CodeGenPrepare.cpp revision f5102a0f088e7c96f7028bf7ca1c24975c314fff
1//===- CodeGenPrepare.cpp - Prepare a function for code generation --------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass munges the code in the input function to better prepare it for 11// SelectionDAG-based code generation. This works around limitations in it's 12// basic-block-at-a-time approach. It should eventually be removed. 13// 14//===----------------------------------------------------------------------===// 15 16#define DEBUG_TYPE "codegenprepare" 17#include "llvm/Transforms/Scalar.h" 18#include "llvm/Constants.h" 19#include "llvm/DerivedTypes.h" 20#include "llvm/Function.h" 21#include "llvm/InlineAsm.h" 22#include "llvm/Instructions.h" 23#include "llvm/Pass.h" 24#include "llvm/Target/TargetAsmInfo.h" 25#include "llvm/Target/TargetData.h" 26#include "llvm/Target/TargetLowering.h" 27#include "llvm/Target/TargetMachine.h" 28#include "llvm/Transforms/Utils/BasicBlockUtils.h" 29#include "llvm/Transforms/Utils/Local.h" 30#include "llvm/ADT/DenseMap.h" 31#include "llvm/ADT/SmallSet.h" 32#include "llvm/Support/CallSite.h" 33#include "llvm/Support/Compiler.h" 34#include "llvm/Support/Debug.h" 35#include "llvm/Support/GetElementPtrTypeIterator.h" 36#include "llvm/Support/PatternMatch.h" 37using namespace llvm; 38using namespace llvm::PatternMatch; 39 40namespace { 41 class VISIBILITY_HIDDEN CodeGenPrepare : public FunctionPass { 42 /// TLI - Keep a pointer of a TargetLowering to consult for determining 43 /// transformation profitability. 44 const TargetLowering *TLI; 45 public: 46 static char ID; // Pass identification, replacement for typeid 47 explicit CodeGenPrepare(const TargetLowering *tli = 0) 48 : FunctionPass(&ID), TLI(tli) {} 49 bool runOnFunction(Function &F); 50 51 private: 52 bool EliminateMostlyEmptyBlocks(Function &F); 53 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; 54 void EliminateMostlyEmptyBlock(BasicBlock *BB); 55 bool OptimizeBlock(BasicBlock &BB); 56 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy, 57 DenseMap<Value*,Value*> &SunkAddrs); 58 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS, 59 DenseMap<Value*,Value*> &SunkAddrs); 60 bool OptimizeExtUses(Instruction *I); 61 }; 62} 63 64char CodeGenPrepare::ID = 0; 65static RegisterPass<CodeGenPrepare> X("codegenprepare", 66 "Optimize for code generation"); 67 68FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) { 69 return new CodeGenPrepare(TLI); 70} 71 72 73bool CodeGenPrepare::runOnFunction(Function &F) { 74 bool EverMadeChange = false; 75 76 // First pass, eliminate blocks that contain only PHI nodes and an 77 // unconditional branch. 78 EverMadeChange |= EliminateMostlyEmptyBlocks(F); 79 80 bool MadeChange = true; 81 while (MadeChange) { 82 MadeChange = false; 83 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 84 MadeChange |= OptimizeBlock(*BB); 85 EverMadeChange |= MadeChange; 86 } 87 return EverMadeChange; 88} 89 90/// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes 91/// and an unconditional branch. Passes before isel (e.g. LSR/loopsimplify) 92/// often split edges in ways that are non-optimal for isel. Start by 93/// eliminating these blocks so we can split them the way we want them. 94bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) { 95 bool MadeChange = false; 96 // Note that this intentionally skips the entry block. 97 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) { 98 BasicBlock *BB = I++; 99 100 // If this block doesn't end with an uncond branch, ignore it. 101 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 102 if (!BI || !BI->isUnconditional()) 103 continue; 104 105 // If the instruction before the branch isn't a phi node, then other stuff 106 // is happening here. 107 BasicBlock::iterator BBI = BI; 108 if (BBI != BB->begin()) { 109 --BBI; 110 if (!isa<PHINode>(BBI)) continue; 111 } 112 113 // Do not break infinite loops. 114 BasicBlock *DestBB = BI->getSuccessor(0); 115 if (DestBB == BB) 116 continue; 117 118 if (!CanMergeBlocks(BB, DestBB)) 119 continue; 120 121 EliminateMostlyEmptyBlock(BB); 122 MadeChange = true; 123 } 124 return MadeChange; 125} 126 127/// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a 128/// single uncond branch between them, and BB contains no other non-phi 129/// instructions. 130bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB, 131 const BasicBlock *DestBB) const { 132 // We only want to eliminate blocks whose phi nodes are used by phi nodes in 133 // the successor. If there are more complex condition (e.g. preheaders), 134 // don't mess around with them. 135 BasicBlock::const_iterator BBI = BB->begin(); 136 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 137 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end(); 138 UI != E; ++UI) { 139 const Instruction *User = cast<Instruction>(*UI); 140 if (User->getParent() != DestBB || !isa<PHINode>(User)) 141 return false; 142 // If User is inside DestBB block and it is a PHINode then check 143 // incoming value. If incoming value is not from BB then this is 144 // a complex condition (e.g. preheaders) we want to avoid here. 145 if (User->getParent() == DestBB) { 146 if (const PHINode *UPN = dyn_cast<PHINode>(User)) 147 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { 148 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); 149 if (Insn && Insn->getParent() == BB && 150 Insn->getParent() != UPN->getIncomingBlock(I)) 151 return false; 152 } 153 } 154 } 155 } 156 157 // If BB and DestBB contain any common predecessors, then the phi nodes in BB 158 // and DestBB may have conflicting incoming values for the block. If so, we 159 // can't merge the block. 160 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); 161 if (!DestBBPN) return true; // no conflict. 162 163 // Collect the preds of BB. 164 SmallPtrSet<const BasicBlock*, 16> BBPreds; 165 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 166 // It is faster to get preds from a PHI than with pred_iterator. 167 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 168 BBPreds.insert(BBPN->getIncomingBlock(i)); 169 } else { 170 BBPreds.insert(pred_begin(BB), pred_end(BB)); 171 } 172 173 // Walk the preds of DestBB. 174 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { 175 BasicBlock *Pred = DestBBPN->getIncomingBlock(i); 176 if (BBPreds.count(Pred)) { // Common predecessor? 177 BBI = DestBB->begin(); 178 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 179 const Value *V1 = PN->getIncomingValueForBlock(Pred); 180 const Value *V2 = PN->getIncomingValueForBlock(BB); 181 182 // If V2 is a phi node in BB, look up what the mapped value will be. 183 if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) 184 if (V2PN->getParent() == BB) 185 V2 = V2PN->getIncomingValueForBlock(Pred); 186 187 // If there is a conflict, bail out. 188 if (V1 != V2) return false; 189 } 190 } 191 } 192 193 return true; 194} 195 196 197/// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and 198/// an unconditional branch in it. 199void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { 200 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 201 BasicBlock *DestBB = BI->getSuccessor(0); 202 203 DOUT << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB; 204 205 // If the destination block has a single pred, then this is a trivial edge, 206 // just collapse it. 207 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) { 208 if (SinglePred != DestBB) { 209 // Remember if SinglePred was the entry block of the function. If so, we 210 // will need to move BB back to the entry position. 211 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 212 MergeBasicBlockIntoOnlyPred(DestBB); 213 214 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 215 BB->moveBefore(&BB->getParent()->getEntryBlock()); 216 217 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 218 return; 219 } 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 its 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 // As a hack, never split backedges of loops. Even though the copy for any 272 // PHIs inserted on the backedge would be dead for exits from the loop, we 273 // assume that the cost of *splitting* the backedge would be too high. 274 if (Dest == TIBB) 275 return; 276 277 /// TIPHIValues - This array is lazily computed to determine the values of 278 /// PHIs in Dest that TI would provide. 279 SmallVector<Value*, 32> TIPHIValues; 280 281 // Check to see if Dest has any blocks that can be used as a split edge for 282 // this terminator. 283 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { 284 BasicBlock *Pred = *PI; 285 // To be usable, the pred has to end with an uncond branch to the dest. 286 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); 287 if (!PredBr || !PredBr->isUnconditional() || 288 // Must be empty other than the branch. 289 &Pred->front() != PredBr || 290 // Cannot be the entry block; its label does not get emitted. 291 Pred == &(Dest->getParent()->getEntryBlock())) 292 continue; 293 294 // Finally, since we know that Dest has phi nodes in it, we have to make 295 // sure that jumping to Pred will have the same affect as going to Dest in 296 // terms of PHI values. 297 PHINode *PN; 298 unsigned PHINo = 0; 299 bool FoundMatch = true; 300 for (BasicBlock::iterator I = Dest->begin(); 301 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { 302 if (PHINo == TIPHIValues.size()) 303 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); 304 305 // If the PHI entry doesn't work, we can't use this pred. 306 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { 307 FoundMatch = false; 308 break; 309 } 310 } 311 312 // If we found a workable predecessor, change TI to branch to Succ. 313 if (FoundMatch) { 314 Dest->removePredecessor(TIBB); 315 TI->setSuccessor(SuccNum, Pred); 316 return; 317 } 318 } 319 320 SplitCriticalEdge(TI, SuccNum, P, true); 321} 322 323/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop 324/// copy (e.g. it's casting from one pointer type to another, int->uint, or 325/// int->sbyte on PPC), sink it into user blocks to reduce the number of virtual 326/// registers that must be created and coalesced. 327/// 328/// Return true if any changes are made. 329/// 330static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ 331 // If this is a noop copy, 332 MVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 333 MVT DstVT = TLI.getValueType(CI->getType()); 334 335 // This is an fp<->int conversion? 336 if (SrcVT.isInteger() != DstVT.isInteger()) 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.bitsLT(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->getFirstNonPHI(); 385 386 InsertedCast = 387 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 388 InsertPt); 389 MadeChange = true; 390 } 391 392 // Replace a use of the cast with a use of the new cast. 393 TheUse = InsertedCast; 394 } 395 396 // If we removed all uses, nuke the cast. 397 if (CI->use_empty()) { 398 CI->eraseFromParent(); 399 MadeChange = true; 400 } 401 402 return MadeChange; 403} 404 405/// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce 406/// the number of virtual registers that must be created and coalesced. This is 407/// a clear win except on targets with multiple condition code registers 408/// (PowerPC), where it might lose; some adjustment may be wanted there. 409/// 410/// Return true if any changes are made. 411static bool OptimizeCmpExpression(CmpInst *CI) { 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->getFirstNonPHI(); 441 442 InsertedCmp = 443 CmpInst::Create(CI->getOpcode(), CI->getPredicate(), CI->getOperand(0), 444 CI->getOperand(1), "", InsertPt); 445 MadeChange = true; 446 } 447 448 // Replace a use of the cmp with a use of the new cmp. 449 TheUse = InsertedCmp; 450 } 451 452 // If we removed all uses, nuke the cmp. 453 if (CI->use_empty()) 454 CI->eraseFromParent(); 455 456 return MadeChange; 457} 458 459//===----------------------------------------------------------------------===// 460// Addressing Mode Analysis and Optimization 461//===----------------------------------------------------------------------===// 462 463namespace { 464 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode 465 /// which holds actual Value*'s for register values. 466 struct ExtAddrMode : public TargetLowering::AddrMode { 467 Value *BaseReg; 468 Value *ScaledReg; 469 ExtAddrMode() : BaseReg(0), ScaledReg(0) {} 470 void print(OStream &OS) const; 471 void dump() const { 472 print(cerr); 473 cerr << '\n'; 474 } 475 }; 476} // end anonymous namespace 477 478static inline OStream &operator<<(OStream &OS, const ExtAddrMode &AM) { 479 AM.print(OS); 480 return OS; 481} 482 483void ExtAddrMode::print(OStream &OS) const { 484 bool NeedPlus = false; 485 OS << "["; 486 if (BaseGV) 487 OS << (NeedPlus ? " + " : "") 488 << "GV:%" << BaseGV->getName(), NeedPlus = true; 489 490 if (BaseOffs) 491 OS << (NeedPlus ? " + " : "") << BaseOffs, NeedPlus = true; 492 493 if (BaseReg) 494 OS << (NeedPlus ? " + " : "") 495 << "Base:%" << BaseReg->getName(), NeedPlus = true; 496 if (Scale) 497 OS << (NeedPlus ? " + " : "") 498 << Scale << "*%" << ScaledReg->getName(), NeedPlus = true; 499 500 OS << ']'; 501} 502 503namespace { 504/// AddressingModeMatcher - This class exposes a single public method, which is 505/// used to construct a "maximal munch" of the addressing mode for the target 506/// specified by TLI for an access to "V" with an access type of AccessTy. This 507/// returns the addressing mode that is actually matched by value, but also 508/// returns the list of instructions involved in that addressing computation in 509/// AddrModeInsts. 510class AddressingModeMatcher { 511 SmallVectorImpl<Instruction*> &AddrModeInsts; 512 const TargetLowering &TLI; 513 514 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and 515 /// the memory instruction that we're computing this address for. 516 const Type *AccessTy; 517 Instruction *MemoryInst; 518 519 /// AddrMode - This is the addressing mode that we're building up. This is 520 /// part of the return value of this addressing mode matching stuff. 521 ExtAddrMode &AddrMode; 522 523 /// IgnoreProfitability - This is set to true when we should not do 524 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode 525 /// always returns true. 526 bool IgnoreProfitability; 527 528 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI, 529 const TargetLowering &T, const Type *AT, 530 Instruction *MI, ExtAddrMode &AM) 531 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM) { 532 IgnoreProfitability = false; 533 } 534public: 535 536 /// Match - Find the maximal addressing mode that a load/store of V can fold, 537 /// give an access type of AccessTy. This returns a list of involved 538 /// instructions in AddrModeInsts. 539 static ExtAddrMode Match(Value *V, const Type *AccessTy, 540 Instruction *MemoryInst, 541 SmallVectorImpl<Instruction*> &AddrModeInsts, 542 const TargetLowering &TLI) { 543 ExtAddrMode Result; 544 545 bool Success = 546 AddressingModeMatcher(AddrModeInsts, TLI, AccessTy, 547 MemoryInst, Result).MatchAddr(V, 0); 548 Success = Success; assert(Success && "Couldn't select *anything*?"); 549 return Result; 550 } 551private: 552 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth); 553 bool MatchAddr(Value *V, unsigned Depth); 554 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth); 555 bool IsProfitableToFoldIntoAddressingMode(Instruction *I, 556 ExtAddrMode &AMBefore, 557 ExtAddrMode &AMAfter); 558 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2); 559}; 560} // end anonymous namespace 561 562/// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode. 563/// Return true and update AddrMode if this addr mode is legal for the target, 564/// false if not. 565bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale, 566 unsigned Depth) { 567 // If Scale is 1, then this is the same as adding ScaleReg to the addressing 568 // mode. Just process that directly. 569 if (Scale == 1) 570 return MatchAddr(ScaleReg, Depth); 571 572 // If the scale is 0, it takes nothing to add this. 573 if (Scale == 0) 574 return true; 575 576 // If we already have a scale of this value, we can add to it, otherwise, we 577 // need an available scale field. 578 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg) 579 return false; 580 581 ExtAddrMode TestAddrMode = AddrMode; 582 583 // Add scale to turn X*4+X*3 -> X*7. This could also do things like 584 // [A+B + A*7] -> [B+A*8]. 585 TestAddrMode.Scale += Scale; 586 TestAddrMode.ScaledReg = ScaleReg; 587 588 // If the new address isn't legal, bail out. 589 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) 590 return false; 591 592 // It was legal, so commit it. 593 AddrMode = TestAddrMode; 594 595 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now 596 // to see if ScaleReg is actually X+C. If so, we can turn this into adding 597 // X*Scale + C*Scale to addr mode. 598 ConstantInt *CI; Value *AddLHS; 599 if (match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) { 600 TestAddrMode.ScaledReg = AddLHS; 601 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale; 602 603 // If this addressing mode is legal, commit it and remember that we folded 604 // this instruction. 605 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) { 606 AddrModeInsts.push_back(cast<Instruction>(ScaleReg)); 607 AddrMode = TestAddrMode; 608 return true; 609 } 610 } 611 612 // Otherwise, not (x+c)*scale, just return what we have. 613 return true; 614} 615 616/// MightBeFoldableInst - This is a little filter, which returns true if an 617/// addressing computation involving I might be folded into a load/store 618/// accessing it. This doesn't need to be perfect, but needs to accept at least 619/// the set of instructions that MatchOperationAddr can. 620static bool MightBeFoldableInst(Instruction *I) { 621 switch (I->getOpcode()) { 622 case Instruction::BitCast: 623 // Don't touch identity bitcasts. 624 if (I->getType() == I->getOperand(0)->getType()) 625 return false; 626 return isa<PointerType>(I->getType()) || isa<IntegerType>(I->getType()); 627 case Instruction::PtrToInt: 628 // PtrToInt is always a noop, as we know that the int type is pointer sized. 629 return true; 630 case Instruction::IntToPtr: 631 // We know the input is intptr_t, so this is foldable. 632 return true; 633 case Instruction::Add: 634 return true; 635 case Instruction::Mul: 636 case Instruction::Shl: 637 // Can only handle X*C and X << C. 638 return isa<ConstantInt>(I->getOperand(1)); 639 case Instruction::GetElementPtr: 640 return true; 641 default: 642 return false; 643 } 644} 645 646 647/// MatchOperationAddr - Given an instruction or constant expr, see if we can 648/// fold the operation into the addressing mode. If so, update the addressing 649/// mode and return true, otherwise return false without modifying AddrMode. 650bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode, 651 unsigned Depth) { 652 // Avoid exponential behavior on extremely deep expression trees. 653 if (Depth >= 5) return false; 654 655 switch (Opcode) { 656 case Instruction::PtrToInt: 657 // PtrToInt is always a noop, as we know that the int type is pointer sized. 658 return MatchAddr(AddrInst->getOperand(0), Depth); 659 case Instruction::IntToPtr: 660 // This inttoptr is a no-op if the integer type is pointer sized. 661 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) == 662 TLI.getPointerTy()) 663 return MatchAddr(AddrInst->getOperand(0), Depth); 664 return false; 665 case Instruction::BitCast: 666 // BitCast is always a noop, and we can handle it as long as it is 667 // int->int or pointer->pointer (we don't want int<->fp or something). 668 if ((isa<PointerType>(AddrInst->getOperand(0)->getType()) || 669 isa<IntegerType>(AddrInst->getOperand(0)->getType())) && 670 // Don't touch identity bitcasts. These were probably put here by LSR, 671 // and we don't want to mess around with them. Assume it knows what it 672 // is doing. 673 AddrInst->getOperand(0)->getType() != AddrInst->getType()) 674 return MatchAddr(AddrInst->getOperand(0), Depth); 675 return false; 676 case Instruction::Add: { 677 // Check to see if we can merge in the RHS then the LHS. If so, we win. 678 ExtAddrMode BackupAddrMode = AddrMode; 679 unsigned OldSize = AddrModeInsts.size(); 680 if (MatchAddr(AddrInst->getOperand(1), Depth+1) && 681 MatchAddr(AddrInst->getOperand(0), Depth+1)) 682 return true; 683 684 // Restore the old addr mode info. 685 AddrMode = BackupAddrMode; 686 AddrModeInsts.resize(OldSize); 687 688 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS. 689 if (MatchAddr(AddrInst->getOperand(0), Depth+1) && 690 MatchAddr(AddrInst->getOperand(1), Depth+1)) 691 return true; 692 693 // Otherwise we definitely can't merge the ADD in. 694 AddrMode = BackupAddrMode; 695 AddrModeInsts.resize(OldSize); 696 break; 697 } 698 //case Instruction::Or: 699 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD. 700 //break; 701 case Instruction::Mul: 702 case Instruction::Shl: { 703 // Can only handle X*C and X << C. 704 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); 705 if (!RHS) return false; 706 int64_t Scale = RHS->getSExtValue(); 707 if (Opcode == Instruction::Shl) 708 Scale = 1 << Scale; 709 710 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth); 711 } 712 case Instruction::GetElementPtr: { 713 // Scan the GEP. We check it if it contains constant offsets and at most 714 // one variable offset. 715 int VariableOperand = -1; 716 unsigned VariableScale = 0; 717 718 int64_t ConstantOffset = 0; 719 const TargetData *TD = TLI.getTargetData(); 720 gep_type_iterator GTI = gep_type_begin(AddrInst); 721 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) { 722 if (const StructType *STy = dyn_cast<StructType>(*GTI)) { 723 const StructLayout *SL = TD->getStructLayout(STy); 724 unsigned Idx = 725 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue(); 726 ConstantOffset += SL->getElementOffset(Idx); 727 } else { 728 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType()); 729 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) { 730 ConstantOffset += CI->getSExtValue()*TypeSize; 731 } else if (TypeSize) { // Scales of zero don't do anything. 732 // We only allow one variable index at the moment. 733 if (VariableOperand != -1) 734 return false; 735 736 // Remember the variable index. 737 VariableOperand = i; 738 VariableScale = TypeSize; 739 } 740 } 741 } 742 743 // A common case is for the GEP to only do a constant offset. In this case, 744 // just add it to the disp field and check validity. 745 if (VariableOperand == -1) { 746 AddrMode.BaseOffs += ConstantOffset; 747 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){ 748 // Check to see if we can fold the base pointer in too. 749 if (MatchAddr(AddrInst->getOperand(0), Depth+1)) 750 return true; 751 } 752 AddrMode.BaseOffs -= ConstantOffset; 753 return false; 754 } 755 756 // Save the valid addressing mode in case we can't match. 757 ExtAddrMode BackupAddrMode = AddrMode; 758 759 // Check that this has no base reg yet. If so, we won't have a place to 760 // put the base of the GEP (assuming it is not a null ptr). 761 bool SetBaseReg = true; 762 if (isa<ConstantPointerNull>(AddrInst->getOperand(0))) 763 SetBaseReg = false; // null pointer base doesn't need representation. 764 else if (AddrMode.HasBaseReg) 765 return false; // Base register already specified, can't match GEP. 766 else { 767 // Otherwise, we'll use the GEP base as the BaseReg. 768 AddrMode.HasBaseReg = true; 769 AddrMode.BaseReg = AddrInst->getOperand(0); 770 } 771 772 // See if the scale and offset amount is valid for this target. 773 AddrMode.BaseOffs += ConstantOffset; 774 775 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale, 776 Depth)) { 777 AddrMode = BackupAddrMode; 778 return false; 779 } 780 781 // If we have a null as the base of the GEP, folding in the constant offset 782 // plus variable scale is all we can do. 783 if (!SetBaseReg) return true; 784 785 // If this match succeeded, we know that we can form an address with the 786 // GepBase as the basereg. Match the base pointer of the GEP more 787 // aggressively by zeroing out BaseReg and rematching. If the base is 788 // (for example) another GEP, this allows merging in that other GEP into 789 // the addressing mode we're forming. 790 AddrMode.HasBaseReg = false; 791 AddrMode.BaseReg = 0; 792 bool Success = MatchAddr(AddrInst->getOperand(0), Depth+1); 793 assert(Success && "MatchAddr should be able to fill in BaseReg!"); 794 Success=Success; 795 return true; 796 } 797 } 798 return false; 799} 800 801/// MatchAddr - If we can, try to add the value of 'Addr' into the current 802/// addressing mode. If Addr can't be added to AddrMode this returns false and 803/// leaves AddrMode unmodified. This assumes that Addr is either a pointer type 804/// or intptr_t for the target. 805/// 806bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) { 807 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) { 808 // Fold in immediates if legal for the target. 809 AddrMode.BaseOffs += CI->getSExtValue(); 810 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 811 return true; 812 AddrMode.BaseOffs -= CI->getSExtValue(); 813 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) { 814 // If this is a global variable, try to fold it into the addressing mode. 815 if (AddrMode.BaseGV == 0) { 816 AddrMode.BaseGV = GV; 817 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 818 return true; 819 AddrMode.BaseGV = 0; 820 } 821 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) { 822 ExtAddrMode BackupAddrMode = AddrMode; 823 unsigned OldSize = AddrModeInsts.size(); 824 825 // Check to see if it is possible to fold this operation. 826 if (MatchOperationAddr(I, I->getOpcode(), Depth)) { 827 // Okay, it's possible to fold this. Check to see if it is actually 828 // *profitable* to do so. We use a simple cost model to avoid increasing 829 // register pressure too much. 830 if (I->hasOneUse() || 831 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) { 832 AddrModeInsts.push_back(I); 833 return true; 834 } 835 836 // It isn't profitable to do this, roll back. 837 //cerr << "NOT FOLDING: " << *I; 838 AddrMode = BackupAddrMode; 839 AddrModeInsts.resize(OldSize); 840 } 841 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) { 842 if (MatchOperationAddr(CE, CE->getOpcode(), Depth)) 843 return true; 844 } else if (isa<ConstantPointerNull>(Addr)) { 845 // Null pointer gets folded without affecting the addressing mode. 846 return true; 847 } 848 849 // Worse case, the target should support [reg] addressing modes. :) 850 if (!AddrMode.HasBaseReg) { 851 AddrMode.HasBaseReg = true; 852 AddrMode.BaseReg = Addr; 853 // Still check for legality in case the target supports [imm] but not [i+r]. 854 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 855 return true; 856 AddrMode.HasBaseReg = false; 857 AddrMode.BaseReg = 0; 858 } 859 860 // If the base register is already taken, see if we can do [r+r]. 861 if (AddrMode.Scale == 0) { 862 AddrMode.Scale = 1; 863 AddrMode.ScaledReg = Addr; 864 if (TLI.isLegalAddressingMode(AddrMode, AccessTy)) 865 return true; 866 AddrMode.Scale = 0; 867 AddrMode.ScaledReg = 0; 868 } 869 // Couldn't match. 870 return false; 871} 872 873 874/// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified 875/// inline asm call are due to memory operands. If so, return true, otherwise 876/// return false. 877static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal, 878 const TargetLowering &TLI) { 879 std::vector<InlineAsm::ConstraintInfo> 880 Constraints = IA->ParseConstraints(); 881 882 unsigned ArgNo = 1; // ArgNo - The operand of the CallInst. 883 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) { 884 TargetLowering::AsmOperandInfo OpInfo(Constraints[i]); 885 886 // Compute the value type for each operand. 887 switch (OpInfo.Type) { 888 case InlineAsm::isOutput: 889 if (OpInfo.isIndirect) 890 OpInfo.CallOperandVal = CI->getOperand(ArgNo++); 891 break; 892 case InlineAsm::isInput: 893 OpInfo.CallOperandVal = CI->getOperand(ArgNo++); 894 break; 895 case InlineAsm::isClobber: 896 // Nothing to do. 897 break; 898 } 899 900 // Compute the constraint code and ConstraintType to use. 901 TLI.ComputeConstraintToUse(OpInfo, SDValue(), 902 OpInfo.ConstraintType == TargetLowering::C_Memory); 903 904 // If this asm operand is our Value*, and if it isn't an indirect memory 905 // operand, we can't fold it! 906 if (OpInfo.CallOperandVal == OpVal && 907 (OpInfo.ConstraintType != TargetLowering::C_Memory || 908 !OpInfo.isIndirect)) 909 return false; 910 } 911 912 return true; 913} 914 915 916/// FindAllMemoryUses - Recursively walk all the uses of I until we find a 917/// memory use. If we find an obviously non-foldable instruction, return true. 918/// Add the ultimately found memory instructions to MemoryUses. 919static bool FindAllMemoryUses(Instruction *I, 920 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses, 921 SmallPtrSet<Instruction*, 16> &ConsideredInsts, 922 const TargetLowering &TLI) { 923 // If we already considered this instruction, we're done. 924 if (!ConsideredInsts.insert(I)) 925 return false; 926 927 // If this is an obviously unfoldable instruction, bail out. 928 if (!MightBeFoldableInst(I)) 929 return true; 930 931 // Loop over all the uses, recursively processing them. 932 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 933 UI != E; ++UI) { 934 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) { 935 MemoryUses.push_back(std::make_pair(LI, UI.getOperandNo())); 936 continue; 937 } 938 939 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) { 940 if (UI.getOperandNo() == 0) return true; // Storing addr, not into addr. 941 MemoryUses.push_back(std::make_pair(SI, UI.getOperandNo())); 942 continue; 943 } 944 945 if (CallInst *CI = dyn_cast<CallInst>(*UI)) { 946 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue()); 947 if (IA == 0) return true; 948 949 // If this is a memory operand, we're cool, otherwise bail out. 950 if (!IsOperandAMemoryOperand(CI, IA, I, TLI)) 951 return true; 952 continue; 953 } 954 955 if (FindAllMemoryUses(cast<Instruction>(*UI), MemoryUses, ConsideredInsts, 956 TLI)) 957 return true; 958 } 959 960 return false; 961} 962 963 964/// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at 965/// the use site that we're folding it into. If so, there is no cost to 966/// include it in the addressing mode. KnownLive1 and KnownLive2 are two values 967/// that we know are live at the instruction already. 968bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1, 969 Value *KnownLive2) { 970 // If Val is either of the known-live values, we know it is live! 971 if (Val == 0 || Val == KnownLive1 || Val == KnownLive2) 972 return true; 973 974 // All values other than instructions and arguments (e.g. constants) are live. 975 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true; 976 977 // If Val is a constant sized alloca in the entry block, it is live, this is 978 // true because it is just a reference to the stack/frame pointer, which is 979 // live for the whole function. 980 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val)) 981 if (AI->isStaticAlloca()) 982 return true; 983 984 // Check to see if this value is already used in the memory instruction's 985 // block. If so, it's already live into the block at the very least, so we 986 // can reasonably fold it. 987 BasicBlock *MemBB = MemoryInst->getParent(); 988 for (Value::use_iterator UI = Val->use_begin(), E = Val->use_end(); 989 UI != E; ++UI) 990 // We know that uses of arguments and instructions have to be instructions. 991 if (cast<Instruction>(*UI)->getParent() == MemBB) 992 return true; 993 994 return false; 995} 996 997 998 999/// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing 1000/// mode of the machine to fold the specified instruction into a load or store 1001/// that ultimately uses it. However, the specified instruction has multiple 1002/// uses. Given this, it may actually increase register pressure to fold it 1003/// into the load. For example, consider this code: 1004/// 1005/// X = ... 1006/// Y = X+1 1007/// use(Y) -> nonload/store 1008/// Z = Y+1 1009/// load Z 1010/// 1011/// In this case, Y has multiple uses, and can be folded into the load of Z 1012/// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to 1013/// be live at the use(Y) line. If we don't fold Y into load Z, we use one 1014/// fewer register. Since Y can't be folded into "use(Y)" we don't increase the 1015/// number of computations either. 1016/// 1017/// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If 1018/// X was live across 'load Z' for other reasons, we actually *would* want to 1019/// fold the addressing mode in the Z case. This would make Y die earlier. 1020bool AddressingModeMatcher:: 1021IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore, 1022 ExtAddrMode &AMAfter) { 1023 if (IgnoreProfitability) return true; 1024 1025 // AMBefore is the addressing mode before this instruction was folded into it, 1026 // and AMAfter is the addressing mode after the instruction was folded. Get 1027 // the set of registers referenced by AMAfter and subtract out those 1028 // referenced by AMBefore: this is the set of values which folding in this 1029 // address extends the lifetime of. 1030 // 1031 // Note that there are only two potential values being referenced here, 1032 // BaseReg and ScaleReg (global addresses are always available, as are any 1033 // folded immediates). 1034 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg; 1035 1036 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their 1037 // lifetime wasn't extended by adding this instruction. 1038 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg)) 1039 BaseReg = 0; 1040 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg)) 1041 ScaledReg = 0; 1042 1043 // If folding this instruction (and it's subexprs) didn't extend any live 1044 // ranges, we're ok with it. 1045 if (BaseReg == 0 && ScaledReg == 0) 1046 return true; 1047 1048 // If all uses of this instruction are ultimately load/store/inlineasm's, 1049 // check to see if their addressing modes will include this instruction. If 1050 // so, we can fold it into all uses, so it doesn't matter if it has multiple 1051 // uses. 1052 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses; 1053 SmallPtrSet<Instruction*, 16> ConsideredInsts; 1054 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI)) 1055 return false; // Has a non-memory, non-foldable use! 1056 1057 // Now that we know that all uses of this instruction are part of a chain of 1058 // computation involving only operations that could theoretically be folded 1059 // into a memory use, loop over each of these uses and see if they could 1060 // *actually* fold the instruction. 1061 SmallVector<Instruction*, 32> MatchedAddrModeInsts; 1062 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) { 1063 Instruction *User = MemoryUses[i].first; 1064 unsigned OpNo = MemoryUses[i].second; 1065 1066 // Get the access type of this use. If the use isn't a pointer, we don't 1067 // know what it accesses. 1068 Value *Address = User->getOperand(OpNo); 1069 if (!isa<PointerType>(Address->getType())) 1070 return false; 1071 const Type *AddressAccessTy = 1072 cast<PointerType>(Address->getType())->getElementType(); 1073 1074 // Do a match against the root of this address, ignoring profitability. This 1075 // will tell us if the addressing mode for the memory operation will 1076 // *actually* cover the shared instruction. 1077 ExtAddrMode Result; 1078 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy, 1079 MemoryInst, Result); 1080 Matcher.IgnoreProfitability = true; 1081 bool Success = Matcher.MatchAddr(Address, 0); 1082 Success = Success; assert(Success && "Couldn't select *anything*?"); 1083 1084 // If the match didn't cover I, then it won't be shared by it. 1085 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(), 1086 I) == MatchedAddrModeInsts.end()) 1087 return false; 1088 1089 MatchedAddrModeInsts.clear(); 1090 } 1091 1092 return true; 1093} 1094 1095 1096//===----------------------------------------------------------------------===// 1097// Memory Optimization 1098//===----------------------------------------------------------------------===// 1099 1100/// IsNonLocalValue - Return true if the specified values are defined in a 1101/// different basic block than BB. 1102static bool IsNonLocalValue(Value *V, BasicBlock *BB) { 1103 if (Instruction *I = dyn_cast<Instruction>(V)) 1104 return I->getParent() != BB; 1105 return false; 1106} 1107 1108/// OptimizeMemoryInst - Load and Store Instructions have often have 1109/// addressing modes that can do significant amounts of computation. As such, 1110/// instruction selection will try to get the load or store to do as much 1111/// computation as possible for the program. The problem is that isel can only 1112/// see within a single block. As such, we sink as much legal addressing mode 1113/// stuff into the block as possible. 1114/// 1115/// This method is used to optimize both load/store and inline asms with memory 1116/// operands. 1117bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr, 1118 const Type *AccessTy, 1119 DenseMap<Value*,Value*> &SunkAddrs) { 1120 // Figure out what addressing mode will be built up for this operation. 1121 SmallVector<Instruction*, 16> AddrModeInsts; 1122 ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst, 1123 AddrModeInsts, *TLI); 1124 1125 // Check to see if any of the instructions supersumed by this addr mode are 1126 // non-local to I's BB. 1127 bool AnyNonLocal = false; 1128 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { 1129 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) { 1130 AnyNonLocal = true; 1131 break; 1132 } 1133 } 1134 1135 // If all the instructions matched are already in this BB, don't do anything. 1136 if (!AnyNonLocal) { 1137 DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n"); 1138 return false; 1139 } 1140 1141 // Insert this computation right after this user. Since our caller is 1142 // scanning from the top of the BB to the bottom, reuse of the expr are 1143 // guaranteed to happen later. 1144 BasicBlock::iterator InsertPt = MemoryInst; 1145 1146 // Now that we determined the addressing expression we want to use and know 1147 // that we have to sink it into this block. Check to see if we have already 1148 // done this for some other load/store instr in this block. If so, reuse the 1149 // computation. 1150 Value *&SunkAddr = SunkAddrs[Addr]; 1151 if (SunkAddr) { 1152 DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << "\n"); 1153 if (SunkAddr->getType() != Addr->getType()) 1154 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt); 1155 } else { 1156 DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << "\n"); 1157 const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType(); 1158 1159 Value *Result = 0; 1160 // Start with the scale value. 1161 if (AddrMode.Scale) { 1162 Value *V = AddrMode.ScaledReg; 1163 if (V->getType() == IntPtrTy) { 1164 // done. 1165 } else if (isa<PointerType>(V->getType())) { 1166 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 1167 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 1168 cast<IntegerType>(V->getType())->getBitWidth()) { 1169 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt); 1170 } else { 1171 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt); 1172 } 1173 if (AddrMode.Scale != 1) 1174 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy, 1175 AddrMode.Scale), 1176 "sunkaddr", InsertPt); 1177 Result = V; 1178 } 1179 1180 // Add in the base register. 1181 if (AddrMode.BaseReg) { 1182 Value *V = AddrMode.BaseReg; 1183 if (V->getType() != IntPtrTy) 1184 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 1185 if (Result) 1186 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 1187 else 1188 Result = V; 1189 } 1190 1191 // Add in the BaseGV if present. 1192 if (AddrMode.BaseGV) { 1193 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr", 1194 InsertPt); 1195 if (Result) 1196 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 1197 else 1198 Result = V; 1199 } 1200 1201 // Add in the Base Offset if present. 1202 if (AddrMode.BaseOffs) { 1203 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 1204 if (Result) 1205 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 1206 else 1207 Result = V; 1208 } 1209 1210 if (Result == 0) 1211 SunkAddr = Constant::getNullValue(Addr->getType()); 1212 else 1213 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt); 1214 } 1215 1216 MemoryInst->replaceUsesOfWith(Addr, SunkAddr); 1217 1218 if (Addr->use_empty()) 1219 RecursivelyDeleteTriviallyDeadInstructions(Addr); 1220 return true; 1221} 1222 1223/// OptimizeInlineAsmInst - If there are any memory operands, use 1224/// OptimizeMemoryInst to sink their address computing into the block when 1225/// possible / profitable. 1226bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS, 1227 DenseMap<Value*,Value*> &SunkAddrs) { 1228 bool MadeChange = false; 1229 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 1230 1231 // Do a prepass over the constraints, canonicalizing them, and building up the 1232 // ConstraintOperands list. 1233 std::vector<InlineAsm::ConstraintInfo> 1234 ConstraintInfos = IA->ParseConstraints(); 1235 1236 /// ConstraintOperands - Information about all of the constraints. 1237 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands; 1238 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 1239 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 1240 ConstraintOperands. 1241 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i])); 1242 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back(); 1243 1244 // Compute the value type for each operand. 1245 switch (OpInfo.Type) { 1246 case InlineAsm::isOutput: 1247 if (OpInfo.isIndirect) 1248 OpInfo.CallOperandVal = CS.getArgument(ArgNo++); 1249 break; 1250 case InlineAsm::isInput: 1251 OpInfo.CallOperandVal = CS.getArgument(ArgNo++); 1252 break; 1253 case InlineAsm::isClobber: 1254 // Nothing to do. 1255 break; 1256 } 1257 1258 // Compute the constraint code and ConstraintType to use. 1259 TLI->ComputeConstraintToUse(OpInfo, SDValue(), 1260 OpInfo.ConstraintType == TargetLowering::C_Memory); 1261 1262 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 1263 OpInfo.isIndirect) { 1264 Value *OpVal = OpInfo.CallOperandVal; 1265 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs); 1266 } 1267 } 1268 1269 return MadeChange; 1270} 1271 1272bool CodeGenPrepare::OptimizeExtUses(Instruction *I) { 1273 BasicBlock *DefBB = I->getParent(); 1274 1275 // If both result of the {s|z}xt and its source are live out, rewrite all 1276 // other uses of the source with result of extension. 1277 Value *Src = I->getOperand(0); 1278 if (Src->hasOneUse()) 1279 return false; 1280 1281 // Only do this xform if truncating is free. 1282 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType())) 1283 return false; 1284 1285 // Only safe to perform the optimization if the source is also defined in 1286 // this block. 1287 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) 1288 return false; 1289 1290 bool DefIsLiveOut = false; 1291 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 1292 UI != E; ++UI) { 1293 Instruction *User = cast<Instruction>(*UI); 1294 1295 // Figure out which BB this ext is used in. 1296 BasicBlock *UserBB = User->getParent(); 1297 if (UserBB == DefBB) continue; 1298 DefIsLiveOut = true; 1299 break; 1300 } 1301 if (!DefIsLiveOut) 1302 return false; 1303 1304 // Make sure non of the uses are PHI nodes. 1305 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 1306 UI != E; ++UI) { 1307 Instruction *User = cast<Instruction>(*UI); 1308 BasicBlock *UserBB = User->getParent(); 1309 if (UserBB == DefBB) continue; 1310 // Be conservative. We don't want this xform to end up introducing 1311 // reloads just before load / store instructions. 1312 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User)) 1313 return false; 1314 } 1315 1316 // InsertedTruncs - Only insert one trunc in each block once. 1317 DenseMap<BasicBlock*, Instruction*> InsertedTruncs; 1318 1319 bool MadeChange = false; 1320 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 1321 UI != E; ++UI) { 1322 Use &TheUse = UI.getUse(); 1323 Instruction *User = cast<Instruction>(*UI); 1324 1325 // Figure out which BB this ext is used in. 1326 BasicBlock *UserBB = User->getParent(); 1327 if (UserBB == DefBB) continue; 1328 1329 // Both src and def are live in this block. Rewrite the use. 1330 Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; 1331 1332 if (!InsertedTrunc) { 1333 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 1334 1335 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt); 1336 } 1337 1338 // Replace a use of the {s|z}ext source with a use of the result. 1339 TheUse = InsertedTrunc; 1340 1341 MadeChange = true; 1342 } 1343 1344 return MadeChange; 1345} 1346 1347// In this pass we look for GEP and cast instructions that are used 1348// across basic blocks and rewrite them to improve basic-block-at-a-time 1349// selection. 1350bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { 1351 bool MadeChange = false; 1352 1353 // Split all critical edges where the dest block has a PHI and where the phi 1354 // has shared immediate operands. 1355 TerminatorInst *BBTI = BB.getTerminator(); 1356 if (BBTI->getNumSuccessors() > 1) { 1357 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) 1358 if (isa<PHINode>(BBTI->getSuccessor(i)->begin()) && 1359 isCriticalEdge(BBTI, i, true)) 1360 SplitEdgeNicely(BBTI, i, this); 1361 } 1362 1363 1364 // Keep track of non-local addresses that have been sunk into this block. 1365 // This allows us to avoid inserting duplicate code for blocks with multiple 1366 // load/stores of the same address. 1367 DenseMap<Value*, Value*> SunkAddrs; 1368 1369 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) { 1370 Instruction *I = BBI++; 1371 1372 if (CastInst *CI = dyn_cast<CastInst>(I)) { 1373 // If the source of the cast is a constant, then this should have 1374 // already been constant folded. The only reason NOT to constant fold 1375 // it is if something (e.g. LSR) was careful to place the constant 1376 // evaluation in a block other than then one that uses it (e.g. to hoist 1377 // the address of globals out of a loop). If this is the case, we don't 1378 // want to forward-subst the cast. 1379 if (isa<Constant>(CI->getOperand(0))) 1380 continue; 1381 1382 bool Change = false; 1383 if (TLI) { 1384 Change = OptimizeNoopCopyExpression(CI, *TLI); 1385 MadeChange |= Change; 1386 } 1387 1388 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) 1389 MadeChange |= OptimizeExtUses(I); 1390 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 1391 MadeChange |= OptimizeCmpExpression(CI); 1392 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1393 if (TLI) 1394 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(), 1395 SunkAddrs); 1396 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 1397 if (TLI) 1398 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1), 1399 SI->getOperand(0)->getType(), 1400 SunkAddrs); 1401 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 1402 if (GEPI->hasAllZeroIndices()) { 1403 /// The GEP operand must be a pointer, so must its result -> BitCast 1404 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 1405 GEPI->getName(), GEPI); 1406 GEPI->replaceAllUsesWith(NC); 1407 GEPI->eraseFromParent(); 1408 MadeChange = true; 1409 BBI = NC; 1410 } 1411 } else if (CallInst *CI = dyn_cast<CallInst>(I)) { 1412 // If we found an inline asm expession, and if the target knows how to 1413 // lower it to normal LLVM code, do so now. 1414 if (TLI && isa<InlineAsm>(CI->getCalledValue())) 1415 if (const TargetAsmInfo *TAI = 1416 TLI->getTargetMachine().getTargetAsmInfo()) { 1417 if (TAI->ExpandInlineAsm(CI)) 1418 BBI = BB.begin(); 1419 else 1420 // Sink address computing for memory operands into the block. 1421 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs); 1422 } 1423 } 1424 } 1425 1426 return MadeChange; 1427} 1428