lib/Transforms/Scalar/CodeGenPrepare.cpp

   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/LLVMContext.h"
  25 #include "llvm/Pass.h"
  26 #include "llvm/Analysis/ProfileInfo.h"
  27 #include "llvm/Target/TargetData.h"
  28 #include "llvm/Target/TargetLowering.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/Debug.h"
  38 #include "llvm/Support/GetElementPtrTypeIterator.h"
  39 #include "llvm/Support/PatternMatch.h"
  40 #include "llvm/Support/raw_ostream.h"
  41 using namespace llvm;
  42 using namespace llvm::PatternMatch;
  43
  44 static cl::opt<bool> FactorCommonPreds("split-critical-paths-tweak",
  45                                        cl::init(false), cl::Hidden);
  46
  47 namespace {
  48   class CodeGenPrepare : public FunctionPass {
  49     /// TLI - Keep a pointer of a TargetLowering to consult for determining
  50     /// transformation profitability.
  51     const TargetLowering *TLI;
  52     ProfileInfo *PI;
  53
  54     /// BackEdges - Keep a set of all the loop back edges.
  55     ///
  56     SmallSet<std::pair<const BasicBlock*, const BasicBlock*>, 8> BackEdges;
  57   public:
  58     static char ID; // Pass identification, replacement for typeid
  59     explicit CodeGenPrepare(const TargetLowering *tli = 0)
  60       : FunctionPass(&ID), TLI(tli) {}
  61     bool runOnFunction(Function &F);
  62
  63     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
  64       AU.addPreserved<ProfileInfo>();
  65     }
  66
  67   private:
  68     bool EliminateMostlyEmptyBlocks(Function &F);
  69     bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
  70     void EliminateMostlyEmptyBlock(BasicBlock *BB);
  71     bool OptimizeBlock(BasicBlock &BB);
  72     bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy,
  73                             DenseMap<Value*,Value*> &SunkAddrs);
  74     bool OptimizeInlineAsmInst(Instruction *I, CallSite CS,
  75                                DenseMap<Value*,Value*> &SunkAddrs);
  76     bool OptimizeExtUses(Instruction *I);
  77     void findLoopBackEdges(const Function &F);
  78   };
  79 }
  80
  81 char CodeGenPrepare::ID = 0;
  82 static RegisterPass<CodeGenPrepare> X("codegenprepare",
  83                                       "Optimize for code generation");
  84
  85 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
  86   return new CodeGenPrepare(TLI);
  87 }
  88
  89 /// findLoopBackEdges - Do a DFS walk to find loop back edges.
  90 ///
  91 void CodeGenPrepare::findLoopBackEdges(const Function &F) {
  92   SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
  93   FindFunctionBackedges(F, Edges);
  94
  95   BackEdges.insert(Edges.begin(), Edges.end());
  96 }
  97
  98
  99 bool CodeGenPrepare::runOnFunction(Function &F) {
 100   bool EverMadeChange = false;
 101
 102   PI = getAnalysisIfAvailable<ProfileInfo>();
 103   // First pass, eliminate blocks that contain only PHI nodes and an
 104   // unconditional branch.
 105   EverMadeChange |= EliminateMostlyEmptyBlocks(F);
 106
 107   // Now find loop back edges.
 108   findLoopBackEdges(F);
 109
 110   bool MadeChange = true;
 111   while (MadeChange) {
 112     MadeChange = false;
 113     for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
 114       MadeChange |= OptimizeBlock(*BB);
 115     EverMadeChange |= MadeChange;
 116   }
 117   return EverMadeChange;
 118 }
 119
 120 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
 121 /// debug info directives, and an unconditional branch.  Passes before isel
 122 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
 123 /// isel.  Start by eliminating these blocks so we can split them the way we
 124 /// want them.
 125 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
 126   bool MadeChange = false;
 127   // Note that this intentionally skips the entry block.
 128   for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
 129     BasicBlock *BB = I++;
 130
 131     // If this block doesn't end with an uncond branch, ignore it.
 132     BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
 133     if (!BI || !BI->isUnconditional())
 134       continue;
 135
 136     // If the instruction before the branch (skipping debug info) isn't a phi
 137     // node, then other stuff is happening here.
 138     BasicBlock::iterator BBI = BI;
 139     if (BBI != BB->begin()) {
 140       --BBI;
 141       while (isa<DbgInfoIntrinsic>(BBI)) {
 142         if (BBI == BB->begin())
 143           break;
 144         --BBI;
 145       }
 146       if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
 147         continue;
 148     }
 149
 150     // Do not break infinite loops.
 151     BasicBlock *DestBB = BI->getSuccessor(0);
 152     if (DestBB == BB)
 153       continue;
 154
 155     if (!CanMergeBlocks(BB, DestBB))
 156       continue;
 157
 158     EliminateMostlyEmptyBlock(BB);
 159     MadeChange = true;
 160   }
 161   return MadeChange;
 162 }
 163
 164 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
 165 /// single uncond branch between them, and BB contains no other non-phi
 166 /// instructions.
 167 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
 168                                     const BasicBlock *DestBB) const {
 169   // We only want to eliminate blocks whose phi nodes are used by phi nodes in
 170   // the successor.  If there are more complex condition (e.g. preheaders),
 171   // don't mess around with them.
 172   BasicBlock::const_iterator BBI = BB->begin();
 173   while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
 174     for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end();
 175          UI != E; ++UI) {
 176       const Instruction *User = cast<Instruction>(*UI);
 177       if (User->getParent() != DestBB || !isa<PHINode>(User))
 178         return false;
 179       // If User is inside DestBB block and it is a PHINode then check
 180       // incoming value. If incoming value is not from BB then this is
 181       // a complex condition (e.g. preheaders) we want to avoid here.
 182       if (User->getParent() == DestBB) {
 183         if (const PHINode *UPN = dyn_cast<PHINode>(User))
 184           for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
 185             Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
 186             if (Insn && Insn->getParent() == BB &&
 187                 Insn->getParent() != UPN->getIncomingBlock(I))
 188               return false;
 189           }
 190       }
 191     }
 192   }
 193
 194   // If BB and DestBB contain any common predecessors, then the phi nodes in BB
 195   // and DestBB may have conflicting incoming values for the block.  If so, we
 196   // can't merge the block.
 197   const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
 198   if (!DestBBPN) return true;  // no conflict.
 199
 200   // Collect the preds of BB.
 201   SmallPtrSet<const BasicBlock*, 16> BBPreds;
 202   if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
 203     // It is faster to get preds from a PHI than with pred_iterator.
 204     for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
 205       BBPreds.insert(BBPN->getIncomingBlock(i));
 206   } else {
 207     BBPreds.insert(pred_begin(BB), pred_end(BB));
 208   }
 209
 210   // Walk the preds of DestBB.
 211   for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
 212     BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
 213     if (BBPreds.count(Pred)) {   // Common predecessor?
 214       BBI = DestBB->begin();
 215       while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
 216         const Value *V1 = PN->getIncomingValueForBlock(Pred);
 217         const Value *V2 = PN->getIncomingValueForBlock(BB);
 218
 219         // If V2 is a phi node in BB, look up what the mapped value will be.
 220         if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
 221           if (V2PN->getParent() == BB)
 222             V2 = V2PN->getIncomingValueForBlock(Pred);
 223
 224         // If there is a conflict, bail out.
 225         if (V1 != V2) return false;
 226       }
 227     }
 228   }
 229
 230   return true;
 231 }
 232
 233
 234 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
 235 /// an unconditional branch in it.
 236 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
 237   BranchInst *BI = cast<BranchInst>(BB->getTerminator());
 238   BasicBlock *DestBB = BI->getSuccessor(0);
 239
 240   DEBUG(errs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
 241
 242   // If the destination block has a single pred, then this is a trivial edge,
 243   // just collapse it.
 244   if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
 245     if (SinglePred != DestBB) {
 246       // Remember if SinglePred was the entry block of the function.  If so, we
 247       // will need to move BB back to the entry position.
 248       bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
 249       MergeBasicBlockIntoOnlyPred(DestBB, this);
 250
 251       if (isEntry && BB != &BB->getParent()->getEntryBlock())
 252         BB->moveBefore(&BB->getParent()->getEntryBlock());
 253
 254       DEBUG(errs() << "AFTER:\n" << *DestBB << "\n\n\n");
 255       return;
 256     }
 257   }
 258
 259   // Otherwise, we have multiple predecessors of BB.  Update the PHIs in DestBB
 260   // to handle the new incoming edges it is about to have.
 261   PHINode *PN;
 262   for (BasicBlock::iterator BBI = DestBB->begin();
 263        (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
 264     // Remove the incoming value for BB, and remember it.
 265     Value *InVal = PN->removeIncomingValue(BB, false);
 266
 267     // Two options: either the InVal is a phi node defined in BB or it is some
 268     // value that dominates BB.
 269     PHINode *InValPhi = dyn_cast<PHINode>(InVal);
 270     if (InValPhi && InValPhi->getParent() == BB) {
 271       // Add all of the input values of the input PHI as inputs of this phi.
 272       for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
 273         PN->addIncoming(InValPhi->getIncomingValue(i),
 274                         InValPhi->getIncomingBlock(i));
 275     } else {
 276       // Otherwise, add one instance of the dominating value for each edge that
 277       // we will be adding.
 278       if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
 279         for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
 280           PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
 281       } else {
 282         for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
 283           PN->addIncoming(InVal, *PI);
 284       }
 285     }
 286   }
 287
 288   // The PHIs are now updated, change everything that refers to BB to use
 289   // DestBB and remove BB.
 290   BB->replaceAllUsesWith(DestBB);
 291   if (PI) {
 292     PI->replaceAllUses(BB, DestBB);
 293     PI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
 294   }
 295   BB->eraseFromParent();
 296
 297   DEBUG(errs() << "AFTER:\n" << *DestBB << "\n\n\n");
 298 }
 299
 300
 301 /// SplitEdgeNicely - Split the critical edge from TI to its specified
 302 /// successor if it will improve codegen.  We only do this if the successor has
 303 /// phi nodes (otherwise critical edges are ok).  If there is already another
 304 /// predecessor of the succ that is empty (and thus has no phi nodes), use it
 305 /// instead of introducing a new block.
 306 static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum,
 307                      SmallSet<std::pair<const BasicBlock*,
 308                                         const BasicBlock*>, 8> &BackEdges,
 309                              Pass *P) {
 310   BasicBlock *TIBB = TI->getParent();
 311   BasicBlock *Dest = TI->getSuccessor(SuccNum);
 312   assert(isa<PHINode>(Dest->begin()) &&
 313          "This should only be called if Dest has a PHI!");
 314
 315   // Do not split edges to EH landing pads.
 316   if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI)) {
 317     if (Invoke->getSuccessor(1) == Dest)
 318       return;
 319   }
 320
 321   // As a hack, never split backedges of loops.  Even though the copy for any
 322   // PHIs inserted on the backedge would be dead for exits from the loop, we
 323   // assume that the cost of *splitting* the backedge would be too high.
 324   if (BackEdges.count(std::make_pair(TIBB, Dest)))
 325     return;
 326
 327   if (!FactorCommonPreds) {
 328     /// TIPHIValues - This array is lazily computed to determine the values of
 329     /// PHIs in Dest that TI would provide.
 330     SmallVector<Value*, 32> TIPHIValues;
 331
 332     // Check to see if Dest has any blocks that can be used as a split edge for
 333     // this terminator.
 334     for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) {
 335       BasicBlock *Pred = *PI;
 336       // To be usable, the pred has to end with an uncond branch to the dest.
 337       BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator());
 338       if (!PredBr || !PredBr->isUnconditional())
 339         continue;
 340       // Must be empty other than the branch and debug info.
 341       BasicBlock::iterator I = Pred->begin();
 342       while (isa<DbgInfoIntrinsic>(I))
 343         I++;
 344       if (dyn_cast<Instruction>(I) != PredBr)
 345         continue;
 346       // Cannot be the entry block; its label does not get emitted.
 347       if (Pred == &(Dest->getParent()->getEntryBlock()))
 348         continue;
 349
 350       // Finally, since we know that Dest has phi nodes in it, we have to make
 351       // sure that jumping to Pred will have the same effect as going to Dest in
 352       // terms of PHI values.
 353       PHINode *PN;
 354       unsigned PHINo = 0;
 355       bool FoundMatch = true;
 356       for (BasicBlock::iterator I = Dest->begin();
 357            (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) {
 358         if (PHINo == TIPHIValues.size())
 359           TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB));
 360
 361         // If the PHI entry doesn't work, we can't use this pred.
 362         if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) {
 363           FoundMatch = false;
 364           break;
 365         }
 366       }
 367
 368       // If we found a workable predecessor, change TI to branch to Succ.
 369       if (FoundMatch) {
 370         ProfileInfo *PI = P->getAnalysisIfAvailable<ProfileInfo>();
 371         if (PI)
 372           PI->splitEdge(TIBB, Dest, Pred);
 373         Dest->removePredecessor(TIBB);
 374         TI->setSuccessor(SuccNum, Pred);
 375         return;
 376       }
 377     }
 378
 379     SplitCriticalEdge(TI, SuccNum, P, true);
 380     return;
 381   }
 382
 383   PHINode *PN;
 384   SmallVector<Value*, 8> TIPHIValues;
 385   for (BasicBlock::iterator I = Dest->begin();
 386        (PN = dyn_cast<PHINode>(I)); ++I)
 387     TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB));
 388
 389   SmallVector<BasicBlock*, 8> IdenticalPreds;
 390   for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) {
 391     BasicBlock *Pred = *PI;
 392     if (BackEdges.count(std::make_pair(Pred, Dest)))
 393       continue;
 394     if (PI == TIBB)
 395       IdenticalPreds.push_back(Pred);
 396     else {
 397       bool Identical = true;
 398       unsigned PHINo = 0;
 399       for (BasicBlock::iterator I = Dest->begin();
 400            (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo)
 401         if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) {
 402           Identical = false;
 403           break;
 404         }
 405       if (Identical)
 406         IdenticalPreds.push_back(Pred);
 407     }
 408   }
 409
 410   assert(!IdenticalPreds.empty());
 411   SplitBlockPredecessors(Dest, &IdenticalPreds[0], IdenticalPreds.size(),
 412                          ".critedge", P);
 413 }
 414
 415
 416 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
 417 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
 418 /// sink it into user blocks to reduce the number of virtual
 419 /// registers that must be created and coalesced.
 420 ///
 421 /// Return true if any changes are made.
 422 ///
 423 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
 424   // If this is a noop copy,
 425   EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
 426   EVT DstVT = TLI.getValueType(CI->getType());
 427
 428   // This is an fp<->int conversion?
 429   if (SrcVT.isInteger() != DstVT.isInteger())
 430     return false;
 431
 432   // If this is an extension, it will be a zero or sign extension, which
 433   // isn't a noop.
 434   if (SrcVT.bitsLT(DstVT)) return false;
 435
 436   // If these values will be promoted, find out what they will be promoted
 437   // to.  This helps us consider truncates on PPC as noop copies when they
 438   // are.
 439   if (TLI.getTypeAction(CI->getContext(), SrcVT) == TargetLowering::Promote)
 440     SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
 441   if (TLI.getTypeAction(CI->getContext(), DstVT) == TargetLowering::Promote)
 442     DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
 443
 444   // If, after promotion, these are the same types, this is a noop copy.
 445   if (SrcVT != DstVT)
 446     return false;
 447
 448   BasicBlock *DefBB = CI->getParent();
 449
 450   /// InsertedCasts - Only insert a cast in each block once.
 451   DenseMap<BasicBlock*, CastInst*> InsertedCasts;
 452
 453   bool MadeChange = false;
 454   for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
 455        UI != E; ) {
 456     Use &TheUse = UI.getUse();
 457     Instruction *User = cast<Instruction>(*UI);
 458
 459     // Figure out which BB this cast is used in.  For PHI's this is the
 460     // appropriate predecessor block.
 461     BasicBlock *UserBB = User->getParent();
 462     if (PHINode *PN = dyn_cast<PHINode>(User)) {
 463       UserBB = PN->getIncomingBlock(UI);
 464     }
 465
 466     // Preincrement use iterator so we don't invalidate it.
 467     ++UI;
 468
 469     // If this user is in the same block as the cast, don't change the cast.
 470     if (UserBB == DefBB) continue;
 471
 472     // If we have already inserted a cast into this block, use it.
 473     CastInst *&InsertedCast = InsertedCasts[UserBB];
 474
 475     if (!InsertedCast) {
 476       BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
 477
 478       InsertedCast =
 479         CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
 480                          InsertPt);
 481       MadeChange = true;
 482     }
 483
 484     // Replace a use of the cast with a use of the new cast.
 485     TheUse = InsertedCast;
 486   }
 487
 488   // If we removed all uses, nuke the cast.
 489   if (CI->use_empty()) {
 490     CI->eraseFromParent();
 491     MadeChange = true;
 492   }
 493
 494   return MadeChange;
 495 }
 496
 497 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
 498 /// the number of virtual registers that must be created and coalesced.  This is
 499 /// a clear win except on targets with multiple condition code registers
 500 ///  (PowerPC), where it might lose; some adjustment may be wanted there.
 501 ///
 502 /// Return true if any changes are made.
 503 static bool OptimizeCmpExpression(CmpInst *CI) {
 504   BasicBlock *DefBB = CI->getParent();
 505
 506   /// InsertedCmp - Only insert a cmp in each block once.
 507   DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
 508
 509   bool MadeChange = false;
 510   for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
 511        UI != E; ) {
 512     Use &TheUse = UI.getUse();
 513     Instruction *User = cast<Instruction>(*UI);
 514
 515     // Preincrement use iterator so we don't invalidate it.
 516     ++UI;
 517
 518     // Don't bother for PHI nodes.
 519     if (isa<PHINode>(User))
 520       continue;
 521
 522     // Figure out which BB this cmp is used in.
 523     BasicBlock *UserBB = User->getParent();
 524
 525     // If this user is in the same block as the cmp, don't change the cmp.
 526     if (UserBB == DefBB) continue;
 527
 528     // If we have already inserted a cmp into this block, use it.
 529     CmpInst *&InsertedCmp = InsertedCmps[UserBB];
 530
 531     if (!InsertedCmp) {
 532       BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
 533
 534       InsertedCmp =
 535         CmpInst::Create(CI->getOpcode(),
 536                         CI->getPredicate(),  CI->getOperand(0),
 537                         CI->getOperand(1), "", InsertPt);
 538       MadeChange = true;
 539     }
 540
 541     // Replace a use of the cmp with a use of the new cmp.
 542     TheUse = InsertedCmp;
 543   }
 544
 545   // If we removed all uses, nuke the cmp.
 546   if (CI->use_empty())
 547     CI->eraseFromParent();
 548
 549   return MadeChange;
 550 }
 551
 552 //===----------------------------------------------------------------------===//
 553 // Memory Optimization
 554 //===----------------------------------------------------------------------===//
 555
 556 /// IsNonLocalValue - Return true if the specified values are defined in a
 557 /// different basic block than BB.
 558 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
 559   if (Instruction *I = dyn_cast<Instruction>(V))
 560     return I->getParent() != BB;
 561   return false;
 562 }
 563
 564 /// OptimizeMemoryInst - Load and Store Instructions have often have
 565 /// addressing modes that can do significant amounts of computation.  As such,
 566 /// instruction selection will try to get the load or store to do as much
 567 /// computation as possible for the program.  The problem is that isel can only
 568 /// see within a single block.  As such, we sink as much legal addressing mode
 569 /// stuff into the block as possible.
 570 ///
 571 /// This method is used to optimize both load/store and inline asms with memory
 572 /// operands.
 573 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
 574                                         const Type *AccessTy,
 575                                         DenseMap<Value*,Value*> &SunkAddrs) {
 576   // Figure out what addressing mode will be built up for this operation.
 577   SmallVector<Instruction*, 16> AddrModeInsts;
 578   ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst,
 579                                                       AddrModeInsts, *TLI);
 580
 581   // Check to see if any of the instructions supersumed by this addr mode are
 582   // non-local to I's BB.
 583   bool AnyNonLocal = false;
 584   for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
 585     if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
 586       AnyNonLocal = true;
 587       break;
 588     }
 589   }
 590
 591   // If all the instructions matched are already in this BB, don't do anything.
 592   if (!AnyNonLocal) {
 593     DEBUG(errs() << "CGP: Found      local addrmode: " << AddrMode << "\n");
 594     return false;
 595   }
 596
 597   // Insert this computation right after this user.  Since our caller is
 598   // scanning from the top of the BB to the bottom, reuse of the expr are
 599   // guaranteed to happen later.
 600   BasicBlock::iterator InsertPt = MemoryInst;
 601
 602   // Now that we determined the addressing expression we want to use and know
 603   // that we have to sink it into this block.  Check to see if we have already
 604   // done this for some other load/store instr in this block.  If so, reuse the
 605   // computation.
 606   Value *&SunkAddr = SunkAddrs[Addr];
 607   if (SunkAddr) {
 608     DEBUG(errs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
 609                  << *MemoryInst);
 610     if (SunkAddr->getType() != Addr->getType())
 611       SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
 612   } else {
 613     DEBUG(errs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
 614                  << *MemoryInst);
 615     const Type *IntPtrTy =
 616           TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
 617
 618     Value *Result = 0;
 619     // Start with the scale value.
 620     if (AddrMode.Scale) {
 621       Value *V = AddrMode.ScaledReg;
 622       if (V->getType() == IntPtrTy) {
 623         // done.
 624       } else if (isa<PointerType>(V->getType())) {
 625         V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
 626       } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
 627                  cast<IntegerType>(V->getType())->getBitWidth()) {
 628         V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
 629       } else {
 630         V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
 631       }
 632       if (AddrMode.Scale != 1)
 633         V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
 634                                                                 AddrMode.Scale),
 635                                       "sunkaddr", InsertPt);
 636       Result = V;
 637     }
 638
 639     // Add in the base register.
 640     if (AddrMode.BaseReg) {
 641       Value *V = AddrMode.BaseReg;
 642       if (isa<PointerType>(V->getType()))
 643         V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
 644       if (V->getType() != IntPtrTy)
 645         V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true,
 646                                         "sunkaddr", InsertPt);
 647       if (Result)
 648         Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
 649       else
 650         Result = V;
 651     }
 652
 653     // Add in the BaseGV if present.
 654     if (AddrMode.BaseGV) {
 655       Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
 656                                   InsertPt);
 657       if (Result)
 658         Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
 659       else
 660         Result = V;
 661     }
 662
 663     // Add in the Base Offset if present.
 664     if (AddrMode.BaseOffs) {
 665       Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
 666       if (Result)
 667         Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
 668       else
 669         Result = V;
 670     }
 671
 672     if (Result == 0)
 673       SunkAddr = Constant::getNullValue(Addr->getType());
 674     else
 675       SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
 676   }
 677
 678   MemoryInst->replaceUsesOfWith(Addr, SunkAddr);
 679
 680   if (Addr->use_empty())
 681     RecursivelyDeleteTriviallyDeadInstructions(Addr);
 682   return true;
 683 }
 684
 685 /// OptimizeInlineAsmInst - If there are any memory operands, use
 686 /// OptimizeMemoryInst to sink their address computing into the block when
 687 /// possible / profitable.
 688 bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS,
 689                                            DenseMap<Value*,Value*> &SunkAddrs) {
 690   bool MadeChange = false;
 691   InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
 692
 693   // Do a prepass over the constraints, canonicalizing them, and building up the
 694   // ConstraintOperands list.
 695   std::vector<InlineAsm::ConstraintInfo>
 696     ConstraintInfos = IA->ParseConstraints();
 697
 698   /// ConstraintOperands - Information about all of the constraints.
 699   std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands;
 700   unsigned ArgNo = 0;   // ArgNo - The argument of the CallInst.
 701   for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
 702     ConstraintOperands.
 703       push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i]));
 704     TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back();
 705
 706     // Compute the value type for each operand.
 707     switch (OpInfo.Type) {
 708     case InlineAsm::isOutput:
 709       if (OpInfo.isIndirect)
 710         OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
 711       break;
 712     case InlineAsm::isInput:
 713       OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
 714       break;
 715     case InlineAsm::isClobber:
 716       // Nothing to do.
 717       break;
 718     }
 719
 720     // Compute the constraint code and ConstraintType to use.
 721     TLI->ComputeConstraintToUse(OpInfo, SDValue(),
 722                              OpInfo.ConstraintType == TargetLowering::C_Memory);
 723
 724     if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
 725         OpInfo.isIndirect) {
 726       Value *OpVal = OpInfo.CallOperandVal;
 727       MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs);
 728     }
 729   }
 730
 731   return MadeChange;
 732 }
 733
 734 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
 735   BasicBlock *DefBB = I->getParent();
 736
 737   // If both result of the {s|z}xt and its source are live out, rewrite all
 738   // other uses of the source with result of extension.
 739   Value *Src = I->getOperand(0);
 740   if (Src->hasOneUse())
 741     return false;
 742
 743   // Only do this xform if truncating is free.
 744   if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
 745     return false;
 746
 747   // Only safe to perform the optimization if the source is also defined in
 748   // this block.
 749   if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
 750     return false;
 751
 752   bool DefIsLiveOut = false;
 753   for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
 754        UI != E; ++UI) {
 755     Instruction *User = cast<Instruction>(*UI);
 756
 757     // Figure out which BB this ext is used in.
 758     BasicBlock *UserBB = User->getParent();
 759     if (UserBB == DefBB) continue;
 760     DefIsLiveOut = true;
 761     break;
 762   }
 763   if (!DefIsLiveOut)
 764     return false;
 765
 766   // Make sure non of the uses are PHI nodes.
 767   for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
 768        UI != E; ++UI) {
 769     Instruction *User = cast<Instruction>(*UI);
 770     BasicBlock *UserBB = User->getParent();
 771     if (UserBB == DefBB) continue;
 772     // Be conservative. We don't want this xform to end up introducing
 773     // reloads just before load / store instructions.
 774     if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
 775       return false;
 776   }
 777
 778   // InsertedTruncs - Only insert one trunc in each block once.
 779   DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
 780
 781   bool MadeChange = false;
 782   for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
 783        UI != E; ++UI) {
 784     Use &TheUse = UI.getUse();
 785     Instruction *User = cast<Instruction>(*UI);
 786
 787     // Figure out which BB this ext is used in.
 788     BasicBlock *UserBB = User->getParent();
 789     if (UserBB == DefBB) continue;
 790
 791     // Both src and def are live in this block. Rewrite the use.
 792     Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
 793
 794     if (!InsertedTrunc) {
 795       BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
 796
 797       InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
 798     }
 799
 800     // Replace a use of the {s|z}ext source with a use of the result.
 801     TheUse = InsertedTrunc;
 802
 803     MadeChange = true;
 804   }
 805
 806   return MadeChange;
 807 }
 808
 809 // In this pass we look for GEP and cast instructions that are used
 810 // across basic blocks and rewrite them to improve basic-block-at-a-time
 811 // selection.
 812 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
 813   bool MadeChange = false;
 814
 815   // Split all critical edges where the dest block has a PHI.
 816   TerminatorInst *BBTI = BB.getTerminator();
 817   if (BBTI->getNumSuccessors() > 1) {
 818     for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) {
 819       BasicBlock *SuccBB = BBTI->getSuccessor(i);
 820       if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true))
 821         SplitEdgeNicely(BBTI, i, BackEdges, this);
 822     }
 823   }
 824
 825   // Keep track of non-local addresses that have been sunk into this block.
 826   // This allows us to avoid inserting duplicate code for blocks with multiple
 827   // load/stores of the same address.
 828   DenseMap<Value*, Value*> SunkAddrs;
 829
 830   for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) {
 831     Instruction *I = BBI++;
 832
 833     if (CastInst *CI = dyn_cast<CastInst>(I)) {
 834       // If the source of the cast is a constant, then this should have
 835       // already been constant folded.  The only reason NOT to constant fold
 836       // it is if something (e.g. LSR) was careful to place the constant
 837       // evaluation in a block other than then one that uses it (e.g. to hoist
 838       // the address of globals out of a loop).  If this is the case, we don't
 839       // want to forward-subst the cast.
 840       if (isa<Constant>(CI->getOperand(0)))
 841         continue;
 842
 843       bool Change = false;
 844       if (TLI) {
 845         Change = OptimizeNoopCopyExpression(CI, *TLI);
 846         MadeChange |= Change;
 847       }
 848
 849       if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I)))
 850         MadeChange |= OptimizeExtUses(I);
 851     } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
 852       MadeChange |= OptimizeCmpExpression(CI);
 853     } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
 854       if (TLI)
 855         MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(),
 856                                          SunkAddrs);
 857     } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
 858       if (TLI)
 859         MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1),
 860                                          SI->getOperand(0)->getType(),
 861                                          SunkAddrs);
 862     } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
 863       if (GEPI->hasAllZeroIndices()) {
 864         /// The GEP operand must be a pointer, so must its result -> BitCast
 865         Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
 866                                           GEPI->getName(), GEPI);
 867         GEPI->replaceAllUsesWith(NC);
 868         GEPI->eraseFromParent();
 869         MadeChange = true;
 870         BBI = NC;
 871       }
 872     } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
 873       // If we found an inline asm expession, and if the target knows how to
 874       // lower it to normal LLVM code, do so now.
 875       if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
 876         if (TLI->ExpandInlineAsm(CI)) {
 877           BBI = BB.begin();
 878           // Avoid processing instructions out of order, which could cause
 879           // reuse before a value is defined.
 880           SunkAddrs.clear();
 881         } else
 882           // Sink address computing for memory operands into the block.
 883           MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs);
 884       }
 885     }
 886   }
 887
 888   return MadeChange;
 889 }