CodeGenPrepare.cpp revision ecd94c804a563f2a86572dcf1d2e81f397e19daa
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/Debug.h" 32#include "llvm/Support/Compiler.h" 33#include "llvm/Support/GetElementPtrTypeIterator.h" 34using namespace llvm; 35 36namespace { 37 class VISIBILITY_HIDDEN CodeGenPrepare : public FunctionPass { 38 /// TLI - Keep a pointer of a TargetLowering to consult for determining 39 /// transformation profitability. 40 const TargetLowering *TLI; 41 public: 42 static char ID; // Pass identification, replacement for typeid 43 CodeGenPrepare(const TargetLowering *tli = 0) : FunctionPass((intptr_t)&ID), 44 TLI(tli) {} 45 bool runOnFunction(Function &F); 46 47 private: 48 bool EliminateMostlyEmptyBlocks(Function &F); 49 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; 50 void EliminateMostlyEmptyBlock(BasicBlock *BB); 51 bool OptimizeBlock(BasicBlock &BB); 52 bool OptimizeLoadStoreInst(Instruction *I, Value *Addr, 53 const Type *AccessTy, 54 DenseMap<Value*,Value*> &SunkAddrs); 55 }; 56} 57 58char CodeGenPrepare::ID = 0; 59static RegisterPass<CodeGenPrepare> X("codegenprepare", 60 "Optimize for code generation"); 61 62FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) { 63 return new CodeGenPrepare(TLI); 64} 65 66 67bool CodeGenPrepare::runOnFunction(Function &F) { 68 bool EverMadeChange = false; 69 70 // First pass, eliminate blocks that contain only PHI nodes and an 71 // unconditional branch. 72 EverMadeChange |= EliminateMostlyEmptyBlocks(F); 73 74 bool MadeChange = true; 75 while (MadeChange) { 76 MadeChange = false; 77 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 78 MadeChange |= OptimizeBlock(*BB); 79 EverMadeChange |= MadeChange; 80 } 81 return EverMadeChange; 82} 83 84/// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes 85/// and an unconditional branch. Passes before isel (e.g. LSR/loopsimplify) 86/// often split edges in ways that are non-optimal for isel. Start by 87/// eliminating these blocks so we can split them the way we want them. 88bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) { 89 bool MadeChange = false; 90 // Note that this intentionally skips the entry block. 91 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) { 92 BasicBlock *BB = I++; 93 94 // If this block doesn't end with an uncond branch, ignore it. 95 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 96 if (!BI || !BI->isUnconditional()) 97 continue; 98 99 // If the instruction before the branch isn't a phi node, then other stuff 100 // is happening here. 101 BasicBlock::iterator BBI = BI; 102 if (BBI != BB->begin()) { 103 --BBI; 104 if (!isa<PHINode>(BBI)) continue; 105 } 106 107 // Do not break infinite loops. 108 BasicBlock *DestBB = BI->getSuccessor(0); 109 if (DestBB == BB) 110 continue; 111 112 if (!CanMergeBlocks(BB, DestBB)) 113 continue; 114 115 EliminateMostlyEmptyBlock(BB); 116 MadeChange = true; 117 } 118 return MadeChange; 119} 120 121/// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a 122/// single uncond branch between them, and BB contains no other non-phi 123/// instructions. 124bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB, 125 const BasicBlock *DestBB) const { 126 // We only want to eliminate blocks whose phi nodes are used by phi nodes in 127 // the successor. If there are more complex condition (e.g. preheaders), 128 // don't mess around with them. 129 BasicBlock::const_iterator BBI = BB->begin(); 130 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 131 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end(); 132 UI != E; ++UI) { 133 const Instruction *User = cast<Instruction>(*UI); 134 if (User->getParent() != DestBB || !isa<PHINode>(User)) 135 return false; 136 // If User is inside DestBB block and it is a PHINode then check 137 // incoming value. If incoming value is not from BB then this is 138 // a complex condition (e.g. preheaders) we want to avoid here. 139 if (User->getParent() == DestBB) { 140 if (const PHINode *UPN = dyn_cast<PHINode>(User)) 141 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { 142 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); 143 if (Insn && Insn->getParent() == BB && 144 Insn->getParent() != UPN->getIncomingBlock(I)) 145 return false; 146 } 147 } 148 } 149 } 150 151 // If BB and DestBB contain any common predecessors, then the phi nodes in BB 152 // and DestBB may have conflicting incoming values for the block. If so, we 153 // can't merge the block. 154 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); 155 if (!DestBBPN) return true; // no conflict. 156 157 // Collect the preds of BB. 158 SmallPtrSet<BasicBlock*, 16> BBPreds; 159 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 160 // It is faster to get preds from a PHI than with pred_iterator. 161 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 162 BBPreds.insert(BBPN->getIncomingBlock(i)); 163 } else { 164 BBPreds.insert(pred_begin(BB), pred_end(BB)); 165 } 166 167 // Walk the preds of DestBB. 168 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { 169 BasicBlock *Pred = DestBBPN->getIncomingBlock(i); 170 if (BBPreds.count(Pred)) { // Common predecessor? 171 BBI = DestBB->begin(); 172 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 173 const Value *V1 = PN->getIncomingValueForBlock(Pred); 174 const Value *V2 = PN->getIncomingValueForBlock(BB); 175 176 // If V2 is a phi node in BB, look up what the mapped value will be. 177 if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) 178 if (V2PN->getParent() == BB) 179 V2 = V2PN->getIncomingValueForBlock(Pred); 180 181 // If there is a conflict, bail out. 182 if (V1 != V2) return false; 183 } 184 } 185 } 186 187 return true; 188} 189 190 191/// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and 192/// an unconditional branch in it. 193void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { 194 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 195 BasicBlock *DestBB = BI->getSuccessor(0); 196 197 DOUT << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB; 198 199 // If the destination block has a single pred, then this is a trivial edge, 200 // just collapse it. 201 if (DestBB->getSinglePredecessor()) { 202 // If DestBB has single-entry PHI nodes, fold them. 203 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { 204 PN->replaceAllUsesWith(PN->getIncomingValue(0)); 205 PN->eraseFromParent(); 206 } 207 208 // Splice all the PHI nodes from BB over to DestBB. 209 DestBB->getInstList().splice(DestBB->begin(), BB->getInstList(), 210 BB->begin(), BI); 211 212 // Anything that branched to BB now branches to DestBB. 213 BB->replaceAllUsesWith(DestBB); 214 215 // Nuke BB. 216 BB->eraseFromParent(); 217 218 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 219 return; 220 } 221 222 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB 223 // to handle the new incoming edges it is about to have. 224 PHINode *PN; 225 for (BasicBlock::iterator BBI = DestBB->begin(); 226 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 227 // Remove the incoming value for BB, and remember it. 228 Value *InVal = PN->removeIncomingValue(BB, false); 229 230 // Two options: either the InVal is a phi node defined in BB or it is some 231 // value that dominates BB. 232 PHINode *InValPhi = dyn_cast<PHINode>(InVal); 233 if (InValPhi && InValPhi->getParent() == BB) { 234 // Add all of the input values of the input PHI as inputs of this phi. 235 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) 236 PN->addIncoming(InValPhi->getIncomingValue(i), 237 InValPhi->getIncomingBlock(i)); 238 } else { 239 // Otherwise, add one instance of the dominating value for each edge that 240 // we will be adding. 241 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 242 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 243 PN->addIncoming(InVal, BBPN->getIncomingBlock(i)); 244 } else { 245 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 246 PN->addIncoming(InVal, *PI); 247 } 248 } 249 } 250 251 // The PHIs are now updated, change everything that refers to BB to use 252 // DestBB and remove BB. 253 BB->replaceAllUsesWith(DestBB); 254 BB->eraseFromParent(); 255 256 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 257} 258 259 260/// SplitEdgeNicely - Split the critical edge from TI to it's specified 261/// successor if it will improve codegen. We only do this if the successor has 262/// phi nodes (otherwise critical edges are ok). If there is already another 263/// predecessor of the succ that is empty (and thus has no phi nodes), use it 264/// instead of introducing a new block. 265static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, Pass *P) { 266 BasicBlock *TIBB = TI->getParent(); 267 BasicBlock *Dest = TI->getSuccessor(SuccNum); 268 assert(isa<PHINode>(Dest->begin()) && 269 "This should only be called if Dest has a PHI!"); 270 271 /// TIPHIValues - This array is lazily computed to determine the values of 272 /// PHIs in Dest that TI would provide. 273 std::vector<Value*> TIPHIValues; 274 275 // Check to see if Dest has any blocks that can be used as a split edge for 276 // this terminator. 277 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { 278 BasicBlock *Pred = *PI; 279 // To be usable, the pred has to end with an uncond branch to the dest. 280 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); 281 if (!PredBr || !PredBr->isUnconditional() || 282 // Must be empty other than the branch. 283 &Pred->front() != PredBr) 284 continue; 285 286 // Finally, since we know that Dest has phi nodes in it, we have to make 287 // sure that jumping to Pred will have the same affect as going to Dest in 288 // terms of PHI values. 289 PHINode *PN; 290 unsigned PHINo = 0; 291 bool FoundMatch = true; 292 for (BasicBlock::iterator I = Dest->begin(); 293 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { 294 if (PHINo == TIPHIValues.size()) 295 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); 296 297 // If the PHI entry doesn't work, we can't use this pred. 298 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { 299 FoundMatch = false; 300 break; 301 } 302 } 303 304 // If we found a workable predecessor, change TI to branch to Succ. 305 if (FoundMatch) { 306 Dest->removePredecessor(TIBB); 307 TI->setSuccessor(SuccNum, Pred); 308 return; 309 } 310 } 311 312 SplitCriticalEdge(TI, SuccNum, P, true); 313} 314 315/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop 316/// copy (e.g. it's casting from one pointer type to another, int->uint, or 317/// int->sbyte on PPC), sink it into user blocks to reduce the number of virtual 318/// registers that must be created and coallesced. 319/// 320/// Return true if any changes are made. 321static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ 322 // If this is a noop copy, 323 MVT::ValueType SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 324 MVT::ValueType DstVT = TLI.getValueType(CI->getType()); 325 326 // This is an fp<->int conversion? 327 if (MVT::isInteger(SrcVT) != MVT::isInteger(DstVT)) 328 return false; 329 330 // If this is an extension, it will be a zero or sign extension, which 331 // isn't a noop. 332 if (SrcVT < DstVT) return false; 333 334 // If these values will be promoted, find out what they will be promoted 335 // to. This helps us consider truncates on PPC as noop copies when they 336 // are. 337 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote) 338 SrcVT = TLI.getTypeToTransformTo(SrcVT); 339 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote) 340 DstVT = TLI.getTypeToTransformTo(DstVT); 341 342 // If, after promotion, these are the same types, this is a noop copy. 343 if (SrcVT != DstVT) 344 return false; 345 346 BasicBlock *DefBB = CI->getParent(); 347 348 /// InsertedCasts - Only insert a cast in each block once. 349 std::map<BasicBlock*, CastInst*> InsertedCasts; 350 351 bool MadeChange = false; 352 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 353 UI != E; ) { 354 Use &TheUse = UI.getUse(); 355 Instruction *User = cast<Instruction>(*UI); 356 357 // Figure out which BB this cast is used in. For PHI's this is the 358 // appropriate predecessor block. 359 BasicBlock *UserBB = User->getParent(); 360 if (PHINode *PN = dyn_cast<PHINode>(User)) { 361 unsigned OpVal = UI.getOperandNo()/2; 362 UserBB = PN->getIncomingBlock(OpVal); 363 } 364 365 // Preincrement use iterator so we don't invalidate it. 366 ++UI; 367 368 // If this user is in the same block as the cast, don't change the cast. 369 if (UserBB == DefBB) continue; 370 371 // If we have already inserted a cast into this block, use it. 372 CastInst *&InsertedCast = InsertedCasts[UserBB]; 373 374 if (!InsertedCast) { 375 BasicBlock::iterator InsertPt = UserBB->begin(); 376 while (isa<PHINode>(InsertPt)) ++InsertPt; 377 378 InsertedCast = 379 CastInst::create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 380 InsertPt); 381 MadeChange = true; 382 } 383 384 // Replace a use of the cast with a use of the new casat. 385 TheUse = InsertedCast; 386 } 387 388 // If we removed all uses, nuke the cast. 389 if (CI->use_empty()) 390 CI->eraseFromParent(); 391 392 return MadeChange; 393} 394 395/// EraseDeadInstructions - Erase any dead instructions 396static void EraseDeadInstructions(Value *V) { 397 Instruction *I = dyn_cast<Instruction>(V); 398 if (!I || !I->use_empty()) return; 399 400 SmallPtrSet<Instruction*, 16> Insts; 401 Insts.insert(I); 402 403 while (!Insts.empty()) { 404 I = *Insts.begin(); 405 Insts.erase(I); 406 if (isInstructionTriviallyDead(I)) { 407 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 408 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i))) 409 Insts.insert(U); 410 I->eraseFromParent(); 411 } 412 } 413} 414 415 416/// ExtAddrMode - This is an extended version of TargetLowering::AddrMode which 417/// holds actual Value*'s for register values. 418struct ExtAddrMode : public TargetLowering::AddrMode { 419 Value *BaseReg; 420 Value *ScaledReg; 421 ExtAddrMode() : BaseReg(0), ScaledReg(0) {} 422 void dump() const; 423}; 424 425static std::ostream &operator<<(std::ostream &OS, const ExtAddrMode &AM) { 426 bool NeedPlus = false; 427 OS << "["; 428 if (AM.BaseGV) 429 OS << (NeedPlus ? " + " : "") 430 << "GV:%" << AM.BaseGV->getName(), NeedPlus = true; 431 432 if (AM.BaseOffs) 433 OS << (NeedPlus ? " + " : "") << AM.BaseOffs, NeedPlus = true; 434 435 if (AM.BaseReg) 436 OS << (NeedPlus ? " + " : "") 437 << "Base:%" << AM.BaseReg->getName(), NeedPlus = true; 438 if (AM.Scale) 439 OS << (NeedPlus ? " + " : "") 440 << AM.Scale << "*%" << AM.ScaledReg->getName(), NeedPlus = true; 441 442 return OS << "]"; 443} 444 445void ExtAddrMode::dump() const { 446 cerr << *this << "\n"; 447} 448 449static bool TryMatchingScaledValue(Value *ScaleReg, int64_t Scale, 450 const Type *AccessTy, ExtAddrMode &AddrMode, 451 SmallVector<Instruction*, 16> &AddrModeInsts, 452 const TargetLowering &TLI, unsigned Depth); 453 454/// FindMaximalLegalAddressingMode - If we can, try to merge the computation of 455/// Addr into the specified addressing mode. If Addr can't be added to AddrMode 456/// this returns false. This assumes that Addr is either a pointer type or 457/// intptr_t for the target. 458static bool FindMaximalLegalAddressingMode(Value *Addr, const Type *AccessTy, 459 ExtAddrMode &AddrMode, 460 SmallVector<Instruction*, 16> &AddrModeInsts, 461 const TargetLowering &TLI, 462 unsigned Depth) { 463 464 // If this is a global variable, fold it into the addressing mode if possible. 465 if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) { 466 if (AddrMode.BaseGV == 0) { 467 AddrMode.BaseGV = GV; 468 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 469 return true; 470 AddrMode.BaseGV = 0; 471 } 472 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) { 473 AddrMode.BaseOffs += CI->getSExtValue(); 474 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 475 return true; 476 AddrMode.BaseOffs -= CI->getSExtValue(); 477 } else if (isa<ConstantPointerNull>(Addr)) { 478 return true; 479 } 480 481 // Look through constant exprs and instructions. 482 unsigned Opcode = ~0U; 483 User *AddrInst = 0; 484 if (Instruction *I = dyn_cast<Instruction>(Addr)) { 485 Opcode = I->getOpcode(); 486 AddrInst = I; 487 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) { 488 Opcode = CE->getOpcode(); 489 AddrInst = CE; 490 } 491 492 // Limit recursion to avoid exponential behavior. 493 if (Depth == 5) { AddrInst = 0; Opcode = ~0U; } 494 495 // If this is really an instruction, add it to our list of related 496 // instructions. 497 if (Instruction *I = dyn_cast_or_null<Instruction>(AddrInst)) 498 AddrModeInsts.push_back(I); 499 500 switch (Opcode) { 501 case Instruction::PtrToInt: 502 // PtrToInt is always a noop, as we know that the int type is pointer sized. 503 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 504 AddrMode, AddrModeInsts, TLI, Depth)) 505 return true; 506 break; 507 case Instruction::IntToPtr: 508 // This inttoptr is a no-op if the integer type is pointer sized. 509 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) == 510 TLI.getPointerTy()) { 511 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 512 AddrMode, AddrModeInsts, TLI, Depth)) 513 return true; 514 } 515 break; 516 case Instruction::Add: { 517 // Check to see if we can merge in the RHS then the LHS. If so, we win. 518 ExtAddrMode BackupAddrMode = AddrMode; 519 unsigned OldSize = AddrModeInsts.size(); 520 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(1), AccessTy, 521 AddrMode, AddrModeInsts, TLI, Depth+1) && 522 FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 523 AddrMode, AddrModeInsts, TLI, Depth+1)) 524 return true; 525 526 // Restore the old addr mode info. 527 AddrMode = BackupAddrMode; 528 AddrModeInsts.resize(OldSize); 529 530 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS. 531 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 532 AddrMode, AddrModeInsts, TLI, Depth+1) && 533 FindMaximalLegalAddressingMode(AddrInst->getOperand(1), AccessTy, 534 AddrMode, AddrModeInsts, TLI, Depth+1)) 535 return true; 536 537 // Otherwise we definitely can't merge the ADD in. 538 AddrMode = BackupAddrMode; 539 AddrModeInsts.resize(OldSize); 540 break; 541 } 542 case Instruction::Or: { 543 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); 544 if (!RHS) break; 545 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD. 546 break; 547 } 548 case Instruction::Mul: 549 case Instruction::Shl: { 550 // Can only handle X*C and X << C, and can only handle this when the scale 551 // field is available. 552 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); 553 if (!RHS) break; 554 int64_t Scale = RHS->getSExtValue(); 555 if (Opcode == Instruction::Shl) 556 Scale = 1 << Scale; 557 558 if (TryMatchingScaledValue(AddrInst->getOperand(0), Scale, AccessTy, 559 AddrMode, AddrModeInsts, TLI, Depth)) 560 return true; 561 break; 562 } 563 case Instruction::GetElementPtr: { 564 // Scan the GEP. We check it if it contains constant offsets and at most 565 // one variable offset. 566 int VariableOperand = -1; 567 unsigned VariableScale = 0; 568 569 int64_t ConstantOffset = 0; 570 const TargetData *TD = TLI.getTargetData(); 571 gep_type_iterator GTI = gep_type_begin(AddrInst); 572 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) { 573 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 574 const StructLayout *SL = TD->getStructLayout(STy); 575 unsigned Idx = 576 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue(); 577 ConstantOffset += SL->getElementOffset(Idx); 578 } else { 579 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType()); 580 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) { 581 ConstantOffset += CI->getSExtValue()*TypeSize; 582 } else if (TypeSize) { // Scales of zero don't do anything. 583 // We only allow one variable index at the moment. 584 if (VariableOperand != -1) { 585 VariableOperand = -2; 586 break; 587 } 588 589 // Remember the variable index. 590 VariableOperand = i; 591 VariableScale = TypeSize; 592 } 593 } 594 } 595 596 // If the GEP had multiple variable indices, punt. 597 if (VariableOperand == -2) 598 break; 599 600 // A common case is for the GEP to only do a constant offset. In this case, 601 // just add it to the disp field and check validity. 602 if (VariableOperand == -1) { 603 AddrMode.BaseOffs += ConstantOffset; 604 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){ 605 // Check to see if we can fold the base pointer in too. 606 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 607 AddrMode, AddrModeInsts, TLI, 608 Depth+1)) 609 return true; 610 } 611 AddrMode.BaseOffs -= ConstantOffset; 612 } else { 613 // Check that this has no base reg yet. If so, we won't have a place to 614 // put the base of the GEP (assuming it is not a null ptr). 615 bool SetBaseReg = false; 616 if (AddrMode.HasBaseReg) { 617 if (!isa<ConstantPointerNull>(AddrInst->getOperand(0))) 618 break; 619 } else { 620 AddrMode.HasBaseReg = true; 621 AddrMode.BaseReg = AddrInst->getOperand(0); 622 SetBaseReg = true; 623 } 624 625 // See if the scale amount is valid for this target. 626 AddrMode.BaseOffs += ConstantOffset; 627 if (TryMatchingScaledValue(AddrInst->getOperand(VariableOperand), 628 VariableScale, AccessTy, AddrMode, 629 AddrModeInsts, TLI, Depth)) { 630 if (!SetBaseReg) return true; 631 632 // If this match succeeded, we know that we can form an address with the 633 // GepBase as the basereg. See if we can match *more*. 634 AddrMode.HasBaseReg = false; 635 AddrMode.BaseReg = 0; 636 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 637 AddrMode, AddrModeInsts, TLI, 638 Depth+1)) 639 return true; 640 // Strange, shouldn't happen. Restore the base reg and succeed the easy 641 // way. 642 AddrMode.HasBaseReg = true; 643 AddrMode.BaseReg = AddrInst->getOperand(0); 644 return true; 645 } 646 647 AddrMode.BaseOffs -= ConstantOffset; 648 if (SetBaseReg) { 649 AddrMode.HasBaseReg = false; 650 AddrMode.BaseReg = 0; 651 } 652 } 653 break; 654 } 655 } 656 657 if (Instruction *I = dyn_cast_or_null<Instruction>(AddrInst)) { 658 assert(AddrModeInsts.back() == I && "Stack imbalance"); 659 AddrModeInsts.pop_back(); 660 } 661 662 // Worse case, the target should support [reg] addressing modes. :) 663 if (!AddrMode.HasBaseReg) { 664 AddrMode.HasBaseReg = true; 665 // Still check for legality in case the target supports [imm] but not [i+r]. 666 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { 667 AddrMode.BaseReg = Addr; 668 return true; 669 } 670 AddrMode.HasBaseReg = false; 671 } 672 673 // If the base register is already taken, see if we can do [r+r]. 674 if (AddrMode.Scale == 0) { 675 AddrMode.Scale = 1; 676 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { 677 AddrMode.ScaledReg = Addr; 678 return true; 679 } 680 AddrMode.Scale = 0; 681 } 682 // Couldn't match. 683 return false; 684} 685 686/// TryMatchingScaledValue - Try adding ScaleReg*Scale to the specified 687/// addressing mode. Return true if this addr mode is legal for the target, 688/// false if not. 689static bool TryMatchingScaledValue(Value *ScaleReg, int64_t Scale, 690 const Type *AccessTy, ExtAddrMode &AddrMode, 691 SmallVector<Instruction*, 16> &AddrModeInsts, 692 const TargetLowering &TLI, unsigned Depth) { 693 // If we already have a scale of this value, we can add to it, otherwise, we 694 // need an available scale field. 695 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg) 696 return false; 697 698 ExtAddrMode InputAddrMode = AddrMode; 699 700 // Add scale to turn X*4+X*3 -> X*7. This could also do things like 701 // [A+B + A*7] -> [B+A*8]. 702 AddrMode.Scale += Scale; 703 AddrMode.ScaledReg = ScaleReg; 704 705 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { 706 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now 707 // to see if ScaleReg is actually X+C. If so, we can turn this into adding 708 // X*Scale + C*Scale to addr mode. 709 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(ScaleReg); 710 if (BinOp && BinOp->getOpcode() == Instruction::Add && 711 isa<ConstantInt>(BinOp->getOperand(1)) && InputAddrMode.ScaledReg ==0) { 712 713 InputAddrMode.Scale = Scale; 714 InputAddrMode.ScaledReg = BinOp->getOperand(0); 715 InputAddrMode.BaseOffs += 716 cast<ConstantInt>(BinOp->getOperand(1))->getSExtValue()*Scale; 717 if (TLI.isLegalAddressingMode(InputAddrMode, AccessTy)) { 718 AddrModeInsts.push_back(BinOp); 719 AddrMode = InputAddrMode; 720 return true; 721 } 722 } 723 724 // Otherwise, not (x+c)*scale, just return what we have. 725 return true; 726 } 727 728 // Otherwise, back this attempt out. 729 AddrMode.Scale -= Scale; 730 if (AddrMode.Scale == 0) AddrMode.ScaledReg = 0; 731 732 return false; 733} 734 735 736/// IsNonLocalValue - Return true if the specified values are defined in a 737/// different basic block than BB. 738static bool IsNonLocalValue(Value *V, BasicBlock *BB) { 739 if (Instruction *I = dyn_cast<Instruction>(V)) 740 return I->getParent() != BB; 741 return false; 742} 743 744/// OptimizeLoadStoreInst - Load and Store Instructions have often have 745/// addressing modes that can do significant amounts of computation. As such, 746/// instruction selection will try to get the load or store to do as much 747/// computation as possible for the program. The problem is that isel can only 748/// see within a single block. As such, we sink as much legal addressing mode 749/// stuff into the block as possible. 750bool CodeGenPrepare::OptimizeLoadStoreInst(Instruction *LdStInst, Value *Addr, 751 const Type *AccessTy, 752 DenseMap<Value*,Value*> &SunkAddrs) { 753 // Figure out what addressing mode will be built up for this operation. 754 SmallVector<Instruction*, 16> AddrModeInsts; 755 ExtAddrMode AddrMode; 756 bool Success = FindMaximalLegalAddressingMode(Addr, AccessTy, AddrMode, 757 AddrModeInsts, *TLI, 0); 758 Success = Success; assert(Success && "Couldn't select *anything*?"); 759 760 // Check to see if any of the instructions supersumed by this addr mode are 761 // non-local to I's BB. 762 bool AnyNonLocal = false; 763 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { 764 if (IsNonLocalValue(AddrModeInsts[i], LdStInst->getParent())) { 765 AnyNonLocal = true; 766 break; 767 } 768 } 769 770 // If all the instructions matched are already in this BB, don't do anything. 771 if (!AnyNonLocal) { 772 DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n"); 773 return false; 774 } 775 776 // Insert this computation right after this user. Since our caller is 777 // scanning from the top of the BB to the bottom, reuse of the expr are 778 // guaranteed to happen later. 779 BasicBlock::iterator InsertPt = LdStInst; 780 781 // Now that we determined the addressing expression we want to use and know 782 // that we have to sink it into this block. Check to see if we have already 783 // done this for some other load/store instr in this block. If so, reuse the 784 // computation. 785 Value *&SunkAddr = SunkAddrs[Addr]; 786 if (SunkAddr) { 787 DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << "\n"); 788 if (SunkAddr->getType() != Addr->getType()) 789 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt); 790 } else { 791 DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << "\n"); 792 const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType(); 793 794 Value *Result = 0; 795 // Start with the scale value. 796 if (AddrMode.Scale) { 797 Value *V = AddrMode.ScaledReg; 798 if (V->getType() == IntPtrTy) { 799 // done. 800 } else if (isa<PointerType>(V->getType())) { 801 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 802 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 803 cast<IntegerType>(V->getType())->getBitWidth()) { 804 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt); 805 } else { 806 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt); 807 } 808 if (AddrMode.Scale != 1) 809 V = BinaryOperator::createMul(V, ConstantInt::get(IntPtrTy, 810 AddrMode.Scale), 811 "sunkaddr", InsertPt); 812 Result = V; 813 } 814 815 // Add in the base register. 816 if (AddrMode.BaseReg) { 817 Value *V = AddrMode.BaseReg; 818 if (V->getType() != IntPtrTy) 819 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 820 if (Result) 821 Result = BinaryOperator::createAdd(Result, V, "sunkaddr", InsertPt); 822 else 823 Result = V; 824 } 825 826 // Add in the BaseGV if present. 827 if (AddrMode.BaseGV) { 828 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr", 829 InsertPt); 830 if (Result) 831 Result = BinaryOperator::createAdd(Result, V, "sunkaddr", InsertPt); 832 else 833 Result = V; 834 } 835 836 // Add in the Base Offset if present. 837 if (AddrMode.BaseOffs) { 838 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 839 if (Result) 840 Result = BinaryOperator::createAdd(Result, V, "sunkaddr", InsertPt); 841 else 842 Result = V; 843 } 844 845 if (Result == 0) 846 SunkAddr = Constant::getNullValue(Addr->getType()); 847 else 848 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt); 849 } 850 851 LdStInst->replaceUsesOfWith(Addr, SunkAddr); 852 853 if (Addr->use_empty()) 854 EraseDeadInstructions(Addr); 855 return true; 856} 857 858// In this pass we look for GEP and cast instructions that are used 859// across basic blocks and rewrite them to improve basic-block-at-a-time 860// selection. 861bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { 862 bool MadeChange = false; 863 864 // Split all critical edges where the dest block has a PHI and where the phi 865 // has shared immediate operands. 866 TerminatorInst *BBTI = BB.getTerminator(); 867 if (BBTI->getNumSuccessors() > 1) { 868 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) 869 if (isa<PHINode>(BBTI->getSuccessor(i)->begin()) && 870 isCriticalEdge(BBTI, i, true)) 871 SplitEdgeNicely(BBTI, i, this); 872 } 873 874 875 // Keep track of non-local addresses that have been sunk into this block. 876 // This allows us to avoid inserting duplicate code for blocks with multiple 877 // load/stores of the same address. 878 DenseMap<Value*, Value*> SunkAddrs; 879 880 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) { 881 Instruction *I = BBI++; 882 883 if (CastInst *CI = dyn_cast<CastInst>(I)) { 884 // If the source of the cast is a constant, then this should have 885 // already been constant folded. The only reason NOT to constant fold 886 // it is if something (e.g. LSR) was careful to place the constant 887 // evaluation in a block other than then one that uses it (e.g. to hoist 888 // the address of globals out of a loop). If this is the case, we don't 889 // want to forward-subst the cast. 890 if (isa<Constant>(CI->getOperand(0))) 891 continue; 892 893 if (TLI) 894 MadeChange |= OptimizeNoopCopyExpression(CI, *TLI); 895 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 896 if (TLI) 897 MadeChange |= OptimizeLoadStoreInst(I, I->getOperand(0), LI->getType(), 898 SunkAddrs); 899 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 900 if (TLI) 901 MadeChange |= OptimizeLoadStoreInst(I, SI->getOperand(1), 902 SI->getOperand(0)->getType(), 903 SunkAddrs); 904 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 905 if (GEPI->hasAllZeroIndices()) { 906 /// The GEP operand must be a pointer, so must its result -> BitCast 907 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 908 GEPI->getName(), GEPI); 909 GEPI->replaceAllUsesWith(NC); 910 GEPI->eraseFromParent(); 911 MadeChange = true; 912 BBI = NC; 913 } 914 } else if (CallInst *CI = dyn_cast<CallInst>(I)) { 915 // If we found an inline asm expession, and if the target knows how to 916 // lower it to normal LLVM code, do so now. 917 if (TLI && isa<InlineAsm>(CI->getCalledValue())) 918 if (const TargetAsmInfo *TAI = 919 TLI->getTargetMachine().getTargetAsmInfo()) { 920 if (TAI->ExpandInlineAsm(CI)) 921 BBI = BB.begin(); 922 } 923 } 924 } 925 926 return MadeChange; 927} 928 929