CodeGenPrepare.cpp revision 75abc1ed0618048c3cf6c5b71c9868c10d6c1478
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 CodeGenPrepare(const TargetLowering *tli = 0) : TLI(tli) {} 43 bool runOnFunction(Function &F); 44 45 private: 46 bool EliminateMostlyEmptyBlocks(Function &F); 47 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; 48 void EliminateMostlyEmptyBlock(BasicBlock *BB); 49 bool OptimizeBlock(BasicBlock &BB); 50 bool OptimizeLoadStoreInst(Instruction *I, Value *Addr, 51 const Type *AccessTy, 52 DenseMap<Value*,Value*> &SunkAddrs); 53 }; 54} 55static RegisterPass<CodeGenPrepare> X("codegenprepare", 56 "Optimize for code generation"); 57 58FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) { 59 return new CodeGenPrepare(TLI); 60} 61 62 63bool CodeGenPrepare::runOnFunction(Function &F) { 64 bool EverMadeChange = false; 65 66 // First pass, eliminate blocks that contain only PHI nodes and an 67 // unconditional branch. 68 EverMadeChange |= EliminateMostlyEmptyBlocks(F); 69 70 bool MadeChange = true; 71 while (MadeChange) { 72 MadeChange = false; 73 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 74 MadeChange |= OptimizeBlock(*BB); 75 EverMadeChange |= MadeChange; 76 } 77 return EverMadeChange; 78} 79 80/// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes 81/// and an unconditional branch. Passes before isel (e.g. LSR/loopsimplify) 82/// often split edges in ways that are non-optimal for isel. Start by 83/// eliminating these blocks so we can split them the way we want them. 84bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) { 85 bool MadeChange = false; 86 // Note that this intentionally skips the entry block. 87 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) { 88 BasicBlock *BB = I++; 89 90 // If this block doesn't end with an uncond branch, ignore it. 91 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 92 if (!BI || !BI->isUnconditional()) 93 continue; 94 95 // If the instruction before the branch isn't a phi node, then other stuff 96 // is happening here. 97 BasicBlock::iterator BBI = BI; 98 if (BBI != BB->begin()) { 99 --BBI; 100 if (!isa<PHINode>(BBI)) continue; 101 } 102 103 // Do not break infinite loops. 104 BasicBlock *DestBB = BI->getSuccessor(0); 105 if (DestBB == BB) 106 continue; 107 108 if (!CanMergeBlocks(BB, DestBB)) 109 continue; 110 111 EliminateMostlyEmptyBlock(BB); 112 MadeChange = true; 113 } 114 return MadeChange; 115} 116 117/// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a 118/// single uncond branch between them, and BB contains no other non-phi 119/// instructions. 120bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB, 121 const BasicBlock *DestBB) const { 122 // We only want to eliminate blocks whose phi nodes are used by phi nodes in 123 // the successor. If there are more complex condition (e.g. preheaders), 124 // don't mess around with them. 125 BasicBlock::const_iterator BBI = BB->begin(); 126 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 127 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end(); 128 UI != E; ++UI) { 129 const Instruction *User = cast<Instruction>(*UI); 130 if (User->getParent() != DestBB || !isa<PHINode>(User)) 131 return false; 132 // If User is inside DestBB block and it is a PHINode then check 133 // incoming value. If incoming value is not from BB then this is 134 // a complex condition (e.g. preheaders) we want to avoid here. 135 if (User->getParent() == DestBB) { 136 if (const PHINode *UPN = dyn_cast<PHINode>(User)) 137 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { 138 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); 139 if (Insn && Insn->getParent() == BB && 140 Insn->getParent() != UPN->getIncomingBlock(I)) 141 return false; 142 } 143 } 144 } 145 } 146 147 // If BB and DestBB contain any common predecessors, then the phi nodes in BB 148 // and DestBB may have conflicting incoming values for the block. If so, we 149 // can't merge the block. 150 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); 151 if (!DestBBPN) return true; // no conflict. 152 153 // Collect the preds of BB. 154 SmallPtrSet<BasicBlock*, 16> BBPreds; 155 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 156 // It is faster to get preds from a PHI than with pred_iterator. 157 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 158 BBPreds.insert(BBPN->getIncomingBlock(i)); 159 } else { 160 BBPreds.insert(pred_begin(BB), pred_end(BB)); 161 } 162 163 // Walk the preds of DestBB. 164 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { 165 BasicBlock *Pred = DestBBPN->getIncomingBlock(i); 166 if (BBPreds.count(Pred)) { // Common predecessor? 167 BBI = DestBB->begin(); 168 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 169 const Value *V1 = PN->getIncomingValueForBlock(Pred); 170 const Value *V2 = PN->getIncomingValueForBlock(BB); 171 172 // If V2 is a phi node in BB, look up what the mapped value will be. 173 if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) 174 if (V2PN->getParent() == BB) 175 V2 = V2PN->getIncomingValueForBlock(Pred); 176 177 // If there is a conflict, bail out. 178 if (V1 != V2) return false; 179 } 180 } 181 } 182 183 return true; 184} 185 186 187/// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and 188/// an unconditional branch in it. 189void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { 190 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 191 BasicBlock *DestBB = BI->getSuccessor(0); 192 193 DOUT << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB; 194 195 // If the destination block has a single pred, then this is a trivial edge, 196 // just collapse it. 197 if (DestBB->getSinglePredecessor()) { 198 // If DestBB has single-entry PHI nodes, fold them. 199 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { 200 PN->replaceAllUsesWith(PN->getIncomingValue(0)); 201 PN->eraseFromParent(); 202 } 203 204 // Splice all the PHI nodes from BB over to DestBB. 205 DestBB->getInstList().splice(DestBB->begin(), BB->getInstList(), 206 BB->begin(), BI); 207 208 // Anything that branched to BB now branches to DestBB. 209 BB->replaceAllUsesWith(DestBB); 210 211 // Nuke BB. 212 BB->eraseFromParent(); 213 214 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 215 return; 216 } 217 218 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB 219 // to handle the new incoming edges it is about to have. 220 PHINode *PN; 221 for (BasicBlock::iterator BBI = DestBB->begin(); 222 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 223 // Remove the incoming value for BB, and remember it. 224 Value *InVal = PN->removeIncomingValue(BB, false); 225 226 // Two options: either the InVal is a phi node defined in BB or it is some 227 // value that dominates BB. 228 PHINode *InValPhi = dyn_cast<PHINode>(InVal); 229 if (InValPhi && InValPhi->getParent() == BB) { 230 // Add all of the input values of the input PHI as inputs of this phi. 231 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) 232 PN->addIncoming(InValPhi->getIncomingValue(i), 233 InValPhi->getIncomingBlock(i)); 234 } else { 235 // Otherwise, add one instance of the dominating value for each edge that 236 // we will be adding. 237 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 238 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 239 PN->addIncoming(InVal, BBPN->getIncomingBlock(i)); 240 } else { 241 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 242 PN->addIncoming(InVal, *PI); 243 } 244 } 245 } 246 247 // The PHIs are now updated, change everything that refers to BB to use 248 // DestBB and remove BB. 249 BB->replaceAllUsesWith(DestBB); 250 BB->eraseFromParent(); 251 252 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 253} 254 255 256/// SplitEdgeNicely - Split the critical edge from TI to it's specified 257/// successor if it will improve codegen. We only do this if the successor has 258/// phi nodes (otherwise critical edges are ok). If there is already another 259/// predecessor of the succ that is empty (and thus has no phi nodes), use it 260/// instead of introducing a new block. 261static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, Pass *P) { 262 BasicBlock *TIBB = TI->getParent(); 263 BasicBlock *Dest = TI->getSuccessor(SuccNum); 264 assert(isa<PHINode>(Dest->begin()) && 265 "This should only be called if Dest has a PHI!"); 266 267 /// TIPHIValues - This array is lazily computed to determine the values of 268 /// PHIs in Dest that TI would provide. 269 std::vector<Value*> TIPHIValues; 270 271 // Check to see if Dest has any blocks that can be used as a split edge for 272 // this terminator. 273 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { 274 BasicBlock *Pred = *PI; 275 // To be usable, the pred has to end with an uncond branch to the dest. 276 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); 277 if (!PredBr || !PredBr->isUnconditional() || 278 // Must be empty other than the branch. 279 &Pred->front() != PredBr) 280 continue; 281 282 // Finally, since we know that Dest has phi nodes in it, we have to make 283 // sure that jumping to Pred will have the same affect as going to Dest in 284 // terms of PHI values. 285 PHINode *PN; 286 unsigned PHINo = 0; 287 bool FoundMatch = true; 288 for (BasicBlock::iterator I = Dest->begin(); 289 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { 290 if (PHINo == TIPHIValues.size()) 291 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); 292 293 // If the PHI entry doesn't work, we can't use this pred. 294 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { 295 FoundMatch = false; 296 break; 297 } 298 } 299 300 // If we found a workable predecessor, change TI to branch to Succ. 301 if (FoundMatch) { 302 Dest->removePredecessor(TIBB); 303 TI->setSuccessor(SuccNum, Pred); 304 return; 305 } 306 } 307 308 SplitCriticalEdge(TI, SuccNum, P, true); 309} 310 311/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop 312/// copy (e.g. it's casting from one pointer type to another, int->uint, or 313/// int->sbyte on PPC), sink it into user blocks to reduce the number of virtual 314/// registers that must be created and coallesced. 315/// 316/// Return true if any changes are made. 317static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ 318 // If this is a noop copy, 319 MVT::ValueType SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 320 MVT::ValueType DstVT = TLI.getValueType(CI->getType()); 321 322 // This is an fp<->int conversion? 323 if (MVT::isInteger(SrcVT) != MVT::isInteger(DstVT)) 324 return false; 325 326 // If this is an extension, it will be a zero or sign extension, which 327 // isn't a noop. 328 if (SrcVT < DstVT) return false; 329 330 // If these values will be promoted, find out what they will be promoted 331 // to. This helps us consider truncates on PPC as noop copies when they 332 // are. 333 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote) 334 SrcVT = TLI.getTypeToTransformTo(SrcVT); 335 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote) 336 DstVT = TLI.getTypeToTransformTo(DstVT); 337 338 // If, after promotion, these are the same types, this is a noop copy. 339 if (SrcVT != DstVT) 340 return false; 341 342 BasicBlock *DefBB = CI->getParent(); 343 344 /// InsertedCasts - Only insert a cast in each block once. 345 std::map<BasicBlock*, CastInst*> InsertedCasts; 346 347 bool MadeChange = false; 348 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 349 UI != E; ) { 350 Use &TheUse = UI.getUse(); 351 Instruction *User = cast<Instruction>(*UI); 352 353 // Figure out which BB this cast is used in. For PHI's this is the 354 // appropriate predecessor block. 355 BasicBlock *UserBB = User->getParent(); 356 if (PHINode *PN = dyn_cast<PHINode>(User)) { 357 unsigned OpVal = UI.getOperandNo()/2; 358 UserBB = PN->getIncomingBlock(OpVal); 359 } 360 361 // Preincrement use iterator so we don't invalidate it. 362 ++UI; 363 364 // If this user is in the same block as the cast, don't change the cast. 365 if (UserBB == DefBB) continue; 366 367 // If we have already inserted a cast into this block, use it. 368 CastInst *&InsertedCast = InsertedCasts[UserBB]; 369 370 if (!InsertedCast) { 371 BasicBlock::iterator InsertPt = UserBB->begin(); 372 while (isa<PHINode>(InsertPt)) ++InsertPt; 373 374 InsertedCast = 375 CastInst::create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 376 InsertPt); 377 MadeChange = true; 378 } 379 380 // Replace a use of the cast with a use of the new casat. 381 TheUse = InsertedCast; 382 } 383 384 // If we removed all uses, nuke the cast. 385 if (CI->use_empty()) 386 CI->eraseFromParent(); 387 388 return MadeChange; 389} 390 391/// EraseDeadInstructions - Erase any dead instructions 392static void EraseDeadInstructions(Value *V) { 393 Instruction *I = dyn_cast<Instruction>(V); 394 if (!I || !I->use_empty()) return; 395 396 SmallPtrSet<Instruction*, 16> Insts; 397 Insts.insert(I); 398 399 while (!Insts.empty()) { 400 I = *Insts.begin(); 401 Insts.erase(I); 402 if (isInstructionTriviallyDead(I)) { 403 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) 404 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i))) 405 Insts.insert(U); 406 I->eraseFromParent(); 407 } 408 } 409} 410 411 412/// ExtAddrMode - This is an extended version of TargetLowering::AddrMode which 413/// holds actual Value*'s for register values. 414struct ExtAddrMode : public TargetLowering::AddrMode { 415 Value *BaseReg; 416 Value *ScaledReg; 417 ExtAddrMode() : BaseReg(0), ScaledReg(0) {} 418 void dump() const; 419}; 420 421static std::ostream &operator<<(std::ostream &OS, const ExtAddrMode &AM) { 422 bool NeedPlus = false; 423 OS << "["; 424 if (AM.BaseGV) 425 OS << (NeedPlus ? " + " : "") 426 << "GV:%" << AM.BaseGV->getName(), NeedPlus = true; 427 428 if (AM.BaseOffs) 429 OS << (NeedPlus ? " + " : "") << AM.BaseOffs, NeedPlus = true; 430 431 if (AM.BaseReg) 432 OS << (NeedPlus ? " + " : "") 433 << "Base:%" << AM.BaseReg->getName(), NeedPlus = true; 434 if (AM.Scale) 435 OS << (NeedPlus ? " + " : "") 436 << AM.Scale << "*%" << AM.ScaledReg->getName(), NeedPlus = true; 437 438 return OS << "]"; 439} 440 441void ExtAddrMode::dump() const { 442 cerr << *this << "\n"; 443} 444 445static bool TryMatchingScaledValue(Value *ScaleReg, int64_t Scale, 446 const Type *AccessTy, ExtAddrMode &AddrMode, 447 SmallVector<Instruction*, 16> &AddrModeInsts, 448 const TargetLowering &TLI, unsigned Depth); 449 450/// FindMaximalLegalAddressingMode - If we can, try to merge the computation of 451/// Addr into the specified addressing mode. If Addr can't be added to AddrMode 452/// this returns false. This assumes that Addr is either a pointer type or 453/// intptr_t for the target. 454static bool FindMaximalLegalAddressingMode(Value *Addr, const Type *AccessTy, 455 ExtAddrMode &AddrMode, 456 SmallVector<Instruction*, 16> &AddrModeInsts, 457 const TargetLowering &TLI, 458 unsigned Depth) { 459 460 // If this is a global variable, fold it into the addressing mode if possible. 461 if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) { 462 if (AddrMode.BaseGV == 0) { 463 AddrMode.BaseGV = GV; 464 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 465 return true; 466 AddrMode.BaseGV = 0; 467 } 468 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) { 469 AddrMode.BaseOffs += CI->getSExtValue(); 470 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 471 return true; 472 AddrMode.BaseOffs -= CI->getSExtValue(); 473 } else if (isa<ConstantPointerNull>(Addr)) { 474 return true; 475 } 476 477 // Look through constant exprs and instructions. 478 unsigned Opcode = ~0U; 479 User *AddrInst = 0; 480 if (Instruction *I = dyn_cast<Instruction>(Addr)) { 481 Opcode = I->getOpcode(); 482 AddrInst = I; 483 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) { 484 Opcode = CE->getOpcode(); 485 AddrInst = CE; 486 } 487 488 // Limit recursion to avoid exponential behavior. 489 if (Depth == 5) { AddrInst = 0; Opcode = ~0U; } 490 491 // If this is really an instruction, add it to our list of related 492 // instructions. 493 if (Instruction *I = dyn_cast_or_null<Instruction>(AddrInst)) 494 AddrModeInsts.push_back(I); 495 496 switch (Opcode) { 497 case Instruction::PtrToInt: 498 // PtrToInt is always a noop, as we know that the int type is pointer sized. 499 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 500 AddrMode, AddrModeInsts, TLI, Depth)) 501 return true; 502 break; 503 case Instruction::IntToPtr: 504 // This inttoptr is a no-op if the integer type is pointer sized. 505 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) == 506 TLI.getPointerTy()) { 507 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 508 AddrMode, AddrModeInsts, TLI, Depth)) 509 return true; 510 } 511 break; 512 case Instruction::Add: { 513 // Check to see if we can merge in the RHS then the LHS. If so, we win. 514 ExtAddrMode BackupAddrMode = AddrMode; 515 unsigned OldSize = AddrModeInsts.size(); 516 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(1), AccessTy, 517 AddrMode, AddrModeInsts, TLI, Depth+1) && 518 FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 519 AddrMode, AddrModeInsts, TLI, Depth+1)) 520 return true; 521 522 // Restore the old addr mode info. 523 AddrMode = BackupAddrMode; 524 AddrModeInsts.resize(OldSize); 525 526 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS. 527 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 528 AddrMode, AddrModeInsts, TLI, Depth+1) && 529 FindMaximalLegalAddressingMode(AddrInst->getOperand(1), AccessTy, 530 AddrMode, AddrModeInsts, TLI, Depth+1)) 531 return true; 532 533 // Otherwise we definitely can't merge the ADD in. 534 AddrMode = BackupAddrMode; 535 AddrModeInsts.resize(OldSize); 536 break; 537 } 538 case Instruction::Or: { 539 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); 540 if (!RHS) break; 541 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD. 542 break; 543 } 544 case Instruction::Mul: 545 case Instruction::Shl: { 546 // Can only handle X*C and X << C, and can only handle this when the scale 547 // field is available. 548 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); 549 if (!RHS) break; 550 int64_t Scale = RHS->getSExtValue(); 551 if (Opcode == Instruction::Shl) 552 Scale = 1 << Scale; 553 554 if (TryMatchingScaledValue(AddrInst->getOperand(0), Scale, AccessTy, 555 AddrMode, AddrModeInsts, TLI, Depth)) 556 return true; 557 break; 558 } 559 case Instruction::GetElementPtr: { 560 // Scan the GEP. We check it if it contains constant offsets and at most 561 // one variable offset. 562 int VariableOperand = -1; 563 unsigned VariableScale = 0; 564 565 int64_t ConstantOffset = 0; 566 const TargetData *TD = TLI.getTargetData(); 567 gep_type_iterator GTI = gep_type_begin(AddrInst); 568 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) { 569 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 570 const StructLayout *SL = TD->getStructLayout(STy); 571 unsigned Idx = 572 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue(); 573 ConstantOffset += SL->getElementOffset(Idx); 574 } else { 575 uint64_t TypeSize = TD->getTypeSize(GTI.getIndexedType()); 576 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) { 577 ConstantOffset += CI->getSExtValue()*TypeSize; 578 } else if (TypeSize) { // Scales of zero don't do anything. 579 // We only allow one variable index at the moment. 580 if (VariableOperand != -1) { 581 VariableOperand = -2; 582 break; 583 } 584 585 // Remember the variable index. 586 VariableOperand = i; 587 VariableScale = TypeSize; 588 } 589 } 590 } 591 592 // If the GEP had multiple variable indices, punt. 593 if (VariableOperand == -2) 594 break; 595 596 // A common case is for the GEP to only do a constant offset. In this case, 597 // just add it to the disp field and check validity. 598 if (VariableOperand == -1) { 599 AddrMode.BaseOffs += ConstantOffset; 600 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){ 601 // Check to see if we can fold the base pointer in too. 602 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 603 AddrMode, AddrModeInsts, TLI, 604 Depth+1)) 605 return true; 606 } 607 AddrMode.BaseOffs -= ConstantOffset; 608 } else { 609 // Check that this has no base reg yet. If so, we won't have a place to 610 // put the base of the GEP (assuming it is not a null ptr). 611 bool SetBaseReg = false; 612 if (AddrMode.HasBaseReg) { 613 if (!isa<ConstantPointerNull>(AddrInst->getOperand(0))) 614 break; 615 } else { 616 AddrMode.HasBaseReg = true; 617 AddrMode.BaseReg = AddrInst->getOperand(0); 618 SetBaseReg = true; 619 } 620 621 // See if the scale amount is valid for this target. 622 AddrMode.BaseOffs += ConstantOffset; 623 if (TryMatchingScaledValue(AddrInst->getOperand(VariableOperand), 624 VariableScale, AccessTy, AddrMode, 625 AddrModeInsts, TLI, Depth)) { 626 if (!SetBaseReg) return true; 627 628 // If this match succeeded, we know that we can form an address with the 629 // GepBase as the basereg. See if we can match *more*. 630 AddrMode.HasBaseReg = false; 631 AddrMode.BaseReg = 0; 632 if (FindMaximalLegalAddressingMode(AddrInst->getOperand(0), AccessTy, 633 AddrMode, AddrModeInsts, TLI, 634 Depth+1)) 635 return true; 636 // Strange, shouldn't happen. Restore the base reg and succeed the easy 637 // way. 638 AddrMode.HasBaseReg = true; 639 AddrMode.BaseReg = AddrInst->getOperand(0); 640 return true; 641 } 642 643 AddrMode.BaseOffs -= ConstantOffset; 644 if (SetBaseReg) { 645 AddrMode.HasBaseReg = false; 646 AddrMode.BaseReg = 0; 647 } 648 } 649 break; 650 } 651 } 652 653 if (Instruction *I = dyn_cast_or_null<Instruction>(AddrInst)) { 654 assert(AddrModeInsts.back() == I && "Stack imbalance"); 655 AddrModeInsts.pop_back(); 656 } 657 658 // Worse case, the target should support [reg] addressing modes. :) 659 if (!AddrMode.HasBaseReg) { 660 AddrMode.HasBaseReg = true; 661 // Still check for legality in case the target supports [imm] but not [i+r]. 662 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { 663 AddrMode.BaseReg = Addr; 664 return true; 665 } 666 AddrMode.HasBaseReg = false; 667 } 668 669 // If the base register is already taken, see if we can do [r+r]. 670 if (AddrMode.Scale == 0) { 671 AddrMode.Scale = 1; 672 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { 673 AddrMode.ScaledReg = Addr; 674 return true; 675 } 676 AddrMode.Scale = 0; 677 } 678 // Couldn't match. 679 return false; 680} 681 682/// TryMatchingScaledValue - Try adding ScaleReg*Scale to the specified 683/// addressing mode. Return true if this addr mode is legal for the target, 684/// false if not. 685static bool TryMatchingScaledValue(Value *ScaleReg, int64_t Scale, 686 const Type *AccessTy, ExtAddrMode &AddrMode, 687 SmallVector<Instruction*, 16> &AddrModeInsts, 688 const TargetLowering &TLI, unsigned Depth) { 689 // If we already have a scale of this value, we can add to it, otherwise, we 690 // need an available scale field. 691 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg) 692 return false; 693 694 ExtAddrMode InputAddrMode = AddrMode; 695 696 // Add scale to turn X*4+X*3 -> X*7. This could also do things like 697 // [A+B + A*7] -> [B+A*8]. 698 AddrMode.Scale += Scale; 699 AddrMode.ScaledReg = ScaleReg; 700 701 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) { 702 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now 703 // to see if ScaleReg is actually X+C. If so, we can turn this into adding 704 // X*Scale + C*Scale to addr mode. 705 BinaryOperator *BinOp = dyn_cast<BinaryOperator>(ScaleReg); 706 if (BinOp && BinOp->getOpcode() == Instruction::Add && 707 isa<ConstantInt>(BinOp->getOperand(1)) && InputAddrMode.ScaledReg ==0) { 708 709 InputAddrMode.Scale = Scale; 710 InputAddrMode.ScaledReg = BinOp->getOperand(0); 711 InputAddrMode.BaseOffs += 712 cast<ConstantInt>(BinOp->getOperand(1))->getSExtValue()*Scale; 713 if (TLI.isLegalAddressingMode(InputAddrMode, AccessTy)) { 714 AddrModeInsts.push_back(BinOp); 715 AddrMode = InputAddrMode; 716 return true; 717 } 718 } 719 720 // Otherwise, not (x+c)*scale, just return what we have. 721 return true; 722 } 723 724 // Otherwise, back this attempt out. 725 AddrMode.Scale -= Scale; 726 if (AddrMode.Scale == 0) AddrMode.ScaledReg = 0; 727 728 return false; 729} 730 731 732/// IsNonLocalValue - Return true if the specified values are defined in a 733/// different basic block than BB. 734static bool IsNonLocalValue(Value *V, BasicBlock *BB) { 735 if (Instruction *I = dyn_cast<Instruction>(V)) 736 return I->getParent() != BB; 737 return false; 738} 739 740/// OptimizeLoadStoreInst - Load and Store Instructions have often have 741/// addressing modes that can do significant amounts of computation. As such, 742/// instruction selection will try to get the load or store to do as much 743/// computation as possible for the program. The problem is that isel can only 744/// see within a single block. As such, we sink as much legal addressing mode 745/// stuff into the block as possible. 746bool CodeGenPrepare::OptimizeLoadStoreInst(Instruction *LdStInst, Value *Addr, 747 const Type *AccessTy, 748 DenseMap<Value*,Value*> &SunkAddrs) { 749 // Figure out what addressing mode will be built up for this operation. 750 SmallVector<Instruction*, 16> AddrModeInsts; 751 ExtAddrMode AddrMode; 752 bool Success = FindMaximalLegalAddressingMode(Addr, AccessTy, AddrMode, 753 AddrModeInsts, *TLI, 0); 754 Success = Success; assert(Success && "Couldn't select *anything*?"); 755 756 // Check to see if any of the instructions supersumed by this addr mode are 757 // non-local to I's BB. 758 bool AnyNonLocal = false; 759 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { 760 if (IsNonLocalValue(AddrModeInsts[i], LdStInst->getParent())) { 761 AnyNonLocal = true; 762 break; 763 } 764 } 765 766 // If all the instructions matched are already in this BB, don't do anything. 767 if (!AnyNonLocal) { 768 DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n"); 769 return false; 770 } 771 772 // Insert this computation right after this user. Since our caller is 773 // scanning from the top of the BB to the bottom, reuse of the expr are 774 // guaranteed to happen later. 775 BasicBlock::iterator InsertPt = LdStInst; 776 777 // Now that we determined the addressing expression we want to use and know 778 // that we have to sink it into this block. Check to see if we have already 779 // done this for some other load/store instr in this block. If so, reuse the 780 // computation. 781 Value *&SunkAddr = SunkAddrs[Addr]; 782 if (SunkAddr) { 783 DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << "\n"); 784 if (SunkAddr->getType() != Addr->getType()) 785 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt); 786 } else { 787 DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << "\n"); 788 const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType(); 789 790 Value *Result = 0; 791 // Start with the scale value. 792 if (AddrMode.Scale) { 793 Value *V = AddrMode.ScaledReg; 794 if (V->getType() == IntPtrTy) { 795 // done. 796 } else if (isa<PointerType>(V->getType())) { 797 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 798 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 799 cast<IntegerType>(V->getType())->getBitWidth()) { 800 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt); 801 } else { 802 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt); 803 } 804 if (AddrMode.Scale != 1) 805 V = BinaryOperator::createMul(V, ConstantInt::get(IntPtrTy, 806 AddrMode.Scale), 807 "sunkaddr", InsertPt); 808 Result = V; 809 } 810 811 // Add in the base register. 812 if (AddrMode.BaseReg) { 813 Value *V = AddrMode.BaseReg; 814 if (V->getType() != IntPtrTy) 815 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 816 if (Result) 817 Result = BinaryOperator::createAdd(Result, V, "sunkaddr", InsertPt); 818 else 819 Result = V; 820 } 821 822 // Add in the BaseGV if present. 823 if (AddrMode.BaseGV) { 824 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr", 825 InsertPt); 826 if (Result) 827 Result = BinaryOperator::createAdd(Result, V, "sunkaddr", InsertPt); 828 else 829 Result = V; 830 } 831 832 // Add in the Base Offset if present. 833 if (AddrMode.BaseOffs) { 834 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 835 if (Result) 836 Result = BinaryOperator::createAdd(Result, V, "sunkaddr", InsertPt); 837 else 838 Result = V; 839 } 840 841 if (Result == 0) 842 SunkAddr = Constant::getNullValue(Addr->getType()); 843 else 844 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt); 845 } 846 847 LdStInst->replaceUsesOfWith(Addr, SunkAddr); 848 849 if (Addr->use_empty()) 850 EraseDeadInstructions(Addr); 851 return true; 852} 853 854// In this pass we look for GEP and cast instructions that are used 855// across basic blocks and rewrite them to improve basic-block-at-a-time 856// selection. 857bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { 858 bool MadeChange = false; 859 860 // Split all critical edges where the dest block has a PHI and where the phi 861 // has shared immediate operands. 862 TerminatorInst *BBTI = BB.getTerminator(); 863 if (BBTI->getNumSuccessors() > 1) { 864 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) 865 if (isa<PHINode>(BBTI->getSuccessor(i)->begin()) && 866 isCriticalEdge(BBTI, i, true)) 867 SplitEdgeNicely(BBTI, i, this); 868 } 869 870 871 // Keep track of non-local addresses that have been sunk into this block. 872 // This allows us to avoid inserting duplicate code for blocks with multiple 873 // load/stores of the same address. 874 DenseMap<Value*, Value*> SunkAddrs; 875 876 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) { 877 Instruction *I = BBI++; 878 879 if (CastInst *CI = dyn_cast<CastInst>(I)) { 880 // If the source of the cast is a constant, then this should have 881 // already been constant folded. The only reason NOT to constant fold 882 // it is if something (e.g. LSR) was careful to place the constant 883 // evaluation in a block other than then one that uses it (e.g. to hoist 884 // the address of globals out of a loop). If this is the case, we don't 885 // want to forward-subst the cast. 886 if (isa<Constant>(CI->getOperand(0))) 887 continue; 888 889 if (TLI) 890 MadeChange |= OptimizeNoopCopyExpression(CI, *TLI); 891 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 892 if (TLI) 893 MadeChange |= OptimizeLoadStoreInst(I, I->getOperand(0), LI->getType(), 894 SunkAddrs); 895 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 896 if (TLI) 897 MadeChange |= OptimizeLoadStoreInst(I, SI->getOperand(1), 898 SI->getOperand(0)->getType(), 899 SunkAddrs); 900 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 901 if (GEPI->hasAllZeroIndices()) { 902 /// The GEP operand must be a pointer, so must its result -> BitCast 903 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 904 GEPI->getName(), GEPI); 905 GEPI->replaceAllUsesWith(NC); 906 GEPI->eraseFromParent(); 907 MadeChange = true; 908 BBI = NC; 909 } 910 } else if (CallInst *CI = dyn_cast<CallInst>(I)) { 911 // If we found an inline asm expession, and if the target knows how to 912 // lower it to normal LLVM code, do so now. 913 if (TLI && isa<InlineAsm>(CI->getCalledValue())) 914 if (const TargetAsmInfo *TAI = 915 TLI->getTargetMachine().getTargetAsmInfo()) { 916 if (TAI->ExpandInlineAsm(CI)) 917 BBI = BB.begin(); 918 } 919 } 920 } 921 922 return MadeChange; 923} 924 925