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/Constants.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Function.h"
21 #include "llvm/InlineAsm.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/IntrinsicInst.h"
24 #include "llvm/Pass.h"
25 #include "llvm/Analysis/Dominators.h"
26 #include "llvm/Analysis/InstructionSimplify.h"
27 #include "llvm/Analysis/ProfileInfo.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Target/TargetLowering.h"
30 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
31 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
32 #include "llvm/Transforms/Utils/Local.h"
33 #include "llvm/Transforms/Utils/BuildLibCalls.h"
34 #include "llvm/ADT/DenseMap.h"
35 #include "llvm/ADT/SmallSet.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/Assembly/Writer.h"
38 #include "llvm/Support/CallSite.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/GetElementPtrTypeIterator.h"
42 #include "llvm/Support/PatternMatch.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Support/IRBuilder.h"
45 #include "llvm/Support/ValueHandle.h"
47 using namespace llvm::PatternMatch;
49 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
50 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
51 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
52 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
54 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
56 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
57 "computations were sunk");
58 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
59 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
60 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
62 static cl::opt<bool> DisableBranchOpts(
63 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
64 cl::desc("Disable branch optimizations in CodeGenPrepare"));
67 class CodeGenPrepare : public FunctionPass {
68 /// TLI - Keep a pointer of a TargetLowering to consult for determining
69 /// transformation profitability.
70 const TargetLowering *TLI;
74 /// CurInstIterator - As we scan instructions optimizing them, this is the
75 /// next instruction to optimize. Xforms that can invalidate this should
77 BasicBlock::iterator CurInstIterator;
79 /// Keeps track of non-local addresses that have been sunk into a block.
80 /// This allows us to avoid inserting duplicate code for blocks with
81 /// multiple load/stores of the same address.
82 DenseMap<Value*, Value*> SunkAddrs;
84 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
89 static char ID; // Pass identification, replacement for typeid
90 explicit CodeGenPrepare(const TargetLowering *tli = 0)
91 : FunctionPass(ID), TLI(tli) {
92 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
94 bool runOnFunction(Function &F);
96 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
97 AU.addPreserved<DominatorTree>();
98 AU.addPreserved<ProfileInfo>();
102 bool EliminateMostlyEmptyBlocks(Function &F);
103 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
104 void EliminateMostlyEmptyBlock(BasicBlock *BB);
105 bool OptimizeBlock(BasicBlock &BB);
106 bool OptimizeInst(Instruction *I);
107 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
108 bool OptimizeInlineAsmInst(CallInst *CS);
109 bool OptimizeCallInst(CallInst *CI);
110 bool MoveExtToFormExtLoad(Instruction *I);
111 bool OptimizeExtUses(Instruction *I);
112 bool DupRetToEnableTailCallOpts(ReturnInst *RI);
116 char CodeGenPrepare::ID = 0;
117 INITIALIZE_PASS(CodeGenPrepare, "codegenprepare",
118 "Optimize for code generation", false, false)
120 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
121 return new CodeGenPrepare(TLI);
124 bool CodeGenPrepare::runOnFunction(Function &F) {
125 bool EverMadeChange = false;
128 DT = getAnalysisIfAvailable<DominatorTree>();
129 PFI = getAnalysisIfAvailable<ProfileInfo>();
131 // First pass, eliminate blocks that contain only PHI nodes and an
132 // unconditional branch.
133 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
135 bool MadeChange = true;
138 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
139 BasicBlock *BB = I++;
140 MadeChange |= OptimizeBlock(*BB);
142 EverMadeChange |= MadeChange;
147 if (!DisableBranchOpts) {
149 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
150 MadeChange |= ConstantFoldTerminator(BB, true);
154 EverMadeChange |= MadeChange;
157 if (ModifiedDT && DT)
158 DT->DT->recalculate(F);
160 return EverMadeChange;
163 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
164 /// debug info directives, and an unconditional branch. Passes before isel
165 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
166 /// isel. Start by eliminating these blocks so we can split them the way we
168 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
169 bool MadeChange = false;
170 // Note that this intentionally skips the entry block.
171 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
172 BasicBlock *BB = I++;
174 // If this block doesn't end with an uncond branch, ignore it.
175 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
176 if (!BI || !BI->isUnconditional())
179 // If the instruction before the branch (skipping debug info) isn't a phi
180 // node, then other stuff is happening here.
181 BasicBlock::iterator BBI = BI;
182 if (BBI != BB->begin()) {
184 while (isa<DbgInfoIntrinsic>(BBI)) {
185 if (BBI == BB->begin())
189 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
193 // Do not break infinite loops.
194 BasicBlock *DestBB = BI->getSuccessor(0);
198 if (!CanMergeBlocks(BB, DestBB))
201 EliminateMostlyEmptyBlock(BB);
207 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
208 /// single uncond branch between them, and BB contains no other non-phi
210 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
211 const BasicBlock *DestBB) const {
212 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
213 // the successor. If there are more complex condition (e.g. preheaders),
214 // don't mess around with them.
215 BasicBlock::const_iterator BBI = BB->begin();
216 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
217 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
219 const Instruction *User = cast<Instruction>(*UI);
220 if (User->getParent() != DestBB || !isa<PHINode>(User))
222 // If User is inside DestBB block and it is a PHINode then check
223 // incoming value. If incoming value is not from BB then this is
224 // a complex condition (e.g. preheaders) we want to avoid here.
225 if (User->getParent() == DestBB) {
226 if (const PHINode *UPN = dyn_cast<PHINode>(User))
227 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
228 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
229 if (Insn && Insn->getParent() == BB &&
230 Insn->getParent() != UPN->getIncomingBlock(I))
237 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
238 // and DestBB may have conflicting incoming values for the block. If so, we
239 // can't merge the block.
240 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
241 if (!DestBBPN) return true; // no conflict.
243 // Collect the preds of BB.
244 SmallPtrSet<const BasicBlock*, 16> BBPreds;
245 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
246 // It is faster to get preds from a PHI than with pred_iterator.
247 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
248 BBPreds.insert(BBPN->getIncomingBlock(i));
250 BBPreds.insert(pred_begin(BB), pred_end(BB));
253 // Walk the preds of DestBB.
254 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
255 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
256 if (BBPreds.count(Pred)) { // Common predecessor?
257 BBI = DestBB->begin();
258 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
259 const Value *V1 = PN->getIncomingValueForBlock(Pred);
260 const Value *V2 = PN->getIncomingValueForBlock(BB);
262 // If V2 is a phi node in BB, look up what the mapped value will be.
263 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
264 if (V2PN->getParent() == BB)
265 V2 = V2PN->getIncomingValueForBlock(Pred);
267 // If there is a conflict, bail out.
268 if (V1 != V2) return false;
277 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
278 /// an unconditional branch in it.
279 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
280 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
281 BasicBlock *DestBB = BI->getSuccessor(0);
283 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
285 // If the destination block has a single pred, then this is a trivial edge,
287 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
288 if (SinglePred != DestBB) {
289 // Remember if SinglePred was the entry block of the function. If so, we
290 // will need to move BB back to the entry position.
291 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
292 MergeBasicBlockIntoOnlyPred(DestBB, this);
294 if (isEntry && BB != &BB->getParent()->getEntryBlock())
295 BB->moveBefore(&BB->getParent()->getEntryBlock());
297 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
302 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
303 // to handle the new incoming edges it is about to have.
305 for (BasicBlock::iterator BBI = DestBB->begin();
306 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
307 // Remove the incoming value for BB, and remember it.
308 Value *InVal = PN->removeIncomingValue(BB, false);
310 // Two options: either the InVal is a phi node defined in BB or it is some
311 // value that dominates BB.
312 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
313 if (InValPhi && InValPhi->getParent() == BB) {
314 // Add all of the input values of the input PHI as inputs of this phi.
315 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
316 PN->addIncoming(InValPhi->getIncomingValue(i),
317 InValPhi->getIncomingBlock(i));
319 // Otherwise, add one instance of the dominating value for each edge that
320 // we will be adding.
321 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
322 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
323 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
325 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
326 PN->addIncoming(InVal, *PI);
331 // The PHIs are now updated, change everything that refers to BB to use
332 // DestBB and remove BB.
333 BB->replaceAllUsesWith(DestBB);
334 if (DT && !ModifiedDT) {
335 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
336 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
337 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
338 DT->changeImmediateDominator(DestBB, NewIDom);
342 PFI->replaceAllUses(BB, DestBB);
343 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
345 BB->eraseFromParent();
348 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
351 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
352 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
353 /// sink it into user blocks to reduce the number of virtual
354 /// registers that must be created and coalesced.
356 /// Return true if any changes are made.
358 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
359 // If this is a noop copy,
360 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
361 EVT DstVT = TLI.getValueType(CI->getType());
363 // This is an fp<->int conversion?
364 if (SrcVT.isInteger() != DstVT.isInteger())
367 // If this is an extension, it will be a zero or sign extension, which
369 if (SrcVT.bitsLT(DstVT)) return false;
371 // If these values will be promoted, find out what they will be promoted
372 // to. This helps us consider truncates on PPC as noop copies when they
374 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
375 TargetLowering::TypePromoteInteger)
376 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
377 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
378 TargetLowering::TypePromoteInteger)
379 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
381 // If, after promotion, these are the same types, this is a noop copy.
385 BasicBlock *DefBB = CI->getParent();
387 /// InsertedCasts - Only insert a cast in each block once.
388 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
390 bool MadeChange = false;
391 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
393 Use &TheUse = UI.getUse();
394 Instruction *User = cast<Instruction>(*UI);
396 // Figure out which BB this cast is used in. For PHI's this is the
397 // appropriate predecessor block.
398 BasicBlock *UserBB = User->getParent();
399 if (PHINode *PN = dyn_cast<PHINode>(User)) {
400 UserBB = PN->getIncomingBlock(UI);
403 // Preincrement use iterator so we don't invalidate it.
406 // If this user is in the same block as the cast, don't change the cast.
407 if (UserBB == DefBB) continue;
409 // If we have already inserted a cast into this block, use it.
410 CastInst *&InsertedCast = InsertedCasts[UserBB];
413 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
415 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
420 // Replace a use of the cast with a use of the new cast.
421 TheUse = InsertedCast;
425 // If we removed all uses, nuke the cast.
426 if (CI->use_empty()) {
427 CI->eraseFromParent();
434 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
435 /// the number of virtual registers that must be created and coalesced. This is
436 /// a clear win except on targets with multiple condition code registers
437 /// (PowerPC), where it might lose; some adjustment may be wanted there.
439 /// Return true if any changes are made.
440 static bool OptimizeCmpExpression(CmpInst *CI) {
441 BasicBlock *DefBB = CI->getParent();
443 /// InsertedCmp - Only insert a cmp in each block once.
444 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
446 bool MadeChange = false;
447 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
449 Use &TheUse = UI.getUse();
450 Instruction *User = cast<Instruction>(*UI);
452 // Preincrement use iterator so we don't invalidate it.
455 // Don't bother for PHI nodes.
456 if (isa<PHINode>(User))
459 // Figure out which BB this cmp is used in.
460 BasicBlock *UserBB = User->getParent();
462 // If this user is in the same block as the cmp, don't change the cmp.
463 if (UserBB == DefBB) continue;
465 // If we have already inserted a cmp into this block, use it.
466 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
469 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
471 CmpInst::Create(CI->getOpcode(),
472 CI->getPredicate(), CI->getOperand(0),
473 CI->getOperand(1), "", InsertPt);
477 // Replace a use of the cmp with a use of the new cmp.
478 TheUse = InsertedCmp;
482 // If we removed all uses, nuke the cmp.
484 CI->eraseFromParent();
490 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
492 void replaceCall(Value *With) {
493 CI->replaceAllUsesWith(With);
494 CI->eraseFromParent();
496 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
497 if (ConstantInt *SizeCI =
498 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
499 return SizeCI->isAllOnesValue();
503 } // end anonymous namespace
505 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
506 BasicBlock *BB = CI->getParent();
508 // Lower inline assembly if we can.
509 // If we found an inline asm expession, and if the target knows how to
510 // lower it to normal LLVM code, do so now.
511 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
512 if (TLI->ExpandInlineAsm(CI)) {
513 // Avoid invalidating the iterator.
514 CurInstIterator = BB->begin();
515 // Avoid processing instructions out of order, which could cause
516 // reuse before a value is defined.
520 // Sink address computing for memory operands into the block.
521 if (OptimizeInlineAsmInst(CI))
525 // Lower all uses of llvm.objectsize.*
526 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
527 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
528 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
529 Type *ReturnTy = CI->getType();
530 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
532 // Substituting this can cause recursive simplifications, which can
533 // invalidate our iterator. Use a WeakVH to hold onto it in case this
535 WeakVH IterHandle(CurInstIterator);
537 ReplaceAndSimplifyAllUses(CI, RetVal, TLI ? TLI->getTargetData() : 0,
538 ModifiedDT ? 0 : DT);
540 // If the iterator instruction was recursively deleted, start over at the
541 // start of the block.
542 if (IterHandle != CurInstIterator) {
543 CurInstIterator = BB->begin();
549 // From here on out we're working with named functions.
550 if (CI->getCalledFunction() == 0) return false;
552 // llvm.dbg.value is far away from the value then iSel may not be able
553 // handle it properly. iSel will drop llvm.dbg.value if it can not
554 // find a node corresponding to the value.
555 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(CI))
556 if (Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue()))
557 if (!VI->isTerminator() &&
558 (DVI->getParent() != VI->getParent() || DT->dominates(DVI, VI))) {
559 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
560 DVI->removeFromParent();
561 if (isa<PHINode>(VI))
562 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
564 DVI->insertAfter(VI);
568 // We'll need TargetData from here on out.
569 const TargetData *TD = TLI ? TLI->getTargetData() : 0;
570 if (!TD) return false;
572 // Lower all default uses of _chk calls. This is very similar
573 // to what InstCombineCalls does, but here we are only lowering calls
574 // that have the default "don't know" as the objectsize. Anything else
575 // should be left alone.
576 CodeGenPrepareFortifiedLibCalls Simplifier;
577 return Simplifier.fold(CI, TD);
580 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
581 /// instructions to the predecessor to enable tail call optimizations. The
582 /// case it is currently looking for is:
584 /// %tmp0 = tail call i32 @f0()
587 /// %tmp1 = tail call i32 @f1()
590 /// %tmp2 = tail call i32 @f2()
593 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
599 /// %tmp0 = tail call i32 @f0()
602 /// %tmp1 = tail call i32 @f1()
605 /// %tmp2 = tail call i32 @f2()
608 bool CodeGenPrepare::DupRetToEnableTailCallOpts(ReturnInst *RI) {
612 Value *V = RI->getReturnValue();
613 PHINode *PN = V ? dyn_cast<PHINode>(V) : NULL;
617 BasicBlock *BB = RI->getParent();
618 if (PN && PN->getParent() != BB)
621 // It's not safe to eliminate the sign / zero extension of the return value.
622 // See llvm::isInTailCallPosition().
623 const Function *F = BB->getParent();
624 unsigned CallerRetAttr = F->getAttributes().getRetAttributes();
625 if ((CallerRetAttr & Attribute::ZExt) || (CallerRetAttr & Attribute::SExt))
628 // Make sure there are no instructions between the PHI and return, or that the
629 // return is the first instruction in the block.
631 BasicBlock::iterator BI = BB->begin();
632 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
636 BasicBlock::iterator BI = BB->begin();
637 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
642 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
644 SmallVector<CallInst*, 4> TailCalls;
646 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
647 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
648 // Make sure the phi value is indeed produced by the tail call.
649 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
650 TLI->mayBeEmittedAsTailCall(CI))
651 TailCalls.push_back(CI);
654 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
655 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
656 if (!VisitedBBs.insert(*PI))
659 BasicBlock::InstListType &InstList = (*PI)->getInstList();
660 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
661 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
662 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
666 CallInst *CI = dyn_cast<CallInst>(&*RI);
667 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
668 TailCalls.push_back(CI);
672 bool Changed = false;
673 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
674 CallInst *CI = TailCalls[i];
677 // Conservatively require the attributes of the call to match those of the
678 // return. Ignore noalias because it doesn't affect the call sequence.
679 unsigned CalleeRetAttr = CS.getAttributes().getRetAttributes();
680 if ((CalleeRetAttr ^ CallerRetAttr) & ~Attribute::NoAlias)
683 // Make sure the call instruction is followed by an unconditional branch to
685 BasicBlock *CallBB = CI->getParent();
686 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
687 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
690 // Duplicate the return into CallBB.
691 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
692 ModifiedDT = Changed = true;
696 // If we eliminated all predecessors of the block, delete the block now.
697 if (Changed && pred_begin(BB) == pred_end(BB))
698 BB->eraseFromParent();
703 //===----------------------------------------------------------------------===//
704 // Memory Optimization
705 //===----------------------------------------------------------------------===//
707 /// IsNonLocalValue - Return true if the specified values are defined in a
708 /// different basic block than BB.
709 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
710 if (Instruction *I = dyn_cast<Instruction>(V))
711 return I->getParent() != BB;
715 /// OptimizeMemoryInst - Load and Store Instructions often have
716 /// addressing modes that can do significant amounts of computation. As such,
717 /// instruction selection will try to get the load or store to do as much
718 /// computation as possible for the program. The problem is that isel can only
719 /// see within a single block. As such, we sink as much legal addressing mode
720 /// stuff into the block as possible.
722 /// This method is used to optimize both load/store and inline asms with memory
724 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
728 // Try to collapse single-value PHI nodes. This is necessary to undo
729 // unprofitable PRE transformations.
730 SmallVector<Value*, 8> worklist;
731 SmallPtrSet<Value*, 16> Visited;
732 worklist.push_back(Addr);
734 // Use a worklist to iteratively look through PHI nodes, and ensure that
735 // the addressing mode obtained from the non-PHI roots of the graph
737 Value *Consensus = 0;
738 unsigned NumUsesConsensus = 0;
739 bool IsNumUsesConsensusValid = false;
740 SmallVector<Instruction*, 16> AddrModeInsts;
741 ExtAddrMode AddrMode;
742 while (!worklist.empty()) {
743 Value *V = worklist.back();
746 // Break use-def graph loops.
747 if (Visited.count(V)) {
754 // For a PHI node, push all of its incoming values.
755 if (PHINode *P = dyn_cast<PHINode>(V)) {
756 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
757 worklist.push_back(P->getIncomingValue(i));
761 // For non-PHIs, determine the addressing mode being computed.
762 SmallVector<Instruction*, 16> NewAddrModeInsts;
763 ExtAddrMode NewAddrMode =
764 AddressingModeMatcher::Match(V, AccessTy,MemoryInst,
765 NewAddrModeInsts, *TLI);
767 // This check is broken into two cases with very similar code to avoid using
768 // getNumUses() as much as possible. Some values have a lot of uses, so
769 // calling getNumUses() unconditionally caused a significant compile-time
773 AddrMode = NewAddrMode;
774 AddrModeInsts = NewAddrModeInsts;
776 } else if (NewAddrMode == AddrMode) {
777 if (!IsNumUsesConsensusValid) {
778 NumUsesConsensus = Consensus->getNumUses();
779 IsNumUsesConsensusValid = true;
782 // Ensure that the obtained addressing mode is equivalent to that obtained
783 // for all other roots of the PHI traversal. Also, when choosing one
784 // such root as representative, select the one with the most uses in order
785 // to keep the cost modeling heuristics in AddressingModeMatcher
787 unsigned NumUses = V->getNumUses();
788 if (NumUses > NumUsesConsensus) {
790 NumUsesConsensus = NumUses;
791 AddrModeInsts = NewAddrModeInsts;
800 // If the addressing mode couldn't be determined, or if multiple different
801 // ones were determined, bail out now.
802 if (!Consensus) return false;
804 // Check to see if any of the instructions supersumed by this addr mode are
805 // non-local to I's BB.
806 bool AnyNonLocal = false;
807 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
808 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
814 // If all the instructions matched are already in this BB, don't do anything.
816 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
820 // Insert this computation right after this user. Since our caller is
821 // scanning from the top of the BB to the bottom, reuse of the expr are
822 // guaranteed to happen later.
823 BasicBlock::iterator InsertPt = MemoryInst;
825 // Now that we determined the addressing expression we want to use and know
826 // that we have to sink it into this block. Check to see if we have already
827 // done this for some other load/store instr in this block. If so, reuse the
829 Value *&SunkAddr = SunkAddrs[Addr];
831 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
833 if (SunkAddr->getType() != Addr->getType())
834 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
836 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
839 TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
843 // Start with the base register. Do this first so that subsequent address
844 // matching finds it last, which will prevent it from trying to match it
845 // as the scaled value in case it happens to be a mul. That would be
846 // problematic if we've sunk a different mul for the scale, because then
847 // we'd end up sinking both muls.
848 if (AddrMode.BaseReg) {
849 Value *V = AddrMode.BaseReg;
850 if (V->getType()->isPointerTy())
851 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
852 if (V->getType() != IntPtrTy)
853 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true,
854 "sunkaddr", InsertPt);
858 // Add the scale value.
859 if (AddrMode.Scale) {
860 Value *V = AddrMode.ScaledReg;
861 if (V->getType() == IntPtrTy) {
863 } else if (V->getType()->isPointerTy()) {
864 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
865 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
866 cast<IntegerType>(V->getType())->getBitWidth()) {
867 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
869 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
871 if (AddrMode.Scale != 1)
872 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
874 "sunkaddr", InsertPt);
876 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
881 // Add in the BaseGV if present.
882 if (AddrMode.BaseGV) {
883 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
886 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
891 // Add in the Base Offset if present.
892 if (AddrMode.BaseOffs) {
893 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
895 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
901 SunkAddr = Constant::getNullValue(Addr->getType());
903 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
906 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
908 // If we have no uses, recursively delete the value and all dead instructions
910 if (Repl->use_empty()) {
911 // This can cause recursive deletion, which can invalidate our iterator.
912 // Use a WeakVH to hold onto it in case this happens.
913 WeakVH IterHandle(CurInstIterator);
914 BasicBlock *BB = CurInstIterator->getParent();
916 RecursivelyDeleteTriviallyDeadInstructions(Repl);
918 if (IterHandle != CurInstIterator) {
919 // If the iterator instruction was recursively deleted, start over at the
920 // start of the block.
921 CurInstIterator = BB->begin();
924 // This address is now available for reassignment, so erase the table
925 // entry; we don't want to match some completely different instruction.
933 /// OptimizeInlineAsmInst - If there are any memory operands, use
934 /// OptimizeMemoryInst to sink their address computing into the block when
935 /// possible / profitable.
936 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
937 bool MadeChange = false;
939 TargetLowering::AsmOperandInfoVector
940 TargetConstraints = TLI->ParseConstraints(CS);
942 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
943 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
945 // Compute the constraint code and ConstraintType to use.
946 TLI->ComputeConstraintToUse(OpInfo, SDValue());
948 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
950 Value *OpVal = CS->getArgOperand(ArgNo++);
951 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
952 } else if (OpInfo.Type == InlineAsm::isInput)
959 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
960 /// basic block as the load, unless conditions are unfavorable. This allows
961 /// SelectionDAG to fold the extend into the load.
963 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
964 // Look for a load being extended.
965 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
966 if (!LI) return false;
968 // If they're already in the same block, there's nothing to do.
969 if (LI->getParent() == I->getParent())
972 // If the load has other users and the truncate is not free, this probably
974 if (!LI->hasOneUse() &&
975 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
976 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
977 !TLI->isTruncateFree(I->getType(), LI->getType()))
980 // Check whether the target supports casts folded into loads.
982 if (isa<ZExtInst>(I))
983 LType = ISD::ZEXTLOAD;
985 assert(isa<SExtInst>(I) && "Unexpected ext type!");
986 LType = ISD::SEXTLOAD;
988 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
991 // Move the extend into the same block as the load, so that SelectionDAG
993 I->removeFromParent();
999 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
1000 BasicBlock *DefBB = I->getParent();
1002 // If the result of a {s|z}ext and its source are both live out, rewrite all
1003 // other uses of the source with result of extension.
1004 Value *Src = I->getOperand(0);
1005 if (Src->hasOneUse())
1008 // Only do this xform if truncating is free.
1009 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
1012 // Only safe to perform the optimization if the source is also defined in
1014 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
1017 bool DefIsLiveOut = false;
1018 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1020 Instruction *User = cast<Instruction>(*UI);
1022 // Figure out which BB this ext is used in.
1023 BasicBlock *UserBB = User->getParent();
1024 if (UserBB == DefBB) continue;
1025 DefIsLiveOut = true;
1031 // Make sure non of the uses are PHI nodes.
1032 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1034 Instruction *User = cast<Instruction>(*UI);
1035 BasicBlock *UserBB = User->getParent();
1036 if (UserBB == DefBB) continue;
1037 // Be conservative. We don't want this xform to end up introducing
1038 // reloads just before load / store instructions.
1039 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
1043 // InsertedTruncs - Only insert one trunc in each block once.
1044 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
1046 bool MadeChange = false;
1047 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1049 Use &TheUse = UI.getUse();
1050 Instruction *User = cast<Instruction>(*UI);
1052 // Figure out which BB this ext is used in.
1053 BasicBlock *UserBB = User->getParent();
1054 if (UserBB == DefBB) continue;
1056 // Both src and def are live in this block. Rewrite the use.
1057 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
1059 if (!InsertedTrunc) {
1060 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1061 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
1064 // Replace a use of the {s|z}ext source with a use of the result.
1065 TheUse = InsertedTrunc;
1073 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
1074 if (PHINode *P = dyn_cast<PHINode>(I)) {
1075 // It is possible for very late stage optimizations (such as SimplifyCFG)
1076 // to introduce PHI nodes too late to be cleaned up. If we detect such a
1077 // trivial PHI, go ahead and zap it here.
1078 if (Value *V = SimplifyInstruction(P)) {
1079 P->replaceAllUsesWith(V);
1080 P->eraseFromParent();
1087 if (CastInst *CI = dyn_cast<CastInst>(I)) {
1088 // If the source of the cast is a constant, then this should have
1089 // already been constant folded. The only reason NOT to constant fold
1090 // it is if something (e.g. LSR) was careful to place the constant
1091 // evaluation in a block other than then one that uses it (e.g. to hoist
1092 // the address of globals out of a loop). If this is the case, we don't
1093 // want to forward-subst the cast.
1094 if (isa<Constant>(CI->getOperand(0)))
1097 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
1100 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
1101 bool MadeChange = MoveExtToFormExtLoad(I);
1102 return MadeChange | OptimizeExtUses(I);
1107 if (CmpInst *CI = dyn_cast<CmpInst>(I))
1108 return OptimizeCmpExpression(CI);
1110 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1112 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
1116 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1118 return OptimizeMemoryInst(I, SI->getOperand(1),
1119 SI->getOperand(0)->getType());
1123 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1124 if (GEPI->hasAllZeroIndices()) {
1125 /// The GEP operand must be a pointer, so must its result -> BitCast
1126 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1127 GEPI->getName(), GEPI);
1128 GEPI->replaceAllUsesWith(NC);
1129 GEPI->eraseFromParent();
1137 if (CallInst *CI = dyn_cast<CallInst>(I))
1138 return OptimizeCallInst(CI);
1140 if (ReturnInst *RI = dyn_cast<ReturnInst>(I))
1141 return DupRetToEnableTailCallOpts(RI);
1146 // In this pass we look for GEP and cast instructions that are used
1147 // across basic blocks and rewrite them to improve basic-block-at-a-time
1149 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
1151 bool MadeChange = false;
1153 CurInstIterator = BB.begin();
1154 for (BasicBlock::iterator E = BB.end(); CurInstIterator != E; )
1155 MadeChange |= OptimizeInst(CurInstIterator++);