CodeGenPrepare.cpp revision eed707b1e6097aac2bb6b3d47271f6300ace7f2e
1a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block//===- CodeGenPrepare.cpp - Prepare a function for code generation --------===// 2a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block// 3a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block// The LLVM Compiler Infrastructure 4a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block// 5a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block// This file is distributed under the University of Illinois Open Source 6ab9e7a118cf1ea2e3a93dce683b2ded3e7291ddbBen Murdoch// License. See LICENSE.TXT for details. 7ab9e7a118cf1ea2e3a93dce683b2ded3e7291ddbBen Murdoch// 8a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block//===----------------------------------------------------------------------===// 9a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block// 10a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block// This pass munges the code in the input function to better prepare it for 11a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block// SelectionDAG-based code generation. This works around limitations in it's 12a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block// basic-block-at-a-time approach. It should eventually be removed. 13a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block// 14a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block//===----------------------------------------------------------------------===// 15a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block 16a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#define DEBUG_TYPE "codegenprepare" 17a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Transforms/Scalar.h" 18a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Constants.h" 19a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/DerivedTypes.h" 20a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Function.h" 21a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/InlineAsm.h" 22a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Instructions.h" 23a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/IntrinsicInst.h" 24a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/LLVMContext.h" 25a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Pass.h" 26a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Target/TargetData.h" 27a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Target/TargetLowering.h" 28a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Transforms/Utils/AddrModeMatcher.h" 29a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Transforms/Utils/BasicBlockUtils.h" 30a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Transforms/Utils/Local.h" 31a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/ADT/DenseMap.h" 32a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/ADT/SmallSet.h" 33a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Assembly/Writer.h" 34a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Support/CallSite.h" 35a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Support/CommandLine.h" 36a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Support/Compiler.h" 37a9bfd6c4a32dfd9cc032cb67c6ccb8d09c16f579Steve Block#include "llvm/Support/Debug.h" 38#include "llvm/Support/GetElementPtrTypeIterator.h" 39#include "llvm/Support/PatternMatch.h" 40using namespace llvm; 41using namespace llvm::PatternMatch; 42 43static cl::opt<bool> FactorCommonPreds("split-critical-paths-tweak", 44 cl::init(false), cl::Hidden); 45 46namespace { 47 class VISIBILITY_HIDDEN CodeGenPrepare : public FunctionPass { 48 /// TLI - Keep a pointer of a TargetLowering to consult for determining 49 /// transformation profitability. 50 const TargetLowering *TLI; 51 52 /// BackEdges - Keep a set of all the loop back edges. 53 /// 54 SmallSet<std::pair<const BasicBlock*, const BasicBlock*>, 8> BackEdges; 55 public: 56 static char ID; // Pass identification, replacement for typeid 57 explicit CodeGenPrepare(const TargetLowering *tli = 0) 58 : FunctionPass(&ID), TLI(tli) {} 59 bool runOnFunction(Function &F); 60 61 private: 62 bool EliminateMostlyEmptyBlocks(Function &F); 63 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; 64 void EliminateMostlyEmptyBlock(BasicBlock *BB); 65 bool OptimizeBlock(BasicBlock &BB); 66 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy, 67 DenseMap<Value*,Value*> &SunkAddrs); 68 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS, 69 DenseMap<Value*,Value*> &SunkAddrs); 70 bool OptimizeExtUses(Instruction *I); 71 void findLoopBackEdges(const Function &F); 72 }; 73} 74 75char CodeGenPrepare::ID = 0; 76static RegisterPass<CodeGenPrepare> X("codegenprepare", 77 "Optimize for code generation"); 78 79FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) { 80 return new CodeGenPrepare(TLI); 81} 82 83/// findLoopBackEdges - Do a DFS walk to find loop back edges. 84/// 85void CodeGenPrepare::findLoopBackEdges(const Function &F) { 86 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 87 FindFunctionBackedges(F, Edges); 88 89 BackEdges.insert(Edges.begin(), Edges.end()); 90} 91 92 93bool CodeGenPrepare::runOnFunction(Function &F) { 94 bool EverMadeChange = false; 95 96 // First pass, eliminate blocks that contain only PHI nodes and an 97 // unconditional branch. 98 EverMadeChange |= EliminateMostlyEmptyBlocks(F); 99 100 // Now find loop back edges. 101 findLoopBackEdges(F); 102 103 bool MadeChange = true; 104 while (MadeChange) { 105 MadeChange = false; 106 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 107 MadeChange |= OptimizeBlock(*BB); 108 EverMadeChange |= MadeChange; 109 } 110 return EverMadeChange; 111} 112 113/// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes, 114/// debug info directives, and an unconditional branch. Passes before isel 115/// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for 116/// isel. Start by eliminating these blocks so we can split them the way we 117/// want them. 118bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) { 119 bool MadeChange = false; 120 // Note that this intentionally skips the entry block. 121 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) { 122 BasicBlock *BB = I++; 123 124 // If this block doesn't end with an uncond branch, ignore it. 125 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); 126 if (!BI || !BI->isUnconditional()) 127 continue; 128 129 // If the instruction before the branch (skipping debug info) isn't a phi 130 // node, then other stuff is happening here. 131 BasicBlock::iterator BBI = BI; 132 if (BBI != BB->begin()) { 133 --BBI; 134 while (isa<DbgInfoIntrinsic>(BBI)) { 135 if (BBI == BB->begin()) 136 break; 137 --BBI; 138 } 139 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI)) 140 continue; 141 } 142 143 // Do not break infinite loops. 144 BasicBlock *DestBB = BI->getSuccessor(0); 145 if (DestBB == BB) 146 continue; 147 148 if (!CanMergeBlocks(BB, DestBB)) 149 continue; 150 151 EliminateMostlyEmptyBlock(BB); 152 MadeChange = true; 153 } 154 return MadeChange; 155} 156 157/// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a 158/// single uncond branch between them, and BB contains no other non-phi 159/// instructions. 160bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB, 161 const BasicBlock *DestBB) const { 162 // We only want to eliminate blocks whose phi nodes are used by phi nodes in 163 // the successor. If there are more complex condition (e.g. preheaders), 164 // don't mess around with them. 165 BasicBlock::const_iterator BBI = BB->begin(); 166 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 167 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end(); 168 UI != E; ++UI) { 169 const Instruction *User = cast<Instruction>(*UI); 170 if (User->getParent() != DestBB || !isa<PHINode>(User)) 171 return false; 172 // If User is inside DestBB block and it is a PHINode then check 173 // incoming value. If incoming value is not from BB then this is 174 // a complex condition (e.g. preheaders) we want to avoid here. 175 if (User->getParent() == DestBB) { 176 if (const PHINode *UPN = dyn_cast<PHINode>(User)) 177 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { 178 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); 179 if (Insn && Insn->getParent() == BB && 180 Insn->getParent() != UPN->getIncomingBlock(I)) 181 return false; 182 } 183 } 184 } 185 } 186 187 // If BB and DestBB contain any common predecessors, then the phi nodes in BB 188 // and DestBB may have conflicting incoming values for the block. If so, we 189 // can't merge the block. 190 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); 191 if (!DestBBPN) return true; // no conflict. 192 193 // Collect the preds of BB. 194 SmallPtrSet<const BasicBlock*, 16> BBPreds; 195 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 196 // It is faster to get preds from a PHI than with pred_iterator. 197 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 198 BBPreds.insert(BBPN->getIncomingBlock(i)); 199 } else { 200 BBPreds.insert(pred_begin(BB), pred_end(BB)); 201 } 202 203 // Walk the preds of DestBB. 204 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { 205 BasicBlock *Pred = DestBBPN->getIncomingBlock(i); 206 if (BBPreds.count(Pred)) { // Common predecessor? 207 BBI = DestBB->begin(); 208 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) { 209 const Value *V1 = PN->getIncomingValueForBlock(Pred); 210 const Value *V2 = PN->getIncomingValueForBlock(BB); 211 212 // If V2 is a phi node in BB, look up what the mapped value will be. 213 if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) 214 if (V2PN->getParent() == BB) 215 V2 = V2PN->getIncomingValueForBlock(Pred); 216 217 // If there is a conflict, bail out. 218 if (V1 != V2) return false; 219 } 220 } 221 } 222 223 return true; 224} 225 226 227/// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and 228/// an unconditional branch in it. 229void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) { 230 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 231 BasicBlock *DestBB = BI->getSuccessor(0); 232 233 DOUT << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB; 234 235 // If the destination block has a single pred, then this is a trivial edge, 236 // just collapse it. 237 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) { 238 if (SinglePred != DestBB) { 239 // Remember if SinglePred was the entry block of the function. If so, we 240 // will need to move BB back to the entry position. 241 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 242 MergeBasicBlockIntoOnlyPred(DestBB); 243 244 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 245 BB->moveBefore(&BB->getParent()->getEntryBlock()); 246 247 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 248 return; 249 } 250 } 251 252 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB 253 // to handle the new incoming edges it is about to have. 254 PHINode *PN; 255 for (BasicBlock::iterator BBI = DestBB->begin(); 256 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 257 // Remove the incoming value for BB, and remember it. 258 Value *InVal = PN->removeIncomingValue(BB, false); 259 260 // Two options: either the InVal is a phi node defined in BB or it is some 261 // value that dominates BB. 262 PHINode *InValPhi = dyn_cast<PHINode>(InVal); 263 if (InValPhi && InValPhi->getParent() == BB) { 264 // Add all of the input values of the input PHI as inputs of this phi. 265 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) 266 PN->addIncoming(InValPhi->getIncomingValue(i), 267 InValPhi->getIncomingBlock(i)); 268 } else { 269 // Otherwise, add one instance of the dominating value for each edge that 270 // we will be adding. 271 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { 272 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) 273 PN->addIncoming(InVal, BBPN->getIncomingBlock(i)); 274 } else { 275 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 276 PN->addIncoming(InVal, *PI); 277 } 278 } 279 } 280 281 // The PHIs are now updated, change everything that refers to BB to use 282 // DestBB and remove BB. 283 BB->replaceAllUsesWith(DestBB); 284 BB->eraseFromParent(); 285 286 DOUT << "AFTER:\n" << *DestBB << "\n\n\n"; 287} 288 289 290/// SplitEdgeNicely - Split the critical edge from TI to its specified 291/// successor if it will improve codegen. We only do this if the successor has 292/// phi nodes (otherwise critical edges are ok). If there is already another 293/// predecessor of the succ that is empty (and thus has no phi nodes), use it 294/// instead of introducing a new block. 295static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, 296 SmallSet<std::pair<const BasicBlock*, 297 const BasicBlock*>, 8> &BackEdges, 298 Pass *P) { 299 BasicBlock *TIBB = TI->getParent(); 300 BasicBlock *Dest = TI->getSuccessor(SuccNum); 301 assert(isa<PHINode>(Dest->begin()) && 302 "This should only be called if Dest has a PHI!"); 303 304 // Do not split edges to EH landing pads. 305 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI)) { 306 if (Invoke->getSuccessor(1) == Dest) 307 return; 308 } 309 310 // As a hack, never split backedges of loops. Even though the copy for any 311 // PHIs inserted on the backedge would be dead for exits from the loop, we 312 // assume that the cost of *splitting* the backedge would be too high. 313 if (BackEdges.count(std::make_pair(TIBB, Dest))) 314 return; 315 316 if (!FactorCommonPreds) { 317 /// TIPHIValues - This array is lazily computed to determine the values of 318 /// PHIs in Dest that TI would provide. 319 SmallVector<Value*, 32> TIPHIValues; 320 321 // Check to see if Dest has any blocks that can be used as a split edge for 322 // this terminator. 323 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { 324 BasicBlock *Pred = *PI; 325 // To be usable, the pred has to end with an uncond branch to the dest. 326 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); 327 if (!PredBr || !PredBr->isUnconditional()) 328 continue; 329 // Must be empty other than the branch and debug info. 330 BasicBlock::iterator I = Pred->begin(); 331 while (isa<DbgInfoIntrinsic>(I)) 332 I++; 333 if (dyn_cast<Instruction>(I) != PredBr) 334 continue; 335 // Cannot be the entry block; its label does not get emitted. 336 if (Pred == &(Dest->getParent()->getEntryBlock())) 337 continue; 338 339 // Finally, since we know that Dest has phi nodes in it, we have to make 340 // sure that jumping to Pred will have the same effect as going to Dest in 341 // terms of PHI values. 342 PHINode *PN; 343 unsigned PHINo = 0; 344 bool FoundMatch = true; 345 for (BasicBlock::iterator I = Dest->begin(); 346 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { 347 if (PHINo == TIPHIValues.size()) 348 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); 349 350 // If the PHI entry doesn't work, we can't use this pred. 351 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { 352 FoundMatch = false; 353 break; 354 } 355 } 356 357 // If we found a workable predecessor, change TI to branch to Succ. 358 if (FoundMatch) { 359 Dest->removePredecessor(TIBB); 360 TI->setSuccessor(SuccNum, Pred); 361 return; 362 } 363 } 364 365 SplitCriticalEdge(TI, SuccNum, P, true); 366 return; 367 } 368 369 PHINode *PN; 370 SmallVector<Value*, 8> TIPHIValues; 371 for (BasicBlock::iterator I = Dest->begin(); 372 (PN = dyn_cast<PHINode>(I)); ++I) 373 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); 374 375 SmallVector<BasicBlock*, 8> IdenticalPreds; 376 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { 377 BasicBlock *Pred = *PI; 378 if (BackEdges.count(std::make_pair(Pred, Dest))) 379 continue; 380 if (PI == TIBB) 381 IdenticalPreds.push_back(Pred); 382 else { 383 bool Identical = true; 384 unsigned PHINo = 0; 385 for (BasicBlock::iterator I = Dest->begin(); 386 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) 387 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { 388 Identical = false; 389 break; 390 } 391 if (Identical) 392 IdenticalPreds.push_back(Pred); 393 } 394 } 395 396 assert(!IdenticalPreds.empty()); 397 SplitBlockPredecessors(Dest, &IdenticalPreds[0], IdenticalPreds.size(), 398 ".critedge", P); 399} 400 401 402/// OptimizeNoopCopyExpression - If the specified cast instruction is a noop 403/// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC), 404/// sink it into user blocks to reduce the number of virtual 405/// registers that must be created and coalesced. 406/// 407/// Return true if any changes are made. 408/// 409static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){ 410 // If this is a noop copy, 411 MVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 412 MVT DstVT = TLI.getValueType(CI->getType()); 413 414 // This is an fp<->int conversion? 415 if (SrcVT.isInteger() != DstVT.isInteger()) 416 return false; 417 418 // If this is an extension, it will be a zero or sign extension, which 419 // isn't a noop. 420 if (SrcVT.bitsLT(DstVT)) return false; 421 422 // If these values will be promoted, find out what they will be promoted 423 // to. This helps us consider truncates on PPC as noop copies when they 424 // are. 425 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote) 426 SrcVT = TLI.getTypeToTransformTo(SrcVT); 427 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote) 428 DstVT = TLI.getTypeToTransformTo(DstVT); 429 430 // If, after promotion, these are the same types, this is a noop copy. 431 if (SrcVT != DstVT) 432 return false; 433 434 BasicBlock *DefBB = CI->getParent(); 435 436 /// InsertedCasts - Only insert a cast in each block once. 437 DenseMap<BasicBlock*, CastInst*> InsertedCasts; 438 439 bool MadeChange = false; 440 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 441 UI != E; ) { 442 Use &TheUse = UI.getUse(); 443 Instruction *User = cast<Instruction>(*UI); 444 445 // Figure out which BB this cast is used in. For PHI's this is the 446 // appropriate predecessor block. 447 BasicBlock *UserBB = User->getParent(); 448 if (PHINode *PN = dyn_cast<PHINode>(User)) { 449 UserBB = PN->getIncomingBlock(UI); 450 } 451 452 // Preincrement use iterator so we don't invalidate it. 453 ++UI; 454 455 // If this user is in the same block as the cast, don't change the cast. 456 if (UserBB == DefBB) continue; 457 458 // If we have already inserted a cast into this block, use it. 459 CastInst *&InsertedCast = InsertedCasts[UserBB]; 460 461 if (!InsertedCast) { 462 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 463 464 InsertedCast = 465 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 466 InsertPt); 467 MadeChange = true; 468 } 469 470 // Replace a use of the cast with a use of the new cast. 471 TheUse = InsertedCast; 472 } 473 474 // If we removed all uses, nuke the cast. 475 if (CI->use_empty()) { 476 CI->eraseFromParent(); 477 MadeChange = true; 478 } 479 480 return MadeChange; 481} 482 483/// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce 484/// the number of virtual registers that must be created and coalesced. This is 485/// a clear win except on targets with multiple condition code registers 486/// (PowerPC), where it might lose; some adjustment may be wanted there. 487/// 488/// Return true if any changes are made. 489static bool OptimizeCmpExpression(CmpInst *CI) { 490 BasicBlock *DefBB = CI->getParent(); 491 492 /// InsertedCmp - Only insert a cmp in each block once. 493 DenseMap<BasicBlock*, CmpInst*> InsertedCmps; 494 495 bool MadeChange = false; 496 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 497 UI != E; ) { 498 Use &TheUse = UI.getUse(); 499 Instruction *User = cast<Instruction>(*UI); 500 501 // Preincrement use iterator so we don't invalidate it. 502 ++UI; 503 504 // Don't bother for PHI nodes. 505 if (isa<PHINode>(User)) 506 continue; 507 508 // Figure out which BB this cmp is used in. 509 BasicBlock *UserBB = User->getParent(); 510 511 // If this user is in the same block as the cmp, don't change the cmp. 512 if (UserBB == DefBB) continue; 513 514 // If we have already inserted a cmp into this block, use it. 515 CmpInst *&InsertedCmp = InsertedCmps[UserBB]; 516 517 if (!InsertedCmp) { 518 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 519 520 InsertedCmp = 521 CmpInst::Create(DefBB->getContext(), CI->getOpcode(), 522 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 LLVMContext &Context = MemoryInst->getContext(); 563 564 // Figure out what addressing mode will be built up for this operation. 565 SmallVector<Instruction*, 16> AddrModeInsts; 566 ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst, 567 AddrModeInsts, *TLI); 568 569 // Check to see if any of the instructions supersumed by this addr mode are 570 // non-local to I's BB. 571 bool AnyNonLocal = false; 572 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) { 573 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) { 574 AnyNonLocal = true; 575 break; 576 } 577 } 578 579 // If all the instructions matched are already in this BB, don't do anything. 580 if (!AnyNonLocal) { 581 DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n"); 582 return false; 583 } 584 585 // Insert this computation right after this user. Since our caller is 586 // scanning from the top of the BB to the bottom, reuse of the expr are 587 // guaranteed to happen later. 588 BasicBlock::iterator InsertPt = MemoryInst; 589 590 // Now that we determined the addressing expression we want to use and know 591 // that we have to sink it into this block. Check to see if we have already 592 // done this for some other load/store instr in this block. If so, reuse the 593 // computation. 594 Value *&SunkAddr = SunkAddrs[Addr]; 595 if (SunkAddr) { 596 DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for " 597 << *MemoryInst); 598 if (SunkAddr->getType() != Addr->getType()) 599 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt); 600 } else { 601 DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for " 602 << *MemoryInst); 603 const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType(); 604 605 Value *Result = 0; 606 // Start with the scale value. 607 if (AddrMode.Scale) { 608 Value *V = AddrMode.ScaledReg; 609 if (V->getType() == IntPtrTy) { 610 // done. 611 } else if (isa<PointerType>(V->getType())) { 612 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 613 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < 614 cast<IntegerType>(V->getType())->getBitWidth()) { 615 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt); 616 } else { 617 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt); 618 } 619 if (AddrMode.Scale != 1) 620 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy, 621 AddrMode.Scale), 622 "sunkaddr", InsertPt); 623 Result = V; 624 } 625 626 // Add in the base register. 627 if (AddrMode.BaseReg) { 628 Value *V = AddrMode.BaseReg; 629 if (isa<PointerType>(V->getType())) 630 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt); 631 if (V->getType() != IntPtrTy) 632 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true, 633 "sunkaddr", InsertPt); 634 if (Result) 635 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 636 else 637 Result = V; 638 } 639 640 // Add in the BaseGV if present. 641 if (AddrMode.BaseGV) { 642 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr", 643 InsertPt); 644 if (Result) 645 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 646 else 647 Result = V; 648 } 649 650 // Add in the Base Offset if present. 651 if (AddrMode.BaseOffs) { 652 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); 653 if (Result) 654 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt); 655 else 656 Result = V; 657 } 658 659 if (Result == 0) 660 SunkAddr = Context.getNullValue(Addr->getType()); 661 else 662 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt); 663 } 664 665 MemoryInst->replaceUsesOfWith(Addr, SunkAddr); 666 667 if (Addr->use_empty()) 668 RecursivelyDeleteTriviallyDeadInstructions(Addr); 669 return true; 670} 671 672/// OptimizeInlineAsmInst - If there are any memory operands, use 673/// OptimizeMemoryInst to sink their address computing into the block when 674/// possible / profitable. 675bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS, 676 DenseMap<Value*,Value*> &SunkAddrs) { 677 bool MadeChange = false; 678 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 679 680 // Do a prepass over the constraints, canonicalizing them, and building up the 681 // ConstraintOperands list. 682 std::vector<InlineAsm::ConstraintInfo> 683 ConstraintInfos = IA->ParseConstraints(); 684 685 /// ConstraintOperands - Information about all of the constraints. 686 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands; 687 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 688 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 689 ConstraintOperands. 690 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i])); 691 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back(); 692 693 // Compute the value type for each operand. 694 switch (OpInfo.Type) { 695 case InlineAsm::isOutput: 696 if (OpInfo.isIndirect) 697 OpInfo.CallOperandVal = CS.getArgument(ArgNo++); 698 break; 699 case InlineAsm::isInput: 700 OpInfo.CallOperandVal = CS.getArgument(ArgNo++); 701 break; 702 case InlineAsm::isClobber: 703 // Nothing to do. 704 break; 705 } 706 707 // Compute the constraint code and ConstraintType to use. 708 TLI->ComputeConstraintToUse(OpInfo, SDValue(), 709 OpInfo.ConstraintType == TargetLowering::C_Memory); 710 711 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 712 OpInfo.isIndirect) { 713 Value *OpVal = OpInfo.CallOperandVal; 714 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs); 715 } 716 } 717 718 return MadeChange; 719} 720 721bool CodeGenPrepare::OptimizeExtUses(Instruction *I) { 722 BasicBlock *DefBB = I->getParent(); 723 724 // If both result of the {s|z}xt and its source are live out, rewrite all 725 // other uses of the source with result of extension. 726 Value *Src = I->getOperand(0); 727 if (Src->hasOneUse()) 728 return false; 729 730 // Only do this xform if truncating is free. 731 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType())) 732 return false; 733 734 // Only safe to perform the optimization if the source is also defined in 735 // this block. 736 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) 737 return false; 738 739 bool DefIsLiveOut = false; 740 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 741 UI != E; ++UI) { 742 Instruction *User = cast<Instruction>(*UI); 743 744 // Figure out which BB this ext is used in. 745 BasicBlock *UserBB = User->getParent(); 746 if (UserBB == DefBB) continue; 747 DefIsLiveOut = true; 748 break; 749 } 750 if (!DefIsLiveOut) 751 return false; 752 753 // Make sure non of the uses are PHI nodes. 754 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 755 UI != E; ++UI) { 756 Instruction *User = cast<Instruction>(*UI); 757 BasicBlock *UserBB = User->getParent(); 758 if (UserBB == DefBB) continue; 759 // Be conservative. We don't want this xform to end up introducing 760 // reloads just before load / store instructions. 761 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User)) 762 return false; 763 } 764 765 // InsertedTruncs - Only insert one trunc in each block once. 766 DenseMap<BasicBlock*, Instruction*> InsertedTruncs; 767 768 bool MadeChange = false; 769 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end(); 770 UI != E; ++UI) { 771 Use &TheUse = UI.getUse(); 772 Instruction *User = cast<Instruction>(*UI); 773 774 // Figure out which BB this ext is used in. 775 BasicBlock *UserBB = User->getParent(); 776 if (UserBB == DefBB) continue; 777 778 // Both src and def are live in this block. Rewrite the use. 779 Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; 780 781 if (!InsertedTrunc) { 782 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI(); 783 784 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt); 785 } 786 787 // Replace a use of the {s|z}ext source with a use of the result. 788 TheUse = InsertedTrunc; 789 790 MadeChange = true; 791 } 792 793 return MadeChange; 794} 795 796// In this pass we look for GEP and cast instructions that are used 797// across basic blocks and rewrite them to improve basic-block-at-a-time 798// selection. 799bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) { 800 bool MadeChange = false; 801 802 // Split all critical edges where the dest block has a PHI. 803 TerminatorInst *BBTI = BB.getTerminator(); 804 if (BBTI->getNumSuccessors() > 1) { 805 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) { 806 BasicBlock *SuccBB = BBTI->getSuccessor(i); 807 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true)) 808 SplitEdgeNicely(BBTI, i, BackEdges, this); 809 } 810 } 811 812 // Keep track of non-local addresses that have been sunk into this block. 813 // This allows us to avoid inserting duplicate code for blocks with multiple 814 // load/stores of the same address. 815 DenseMap<Value*, Value*> SunkAddrs; 816 817 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) { 818 Instruction *I = BBI++; 819 820 if (CastInst *CI = dyn_cast<CastInst>(I)) { 821 // If the source of the cast is a constant, then this should have 822 // already been constant folded. The only reason NOT to constant fold 823 // it is if something (e.g. LSR) was careful to place the constant 824 // evaluation in a block other than then one that uses it (e.g. to hoist 825 // the address of globals out of a loop). If this is the case, we don't 826 // want to forward-subst the cast. 827 if (isa<Constant>(CI->getOperand(0))) 828 continue; 829 830 bool Change = false; 831 if (TLI) { 832 Change = OptimizeNoopCopyExpression(CI, *TLI); 833 MadeChange |= Change; 834 } 835 836 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) 837 MadeChange |= OptimizeExtUses(I); 838 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) { 839 MadeChange |= OptimizeCmpExpression(CI); 840 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 841 if (TLI) 842 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(), 843 SunkAddrs); 844 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 845 if (TLI) 846 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1), 847 SI->getOperand(0)->getType(), 848 SunkAddrs); 849 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 850 if (GEPI->hasAllZeroIndices()) { 851 /// The GEP operand must be a pointer, so must its result -> BitCast 852 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 853 GEPI->getName(), GEPI); 854 GEPI->replaceAllUsesWith(NC); 855 GEPI->eraseFromParent(); 856 MadeChange = true; 857 BBI = NC; 858 } 859 } else if (CallInst *CI = dyn_cast<CallInst>(I)) { 860 // If we found an inline asm expession, and if the target knows how to 861 // lower it to normal LLVM code, do so now. 862 if (TLI && isa<InlineAsm>(CI->getCalledValue())) { 863 if (TLI->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