CodeGenPrepare.cpp revision 47f5751c8092f96b989318faab991b23e558babf
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" 36using namespace llvm; 37 38namespace { 39 class VISIBILITY_HIDDEN CodeGenPrepare : public FunctionPass { 40 /// TLI - Keep a pointer of a TargetLowering to consult for determining 41 /// transformation profitability. 42 const TargetLowering *TLI; 43 public: 44 static char ID; // Pass identification, replacement for typeid 45 explicit CodeGenPrepare(const TargetLowering *tli = 0) 46 : FunctionPass(&ID), TLI(tli) {} 47 bool runOnFunction(Function &F); 48 49 private: 50 bool EliminateMostlyEmptyBlocks(Function &F); 51 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; 52 void EliminateMostlyEmptyBlock(BasicBlock *BB); 53 bool OptimizeBlock(BasicBlock &BB); 54 bool OptimizeLoadStoreInst(Instruction *I, Value *Addr, 55 const Type *AccessTy, 56 DenseMap<Value*,Value*> &SunkAddrs); 57 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS, 58 DenseMap<Value*,Value*> &SunkAddrs); 59 bool OptimizeExtUses(Instruction *I); 60 }; 61} 62 63char CodeGenPrepare::ID = 0; 64static RegisterPass<CodeGenPrepare> X("codegenprepare", 65 "Optimize for code generation"); 66 67FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) { 68 return new CodeGenPrepare(TLI); 69} 70 71 72bool CodeGenPrepare::runOnFunction(Function &F) { 73 bool EverMadeChange = false; 74 75 // First pass, eliminate blocks that contain only PHI nodes and an 76 // unconditional branch. 77 EverMadeChange |= EliminateMostlyEmptyBlocks(F); 78 79 bool MadeChange = true; 80 while (MadeChange) { 81 MadeChange = false; 82 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 83 MadeChange |= OptimizeBlock(*BB); 84 EverMadeChange |= MadeChange; 85 } 86 return EverMadeChange; 87} 88 89/// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes 90/// and an unconditional branch. Passes before isel (e.g. LSR/loopsimplify) 91/// often split edges in ways that are non-optimal for isel. Start by 92/// eliminating these blocks so we can split them the way we want them. 93bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) { 94 bool MadeChange = false; 95 // Note that this intentionally skips the entry block. 96 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) { 97 BasicBlock *BB = I++; 98 99 // If this block doesn't end with an uncond branch, ignore it. 100 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 101 if (!BI || !BI->isUnconditional()) 102 continue; 103 104 // If the instruction before the branch isn't a phi node, then other stuff 105 // is happening here. 106 BasicBlock::iterator BBI = BI; 107 if (BBI != BB->begin()) { 108 --BBI; 109 if (!isa<PHINode>(BBI)) continue; 110 } 111 112 // Do not break infinite loops. 113 BasicBlock *DestBB = BI->getSuccessor(0); 114 if (DestBB == BB) 115 continue; 116 117 if (!CanMergeBlocks(BB, DestBB)) 118 continue; 119 120 EliminateMostlyEmptyBlock(BB); 121 MadeChange = true; 122 } 123 return MadeChange; 124} 125 126/// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a 127/// single uncond branch between them, and BB contains no other non-phi 128/// instructions. 129bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB, 130 const BasicBlock *DestBB) const { 131 // We only want to eliminate blocks whose phi nodes are used by phi nodes in 132 // the successor. If there are more complex condition (e.g. preheaders), 133 // don't mess around with them. 134 BasicBlock::const_iterator BBI = BB->begin(); 135 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 136 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end(); 137 UI != E; ++UI) { 138 const Instruction *User = cast<Instruction>(*UI); 139 if (User->getParent() != DestBB || !isa<PHINode>(User)) 140 return false; 141 // If User is inside DestBB block and it is a PHINode then check 142 // incoming value. If incoming value is not from BB then this is 143 // a complex condition (e.g. preheaders) we want to avoid here. 144 if (User->getParent() == DestBB) { 145 if (const PHINode *UPN = dyn_cast<PHINode>(User)) 146 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { 147 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); 148 if (Insn && Insn->getParent() == BB && 149 Insn->getParent() != UPN->getIncomingBlock(I)) 150 return false; 151 } 152 } 153 } 154 } 155 156 // If BB and DestBB contain any common predecessors, then the phi nodes in BB 157 // and DestBB may have conflicting incoming values for the block. If so, we 158 // can't merge the block. 159 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); 160 if (!DestBBPN) return true; // no conflict. 161 162 // Collect the preds of BB. 163 SmallPtrSet<const BasicBlock*, 16> BBPreds; 164 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 165 // It is faster to get preds from a PHI than with pred_iterator. 166 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 167 BBPreds.insert(BBPN->getIncomingBlock(i)); 168 } else { 169 BBPreds.insert(pred_begin(BB), pred_end(BB)); 170 } 171 172 // Walk the preds of DestBB. 173 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { 174 BasicBlock *Pred = DestBBPN->getIncomingBlock(i); 175 if (BBPreds.count(Pred)) { // Common predecessor? 176 BBI = DestBB->begin(); 177 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 178 const Value *V1 = PN->getIncomingValueForBlock(Pred); 179 const Value *V2 = PN->getIncomingValueForBlock(BB); 180 181 // If V2 is a phi node in BB, look up what the mapped value will be. 182 if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) 183 if (V2PN->getParent() == BB) 184 V2 = V2PN->getIncomingValueForBlock(Pred); 185 186 // If there is a conflict, bail out. 187 if (V1 != V2) return false; 188 } 189 } 190 } 191 192 return true; 193} 194 195 196/// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and 197/// an unconditional branch in it. 198void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { 199 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 200 BasicBlock *DestBB = BI->getSuccessor(0); 201 202 DOUT << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB; 203 204 // If the destination block has a single pred, then this is a trivial edge, 205 // just collapse it. 206 if (DestBB->getSinglePredecessor()) { 207 // If DestBB has single-entry PHI nodes, fold them. 208 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { 209 Value *NewVal = PN->getIncomingValue(0); 210 // Replace self referencing PHI with undef, it must be dead. 211 if (NewVal == PN) NewVal = UndefValue::get(PN->getType()); 212 PN->replaceAllUsesWith(NewVal); 213 PN->eraseFromParent(); 214 } 215 216 // Splice all the PHI nodes from BB over to DestBB. 217 DestBB->getInstList().splice(DestBB->begin(), BB->getInstList(), 218 BB->begin(), BI); 219 220 // Anything that branched to BB now branches to DestBB. 221 BB->replaceAllUsesWith(DestBB); 222 223 // Nuke BB. 224 BB->eraseFromParent(); 225 226 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 227 return; 228 } 229 230 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB 231 // to handle the new incoming edges it is about to have. 232 PHINode *PN; 233 for (BasicBlock::iterator BBI = DestBB->begin(); 234 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 235 // Remove the incoming value for BB, and remember it. 236 Value *InVal = PN->removeIncomingValue(BB, false); 237 238 // Two options: either the InVal is a phi node defined in BB or it is some 239 // value that dominates BB. 240 PHINode *InValPhi = dyn_cast<PHINode>(InVal); 241 if (InValPhi && InValPhi->getParent() == BB) { 242 // Add all of the input values of the input PHI as inputs of this phi. 243 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) 244 PN->addIncoming(InValPhi->getIncomingValue(i), 245 InValPhi->getIncomingBlock(i)); 246 } else { 247 // Otherwise, add one instance of the dominating value for each edge that 248 // we will be adding. 249 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 250 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 251 PN->addIncoming(InVal, BBPN->getIncomingBlock(i)); 252 } else { 253 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 254 PN->addIncoming(InVal, *PI); 255 } 256 } 257 } 258 259 // The PHIs are now updated, change everything that refers to BB to use 260 // DestBB and remove BB. 261 BB->replaceAllUsesWith(DestBB); 262 BB->eraseFromParent(); 263 264 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 265} 266 267 268/// SplitEdgeNicely - Split the critical edge from TI to its specified 269/// successor if it will improve codegen. We only do this if the successor has 270/// phi nodes (otherwise critical edges are ok). If there is already another 271/// predecessor of the succ that is empty (and thus has no phi nodes), use it 272/// instead of introducing a new block. 273static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, Pass *P) { 274 BasicBlock *TIBB = TI->getParent(); 275 BasicBlock *Dest = TI->getSuccessor(SuccNum); 276 assert(isa<PHINode>(Dest->begin()) && 277 "This should only be called if Dest has a PHI!"); 278 279 // As a hack, never split backedges of loops. Even though the copy for any 280 // PHIs inserted on the backedge would be dead for exits from the loop, we 281 // assume that the cost of *splitting* the backedge would be too high. 282 if (Dest == TIBB) 283 return; 284 285 /// TIPHIValues - This array is lazily computed to determine the values of 286 /// PHIs in Dest that TI would provide. 287 SmallVector<Value*, 32> TIPHIValues; 288 289 // Check to see if Dest has any blocks that can be used as a split edge for 290 // this terminator. 291 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { 292 BasicBlock *Pred = *PI; 293 // To be usable, the pred has to end with an uncond branch to the dest. 294 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); 295 if (!PredBr || !PredBr->isUnconditional() || 296 // Must be empty other than the branch. 297 &Pred->front() != PredBr || 298 // Cannot be the entry block; its label does not get emitted. 299 Pred == &(Dest->getParent()->getEntryBlock())) 300 continue; 301 302 // Finally, since we know that Dest has phi nodes in it, we have to make 303 // sure that jumping to Pred will have the same affect as going to Dest in 304 // terms of PHI values. 305 PHINode *PN; 306 unsigned PHINo = 0; 307 bool FoundMatch = true; 308 for (BasicBlock::iterator I = Dest->begin(); 309 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { 310 if (PHINo == TIPHIValues.size()) 311 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); 312 313 // If the PHI entry doesn't work, we can't use this pred. 314 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { 315 FoundMatch = false; 316 break; 317 } 318 } 319 320 // If we found a workable predecessor, change TI to branch to Succ. 321 if (FoundMatch) { 322 Dest->removePredecessor(TIBB); 323 TI->setSuccessor(SuccNum, Pred); 324 return; 325 } 326 } 327 328 SplitCriticalEdge(TI, SuccNum, P, true); 329} 330 331/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop 332/// copy (e.g. it's casting from one pointer type to another, int->uint, or 333/// int->sbyte on PPC), sink it into user blocks to reduce the number of virtual 334/// registers that must be created and coalesced. 335/// 336/// Return true if any changes are made. 337static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ 338 // If this is a noop copy, 339 MVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 340 MVT DstVT = TLI.getValueType(CI->getType()); 341 342 // This is an fp<->int conversion? 343 if (SrcVT.isInteger() != DstVT.isInteger()) 344 return false; 345 346 // If this is an extension, it will be a zero or sign extension, which 347 // isn't a noop. 348 if (SrcVT.bitsLT(DstVT)) return false; 349 350 // If these values will be promoted, find out what they will be promoted 351 // to. This helps us consider truncates on PPC as noop copies when they 352 // are. 353 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote) 354 SrcVT = TLI.getTypeToTransformTo(SrcVT); 355 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote) 356 DstVT = TLI.getTypeToTransformTo(DstVT); 357 358 // If, after promotion, these are the same types, this is a noop copy. 359 if (SrcVT != DstVT) 360 return false; 361 362 BasicBlock *DefBB = CI->getParent(); 363 364 /// InsertedCasts - Only insert a cast in each block once. 365 DenseMap<BasicBlock*, CastInst*> InsertedCasts; 366 367 bool MadeChange = false; 368 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 369 UI != E; ) { 370 Use &TheUse = UI.getUse(); 371 Instruction *User = cast<Instruction>(*UI); 372 373 // Figure out which BB this cast is used in. For PHI's this is the 374 // appropriate predecessor block. 375 BasicBlock *UserBB = User->getParent(); 376 if (PHINode *PN = dyn_cast<PHINode>(User)) { 377 unsigned OpVal = UI.getOperandNo()/2; 378 UserBB = PN->getIncomingBlock(OpVal); 379 } 380 381 // Preincrement use iterator so we don't invalidate it. 382 ++UI; 383 384 // If this user is in the same block as the cast, don't change the cast. 385 if (UserBB == DefBB) continue; 386 387 // If we have already inserted a cast into this block, use it. 388 CastInst *&InsertedCast = InsertedCasts[UserBB]; 389 390 if (!InsertedCast) { 391 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 392 393 InsertedCast = 394 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 395 InsertPt); 396 MadeChange = true; 397 } 398 399 // Replace a use of the cast with a use of the new cast. 400 TheUse = InsertedCast; 401 } 402 403 // If we removed all uses, nuke the cast. 404 if (CI->use_empty()) { 405 CI->eraseFromParent(); 406 MadeChange = true; 407 } 408 409 return MadeChange; 410} 411 412/// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce 413/// the number of virtual registers that must be created and coalesced. This is 414/// a clear win except on targets with multiple condition code registers 415/// (PowerPC), where it might lose; some adjustment may be wanted there. 416/// 417/// Return true if any changes are made. 418static bool OptimizeCmpExpression(CmpInst *CI){ 419 420 BasicBlock *DefBB = CI->getParent(); 421 422 /// InsertedCmp - Only insert a cmp in each block once. 423 DenseMap<BasicBlock*, CmpInst*> InsertedCmps; 424 425 bool MadeChange = false; 426 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 427 UI != E; ) { 428 Use &TheUse = UI.getUse(); 429 Instruction *User = cast<Instruction>(*UI); 430 431 // Preincrement use iterator so we don't invalidate it. 432 ++UI; 433 434 // Don't bother for PHI nodes. 435 if (isa<PHINode>(User)) 436 continue; 437 438 // Figure out which BB this cmp is used in. 439 BasicBlock *UserBB = User->getParent(); 440 441 // If this user is in the same block as the cmp, don't change the cmp. 442 if (UserBB == DefBB) continue; 443 444 // If we have already inserted a cmp into this block, use it. 445 CmpInst *&InsertedCmp = InsertedCmps[UserBB]; 446 447 if (!InsertedCmp) { 448 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 449 450 InsertedCmp = 451 CmpInst::Create(CI->getOpcode(), CI->getPredicate(), CI->getOperand(0), 452 CI->getOperand(1), "", InsertPt); 453 MadeChange = true; 454 } 455 456 // Replace a use of the cmp with a use of the new cmp. 457 TheUse = InsertedCmp; 458 } 459 460 // If we removed all uses, nuke the cmp. 461 if (CI->use_empty()) 462 CI->eraseFromParent(); 463 464 return MadeChange; 465} 466 467/// EraseDeadInstructions - Erase any dead instructions 468static void EraseDeadInstructions(Value *V) { 469 Instruction *I = dyn_cast<Instruction>(V); 470 if (!I || !I->use_empty()) return; 471 472 SmallPtrSet<Instruction*, 16> Insts; 473 Insts.insert(I); 474 475 while (!Insts.empty()) { 476 I = *Insts.begin(); 477 Insts.erase(I); 478 if (isInstructionTriviallyDead(I)) { 479 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 480 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i))) 481 Insts.insert(U); 482 I->eraseFromParent(); 483 } 484 } 485} 486 487namespace { 488 489/// ExtAddrMode - This is an extended version of TargetLowering::AddrMode which 490/// holds actual Value*'s for register values. 491struct ExtAddrMode : public TargetLowering::AddrMode { 492 Value *BaseReg; 493 Value *ScaledReg; 494 ExtAddrMode() : BaseReg(0), ScaledReg(0) {} 495 void dump() const; 496}; 497 498static std::ostream &operator<<(std::ostream &OS, const ExtAddrMode &AM) { 499 bool NeedPlus = false; 500 OS << "["; 501 if (AM.BaseGV) 502 OS << (NeedPlus ? " + " : "") 503 << "GV:%" << AM.BaseGV->getName(), NeedPlus = true; 504 505 if (AM.BaseOffs) 506 OS << (NeedPlus ? " + " : "") << AM.BaseOffs, NeedPlus = true; 507 508 if (AM.BaseReg) 509 OS << (NeedPlus ? " + " : "") 510 << "Base:%" << AM.BaseReg->getName(), NeedPlus = true; 511 if (AM.Scale) 512 OS << (NeedPlus ? " + " : "") 513 << AM.Scale << "*%" << AM.ScaledReg->getName(), NeedPlus = true; 514 515 return OS << "]"; 516} 517 518void ExtAddrMode::dump() const { 519 cerr << *this << "\n"; 520} 521 522} 523 524static bool TryMatchingScaledValue(Value *ScaleReg, int64_t Scale, 525 const Type *AccessTy, ExtAddrMode &AddrMode, 526 SmallVector<Instruction*, 16> &AddrModeInsts, 527 const TargetLowering &TLI, unsigned Depth); 528 529/// FindMaximalLegalAddressingMode - If we can, try to merge the computation of 530/// Addr into the specified addressing mode. If Addr can't be added to AddrMode 531/// this returns false. This assumes that Addr is either a pointer type or 532/// intptr_t for the target. 533static bool FindMaximalLegalAddressingMode(Value *Addr, const Type *AccessTy, 534 ExtAddrMode &AddrMode, 535 SmallVector<Instruction*, 16> &AddrModeInsts, 536 const TargetLowering &TLI, 537 unsigned Depth) { 538 539 // If this is a global variable, fold it into the addressing mode if possible. 540 if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) { 541 if (AddrMode.BaseGV == 0) { 542 AddrMode.BaseGV = GV; 543 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 544 return true; 545 AddrMode.BaseGV = 0; 546 } 547 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) { 548 AddrMode.BaseOffs += CI->getSExtValue(); 549 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 550 return true; 551 AddrMode.BaseOffs -= CI->getSExtValue(); 552 } else if (isa<ConstantPointerNull>(Addr)) { 553 return true; 554 } 555 556 // Look through constant exprs and instructions. 557 unsigned Opcode = ~0U; 558 User *AddrInst = 0; 559 if (Instruction *I = dyn_cast<Instruction>(Addr)) { 560 Opcode = I->getOpcode(); 561 AddrInst = I; 562 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) { 563 Opcode = CE->getOpcode(); 564 AddrInst = CE; 565 } 566 567 // Limit recursion to avoid exponential behavior. 568 if (Depth == 5) { AddrInst = 0; Opcode = ~0U; } 569 570 // If this is really an instruction, add it to our list of related 571 // instructions. 572 if (Instruction *I = dyn_cast_or_null<Instruction>(AddrInst)) 573 AddrModeInsts.push_back(I); 574 575 if (AddrInst && !AddrInst->hasOneUse()) 576 ; 577 else 578 switch (Opcode) { 579 case Instruction::PtrToInt: 580 // PtrToInt is always a noop, as we know that the int type is pointer sized. 581 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 582 AddrMode, AddrModeInsts, TLI, Depth)) 583 return true; 584 break; 585 case Instruction::IntToPtr: 586 // This inttoptr is a no-op if the integer type is pointer sized. 587 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) == 588 TLI.getPointerTy()) { 589 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 590 AddrMode, AddrModeInsts, TLI, Depth)) 591 return true; 592 } 593 break; 594 case Instruction::Add: { 595 // Check to see if we can merge in the RHS then the LHS. If so, we win. 596 ExtAddrMode BackupAddrMode = AddrMode; 597 unsigned OldSize = AddrModeInsts.size(); 598 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(1), AccessTy, 599 AddrMode, AddrModeInsts, TLI, Depth+1) && 600 FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 601 AddrMode, AddrModeInsts, TLI, Depth+1)) 602 return true; 603 604 // Restore the old addr mode info. 605 AddrMode = BackupAddrMode; 606 AddrModeInsts.resize(OldSize); 607 608 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS. 609 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 610 AddrMode, AddrModeInsts, TLI, Depth+1) && 611 FindMaximalLegalAddressingMode(AddrInst->getOperand(1), AccessTy, 612 AddrMode, AddrModeInsts, TLI, Depth+1)) 613 return true; 614 615 // Otherwise we definitely can't merge the ADD in. 616 AddrMode = BackupAddrMode; 617 AddrModeInsts.resize(OldSize); 618 break; 619 } 620 case Instruction::Or: { 621 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); 622 if (!RHS) break; 623 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD. 624 break; 625 } 626 case Instruction::Mul: 627 case Instruction::Shl: { 628 // Can only handle X*C and X << C, and can only handle this when the scale 629 // field is available. 630 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); 631 if (!RHS) break; 632 int64_t Scale = RHS->getSExtValue(); 633 if (Opcode == Instruction::Shl) 634 Scale = 1 << Scale; 635 636 if (TryMatchingScaledValue(AddrInst->getOperand(0), Scale, AccessTy, 637 AddrMode, AddrModeInsts, TLI, Depth)) 638 return true; 639 break; 640 } 641 case Instruction::GetElementPtr: { 642 // Scan the GEP. We check it if it contains constant offsets and at most 643 // one variable offset. 644 int VariableOperand = -1; 645 unsigned VariableScale = 0; 646 647 int64_t ConstantOffset = 0; 648 const TargetData *TD = TLI.getTargetData(); 649 gep_type_iterator GTI = gep_type_begin(AddrInst); 650 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) { 651 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 652 const StructLayout *SL = TD->getStructLayout(STy); 653 unsigned Idx = 654 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue(); 655 ConstantOffset += SL->getElementOffset(Idx); 656 } else { 657 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType()); 658 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) { 659 ConstantOffset += CI->getSExtValue()*TypeSize; 660 } else if (TypeSize) { // Scales of zero don't do anything. 661 // We only allow one variable index at the moment. 662 if (VariableOperand != -1) { 663 VariableOperand = -2; 664 break; 665 } 666 667 // Remember the variable index. 668 VariableOperand = i; 669 VariableScale = TypeSize; 670 } 671 } 672 } 673 674 // If the GEP had multiple variable indices, punt. 675 if (VariableOperand == -2) 676 break; 677 678 // A common case is for the GEP to only do a constant offset. In this case, 679 // just add it to the disp field and check validity. 680 if (VariableOperand == -1) { 681 AddrMode.BaseOffs += ConstantOffset; 682 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){ 683 // Check to see if we can fold the base pointer in too. 684 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 685 AddrMode, AddrModeInsts, TLI, 686 Depth+1)) 687 return true; 688 } 689 AddrMode.BaseOffs -= ConstantOffset; 690 } else { 691 // Check that this has no base reg yet. If so, we won't have a place to 692 // put the base of the GEP (assuming it is not a null ptr). 693 bool SetBaseReg = false; 694 if (AddrMode.HasBaseReg) { 695 if (!isa<ConstantPointerNull>(AddrInst->getOperand(0))) 696 break; 697 } else { 698 AddrMode.HasBaseReg = true; 699 AddrMode.BaseReg = AddrInst->getOperand(0); 700 SetBaseReg = true; 701 } 702 703 // See if the scale amount is valid for this target. 704 AddrMode.BaseOffs += ConstantOffset; 705 if (TryMatchingScaledValue(AddrInst->getOperand(VariableOperand), 706 VariableScale, AccessTy, AddrMode, 707 AddrModeInsts, TLI, Depth)) { 708 if (!SetBaseReg) return true; 709 710 // If this match succeeded, we know that we can form an address with the 711 // GepBase as the basereg. See if we can match *more*. 712 AddrMode.HasBaseReg = false; 713 AddrMode.BaseReg = 0; 714 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 715 AddrMode, AddrModeInsts, TLI, 716 Depth+1)) 717 return true; 718 // Strange, shouldn't happen. Restore the base reg and succeed the easy 719 // way. 720 AddrMode.HasBaseReg = true; 721 AddrMode.BaseReg = AddrInst->getOperand(0); 722 return true; 723 } 724 725 AddrMode.BaseOffs -= ConstantOffset; 726 if (SetBaseReg) { 727 AddrMode.HasBaseReg = false; 728 AddrMode.BaseReg = 0; 729 } 730 } 731 break; 732 } 733 } 734 735 if (Instruction *I = dyn_cast_or_null<Instruction>(AddrInst)) { 736 assert(AddrModeInsts.back() == I && "Stack imbalance"); I = I; 737 AddrModeInsts.pop_back(); 738 } 739 740 // Worse case, the target should support [reg] addressing modes. :) 741 if (!AddrMode.HasBaseReg) { 742 AddrMode.HasBaseReg = true; 743 // Still check for legality in case the target supports [imm] but not [i+r]. 744 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { 745 AddrMode.BaseReg = Addr; 746 return true; 747 } 748 AddrMode.HasBaseReg = false; 749 } 750 751 // If the base register is already taken, see if we can do [r+r]. 752 if (AddrMode.Scale == 0) { 753 AddrMode.Scale = 1; 754 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { 755 AddrMode.ScaledReg = Addr; 756 return true; 757 } 758 AddrMode.Scale = 0; 759 } 760 // Couldn't match. 761 return false; 762} 763 764/// TryMatchingScaledValue - Try adding ScaleReg*Scale to the specified 765/// addressing mode. Return true if this addr mode is legal for the target, 766/// false if not. 767static bool TryMatchingScaledValue(Value *ScaleReg, int64_t Scale, 768 const Type *AccessTy, ExtAddrMode &AddrMode, 769 SmallVector<Instruction*, 16> &AddrModeInsts, 770 const TargetLowering &TLI, unsigned Depth) { 771 // If we already have a scale of this value, we can add to it, otherwise, we 772 // need an available scale field. 773 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg) 774 return false; 775 776 ExtAddrMode InputAddrMode = AddrMode; 777 778 // Add scale to turn X*4+X*3 -> X*7. This could also do things like 779 // [A+B + A*7] -> [B+A*8]. 780 AddrMode.Scale += Scale; 781 AddrMode.ScaledReg = ScaleReg; 782 783 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { 784 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now 785 // to see if ScaleReg is actually X+C. If so, we can turn this into adding 786 // X*Scale + C*Scale to addr mode. 787 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(ScaleReg); 788 if (BinOp && BinOp->getOpcode() == Instruction::Add && 789 isa<ConstantInt>(BinOp->getOperand(1)) && InputAddrMode.ScaledReg ==0) { 790 791 InputAddrMode.Scale = Scale; 792 InputAddrMode.ScaledReg = BinOp->getOperand(0); 793 InputAddrMode.BaseOffs += 794 cast<ConstantInt>(BinOp->getOperand(1))->getSExtValue()*Scale; 795 if (TLI.isLegalAddressingMode(InputAddrMode, AccessTy)) { 796 AddrModeInsts.push_back(BinOp); 797 AddrMode = InputAddrMode; 798 return true; 799 } 800 } 801 802 // Otherwise, not (x+c)*scale, just return what we have. 803 return true; 804 } 805 806 // Otherwise, back this attempt out. 807 AddrMode.Scale -= Scale; 808 if (AddrMode.Scale == 0) AddrMode.ScaledReg = 0; 809 810 return false; 811} 812 813 814/// IsNonLocalValue - Return true if the specified values are defined in a 815/// different basic block than BB. 816static bool IsNonLocalValue(Value *V, BasicBlock *BB) { 817 if (Instruction *I = dyn_cast<Instruction>(V)) 818 return I->getParent() != BB; 819 return false; 820} 821 822/// OptimizeLoadStoreInst - Load and Store Instructions have often have 823/// addressing modes that can do significant amounts of computation. As such, 824/// instruction selection will try to get the load or store to do as much 825/// computation as possible for the program. The problem is that isel can only 826/// see within a single block. As such, we sink as much legal addressing mode 827/// stuff into the block as possible. 828bool CodeGenPrepare::OptimizeLoadStoreInst(Instruction *LdStInst, Value *Addr, 829 const Type *AccessTy, 830 DenseMap<Value*,Value*> &SunkAddrs) { 831 // Figure out what addressing mode will be built up for this operation. 832 SmallVector<Instruction*, 16> AddrModeInsts; 833 ExtAddrMode AddrMode; 834 bool Success = FindMaximalLegalAddressingMode(Addr, AccessTy, AddrMode, 835 AddrModeInsts, *TLI, 0); 836 Success = Success; assert(Success && "Couldn't select *anything*?"); 837 838 // Check to see if any of the instructions supersumed by this addr mode are 839 // non-local to I's BB. 840 bool AnyNonLocal = false; 841 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { 842 if (IsNonLocalValue(AddrModeInsts[i], LdStInst->getParent())) { 843 AnyNonLocal = true; 844 break; 845 } 846 } 847 848 // If all the instructions matched are already in this BB, don't do anything. 849 if (!AnyNonLocal) { 850 DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n"); 851 return false; 852 } 853 854 // Insert this computation right after this user. Since our caller is 855 // scanning from the top of the BB to the bottom, reuse of the expr are 856 // guaranteed to happen later. 857 BasicBlock::iterator InsertPt = LdStInst; 858 859 // Now that we determined the addressing expression we want to use and know 860 // that we have to sink it into this block. Check to see if we have already 861 // done this for some other load/store instr in this block. If so, reuse the 862 // computation. 863 Value *&SunkAddr = SunkAddrs[Addr]; 864 if (SunkAddr) { 865 DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << "\n"); 866 if (SunkAddr->getType() != Addr->getType()) 867 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt); 868 } else { 869 DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << "\n"); 870 const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType(); 871 872 Value *Result = 0; 873 // Start with the scale value. 874 if (AddrMode.Scale) { 875 Value *V = AddrMode.ScaledReg; 876 if (V->getType() == IntPtrTy) { 877 // done. 878 } else if (isa<PointerType>(V->getType())) { 879 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 880 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 881 cast<IntegerType>(V->getType())->getBitWidth()) { 882 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt); 883 } else { 884 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt); 885 } 886 if (AddrMode.Scale != 1) 887 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy, 888 AddrMode.Scale), 889 "sunkaddr", InsertPt); 890 Result = V; 891 } 892 893 // Add in the base register. 894 if (AddrMode.BaseReg) { 895 Value *V = AddrMode.BaseReg; 896 if (V->getType() != IntPtrTy) 897 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 898 if (Result) 899 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 900 else 901 Result = V; 902 } 903 904 // Add in the BaseGV if present. 905 if (AddrMode.BaseGV) { 906 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr", 907 InsertPt); 908 if (Result) 909 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 910 else 911 Result = V; 912 } 913 914 // Add in the Base Offset if present. 915 if (AddrMode.BaseOffs) { 916 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 917 if (Result) 918 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 919 else 920 Result = V; 921 } 922 923 if (Result == 0) 924 SunkAddr = Constant::getNullValue(Addr->getType()); 925 else 926 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt); 927 } 928 929 LdStInst->replaceUsesOfWith(Addr, SunkAddr); 930 931 if (Addr->use_empty()) 932 EraseDeadInstructions(Addr); 933 return true; 934} 935 936/// OptimizeInlineAsmInst - If there are any memory operands, use 937/// OptimizeLoadStoreInt to sink their address computing into the block when 938/// possible / profitable. 939bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS, 940 DenseMap<Value*,Value*> &SunkAddrs) { 941 bool MadeChange = false; 942 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 943 944 // Do a prepass over the constraints, canonicalizing them, and building up the 945 // ConstraintOperands list. 946 std::vector<InlineAsm::ConstraintInfo> 947 ConstraintInfos = IA->ParseConstraints(); 948 949 /// ConstraintOperands - Information about all of the constraints. 950 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands; 951 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 952 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 953 ConstraintOperands. 954 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i])); 955 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back(); 956 957 // Compute the value type for each operand. 958 switch (OpInfo.Type) { 959 case InlineAsm::isOutput: 960 if (OpInfo.isIndirect) 961 OpInfo.CallOperandVal = CS.getArgument(ArgNo++); 962 break; 963 case InlineAsm::isInput: 964 OpInfo.CallOperandVal = CS.getArgument(ArgNo++); 965 break; 966 case InlineAsm::isClobber: 967 // Nothing to do. 968 break; 969 } 970 971 // Compute the constraint code and ConstraintType to use. 972 TLI->ComputeConstraintToUse(OpInfo, SDValue(), 973 OpInfo.ConstraintType == TargetLowering::C_Memory); 974 975 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 976 OpInfo.isIndirect) { 977 Value *OpVal = OpInfo.CallOperandVal; 978 MadeChange |= OptimizeLoadStoreInst(I, OpVal, OpVal->getType(), 979 SunkAddrs); 980 } 981 } 982 983 return MadeChange; 984} 985 986bool CodeGenPrepare::OptimizeExtUses(Instruction *I) { 987 BasicBlock *DefBB = I->getParent(); 988 989 // If both result of the {s|z}xt and its source are live out, rewrite all 990 // other uses of the source with result of extension. 991 Value *Src = I->getOperand(0); 992 if (Src->hasOneUse()) 993 return false; 994 995 // Only do this xform if truncating is free. 996 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType())) 997 return false; 998 999 // Only safe to perform the optimization if the source is also defined in 1000 // this block. 1001 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) 1002 return false; 1003 1004 bool DefIsLiveOut = false; 1005 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 1006 UI != E; ++UI) { 1007 Instruction *User = cast<Instruction>(*UI); 1008 1009 // Figure out which BB this ext is used in. 1010 BasicBlock *UserBB = User->getParent(); 1011 if (UserBB == DefBB) continue; 1012 DefIsLiveOut = true; 1013 break; 1014 } 1015 if (!DefIsLiveOut) 1016 return false; 1017 1018 // Make sure non of the uses are PHI nodes. 1019 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 1020 UI != E; ++UI) { 1021 Instruction *User = cast<Instruction>(*UI); 1022 BasicBlock *UserBB = User->getParent(); 1023 if (UserBB == DefBB) continue; 1024 // Be conservative. We don't want this xform to end up introducing 1025 // reloads just before load / store instructions. 1026 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User)) 1027 return false; 1028 } 1029 1030 // InsertedTruncs - Only insert one trunc in each block once. 1031 DenseMap<BasicBlock*, Instruction*> InsertedTruncs; 1032 1033 bool MadeChange = false; 1034 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 1035 UI != E; ++UI) { 1036 Use &TheUse = UI.getUse(); 1037 Instruction *User = cast<Instruction>(*UI); 1038 1039 // Figure out which BB this ext is used in. 1040 BasicBlock *UserBB = User->getParent(); 1041 if (UserBB == DefBB) continue; 1042 1043 // Both src and def are live in this block. Rewrite the use. 1044 Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; 1045 1046 if (!InsertedTrunc) { 1047 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 1048 1049 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt); 1050 } 1051 1052 // Replace a use of the {s|z}ext source with a use of the result. 1053 TheUse = InsertedTrunc; 1054 1055 MadeChange = true; 1056 } 1057 1058 return MadeChange; 1059} 1060 1061// In this pass we look for GEP and cast instructions that are used 1062// across basic blocks and rewrite them to improve basic-block-at-a-time 1063// selection. 1064bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { 1065 bool MadeChange = false; 1066 1067 // Split all critical edges where the dest block has a PHI and where the phi 1068 // has shared immediate operands. 1069 TerminatorInst *BBTI = BB.getTerminator(); 1070 if (BBTI->getNumSuccessors() > 1) { 1071 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) 1072 if (isa<PHINode>(BBTI->getSuccessor(i)->begin()) && 1073 isCriticalEdge(BBTI, i, true)) 1074 SplitEdgeNicely(BBTI, i, this); 1075 } 1076 1077 1078 // Keep track of non-local addresses that have been sunk into this block. 1079 // This allows us to avoid inserting duplicate code for blocks with multiple 1080 // load/stores of the same address. 1081 DenseMap<Value*, Value*> SunkAddrs; 1082 1083 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) { 1084 Instruction *I = BBI++; 1085 1086 if (CastInst *CI = dyn_cast<CastInst>(I)) { 1087 // If the source of the cast is a constant, then this should have 1088 // already been constant folded. The only reason NOT to constant fold 1089 // it is if something (e.g. LSR) was careful to place the constant 1090 // evaluation in a block other than then one that uses it (e.g. to hoist 1091 // the address of globals out of a loop). If this is the case, we don't 1092 // want to forward-subst the cast. 1093 if (isa<Constant>(CI->getOperand(0))) 1094 continue; 1095 1096 bool Change = false; 1097 if (TLI) { 1098 Change = OptimizeNoopCopyExpression(CI, *TLI); 1099 MadeChange |= Change; 1100 } 1101 1102 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) 1103 MadeChange |= OptimizeExtUses(I); 1104 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 1105 MadeChange |= OptimizeCmpExpression(CI); 1106 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1107 if (TLI) 1108 MadeChange |= OptimizeLoadStoreInst(I, I->getOperand(0), LI->getType(), 1109 SunkAddrs); 1110 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 1111 if (TLI) 1112 MadeChange |= OptimizeLoadStoreInst(I, SI->getOperand(1), 1113 SI->getOperand(0)->getType(), 1114 SunkAddrs); 1115 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 1116 if (GEPI->hasAllZeroIndices()) { 1117 /// The GEP operand must be a pointer, so must its result -> BitCast 1118 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 1119 GEPI->getName(), GEPI); 1120 GEPI->replaceAllUsesWith(NC); 1121 GEPI->eraseFromParent(); 1122 MadeChange = true; 1123 BBI = NC; 1124 } 1125 } else if (CallInst *CI = dyn_cast<CallInst>(I)) { 1126 // If we found an inline asm expession, and if the target knows how to 1127 // lower it to normal LLVM code, do so now. 1128 if (TLI && isa<InlineAsm>(CI->getCalledValue())) 1129 if (const TargetAsmInfo *TAI = 1130 TLI->getTargetMachine().getTargetAsmInfo()) { 1131 if (TAI->ExpandInlineAsm(CI)) 1132 BBI = BB.begin(); 1133 else 1134 // Sink address computing for memory operands into the block. 1135 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs); 1136 } 1137 } 1138 } 1139 1140 return MadeChange; 1141} 1142