CodeGenPrepare.cpp revision 8b0d4f61bbe07060a4638ae1d3731dec09d13854
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/IntrinsicInst.h" 24#include "llvm/Pass.h" 25#include "llvm/Target/TargetAsmInfo.h" 26#include "llvm/Target/TargetData.h" 27#include "llvm/Target/TargetLowering.h" 28#include "llvm/Target/TargetMachine.h" 29#include "llvm/Transforms/Utils/AddrModeMatcher.h" 30#include "llvm/Transforms/Utils/BasicBlockUtils.h" 31#include "llvm/Transforms/Utils/Local.h" 32#include "llvm/ADT/DenseMap.h" 33#include "llvm/ADT/SmallSet.h" 34#include "llvm/Assembly/Writer.h" 35#include "llvm/Support/CallSite.h" 36#include "llvm/Support/CommandLine.h" 37#include "llvm/Support/Compiler.h" 38#include "llvm/Support/Debug.h" 39#include "llvm/Support/GetElementPtrTypeIterator.h" 40#include "llvm/Support/PatternMatch.h" 41using namespace llvm; 42using namespace llvm::PatternMatch; 43 44static cl::opt<bool> FactorCommonPreds("split-critical-paths-tweak", 45 cl::init(false), cl::Hidden); 46 47namespace { 48 class VISIBILITY_HIDDEN CodeGenPrepare : public FunctionPass { 49 /// TLI - Keep a pointer of a TargetLowering to consult for determining 50 /// transformation profitability. 51 const TargetLowering *TLI; 52 53 /// BackEdges - Keep a set of all the loop back edges. 54 /// 55 SmallSet<std::pair<const BasicBlock*, const BasicBlock*>, 8> BackEdges; 56 public: 57 static char ID; // Pass identification, replacement for typeid 58 explicit CodeGenPrepare(const TargetLowering *tli = 0) 59 : FunctionPass(&ID), TLI(tli) {} 60 bool runOnFunction(Function &F); 61 62 private: 63 bool EliminateMostlyEmptyBlocks(Function &F); 64 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; 65 void EliminateMostlyEmptyBlock(BasicBlock *BB); 66 bool OptimizeBlock(BasicBlock &BB); 67 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy, 68 DenseMap<Value*,Value*> &SunkAddrs); 69 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS, 70 DenseMap<Value*,Value*> &SunkAddrs); 71 bool OptimizeExtUses(Instruction *I); 72 void findLoopBackEdges(const Function &F); 73 }; 74} 75 76char CodeGenPrepare::ID = 0; 77static RegisterPass<CodeGenPrepare> X("codegenprepare", 78 "Optimize for code generation"); 79 80FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) { 81 return new CodeGenPrepare(TLI); 82} 83 84/// findLoopBackEdges - Do a DFS walk to find loop back edges. 85/// 86void CodeGenPrepare::findLoopBackEdges(const Function &F) { 87 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 88 FindFunctionBackedges(F, Edges); 89 90 BackEdges.insert(Edges.begin(), Edges.end()); 91} 92 93 94bool CodeGenPrepare::runOnFunction(Function &F) { 95 bool EverMadeChange = false; 96 97 // First pass, eliminate blocks that contain only PHI nodes and an 98 // unconditional branch. 99 EverMadeChange |= EliminateMostlyEmptyBlocks(F); 100 101 // Now find loop back edges. 102 findLoopBackEdges(F); 103 104 bool MadeChange = true; 105 while (MadeChange) { 106 MadeChange = false; 107 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 108 MadeChange |= OptimizeBlock(*BB); 109 EverMadeChange |= MadeChange; 110 } 111 return EverMadeChange; 112} 113 114/// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes, 115/// debug info directives, and an unconditional branch. Passes before isel 116/// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for 117/// isel. Start by eliminating these blocks so we can split them the way we 118/// want them. 119bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) { 120 bool MadeChange = false; 121 // Note that this intentionally skips the entry block. 122 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) { 123 BasicBlock *BB = I++; 124 125 // If this block doesn't end with an uncond branch, ignore it. 126 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 127 if (!BI || !BI->isUnconditional()) 128 continue; 129 130 // If the instruction before the branch (skipping debug info) isn't a phi 131 // node, then other stuff is happening here. 132 BasicBlock::iterator BBI = BI; 133 if (BBI != BB->begin()) { 134 --BBI; 135 while (isa<DbgInfoIntrinsic>(BBI)) { 136 if (BBI == BB->begin()) 137 break; 138 --BBI; 139 } 140 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI)) 141 continue; 142 } 143 144 // Do not break infinite loops. 145 BasicBlock *DestBB = BI->getSuccessor(0); 146 if (DestBB == BB) 147 continue; 148 149 if (!CanMergeBlocks(BB, DestBB)) 150 continue; 151 152 EliminateMostlyEmptyBlock(BB); 153 MadeChange = true; 154 } 155 return MadeChange; 156} 157 158/// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a 159/// single uncond branch between them, and BB contains no other non-phi 160/// instructions. 161bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB, 162 const BasicBlock *DestBB) const { 163 // We only want to eliminate blocks whose phi nodes are used by phi nodes in 164 // the successor. If there are more complex condition (e.g. preheaders), 165 // don't mess around with them. 166 BasicBlock::const_iterator BBI = BB->begin(); 167 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 168 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end(); 169 UI != E; ++UI) { 170 const Instruction *User = cast<Instruction>(*UI); 171 if (User->getParent() != DestBB || !isa<PHINode>(User)) 172 return false; 173 // If User is inside DestBB block and it is a PHINode then check 174 // incoming value. If incoming value is not from BB then this is 175 // a complex condition (e.g. preheaders) we want to avoid here. 176 if (User->getParent() == DestBB) { 177 if (const PHINode *UPN = dyn_cast<PHINode>(User)) 178 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { 179 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); 180 if (Insn && Insn->getParent() == BB && 181 Insn->getParent() != UPN->getIncomingBlock(I)) 182 return false; 183 } 184 } 185 } 186 } 187 188 // If BB and DestBB contain any common predecessors, then the phi nodes in BB 189 // and DestBB may have conflicting incoming values for the block. If so, we 190 // can't merge the block. 191 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); 192 if (!DestBBPN) return true; // no conflict. 193 194 // Collect the preds of BB. 195 SmallPtrSet<const BasicBlock*, 16> BBPreds; 196 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 197 // It is faster to get preds from a PHI than with pred_iterator. 198 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 199 BBPreds.insert(BBPN->getIncomingBlock(i)); 200 } else { 201 BBPreds.insert(pred_begin(BB), pred_end(BB)); 202 } 203 204 // Walk the preds of DestBB. 205 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { 206 BasicBlock *Pred = DestBBPN->getIncomingBlock(i); 207 if (BBPreds.count(Pred)) { // Common predecessor? 208 BBI = DestBB->begin(); 209 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 210 const Value *V1 = PN->getIncomingValueForBlock(Pred); 211 const Value *V2 = PN->getIncomingValueForBlock(BB); 212 213 // If V2 is a phi node in BB, look up what the mapped value will be. 214 if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) 215 if (V2PN->getParent() == BB) 216 V2 = V2PN->getIncomingValueForBlock(Pred); 217 218 // If there is a conflict, bail out. 219 if (V1 != V2) return false; 220 } 221 } 222 } 223 224 return true; 225} 226 227 228/// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and 229/// an unconditional branch in it. 230void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { 231 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 232 BasicBlock *DestBB = BI->getSuccessor(0); 233 234 DOUT << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB; 235 236 // If the destination block has a single pred, then this is a trivial edge, 237 // just collapse it. 238 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) { 239 if (SinglePred != DestBB) { 240 // Remember if SinglePred was the entry block of the function. If so, we 241 // will need to move BB back to the entry position. 242 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 243 MergeBasicBlockIntoOnlyPred(DestBB); 244 245 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 246 BB->moveBefore(&BB->getParent()->getEntryBlock()); 247 248 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 249 return; 250 } 251 } 252 253 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB 254 // to handle the new incoming edges it is about to have. 255 PHINode *PN; 256 for (BasicBlock::iterator BBI = DestBB->begin(); 257 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 258 // Remove the incoming value for BB, and remember it. 259 Value *InVal = PN->removeIncomingValue(BB, false); 260 261 // Two options: either the InVal is a phi node defined in BB or it is some 262 // value that dominates BB. 263 PHINode *InValPhi = dyn_cast<PHINode>(InVal); 264 if (InValPhi && InValPhi->getParent() == BB) { 265 // Add all of the input values of the input PHI as inputs of this phi. 266 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) 267 PN->addIncoming(InValPhi->getIncomingValue(i), 268 InValPhi->getIncomingBlock(i)); 269 } else { 270 // Otherwise, add one instance of the dominating value for each edge that 271 // we will be adding. 272 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 273 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 274 PN->addIncoming(InVal, BBPN->getIncomingBlock(i)); 275 } else { 276 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 277 PN->addIncoming(InVal, *PI); 278 } 279 } 280 } 281 282 // The PHIs are now updated, change everything that refers to BB to use 283 // DestBB and remove BB. 284 BB->replaceAllUsesWith(DestBB); 285 BB->eraseFromParent(); 286 287 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 288} 289 290 291/// SplitEdgeNicely - Split the critical edge from TI to its specified 292/// successor if it will improve codegen. We only do this if the successor has 293/// phi nodes (otherwise critical edges are ok). If there is already another 294/// predecessor of the succ that is empty (and thus has no phi nodes), use it 295/// instead of introducing a new block. 296static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, 297 SmallSet<std::pair<const BasicBlock*, 298 const BasicBlock*>, 8> &BackEdges, 299 Pass *P) { 300 BasicBlock *TIBB = TI->getParent(); 301 BasicBlock *Dest = TI->getSuccessor(SuccNum); 302 assert(isa<PHINode>(Dest->begin()) && 303 "This should only be called if Dest has a PHI!"); 304 305 // Do not split edges to EH landing pads. 306 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI)) { 307 if (Invoke->getSuccessor(1) == Dest) 308 return; 309 } 310 311 // As a hack, never split backedges of loops. Even though the copy for any 312 // PHIs inserted on the backedge would be dead for exits from the loop, we 313 // assume that the cost of *splitting* the backedge would be too high. 314 if (BackEdges.count(std::make_pair(TIBB, Dest))) 315 return; 316 317 if (!FactorCommonPreds) { 318 /// TIPHIValues - This array is lazily computed to determine the values of 319 /// PHIs in Dest that TI would provide. 320 SmallVector<Value*, 32> TIPHIValues; 321 322 // Check to see if Dest has any blocks that can be used as a split edge for 323 // this terminator. 324 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { 325 BasicBlock *Pred = *PI; 326 // To be usable, the pred has to end with an uncond branch to the dest. 327 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); 328 if (!PredBr || !PredBr->isUnconditional()) 329 continue; 330 // Must be empty other than the branch and debug info. 331 BasicBlock::iterator I = Pred->begin(); 332 while (isa<DbgInfoIntrinsic>(I)) 333 I++; 334 if (dyn_cast<Instruction>(I) != PredBr) 335 continue; 336 // Cannot be the entry block; its label does not get emitted. 337 if (Pred == &(Dest->getParent()->getEntryBlock())) 338 continue; 339 340 // Finally, since we know that Dest has phi nodes in it, we have to make 341 // sure that jumping to Pred will have the same effect as going to Dest in 342 // terms of PHI values. 343 PHINode *PN; 344 unsigned PHINo = 0; 345 bool FoundMatch = true; 346 for (BasicBlock::iterator I = Dest->begin(); 347 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { 348 if (PHINo == TIPHIValues.size()) 349 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); 350 351 // If the PHI entry doesn't work, we can't use this pred. 352 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { 353 FoundMatch = false; 354 break; 355 } 356 } 357 358 // If we found a workable predecessor, change TI to branch to Succ. 359 if (FoundMatch) { 360 Dest->removePredecessor(TIBB); 361 TI->setSuccessor(SuccNum, Pred); 362 return; 363 } 364 } 365 366 SplitCriticalEdge(TI, SuccNum, P, true); 367 return; 368 } 369 370 PHINode *PN; 371 SmallVector<Value*, 8> TIPHIValues; 372 for (BasicBlock::iterator I = Dest->begin(); 373 (PN = dyn_cast<PHINode>(I)); ++I) 374 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); 375 376 SmallVector<BasicBlock*, 8> IdenticalPreds; 377 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { 378 BasicBlock *Pred = *PI; 379 if (BackEdges.count(std::make_pair(Pred, Dest))) 380 continue; 381 if (PI == TIBB) 382 IdenticalPreds.push_back(Pred); 383 else { 384 bool Identical = true; 385 unsigned PHINo = 0; 386 for (BasicBlock::iterator I = Dest->begin(); 387 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) 388 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { 389 Identical = false; 390 break; 391 } 392 if (Identical) 393 IdenticalPreds.push_back(Pred); 394 } 395 } 396 397 assert(!IdenticalPreds.empty()); 398 SplitBlockPredecessors(Dest, &IdenticalPreds[0], IdenticalPreds.size(), 399 ".critedge", P); 400} 401 402 403/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop 404/// copy (e.g. it's casting from one pointer type to another, int->uint, or 405/// int->sbyte on PPC), sink it into user blocks to reduce the number of virtual 406/// registers that must be created and coalesced. 407/// 408/// Return true if any changes are made. 409/// 410static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ 411 // If this is a noop copy, 412 MVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 413 MVT DstVT = TLI.getValueType(CI->getType()); 414 415 // This is an fp<->int conversion? 416 if (SrcVT.isInteger() != DstVT.isInteger()) 417 return false; 418 419 // If this is an extension, it will be a zero or sign extension, which 420 // isn't a noop. 421 if (SrcVT.bitsLT(DstVT)) return false; 422 423 // If these values will be promoted, find out what they will be promoted 424 // to. This helps us consider truncates on PPC as noop copies when they 425 // are. 426 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote) 427 SrcVT = TLI.getTypeToTransformTo(SrcVT); 428 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote) 429 DstVT = TLI.getTypeToTransformTo(DstVT); 430 431 // If, after promotion, these are the same types, this is a noop copy. 432 if (SrcVT != DstVT) 433 return false; 434 435 BasicBlock *DefBB = CI->getParent(); 436 437 /// InsertedCasts - Only insert a cast in each block once. 438 DenseMap<BasicBlock*, CastInst*> InsertedCasts; 439 440 bool MadeChange = false; 441 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 442 UI != E; ) { 443 Use &TheUse = UI.getUse(); 444 Instruction *User = cast<Instruction>(*UI); 445 446 // Figure out which BB this cast is used in. For PHI's this is the 447 // appropriate predecessor block. 448 BasicBlock *UserBB = User->getParent(); 449 if (PHINode *PN = dyn_cast<PHINode>(User)) { 450 UserBB = PN->getIncomingBlock(UI); 451 } 452 453 // Preincrement use iterator so we don't invalidate it. 454 ++UI; 455 456 // If this user is in the same block as the cast, don't change the cast. 457 if (UserBB == DefBB) continue; 458 459 // If we have already inserted a cast into this block, use it. 460 CastInst *&InsertedCast = InsertedCasts[UserBB]; 461 462 if (!InsertedCast) { 463 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 464 465 InsertedCast = 466 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 467 InsertPt); 468 MadeChange = true; 469 } 470 471 // Replace a use of the cast with a use of the new cast. 472 TheUse = InsertedCast; 473 } 474 475 // If we removed all uses, nuke the cast. 476 if (CI->use_empty()) { 477 CI->eraseFromParent(); 478 MadeChange = true; 479 } 480 481 return MadeChange; 482} 483 484/// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce 485/// the number of virtual registers that must be created and coalesced. This is 486/// a clear win except on targets with multiple condition code registers 487/// (PowerPC), where it might lose; some adjustment may be wanted there. 488/// 489/// Return true if any changes are made. 490static bool OptimizeCmpExpression(CmpInst *CI) { 491 BasicBlock *DefBB = CI->getParent(); 492 493 /// InsertedCmp - Only insert a cmp in each block once. 494 DenseMap<BasicBlock*, CmpInst*> InsertedCmps; 495 496 bool MadeChange = false; 497 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 498 UI != E; ) { 499 Use &TheUse = UI.getUse(); 500 Instruction *User = cast<Instruction>(*UI); 501 502 // Preincrement use iterator so we don't invalidate it. 503 ++UI; 504 505 // Don't bother for PHI nodes. 506 if (isa<PHINode>(User)) 507 continue; 508 509 // Figure out which BB this cmp is used in. 510 BasicBlock *UserBB = User->getParent(); 511 512 // If this user is in the same block as the cmp, don't change the cmp. 513 if (UserBB == DefBB) continue; 514 515 // If we have already inserted a cmp into this block, use it. 516 CmpInst *&InsertedCmp = InsertedCmps[UserBB]; 517 518 if (!InsertedCmp) { 519 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 520 521 InsertedCmp = 522 CmpInst::Create(CI->getOpcode(), CI->getPredicate(), CI->getOperand(0), 523 CI->getOperand(1), "", InsertPt); 524 MadeChange = true; 525 } 526 527 // Replace a use of the cmp with a use of the new cmp. 528 TheUse = InsertedCmp; 529 } 530 531 // If we removed all uses, nuke the cmp. 532 if (CI->use_empty()) 533 CI->eraseFromParent(); 534 535 return MadeChange; 536} 537 538//===----------------------------------------------------------------------===// 539// Memory Optimization 540//===----------------------------------------------------------------------===// 541 542/// IsNonLocalValue - Return true if the specified values are defined in a 543/// different basic block than BB. 544static bool IsNonLocalValue(Value *V, BasicBlock *BB) { 545 if (Instruction *I = dyn_cast<Instruction>(V)) 546 return I->getParent() != BB; 547 return false; 548} 549 550/// OptimizeMemoryInst - Load and Store Instructions have often have 551/// addressing modes that can do significant amounts of computation. As such, 552/// instruction selection will try to get the load or store to do as much 553/// computation as possible for the program. The problem is that isel can only 554/// see within a single block. As such, we sink as much legal addressing mode 555/// stuff into the block as possible. 556/// 557/// This method is used to optimize both load/store and inline asms with memory 558/// operands. 559bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr, 560 const Type *AccessTy, 561 DenseMap<Value*,Value*> &SunkAddrs) { 562 // Figure out what addressing mode will be built up for this operation. 563 SmallVector<Instruction*, 16> AddrModeInsts; 564 ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst, 565 AddrModeInsts, *TLI); 566 567 // Check to see if any of the instructions supersumed by this addr mode are 568 // non-local to I's BB. 569 bool AnyNonLocal = false; 570 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { 571 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) { 572 AnyNonLocal = true; 573 break; 574 } 575 } 576 577 // If all the instructions matched are already in this BB, don't do anything. 578 if (!AnyNonLocal) { 579 DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n"); 580 return false; 581 } 582 583 // Insert this computation right after this user. Since our caller is 584 // scanning from the top of the BB to the bottom, reuse of the expr are 585 // guaranteed to happen later. 586 BasicBlock::iterator InsertPt = MemoryInst; 587 588 // Now that we determined the addressing expression we want to use and know 589 // that we have to sink it into this block. Check to see if we have already 590 // done this for some other load/store instr in this block. If so, reuse the 591 // computation. 592 Value *&SunkAddr = SunkAddrs[Addr]; 593 if (SunkAddr) { 594 DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for " 595 << *MemoryInst); 596 if (SunkAddr->getType() != Addr->getType()) 597 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt); 598 } else { 599 DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " 600 << *MemoryInst); 601 const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType(); 602 603 Value *Result = 0; 604 // Start with the scale value. 605 if (AddrMode.Scale) { 606 Value *V = AddrMode.ScaledReg; 607 if (V->getType() == IntPtrTy) { 608 // done. 609 } else if (isa<PointerType>(V->getType())) { 610 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 611 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 612 cast<IntegerType>(V->getType())->getBitWidth()) { 613 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt); 614 } else { 615 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt); 616 } 617 if (AddrMode.Scale != 1) 618 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy, 619 AddrMode.Scale), 620 "sunkaddr", InsertPt); 621 Result = V; 622 } 623 624 // Add in the base register. 625 if (AddrMode.BaseReg) { 626 Value *V = AddrMode.BaseReg; 627 if (isa<PointerType>(V->getType())) 628 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 629 if (V->getType() != IntPtrTy) 630 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true, 631 "sunkaddr", InsertPt); 632 if (Result) 633 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 634 else 635 Result = V; 636 } 637 638 // Add in the BaseGV if present. 639 if (AddrMode.BaseGV) { 640 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr", 641 InsertPt); 642 if (Result) 643 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 644 else 645 Result = V; 646 } 647 648 // Add in the Base Offset if present. 649 if (AddrMode.BaseOffs) { 650 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 651 if (Result) 652 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 653 else 654 Result = V; 655 } 656 657 if (Result == 0) 658 SunkAddr = Constant::getNullValue(Addr->getType()); 659 else 660 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt); 661 } 662 663 MemoryInst->replaceUsesOfWith(Addr, SunkAddr); 664 665 if (Addr->use_empty()) 666 RecursivelyDeleteTriviallyDeadInstructions(Addr); 667 return true; 668} 669 670/// OptimizeInlineAsmInst - If there are any memory operands, use 671/// OptimizeMemoryInst to sink their address computing into the block when 672/// possible / profitable. 673bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS, 674 DenseMap<Value*,Value*> &SunkAddrs) { 675 bool MadeChange = false; 676 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 677 678 // Do a prepass over the constraints, canonicalizing them, and building up the 679 // ConstraintOperands list. 680 std::vector<InlineAsm::ConstraintInfo> 681 ConstraintInfos = IA->ParseConstraints(); 682 683 /// ConstraintOperands - Information about all of the constraints. 684 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands; 685 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 686 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 687 ConstraintOperands. 688 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i])); 689 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back(); 690 691 // Compute the value type for each operand. 692 switch (OpInfo.Type) { 693 case InlineAsm::isOutput: 694 if (OpInfo.isIndirect) 695 OpInfo.CallOperandVal = CS.getArgument(ArgNo++); 696 break; 697 case InlineAsm::isInput: 698 OpInfo.CallOperandVal = CS.getArgument(ArgNo++); 699 break; 700 case InlineAsm::isClobber: 701 // Nothing to do. 702 break; 703 } 704 705 // Compute the constraint code and ConstraintType to use. 706 TLI->ComputeConstraintToUse(OpInfo, SDValue(), 707 OpInfo.ConstraintType == TargetLowering::C_Memory); 708 709 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 710 OpInfo.isIndirect) { 711 Value *OpVal = OpInfo.CallOperandVal; 712 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs); 713 } 714 } 715 716 return MadeChange; 717} 718 719bool CodeGenPrepare::OptimizeExtUses(Instruction *I) { 720 BasicBlock *DefBB = I->getParent(); 721 722 // If both result of the {s|z}xt and its source are live out, rewrite all 723 // other uses of the source with result of extension. 724 Value *Src = I->getOperand(0); 725 if (Src->hasOneUse()) 726 return false; 727 728 // Only do this xform if truncating is free. 729 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType())) 730 return false; 731 732 // Only safe to perform the optimization if the source is also defined in 733 // this block. 734 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) 735 return false; 736 737 bool DefIsLiveOut = false; 738 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 739 UI != E; ++UI) { 740 Instruction *User = cast<Instruction>(*UI); 741 742 // Figure out which BB this ext is used in. 743 BasicBlock *UserBB = User->getParent(); 744 if (UserBB == DefBB) continue; 745 DefIsLiveOut = true; 746 break; 747 } 748 if (!DefIsLiveOut) 749 return false; 750 751 // Make sure non of the uses are PHI nodes. 752 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 753 UI != E; ++UI) { 754 Instruction *User = cast<Instruction>(*UI); 755 BasicBlock *UserBB = User->getParent(); 756 if (UserBB == DefBB) continue; 757 // Be conservative. We don't want this xform to end up introducing 758 // reloads just before load / store instructions. 759 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User)) 760 return false; 761 } 762 763 // InsertedTruncs - Only insert one trunc in each block once. 764 DenseMap<BasicBlock*, Instruction*> InsertedTruncs; 765 766 bool MadeChange = false; 767 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 768 UI != E; ++UI) { 769 Use &TheUse = UI.getUse(); 770 Instruction *User = cast<Instruction>(*UI); 771 772 // Figure out which BB this ext is used in. 773 BasicBlock *UserBB = User->getParent(); 774 if (UserBB == DefBB) continue; 775 776 // Both src and def are live in this block. Rewrite the use. 777 Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; 778 779 if (!InsertedTrunc) { 780 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 781 782 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt); 783 } 784 785 // Replace a use of the {s|z}ext source with a use of the result. 786 TheUse = InsertedTrunc; 787 788 MadeChange = true; 789 } 790 791 return MadeChange; 792} 793 794// In this pass we look for GEP and cast instructions that are used 795// across basic blocks and rewrite them to improve basic-block-at-a-time 796// selection. 797bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { 798 bool MadeChange = false; 799 800 // Split all critical edges where the dest block has a PHI. 801 TerminatorInst *BBTI = BB.getTerminator(); 802 if (BBTI->getNumSuccessors() > 1) { 803 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) { 804 BasicBlock *SuccBB = BBTI->getSuccessor(i); 805 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true)) 806 SplitEdgeNicely(BBTI, i, BackEdges, this); 807 } 808 } 809 810 // Keep track of non-local addresses that have been sunk into this block. 811 // This allows us to avoid inserting duplicate code for blocks with multiple 812 // load/stores of the same address. 813 DenseMap<Value*, Value*> SunkAddrs; 814 815 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) { 816 Instruction *I = BBI++; 817 818 if (CastInst *CI = dyn_cast<CastInst>(I)) { 819 // If the source of the cast is a constant, then this should have 820 // already been constant folded. The only reason NOT to constant fold 821 // it is if something (e.g. LSR) was careful to place the constant 822 // evaluation in a block other than then one that uses it (e.g. to hoist 823 // the address of globals out of a loop). If this is the case, we don't 824 // want to forward-subst the cast. 825 if (isa<Constant>(CI->getOperand(0))) 826 continue; 827 828 bool Change = false; 829 if (TLI) { 830 Change = OptimizeNoopCopyExpression(CI, *TLI); 831 MadeChange |= Change; 832 } 833 834 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) 835 MadeChange |= OptimizeExtUses(I); 836 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 837 MadeChange |= OptimizeCmpExpression(CI); 838 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 839 if (TLI) 840 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(), 841 SunkAddrs); 842 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 843 if (TLI) 844 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1), 845 SI->getOperand(0)->getType(), 846 SunkAddrs); 847 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 848 if (GEPI->hasAllZeroIndices()) { 849 /// The GEP operand must be a pointer, so must its result -> BitCast 850 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 851 GEPI->getName(), GEPI); 852 GEPI->replaceAllUsesWith(NC); 853 GEPI->eraseFromParent(); 854 MadeChange = true; 855 BBI = NC; 856 } 857 } else if (CallInst *CI = dyn_cast<CallInst>(I)) { 858 // If we found an inline asm expession, and if the target knows how to 859 // lower it to normal LLVM code, do so now. 860 if (TLI && isa<InlineAsm>(CI->getCalledValue())) 861 if (const TargetAsmInfo *TAI = 862 TLI->getTargetMachine().getTargetAsmInfo()) { 863 if (TAI->ExpandInlineAsm(CI)) { 864 BBI = BB.begin(); 865 // Avoid processing instructions out of order, which could cause 866 // reuse before a value is defined. 867 SunkAddrs.clear(); 868 } else 869 // Sink address computing for memory operands into the block. 870 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs); 871 } 872 } 873 } 874 875 return MadeChange; 876} 877