analysis and Optimization 490//===----------------------------------------------------------------------===// 491 492namespace { 493 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode 494 /// which holds actual Value*'s for register values. 495 struct ExtAddrMode : public TargetLowering::AddrMode { 496 Value *BaseReg; 497 Value *ScaledReg; 498 ExtAddrMode() : BaseReg(0), ScaledReg(0) {} 499 void print(OStream &OS) const; 500 void dump() const { 501 print(cerr); 502 cerr << '\n'; 503 } 504 }; 505} // end anonymous namespace 506 507static inline OStream &operator<<(OStream &OS, const ExtAddrMode &AM) { 508 AM.print(OS); 509 return OS; 510} 511 512void ExtAddrMode::print(OStream &OS) const { 513 bool NeedPlus = false; 514 OS << "["; 515 if (BaseGV) 516 OS << (NeedPlus ? " + " : "") 517 << "GV:%" << BaseGV->getName(), NeedPlus = true; 518 519 if (BaseOffs) 520 OS << (NeedPlus ? " + " : "") << BaseOffs, NeedPlus = true; 521 522 if (BaseReg) 523 OS << (NeedPlus ? " + " : "") 524 << "Base:%" << BaseReg->getName(), NeedPlus = true; 525 if (Scale) 526 OS << (NeedPlus ? " + " : "") 527 << Scale << "*%" << ScaledReg->getName(), NeedPlus = true; 528 529 OS << ']'; 530} 531 532namespace { 533/// AddressingModeMatcher - This class exposes a single public method, which is 534/// used to construct a "maximal munch" of the addressing mode for the target 535/// specified by TLI for an access to "V" with an access type of AccessTy. This 536/// returns the addressing mode that is actually matched by value, but also 537/// returns the list of instructions involved in that addressing computation in 538/// AddrModeInsts. 539class AddressingModeMatcher { 540 SmallVectorImpl<Instruction*> &AddrModeInsts; 541 const TargetLowering &TLI; 542 const Type *AccessTy; 543 ExtAddrMode &AddrMode; 544 545 /// IgnoreProfitability - This is set to true when we should not do 546 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode 547 /// always returns true. 548 bool IgnoreProfitability; 549 550 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI, 551 const TargetLowering &T, const Type *AT,ExtAddrMode &AM) 552 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), AddrMode(AM) { 553 IgnoreProfitability = false; 554 } 555public: 556 557 /// Match - Find the maximal addressing mode that a load/store of V can fold, 558 /// give an access type of AccessTy. This returns a list of involved 559 /// instructions in AddrModeInsts. 560 static ExtAddrMode Match(Value *V, const Type *AccessTy, 561 SmallVectorImpl<Instruction*> &AddrModeInsts, 562 const TargetLowering &TLI) { 563 ExtAddrMode Result; 564 565 bool Success = 566 AddressingModeMatcher(AddrModeInsts,TLI,AccessTy,Result).MatchAddr(V, 0); 567 Success = Success; assert(Success && "Couldn't select *anything*?"); 568 return Result; 569 } 570private: 571 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth); 572 bool MatchAddr(Value *V, unsigned Depth); 573 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth); 574 bool IsProfitableToFoldIntoAddressingMode(Instruction *I, 575 ExtAddrMode &AMBefore, 576 ExtAddrMode &AMAfter); 577 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2); 578}; 579} // end anonymous namespace 580 581/// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode. 582/// Return true and update AddrMode if this addr mode is legal for the target, 583/// false if not. 584bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale, 585 unsigned Depth) { 586 // If Scale is 1, then this is the same as adding ScaleReg to the addressing 587 // mode. Just process that directly. 588 if (Scale == 1) 589 return MatchAddr(ScaleReg, Depth); 590 591 // If the scale is 0, it takes nothing to add this. 592 if (Scale == 0) 593 return true; 594 595 // If we already have a scale of this value, we can add to it, otherwise, we 596 // need an available scale field. 597 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg) 598 return false; 599 600 ExtAddrMode TestAddrMode = AddrMode; 601 602 // Add scale to turn X*4+X*3 -> X*7. This could also do things like 603 // [A+B + A*7] -> [B+A*8]. 604 TestAddrMode.Scale += Scale; 605 TestAddrMode.ScaledReg = ScaleReg; 606 607 // If the new address isn't legal, bail out. 608 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) 609 return false; 610 611 // It was legal, so commit it. 612 AddrMode = TestAddrMode; 613 614 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now 615 // to see if ScaleReg is actually X+C. If so, we can turn this into adding 616 // X*Scale + C*Scale to addr mode. 617 ConstantInt *CI; Value *AddLHS; 618 if (match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) { 619 TestAddrMode.ScaledReg = AddLHS; 620 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale; 621 622 // If this addressing mode is legal, commit it and remember that we folded 623 // this instruction. 624 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) { 625 AddrModeInsts.push_back(cast<Instruction>(ScaleReg)); 626 AddrMode = TestAddrMode; 627 return true; 628 } 629 } 630 631 // Otherwise, not (x+c)*scale, just return what we have. 632 return true; 633} 634 635/// MightBeFoldableInst - This is a little filter, which returns true if an 636/// addressing computation involving I might be folded into a load/store 637/// accessing it. This doesn't need to be perfect, but needs to accept at least 638/// the set of instructions that MatchOperationAddr can. 639static bool MightBeFoldableInst(Instruction *I) { 640 switch (I->getOpcode()) { 641 case Instruction::BitCast: 642 // Don't touch identity bitcasts. 643 if (I->getType() == I->getOperand(0)->getType()) 644 return false; 645 return isa<PointerType>(I->getType()) || isa<IntegerType>(I->getType()); 646 case Instruction::PtrToInt: 647 // PtrToInt is always a noop, as we know that the int type is pointer sized. 648 return true; 649 case Instruction::IntToPtr: 650 // We know the input is intptr_t, so this is foldable. 651 return true; 652 case Instruction::Add: 653 return true; 654 case Instruction::Mul: 655 case Instruction::Shl: 656 // Can only handle X*C and X << C. 657 return isa<ConstantInt>(I->getOperand(1)); 658 case Instruction::GetElementPtr: 659 return true; 660 default: 661 return false; 662 } 663} 664 665 666/// MatchOperationAddr - Given an instruction or constant expr, see if we can 667/// fold the operation into the addressing mode. If so, update the addressing 668/// mode and return true, otherwise return false without modifying AddrMode. 669bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode, 670 unsigned Depth) { 671 // Avoid exponential behavior on extremely deep expression trees. 672 if (Depth >= 5) return false; 673 674 switch (Opcode) { 675 case Instruction::PtrToInt: 676 // PtrToInt is always a noop, as we know that the int type is pointer sized. 677 return MatchAddr(AddrInst->getOperand(0), Depth); 678 case Instruction::IntToPtr: 679 // This inttoptr is a no-op if the integer type is pointer sized. 680 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) == 681 TLI.getPointerTy()) 682 return MatchAddr(AddrInst->getOperand(0), Depth); 683 return false; 684 case Instruction::BitCast: 685 // BitCast is always a noop, and we can handle it as long as it is 686 // int->int or pointer->pointer (we don't want int<->fp or something). 687 if ((isa<PointerType>(AddrInst->getOperand(0)->getType()) || 688 isa<IntegerType>(AddrInst->getOperand(0)->getType())) && 689 // Don't touch identity bitcasts. These were probably put here by LSR, 690 // and we don't want to mess around with them. Assume it knows what it 691 // is doing. 692 AddrInst->getOperand(0)->getType() != AddrInst->getType()) 693 return MatchAddr(AddrInst->getOperand(0), Depth); 694 return false; 695 case Instruction::Add: { 696 // Check to see if we can merge in the RHS then the LHS. If so, we win. 697 ExtAddrMode BackupAddrMode = AddrMode; 698 unsigned OldSize = AddrModeInsts.size(); 699 if (MatchAddr(AddrInst->getOperand(1), Depth+1) && 700 MatchAddr(AddrInst->getOperand(0), Depth+1)) 701 return true; 702 703 // Restore the old addr mode info. 704 AddrMode = BackupAddrMode; 705 AddrModeInsts.resize(OldSize); 706 707 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS. 708 if (MatchAddr(AddrInst->getOperand(0), Depth+1) && 709 MatchAddr(AddrInst->getOperand(1), Depth+1)) 710 return true; 711 712 // Otherwise we definitely can't merge the ADD in. 713 AddrMode = BackupAddrMode; 714 AddrModeInsts.resize(OldSize); 715 break; 716 } 717 //case Instruction::Or: 718 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD. 719 //break; 720 case Instruction::Mul: 721 case Instruction::Shl: { 722 // Can only handle X*C and X << C. 723 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); 724 if (!RHS) return false; 725 int64_t Scale = RHS->getSExtValue(); 726 if (Opcode == Instruction::Shl) 727 Scale = 1 << Scale; 728 729 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth); 730 } 731 case Instruction::GetElementPtr: { 732 // Scan the GEP. We check it if it contains constant offsets and at most 733 // one variable offset. 734 int VariableOperand = -1; 735 unsigned VariableScale = 0; 736 737 int64_t ConstantOffset = 0; 738 const TargetData *TD = TLI.getTargetData(); 739 gep_type_iterator GTI = gep_type_begin(AddrInst); 740 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) { 741 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 742 const StructLayout *SL = TD->getStructLayout(STy); 743 unsigned Idx = 744 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue(); 745 ConstantOffset += SL->getElementOffset(Idx); 746 } else { 747 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType()); 748 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) { 749 ConstantOffset += CI->getSExtValue()*TypeSize; 750 } else if (TypeSize) { // Scales of zero don't do anything. 751 // We only allow one variable index at the moment. 752 if (VariableOperand != -1) 753 return false; 754 755 // Remember the variable index. 756 VariableOperand = i; 757 VariableScale = TypeSize; 758 } 759 } 760 } 761 762 // A common case is for the GEP to only do a constant offset. In this case, 763 // just add it to the disp field and check validity. 764 if (VariableOperand == -1) { 765 AddrMode.BaseOffs += ConstantOffset; 766 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){ 767 // Check to see if we can fold the base pointer in too. 768 if (MatchAddr(AddrInst->getOperand(0), Depth+1)) 769 return true; 770 } 771 AddrMode.BaseOffs -= ConstantOffset; 772 return false; 773 } 774 775 // Save the valid addressing mode in case we can't match. 776 ExtAddrMode BackupAddrMode = AddrMode; 777 778 // Check that this has no base reg yet. If so, we won't have a place to 779 // put the base of the GEP (assuming it is not a null ptr). 780 bool SetBaseReg = true; 781 if (isa<ConstantPointerNull>(AddrInst->getOperand(0))) 782 SetBaseReg = false; // null pointer base doesn't need representation. 783 else if (AddrMode.HasBaseReg) 784 return false; // Base register already specified, can't match GEP. 785 else { 786 // Otherwise, we'll use the GEP base as the BaseReg. 787 AddrMode.HasBaseReg = true; 788 AddrMode.BaseReg = AddrInst->getOperand(0); 789 } 790 791 // See if the scale and offset amount is valid for this target. 792 AddrMode.BaseOffs += ConstantOffset; 793 794 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale, 795 Depth)) { 796 AddrMode = BackupAddrMode; 797 return false; 798 } 799 800 // If we have a null as the base of the GEP, folding in the constant offset 801 // plus variable scale is all we can do. 802 if (!SetBaseReg) return true; 803 804 // If this match succeeded, we know that we can form an address with the 805 // GepBase as the basereg. Match the base pointer of the GEP more 806 // aggressively by zeroing out BaseReg and rematching. If the base is 807 // (for example) another GEP, this allows merging in that other GEP into 808 // the addressing mode we're forming. 809 AddrMode.HasBaseReg = false; 810 AddrMode.BaseReg = 0; 811 bool Success = MatchAddr(AddrInst->getOperand(0), Depth+1); 812 assert(Success && "MatchAddr should be able to fill in BaseReg!"); 813 Success=Success; 814 return true; 815 } 816 } 817 return false; 818} 819 820/// MatchAddr - If we can, try to add the value of 'Addr' into the current 821/// addressing mode. If Addr can't be added to AddrMode this returns false and 822/// leaves AddrMode unmodified. This assumes that Addr is either a pointer type 823/// or intptr_t for the target. 824/// 825bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) { 826 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) { 827 // Fold in immediates if legal for the target. 828 AddrMode.BaseOffs += CI->getSExtValue(); 829 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 830 return true; 831 AddrMode.BaseOffs -= CI->getSExtValue(); 832 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) { 833 // If this is a global variable, try to fold it into the addressing mode. 834 if (AddrMode.BaseGV == 0) { 835 AddrMode.BaseGV = GV; 836 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 837 return true; 838 AddrMode.BaseGV = 0; 839 } 840 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) { 841 ExtAddrMode BackupAddrMode = AddrMode; 842 unsigned OldSize = AddrModeInsts.size(); 843 844 // Check to see if it is possible to fold this operation. 845 if (MatchOperationAddr(I, I->getOpcode(), Depth)) { 846 // Okay, it's possible to fold this. Check to see if it is actually 847 // *profitable* to do so. We use a simple cost model to avoid increasing 848 // register pressure too much. 849 if (I->hasOneUse() || 850 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) { 851 AddrModeInsts.push_back(I); 852 return true; 853 } 854 855 // It isn't profitable to do this, roll back. 856 //cerr << "NOT FOLDING: " << *I; 857 AddrMode = BackupAddrMode; 858 AddrModeInsts.resize(OldSize); 859 } 860 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) { 861 if (MatchOperationAddr(CE, CE->getOpcode(), Depth)) 862 return true; 863 } else if (isa<ConstantPointerNull>(Addr)) { 864 // Null pointer gets folded without affecting the addressing mode. 865 return true; 866 } 867 868 // Worse case, the target should support [reg] addressing modes. :) 869 if (!AddrMode.HasBaseReg) { 870 AddrMode.HasBaseReg = true; 871 AddrMode.BaseReg = Addr; 872 // Still check for legality in case the target supports [imm] but not [i+r]. 873 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 874 return true; 875 AddrMode.HasBaseReg = false; 876 AddrMode.BaseReg = 0; 877 } 878 879 // If the base register is already taken, see if we can do [r+r]. 880 if (AddrMode.Scale == 0) { 881 AddrMode.Scale = 1; 882 AddrMode.ScaledReg = Addr; 883 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 884 return true; 885 AddrMode.Scale = 0; 886 AddrMode.ScaledReg = 0; 887 } 888 // Couldn't match. 889 return false; 890} 891 892/// FindAllMemoryUses - Recursively walk all the uses of I until we find a 893/// memory use. If we find an obviously non-foldable instruction, return true. 894/// Add the ultimately found memory instructions to MemoryUses. 895static bool FindAllMemoryUses(Instruction *I, 896 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses, 897 SmallPtrSet<Instruction*, 16> &ConsideredInsts) { 898 // If we already considered this instruction, we're done. 899 if (!ConsideredInsts.insert(I)) 900 return false; 901 902 // If this is an obviously unfoldable instruction, bail out. 903 if (!MightBeFoldableInst(I)) 904 return true; 905 906 // Loop over all the uses, recursively processing them. 907 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 908 UI != E; ++UI) { 909 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 910 MemoryUses.push_back(std::make_pair(LI, UI.getOperandNo())); 911 continue; 912 } 913 914 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 915 if (UI.getOperandNo() == 0) return true; // Storing addr, not into addr. 916 MemoryUses.push_back(std::make_pair(SI, UI.getOperandNo())); 917 continue; 918 } 919 920 if (CallInst *CI = dyn_cast<CallInst>(*UI)) { 921 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue()); 922 if (IA == 0) return true; 923 924 925 // FIXME: HANDLE MEM OPS 926 //MemoryUses.push_back(std::make_pair(CI, UI.getOperandNo())); 927 return true; 928 } 929 930 if (FindAllMemoryUses(cast<Instruction>(*UI), MemoryUses, ConsideredInsts)) 931 return true; 932 } 933 934 return false; 935} 936 937 938/// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at 939/// the use site that we're folding it into. If so, there is no cost to 940/// include it in the addressing mode. KnownLive1 and KnownLive2 are two values 941/// that we know are live at the instruction already. 942bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1, 943 Value *KnownLive2) { 944 // If Val is either of the known-live values, we know it is live! 945 if (Val == 0 || Val == KnownLive1 || Val == KnownLive2) 946 return true; 947 948 // All non-instruction values other than arguments (constants) are live. 949 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true; 950 951 // If Val is a constant sized alloca in the entry block, it is live, this is 952 // true because it is just a reference to the stack/frame pointer, which is 953 // live for the whole function. 954 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val)) 955 if (AI->isStaticAlloca()) 956 return true; 957 958 return false; 959} 960 961 962 963#include "llvm/Support/CommandLine.h" 964cl::opt<bool> ENABLECRAZYHACK("enable-smarter-addr-folding", cl::Hidden); 965 966 967/// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing 968/// mode of the machine to fold the specified instruction into a load or store 969/// that ultimately uses it. However, the specified instruction has multiple 970/// uses. Given this, it may actually increase register pressure to fold it 971/// into the load. For example, consider this code: 972/// 973/// X = ... 974/// Y = X+1 975/// use(Y) -> nonload/store 976/// Z = Y+1 977/// load Z 978/// 979/// In this case, Y has multiple uses, and can be folded into the load of Z 980/// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to 981/// be live at the use(Y) line. If we don't fold Y into load Z, we use one 982/// fewer register. Since Y can't be folded into "use(Y)" we don't increase the 983/// number of computations either. 984/// 985/// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If 986/// X was live across 'load Z' for other reasons, we actually *would* want to 987/// fold the addressing mode in the Z case. This would make Y die earlier. 988bool AddressingModeMatcher:: 989IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore, 990 ExtAddrMode &AMAfter) { 991 if (IgnoreProfitability || !ENABLECRAZYHACK) return true; 992 993 // AMBefore is the addressing mode before this instruction was folded into it, 994 // and AMAfter is the addressing mode after the instruction was folded. Get 995 // the set of registers referenced by AMAfter and subtract out those 996 // referenced by AMBefore: this is the set of values which folding in this 997 // address extends the lifetime of. 998 // 999 // Note that there are only two potential values being referenced here, 1000 // BaseReg and ScaleReg (global addresses are always available, as are any 1001 // folded immediates). 1002 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg; 1003 1004 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their 1005 // lifetime wasn't extended by adding this instruction. 1006 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg)) 1007 BaseReg = 0; 1008 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg)) 1009 ScaledReg = 0; 1010 1011 // If folding this instruction (and it's subexprs) didn't extend any live 1012 // ranges, we're ok with it. 1013 if (BaseReg == 0 && ScaledReg == 0) 1014 return true; 1015 1016 // If all uses of this instruction are ultimately load/store/inlineasm's, 1017 // check to see if their addressing modes will include this instruction. If 1018 // so, we can fold it into all uses, so it doesn't matter if it has multiple 1019 // uses. 1020 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses; 1021 SmallPtrSet<Instruction*, 16> ConsideredInsts; 1022 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts)) 1023 return false; // Has a non-memory, non-foldable use! 1024 1025 // Now that we know that all uses of this instruction are part of a chain of 1026 // computation involving only operations that could theoretically be folded 1027 // into a memory use, loop over each of these uses and see if they could 1028 // *actually* fold the instruction. 1029 SmallVector<Instruction*, 32> MatchedAddrModeInsts; 1030 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) { 1031 Instruction *User = MemoryUses[i].first; 1032 unsigned OpNo = MemoryUses[i].second; 1033 1034 // Get the access type of this use. If the use isn't a pointer, we don't 1035 // know what it accesses. 1036 Value *Address = User->getOperand(OpNo); 1037 if (!isa<PointerType>(Address->getType())) 1038 return false; 1039 const Type *AddressAccessTy = 1040 cast<PointerType>(Address->getType())->getElementType(); 1041 1042 // Do a match against the root of this address, ignoring profitability. This 1043 // will tell us if the addressing mode for the memory operation will 1044 // *actually* cover the shared instruction. 1045 ExtAddrMode Result; 1046 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy, 1047 Result); 1048 Matcher.IgnoreProfitability = true; 1049 bool Success = Matcher.MatchAddr(Address, 0); 1050 Success = Success; assert(Success && "Couldn't select *anything*?"); 1051 1052 // If the match didn't cover I, then it won't be shared by it. 1053 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(), 1054 I) == MatchedAddrModeInsts.end()) 1055 return false; 1056 1057 MatchedAddrModeInsts.clear(); 1058 } 1059 1060 return true; 1061} 1062 1063 1064//===----------------------------------------------------------------------===// 1065// Memory Optimization 1066//===----------------------------------------------------------------------===// 1067 1068/// IsNonLocalValue - Return true if the specified values are defined in a 1069/// different basic block than BB. 1070static bool IsNonLocalValue(Value *V, BasicBlock *BB) { 1071 if (Instruction *I = dyn_cast<Instruction>(V)) 1072 return I->getParent() != BB; 1073 return false; 1074} 1075 1076/// OptimizeMemoryInst - Load and Store Instructions have often have 1077/// addressing modes that can do significant amounts of computation. As such, 1078/// instruction selection will try to get the load or store to do as much 1079/// computation as possible for the program. The problem is that isel can only 1080/// see within a single block. As such, we sink as much legal addressing mode 1081/// stuff into the block as possible. 1082/// 1083/// This method is used to optimize both load/store and inline asms with memory 1084/// operands. 1085bool CodeGenPrepare::OptimizeMemoryInst(Instruction *LdStInst, Value *Addr, 1086 const Type *AccessTy, 1087 DenseMap<Value*,Value*> &SunkAddrs) { 1088 // Figure out what addressing mode will be built up for this operation. 1089 SmallVector<Instruction*, 16> AddrModeInsts; 1090 ExtAddrMode AddrMode = 1091 AddressingModeMatcher::Match(Addr, AccessTy, AddrModeInsts, *TLI); 1092 1093 // Check to see if any of the instructions supersumed by this addr mode are 1094 // non-local to I's BB. 1095 bool AnyNonLocal = false; 1096 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { 1097 if (IsNonLocalValue(AddrModeInsts[i], LdStInst->getParent())) { 1098 AnyNonLocal = true; 1099 break; 1100 } 1101 } 1102 1103 // If all the instructions matched are already in this BB, don't do anything. 1104 if (!AnyNonLocal) { 1105 DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n"); 1106 return false; 1107 } 1108 1109 // Insert this computation right after this user. Since our caller is 1110 // scanning from the top of the BB to the bottom, reuse of the expr are 1111 // guaranteed to happen later. 1112 BasicBlock::iterator InsertPt = LdStInst; 1113 1114 // Now that we determined the addressing expression we want to use and know 1115 // that we have to sink it into this block. Check to see if we have already 1116 // done this for some other load/store instr in this block. If so, reuse the 1117 // computation. 1118 Value *&SunkAddr = SunkAddrs[Addr]; 1119 if (SunkAddr) { 1120 DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << "\n"); 1121 if (SunkAddr->getType() != Addr->getType()) 1122 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt); 1123 } else { 1124 DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << "\n"); 1125 const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType(); 1126 1127 Value *Result = 0; 1128 // Start with the scale value. 1129 if (AddrMode.Scale) { 1130 Value *V = AddrMode.ScaledReg; 1131 if (V->getType() == IntPtrTy) { 1132 // done. 1133 } else if (isa<PointerType>(V->getType())) { 1134 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 1135 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 1136 cast<IntegerType>(V->getType())->getBitWidth()) { 1137 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt); 1138 } else { 1139 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt); 1140 } 1141 if (AddrMode.Scale != 1) 1142 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy, 1143 AddrMode.Scale), 1144 "sunkaddr", InsertPt); 1145 Result = V; 1146 } 1147 1148 // Add in the base register. 1149 if (AddrMode.BaseReg) { 1150 Value *V = AddrMode.BaseReg; 1151 if (V->getType() != IntPtrTy) 1152 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 1153 if (Result) 1154 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 1155 else 1156 Result = V; 1157 } 1158 1159 // Add in the BaseGV if present. 1160 if (AddrMode.BaseGV) { 1161 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr", 1162 InsertPt); 1163 if (Result) 1164 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 1165 else 1166 Result = V; 1167 } 1168 1169 // Add in the Base Offset if present. 1170 if (AddrMode.BaseOffs) { 1171 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 1172 if (Result) 1173 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 1174 else 1175 Result = V; 1176 } 1177 1178 if (Result == 0) 1179 SunkAddr = Constant::getNullValue(Addr->getType()); 1180 else 1181 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt); 1182 } 1183 1184 LdStInst->replaceUsesOfWith(Addr, SunkAddr); 1185 1186 if (Addr->use_empty()) 1187 EraseDeadInstructions(Addr); 1188 return true; 1189} 1190 1191/// OptimizeInlineAsmInst - If there are any memory operands, use 1192/// OptimizeMemoryInst to sink their address computing into the block when 1193/// possible / profitable. 1194bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS, 1195 DenseMap<Value*,Value*> &SunkAddrs) { 1196 bool MadeChange = false; 1197 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 1198 1199 // Do a prepass over the constraints, canonicalizing them, and building up the 1200 // ConstraintOperands list. 1201 std::vector<InlineAsm::ConstraintInfo> 1202 ConstraintInfos = IA->ParseConstraints(); 1203 1204 /// ConstraintOperands - Information about all of the constraints. 1205 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands; 1206 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 1207 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 1208 ConstraintOperands. 1209 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i])); 1210 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back(); 1211 1212 // Compute the value type for each operand. 1213 switch (OpInfo.Type) { 1214 case InlineAsm::isOutput: 1215 if (OpInfo.isIndirect) 1216 OpInfo.CallOperandVal = CS.getArgument(ArgNo++); 1217 break; 1218 case InlineAsm::isInput: 1219 OpInfo.CallOperandVal = CS.getArgument(ArgNo++); 1220 break; 1221 case InlineAsm::isClobber: 1222 // Nothing to do. 1223 break; 1224 } 1225 1226 // Compute the constraint code and ConstraintType to use. 1227 TLI->ComputeConstraintToUse(OpInfo, SDValue(), 1228 OpInfo.ConstraintType == TargetLowering::C_Memory); 1229 1230 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 1231 OpInfo.isIndirect) { 1232 Value *OpVal = OpInfo.CallOperandVal; 1233 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs); 1234 } 1235 } 1236 1237 return MadeChange; 1238} 1239 1240bool CodeGenPrepare::OptimizeExtUses(Instruction *I) { 1241 BasicBlock *DefBB = I->getParent(); 1242 1243 // If both result of the {s|z}xt and its source are live out, rewrite all 1244 // other uses of the source with result of extension. 1245 Value *Src = I->getOperand(0); 1246 if (Src->hasOneUse()) 1247 return false; 1248 1249 // Only do this xform if truncating is free. 1250 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType())) 1251 return false; 1252 1253 // Only safe to perform the optimization if the source is also defined in 1254 // this block. 1255 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) 1256 return false; 1257 1258 bool DefIsLiveOut = false; 1259 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 1260 UI != E; ++UI) { 1261 Instruction *User = cast<Instruction>(*UI); 1262 1263 // Figure out which BB this ext is used in. 1264 BasicBlock *UserBB = User->getParent(); 1265 if (UserBB == DefBB) continue; 1266 DefIsLiveOut = true; 1267 break; 1268 } 1269 if (!DefIsLiveOut) 1270 return false; 1271 1272 // Make sure non of the uses are PHI nodes. 1273 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 1274 UI != E; ++UI) { 1275 Instruction *User = cast<Instruction>(*UI); 1276 BasicBlock *UserBB = User->getParent(); 1277 if (UserBB == DefBB) continue; 1278 // Be conservative. We don't want this xform to end up introducing 1279 // reloads just before load / store instructions. 1280 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User)) 1281 return false; 1282 } 1283 1284 // InsertedTruncs - Only insert one trunc in each block once. 1285 DenseMap<BasicBlock*, Instruction*> InsertedTruncs; 1286 1287 bool MadeChange = false; 1288 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 1289 UI != E; ++UI) { 1290 Use &TheUse = UI.getUse(); 1291 Instruction *User = cast<Instruction>(*UI); 1292 1293 // Figure out which BB this ext is used in. 1294 BasicBlock *UserBB = User->getParent(); 1295 if (UserBB == DefBB) continue; 1296 1297 // Both src and def are live in this block. Rewrite the use. 1298 Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; 1299 1300 if (!InsertedTrunc) { 1301 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 1302 1303 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt); 1304 } 1305 1306 // Replace a use of the {s|z}ext source with a use of the result. 1307 TheUse = InsertedTrunc; 1308 1309 MadeChange = true; 1310 } 1311 1312 return MadeChange; 1313} 1314 1315// In this pass we look for GEP and cast instructions that are used 1316// across basic blocks and rewrite them to improve basic-block-at-a-time 1317// selection. 1318bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { 1319 bool MadeChange = false; 1320 1321 // Split all critical edges where the dest block has a PHI and where the phi 1322 // has shared immediate operands. 1323 TerminatorInst *BBTI = BB.getTerminator(); 1324 if (BBTI->getNumSuccessors() > 1) { 1325 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) 1326 if (isa<PHINode>(BBTI->getSuccessor(i)->begin()) && 1327 isCriticalEdge(BBTI, i, true)) 1328 SplitEdgeNicely(BBTI, i, this); 1329 } 1330 1331 1332 // Keep track of non-local addresses that have been sunk into this block. 1333 // This allows us to avoid inserting duplicate code for blocks with multiple 1334 // load/stores of the same address. 1335 DenseMap<Value*, Value*> SunkAddrs; 1336 1337 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) { 1338 Instruction *I = BBI++; 1339 1340 if (CastInst *CI = dyn_cast<CastInst>(I)) { 1341 // If the source of the cast is a constant, then this should have 1342 // already been constant folded. The only reason NOT to constant fold 1343 // it is if something (e.g. LSR) was careful to place the constant 1344 // evaluation in a block other than then one that uses it (e.g. to hoist 1345 // the address of globals out of a loop). If this is the case, we don't 1346 // want to forward-subst the cast. 1347 if (isa<Constant>(CI->getOperand(0))) 1348 continue; 1349 1350 bool Change = false; 1351 if (TLI) { 1352 Change = OptimizeNoopCopyExpression(CI, *TLI); 1353 MadeChange |= Change; 1354 } 1355 1356 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) 1357 MadeChange |= OptimizeExtUses(I); 1358 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 1359 MadeChange |= OptimizeCmpExpression(CI); 1360 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1361 if (TLI) 1362 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(), 1363 SunkAddrs); 1364 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 1365 if (TLI) 1366 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1), 1367 SI->getOperand(0)->getType(), 1368 SunkAddrs); 1369 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 1370 if (GEPI->hasAllZeroIndices()) { 1371 /// The GEP operand must be a pointer, so must its result -> BitCast 1372 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 1373 GEPI->getName(), GEPI); 1374 GEPI->replaceAllUsesWith(NC); 1375 GEPI->eraseFromParent(); 1376 MadeChange = true; 1377 BBI = NC; 1378 } 1379 } else if (CallInst *CI = dyn_cast<CallInst>(I)) { 1380 // If we found an inline asm expession, and if the target knows how to 1381 // lower it to normal LLVM code, do so now. 1382 if (TLI && isa<InlineAsm>(CI->getCalledValue())) 1383 if (const TargetAsmInfo *TAI = 1384 TLI->getTargetMachine().getTargetAsmInfo()) { 1385 if (TAI->ExpandInlineAsm(CI)) 1386 BBI = BB.begin(); 1387 else 1388 // Sink address computing for memory operands into the block. 1389 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs); 1390 } 1391 } 1392 } 1393 1394 return MadeChange; 1395} 1396