CodeGenPrepare.cpp revision f9785f92b630e69262c395b2fc0893451169d68b
1//===- CodeGenPrepare.cpp - Prepare a function for code generation --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by Chris Lattner and is distributed under 6// the University of Illinois Open Source 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/Instructions.h" 22#include "llvm/Pass.h" 23#include "llvm/Target/TargetAsmInfo.h" 24#include "llvm/Target/TargetData.h" 25#include "llvm/Target/TargetLowering.h" 26#include "llvm/Target/TargetMachine.h" 27#include "llvm/Transforms/Utils/BasicBlockUtils.h" 28#include "llvm/Transforms/Utils/Local.h" 29#include "llvm/ADT/DenseMap.h" 30#include "llvm/ADT/SmallSet.h" 31#include "llvm/Support/CommandLine.h" 32#include "llvm/Support/Compiler.h" 33#include "llvm/Support/Debug.h" 34#include "llvm/Support/GetElementPtrTypeIterator.h" 35using namespace llvm; 36 37namespace { 38 cl::opt<bool> OptExtUses("optimize-ext-uses", 39 cl::init(true), cl::Hidden); 40} 41 42namespace { 43 class VISIBILITY_HIDDEN CodeGenPrepare : public FunctionPass { 44 /// TLI - Keep a pointer of a TargetLowering to consult for determining 45 /// transformation profitability. 46 const TargetLowering *TLI; 47 public: 48 static char ID; // Pass identification, replacement for typeid 49 explicit CodeGenPrepare(const TargetLowering *tli = 0) 50 : FunctionPass((intptr_t)&ID), TLI(tli) {} 51 bool runOnFunction(Function &F); 52 53 private: 54 bool EliminateMostlyEmptyBlocks(Function &F); 55 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; 56 void EliminateMostlyEmptyBlock(BasicBlock *BB); 57 bool OptimizeBlock(BasicBlock &BB); 58 bool OptimizeLoadStoreInst(Instruction *I, Value *Addr, 59 const Type *AccessTy, 60 DenseMap<Value*,Value*> &SunkAddrs); 61 bool OptimizeExtUses(Instruction *I); 62 }; 63} 64 65char CodeGenPrepare::ID = 0; 66static RegisterPass<CodeGenPrepare> X("codegenprepare", 67 "Optimize for code generation"); 68 69FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) { 70 return new CodeGenPrepare(TLI); 71} 72 73 74bool CodeGenPrepare::runOnFunction(Function &F) { 75 bool EverMadeChange = false; 76 77 // First pass, eliminate blocks that contain only PHI nodes and an 78 // unconditional branch. 79 EverMadeChange |= EliminateMostlyEmptyBlocks(F); 80 81 bool MadeChange = true; 82 while (MadeChange) { 83 MadeChange = false; 84 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 85 MadeChange |= OptimizeBlock(*BB); 86 EverMadeChange |= MadeChange; 87 } 88 return EverMadeChange; 89} 90 91/// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes 92/// and an unconditional branch. Passes before isel (e.g. LSR/loopsimplify) 93/// often split edges in ways that are non-optimal for isel. Start by 94/// eliminating these blocks so we can split them the way we want them. 95bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) { 96 bool MadeChange = false; 97 // Note that this intentionally skips the entry block. 98 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) { 99 BasicBlock *BB = I++; 100 101 // If this block doesn't end with an uncond branch, ignore it. 102 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 103 if (!BI || !BI->isUnconditional()) 104 continue; 105 106 // If the instruction before the branch isn't a phi node, then other stuff 107 // is happening here. 108 BasicBlock::iterator BBI = BI; 109 if (BBI != BB->begin()) { 110 --BBI; 111 if (!isa<PHINode>(BBI)) continue; 112 } 113 114 // Do not break infinite loops. 115 BasicBlock *DestBB = BI->getSuccessor(0); 116 if (DestBB == BB) 117 continue; 118 119 if (!CanMergeBlocks(BB, DestBB)) 120 continue; 121 122 EliminateMostlyEmptyBlock(BB); 123 MadeChange = true; 124 } 125 return MadeChange; 126} 127 128/// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a 129/// single uncond branch between them, and BB contains no other non-phi 130/// instructions. 131bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB, 132 const BasicBlock *DestBB) const { 133 // We only want to eliminate blocks whose phi nodes are used by phi nodes in 134 // the successor. If there are more complex condition (e.g. preheaders), 135 // don't mess around with them. 136 BasicBlock::const_iterator BBI = BB->begin(); 137 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 138 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end(); 139 UI != E; ++UI) { 140 const Instruction *User = cast<Instruction>(*UI); 141 if (User->getParent() != DestBB || !isa<PHINode>(User)) 142 return false; 143 // If User is inside DestBB block and it is a PHINode then check 144 // incoming value. If incoming value is not from BB then this is 145 // a complex condition (e.g. preheaders) we want to avoid here. 146 if (User->getParent() == DestBB) { 147 if (const PHINode *UPN = dyn_cast<PHINode>(User)) 148 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { 149 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); 150 if (Insn && Insn->getParent() == BB && 151 Insn->getParent() != UPN->getIncomingBlock(I)) 152 return false; 153 } 154 } 155 } 156 } 157 158 // If BB and DestBB contain any common predecessors, then the phi nodes in BB 159 // and DestBB may have conflicting incoming values for the block. If so, we 160 // can't merge the block. 161 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); 162 if (!DestBBPN) return true; // no conflict. 163 164 // Collect the preds of BB. 165 SmallPtrSet<const BasicBlock*, 16> BBPreds; 166 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 167 // It is faster to get preds from a PHI than with pred_iterator. 168 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 169 BBPreds.insert(BBPN->getIncomingBlock(i)); 170 } else { 171 BBPreds.insert(pred_begin(BB), pred_end(BB)); 172 } 173 174 // Walk the preds of DestBB. 175 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { 176 BasicBlock *Pred = DestBBPN->getIncomingBlock(i); 177 if (BBPreds.count(Pred)) { // Common predecessor? 178 BBI = DestBB->begin(); 179 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 180 const Value *V1 = PN->getIncomingValueForBlock(Pred); 181 const Value *V2 = PN->getIncomingValueForBlock(BB); 182 183 // If V2 is a phi node in BB, look up what the mapped value will be. 184 if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) 185 if (V2PN->getParent() == BB) 186 V2 = V2PN->getIncomingValueForBlock(Pred); 187 188 // If there is a conflict, bail out. 189 if (V1 != V2) return false; 190 } 191 } 192 } 193 194 return true; 195} 196 197 198/// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and 199/// an unconditional branch in it. 200void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { 201 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 202 BasicBlock *DestBB = BI->getSuccessor(0); 203 204 DOUT << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB; 205 206 // If the destination block has a single pred, then this is a trivial edge, 207 // just collapse it. 208 if (DestBB->getSinglePredecessor()) { 209 // If DestBB has single-entry PHI nodes, fold them. 210 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { 211 PN->replaceAllUsesWith(PN->getIncomingValue(0)); 212 PN->eraseFromParent(); 213 } 214 215 // Splice all the PHI nodes from BB over to DestBB. 216 DestBB->getInstList().splice(DestBB->begin(), BB->getInstList(), 217 BB->begin(), BI); 218 219 // Anything that branched to BB now branches to DestBB. 220 BB->replaceAllUsesWith(DestBB); 221 222 // Nuke BB. 223 BB->eraseFromParent(); 224 225 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 226 return; 227 } 228 229 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB 230 // to handle the new incoming edges it is about to have. 231 PHINode *PN; 232 for (BasicBlock::iterator BBI = DestBB->begin(); 233 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 234 // Remove the incoming value for BB, and remember it. 235 Value *InVal = PN->removeIncomingValue(BB, false); 236 237 // Two options: either the InVal is a phi node defined in BB or it is some 238 // value that dominates BB. 239 PHINode *InValPhi = dyn_cast<PHINode>(InVal); 240 if (InValPhi && InValPhi->getParent() == BB) { 241 // Add all of the input values of the input PHI as inputs of this phi. 242 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) 243 PN->addIncoming(InValPhi->getIncomingValue(i), 244 InValPhi->getIncomingBlock(i)); 245 } else { 246 // Otherwise, add one instance of the dominating value for each edge that 247 // we will be adding. 248 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 249 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 250 PN->addIncoming(InVal, BBPN->getIncomingBlock(i)); 251 } else { 252 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 253 PN->addIncoming(InVal, *PI); 254 } 255 } 256 } 257 258 // The PHIs are now updated, change everything that refers to BB to use 259 // DestBB and remove BB. 260 BB->replaceAllUsesWith(DestBB); 261 BB->eraseFromParent(); 262 263 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 264} 265 266 267/// SplitEdgeNicely - Split the critical edge from TI to it's specified 268/// successor if it will improve codegen. We only do this if the successor has 269/// phi nodes (otherwise critical edges are ok). If there is already another 270/// predecessor of the succ that is empty (and thus has no phi nodes), use it 271/// instead of introducing a new block. 272static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, Pass *P) { 273 BasicBlock *TIBB = TI->getParent(); 274 BasicBlock *Dest = TI->getSuccessor(SuccNum); 275 assert(isa<PHINode>(Dest->begin()) && 276 "This should only be called if Dest has a PHI!"); 277 278 /// TIPHIValues - This array is lazily computed to determine the values of 279 /// PHIs in Dest that TI would provide. 280 std::vector<Value*> TIPHIValues; 281 282 // Check to see if Dest has any blocks that can be used as a split edge for 283 // this terminator. 284 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { 285 BasicBlock *Pred = *PI; 286 // To be usable, the pred has to end with an uncond branch to the dest. 287 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); 288 if (!PredBr || !PredBr->isUnconditional() || 289 // Must be empty other than the branch. 290 &Pred->front() != PredBr || 291 // Cannot be the entry block; its label does not get emitted. 292 Pred == &(Dest->getParent()->getEntryBlock())) 293 continue; 294 295 // Finally, since we know that Dest has phi nodes in it, we have to make 296 // sure that jumping to Pred will have the same affect as going to Dest in 297 // terms of PHI values. 298 PHINode *PN; 299 unsigned PHINo = 0; 300 bool FoundMatch = true; 301 for (BasicBlock::iterator I = Dest->begin(); 302 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { 303 if (PHINo == TIPHIValues.size()) 304 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); 305 306 // If the PHI entry doesn't work, we can't use this pred. 307 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { 308 FoundMatch = false; 309 break; 310 } 311 } 312 313 // If we found a workable predecessor, change TI to branch to Succ. 314 if (FoundMatch) { 315 Dest->removePredecessor(TIBB); 316 TI->setSuccessor(SuccNum, Pred); 317 return; 318 } 319 } 320 321 SplitCriticalEdge(TI, SuccNum, P, true); 322} 323 324/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop 325/// copy (e.g. it's casting from one pointer type to another, int->uint, or 326/// int->sbyte on PPC), sink it into user blocks to reduce the number of virtual 327/// registers that must be created and coalesced. 328/// 329/// Return true if any changes are made. 330static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ 331 // If this is a noop copy, 332 MVT::ValueType SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 333 MVT::ValueType DstVT = TLI.getValueType(CI->getType()); 334 335 // This is an fp<->int conversion? 336 if (MVT::isInteger(SrcVT) != MVT::isInteger(DstVT)) 337 return false; 338 339 // If this is an extension, it will be a zero or sign extension, which 340 // isn't a noop. 341 if (SrcVT < DstVT) return false; 342 343 // If these values will be promoted, find out what they will be promoted 344 // to. This helps us consider truncates on PPC as noop copies when they 345 // are. 346 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote) 347 SrcVT = TLI.getTypeToTransformTo(SrcVT); 348 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote) 349 DstVT = TLI.getTypeToTransformTo(DstVT); 350 351 // If, after promotion, these are the same types, this is a noop copy. 352 if (SrcVT != DstVT) 353 return false; 354 355 BasicBlock *DefBB = CI->getParent(); 356 357 /// InsertedCasts - Only insert a cast in each block once. 358 DenseMap<BasicBlock*, CastInst*> InsertedCasts; 359 360 bool MadeChange = false; 361 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 362 UI != E; ) { 363 Use &TheUse = UI.getUse(); 364 Instruction *User = cast<Instruction>(*UI); 365 366 // Figure out which BB this cast is used in. For PHI's this is the 367 // appropriate predecessor block. 368 BasicBlock *UserBB = User->getParent(); 369 if (PHINode *PN = dyn_cast<PHINode>(User)) { 370 unsigned OpVal = UI.getOperandNo()/2; 371 UserBB = PN->getIncomingBlock(OpVal); 372 } 373 374 // Preincrement use iterator so we don't invalidate it. 375 ++UI; 376 377 // If this user is in the same block as the cast, don't change the cast. 378 if (UserBB == DefBB) continue; 379 380 // If we have already inserted a cast into this block, use it. 381 CastInst *&InsertedCast = InsertedCasts[UserBB]; 382 383 if (!InsertedCast) { 384 BasicBlock::iterator InsertPt = UserBB->begin(); 385 while (isa<PHINode>(InsertPt)) ++InsertPt; 386 387 InsertedCast = 388 CastInst::create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 389 InsertPt); 390 MadeChange = true; 391 } 392 393 // Replace a use of the cast with a use of the new cast. 394 TheUse = InsertedCast; 395 } 396 397 // If we removed all uses, nuke the cast. 398 if (CI->use_empty()) 399 CI->eraseFromParent(); 400 401 return MadeChange; 402} 403 404/// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce 405/// the number of virtual registers that must be created and coalesced. This is 406/// a clear win except on targets with multiple condition code registers 407/// (PowerPC), where it might lose; some adjustment may be wanted there. 408/// 409/// Return true if any changes are made. 410static bool OptimizeCmpExpression(CmpInst *CI){ 411 412 BasicBlock *DefBB = CI->getParent(); 413 414 /// InsertedCmp - Only insert a cmp in each block once. 415 DenseMap<BasicBlock*, CmpInst*> InsertedCmps; 416 417 bool MadeChange = false; 418 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 419 UI != E; ) { 420 Use &TheUse = UI.getUse(); 421 Instruction *User = cast<Instruction>(*UI); 422 423 // Preincrement use iterator so we don't invalidate it. 424 ++UI; 425 426 // Don't bother for PHI nodes. 427 if (isa<PHINode>(User)) 428 continue; 429 430 // Figure out which BB this cmp is used in. 431 BasicBlock *UserBB = User->getParent(); 432 433 // If this user is in the same block as the cmp, don't change the cmp. 434 if (UserBB == DefBB) continue; 435 436 // If we have already inserted a cmp into this block, use it. 437 CmpInst *&InsertedCmp = InsertedCmps[UserBB]; 438 439 if (!InsertedCmp) { 440 BasicBlock::iterator InsertPt = UserBB->begin(); 441 while (isa<PHINode>(InsertPt)) ++InsertPt; 442 443 InsertedCmp = 444 CmpInst::create(CI->getOpcode(), CI->getPredicate(), CI->getOperand(0), 445 CI->getOperand(1), "", InsertPt); 446 MadeChange = true; 447 } 448 449 // Replace a use of the cmp with a use of the new cmp. 450 TheUse = InsertedCmp; 451 } 452 453 // If we removed all uses, nuke the cmp. 454 if (CI->use_empty()) 455 CI->eraseFromParent(); 456 457 return MadeChange; 458} 459 460/// EraseDeadInstructions - Erase any dead instructions 461static void EraseDeadInstructions(Value *V) { 462 Instruction *I = dyn_cast<Instruction>(V); 463 if (!I || !I->use_empty()) return; 464 465 SmallPtrSet<Instruction*, 16> Insts; 466 Insts.insert(I); 467 468 while (!Insts.empty()) { 469 I = *Insts.begin(); 470 Insts.erase(I); 471 if (isInstructionTriviallyDead(I)) { 472 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 473 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i))) 474 Insts.insert(U); 475 I->eraseFromParent(); 476 } 477 } 478} 479 480 481/// ExtAddrMode - This is an extended version of TargetLowering::AddrMode which 482/// holds actual Value*'s for register values. 483struct ExtAddrMode : public TargetLowering::AddrMode { 484 Value *BaseReg; 485 Value *ScaledReg; 486 ExtAddrMode() : BaseReg(0), ScaledReg(0) {} 487 void dump() const; 488}; 489 490static std::ostream &operator<<(std::ostream &OS, const ExtAddrMode &AM) { 491 bool NeedPlus = false; 492 OS << "["; 493 if (AM.BaseGV) 494 OS << (NeedPlus ? " + " : "") 495 << "GV:%" << AM.BaseGV->getName(), NeedPlus = true; 496 497 if (AM.BaseOffs) 498 OS << (NeedPlus ? " + " : "") << AM.BaseOffs, NeedPlus = true; 499 500 if (AM.BaseReg) 501 OS << (NeedPlus ? " + " : "") 502 << "Base:%" << AM.BaseReg->getName(), NeedPlus = true; 503 if (AM.Scale) 504 OS << (NeedPlus ? " + " : "") 505 << AM.Scale << "*%" << AM.ScaledReg->getName(), NeedPlus = true; 506 507 return OS << "]"; 508} 509 510void ExtAddrMode::dump() const { 511 cerr << *this << "\n"; 512} 513 514static bool TryMatchingScaledValue(Value *ScaleReg, int64_t Scale, 515 const Type *AccessTy, ExtAddrMode &AddrMode, 516 SmallVector<Instruction*, 16> &AddrModeInsts, 517 const TargetLowering &TLI, unsigned Depth); 518 519/// FindMaximalLegalAddressingMode - If we can, try to merge the computation of 520/// Addr into the specified addressing mode. If Addr can't be added to AddrMode 521/// this returns false. This assumes that Addr is either a pointer type or 522/// intptr_t for the target. 523static bool FindMaximalLegalAddressingMode(Value *Addr, const Type *AccessTy, 524 ExtAddrMode &AddrMode, 525 SmallVector<Instruction*, 16> &AddrModeInsts, 526 const TargetLowering &TLI, 527 unsigned Depth) { 528 529 // If this is a global variable, fold it into the addressing mode if possible. 530 if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) { 531 if (AddrMode.BaseGV == 0) { 532 AddrMode.BaseGV = GV; 533 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 534 return true; 535 AddrMode.BaseGV = 0; 536 } 537 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) { 538 AddrMode.BaseOffs += CI->getSExtValue(); 539 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 540 return true; 541 AddrMode.BaseOffs -= CI->getSExtValue(); 542 } else if (isa<ConstantPointerNull>(Addr)) { 543 return true; 544 } 545 546 // Look through constant exprs and instructions. 547 unsigned Opcode = ~0U; 548 User *AddrInst = 0; 549 if (Instruction *I = dyn_cast<Instruction>(Addr)) { 550 Opcode = I->getOpcode(); 551 AddrInst = I; 552 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) { 553 Opcode = CE->getOpcode(); 554 AddrInst = CE; 555 } 556 557 // Limit recursion to avoid exponential behavior. 558 if (Depth == 5) { AddrInst = 0; Opcode = ~0U; } 559 560 // If this is really an instruction, add it to our list of related 561 // instructions. 562 if (Instruction *I = dyn_cast_or_null<Instruction>(AddrInst)) 563 AddrModeInsts.push_back(I); 564 565 switch (Opcode) { 566 case Instruction::PtrToInt: 567 // PtrToInt is always a noop, as we know that the int type is pointer sized. 568 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 569 AddrMode, AddrModeInsts, TLI, Depth)) 570 return true; 571 break; 572 case Instruction::IntToPtr: 573 // This inttoptr is a no-op if the integer type is pointer sized. 574 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) == 575 TLI.getPointerTy()) { 576 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 577 AddrMode, AddrModeInsts, TLI, Depth)) 578 return true; 579 } 580 break; 581 case Instruction::Add: { 582 // Check to see if we can merge in the RHS then the LHS. If so, we win. 583 ExtAddrMode BackupAddrMode = AddrMode; 584 unsigned OldSize = AddrModeInsts.size(); 585 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(1), AccessTy, 586 AddrMode, AddrModeInsts, TLI, Depth+1) && 587 FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 588 AddrMode, AddrModeInsts, TLI, Depth+1)) 589 return true; 590 591 // Restore the old addr mode info. 592 AddrMode = BackupAddrMode; 593 AddrModeInsts.resize(OldSize); 594 595 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS. 596 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 597 AddrMode, AddrModeInsts, TLI, Depth+1) && 598 FindMaximalLegalAddressingMode(AddrInst->getOperand(1), AccessTy, 599 AddrMode, AddrModeInsts, TLI, Depth+1)) 600 return true; 601 602 // Otherwise we definitely can't merge the ADD in. 603 AddrMode = BackupAddrMode; 604 AddrModeInsts.resize(OldSize); 605 break; 606 } 607 case Instruction::Or: { 608 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); 609 if (!RHS) break; 610 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD. 611 break; 612 } 613 case Instruction::Mul: 614 case Instruction::Shl: { 615 // Can only handle X*C and X << C, and can only handle this when the scale 616 // field is available. 617 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); 618 if (!RHS) break; 619 int64_t Scale = RHS->getSExtValue(); 620 if (Opcode == Instruction::Shl) 621 Scale = 1 << Scale; 622 623 if (TryMatchingScaledValue(AddrInst->getOperand(0), Scale, AccessTy, 624 AddrMode, AddrModeInsts, TLI, Depth)) 625 return true; 626 break; 627 } 628 case Instruction::GetElementPtr: { 629 // Scan the GEP. We check it if it contains constant offsets and at most 630 // one variable offset. 631 int VariableOperand = -1; 632 unsigned VariableScale = 0; 633 634 int64_t ConstantOffset = 0; 635 const TargetData *TD = TLI.getTargetData(); 636 gep_type_iterator GTI = gep_type_begin(AddrInst); 637 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) { 638 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 639 const StructLayout *SL = TD->getStructLayout(STy); 640 unsigned Idx = 641 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue(); 642 ConstantOffset += SL->getElementOffset(Idx); 643 } else { 644 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType()); 645 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) { 646 ConstantOffset += CI->getSExtValue()*TypeSize; 647 } else if (TypeSize) { // Scales of zero don't do anything. 648 // We only allow one variable index at the moment. 649 if (VariableOperand != -1) { 650 VariableOperand = -2; 651 break; 652 } 653 654 // Remember the variable index. 655 VariableOperand = i; 656 VariableScale = TypeSize; 657 } 658 } 659 } 660 661 // If the GEP had multiple variable indices, punt. 662 if (VariableOperand == -2) 663 break; 664 665 // A common case is for the GEP to only do a constant offset. In this case, 666 // just add it to the disp field and check validity. 667 if (VariableOperand == -1) { 668 AddrMode.BaseOffs += ConstantOffset; 669 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){ 670 // Check to see if we can fold the base pointer in too. 671 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 672 AddrMode, AddrModeInsts, TLI, 673 Depth+1)) 674 return true; 675 } 676 AddrMode.BaseOffs -= ConstantOffset; 677 } else { 678 // Check that this has no base reg yet. If so, we won't have a place to 679 // put the base of the GEP (assuming it is not a null ptr). 680 bool SetBaseReg = false; 681 if (AddrMode.HasBaseReg) { 682 if (!isa<ConstantPointerNull>(AddrInst->getOperand(0))) 683 break; 684 } else { 685 AddrMode.HasBaseReg = true; 686 AddrMode.BaseReg = AddrInst->getOperand(0); 687 SetBaseReg = true; 688 } 689 690 // See if the scale amount is valid for this target. 691 AddrMode.BaseOffs += ConstantOffset; 692 if (TryMatchingScaledValue(AddrInst->getOperand(VariableOperand), 693 VariableScale, AccessTy, AddrMode, 694 AddrModeInsts, TLI, Depth)) { 695 if (!SetBaseReg) return true; 696 697 // If this match succeeded, we know that we can form an address with the 698 // GepBase as the basereg. See if we can match *more*. 699 AddrMode.HasBaseReg = false; 700 AddrMode.BaseReg = 0; 701 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 702 AddrMode, AddrModeInsts, TLI, 703 Depth+1)) 704 return true; 705 // Strange, shouldn't happen. Restore the base reg and succeed the easy 706 // way. 707 AddrMode.HasBaseReg = true; 708 AddrMode.BaseReg = AddrInst->getOperand(0); 709 return true; 710 } 711 712 AddrMode.BaseOffs -= ConstantOffset; 713 if (SetBaseReg) { 714 AddrMode.HasBaseReg = false; 715 AddrMode.BaseReg = 0; 716 } 717 } 718 break; 719 } 720 } 721 722 if (Instruction *I = dyn_cast_or_null<Instruction>(AddrInst)) { 723 assert(AddrModeInsts.back() == I && "Stack imbalance"); 724 AddrModeInsts.pop_back(); 725 } 726 727 // Worse case, the target should support [reg] addressing modes. :) 728 if (!AddrMode.HasBaseReg) { 729 AddrMode.HasBaseReg = true; 730 // Still check for legality in case the target supports [imm] but not [i+r]. 731 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { 732 AddrMode.BaseReg = Addr; 733 return true; 734 } 735 AddrMode.HasBaseReg = false; 736 } 737 738 // If the base register is already taken, see if we can do [r+r]. 739 if (AddrMode.Scale == 0) { 740 AddrMode.Scale = 1; 741 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { 742 AddrMode.ScaledReg = Addr; 743 return true; 744 } 745 AddrMode.Scale = 0; 746 } 747 // Couldn't match. 748 return false; 749} 750 751/// TryMatchingScaledValue - Try adding ScaleReg*Scale to the specified 752/// addressing mode. Return true if this addr mode is legal for the target, 753/// false if not. 754static bool TryMatchingScaledValue(Value *ScaleReg, int64_t Scale, 755 const Type *AccessTy, ExtAddrMode &AddrMode, 756 SmallVector<Instruction*, 16> &AddrModeInsts, 757 const TargetLowering &TLI, unsigned Depth) { 758 // If we already have a scale of this value, we can add to it, otherwise, we 759 // need an available scale field. 760 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg) 761 return false; 762 763 ExtAddrMode InputAddrMode = AddrMode; 764 765 // Add scale to turn X*4+X*3 -> X*7. This could also do things like 766 // [A+B + A*7] -> [B+A*8]. 767 AddrMode.Scale += Scale; 768 AddrMode.ScaledReg = ScaleReg; 769 770 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { 771 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now 772 // to see if ScaleReg is actually X+C. If so, we can turn this into adding 773 // X*Scale + C*Scale to addr mode. 774 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(ScaleReg); 775 if (BinOp && BinOp->getOpcode() == Instruction::Add && 776 isa<ConstantInt>(BinOp->getOperand(1)) && InputAddrMode.ScaledReg ==0) { 777 778 InputAddrMode.Scale = Scale; 779 InputAddrMode.ScaledReg = BinOp->getOperand(0); 780 InputAddrMode.BaseOffs += 781 cast<ConstantInt>(BinOp->getOperand(1))->getSExtValue()*Scale; 782 if (TLI.isLegalAddressingMode(InputAddrMode, AccessTy)) { 783 AddrModeInsts.push_back(BinOp); 784 AddrMode = InputAddrMode; 785 return true; 786 } 787 } 788 789 // Otherwise, not (x+c)*scale, just return what we have. 790 return true; 791 } 792 793 // Otherwise, back this attempt out. 794 AddrMode.Scale -= Scale; 795 if (AddrMode.Scale == 0) AddrMode.ScaledReg = 0; 796 797 return false; 798} 799 800 801/// IsNonLocalValue - Return true if the specified values are defined in a 802/// different basic block than BB. 803static bool IsNonLocalValue(Value *V, BasicBlock *BB) { 804 if (Instruction *I = dyn_cast<Instruction>(V)) 805 return I->getParent() != BB; 806 return false; 807} 808 809/// OptimizeLoadStoreInst - Load and Store Instructions have often have 810/// addressing modes that can do significant amounts of computation. As such, 811/// instruction selection will try to get the load or store to do as much 812/// computation as possible for the program. The problem is that isel can only 813/// see within a single block. As such, we sink as much legal addressing mode 814/// stuff into the block as possible. 815bool CodeGenPrepare::OptimizeLoadStoreInst(Instruction *LdStInst, Value *Addr, 816 const Type *AccessTy, 817 DenseMap<Value*,Value*> &SunkAddrs) { 818 // Figure out what addressing mode will be built up for this operation. 819 SmallVector<Instruction*, 16> AddrModeInsts; 820 ExtAddrMode AddrMode; 821 bool Success = FindMaximalLegalAddressingMode(Addr, AccessTy, AddrMode, 822 AddrModeInsts, *TLI, 0); 823 Success = Success; assert(Success && "Couldn't select *anything*?"); 824 825 // Check to see if any of the instructions supersumed by this addr mode are 826 // non-local to I's BB. 827 bool AnyNonLocal = false; 828 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { 829 if (IsNonLocalValue(AddrModeInsts[i], LdStInst->getParent())) { 830 AnyNonLocal = true; 831 break; 832 } 833 } 834 835 // If all the instructions matched are already in this BB, don't do anything. 836 if (!AnyNonLocal) { 837 DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n"); 838 return false; 839 } 840 841 // Insert this computation right after this user. Since our caller is 842 // scanning from the top of the BB to the bottom, reuse of the expr are 843 // guaranteed to happen later. 844 BasicBlock::iterator InsertPt = LdStInst; 845 846 // Now that we determined the addressing expression we want to use and know 847 // that we have to sink it into this block. Check to see if we have already 848 // done this for some other load/store instr in this block. If so, reuse the 849 // computation. 850 Value *&SunkAddr = SunkAddrs[Addr]; 851 if (SunkAddr) { 852 DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << "\n"); 853 if (SunkAddr->getType() != Addr->getType()) 854 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt); 855 } else { 856 DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << "\n"); 857 const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType(); 858 859 Value *Result = 0; 860 // Start with the scale value. 861 if (AddrMode.Scale) { 862 Value *V = AddrMode.ScaledReg; 863 if (V->getType() == IntPtrTy) { 864 // done. 865 } else if (isa<PointerType>(V->getType())) { 866 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 867 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 868 cast<IntegerType>(V->getType())->getBitWidth()) { 869 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt); 870 } else { 871 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt); 872 } 873 if (AddrMode.Scale != 1) 874 V = BinaryOperator::createMul(V, ConstantInt::get(IntPtrTy, 875 AddrMode.Scale), 876 "sunkaddr", InsertPt); 877 Result = V; 878 } 879 880 // Add in the base register. 881 if (AddrMode.BaseReg) { 882 Value *V = AddrMode.BaseReg; 883 if (V->getType() != IntPtrTy) 884 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 885 if (Result) 886 Result = BinaryOperator::createAdd(Result, V, "sunkaddr", InsertPt); 887 else 888 Result = V; 889 } 890 891 // Add in the BaseGV if present. 892 if (AddrMode.BaseGV) { 893 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr", 894 InsertPt); 895 if (Result) 896 Result = BinaryOperator::createAdd(Result, V, "sunkaddr", InsertPt); 897 else 898 Result = V; 899 } 900 901 // Add in the Base Offset if present. 902 if (AddrMode.BaseOffs) { 903 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 904 if (Result) 905 Result = BinaryOperator::createAdd(Result, V, "sunkaddr", InsertPt); 906 else 907 Result = V; 908 } 909 910 if (Result == 0) 911 SunkAddr = Constant::getNullValue(Addr->getType()); 912 else 913 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt); 914 } 915 916 LdStInst->replaceUsesOfWith(Addr, SunkAddr); 917 918 if (Addr->use_empty()) 919 EraseDeadInstructions(Addr); 920 return true; 921} 922 923bool CodeGenPrepare::OptimizeExtUses(Instruction *I) { 924 BasicBlock *DefBB = I->getParent(); 925 926 // If both result of the {s|z}xt and its source are live out, rewrite all 927 // other uses of the source with result of extension. 928 Value *Src = I->getOperand(0); 929 if (Src->hasOneUse()) 930 return false; 931 932 // Only do this xform is truncating is free. 933 if (!TLI->isTruncateFree(I->getType(), Src->getType())) 934 return false; 935 936 // Only safe to perform the optimization if the source is also defined in 937 // this block. 938 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) 939 return false; 940 941 bool DefIsLiveOut = false; 942 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 943 UI != E; ++UI) { 944 Instruction *User = cast<Instruction>(*UI); 945 946 // Figure out which BB this ext is used in. 947 BasicBlock *UserBB = User->getParent(); 948 if (UserBB == DefBB) continue; 949 DefIsLiveOut = true; 950 break; 951 } 952 if (!DefIsLiveOut) 953 return false; 954 955 // Make sure non of the uses are PHI nodes. 956 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 957 UI != E; ++UI) { 958 Instruction *User = cast<Instruction>(*UI); 959 BasicBlock *UserBB = User->getParent(); 960 if (UserBB == DefBB) continue; 961 // Be conservative. We don't want this xform to end up introducing 962 // reloads just before load / store instructions. 963 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User)) 964 return false; 965 } 966 967 // InsertedTruncs - Only insert one trunc in each block once. 968 DenseMap<BasicBlock*, Instruction*> InsertedTruncs; 969 970 bool MadeChange = false; 971 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 972 UI != E; ++UI) { 973 Use &TheUse = UI.getUse(); 974 Instruction *User = cast<Instruction>(*UI); 975 976 // Figure out which BB this ext is used in. 977 BasicBlock *UserBB = User->getParent(); 978 if (UserBB == DefBB) continue; 979 980 // Both src and def are live in this block. Rewrite the use. 981 Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; 982 983 if (!InsertedTrunc) { 984 BasicBlock::iterator InsertPt = UserBB->begin(); 985 while (isa<PHINode>(InsertPt)) ++InsertPt; 986 987 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt); 988 } 989 990 // Replace a use of the {s|z}ext source with a use of the result. 991 TheUse = InsertedTrunc; 992 993 MadeChange = true; 994 } 995 996 return MadeChange; 997} 998 999// In this pass we look for GEP and cast instructions that are used 1000// across basic blocks and rewrite them to improve basic-block-at-a-time 1001// selection. 1002bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { 1003 bool MadeChange = false; 1004 1005 // Split all critical edges where the dest block has a PHI and where the phi 1006 // has shared immediate operands. 1007 TerminatorInst *BBTI = BB.getTerminator(); 1008 if (BBTI->getNumSuccessors() > 1) { 1009 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) 1010 if (isa<PHINode>(BBTI->getSuccessor(i)->begin()) && 1011 isCriticalEdge(BBTI, i, true)) 1012 SplitEdgeNicely(BBTI, i, this); 1013 } 1014 1015 1016 // Keep track of non-local addresses that have been sunk into this block. 1017 // This allows us to avoid inserting duplicate code for blocks with multiple 1018 // load/stores of the same address. 1019 DenseMap<Value*, Value*> SunkAddrs; 1020 1021 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) { 1022 Instruction *I = BBI++; 1023 1024 if (CastInst *CI = dyn_cast<CastInst>(I)) { 1025 // If the source of the cast is a constant, then this should have 1026 // already been constant folded. The only reason NOT to constant fold 1027 // it is if something (e.g. LSR) was careful to place the constant 1028 // evaluation in a block other than then one that uses it (e.g. to hoist 1029 // the address of globals out of a loop). If this is the case, we don't 1030 // want to forward-subst the cast. 1031 if (isa<Constant>(CI->getOperand(0))) 1032 continue; 1033 1034 bool Change = false; 1035 if (TLI) { 1036 Change = OptimizeNoopCopyExpression(CI, *TLI); 1037 MadeChange |= Change; 1038 } 1039 1040 if (OptExtUses && !Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) 1041 MadeChange |= OptimizeExtUses(I); 1042 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 1043 MadeChange |= OptimizeCmpExpression(CI); 1044 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1045 if (TLI) 1046 MadeChange |= OptimizeLoadStoreInst(I, I->getOperand(0), LI->getType(), 1047 SunkAddrs); 1048 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 1049 if (TLI) 1050 MadeChange |= OptimizeLoadStoreInst(I, SI->getOperand(1), 1051 SI->getOperand(0)->getType(), 1052 SunkAddrs); 1053 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 1054 if (GEPI->hasAllZeroIndices()) { 1055 /// The GEP operand must be a pointer, so must its result -> BitCast 1056 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 1057 GEPI->getName(), GEPI); 1058 GEPI->replaceAllUsesWith(NC); 1059 GEPI->eraseFromParent(); 1060 MadeChange = true; 1061 BBI = NC; 1062 } 1063 } else if (CallInst *CI = dyn_cast<CallInst>(I)) { 1064 // If we found an inline asm expession, and if the target knows how to 1065 // lower it to normal LLVM code, do so now. 1066 if (TLI && isa<InlineAsm>(CI->getCalledValue())) 1067 if (const TargetAsmInfo *TAI = 1068 TLI->getTargetMachine().getTargetAsmInfo()) { 1069 if (TAI->ExpandInlineAsm(CI)) 1070 BBI = BB.begin(); 1071 } 1072 } 1073 } 1074 1075 return MadeChange; 1076} 1077 1078