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");
62 class CodeGenPrepare : public FunctionPass {
63 /// TLI - Keep a pointer of a TargetLowering to consult for determining
64 /// transformation profitability.
65 const TargetLowering *TLI;
69 /// CurInstIterator - As we scan instructions optimizing them, this is the
70 /// next instruction to optimize. Xforms that can invalidate this should
72 BasicBlock::iterator CurInstIterator;
74 // Keeps track of non-local addresses that have been sunk into a block. This
75 // allows us to avoid inserting duplicate code for blocks with multiple
76 // load/stores of the same address.
77 DenseMap<Value*, Value*> SunkAddrs;
80 static char ID; // Pass identification, replacement for typeid
81 explicit CodeGenPrepare(const TargetLowering *tli = 0)
82 : FunctionPass(ID), TLI(tli) {
83 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
85 bool runOnFunction(Function &F);
87 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
88 AU.addPreserved<DominatorTree>();
89 AU.addPreserved<ProfileInfo>();
93 bool EliminateMostlyEmptyBlocks(Function &F);
94 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
95 void EliminateMostlyEmptyBlock(BasicBlock *BB);
96 bool OptimizeBlock(BasicBlock &BB);
97 bool OptimizeInst(Instruction *I);
98 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy);
99 bool OptimizeInlineAsmInst(CallInst *CS);
100 bool OptimizeCallInst(CallInst *CI);
101 bool MoveExtToFormExtLoad(Instruction *I);
102 bool OptimizeExtUses(Instruction *I);
106 char CodeGenPrepare::ID = 0;
107 INITIALIZE_PASS(CodeGenPrepare, "codegenprepare",
108 "Optimize for code generation", false, false)
110 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
111 return new CodeGenPrepare(TLI);
114 bool CodeGenPrepare::runOnFunction(Function &F) {
115 bool EverMadeChange = false;
117 DT = getAnalysisIfAvailable<DominatorTree>();
118 PFI = getAnalysisIfAvailable<ProfileInfo>();
119 // First pass, eliminate blocks that contain only PHI nodes and an
120 // unconditional branch.
121 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
123 bool MadeChange = true;
126 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
127 MadeChange |= OptimizeBlock(*BB);
128 EverMadeChange |= MadeChange;
133 return EverMadeChange;
136 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
137 /// debug info directives, and an unconditional branch. Passes before isel
138 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
139 /// isel. Start by eliminating these blocks so we can split them the way we
141 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
142 bool MadeChange = false;
143 // Note that this intentionally skips the entry block.
144 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
145 BasicBlock *BB = I++;
147 // If this block doesn't end with an uncond branch, ignore it.
148 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
149 if (!BI || !BI->isUnconditional())
152 // If the instruction before the branch (skipping debug info) isn't a phi
153 // node, then other stuff is happening here.
154 BasicBlock::iterator BBI = BI;
155 if (BBI != BB->begin()) {
157 while (isa<DbgInfoIntrinsic>(BBI)) {
158 if (BBI == BB->begin())
162 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
166 // Do not break infinite loops.
167 BasicBlock *DestBB = BI->getSuccessor(0);
171 if (!CanMergeBlocks(BB, DestBB))
174 EliminateMostlyEmptyBlock(BB);
180 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
181 /// single uncond branch between them, and BB contains no other non-phi
183 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
184 const BasicBlock *DestBB) const {
185 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
186 // the successor. If there are more complex condition (e.g. preheaders),
187 // don't mess around with them.
188 BasicBlock::const_iterator BBI = BB->begin();
189 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
190 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
192 const Instruction *User = cast<Instruction>(*UI);
193 if (User->getParent() != DestBB || !isa<PHINode>(User))
195 // If User is inside DestBB block and it is a PHINode then check
196 // incoming value. If incoming value is not from BB then this is
197 // a complex condition (e.g. preheaders) we want to avoid here.
198 if (User->getParent() == DestBB) {
199 if (const PHINode *UPN = dyn_cast<PHINode>(User))
200 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
201 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
202 if (Insn && Insn->getParent() == BB &&
203 Insn->getParent() != UPN->getIncomingBlock(I))
210 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
211 // and DestBB may have conflicting incoming values for the block. If so, we
212 // can't merge the block.
213 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
214 if (!DestBBPN) return true; // no conflict.
216 // Collect the preds of BB.
217 SmallPtrSet<const BasicBlock*, 16> BBPreds;
218 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
219 // It is faster to get preds from a PHI than with pred_iterator.
220 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
221 BBPreds.insert(BBPN->getIncomingBlock(i));
223 BBPreds.insert(pred_begin(BB), pred_end(BB));
226 // Walk the preds of DestBB.
227 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
228 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
229 if (BBPreds.count(Pred)) { // Common predecessor?
230 BBI = DestBB->begin();
231 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
232 const Value *V1 = PN->getIncomingValueForBlock(Pred);
233 const Value *V2 = PN->getIncomingValueForBlock(BB);
235 // If V2 is a phi node in BB, look up what the mapped value will be.
236 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
237 if (V2PN->getParent() == BB)
238 V2 = V2PN->getIncomingValueForBlock(Pred);
240 // If there is a conflict, bail out.
241 if (V1 != V2) return false;
250 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
251 /// an unconditional branch in it.
252 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
253 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
254 BasicBlock *DestBB = BI->getSuccessor(0);
256 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
258 // If the destination block has a single pred, then this is a trivial edge,
260 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
261 if (SinglePred != DestBB) {
262 // Remember if SinglePred was the entry block of the function. If so, we
263 // will need to move BB back to the entry position.
264 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
265 MergeBasicBlockIntoOnlyPred(DestBB, this);
267 if (isEntry && BB != &BB->getParent()->getEntryBlock())
268 BB->moveBefore(&BB->getParent()->getEntryBlock());
270 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
275 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
276 // to handle the new incoming edges it is about to have.
278 for (BasicBlock::iterator BBI = DestBB->begin();
279 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
280 // Remove the incoming value for BB, and remember it.
281 Value *InVal = PN->removeIncomingValue(BB, false);
283 // Two options: either the InVal is a phi node defined in BB or it is some
284 // value that dominates BB.
285 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
286 if (InValPhi && InValPhi->getParent() == BB) {
287 // Add all of the input values of the input PHI as inputs of this phi.
288 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
289 PN->addIncoming(InValPhi->getIncomingValue(i),
290 InValPhi->getIncomingBlock(i));
292 // Otherwise, add one instance of the dominating value for each edge that
293 // we will be adding.
294 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
295 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
296 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
298 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
299 PN->addIncoming(InVal, *PI);
304 // The PHIs are now updated, change everything that refers to BB to use
305 // DestBB and remove BB.
306 BB->replaceAllUsesWith(DestBB);
308 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
309 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
310 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
311 DT->changeImmediateDominator(DestBB, NewIDom);
315 PFI->replaceAllUses(BB, DestBB);
316 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
318 BB->eraseFromParent();
321 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
324 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
325 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
326 /// sink it into user blocks to reduce the number of virtual
327 /// registers that must be created and coalesced.
329 /// Return true if any changes are made.
331 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
332 // If this is a noop copy,
333 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
334 EVT DstVT = TLI.getValueType(CI->getType());
336 // This is an fp<->int conversion?
337 if (SrcVT.isInteger() != DstVT.isInteger())
340 // If this is an extension, it will be a zero or sign extension, which
342 if (SrcVT.bitsLT(DstVT)) return false;
344 // If these values will be promoted, find out what they will be promoted
345 // to. This helps us consider truncates on PPC as noop copies when they
347 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote)
348 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
349 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote)
350 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
352 // If, after promotion, these are the same types, this is a noop copy.
356 BasicBlock *DefBB = CI->getParent();
358 /// InsertedCasts - Only insert a cast in each block once.
359 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
361 bool MadeChange = false;
362 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
364 Use &TheUse = UI.getUse();
365 Instruction *User = cast<Instruction>(*UI);
367 // Figure out which BB this cast is used in. For PHI's this is the
368 // appropriate predecessor block.
369 BasicBlock *UserBB = User->getParent();
370 if (PHINode *PN = dyn_cast<PHINode>(User)) {
371 UserBB = PN->getIncomingBlock(UI);
374 // Preincrement use iterator so we don't invalidate it.
377 // If this user is in the same block as the cast, don't change the cast.
378 if (UserBB == DefBB) continue;
380 // If we have already inserted a cast into this block, use it.
381 CastInst *&InsertedCast = InsertedCasts[UserBB];
384 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
387 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
392 // Replace a use of the cast with a use of the new cast.
393 TheUse = InsertedCast;
397 // If we removed all uses, nuke the cast.
398 if (CI->use_empty()) {
399 CI->eraseFromParent();
406 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
407 /// the number of virtual registers that must be created and coalesced. This is
408 /// a clear win except on targets with multiple condition code registers
409 /// (PowerPC), where it might lose; some adjustment may be wanted there.
411 /// Return true if any changes are made.
412 static bool OptimizeCmpExpression(CmpInst *CI) {
413 BasicBlock *DefBB = CI->getParent();
415 /// InsertedCmp - Only insert a cmp in each block once.
416 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
418 bool MadeChange = false;
419 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
421 Use &TheUse = UI.getUse();
422 Instruction *User = cast<Instruction>(*UI);
424 // Preincrement use iterator so we don't invalidate it.
427 // Don't bother for PHI nodes.
428 if (isa<PHINode>(User))
431 // Figure out which BB this cmp is used in.
432 BasicBlock *UserBB = User->getParent();
434 // If this user is in the same block as the cmp, don't change the cmp.
435 if (UserBB == DefBB) continue;
437 // If we have already inserted a cmp into this block, use it.
438 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
441 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
444 CmpInst::Create(CI->getOpcode(),
445 CI->getPredicate(), CI->getOperand(0),
446 CI->getOperand(1), "", InsertPt);
450 // Replace a use of the cmp with a use of the new cmp.
451 TheUse = InsertedCmp;
455 // If we removed all uses, nuke the cmp.
457 CI->eraseFromParent();
463 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
465 void replaceCall(Value *With) {
466 CI->replaceAllUsesWith(With);
467 CI->eraseFromParent();
469 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
470 if (ConstantInt *SizeCI =
471 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
472 return SizeCI->isAllOnesValue();
476 } // end anonymous namespace
478 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
479 BasicBlock *BB = CI->getParent();
481 // Lower inline assembly if we can.
482 // If we found an inline asm expession, and if the target knows how to
483 // lower it to normal LLVM code, do so now.
484 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
485 if (TLI->ExpandInlineAsm(CI)) {
486 // Avoid invalidating the iterator.
487 CurInstIterator = BB->begin();
488 // Avoid processing instructions out of order, which could cause
489 // reuse before a value is defined.
493 // Sink address computing for memory operands into the block.
494 if (OptimizeInlineAsmInst(CI))
498 // Lower all uses of llvm.objectsize.*
499 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
500 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
501 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
502 const Type *ReturnTy = CI->getType();
503 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
505 // Substituting this can cause recursive simplifications, which can
506 // invalidate our iterator. Use a WeakVH to hold onto it in case this
508 WeakVH IterHandle(CurInstIterator);
510 ReplaceAndSimplifyAllUses(CI, RetVal, TLI ? TLI->getTargetData() : 0, DT);
512 // If the iterator instruction was recursively deleted, start over at the
513 // start of the block.
514 if (IterHandle != CurInstIterator) {
515 CurInstIterator = BB->begin();
521 // From here on out we're working with named functions.
522 if (CI->getCalledFunction() == 0) return false;
524 // We'll need TargetData from here on out.
525 const TargetData *TD = TLI ? TLI->getTargetData() : 0;
526 if (!TD) return false;
528 // Lower all default uses of _chk calls. This is very similar
529 // to what InstCombineCalls does, but here we are only lowering calls
530 // that have the default "don't know" as the objectsize. Anything else
531 // should be left alone.
532 CodeGenPrepareFortifiedLibCalls Simplifier;
533 return Simplifier.fold(CI, TD);
536 //===----------------------------------------------------------------------===//
537 // Memory Optimization
538 //===----------------------------------------------------------------------===//
540 /// IsNonLocalValue - Return true if the specified values are defined in a
541 /// different basic block than BB.
542 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
543 if (Instruction *I = dyn_cast<Instruction>(V))
544 return I->getParent() != BB;
548 /// OptimizeMemoryInst - Load and Store Instructions often have
549 /// addressing modes that can do significant amounts of computation. As such,
550 /// instruction selection will try to get the load or store to do as much
551 /// computation as possible for the program. The problem is that isel can only
552 /// see within a single block. As such, we sink as much legal addressing mode
553 /// stuff into the block as possible.
555 /// This method is used to optimize both load/store and inline asms with memory
557 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
558 const Type *AccessTy) {
561 // Try to collapse single-value PHI nodes. This is necessary to undo
562 // unprofitable PRE transformations.
563 SmallVector<Value*, 8> worklist;
564 SmallPtrSet<Value*, 16> Visited;
565 worklist.push_back(Addr);
567 // Use a worklist to iteratively look through PHI nodes, and ensure that
568 // the addressing mode obtained from the non-PHI roots of the graph
570 Value *Consensus = 0;
571 unsigned NumUsesConsensus = 0;
572 bool IsNumUsesConsensusValid = false;
573 SmallVector<Instruction*, 16> AddrModeInsts;
574 ExtAddrMode AddrMode;
575 while (!worklist.empty()) {
576 Value *V = worklist.back();
579 // Break use-def graph loops.
580 if (Visited.count(V)) {
587 // For a PHI node, push all of its incoming values.
588 if (PHINode *P = dyn_cast<PHINode>(V)) {
589 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
590 worklist.push_back(P->getIncomingValue(i));
594 // For non-PHIs, determine the addressing mode being computed.
595 SmallVector<Instruction*, 16> NewAddrModeInsts;
596 ExtAddrMode NewAddrMode =
597 AddressingModeMatcher::Match(V, AccessTy,MemoryInst,
598 NewAddrModeInsts, *TLI);
600 // This check is broken into two cases with very similar code to avoid using
601 // getNumUses() as much as possible. Some values have a lot of uses, so
602 // calling getNumUses() unconditionally caused a significant compile-time
606 AddrMode = NewAddrMode;
607 AddrModeInsts = NewAddrModeInsts;
609 } else if (NewAddrMode == AddrMode) {
610 if (!IsNumUsesConsensusValid) {
611 NumUsesConsensus = Consensus->getNumUses();
612 IsNumUsesConsensusValid = true;
615 // Ensure that the obtained addressing mode is equivalent to that obtained
616 // for all other roots of the PHI traversal. Also, when choosing one
617 // such root as representative, select the one with the most uses in order
618 // to keep the cost modeling heuristics in AddressingModeMatcher
620 unsigned NumUses = V->getNumUses();
621 if (NumUses > NumUsesConsensus) {
623 NumUsesConsensus = NumUses;
624 AddrModeInsts = NewAddrModeInsts;
633 // If the addressing mode couldn't be determined, or if multiple different
634 // ones were determined, bail out now.
635 if (!Consensus) return false;
637 // Check to see if any of the instructions supersumed by this addr mode are
638 // non-local to I's BB.
639 bool AnyNonLocal = false;
640 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
641 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
647 // If all the instructions matched are already in this BB, don't do anything.
649 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
653 // Insert this computation right after this user. Since our caller is
654 // scanning from the top of the BB to the bottom, reuse of the expr are
655 // guaranteed to happen later.
656 BasicBlock::iterator InsertPt = MemoryInst;
658 // Now that we determined the addressing expression we want to use and know
659 // that we have to sink it into this block. Check to see if we have already
660 // done this for some other load/store instr in this block. If so, reuse the
662 Value *&SunkAddr = SunkAddrs[Addr];
664 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
666 if (SunkAddr->getType() != Addr->getType())
667 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
669 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
671 const Type *IntPtrTy =
672 TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
676 // Start with the base register. Do this first so that subsequent address
677 // matching finds it last, which will prevent it from trying to match it
678 // as the scaled value in case it happens to be a mul. That would be
679 // problematic if we've sunk a different mul for the scale, because then
680 // we'd end up sinking both muls.
681 if (AddrMode.BaseReg) {
682 Value *V = AddrMode.BaseReg;
683 if (V->getType()->isPointerTy())
684 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
685 if (V->getType() != IntPtrTy)
686 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true,
687 "sunkaddr", InsertPt);
691 // Add the scale value.
692 if (AddrMode.Scale) {
693 Value *V = AddrMode.ScaledReg;
694 if (V->getType() == IntPtrTy) {
696 } else if (V->getType()->isPointerTy()) {
697 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
698 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
699 cast<IntegerType>(V->getType())->getBitWidth()) {
700 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
702 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
704 if (AddrMode.Scale != 1)
705 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
707 "sunkaddr", InsertPt);
709 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
714 // Add in the BaseGV if present.
715 if (AddrMode.BaseGV) {
716 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
719 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
724 // Add in the Base Offset if present.
725 if (AddrMode.BaseOffs) {
726 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
728 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
734 SunkAddr = Constant::getNullValue(Addr->getType());
736 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
739 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
741 if (Repl->use_empty()) {
742 RecursivelyDeleteTriviallyDeadInstructions(Repl);
743 // This address is now available for reassignment, so erase the table entry;
744 // we don't want to match some completely different instruction.
751 /// OptimizeInlineAsmInst - If there are any memory operands, use
752 /// OptimizeMemoryInst to sink their address computing into the block when
753 /// possible / profitable.
754 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
755 bool MadeChange = false;
757 TargetLowering::AsmOperandInfoVector
758 TargetConstraints = TLI->ParseConstraints(CS);
760 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
761 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
763 // Compute the constraint code and ConstraintType to use.
764 TLI->ComputeConstraintToUse(OpInfo, SDValue());
766 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
768 Value *OpVal = CS->getArgOperand(ArgNo++);
769 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
770 } else if (OpInfo.Type == InlineAsm::isInput)
777 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
778 /// basic block as the load, unless conditions are unfavorable. This allows
779 /// SelectionDAG to fold the extend into the load.
781 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
782 // Look for a load being extended.
783 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
784 if (!LI) return false;
786 // If they're already in the same block, there's nothing to do.
787 if (LI->getParent() == I->getParent())
790 // If the load has other users and the truncate is not free, this probably
792 if (!LI->hasOneUse() &&
793 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
794 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
795 !TLI->isTruncateFree(I->getType(), LI->getType()))
798 // Check whether the target supports casts folded into loads.
800 if (isa<ZExtInst>(I))
801 LType = ISD::ZEXTLOAD;
803 assert(isa<SExtInst>(I) && "Unexpected ext type!");
804 LType = ISD::SEXTLOAD;
806 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
809 // Move the extend into the same block as the load, so that SelectionDAG
811 I->removeFromParent();
817 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
818 BasicBlock *DefBB = I->getParent();
820 // If the result of a {s|z}ext and its source are both live out, rewrite all
821 // other uses of the source with result of extension.
822 Value *Src = I->getOperand(0);
823 if (Src->hasOneUse())
826 // Only do this xform if truncating is free.
827 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
830 // Only safe to perform the optimization if the source is also defined in
832 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
835 bool DefIsLiveOut = false;
836 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
838 Instruction *User = cast<Instruction>(*UI);
840 // Figure out which BB this ext is used in.
841 BasicBlock *UserBB = User->getParent();
842 if (UserBB == DefBB) continue;
849 // Make sure non of the uses are PHI nodes.
850 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
852 Instruction *User = cast<Instruction>(*UI);
853 BasicBlock *UserBB = User->getParent();
854 if (UserBB == DefBB) continue;
855 // Be conservative. We don't want this xform to end up introducing
856 // reloads just before load / store instructions.
857 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
861 // InsertedTruncs - Only insert one trunc in each block once.
862 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
864 bool MadeChange = false;
865 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
867 Use &TheUse = UI.getUse();
868 Instruction *User = cast<Instruction>(*UI);
870 // Figure out which BB this ext is used in.
871 BasicBlock *UserBB = User->getParent();
872 if (UserBB == DefBB) continue;
874 // Both src and def are live in this block. Rewrite the use.
875 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
877 if (!InsertedTrunc) {
878 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
880 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
883 // Replace a use of the {s|z}ext source with a use of the result.
884 TheUse = InsertedTrunc;
892 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
893 if (PHINode *P = dyn_cast<PHINode>(I)) {
894 // It is possible for very late stage optimizations (such as SimplifyCFG)
895 // to introduce PHI nodes too late to be cleaned up. If we detect such a
896 // trivial PHI, go ahead and zap it here.
897 if (Value *V = SimplifyInstruction(P)) {
898 P->replaceAllUsesWith(V);
899 P->eraseFromParent();
906 if (CastInst *CI = dyn_cast<CastInst>(I)) {
907 // If the source of the cast is a constant, then this should have
908 // already been constant folded. The only reason NOT to constant fold
909 // it is if something (e.g. LSR) was careful to place the constant
910 // evaluation in a block other than then one that uses it (e.g. to hoist
911 // the address of globals out of a loop). If this is the case, we don't
912 // want to forward-subst the cast.
913 if (isa<Constant>(CI->getOperand(0)))
916 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
919 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
920 bool MadeChange = MoveExtToFormExtLoad(I);
921 return MadeChange | OptimizeExtUses(I);
926 if (CmpInst *CI = dyn_cast<CmpInst>(I))
927 return OptimizeCmpExpression(CI);
929 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
931 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
935 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
937 return OptimizeMemoryInst(I, SI->getOperand(1),
938 SI->getOperand(0)->getType());
942 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
943 if (GEPI->hasAllZeroIndices()) {
944 /// The GEP operand must be a pointer, so must its result -> BitCast
945 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
946 GEPI->getName(), GEPI);
947 GEPI->replaceAllUsesWith(NC);
948 GEPI->eraseFromParent();
956 if (CallInst *CI = dyn_cast<CallInst>(I))
957 return OptimizeCallInst(CI);
962 // In this pass we look for GEP and cast instructions that are used
963 // across basic blocks and rewrite them to improve basic-block-at-a-time
965 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
967 bool MadeChange = false;
969 CurInstIterator = BB.begin();
970 for (BasicBlock::iterator E = BB.end(); CurInstIterator != E; )
971 MadeChange |= OptimizeInst(CurInstIterator++);