1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "codegenprepare"
17 #include "llvm/Transforms/Scalar.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/ADT/ValueMap.h"
22 #include "llvm/Analysis/DominatorInternals.h"
23 #include "llvm/Analysis/Dominators.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Assembly/Writer.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/InlineAsm.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/Pass.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/ValueHandle.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Target/TargetLibraryInfo.h"
43 #include "llvm/Target/TargetLowering.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 #include "llvm/Transforms/Utils/BuildLibCalls.h"
46 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
47 #include "llvm/Transforms/Utils/Local.h"
49 using namespace llvm::PatternMatch;
51 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
52 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
53 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
54 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
56 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
58 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
59 "computations were sunk");
60 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
61 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
62 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
63 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
64 STATISTIC(NumSelectsExpanded, "Number of selects turned into branches");
66 static cl::opt<bool> DisableBranchOpts(
67 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
68 cl::desc("Disable branch optimizations in CodeGenPrepare"));
70 static cl::opt<bool> DisableSelectToBranch(
71 "disable-cgp-select2branch", cl::Hidden, cl::init(false),
72 cl::desc("Disable select to branch conversion."));
75 class CodeGenPrepare : public FunctionPass {
76 /// TLI - Keep a pointer of a TargetLowering to consult for determining
77 /// transformation profitability.
78 const TargetMachine *TM;
79 const TargetLowering *TLI;
80 const TargetLibraryInfo *TLInfo;
83 /// CurInstIterator - As we scan instructions optimizing them, this is the
84 /// next instruction to optimize. Xforms that can invalidate this should
86 BasicBlock::iterator CurInstIterator;
88 /// Keeps track of non-local addresses that have been sunk into a block.
89 /// This allows us to avoid inserting duplicate code for blocks with
90 /// multiple load/stores of the same address.
91 ValueMap<Value*, Value*> SunkAddrs;
93 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
97 /// OptSize - True if optimizing for size.
101 static char ID; // Pass identification, replacement for typeid
102 explicit CodeGenPrepare(const TargetMachine *TM = 0)
103 : FunctionPass(ID), TM(TM), TLI(0) {
104 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
106 bool runOnFunction(Function &F);
108 const char *getPassName() const { return "CodeGen Prepare"; }
110 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
111 AU.addPreserved<DominatorTree>();
112 AU.addRequired<TargetLibraryInfo>();
116 bool EliminateFallThrough(Function &F);
117 bool EliminateMostlyEmptyBlocks(Function &F);
118 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
119 void EliminateMostlyEmptyBlock(BasicBlock *BB);
120 bool OptimizeBlock(BasicBlock &BB);
121 bool OptimizeInst(Instruction *I);
122 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
123 bool OptimizeInlineAsmInst(CallInst *CS);
124 bool OptimizeCallInst(CallInst *CI);
125 bool MoveExtToFormExtLoad(Instruction *I);
126 bool OptimizeExtUses(Instruction *I);
127 bool OptimizeSelectInst(SelectInst *SI);
128 bool DupRetToEnableTailCallOpts(BasicBlock *BB);
129 bool PlaceDbgValues(Function &F);
133 char CodeGenPrepare::ID = 0;
134 INITIALIZE_PASS_BEGIN(CodeGenPrepare, "codegenprepare",
135 "Optimize for code generation", false, false)
136 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
137 INITIALIZE_PASS_END(CodeGenPrepare, "codegenprepare",
138 "Optimize for code generation", false, false)
140 FunctionPass *llvm::createCodeGenPreparePass(const TargetMachine *TM) {
141 return new CodeGenPrepare(TM);
144 bool CodeGenPrepare::runOnFunction(Function &F) {
145 bool EverMadeChange = false;
148 if (TM) TLI = TM->getTargetLowering();
149 TLInfo = &getAnalysis<TargetLibraryInfo>();
150 DT = getAnalysisIfAvailable<DominatorTree>();
151 OptSize = F.getAttributes().hasAttribute(AttributeSet::FunctionIndex,
152 Attribute::OptimizeForSize);
154 /// This optimization identifies DIV instructions that can be
155 /// profitably bypassed and carried out with a shorter, faster divide.
156 if (!OptSize && TLI && TLI->isSlowDivBypassed()) {
157 const DenseMap<unsigned int, unsigned int> &BypassWidths =
158 TLI->getBypassSlowDivWidths();
159 for (Function::iterator I = F.begin(); I != F.end(); I++)
160 EverMadeChange |= bypassSlowDivision(F, I, BypassWidths);
163 // Eliminate blocks that contain only PHI nodes and an
164 // unconditional branch.
165 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
167 // llvm.dbg.value is far away from the value then iSel may not be able
168 // handle it properly. iSel will drop llvm.dbg.value if it can not
169 // find a node corresponding to the value.
170 EverMadeChange |= PlaceDbgValues(F);
172 bool MadeChange = true;
175 for (Function::iterator I = F.begin(); I != F.end(); ) {
176 BasicBlock *BB = I++;
177 MadeChange |= OptimizeBlock(*BB);
179 EverMadeChange |= MadeChange;
184 if (!DisableBranchOpts) {
186 SmallPtrSet<BasicBlock*, 8> WorkList;
187 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
188 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
189 MadeChange |= ConstantFoldTerminator(BB, true);
190 if (!MadeChange) continue;
192 for (SmallVectorImpl<BasicBlock*>::iterator
193 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
194 if (pred_begin(*II) == pred_end(*II))
195 WorkList.insert(*II);
198 // Delete the dead blocks and any of their dead successors.
199 MadeChange |= !WorkList.empty();
200 while (!WorkList.empty()) {
201 BasicBlock *BB = *WorkList.begin();
203 SmallVector<BasicBlock*, 2> Successors(succ_begin(BB), succ_end(BB));
207 for (SmallVectorImpl<BasicBlock*>::iterator
208 II = Successors.begin(), IE = Successors.end(); II != IE; ++II)
209 if (pred_begin(*II) == pred_end(*II))
210 WorkList.insert(*II);
213 // Merge pairs of basic blocks with unconditional branches, connected by
215 if (EverMadeChange || MadeChange)
216 MadeChange |= EliminateFallThrough(F);
220 EverMadeChange |= MadeChange;
223 if (ModifiedDT && DT)
224 DT->DT->recalculate(F);
226 return EverMadeChange;
229 /// EliminateFallThrough - Merge basic blocks which are connected
230 /// by a single edge, where one of the basic blocks has a single successor
231 /// pointing to the other basic block, which has a single predecessor.
232 bool CodeGenPrepare::EliminateFallThrough(Function &F) {
233 bool Changed = false;
234 // Scan all of the blocks in the function, except for the entry block.
235 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
236 BasicBlock *BB = I++;
237 // If the destination block has a single pred, then this is a trivial
238 // edge, just collapse it.
239 BasicBlock *SinglePred = BB->getSinglePredecessor();
241 // Don't merge if BB's address is taken.
242 if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue;
244 BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator());
245 if (Term && !Term->isConditional()) {
247 DEBUG(dbgs() << "To merge:\n"<< *SinglePred << "\n\n\n");
248 // Remember if SinglePred was the entry block of the function.
249 // If so, we will need to move BB back to the entry position.
250 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
251 MergeBasicBlockIntoOnlyPred(BB, this);
253 if (isEntry && BB != &BB->getParent()->getEntryBlock())
254 BB->moveBefore(&BB->getParent()->getEntryBlock());
256 // We have erased a block. Update the iterator.
263 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
264 /// debug info directives, and an unconditional branch. Passes before isel
265 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
266 /// isel. Start by eliminating these blocks so we can split them the way we
268 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
269 bool MadeChange = false;
270 // Note that this intentionally skips the entry block.
271 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
272 BasicBlock *BB = I++;
274 // If this block doesn't end with an uncond branch, ignore it.
275 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
276 if (!BI || !BI->isUnconditional())
279 // If the instruction before the branch (skipping debug info) isn't a phi
280 // node, then other stuff is happening here.
281 BasicBlock::iterator BBI = BI;
282 if (BBI != BB->begin()) {
284 while (isa<DbgInfoIntrinsic>(BBI)) {
285 if (BBI == BB->begin())
289 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
293 // Do not break infinite loops.
294 BasicBlock *DestBB = BI->getSuccessor(0);
298 if (!CanMergeBlocks(BB, DestBB))
301 EliminateMostlyEmptyBlock(BB);
307 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
308 /// single uncond branch between them, and BB contains no other non-phi
310 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
311 const BasicBlock *DestBB) const {
312 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
313 // the successor. If there are more complex condition (e.g. preheaders),
314 // don't mess around with them.
315 BasicBlock::const_iterator BBI = BB->begin();
316 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
317 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
319 const Instruction *User = cast<Instruction>(*UI);
320 if (User->getParent() != DestBB || !isa<PHINode>(User))
322 // If User is inside DestBB block and it is a PHINode then check
323 // incoming value. If incoming value is not from BB then this is
324 // a complex condition (e.g. preheaders) we want to avoid here.
325 if (User->getParent() == DestBB) {
326 if (const PHINode *UPN = dyn_cast<PHINode>(User))
327 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
328 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
329 if (Insn && Insn->getParent() == BB &&
330 Insn->getParent() != UPN->getIncomingBlock(I))
337 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
338 // and DestBB may have conflicting incoming values for the block. If so, we
339 // can't merge the block.
340 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
341 if (!DestBBPN) return true; // no conflict.
343 // Collect the preds of BB.
344 SmallPtrSet<const BasicBlock*, 16> BBPreds;
345 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
346 // It is faster to get preds from a PHI than with pred_iterator.
347 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
348 BBPreds.insert(BBPN->getIncomingBlock(i));
350 BBPreds.insert(pred_begin(BB), pred_end(BB));
353 // Walk the preds of DestBB.
354 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
355 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
356 if (BBPreds.count(Pred)) { // Common predecessor?
357 BBI = DestBB->begin();
358 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
359 const Value *V1 = PN->getIncomingValueForBlock(Pred);
360 const Value *V2 = PN->getIncomingValueForBlock(BB);
362 // If V2 is a phi node in BB, look up what the mapped value will be.
363 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
364 if (V2PN->getParent() == BB)
365 V2 = V2PN->getIncomingValueForBlock(Pred);
367 // If there is a conflict, bail out.
368 if (V1 != V2) return false;
377 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
378 /// an unconditional branch in it.
379 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
380 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
381 BasicBlock *DestBB = BI->getSuccessor(0);
383 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
385 // If the destination block has a single pred, then this is a trivial edge,
387 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
388 if (SinglePred != DestBB) {
389 // Remember if SinglePred was the entry block of the function. If so, we
390 // will need to move BB back to the entry position.
391 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
392 MergeBasicBlockIntoOnlyPred(DestBB, this);
394 if (isEntry && BB != &BB->getParent()->getEntryBlock())
395 BB->moveBefore(&BB->getParent()->getEntryBlock());
397 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
402 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
403 // to handle the new incoming edges it is about to have.
405 for (BasicBlock::iterator BBI = DestBB->begin();
406 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
407 // Remove the incoming value for BB, and remember it.
408 Value *InVal = PN->removeIncomingValue(BB, false);
410 // Two options: either the InVal is a phi node defined in BB or it is some
411 // value that dominates BB.
412 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
413 if (InValPhi && InValPhi->getParent() == BB) {
414 // Add all of the input values of the input PHI as inputs of this phi.
415 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
416 PN->addIncoming(InValPhi->getIncomingValue(i),
417 InValPhi->getIncomingBlock(i));
419 // Otherwise, add one instance of the dominating value for each edge that
420 // we will be adding.
421 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
422 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
423 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
425 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
426 PN->addIncoming(InVal, *PI);
431 // The PHIs are now updated, change everything that refers to BB to use
432 // DestBB and remove BB.
433 BB->replaceAllUsesWith(DestBB);
434 if (DT && !ModifiedDT) {
435 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
436 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
437 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
438 DT->changeImmediateDominator(DestBB, NewIDom);
441 BB->eraseFromParent();
444 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
447 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
448 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
449 /// sink it into user blocks to reduce the number of virtual
450 /// registers that must be created and coalesced.
452 /// Return true if any changes are made.
454 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
455 // If this is a noop copy,
456 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
457 EVT DstVT = TLI.getValueType(CI->getType());
459 // This is an fp<->int conversion?
460 if (SrcVT.isInteger() != DstVT.isInteger())
463 // If this is an extension, it will be a zero or sign extension, which
465 if (SrcVT.bitsLT(DstVT)) return false;
467 // If these values will be promoted, find out what they will be promoted
468 // to. This helps us consider truncates on PPC as noop copies when they
470 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
471 TargetLowering::TypePromoteInteger)
472 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
473 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
474 TargetLowering::TypePromoteInteger)
475 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
477 // If, after promotion, these are the same types, this is a noop copy.
481 BasicBlock *DefBB = CI->getParent();
483 /// InsertedCasts - Only insert a cast in each block once.
484 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
486 bool MadeChange = false;
487 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
489 Use &TheUse = UI.getUse();
490 Instruction *User = cast<Instruction>(*UI);
492 // Figure out which BB this cast is used in. For PHI's this is the
493 // appropriate predecessor block.
494 BasicBlock *UserBB = User->getParent();
495 if (PHINode *PN = dyn_cast<PHINode>(User)) {
496 UserBB = PN->getIncomingBlock(UI);
499 // Preincrement use iterator so we don't invalidate it.
502 // If this user is in the same block as the cast, don't change the cast.
503 if (UserBB == DefBB) continue;
505 // If we have already inserted a cast into this block, use it.
506 CastInst *&InsertedCast = InsertedCasts[UserBB];
509 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
511 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
516 // Replace a use of the cast with a use of the new cast.
517 TheUse = InsertedCast;
521 // If we removed all uses, nuke the cast.
522 if (CI->use_empty()) {
523 CI->eraseFromParent();
530 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
531 /// the number of virtual registers that must be created and coalesced. This is
532 /// a clear win except on targets with multiple condition code registers
533 /// (PowerPC), where it might lose; some adjustment may be wanted there.
535 /// Return true if any changes are made.
536 static bool OptimizeCmpExpression(CmpInst *CI) {
537 BasicBlock *DefBB = CI->getParent();
539 /// InsertedCmp - Only insert a cmp in each block once.
540 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
542 bool MadeChange = false;
543 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
545 Use &TheUse = UI.getUse();
546 Instruction *User = cast<Instruction>(*UI);
548 // Preincrement use iterator so we don't invalidate it.
551 // Don't bother for PHI nodes.
552 if (isa<PHINode>(User))
555 // Figure out which BB this cmp is used in.
556 BasicBlock *UserBB = User->getParent();
558 // If this user is in the same block as the cmp, don't change the cmp.
559 if (UserBB == DefBB) continue;
561 // If we have already inserted a cmp into this block, use it.
562 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
565 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
567 CmpInst::Create(CI->getOpcode(),
568 CI->getPredicate(), CI->getOperand(0),
569 CI->getOperand(1), "", InsertPt);
573 // Replace a use of the cmp with a use of the new cmp.
574 TheUse = InsertedCmp;
578 // If we removed all uses, nuke the cmp.
580 CI->eraseFromParent();
586 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
588 void replaceCall(Value *With) {
589 CI->replaceAllUsesWith(With);
590 CI->eraseFromParent();
592 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
593 if (ConstantInt *SizeCI =
594 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
595 return SizeCI->isAllOnesValue();
599 } // end anonymous namespace
601 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
602 BasicBlock *BB = CI->getParent();
604 // Lower inline assembly if we can.
605 // If we found an inline asm expession, and if the target knows how to
606 // lower it to normal LLVM code, do so now.
607 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
608 if (TLI->ExpandInlineAsm(CI)) {
609 // Avoid invalidating the iterator.
610 CurInstIterator = BB->begin();
611 // Avoid processing instructions out of order, which could cause
612 // reuse before a value is defined.
616 // Sink address computing for memory operands into the block.
617 if (OptimizeInlineAsmInst(CI))
621 // Lower all uses of llvm.objectsize.*
622 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
623 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
624 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
625 Type *ReturnTy = CI->getType();
626 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
628 // Substituting this can cause recursive simplifications, which can
629 // invalidate our iterator. Use a WeakVH to hold onto it in case this
631 WeakVH IterHandle(CurInstIterator);
633 replaceAndRecursivelySimplify(CI, RetVal, TLI ? TLI->getDataLayout() : 0,
634 TLInfo, ModifiedDT ? 0 : DT);
636 // If the iterator instruction was recursively deleted, start over at the
637 // start of the block.
638 if (IterHandle != CurInstIterator) {
639 CurInstIterator = BB->begin();
646 SmallVector<Value*, 2> PtrOps;
648 if (TLI->GetAddrModeArguments(II, PtrOps, AccessTy))
649 while (!PtrOps.empty())
650 if (OptimizeMemoryInst(II, PtrOps.pop_back_val(), AccessTy))
654 // From here on out we're working with named functions.
655 if (CI->getCalledFunction() == 0) return false;
657 // We'll need DataLayout from here on out.
658 const DataLayout *TD = TLI ? TLI->getDataLayout() : 0;
659 if (!TD) return false;
661 // Lower all default uses of _chk calls. This is very similar
662 // to what InstCombineCalls does, but here we are only lowering calls
663 // that have the default "don't know" as the objectsize. Anything else
664 // should be left alone.
665 CodeGenPrepareFortifiedLibCalls Simplifier;
666 return Simplifier.fold(CI, TD, TLInfo);
669 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
670 /// instructions to the predecessor to enable tail call optimizations. The
671 /// case it is currently looking for is:
674 /// %tmp0 = tail call i32 @f0()
677 /// %tmp1 = tail call i32 @f1()
680 /// %tmp2 = tail call i32 @f2()
683 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
691 /// %tmp0 = tail call i32 @f0()
694 /// %tmp1 = tail call i32 @f1()
697 /// %tmp2 = tail call i32 @f2()
700 bool CodeGenPrepare::DupRetToEnableTailCallOpts(BasicBlock *BB) {
704 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator());
709 BitCastInst *BCI = 0;
710 Value *V = RI->getReturnValue();
712 BCI = dyn_cast<BitCastInst>(V);
714 V = BCI->getOperand(0);
716 PN = dyn_cast<PHINode>(V);
721 if (PN && PN->getParent() != BB)
724 // It's not safe to eliminate the sign / zero extension of the return value.
725 // See llvm::isInTailCallPosition().
726 const Function *F = BB->getParent();
727 AttributeSet CallerAttrs = F->getAttributes();
728 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) ||
729 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
732 // Make sure there are no instructions between the PHI and return, or that the
733 // return is the first instruction in the block.
735 BasicBlock::iterator BI = BB->begin();
736 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
738 // Also skip over the bitcast.
743 BasicBlock::iterator BI = BB->begin();
744 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
749 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
751 SmallVector<CallInst*, 4> TailCalls;
753 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
754 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
755 // Make sure the phi value is indeed produced by the tail call.
756 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
757 TLI->mayBeEmittedAsTailCall(CI))
758 TailCalls.push_back(CI);
761 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
762 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
763 if (!VisitedBBs.insert(*PI))
766 BasicBlock::InstListType &InstList = (*PI)->getInstList();
767 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
768 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
769 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
773 CallInst *CI = dyn_cast<CallInst>(&*RI);
774 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
775 TailCalls.push_back(CI);
779 bool Changed = false;
780 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
781 CallInst *CI = TailCalls[i];
784 // Conservatively require the attributes of the call to match those of the
785 // return. Ignore noalias because it doesn't affect the call sequence.
786 AttributeSet CalleeAttrs = CS.getAttributes();
787 if (AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
788 removeAttribute(Attribute::NoAlias) !=
789 AttrBuilder(CalleeAttrs, AttributeSet::ReturnIndex).
790 removeAttribute(Attribute::NoAlias))
793 // Make sure the call instruction is followed by an unconditional branch to
795 BasicBlock *CallBB = CI->getParent();
796 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
797 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
800 // Duplicate the return into CallBB.
801 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
802 ModifiedDT = Changed = true;
806 // If we eliminated all predecessors of the block, delete the block now.
807 if (Changed && !BB->hasAddressTaken() && pred_begin(BB) == pred_end(BB))
808 BB->eraseFromParent();
813 //===----------------------------------------------------------------------===//
814 // Memory Optimization
815 //===----------------------------------------------------------------------===//
819 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
820 /// which holds actual Value*'s for register values.
821 struct ExtAddrMode : public TargetLowering::AddrMode {
824 ExtAddrMode() : BaseReg(0), ScaledReg(0) {}
825 void print(raw_ostream &OS) const;
828 bool operator==(const ExtAddrMode& O) const {
829 return (BaseReg == O.BaseReg) && (ScaledReg == O.ScaledReg) &&
830 (BaseGV == O.BaseGV) && (BaseOffs == O.BaseOffs) &&
831 (HasBaseReg == O.HasBaseReg) && (Scale == O.Scale);
836 static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) {
842 void ExtAddrMode::print(raw_ostream &OS) const {
843 bool NeedPlus = false;
846 OS << (NeedPlus ? " + " : "")
848 WriteAsOperand(OS, BaseGV, /*PrintType=*/false);
853 OS << (NeedPlus ? " + " : "") << BaseOffs, NeedPlus = true;
856 OS << (NeedPlus ? " + " : "")
858 WriteAsOperand(OS, BaseReg, /*PrintType=*/false);
862 OS << (NeedPlus ? " + " : "")
864 WriteAsOperand(OS, ScaledReg, /*PrintType=*/false);
870 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
871 void ExtAddrMode::dump() const {
878 /// \brief A helper class for matching addressing modes.
880 /// This encapsulates the logic for matching the target-legal addressing modes.
881 class AddressingModeMatcher {
882 SmallVectorImpl<Instruction*> &AddrModeInsts;
883 const TargetLowering &TLI;
885 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
886 /// the memory instruction that we're computing this address for.
888 Instruction *MemoryInst;
890 /// AddrMode - This is the addressing mode that we're building up. This is
891 /// part of the return value of this addressing mode matching stuff.
892 ExtAddrMode &AddrMode;
894 /// IgnoreProfitability - This is set to true when we should not do
895 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
896 /// always returns true.
897 bool IgnoreProfitability;
899 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
900 const TargetLowering &T, Type *AT,
901 Instruction *MI, ExtAddrMode &AM)
902 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM) {
903 IgnoreProfitability = false;
907 /// Match - Find the maximal addressing mode that a load/store of V can fold,
908 /// give an access type of AccessTy. This returns a list of involved
909 /// instructions in AddrModeInsts.
910 static ExtAddrMode Match(Value *V, Type *AccessTy,
911 Instruction *MemoryInst,
912 SmallVectorImpl<Instruction*> &AddrModeInsts,
913 const TargetLowering &TLI) {
917 AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
918 MemoryInst, Result).MatchAddr(V, 0);
919 (void)Success; assert(Success && "Couldn't select *anything*?");
923 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
924 bool MatchAddr(Value *V, unsigned Depth);
925 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth);
926 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
927 ExtAddrMode &AMBefore,
928 ExtAddrMode &AMAfter);
929 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
932 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
933 /// Return true and update AddrMode if this addr mode is legal for the target,
935 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
937 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
938 // mode. Just process that directly.
940 return MatchAddr(ScaleReg, Depth);
942 // If the scale is 0, it takes nothing to add this.
946 // If we already have a scale of this value, we can add to it, otherwise, we
947 // need an available scale field.
948 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
951 ExtAddrMode TestAddrMode = AddrMode;
953 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
954 // [A+B + A*7] -> [B+A*8].
955 TestAddrMode.Scale += Scale;
956 TestAddrMode.ScaledReg = ScaleReg;
958 // If the new address isn't legal, bail out.
959 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
962 // It was legal, so commit it.
963 AddrMode = TestAddrMode;
965 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
966 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
967 // X*Scale + C*Scale to addr mode.
968 ConstantInt *CI = 0; Value *AddLHS = 0;
969 if (isa<Instruction>(ScaleReg) && // not a constant expr.
970 match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
971 TestAddrMode.ScaledReg = AddLHS;
972 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
974 // If this addressing mode is legal, commit it and remember that we folded
976 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
977 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
978 AddrMode = TestAddrMode;
983 // Otherwise, not (x+c)*scale, just return what we have.
987 /// MightBeFoldableInst - This is a little filter, which returns true if an
988 /// addressing computation involving I might be folded into a load/store
989 /// accessing it. This doesn't need to be perfect, but needs to accept at least
990 /// the set of instructions that MatchOperationAddr can.
991 static bool MightBeFoldableInst(Instruction *I) {
992 switch (I->getOpcode()) {
993 case Instruction::BitCast:
994 // Don't touch identity bitcasts.
995 if (I->getType() == I->getOperand(0)->getType())
997 return I->getType()->isPointerTy() || I->getType()->isIntegerTy();
998 case Instruction::PtrToInt:
999 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1001 case Instruction::IntToPtr:
1002 // We know the input is intptr_t, so this is foldable.
1004 case Instruction::Add:
1006 case Instruction::Mul:
1007 case Instruction::Shl:
1008 // Can only handle X*C and X << C.
1009 return isa<ConstantInt>(I->getOperand(1));
1010 case Instruction::GetElementPtr:
1017 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
1018 /// fold the operation into the addressing mode. If so, update the addressing
1019 /// mode and return true, otherwise return false without modifying AddrMode.
1020 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
1022 // Avoid exponential behavior on extremely deep expression trees.
1023 if (Depth >= 5) return false;
1026 case Instruction::PtrToInt:
1027 // PtrToInt is always a noop, as we know that the int type is pointer sized.
1028 return MatchAddr(AddrInst->getOperand(0), Depth);
1029 case Instruction::IntToPtr:
1030 // This inttoptr is a no-op if the integer type is pointer sized.
1031 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
1032 TLI.getPointerTy(AddrInst->getType()->getPointerAddressSpace()))
1033 return MatchAddr(AddrInst->getOperand(0), Depth);
1035 case Instruction::BitCast:
1036 // BitCast is always a noop, and we can handle it as long as it is
1037 // int->int or pointer->pointer (we don't want int<->fp or something).
1038 if ((AddrInst->getOperand(0)->getType()->isPointerTy() ||
1039 AddrInst->getOperand(0)->getType()->isIntegerTy()) &&
1040 // Don't touch identity bitcasts. These were probably put here by LSR,
1041 // and we don't want to mess around with them. Assume it knows what it
1043 AddrInst->getOperand(0)->getType() != AddrInst->getType())
1044 return MatchAddr(AddrInst->getOperand(0), Depth);
1046 case Instruction::Add: {
1047 // Check to see if we can merge in the RHS then the LHS. If so, we win.
1048 ExtAddrMode BackupAddrMode = AddrMode;
1049 unsigned OldSize = AddrModeInsts.size();
1050 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
1051 MatchAddr(AddrInst->getOperand(0), Depth+1))
1054 // Restore the old addr mode info.
1055 AddrMode = BackupAddrMode;
1056 AddrModeInsts.resize(OldSize);
1058 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
1059 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
1060 MatchAddr(AddrInst->getOperand(1), Depth+1))
1063 // Otherwise we definitely can't merge the ADD in.
1064 AddrMode = BackupAddrMode;
1065 AddrModeInsts.resize(OldSize);
1068 //case Instruction::Or:
1069 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
1071 case Instruction::Mul:
1072 case Instruction::Shl: {
1073 // Can only handle X*C and X << C.
1074 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
1075 if (!RHS) return false;
1076 int64_t Scale = RHS->getSExtValue();
1077 if (Opcode == Instruction::Shl)
1078 Scale = 1LL << Scale;
1080 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
1082 case Instruction::GetElementPtr: {
1083 // Scan the GEP. We check it if it contains constant offsets and at most
1084 // one variable offset.
1085 int VariableOperand = -1;
1086 unsigned VariableScale = 0;
1088 int64_t ConstantOffset = 0;
1089 const DataLayout *TD = TLI.getDataLayout();
1090 gep_type_iterator GTI = gep_type_begin(AddrInst);
1091 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
1092 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
1093 const StructLayout *SL = TD->getStructLayout(STy);
1095 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
1096 ConstantOffset += SL->getElementOffset(Idx);
1098 uint64_t TypeSize = TD->getTypeAllocSize(GTI.getIndexedType());
1099 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
1100 ConstantOffset += CI->getSExtValue()*TypeSize;
1101 } else if (TypeSize) { // Scales of zero don't do anything.
1102 // We only allow one variable index at the moment.
1103 if (VariableOperand != -1)
1106 // Remember the variable index.
1107 VariableOperand = i;
1108 VariableScale = TypeSize;
1113 // A common case is for the GEP to only do a constant offset. In this case,
1114 // just add it to the disp field and check validity.
1115 if (VariableOperand == -1) {
1116 AddrMode.BaseOffs += ConstantOffset;
1117 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
1118 // Check to see if we can fold the base pointer in too.
1119 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
1122 AddrMode.BaseOffs -= ConstantOffset;
1126 // Save the valid addressing mode in case we can't match.
1127 ExtAddrMode BackupAddrMode = AddrMode;
1128 unsigned OldSize = AddrModeInsts.size();
1130 // See if the scale and offset amount is valid for this target.
1131 AddrMode.BaseOffs += ConstantOffset;
1133 // Match the base operand of the GEP.
1134 if (!MatchAddr(AddrInst->getOperand(0), Depth+1)) {
1135 // If it couldn't be matched, just stuff the value in a register.
1136 if (AddrMode.HasBaseReg) {
1137 AddrMode = BackupAddrMode;
1138 AddrModeInsts.resize(OldSize);
1141 AddrMode.HasBaseReg = true;
1142 AddrMode.BaseReg = AddrInst->getOperand(0);
1145 // Match the remaining variable portion of the GEP.
1146 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
1148 // If it couldn't be matched, try stuffing the base into a register
1149 // instead of matching it, and retrying the match of the scale.
1150 AddrMode = BackupAddrMode;
1151 AddrModeInsts.resize(OldSize);
1152 if (AddrMode.HasBaseReg)
1154 AddrMode.HasBaseReg = true;
1155 AddrMode.BaseReg = AddrInst->getOperand(0);
1156 AddrMode.BaseOffs += ConstantOffset;
1157 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand),
1158 VariableScale, Depth)) {
1159 // If even that didn't work, bail.
1160 AddrMode = BackupAddrMode;
1161 AddrModeInsts.resize(OldSize);
1172 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
1173 /// addressing mode. If Addr can't be added to AddrMode this returns false and
1174 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
1175 /// or intptr_t for the target.
1177 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
1178 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
1179 // Fold in immediates if legal for the target.
1180 AddrMode.BaseOffs += CI->getSExtValue();
1181 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1183 AddrMode.BaseOffs -= CI->getSExtValue();
1184 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
1185 // If this is a global variable, try to fold it into the addressing mode.
1186 if (AddrMode.BaseGV == 0) {
1187 AddrMode.BaseGV = GV;
1188 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1190 AddrMode.BaseGV = 0;
1192 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
1193 ExtAddrMode BackupAddrMode = AddrMode;
1194 unsigned OldSize = AddrModeInsts.size();
1196 // Check to see if it is possible to fold this operation.
1197 if (MatchOperationAddr(I, I->getOpcode(), Depth)) {
1198 // Okay, it's possible to fold this. Check to see if it is actually
1199 // *profitable* to do so. We use a simple cost model to avoid increasing
1200 // register pressure too much.
1201 if (I->hasOneUse() ||
1202 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
1203 AddrModeInsts.push_back(I);
1207 // It isn't profitable to do this, roll back.
1208 //cerr << "NOT FOLDING: " << *I;
1209 AddrMode = BackupAddrMode;
1210 AddrModeInsts.resize(OldSize);
1212 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
1213 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
1215 } else if (isa<ConstantPointerNull>(Addr)) {
1216 // Null pointer gets folded without affecting the addressing mode.
1220 // Worse case, the target should support [reg] addressing modes. :)
1221 if (!AddrMode.HasBaseReg) {
1222 AddrMode.HasBaseReg = true;
1223 AddrMode.BaseReg = Addr;
1224 // Still check for legality in case the target supports [imm] but not [i+r].
1225 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1227 AddrMode.HasBaseReg = false;
1228 AddrMode.BaseReg = 0;
1231 // If the base register is already taken, see if we can do [r+r].
1232 if (AddrMode.Scale == 0) {
1234 AddrMode.ScaledReg = Addr;
1235 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
1238 AddrMode.ScaledReg = 0;
1244 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
1245 /// inline asm call are due to memory operands. If so, return true, otherwise
1247 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
1248 const TargetLowering &TLI) {
1249 TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(ImmutableCallSite(CI));
1250 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
1251 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
1253 // Compute the constraint code and ConstraintType to use.
1254 TLI.ComputeConstraintToUse(OpInfo, SDValue());
1256 // If this asm operand is our Value*, and if it isn't an indirect memory
1257 // operand, we can't fold it!
1258 if (OpInfo.CallOperandVal == OpVal &&
1259 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
1260 !OpInfo.isIndirect))
1267 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
1268 /// memory use. If we find an obviously non-foldable instruction, return true.
1269 /// Add the ultimately found memory instructions to MemoryUses.
1270 static bool FindAllMemoryUses(Instruction *I,
1271 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
1272 SmallPtrSet<Instruction*, 16> &ConsideredInsts,
1273 const TargetLowering &TLI) {
1274 // If we already considered this instruction, we're done.
1275 if (!ConsideredInsts.insert(I))
1278 // If this is an obviously unfoldable instruction, bail out.
1279 if (!MightBeFoldableInst(I))
1282 // Loop over all the uses, recursively processing them.
1283 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1287 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
1288 MemoryUses.push_back(std::make_pair(LI, UI.getOperandNo()));
1292 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
1293 unsigned opNo = UI.getOperandNo();
1294 if (opNo == 0) return true; // Storing addr, not into addr.
1295 MemoryUses.push_back(std::make_pair(SI, opNo));
1299 if (CallInst *CI = dyn_cast<CallInst>(U)) {
1300 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
1301 if (!IA) return true;
1303 // If this is a memory operand, we're cool, otherwise bail out.
1304 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
1309 if (FindAllMemoryUses(cast<Instruction>(U), MemoryUses, ConsideredInsts,
1317 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
1318 /// the use site that we're folding it into. If so, there is no cost to
1319 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
1320 /// that we know are live at the instruction already.
1321 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
1322 Value *KnownLive2) {
1323 // If Val is either of the known-live values, we know it is live!
1324 if (Val == 0 || Val == KnownLive1 || Val == KnownLive2)
1327 // All values other than instructions and arguments (e.g. constants) are live.
1328 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
1330 // If Val is a constant sized alloca in the entry block, it is live, this is
1331 // true because it is just a reference to the stack/frame pointer, which is
1332 // live for the whole function.
1333 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
1334 if (AI->isStaticAlloca())
1337 // Check to see if this value is already used in the memory instruction's
1338 // block. If so, it's already live into the block at the very least, so we
1339 // can reasonably fold it.
1340 return Val->isUsedInBasicBlock(MemoryInst->getParent());
1343 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
1344 /// mode of the machine to fold the specified instruction into a load or store
1345 /// that ultimately uses it. However, the specified instruction has multiple
1346 /// uses. Given this, it may actually increase register pressure to fold it
1347 /// into the load. For example, consider this code:
1351 /// use(Y) -> nonload/store
1355 /// In this case, Y has multiple uses, and can be folded into the load of Z
1356 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
1357 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
1358 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
1359 /// number of computations either.
1361 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
1362 /// X was live across 'load Z' for other reasons, we actually *would* want to
1363 /// fold the addressing mode in the Z case. This would make Y die earlier.
1364 bool AddressingModeMatcher::
1365 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
1366 ExtAddrMode &AMAfter) {
1367 if (IgnoreProfitability) return true;
1369 // AMBefore is the addressing mode before this instruction was folded into it,
1370 // and AMAfter is the addressing mode after the instruction was folded. Get
1371 // the set of registers referenced by AMAfter and subtract out those
1372 // referenced by AMBefore: this is the set of values which folding in this
1373 // address extends the lifetime of.
1375 // Note that there are only two potential values being referenced here,
1376 // BaseReg and ScaleReg (global addresses are always available, as are any
1377 // folded immediates).
1378 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
1380 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
1381 // lifetime wasn't extended by adding this instruction.
1382 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
1384 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
1387 // If folding this instruction (and it's subexprs) didn't extend any live
1388 // ranges, we're ok with it.
1389 if (BaseReg == 0 && ScaledReg == 0)
1392 // If all uses of this instruction are ultimately load/store/inlineasm's,
1393 // check to see if their addressing modes will include this instruction. If
1394 // so, we can fold it into all uses, so it doesn't matter if it has multiple
1396 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
1397 SmallPtrSet<Instruction*, 16> ConsideredInsts;
1398 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
1399 return false; // Has a non-memory, non-foldable use!
1401 // Now that we know that all uses of this instruction are part of a chain of
1402 // computation involving only operations that could theoretically be folded
1403 // into a memory use, loop over each of these uses and see if they could
1404 // *actually* fold the instruction.
1405 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
1406 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
1407 Instruction *User = MemoryUses[i].first;
1408 unsigned OpNo = MemoryUses[i].second;
1410 // Get the access type of this use. If the use isn't a pointer, we don't
1411 // know what it accesses.
1412 Value *Address = User->getOperand(OpNo);
1413 if (!Address->getType()->isPointerTy())
1415 Type *AddressAccessTy = Address->getType()->getPointerElementType();
1417 // Do a match against the root of this address, ignoring profitability. This
1418 // will tell us if the addressing mode for the memory operation will
1419 // *actually* cover the shared instruction.
1421 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
1422 MemoryInst, Result);
1423 Matcher.IgnoreProfitability = true;
1424 bool Success = Matcher.MatchAddr(Address, 0);
1425 (void)Success; assert(Success && "Couldn't select *anything*?");
1427 // If the match didn't cover I, then it won't be shared by it.
1428 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
1429 I) == MatchedAddrModeInsts.end())
1432 MatchedAddrModeInsts.clear();
1438 } // end anonymous namespace
1440 /// IsNonLocalValue - Return true if the specified values are defined in a
1441 /// different basic block than BB.
1442 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
1443 if (Instruction *I = dyn_cast<Instruction>(V))
1444 return I->getParent() != BB;
1448 /// OptimizeMemoryInst - Load and Store Instructions often have
1449 /// addressing modes that can do significant amounts of computation. As such,
1450 /// instruction selection will try to get the load or store to do as much
1451 /// computation as possible for the program. The problem is that isel can only
1452 /// see within a single block. As such, we sink as much legal addressing mode
1453 /// stuff into the block as possible.
1455 /// This method is used to optimize both load/store and inline asms with memory
1457 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
1461 // Try to collapse single-value PHI nodes. This is necessary to undo
1462 // unprofitable PRE transformations.
1463 SmallVector<Value*, 8> worklist;
1464 SmallPtrSet<Value*, 16> Visited;
1465 worklist.push_back(Addr);
1467 // Use a worklist to iteratively look through PHI nodes, and ensure that
1468 // the addressing mode obtained from the non-PHI roots of the graph
1470 Value *Consensus = 0;
1471 unsigned NumUsesConsensus = 0;
1472 bool IsNumUsesConsensusValid = false;
1473 SmallVector<Instruction*, 16> AddrModeInsts;
1474 ExtAddrMode AddrMode;
1475 while (!worklist.empty()) {
1476 Value *V = worklist.back();
1477 worklist.pop_back();
1479 // Break use-def graph loops.
1480 if (!Visited.insert(V)) {
1485 // For a PHI node, push all of its incoming values.
1486 if (PHINode *P = dyn_cast<PHINode>(V)) {
1487 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
1488 worklist.push_back(P->getIncomingValue(i));
1492 // For non-PHIs, determine the addressing mode being computed.
1493 SmallVector<Instruction*, 16> NewAddrModeInsts;
1494 ExtAddrMode NewAddrMode =
1495 AddressingModeMatcher::Match(V, AccessTy, MemoryInst,
1496 NewAddrModeInsts, *TLI);
1498 // This check is broken into two cases with very similar code to avoid using
1499 // getNumUses() as much as possible. Some values have a lot of uses, so
1500 // calling getNumUses() unconditionally caused a significant compile-time
1504 AddrMode = NewAddrMode;
1505 AddrModeInsts = NewAddrModeInsts;
1507 } else if (NewAddrMode == AddrMode) {
1508 if (!IsNumUsesConsensusValid) {
1509 NumUsesConsensus = Consensus->getNumUses();
1510 IsNumUsesConsensusValid = true;
1513 // Ensure that the obtained addressing mode is equivalent to that obtained
1514 // for all other roots of the PHI traversal. Also, when choosing one
1515 // such root as representative, select the one with the most uses in order
1516 // to keep the cost modeling heuristics in AddressingModeMatcher
1518 unsigned NumUses = V->getNumUses();
1519 if (NumUses > NumUsesConsensus) {
1521 NumUsesConsensus = NumUses;
1522 AddrModeInsts = NewAddrModeInsts;
1531 // If the addressing mode couldn't be determined, or if multiple different
1532 // ones were determined, bail out now.
1533 if (!Consensus) return false;
1535 // Check to see if any of the instructions supersumed by this addr mode are
1536 // non-local to I's BB.
1537 bool AnyNonLocal = false;
1538 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
1539 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
1545 // If all the instructions matched are already in this BB, don't do anything.
1547 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
1551 // Insert this computation right after this user. Since our caller is
1552 // scanning from the top of the BB to the bottom, reuse of the expr are
1553 // guaranteed to happen later.
1554 IRBuilder<> Builder(MemoryInst);
1556 // Now that we determined the addressing expression we want to use and know
1557 // that we have to sink it into this block. Check to see if we have already
1558 // done this for some other load/store instr in this block. If so, reuse the
1560 Value *&SunkAddr = SunkAddrs[Addr];
1562 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
1564 if (SunkAddr->getType() != Addr->getType())
1565 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
1567 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
1569 Type *IntPtrTy = TLI->getDataLayout()->getIntPtrType(Addr->getType());
1572 // Start with the base register. Do this first so that subsequent address
1573 // matching finds it last, which will prevent it from trying to match it
1574 // as the scaled value in case it happens to be a mul. That would be
1575 // problematic if we've sunk a different mul for the scale, because then
1576 // we'd end up sinking both muls.
1577 if (AddrMode.BaseReg) {
1578 Value *V = AddrMode.BaseReg;
1579 if (V->getType()->isPointerTy())
1580 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
1581 if (V->getType() != IntPtrTy)
1582 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
1586 // Add the scale value.
1587 if (AddrMode.Scale) {
1588 Value *V = AddrMode.ScaledReg;
1589 if (V->getType() == IntPtrTy) {
1591 } else if (V->getType()->isPointerTy()) {
1592 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
1593 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
1594 cast<IntegerType>(V->getType())->getBitWidth()) {
1595 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
1597 V = Builder.CreateSExt(V, IntPtrTy, "sunkaddr");
1599 if (AddrMode.Scale != 1)
1600 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
1603 Result = Builder.CreateAdd(Result, V, "sunkaddr");
1608 // Add in the BaseGV if present.
1609 if (AddrMode.BaseGV) {
1610 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
1612 Result = Builder.CreateAdd(Result, V, "sunkaddr");
1617 // Add in the Base Offset if present.
1618 if (AddrMode.BaseOffs) {
1619 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
1621 Result = Builder.CreateAdd(Result, V, "sunkaddr");
1627 SunkAddr = Constant::getNullValue(Addr->getType());
1629 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
1632 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
1634 // If we have no uses, recursively delete the value and all dead instructions
1636 if (Repl->use_empty()) {
1637 // This can cause recursive deletion, which can invalidate our iterator.
1638 // Use a WeakVH to hold onto it in case this happens.
1639 WeakVH IterHandle(CurInstIterator);
1640 BasicBlock *BB = CurInstIterator->getParent();
1642 RecursivelyDeleteTriviallyDeadInstructions(Repl, TLInfo);
1644 if (IterHandle != CurInstIterator) {
1645 // If the iterator instruction was recursively deleted, start over at the
1646 // start of the block.
1647 CurInstIterator = BB->begin();
1655 /// OptimizeInlineAsmInst - If there are any memory operands, use
1656 /// OptimizeMemoryInst to sink their address computing into the block when
1657 /// possible / profitable.
1658 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
1659 bool MadeChange = false;
1661 TargetLowering::AsmOperandInfoVector
1662 TargetConstraints = TLI->ParseConstraints(CS);
1664 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
1665 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
1667 // Compute the constraint code and ConstraintType to use.
1668 TLI->ComputeConstraintToUse(OpInfo, SDValue());
1670 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
1671 OpInfo.isIndirect) {
1672 Value *OpVal = CS->getArgOperand(ArgNo++);
1673 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
1674 } else if (OpInfo.Type == InlineAsm::isInput)
1681 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
1682 /// basic block as the load, unless conditions are unfavorable. This allows
1683 /// SelectionDAG to fold the extend into the load.
1685 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
1686 // Look for a load being extended.
1687 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
1688 if (!LI) return false;
1690 // If they're already in the same block, there's nothing to do.
1691 if (LI->getParent() == I->getParent())
1694 // If the load has other users and the truncate is not free, this probably
1695 // isn't worthwhile.
1696 if (!LI->hasOneUse() &&
1697 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
1698 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
1699 !TLI->isTruncateFree(I->getType(), LI->getType()))
1702 // Check whether the target supports casts folded into loads.
1704 if (isa<ZExtInst>(I))
1705 LType = ISD::ZEXTLOAD;
1707 assert(isa<SExtInst>(I) && "Unexpected ext type!");
1708 LType = ISD::SEXTLOAD;
1710 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
1713 // Move the extend into the same block as the load, so that SelectionDAG
1715 I->removeFromParent();
1721 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
1722 BasicBlock *DefBB = I->getParent();
1724 // If the result of a {s|z}ext and its source are both live out, rewrite all
1725 // other uses of the source with result of extension.
1726 Value *Src = I->getOperand(0);
1727 if (Src->hasOneUse())
1730 // Only do this xform if truncating is free.
1731 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
1734 // Only safe to perform the optimization if the source is also defined in
1736 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
1739 bool DefIsLiveOut = false;
1740 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1742 Instruction *User = cast<Instruction>(*UI);
1744 // Figure out which BB this ext is used in.
1745 BasicBlock *UserBB = User->getParent();
1746 if (UserBB == DefBB) continue;
1747 DefIsLiveOut = true;
1753 // Make sure none of the uses are PHI nodes.
1754 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1756 Instruction *User = cast<Instruction>(*UI);
1757 BasicBlock *UserBB = User->getParent();
1758 if (UserBB == DefBB) continue;
1759 // Be conservative. We don't want this xform to end up introducing
1760 // reloads just before load / store instructions.
1761 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
1765 // InsertedTruncs - Only insert one trunc in each block once.
1766 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
1768 bool MadeChange = false;
1769 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1771 Use &TheUse = UI.getUse();
1772 Instruction *User = cast<Instruction>(*UI);
1774 // Figure out which BB this ext is used in.
1775 BasicBlock *UserBB = User->getParent();
1776 if (UserBB == DefBB) continue;
1778 // Both src and def are live in this block. Rewrite the use.
1779 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
1781 if (!InsertedTrunc) {
1782 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1783 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
1786 // Replace a use of the {s|z}ext source with a use of the result.
1787 TheUse = InsertedTrunc;
1795 /// isFormingBranchFromSelectProfitable - Returns true if a SelectInst should be
1796 /// turned into an explicit branch.
1797 static bool isFormingBranchFromSelectProfitable(SelectInst *SI) {
1798 // FIXME: This should use the same heuristics as IfConversion to determine
1799 // whether a select is better represented as a branch. This requires that
1800 // branch probability metadata is preserved for the select, which is not the
1803 CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition());
1805 // If the branch is predicted right, an out of order CPU can avoid blocking on
1806 // the compare. Emit cmovs on compares with a memory operand as branches to
1807 // avoid stalls on the load from memory. If the compare has more than one use
1808 // there's probably another cmov or setcc around so it's not worth emitting a
1813 Value *CmpOp0 = Cmp->getOperand(0);
1814 Value *CmpOp1 = Cmp->getOperand(1);
1816 // We check that the memory operand has one use to avoid uses of the loaded
1817 // value directly after the compare, making branches unprofitable.
1818 return Cmp->hasOneUse() &&
1819 ((isa<LoadInst>(CmpOp0) && CmpOp0->hasOneUse()) ||
1820 (isa<LoadInst>(CmpOp1) && CmpOp1->hasOneUse()));
1824 /// If we have a SelectInst that will likely profit from branch prediction,
1825 /// turn it into a branch.
1826 bool CodeGenPrepare::OptimizeSelectInst(SelectInst *SI) {
1827 bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1);
1829 // Can we convert the 'select' to CF ?
1830 if (DisableSelectToBranch || OptSize || !TLI || VectorCond)
1833 TargetLowering::SelectSupportKind SelectKind;
1835 SelectKind = TargetLowering::VectorMaskSelect;
1836 else if (SI->getType()->isVectorTy())
1837 SelectKind = TargetLowering::ScalarCondVectorVal;
1839 SelectKind = TargetLowering::ScalarValSelect;
1841 // Do we have efficient codegen support for this kind of 'selects' ?
1842 if (TLI->isSelectSupported(SelectKind)) {
1843 // We have efficient codegen support for the select instruction.
1844 // Check if it is profitable to keep this 'select'.
1845 if (!TLI->isPredictableSelectExpensive() ||
1846 !isFormingBranchFromSelectProfitable(SI))
1852 // First, we split the block containing the select into 2 blocks.
1853 BasicBlock *StartBlock = SI->getParent();
1854 BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(SI));
1855 BasicBlock *NextBlock = StartBlock->splitBasicBlock(SplitPt, "select.end");
1857 // Create a new block serving as the landing pad for the branch.
1858 BasicBlock *SmallBlock = BasicBlock::Create(SI->getContext(), "select.mid",
1859 NextBlock->getParent(), NextBlock);
1861 // Move the unconditional branch from the block with the select in it into our
1862 // landing pad block.
1863 StartBlock->getTerminator()->eraseFromParent();
1864 BranchInst::Create(NextBlock, SmallBlock);
1866 // Insert the real conditional branch based on the original condition.
1867 BranchInst::Create(NextBlock, SmallBlock, SI->getCondition(), SI);
1869 // The select itself is replaced with a PHI Node.
1870 PHINode *PN = PHINode::Create(SI->getType(), 2, "", NextBlock->begin());
1872 PN->addIncoming(SI->getTrueValue(), StartBlock);
1873 PN->addIncoming(SI->getFalseValue(), SmallBlock);
1874 SI->replaceAllUsesWith(PN);
1875 SI->eraseFromParent();
1877 // Instruct OptimizeBlock to skip to the next block.
1878 CurInstIterator = StartBlock->end();
1879 ++NumSelectsExpanded;
1883 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
1884 if (PHINode *P = dyn_cast<PHINode>(I)) {
1885 // It is possible for very late stage optimizations (such as SimplifyCFG)
1886 // to introduce PHI nodes too late to be cleaned up. If we detect such a
1887 // trivial PHI, go ahead and zap it here.
1888 if (Value *V = SimplifyInstruction(P, TLI ? TLI->getDataLayout() : 0,
1890 P->replaceAllUsesWith(V);
1891 P->eraseFromParent();
1898 if (CastInst *CI = dyn_cast<CastInst>(I)) {
1899 // If the source of the cast is a constant, then this should have
1900 // already been constant folded. The only reason NOT to constant fold
1901 // it is if something (e.g. LSR) was careful to place the constant
1902 // evaluation in a block other than then one that uses it (e.g. to hoist
1903 // the address of globals out of a loop). If this is the case, we don't
1904 // want to forward-subst the cast.
1905 if (isa<Constant>(CI->getOperand(0)))
1908 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
1911 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
1912 bool MadeChange = MoveExtToFormExtLoad(I);
1913 return MadeChange | OptimizeExtUses(I);
1918 if (CmpInst *CI = dyn_cast<CmpInst>(I))
1919 if (!TLI || !TLI->hasMultipleConditionRegisters())
1920 return OptimizeCmpExpression(CI);
1922 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1924 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
1928 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1930 return OptimizeMemoryInst(I, SI->getOperand(1),
1931 SI->getOperand(0)->getType());
1935 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1936 if (GEPI->hasAllZeroIndices()) {
1937 /// The GEP operand must be a pointer, so must its result -> BitCast
1938 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1939 GEPI->getName(), GEPI);
1940 GEPI->replaceAllUsesWith(NC);
1941 GEPI->eraseFromParent();
1949 if (CallInst *CI = dyn_cast<CallInst>(I))
1950 return OptimizeCallInst(CI);
1952 if (SelectInst *SI = dyn_cast<SelectInst>(I))
1953 return OptimizeSelectInst(SI);
1958 // In this pass we look for GEP and cast instructions that are used
1959 // across basic blocks and rewrite them to improve basic-block-at-a-time
1961 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
1963 bool MadeChange = false;
1965 CurInstIterator = BB.begin();
1966 while (CurInstIterator != BB.end())
1967 MadeChange |= OptimizeInst(CurInstIterator++);
1969 MadeChange |= DupRetToEnableTailCallOpts(&BB);
1974 // llvm.dbg.value is far away from the value then iSel may not be able
1975 // handle it properly. iSel will drop llvm.dbg.value if it can not
1976 // find a node corresponding to the value.
1977 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
1978 bool MadeChange = false;
1979 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
1980 Instruction *PrevNonDbgInst = NULL;
1981 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
1982 Instruction *Insn = BI; ++BI;
1983 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
1985 PrevNonDbgInst = Insn;
1989 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
1990 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
1991 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
1992 DVI->removeFromParent();
1993 if (isa<PHINode>(VI))
1994 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
1996 DVI->insertAfter(VI);