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/Pass.h"
24 #include "llvm/Target/TargetAsmInfo.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/Target/TargetLowering.h"
27 #include "llvm/Target/TargetMachine.h"
28 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
29 #include "llvm/Transforms/Utils/Local.h"
30 #include "llvm/ADT/DenseMap.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/Support/CallSite.h"
33 #include "llvm/Support/Compiler.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/GetElementPtrTypeIterator.h"
36 #include "llvm/Support/PatternMatch.h"
38 using namespace llvm::PatternMatch;
41 class VISIBILITY_HIDDEN CodeGenPrepare : public FunctionPass {
42 /// TLI - Keep a pointer of a TargetLowering to consult for determining
43 /// transformation profitability.
44 const TargetLowering *TLI;
46 static char ID; // Pass identification, replacement for typeid
47 explicit CodeGenPrepare(const TargetLowering *tli = 0)
48 : FunctionPass(&ID), TLI(tli) {}
49 bool runOnFunction(Function &F);
52 bool EliminateMostlyEmptyBlocks(Function &F);
53 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
54 void EliminateMostlyEmptyBlock(BasicBlock *BB);
55 bool OptimizeBlock(BasicBlock &BB);
56 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy,
57 DenseMap<Value*,Value*> &SunkAddrs);
58 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS,
59 DenseMap<Value*,Value*> &SunkAddrs);
60 bool OptimizeExtUses(Instruction *I);
64 char CodeGenPrepare::ID = 0;
65 static RegisterPass<CodeGenPrepare> X("codegenprepare",
66 "Optimize for code generation");
68 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
69 return new CodeGenPrepare(TLI);
73 bool CodeGenPrepare::runOnFunction(Function &F) {
74 bool EverMadeChange = false;
76 // First pass, eliminate blocks that contain only PHI nodes and an
77 // unconditional branch.
78 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
80 bool MadeChange = true;
83 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
84 MadeChange |= OptimizeBlock(*BB);
85 EverMadeChange |= MadeChange;
87 return EverMadeChange;
90 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes
91 /// and an unconditional branch. Passes before isel (e.g. LSR/loopsimplify)
92 /// often split edges in ways that are non-optimal for isel. Start by
93 /// eliminating these blocks so we can split them the way we want them.
94 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
95 bool MadeChange = false;
96 // Note that this intentionally skips the entry block.
97 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
100 // If this block doesn't end with an uncond branch, ignore it.
101 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
102 if (!BI || !BI->isUnconditional())
105 // If the instruction before the branch isn't a phi node, then other stuff
106 // is happening here.
107 BasicBlock::iterator BBI = BI;
108 if (BBI != BB->begin()) {
110 if (!isa<PHINode>(BBI)) continue;
113 // Do not break infinite loops.
114 BasicBlock *DestBB = BI->getSuccessor(0);
118 if (!CanMergeBlocks(BB, DestBB))
121 EliminateMostlyEmptyBlock(BB);
127 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
128 /// single uncond branch between them, and BB contains no other non-phi
130 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
131 const BasicBlock *DestBB) const {
132 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
133 // the successor. If there are more complex condition (e.g. preheaders),
134 // don't mess around with them.
135 BasicBlock::const_iterator BBI = BB->begin();
136 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
137 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end();
139 const Instruction *User = cast<Instruction>(*UI);
140 if (User->getParent() != DestBB || !isa<PHINode>(User))
142 // If User is inside DestBB block and it is a PHINode then check
143 // incoming value. If incoming value is not from BB then this is
144 // a complex condition (e.g. preheaders) we want to avoid here.
145 if (User->getParent() == DestBB) {
146 if (const PHINode *UPN = dyn_cast<PHINode>(User))
147 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
148 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
149 if (Insn && Insn->getParent() == BB &&
150 Insn->getParent() != UPN->getIncomingBlock(I))
157 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
158 // and DestBB may have conflicting incoming values for the block. If so, we
159 // can't merge the block.
160 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
161 if (!DestBBPN) return true; // no conflict.
163 // Collect the preds of BB.
164 SmallPtrSet<const BasicBlock*, 16> BBPreds;
165 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
166 // It is faster to get preds from a PHI than with pred_iterator.
167 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
168 BBPreds.insert(BBPN->getIncomingBlock(i));
170 BBPreds.insert(pred_begin(BB), pred_end(BB));
173 // Walk the preds of DestBB.
174 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
175 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
176 if (BBPreds.count(Pred)) { // Common predecessor?
177 BBI = DestBB->begin();
178 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
179 const Value *V1 = PN->getIncomingValueForBlock(Pred);
180 const Value *V2 = PN->getIncomingValueForBlock(BB);
182 // If V2 is a phi node in BB, look up what the mapped value will be.
183 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
184 if (V2PN->getParent() == BB)
185 V2 = V2PN->getIncomingValueForBlock(Pred);
187 // If there is a conflict, bail out.
188 if (V1 != V2) return false;
197 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
198 /// an unconditional branch in it.
199 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
200 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
201 BasicBlock *DestBB = BI->getSuccessor(0);
203 DOUT << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB;
205 // If the destination block has a single pred, then this is a trivial edge,
207 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
208 if (SinglePred != DestBB) {
209 // Remember if SinglePred was the entry block of the function. If so, we
210 // will need to move BB back to the entry position.
211 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
212 MergeBasicBlockIntoOnlyPred(DestBB);
214 if (isEntry && BB != &BB->getParent()->getEntryBlock())
215 BB->moveBefore(&BB->getParent()->getEntryBlock());
217 DOUT << "AFTER:\n" << *DestBB << "\n\n\n";
222 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
223 // to handle the new incoming edges it is about to have.
225 for (BasicBlock::iterator BBI = DestBB->begin();
226 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
227 // Remove the incoming value for BB, and remember it.
228 Value *InVal = PN->removeIncomingValue(BB, false);
230 // Two options: either the InVal is a phi node defined in BB or it is some
231 // value that dominates BB.
232 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
233 if (InValPhi && InValPhi->getParent() == BB) {
234 // Add all of the input values of the input PHI as inputs of this phi.
235 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
236 PN->addIncoming(InValPhi->getIncomingValue(i),
237 InValPhi->getIncomingBlock(i));
239 // Otherwise, add one instance of the dominating value for each edge that
240 // we will be adding.
241 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
242 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
243 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
245 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
246 PN->addIncoming(InVal, *PI);
251 // The PHIs are now updated, change everything that refers to BB to use
252 // DestBB and remove BB.
253 BB->replaceAllUsesWith(DestBB);
254 BB->eraseFromParent();
256 DOUT << "AFTER:\n" << *DestBB << "\n\n\n";
260 /// SplitEdgeNicely - Split the critical edge from TI to its specified
261 /// successor if it will improve codegen. We only do this if the successor has
262 /// phi nodes (otherwise critical edges are ok). If there is already another
263 /// predecessor of the succ that is empty (and thus has no phi nodes), use it
264 /// instead of introducing a new block.
265 static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, Pass *P) {
266 BasicBlock *TIBB = TI->getParent();
267 BasicBlock *Dest = TI->getSuccessor(SuccNum);
268 assert(isa<PHINode>(Dest->begin()) &&
269 "This should only be called if Dest has a PHI!");
271 // As a hack, never split backedges of loops. Even though the copy for any
272 // PHIs inserted on the backedge would be dead for exits from the loop, we
273 // assume that the cost of *splitting* the backedge would be too high.
277 /// TIPHIValues - This array is lazily computed to determine the values of
278 /// PHIs in Dest that TI would provide.
279 SmallVector<Value*, 32> TIPHIValues;
281 // Check to see if Dest has any blocks that can be used as a split edge for
283 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) {
284 BasicBlock *Pred = *PI;
285 // To be usable, the pred has to end with an uncond branch to the dest.
286 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator());
287 if (!PredBr || !PredBr->isUnconditional() ||
288 // Must be empty other than the branch.
289 &Pred->front() != PredBr ||
290 // Cannot be the entry block; its label does not get emitted.
291 Pred == &(Dest->getParent()->getEntryBlock()))
294 // Finally, since we know that Dest has phi nodes in it, we have to make
295 // sure that jumping to Pred will have the same affect as going to Dest in
296 // terms of PHI values.
299 bool FoundMatch = true;
300 for (BasicBlock::iterator I = Dest->begin();
301 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) {
302 if (PHINo == TIPHIValues.size())
303 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB));
305 // If the PHI entry doesn't work, we can't use this pred.
306 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) {
312 // If we found a workable predecessor, change TI to branch to Succ.
314 Dest->removePredecessor(TIBB);
315 TI->setSuccessor(SuccNum, Pred);
320 SplitCriticalEdge(TI, SuccNum, P, true);
323 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
324 /// copy (e.g. it's casting from one pointer type to another, int->uint, or
325 /// int->sbyte on PPC), sink it into user blocks to reduce the number of virtual
326 /// registers that must be created and coalesced.
328 /// Return true if any changes are made.
330 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
331 // If this is a noop copy,
332 MVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
333 MVT DstVT = TLI.getValueType(CI->getType());
335 // This is an fp<->int conversion?
336 if (SrcVT.isInteger() != DstVT.isInteger())
339 // If this is an extension, it will be a zero or sign extension, which
341 if (SrcVT.bitsLT(DstVT)) return false;
343 // If these values will be promoted, find out what they will be promoted
344 // to. This helps us consider truncates on PPC as noop copies when they
346 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote)
347 SrcVT = TLI.getTypeToTransformTo(SrcVT);
348 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote)
349 DstVT = TLI.getTypeToTransformTo(DstVT);
351 // If, after promotion, these are the same types, this is a noop copy.
355 BasicBlock *DefBB = CI->getParent();
357 /// InsertedCasts - Only insert a cast in each block once.
358 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
360 bool MadeChange = false;
361 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
363 Use &TheUse = UI.getUse();
364 Instruction *User = cast<Instruction>(*UI);
366 // Figure out which BB this cast is used in. For PHI's this is the
367 // appropriate predecessor block.
368 BasicBlock *UserBB = User->getParent();
369 if (PHINode *PN = dyn_cast<PHINode>(User)) {
370 unsigned OpVal = UI.getOperandNo()/2;
371 UserBB = PN->getIncomingBlock(OpVal);
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;
396 // If we removed all uses, nuke the cast.
397 if (CI->use_empty()) {
398 CI->eraseFromParent();
405 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
406 /// the number of virtual registers that must be created and coalesced. This is
407 /// a clear win except on targets with multiple condition code registers
408 /// (PowerPC), where it might lose; some adjustment may be wanted there.
410 /// Return true if any changes are made.
411 static bool OptimizeCmpExpression(CmpInst *CI) {
412 BasicBlock *DefBB = CI->getParent();
414 /// InsertedCmp - Only insert a cmp in each block once.
415 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
417 bool MadeChange = false;
418 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
420 Use &TheUse = UI.getUse();
421 Instruction *User = cast<Instruction>(*UI);
423 // Preincrement use iterator so we don't invalidate it.
426 // Don't bother for PHI nodes.
427 if (isa<PHINode>(User))
430 // Figure out which BB this cmp is used in.
431 BasicBlock *UserBB = User->getParent();
433 // If this user is in the same block as the cmp, don't change the cmp.
434 if (UserBB == DefBB) continue;
436 // If we have already inserted a cmp into this block, use it.
437 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
440 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
443 CmpInst::Create(CI->getOpcode(), CI->getPredicate(), CI->getOperand(0),
444 CI->getOperand(1), "", InsertPt);
448 // Replace a use of the cmp with a use of the new cmp.
449 TheUse = InsertedCmp;
452 // If we removed all uses, nuke the cmp.
454 CI->eraseFromParent();
459 //===----------------------------------------------------------------------===//
460 // Addressing Mode Analysis and Optimization
461 //===----------------------------------------------------------------------===//
464 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
465 /// which holds actual Value*'s for register values.
466 struct ExtAddrMode : public TargetLowering::AddrMode {
469 ExtAddrMode() : BaseReg(0), ScaledReg(0) {}
470 void print(OStream &OS) const;
476 } // end anonymous namespace
478 static inline OStream &operator<<(OStream &OS, const ExtAddrMode &AM) {
483 void ExtAddrMode::print(OStream &OS) const {
484 bool NeedPlus = false;
487 OS << (NeedPlus ? " + " : "")
488 << "GV:%" << BaseGV->getName(), NeedPlus = true;
491 OS << (NeedPlus ? " + " : "") << BaseOffs, NeedPlus = true;
494 OS << (NeedPlus ? " + " : "")
495 << "Base:%" << BaseReg->getName(), NeedPlus = true;
497 OS << (NeedPlus ? " + " : "")
498 << Scale << "*%" << ScaledReg->getName(), NeedPlus = true;
504 /// AddressingModeMatcher - This class exposes a single public method, which is
505 /// used to construct a "maximal munch" of the addressing mode for the target
506 /// specified by TLI for an access to "V" with an access type of AccessTy. This
507 /// returns the addressing mode that is actually matched by value, but also
508 /// returns the list of instructions involved in that addressing computation in
510 class AddressingModeMatcher {
511 SmallVectorImpl<Instruction*> &AddrModeInsts;
512 const TargetLowering &TLI;
514 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
515 /// the memory instruction that we're computing this address for.
516 const Type *AccessTy;
517 Instruction *MemoryInst;
519 /// AddrMode - This is the addressing mode that we're building up. This is
520 /// part of the return value of this addressing mode matching stuff.
521 ExtAddrMode &AddrMode;
523 /// IgnoreProfitability - This is set to true when we should not do
524 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
525 /// always returns true.
526 bool IgnoreProfitability;
528 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
529 const TargetLowering &T, const Type *AT,
530 Instruction *MI, ExtAddrMode &AM)
531 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM) {
532 IgnoreProfitability = false;
536 /// Match - Find the maximal addressing mode that a load/store of V can fold,
537 /// give an access type of AccessTy. This returns a list of involved
538 /// instructions in AddrModeInsts.
539 static ExtAddrMode Match(Value *V, const Type *AccessTy,
540 Instruction *MemoryInst,
541 SmallVectorImpl<Instruction*> &AddrModeInsts,
542 const TargetLowering &TLI) {
546 AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
547 MemoryInst, Result).MatchAddr(V, 0);
548 Success = Success; assert(Success && "Couldn't select *anything*?");
552 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
553 bool MatchAddr(Value *V, unsigned Depth);
554 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth);
555 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
556 ExtAddrMode &AMBefore,
557 ExtAddrMode &AMAfter);
558 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
560 } // end anonymous namespace
562 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
563 /// Return true and update AddrMode if this addr mode is legal for the target,
565 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
567 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
568 // mode. Just process that directly.
570 return MatchAddr(ScaleReg, Depth);
572 // If the scale is 0, it takes nothing to add this.
576 // If we already have a scale of this value, we can add to it, otherwise, we
577 // need an available scale field.
578 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
581 ExtAddrMode TestAddrMode = AddrMode;
583 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
584 // [A+B + A*7] -> [B+A*8].
585 TestAddrMode.Scale += Scale;
586 TestAddrMode.ScaledReg = ScaleReg;
588 // If the new address isn't legal, bail out.
589 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
592 // It was legal, so commit it.
593 AddrMode = TestAddrMode;
595 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
596 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
597 // X*Scale + C*Scale to addr mode.
598 ConstantInt *CI; Value *AddLHS;
599 if (match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
600 TestAddrMode.ScaledReg = AddLHS;
601 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
603 // If this addressing mode is legal, commit it and remember that we folded
605 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
606 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
607 AddrMode = TestAddrMode;
612 // Otherwise, not (x+c)*scale, just return what we have.
616 /// MightBeFoldableInst - This is a little filter, which returns true if an
617 /// addressing computation involving I might be folded into a load/store
618 /// accessing it. This doesn't need to be perfect, but needs to accept at least
619 /// the set of instructions that MatchOperationAddr can.
620 static bool MightBeFoldableInst(Instruction *I) {
621 switch (I->getOpcode()) {
622 case Instruction::BitCast:
623 // Don't touch identity bitcasts.
624 if (I->getType() == I->getOperand(0)->getType())
626 return isa<PointerType>(I->getType()) || isa<IntegerType>(I->getType());
627 case Instruction::PtrToInt:
628 // PtrToInt is always a noop, as we know that the int type is pointer sized.
630 case Instruction::IntToPtr:
631 // We know the input is intptr_t, so this is foldable.
633 case Instruction::Add:
635 case Instruction::Mul:
636 case Instruction::Shl:
637 // Can only handle X*C and X << C.
638 return isa<ConstantInt>(I->getOperand(1));
639 case Instruction::GetElementPtr:
647 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
648 /// fold the operation into the addressing mode. If so, update the addressing
649 /// mode and return true, otherwise return false without modifying AddrMode.
650 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
652 // Avoid exponential behavior on extremely deep expression trees.
653 if (Depth >= 5) return false;
656 case Instruction::PtrToInt:
657 // PtrToInt is always a noop, as we know that the int type is pointer sized.
658 return MatchAddr(AddrInst->getOperand(0), Depth);
659 case Instruction::IntToPtr:
660 // This inttoptr is a no-op if the integer type is pointer sized.
661 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
663 return MatchAddr(AddrInst->getOperand(0), Depth);
665 case Instruction::BitCast:
666 // BitCast is always a noop, and we can handle it as long as it is
667 // int->int or pointer->pointer (we don't want int<->fp or something).
668 if ((isa<PointerType>(AddrInst->getOperand(0)->getType()) ||
669 isa<IntegerType>(AddrInst->getOperand(0)->getType())) &&
670 // Don't touch identity bitcasts. These were probably put here by LSR,
671 // and we don't want to mess around with them. Assume it knows what it
673 AddrInst->getOperand(0)->getType() != AddrInst->getType())
674 return MatchAddr(AddrInst->getOperand(0), Depth);
676 case Instruction::Add: {
677 // Check to see if we can merge in the RHS then the LHS. If so, we win.
678 ExtAddrMode BackupAddrMode = AddrMode;
679 unsigned OldSize = AddrModeInsts.size();
680 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
681 MatchAddr(AddrInst->getOperand(0), Depth+1))
684 // Restore the old addr mode info.
685 AddrMode = BackupAddrMode;
686 AddrModeInsts.resize(OldSize);
688 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
689 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
690 MatchAddr(AddrInst->getOperand(1), Depth+1))
693 // Otherwise we definitely can't merge the ADD in.
694 AddrMode = BackupAddrMode;
695 AddrModeInsts.resize(OldSize);
698 //case Instruction::Or:
699 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
701 case Instruction::Mul:
702 case Instruction::Shl: {
703 // Can only handle X*C and X << C.
704 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
705 if (!RHS) return false;
706 int64_t Scale = RHS->getSExtValue();
707 if (Opcode == Instruction::Shl)
710 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
712 case Instruction::GetElementPtr: {
713 // Scan the GEP. We check it if it contains constant offsets and at most
714 // one variable offset.
715 int VariableOperand = -1;
716 unsigned VariableScale = 0;
718 int64_t ConstantOffset = 0;
719 const TargetData *TD = TLI.getTargetData();
720 gep_type_iterator GTI = gep_type_begin(AddrInst);
721 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
722 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
723 const StructLayout *SL = TD->getStructLayout(STy);
725 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
726 ConstantOffset += SL->getElementOffset(Idx);
728 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
729 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
730 ConstantOffset += CI->getSExtValue()*TypeSize;
731 } else if (TypeSize) { // Scales of zero don't do anything.
732 // We only allow one variable index at the moment.
733 if (VariableOperand != -1)
736 // Remember the variable index.
738 VariableScale = TypeSize;
743 // A common case is for the GEP to only do a constant offset. In this case,
744 // just add it to the disp field and check validity.
745 if (VariableOperand == -1) {
746 AddrMode.BaseOffs += ConstantOffset;
747 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
748 // Check to see if we can fold the base pointer in too.
749 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
752 AddrMode.BaseOffs -= ConstantOffset;
756 // Save the valid addressing mode in case we can't match.
757 ExtAddrMode BackupAddrMode = AddrMode;
759 // Check that this has no base reg yet. If so, we won't have a place to
760 // put the base of the GEP (assuming it is not a null ptr).
761 bool SetBaseReg = true;
762 if (isa<ConstantPointerNull>(AddrInst->getOperand(0)))
763 SetBaseReg = false; // null pointer base doesn't need representation.
764 else if (AddrMode.HasBaseReg)
765 return false; // Base register already specified, can't match GEP.
767 // Otherwise, we'll use the GEP base as the BaseReg.
768 AddrMode.HasBaseReg = true;
769 AddrMode.BaseReg = AddrInst->getOperand(0);
772 // See if the scale and offset amount is valid for this target.
773 AddrMode.BaseOffs += ConstantOffset;
775 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
777 AddrMode = BackupAddrMode;
781 // If we have a null as the base of the GEP, folding in the constant offset
782 // plus variable scale is all we can do.
783 if (!SetBaseReg) return true;
785 // If this match succeeded, we know that we can form an address with the
786 // GepBase as the basereg. Match the base pointer of the GEP more
787 // aggressively by zeroing out BaseReg and rematching. If the base is
788 // (for example) another GEP, this allows merging in that other GEP into
789 // the addressing mode we're forming.
790 AddrMode.HasBaseReg = false;
791 AddrMode.BaseReg = 0;
792 bool Success = MatchAddr(AddrInst->getOperand(0), Depth+1);
793 assert(Success && "MatchAddr should be able to fill in BaseReg!");
801 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
802 /// addressing mode. If Addr can't be added to AddrMode this returns false and
803 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
804 /// or intptr_t for the target.
806 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
807 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
808 // Fold in immediates if legal for the target.
809 AddrMode.BaseOffs += CI->getSExtValue();
810 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
812 AddrMode.BaseOffs -= CI->getSExtValue();
813 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
814 // If this is a global variable, try to fold it into the addressing mode.
815 if (AddrMode.BaseGV == 0) {
816 AddrMode.BaseGV = GV;
817 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
821 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
822 ExtAddrMode BackupAddrMode = AddrMode;
823 unsigned OldSize = AddrModeInsts.size();
825 // Check to see if it is possible to fold this operation.
826 if (MatchOperationAddr(I, I->getOpcode(), Depth)) {
827 // Okay, it's possible to fold this. Check to see if it is actually
828 // *profitable* to do so. We use a simple cost model to avoid increasing
829 // register pressure too much.
830 if (I->hasOneUse() ||
831 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
832 AddrModeInsts.push_back(I);
836 // It isn't profitable to do this, roll back.
837 //cerr << "NOT FOLDING: " << *I;
838 AddrMode = BackupAddrMode;
839 AddrModeInsts.resize(OldSize);
841 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
842 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
844 } else if (isa<ConstantPointerNull>(Addr)) {
845 // Null pointer gets folded without affecting the addressing mode.
849 // Worse case, the target should support [reg] addressing modes. :)
850 if (!AddrMode.HasBaseReg) {
851 AddrMode.HasBaseReg = true;
852 AddrMode.BaseReg = Addr;
853 // Still check for legality in case the target supports [imm] but not [i+r].
854 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
856 AddrMode.HasBaseReg = false;
857 AddrMode.BaseReg = 0;
860 // If the base register is already taken, see if we can do [r+r].
861 if (AddrMode.Scale == 0) {
863 AddrMode.ScaledReg = Addr;
864 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
867 AddrMode.ScaledReg = 0;
874 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
875 /// inline asm call are due to memory operands. If so, return true, otherwise
877 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
878 const TargetLowering &TLI) {
879 std::vector<InlineAsm::ConstraintInfo>
880 Constraints = IA->ParseConstraints();
882 unsigned ArgNo = 1; // ArgNo - The operand of the CallInst.
883 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
884 TargetLowering::AsmOperandInfo OpInfo(Constraints[i]);
886 // Compute the value type for each operand.
887 switch (OpInfo.Type) {
888 case InlineAsm::isOutput:
889 if (OpInfo.isIndirect)
890 OpInfo.CallOperandVal = CI->getOperand(ArgNo++);
892 case InlineAsm::isInput:
893 OpInfo.CallOperandVal = CI->getOperand(ArgNo++);
895 case InlineAsm::isClobber:
900 // Compute the constraint code and ConstraintType to use.
901 TLI.ComputeConstraintToUse(OpInfo, SDValue(),
902 OpInfo.ConstraintType == TargetLowering::C_Memory);
904 // If this asm operand is our Value*, and if it isn't an indirect memory
905 // operand, we can't fold it!
906 if (OpInfo.CallOperandVal == OpVal &&
907 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
916 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
917 /// memory use. If we find an obviously non-foldable instruction, return true.
918 /// Add the ultimately found memory instructions to MemoryUses.
919 static bool FindAllMemoryUses(Instruction *I,
920 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
921 SmallPtrSet<Instruction*, 16> &ConsideredInsts,
922 const TargetLowering &TLI) {
923 // If we already considered this instruction, we're done.
924 if (!ConsideredInsts.insert(I))
927 // If this is an obviously unfoldable instruction, bail out.
928 if (!MightBeFoldableInst(I))
931 // Loop over all the uses, recursively processing them.
932 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
934 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
935 MemoryUses.push_back(std::make_pair(LI, UI.getOperandNo()));
939 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
940 if (UI.getOperandNo() == 0) return true; // Storing addr, not into addr.
941 MemoryUses.push_back(std::make_pair(SI, UI.getOperandNo()));
945 if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
946 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
947 if (IA == 0) return true;
949 // If this is a memory operand, we're cool, otherwise bail out.
950 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
955 if (FindAllMemoryUses(cast<Instruction>(*UI), MemoryUses, ConsideredInsts,
964 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
965 /// the use site that we're folding it into. If so, there is no cost to
966 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
967 /// that we know are live at the instruction already.
968 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
970 // If Val is either of the known-live values, we know it is live!
971 if (Val == 0 || Val == KnownLive1 || Val == KnownLive2)
974 // All values other than instructions and arguments (e.g. constants) are live.
975 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
977 // If Val is a constant sized alloca in the entry block, it is live, this is
978 // true because it is just a reference to the stack/frame pointer, which is
979 // live for the whole function.
980 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
981 if (AI->isStaticAlloca())
984 // Check to see if this value is already used in the memory instruction's
985 // block. If so, it's already live into the block at the very least, so we
986 // can reasonably fold it.
987 BasicBlock *MemBB = MemoryInst->getParent();
988 for (Value::use_iterator UI = Val->use_begin(), E = Val->use_end();
990 // We know that uses of arguments and instructions have to be instructions.
991 if (cast<Instruction>(*UI)->getParent() == MemBB)
999 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
1000 /// mode of the machine to fold the specified instruction into a load or store
1001 /// that ultimately uses it. However, the specified instruction has multiple
1002 /// uses. Given this, it may actually increase register pressure to fold it
1003 /// into the load. For example, consider this code:
1007 /// use(Y) -> nonload/store
1011 /// In this case, Y has multiple uses, and can be folded into the load of Z
1012 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
1013 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
1014 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
1015 /// number of computations either.
1017 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
1018 /// X was live across 'load Z' for other reasons, we actually *would* want to
1019 /// fold the addressing mode in the Z case. This would make Y die earlier.
1020 bool AddressingModeMatcher::
1021 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
1022 ExtAddrMode &AMAfter) {
1023 if (IgnoreProfitability) return true;
1025 // AMBefore is the addressing mode before this instruction was folded into it,
1026 // and AMAfter is the addressing mode after the instruction was folded. Get
1027 // the set of registers referenced by AMAfter and subtract out those
1028 // referenced by AMBefore: this is the set of values which folding in this
1029 // address extends the lifetime of.
1031 // Note that there are only two potential values being referenced here,
1032 // BaseReg and ScaleReg (global addresses are always available, as are any
1033 // folded immediates).
1034 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
1036 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
1037 // lifetime wasn't extended by adding this instruction.
1038 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
1040 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
1043 // If folding this instruction (and it's subexprs) didn't extend any live
1044 // ranges, we're ok with it.
1045 if (BaseReg == 0 && ScaledReg == 0)
1048 // If all uses of this instruction are ultimately load/store/inlineasm's,
1049 // check to see if their addressing modes will include this instruction. If
1050 // so, we can fold it into all uses, so it doesn't matter if it has multiple
1052 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
1053 SmallPtrSet<Instruction*, 16> ConsideredInsts;
1054 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
1055 return false; // Has a non-memory, non-foldable use!
1057 // Now that we know that all uses of this instruction are part of a chain of
1058 // computation involving only operations that could theoretically be folded
1059 // into a memory use, loop over each of these uses and see if they could
1060 // *actually* fold the instruction.
1061 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
1062 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
1063 Instruction *User = MemoryUses[i].first;
1064 unsigned OpNo = MemoryUses[i].second;
1066 // Get the access type of this use. If the use isn't a pointer, we don't
1067 // know what it accesses.
1068 Value *Address = User->getOperand(OpNo);
1069 if (!isa<PointerType>(Address->getType()))
1071 const Type *AddressAccessTy =
1072 cast<PointerType>(Address->getType())->getElementType();
1074 // Do a match against the root of this address, ignoring profitability. This
1075 // will tell us if the addressing mode for the memory operation will
1076 // *actually* cover the shared instruction.
1078 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
1079 MemoryInst, Result);
1080 Matcher.IgnoreProfitability = true;
1081 bool Success = Matcher.MatchAddr(Address, 0);
1082 Success = Success; assert(Success && "Couldn't select *anything*?");
1084 // If the match didn't cover I, then it won't be shared by it.
1085 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
1086 I) == MatchedAddrModeInsts.end())
1089 MatchedAddrModeInsts.clear();
1096 //===----------------------------------------------------------------------===//
1097 // Memory Optimization
1098 //===----------------------------------------------------------------------===//
1100 /// IsNonLocalValue - Return true if the specified values are defined in a
1101 /// different basic block than BB.
1102 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
1103 if (Instruction *I = dyn_cast<Instruction>(V))
1104 return I->getParent() != BB;
1108 /// OptimizeMemoryInst - Load and Store Instructions have often have
1109 /// addressing modes that can do significant amounts of computation. As such,
1110 /// instruction selection will try to get the load or store to do as much
1111 /// computation as possible for the program. The problem is that isel can only
1112 /// see within a single block. As such, we sink as much legal addressing mode
1113 /// stuff into the block as possible.
1115 /// This method is used to optimize both load/store and inline asms with memory
1117 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
1118 const Type *AccessTy,
1119 DenseMap<Value*,Value*> &SunkAddrs) {
1120 // Figure out what addressing mode will be built up for this operation.
1121 SmallVector<Instruction*, 16> AddrModeInsts;
1122 ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst,
1123 AddrModeInsts, *TLI);
1125 // Check to see if any of the instructions supersumed by this addr mode are
1126 // non-local to I's BB.
1127 bool AnyNonLocal = false;
1128 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
1129 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
1135 // If all the instructions matched are already in this BB, don't do anything.
1137 DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n");
1141 // Insert this computation right after this user. Since our caller is
1142 // scanning from the top of the BB to the bottom, reuse of the expr are
1143 // guaranteed to happen later.
1144 BasicBlock::iterator InsertPt = MemoryInst;
1146 // Now that we determined the addressing expression we want to use and know
1147 // that we have to sink it into this block. Check to see if we have already
1148 // done this for some other load/store instr in this block. If so, reuse the
1150 Value *&SunkAddr = SunkAddrs[Addr];
1152 DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << "\n");
1153 if (SunkAddr->getType() != Addr->getType())
1154 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
1156 DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << "\n");
1157 const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType();
1160 // Start with the scale value.
1161 if (AddrMode.Scale) {
1162 Value *V = AddrMode.ScaledReg;
1163 if (V->getType() == IntPtrTy) {
1165 } else if (isa<PointerType>(V->getType())) {
1166 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
1167 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
1168 cast<IntegerType>(V->getType())->getBitWidth()) {
1169 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
1171 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
1173 if (AddrMode.Scale != 1)
1174 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
1176 "sunkaddr", InsertPt);
1180 // Add in the base register.
1181 if (AddrMode.BaseReg) {
1182 Value *V = AddrMode.BaseReg;
1183 if (V->getType() != IntPtrTy)
1184 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
1186 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
1191 // Add in the BaseGV if present.
1192 if (AddrMode.BaseGV) {
1193 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
1196 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
1201 // Add in the Base Offset if present.
1202 if (AddrMode.BaseOffs) {
1203 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
1205 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
1211 SunkAddr = Constant::getNullValue(Addr->getType());
1213 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
1216 MemoryInst->replaceUsesOfWith(Addr, SunkAddr);
1218 if (Addr->use_empty())
1219 RecursivelyDeleteTriviallyDeadInstructions(Addr);
1223 /// OptimizeInlineAsmInst - If there are any memory operands, use
1224 /// OptimizeMemoryInst to sink their address computing into the block when
1225 /// possible / profitable.
1226 bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS,
1227 DenseMap<Value*,Value*> &SunkAddrs) {
1228 bool MadeChange = false;
1229 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
1231 // Do a prepass over the constraints, canonicalizing them, and building up the
1232 // ConstraintOperands list.
1233 std::vector<InlineAsm::ConstraintInfo>
1234 ConstraintInfos = IA->ParseConstraints();
1236 /// ConstraintOperands - Information about all of the constraints.
1237 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands;
1238 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
1239 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
1241 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i]));
1242 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back();
1244 // Compute the value type for each operand.
1245 switch (OpInfo.Type) {
1246 case InlineAsm::isOutput:
1247 if (OpInfo.isIndirect)
1248 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
1250 case InlineAsm::isInput:
1251 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
1253 case InlineAsm::isClobber:
1258 // Compute the constraint code and ConstraintType to use.
1259 TLI->ComputeConstraintToUse(OpInfo, SDValue(),
1260 OpInfo.ConstraintType == TargetLowering::C_Memory);
1262 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
1263 OpInfo.isIndirect) {
1264 Value *OpVal = OpInfo.CallOperandVal;
1265 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs);
1272 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
1273 BasicBlock *DefBB = I->getParent();
1275 // If both result of the {s|z}xt and its source are live out, rewrite all
1276 // other uses of the source with result of extension.
1277 Value *Src = I->getOperand(0);
1278 if (Src->hasOneUse())
1281 // Only do this xform if truncating is free.
1282 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
1285 // Only safe to perform the optimization if the source is also defined in
1287 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
1290 bool DefIsLiveOut = false;
1291 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1293 Instruction *User = cast<Instruction>(*UI);
1295 // Figure out which BB this ext is used in.
1296 BasicBlock *UserBB = User->getParent();
1297 if (UserBB == DefBB) continue;
1298 DefIsLiveOut = true;
1304 // Make sure non of the uses are PHI nodes.
1305 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1307 Instruction *User = cast<Instruction>(*UI);
1308 BasicBlock *UserBB = User->getParent();
1309 if (UserBB == DefBB) continue;
1310 // Be conservative. We don't want this xform to end up introducing
1311 // reloads just before load / store instructions.
1312 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
1316 // InsertedTruncs - Only insert one trunc in each block once.
1317 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
1319 bool MadeChange = false;
1320 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1322 Use &TheUse = UI.getUse();
1323 Instruction *User = cast<Instruction>(*UI);
1325 // Figure out which BB this ext is used in.
1326 BasicBlock *UserBB = User->getParent();
1327 if (UserBB == DefBB) continue;
1329 // Both src and def are live in this block. Rewrite the use.
1330 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
1332 if (!InsertedTrunc) {
1333 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
1335 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
1338 // Replace a use of the {s|z}ext source with a use of the result.
1339 TheUse = InsertedTrunc;
1347 // In this pass we look for GEP and cast instructions that are used
1348 // across basic blocks and rewrite them to improve basic-block-at-a-time
1350 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
1351 bool MadeChange = false;
1353 // Split all critical edges where the dest block has a PHI and where the phi
1354 // has shared immediate operands.
1355 TerminatorInst *BBTI = BB.getTerminator();
1356 if (BBTI->getNumSuccessors() > 1) {
1357 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i)
1358 if (isa<PHINode>(BBTI->getSuccessor(i)->begin()) &&
1359 isCriticalEdge(BBTI, i, true))
1360 SplitEdgeNicely(BBTI, i, this);
1364 // Keep track of non-local addresses that have been sunk into this block.
1365 // This allows us to avoid inserting duplicate code for blocks with multiple
1366 // load/stores of the same address.
1367 DenseMap<Value*, Value*> SunkAddrs;
1369 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) {
1370 Instruction *I = BBI++;
1372 if (CastInst *CI = dyn_cast<CastInst>(I)) {
1373 // If the source of the cast is a constant, then this should have
1374 // already been constant folded. The only reason NOT to constant fold
1375 // it is if something (e.g. LSR) was careful to place the constant
1376 // evaluation in a block other than then one that uses it (e.g. to hoist
1377 // the address of globals out of a loop). If this is the case, we don't
1378 // want to forward-subst the cast.
1379 if (isa<Constant>(CI->getOperand(0)))
1382 bool Change = false;
1384 Change = OptimizeNoopCopyExpression(CI, *TLI);
1385 MadeChange |= Change;
1388 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I)))
1389 MadeChange |= OptimizeExtUses(I);
1390 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
1391 MadeChange |= OptimizeCmpExpression(CI);
1392 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1394 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(),
1396 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1398 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1),
1399 SI->getOperand(0)->getType(),
1401 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1402 if (GEPI->hasAllZeroIndices()) {
1403 /// The GEP operand must be a pointer, so must its result -> BitCast
1404 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1405 GEPI->getName(), GEPI);
1406 GEPI->replaceAllUsesWith(NC);
1407 GEPI->eraseFromParent();
1411 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
1412 // If we found an inline asm expession, and if the target knows how to
1413 // lower it to normal LLVM code, do so now.
1414 if (TLI && isa<InlineAsm>(CI->getCalledValue()))
1415 if (const TargetAsmInfo *TAI =
1416 TLI->getTargetMachine().getTargetAsmInfo()) {
1417 if (TAI->ExpandInlineAsm(CI))
1420 // Sink address computing for memory operands into the block.
1421 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs);