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 (DestBB->getSinglePredecessor()) {
208 // If DestBB has single-entry PHI nodes, fold them.
209 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
210 Value *NewVal = PN->getIncomingValue(0);
211 // Replace self referencing PHI with undef, it must be dead.
212 if (NewVal == PN) NewVal = UndefValue::get(PN->getType());
213 PN->replaceAllUsesWith(NewVal);
214 PN->eraseFromParent();
217 // Splice all the PHI nodes from BB over to DestBB.
218 DestBB->getInstList().splice(DestBB->begin(), BB->getInstList(),
221 // Anything that branched to BB now branches to DestBB.
222 BB->replaceAllUsesWith(DestBB);
225 BB->eraseFromParent();
227 DOUT << "AFTER:\n" << *DestBB << "\n\n\n";
231 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
232 // to handle the new incoming edges it is about to have.
234 for (BasicBlock::iterator BBI = DestBB->begin();
235 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
236 // Remove the incoming value for BB, and remember it.
237 Value *InVal = PN->removeIncomingValue(BB, false);
239 // Two options: either the InVal is a phi node defined in BB or it is some
240 // value that dominates BB.
241 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
242 if (InValPhi && InValPhi->getParent() == BB) {
243 // Add all of the input values of the input PHI as inputs of this phi.
244 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
245 PN->addIncoming(InValPhi->getIncomingValue(i),
246 InValPhi->getIncomingBlock(i));
248 // Otherwise, add one instance of the dominating value for each edge that
249 // we will be adding.
250 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
251 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
252 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
254 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
255 PN->addIncoming(InVal, *PI);
260 // The PHIs are now updated, change everything that refers to BB to use
261 // DestBB and remove BB.
262 BB->replaceAllUsesWith(DestBB);
263 BB->eraseFromParent();
265 DOUT << "AFTER:\n" << *DestBB << "\n\n\n";
269 /// SplitEdgeNicely - Split the critical edge from TI to its specified
270 /// successor if it will improve codegen. We only do this if the successor has
271 /// phi nodes (otherwise critical edges are ok). If there is already another
272 /// predecessor of the succ that is empty (and thus has no phi nodes), use it
273 /// instead of introducing a new block.
274 static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, Pass *P) {
275 BasicBlock *TIBB = TI->getParent();
276 BasicBlock *Dest = TI->getSuccessor(SuccNum);
277 assert(isa<PHINode>(Dest->begin()) &&
278 "This should only be called if Dest has a PHI!");
280 // As a hack, never split backedges of loops. Even though the copy for any
281 // PHIs inserted on the backedge would be dead for exits from the loop, we
282 // assume that the cost of *splitting* the backedge would be too high.
286 /// TIPHIValues - This array is lazily computed to determine the values of
287 /// PHIs in Dest that TI would provide.
288 SmallVector<Value*, 32> TIPHIValues;
290 // Check to see if Dest has any blocks that can be used as a split edge for
292 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) {
293 BasicBlock *Pred = *PI;
294 // To be usable, the pred has to end with an uncond branch to the dest.
295 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator());
296 if (!PredBr || !PredBr->isUnconditional() ||
297 // Must be empty other than the branch.
298 &Pred->front() != PredBr ||
299 // Cannot be the entry block; its label does not get emitted.
300 Pred == &(Dest->getParent()->getEntryBlock()))
303 // Finally, since we know that Dest has phi nodes in it, we have to make
304 // sure that jumping to Pred will have the same affect as going to Dest in
305 // terms of PHI values.
308 bool FoundMatch = true;
309 for (BasicBlock::iterator I = Dest->begin();
310 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) {
311 if (PHINo == TIPHIValues.size())
312 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB));
314 // If the PHI entry doesn't work, we can't use this pred.
315 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) {
321 // If we found a workable predecessor, change TI to branch to Succ.
323 Dest->removePredecessor(TIBB);
324 TI->setSuccessor(SuccNum, Pred);
329 SplitCriticalEdge(TI, SuccNum, P, true);
332 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
333 /// copy (e.g. it's casting from one pointer type to another, int->uint, or
334 /// int->sbyte on PPC), sink it into user blocks to reduce the number of virtual
335 /// registers that must be created and coalesced.
337 /// Return true if any changes are made.
339 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
340 // If this is a noop copy,
341 MVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
342 MVT DstVT = TLI.getValueType(CI->getType());
344 // This is an fp<->int conversion?
345 if (SrcVT.isInteger() != DstVT.isInteger())
348 // If this is an extension, it will be a zero or sign extension, which
350 if (SrcVT.bitsLT(DstVT)) return false;
352 // If these values will be promoted, find out what they will be promoted
353 // to. This helps us consider truncates on PPC as noop copies when they
355 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote)
356 SrcVT = TLI.getTypeToTransformTo(SrcVT);
357 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote)
358 DstVT = TLI.getTypeToTransformTo(DstVT);
360 // If, after promotion, these are the same types, this is a noop copy.
364 BasicBlock *DefBB = CI->getParent();
366 /// InsertedCasts - Only insert a cast in each block once.
367 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
369 bool MadeChange = false;
370 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
372 Use &TheUse = UI.getUse();
373 Instruction *User = cast<Instruction>(*UI);
375 // Figure out which BB this cast is used in. For PHI's this is the
376 // appropriate predecessor block.
377 BasicBlock *UserBB = User->getParent();
378 if (PHINode *PN = dyn_cast<PHINode>(User)) {
379 unsigned OpVal = UI.getOperandNo()/2;
380 UserBB = PN->getIncomingBlock(OpVal);
383 // Preincrement use iterator so we don't invalidate it.
386 // If this user is in the same block as the cast, don't change the cast.
387 if (UserBB == DefBB) continue;
389 // If we have already inserted a cast into this block, use it.
390 CastInst *&InsertedCast = InsertedCasts[UserBB];
393 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
396 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
401 // Replace a use of the cast with a use of the new cast.
402 TheUse = InsertedCast;
405 // If we removed all uses, nuke the cast.
406 if (CI->use_empty()) {
407 CI->eraseFromParent();
414 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
415 /// the number of virtual registers that must be created and coalesced. This is
416 /// a clear win except on targets with multiple condition code registers
417 /// (PowerPC), where it might lose; some adjustment may be wanted there.
419 /// Return true if any changes are made.
420 static bool OptimizeCmpExpression(CmpInst *CI) {
421 BasicBlock *DefBB = CI->getParent();
423 /// InsertedCmp - Only insert a cmp in each block once.
424 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
426 bool MadeChange = false;
427 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
429 Use &TheUse = UI.getUse();
430 Instruction *User = cast<Instruction>(*UI);
432 // Preincrement use iterator so we don't invalidate it.
435 // Don't bother for PHI nodes.
436 if (isa<PHINode>(User))
439 // Figure out which BB this cmp is used in.
440 BasicBlock *UserBB = User->getParent();
442 // If this user is in the same block as the cmp, don't change the cmp.
443 if (UserBB == DefBB) continue;
445 // If we have already inserted a cmp into this block, use it.
446 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
449 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
452 CmpInst::Create(CI->getOpcode(), CI->getPredicate(), CI->getOperand(0),
453 CI->getOperand(1), "", InsertPt);
457 // Replace a use of the cmp with a use of the new cmp.
458 TheUse = InsertedCmp;
461 // If we removed all uses, nuke the cmp.
463 CI->eraseFromParent();
468 /// EraseDeadInstructions - Erase any dead instructions, recursively.
469 static void EraseDeadInstructions(Value *V) {
470 Instruction *I = dyn_cast<Instruction>(V);
471 if (!I || !I->use_empty()) return;
473 SmallPtrSet<Instruction*, 16> Insts;
476 while (!Insts.empty()) {
479 if (isInstructionTriviallyDead(I)) {
480 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
481 if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
483 I->eraseFromParent();
488 //===----------------------------------------------------------------------===//
489 // Addressing Mode Analysis and Optimization
490 //===----------------------------------------------------------------------===//
493 /// ExtAddrMode - This is an extended version of TargetLowering::AddrMode
494 /// which holds actual Value*'s for register values.
495 struct ExtAddrMode : public TargetLowering::AddrMode {
498 ExtAddrMode() : BaseReg(0), ScaledReg(0) {}
499 void print(OStream &OS) const;
505 } // end anonymous namespace
507 static inline OStream &operator<<(OStream &OS, const ExtAddrMode &AM) {
512 void ExtAddrMode::print(OStream &OS) const {
513 bool NeedPlus = false;
516 OS << (NeedPlus ? " + " : "")
517 << "GV:%" << BaseGV->getName(), NeedPlus = true;
520 OS << (NeedPlus ? " + " : "") << BaseOffs, NeedPlus = true;
523 OS << (NeedPlus ? " + " : "")
524 << "Base:%" << BaseReg->getName(), NeedPlus = true;
526 OS << (NeedPlus ? " + " : "")
527 << Scale << "*%" << ScaledReg->getName(), NeedPlus = true;
533 /// AddressingModeMatcher - This class exposes a single public method, which is
534 /// used to construct a "maximal munch" of the addressing mode for the target
535 /// specified by TLI for an access to "V" with an access type of AccessTy. This
536 /// returns the addressing mode that is actually matched by value, but also
537 /// returns the list of instructions involved in that addressing computation in
539 class AddressingModeMatcher {
540 SmallVectorImpl<Instruction*> &AddrModeInsts;
541 const TargetLowering &TLI;
543 /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and
544 /// the memory instruction that we're computing this address for.
545 const Type *AccessTy;
546 Instruction *MemoryInst;
548 /// AddrMode - This is the addressing mode that we're building up. This is
549 /// part of the return value of this addressing mode matching stuff.
550 ExtAddrMode &AddrMode;
552 /// IgnoreProfitability - This is set to true when we should not do
553 /// profitability checks. When true, IsProfitableToFoldIntoAddressingMode
554 /// always returns true.
555 bool IgnoreProfitability;
557 AddressingModeMatcher(SmallVectorImpl<Instruction*> &AMI,
558 const TargetLowering &T, const Type *AT,
559 Instruction *MI, ExtAddrMode &AM)
560 : AddrModeInsts(AMI), TLI(T), AccessTy(AT), MemoryInst(MI), AddrMode(AM) {
561 IgnoreProfitability = false;
565 /// Match - Find the maximal addressing mode that a load/store of V can fold,
566 /// give an access type of AccessTy. This returns a list of involved
567 /// instructions in AddrModeInsts.
568 static ExtAddrMode Match(Value *V, const Type *AccessTy,
569 Instruction *MemoryInst,
570 SmallVectorImpl<Instruction*> &AddrModeInsts,
571 const TargetLowering &TLI) {
575 AddressingModeMatcher(AddrModeInsts, TLI, AccessTy,
576 MemoryInst, Result).MatchAddr(V, 0);
577 Success = Success; assert(Success && "Couldn't select *anything*?");
581 bool MatchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth);
582 bool MatchAddr(Value *V, unsigned Depth);
583 bool MatchOperationAddr(User *Operation, unsigned Opcode, unsigned Depth);
584 bool IsProfitableToFoldIntoAddressingMode(Instruction *I,
585 ExtAddrMode &AMBefore,
586 ExtAddrMode &AMAfter);
587 bool ValueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2);
589 } // end anonymous namespace
591 /// MatchScaledValue - Try adding ScaleReg*Scale to the current addressing mode.
592 /// Return true and update AddrMode if this addr mode is legal for the target,
594 bool AddressingModeMatcher::MatchScaledValue(Value *ScaleReg, int64_t Scale,
596 // If Scale is 1, then this is the same as adding ScaleReg to the addressing
597 // mode. Just process that directly.
599 return MatchAddr(ScaleReg, Depth);
601 // If the scale is 0, it takes nothing to add this.
605 // If we already have a scale of this value, we can add to it, otherwise, we
606 // need an available scale field.
607 if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg)
610 ExtAddrMode TestAddrMode = AddrMode;
612 // Add scale to turn X*4+X*3 -> X*7. This could also do things like
613 // [A+B + A*7] -> [B+A*8].
614 TestAddrMode.Scale += Scale;
615 TestAddrMode.ScaledReg = ScaleReg;
617 // If the new address isn't legal, bail out.
618 if (!TLI.isLegalAddressingMode(TestAddrMode, AccessTy))
621 // It was legal, so commit it.
622 AddrMode = TestAddrMode;
624 // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now
625 // to see if ScaleReg is actually X+C. If so, we can turn this into adding
626 // X*Scale + C*Scale to addr mode.
627 ConstantInt *CI; Value *AddLHS;
628 if (match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI)))) {
629 TestAddrMode.ScaledReg = AddLHS;
630 TestAddrMode.BaseOffs += CI->getSExtValue()*TestAddrMode.Scale;
632 // If this addressing mode is legal, commit it and remember that we folded
634 if (TLI.isLegalAddressingMode(TestAddrMode, AccessTy)) {
635 AddrModeInsts.push_back(cast<Instruction>(ScaleReg));
636 AddrMode = TestAddrMode;
641 // Otherwise, not (x+c)*scale, just return what we have.
645 /// MightBeFoldableInst - This is a little filter, which returns true if an
646 /// addressing computation involving I might be folded into a load/store
647 /// accessing it. This doesn't need to be perfect, but needs to accept at least
648 /// the set of instructions that MatchOperationAddr can.
649 static bool MightBeFoldableInst(Instruction *I) {
650 switch (I->getOpcode()) {
651 case Instruction::BitCast:
652 // Don't touch identity bitcasts.
653 if (I->getType() == I->getOperand(0)->getType())
655 return isa<PointerType>(I->getType()) || isa<IntegerType>(I->getType());
656 case Instruction::PtrToInt:
657 // PtrToInt is always a noop, as we know that the int type is pointer sized.
659 case Instruction::IntToPtr:
660 // We know the input is intptr_t, so this is foldable.
662 case Instruction::Add:
664 case Instruction::Mul:
665 case Instruction::Shl:
666 // Can only handle X*C and X << C.
667 return isa<ConstantInt>(I->getOperand(1));
668 case Instruction::GetElementPtr:
676 /// MatchOperationAddr - Given an instruction or constant expr, see if we can
677 /// fold the operation into the addressing mode. If so, update the addressing
678 /// mode and return true, otherwise return false without modifying AddrMode.
679 bool AddressingModeMatcher::MatchOperationAddr(User *AddrInst, unsigned Opcode,
681 // Avoid exponential behavior on extremely deep expression trees.
682 if (Depth >= 5) return false;
685 case Instruction::PtrToInt:
686 // PtrToInt is always a noop, as we know that the int type is pointer sized.
687 return MatchAddr(AddrInst->getOperand(0), Depth);
688 case Instruction::IntToPtr:
689 // This inttoptr is a no-op if the integer type is pointer sized.
690 if (TLI.getValueType(AddrInst->getOperand(0)->getType()) ==
692 return MatchAddr(AddrInst->getOperand(0), Depth);
694 case Instruction::BitCast:
695 // BitCast is always a noop, and we can handle it as long as it is
696 // int->int or pointer->pointer (we don't want int<->fp or something).
697 if ((isa<PointerType>(AddrInst->getOperand(0)->getType()) ||
698 isa<IntegerType>(AddrInst->getOperand(0)->getType())) &&
699 // Don't touch identity bitcasts. These were probably put here by LSR,
700 // and we don't want to mess around with them. Assume it knows what it
702 AddrInst->getOperand(0)->getType() != AddrInst->getType())
703 return MatchAddr(AddrInst->getOperand(0), Depth);
705 case Instruction::Add: {
706 // Check to see if we can merge in the RHS then the LHS. If so, we win.
707 ExtAddrMode BackupAddrMode = AddrMode;
708 unsigned OldSize = AddrModeInsts.size();
709 if (MatchAddr(AddrInst->getOperand(1), Depth+1) &&
710 MatchAddr(AddrInst->getOperand(0), Depth+1))
713 // Restore the old addr mode info.
714 AddrMode = BackupAddrMode;
715 AddrModeInsts.resize(OldSize);
717 // Otherwise this was over-aggressive. Try merging in the LHS then the RHS.
718 if (MatchAddr(AddrInst->getOperand(0), Depth+1) &&
719 MatchAddr(AddrInst->getOperand(1), Depth+1))
722 // Otherwise we definitely can't merge the ADD in.
723 AddrMode = BackupAddrMode;
724 AddrModeInsts.resize(OldSize);
727 //case Instruction::Or:
728 // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD.
730 case Instruction::Mul:
731 case Instruction::Shl: {
732 // Can only handle X*C and X << C.
733 ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1));
734 if (!RHS) return false;
735 int64_t Scale = RHS->getSExtValue();
736 if (Opcode == Instruction::Shl)
739 return MatchScaledValue(AddrInst->getOperand(0), Scale, Depth);
741 case Instruction::GetElementPtr: {
742 // Scan the GEP. We check it if it contains constant offsets and at most
743 // one variable offset.
744 int VariableOperand = -1;
745 unsigned VariableScale = 0;
747 int64_t ConstantOffset = 0;
748 const TargetData *TD = TLI.getTargetData();
749 gep_type_iterator GTI = gep_type_begin(AddrInst);
750 for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) {
751 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
752 const StructLayout *SL = TD->getStructLayout(STy);
754 cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue();
755 ConstantOffset += SL->getElementOffset(Idx);
757 uint64_t TypeSize = TD->getABITypeSize(GTI.getIndexedType());
758 if (ConstantInt *CI = dyn_cast<ConstantInt>(AddrInst->getOperand(i))) {
759 ConstantOffset += CI->getSExtValue()*TypeSize;
760 } else if (TypeSize) { // Scales of zero don't do anything.
761 // We only allow one variable index at the moment.
762 if (VariableOperand != -1)
765 // Remember the variable index.
767 VariableScale = TypeSize;
772 // A common case is for the GEP to only do a constant offset. In this case,
773 // just add it to the disp field and check validity.
774 if (VariableOperand == -1) {
775 AddrMode.BaseOffs += ConstantOffset;
776 if (ConstantOffset == 0 || TLI.isLegalAddressingMode(AddrMode, AccessTy)){
777 // Check to see if we can fold the base pointer in too.
778 if (MatchAddr(AddrInst->getOperand(0), Depth+1))
781 AddrMode.BaseOffs -= ConstantOffset;
785 // Save the valid addressing mode in case we can't match.
786 ExtAddrMode BackupAddrMode = AddrMode;
788 // Check that this has no base reg yet. If so, we won't have a place to
789 // put the base of the GEP (assuming it is not a null ptr).
790 bool SetBaseReg = true;
791 if (isa<ConstantPointerNull>(AddrInst->getOperand(0)))
792 SetBaseReg = false; // null pointer base doesn't need representation.
793 else if (AddrMode.HasBaseReg)
794 return false; // Base register already specified, can't match GEP.
796 // Otherwise, we'll use the GEP base as the BaseReg.
797 AddrMode.HasBaseReg = true;
798 AddrMode.BaseReg = AddrInst->getOperand(0);
801 // See if the scale and offset amount is valid for this target.
802 AddrMode.BaseOffs += ConstantOffset;
804 if (!MatchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale,
806 AddrMode = BackupAddrMode;
810 // If we have a null as the base of the GEP, folding in the constant offset
811 // plus variable scale is all we can do.
812 if (!SetBaseReg) return true;
814 // If this match succeeded, we know that we can form an address with the
815 // GepBase as the basereg. Match the base pointer of the GEP more
816 // aggressively by zeroing out BaseReg and rematching. If the base is
817 // (for example) another GEP, this allows merging in that other GEP into
818 // the addressing mode we're forming.
819 AddrMode.HasBaseReg = false;
820 AddrMode.BaseReg = 0;
821 bool Success = MatchAddr(AddrInst->getOperand(0), Depth+1);
822 assert(Success && "MatchAddr should be able to fill in BaseReg!");
830 /// MatchAddr - If we can, try to add the value of 'Addr' into the current
831 /// addressing mode. If Addr can't be added to AddrMode this returns false and
832 /// leaves AddrMode unmodified. This assumes that Addr is either a pointer type
833 /// or intptr_t for the target.
835 bool AddressingModeMatcher::MatchAddr(Value *Addr, unsigned Depth) {
836 if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) {
837 // Fold in immediates if legal for the target.
838 AddrMode.BaseOffs += CI->getSExtValue();
839 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
841 AddrMode.BaseOffs -= CI->getSExtValue();
842 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
843 // If this is a global variable, try to fold it into the addressing mode.
844 if (AddrMode.BaseGV == 0) {
845 AddrMode.BaseGV = GV;
846 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
850 } else if (Instruction *I = dyn_cast<Instruction>(Addr)) {
851 ExtAddrMode BackupAddrMode = AddrMode;
852 unsigned OldSize = AddrModeInsts.size();
854 // Check to see if it is possible to fold this operation.
855 if (MatchOperationAddr(I, I->getOpcode(), Depth)) {
856 // Okay, it's possible to fold this. Check to see if it is actually
857 // *profitable* to do so. We use a simple cost model to avoid increasing
858 // register pressure too much.
859 if (I->hasOneUse() ||
860 IsProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) {
861 AddrModeInsts.push_back(I);
865 // It isn't profitable to do this, roll back.
866 //cerr << "NOT FOLDING: " << *I;
867 AddrMode = BackupAddrMode;
868 AddrModeInsts.resize(OldSize);
870 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
871 if (MatchOperationAddr(CE, CE->getOpcode(), Depth))
873 } else if (isa<ConstantPointerNull>(Addr)) {
874 // Null pointer gets folded without affecting the addressing mode.
878 // Worse case, the target should support [reg] addressing modes. :)
879 if (!AddrMode.HasBaseReg) {
880 AddrMode.HasBaseReg = true;
881 AddrMode.BaseReg = Addr;
882 // Still check for legality in case the target supports [imm] but not [i+r].
883 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
885 AddrMode.HasBaseReg = false;
886 AddrMode.BaseReg = 0;
889 // If the base register is already taken, see if we can do [r+r].
890 if (AddrMode.Scale == 0) {
892 AddrMode.ScaledReg = Addr;
893 if (TLI.isLegalAddressingMode(AddrMode, AccessTy))
896 AddrMode.ScaledReg = 0;
903 /// IsOperandAMemoryOperand - Check to see if all uses of OpVal by the specified
904 /// inline asm call are due to memory operands. If so, return true, otherwise
906 static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal,
907 const TargetLowering &TLI) {
908 std::vector<InlineAsm::ConstraintInfo>
909 Constraints = IA->ParseConstraints();
911 unsigned ArgNo = 1; // ArgNo - The operand of the CallInst.
912 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
913 TargetLowering::AsmOperandInfo OpInfo(Constraints[i]);
915 // Compute the value type for each operand.
916 switch (OpInfo.Type) {
917 case InlineAsm::isOutput:
918 if (OpInfo.isIndirect)
919 OpInfo.CallOperandVal = CI->getOperand(ArgNo++);
921 case InlineAsm::isInput:
922 OpInfo.CallOperandVal = CI->getOperand(ArgNo++);
924 case InlineAsm::isClobber:
929 // Compute the constraint code and ConstraintType to use.
930 TLI.ComputeConstraintToUse(OpInfo, SDValue(),
931 OpInfo.ConstraintType == TargetLowering::C_Memory);
933 // If this asm operand is our Value*, and if it isn't an indirect memory
934 // operand, we can't fold it!
935 if (OpInfo.CallOperandVal == OpVal &&
936 (OpInfo.ConstraintType != TargetLowering::C_Memory ||
945 /// FindAllMemoryUses - Recursively walk all the uses of I until we find a
946 /// memory use. If we find an obviously non-foldable instruction, return true.
947 /// Add the ultimately found memory instructions to MemoryUses.
948 static bool FindAllMemoryUses(Instruction *I,
949 SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
950 SmallPtrSet<Instruction*, 16> &ConsideredInsts,
951 const TargetLowering &TLI) {
952 // If we already considered this instruction, we're done.
953 if (!ConsideredInsts.insert(I))
956 // If this is an obviously unfoldable instruction, bail out.
957 if (!MightBeFoldableInst(I))
960 // Loop over all the uses, recursively processing them.
961 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
963 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
964 MemoryUses.push_back(std::make_pair(LI, UI.getOperandNo()));
968 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
969 if (UI.getOperandNo() == 0) return true; // Storing addr, not into addr.
970 MemoryUses.push_back(std::make_pair(SI, UI.getOperandNo()));
974 if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
975 InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
976 if (IA == 0) return true;
978 // If this is a memory operand, we're cool, otherwise bail out.
979 if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
984 if (FindAllMemoryUses(cast<Instruction>(*UI), MemoryUses, ConsideredInsts,
993 /// ValueAlreadyLiveAtInst - Retrn true if Val is already known to be live at
994 /// the use site that we're folding it into. If so, there is no cost to
995 /// include it in the addressing mode. KnownLive1 and KnownLive2 are two values
996 /// that we know are live at the instruction already.
997 bool AddressingModeMatcher::ValueAlreadyLiveAtInst(Value *Val,Value *KnownLive1,
999 // If Val is either of the known-live values, we know it is live!
1000 if (Val == 0 || Val == KnownLive1 || Val == KnownLive2)
1003 // All values other than instructions and arguments (e.g. constants) are live.
1004 if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true;
1006 // If Val is a constant sized alloca in the entry block, it is live, this is
1007 // true because it is just a reference to the stack/frame pointer, which is
1008 // live for the whole function.
1009 if (AllocaInst *AI = dyn_cast<AllocaInst>(Val))
1010 if (AI->isStaticAlloca())
1013 // Check to see if this value is already used in the memory instruction's
1014 // block. If so, it's already live into the block at the very least, so we
1015 // can reasonably fold it.
1016 BasicBlock *MemBB = MemoryInst->getParent();
1017 for (Value::use_iterator UI = Val->use_begin(), E = Val->use_end();
1019 // We know that uses of arguments and instructions have to be instructions.
1020 if (cast<Instruction>(*UI)->getParent() == MemBB)
1028 #include "llvm/Support/CommandLine.h"
1029 cl::opt<bool> ENABLECRAZYHACK("enable-smarter-addr-folding", cl::Hidden);
1032 /// IsProfitableToFoldIntoAddressingMode - It is possible for the addressing
1033 /// mode of the machine to fold the specified instruction into a load or store
1034 /// that ultimately uses it. However, the specified instruction has multiple
1035 /// uses. Given this, it may actually increase register pressure to fold it
1036 /// into the load. For example, consider this code:
1040 /// use(Y) -> nonload/store
1044 /// In this case, Y has multiple uses, and can be folded into the load of Z
1045 /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to
1046 /// be live at the use(Y) line. If we don't fold Y into load Z, we use one
1047 /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the
1048 /// number of computations either.
1050 /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If
1051 /// X was live across 'load Z' for other reasons, we actually *would* want to
1052 /// fold the addressing mode in the Z case. This would make Y die earlier.
1053 bool AddressingModeMatcher::
1054 IsProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore,
1055 ExtAddrMode &AMAfter) {
1056 if (IgnoreProfitability || !ENABLECRAZYHACK) return true;
1058 // AMBefore is the addressing mode before this instruction was folded into it,
1059 // and AMAfter is the addressing mode after the instruction was folded. Get
1060 // the set of registers referenced by AMAfter and subtract out those
1061 // referenced by AMBefore: this is the set of values which folding in this
1062 // address extends the lifetime of.
1064 // Note that there are only two potential values being referenced here,
1065 // BaseReg and ScaleReg (global addresses are always available, as are any
1066 // folded immediates).
1067 Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg;
1069 // If the BaseReg or ScaledReg was referenced by the previous addrmode, their
1070 // lifetime wasn't extended by adding this instruction.
1071 if (ValueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg))
1073 if (ValueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg))
1076 // If folding this instruction (and it's subexprs) didn't extend any live
1077 // ranges, we're ok with it.
1078 if (BaseReg == 0 && ScaledReg == 0)
1081 // If all uses of this instruction are ultimately load/store/inlineasm's,
1082 // check to see if their addressing modes will include this instruction. If
1083 // so, we can fold it into all uses, so it doesn't matter if it has multiple
1085 SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses;
1086 SmallPtrSet<Instruction*, 16> ConsideredInsts;
1087 if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI))
1088 return false; // Has a non-memory, non-foldable use!
1090 // Now that we know that all uses of this instruction are part of a chain of
1091 // computation involving only operations that could theoretically be folded
1092 // into a memory use, loop over each of these uses and see if they could
1093 // *actually* fold the instruction.
1094 SmallVector<Instruction*, 32> MatchedAddrModeInsts;
1095 for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) {
1096 Instruction *User = MemoryUses[i].first;
1097 unsigned OpNo = MemoryUses[i].second;
1099 // Get the access type of this use. If the use isn't a pointer, we don't
1100 // know what it accesses.
1101 Value *Address = User->getOperand(OpNo);
1102 if (!isa<PointerType>(Address->getType()))
1104 const Type *AddressAccessTy =
1105 cast<PointerType>(Address->getType())->getElementType();
1107 // Do a match against the root of this address, ignoring profitability. This
1108 // will tell us if the addressing mode for the memory operation will
1109 // *actually* cover the shared instruction.
1111 AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, AddressAccessTy,
1112 MemoryInst, Result);
1113 Matcher.IgnoreProfitability = true;
1114 bool Success = Matcher.MatchAddr(Address, 0);
1115 Success = Success; assert(Success && "Couldn't select *anything*?");
1117 // If the match didn't cover I, then it won't be shared by it.
1118 if (std::find(MatchedAddrModeInsts.begin(), MatchedAddrModeInsts.end(),
1119 I) == MatchedAddrModeInsts.end())
1122 MatchedAddrModeInsts.clear();
1129 //===----------------------------------------------------------------------===//
1130 // Memory Optimization
1131 //===----------------------------------------------------------------------===//
1133 /// IsNonLocalValue - Return true if the specified values are defined in a
1134 /// different basic block than BB.
1135 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
1136 if (Instruction *I = dyn_cast<Instruction>(V))
1137 return I->getParent() != BB;
1141 /// OptimizeMemoryInst - Load and Store Instructions have often have
1142 /// addressing modes that can do significant amounts of computation. As such,
1143 /// instruction selection will try to get the load or store to do as much
1144 /// computation as possible for the program. The problem is that isel can only
1145 /// see within a single block. As such, we sink as much legal addressing mode
1146 /// stuff into the block as possible.
1148 /// This method is used to optimize both load/store and inline asms with memory
1150 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
1151 const Type *AccessTy,
1152 DenseMap<Value*,Value*> &SunkAddrs) {
1153 // Figure out what addressing mode will be built up for this operation.
1154 SmallVector<Instruction*, 16> AddrModeInsts;
1155 ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst,
1156 AddrModeInsts, *TLI);
1158 // Check to see if any of the instructions supersumed by this addr mode are
1159 // non-local to I's BB.
1160 bool AnyNonLocal = false;
1161 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
1162 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
1168 // If all the instructions matched are already in this BB, don't do anything.
1170 DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n");
1174 // Insert this computation right after this user. Since our caller is
1175 // scanning from the top of the BB to the bottom, reuse of the expr are
1176 // guaranteed to happen later.
1177 BasicBlock::iterator InsertPt = MemoryInst;
1179 // Now that we determined the addressing expression we want to use and know
1180 // that we have to sink it into this block. Check to see if we have already
1181 // done this for some other load/store instr in this block. If so, reuse the
1183 Value *&SunkAddr = SunkAddrs[Addr];
1185 DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << "\n");
1186 if (SunkAddr->getType() != Addr->getType())
1187 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
1189 DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << "\n");
1190 const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType();
1193 // Start with the scale value.
1194 if (AddrMode.Scale) {
1195 Value *V = AddrMode.ScaledReg;
1196 if (V->getType() == IntPtrTy) {
1198 } else if (isa<PointerType>(V->getType())) {
1199 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
1200 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
1201 cast<IntegerType>(V->getType())->getBitWidth()) {
1202 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
1204 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
1206 if (AddrMode.Scale != 1)
1207 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
1209 "sunkaddr", InsertPt);
1213 // Add in the base register.
1214 if (AddrMode.BaseReg) {
1215 Value *V = AddrMode.BaseReg;
1216 if (V->getType() != IntPtrTy)
1217 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
1219 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
1224 // Add in the BaseGV if present.
1225 if (AddrMode.BaseGV) {
1226 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
1229 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
1234 // Add in the Base Offset if present.
1235 if (AddrMode.BaseOffs) {
1236 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
1238 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
1244 SunkAddr = Constant::getNullValue(Addr->getType());
1246 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
1249 MemoryInst->replaceUsesOfWith(Addr, SunkAddr);
1251 if (Addr->use_empty())
1252 EraseDeadInstructions(Addr);
1256 /// OptimizeInlineAsmInst - If there are any memory operands, use
1257 /// OptimizeMemoryInst to sink their address computing into the block when
1258 /// possible / profitable.
1259 bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS,
1260 DenseMap<Value*,Value*> &SunkAddrs) {
1261 bool MadeChange = false;
1262 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
1264 // Do a prepass over the constraints, canonicalizing them, and building up the
1265 // ConstraintOperands list.
1266 std::vector<InlineAsm::ConstraintInfo>
1267 ConstraintInfos = IA->ParseConstraints();
1269 /// ConstraintOperands - Information about all of the constraints.
1270 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands;
1271 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
1272 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
1274 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i]));
1275 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back();
1277 // Compute the value type for each operand.
1278 switch (OpInfo.Type) {
1279 case InlineAsm::isOutput:
1280 if (OpInfo.isIndirect)
1281 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
1283 case InlineAsm::isInput:
1284 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
1286 case InlineAsm::isClobber:
1291 // Compute the constraint code and ConstraintType to use.
1292 TLI->ComputeConstraintToUse(OpInfo, SDValue(),
1293 OpInfo.ConstraintType == TargetLowering::C_Memory);
1295 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
1296 OpInfo.isIndirect) {
1297 Value *OpVal = OpInfo.CallOperandVal;
1298 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs);
1305 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
1306 BasicBlock *DefBB = I->getParent();
1308 // If both result of the {s|z}xt and its source are live out, rewrite all
1309 // other uses of the source with result of extension.
1310 Value *Src = I->getOperand(0);
1311 if (Src->hasOneUse())
1314 // Only do this xform if truncating is free.
1315 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
1318 // Only safe to perform the optimization if the source is also defined in
1320 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
1323 bool DefIsLiveOut = false;
1324 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1326 Instruction *User = cast<Instruction>(*UI);
1328 // Figure out which BB this ext is used in.
1329 BasicBlock *UserBB = User->getParent();
1330 if (UserBB == DefBB) continue;
1331 DefIsLiveOut = true;
1337 // Make sure non of the uses are PHI nodes.
1338 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1340 Instruction *User = cast<Instruction>(*UI);
1341 BasicBlock *UserBB = User->getParent();
1342 if (UserBB == DefBB) continue;
1343 // Be conservative. We don't want this xform to end up introducing
1344 // reloads just before load / store instructions.
1345 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
1349 // InsertedTruncs - Only insert one trunc in each block once.
1350 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
1352 bool MadeChange = false;
1353 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1355 Use &TheUse = UI.getUse();
1356 Instruction *User = cast<Instruction>(*UI);
1358 // Figure out which BB this ext is used in.
1359 BasicBlock *UserBB = User->getParent();
1360 if (UserBB == DefBB) continue;
1362 // Both src and def are live in this block. Rewrite the use.
1363 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
1365 if (!InsertedTrunc) {
1366 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
1368 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
1371 // Replace a use of the {s|z}ext source with a use of the result.
1372 TheUse = InsertedTrunc;
1380 // In this pass we look for GEP and cast instructions that are used
1381 // across basic blocks and rewrite them to improve basic-block-at-a-time
1383 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
1384 bool MadeChange = false;
1386 // Split all critical edges where the dest block has a PHI and where the phi
1387 // has shared immediate operands.
1388 TerminatorInst *BBTI = BB.getTerminator();
1389 if (BBTI->getNumSuccessors() > 1) {
1390 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i)
1391 if (isa<PHINode>(BBTI->getSuccessor(i)->begin()) &&
1392 isCriticalEdge(BBTI, i, true))
1393 SplitEdgeNicely(BBTI, i, this);
1397 // Keep track of non-local addresses that have been sunk into this block.
1398 // This allows us to avoid inserting duplicate code for blocks with multiple
1399 // load/stores of the same address.
1400 DenseMap<Value*, Value*> SunkAddrs;
1402 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) {
1403 Instruction *I = BBI++;
1405 if (CastInst *CI = dyn_cast<CastInst>(I)) {
1406 // If the source of the cast is a constant, then this should have
1407 // already been constant folded. The only reason NOT to constant fold
1408 // it is if something (e.g. LSR) was careful to place the constant
1409 // evaluation in a block other than then one that uses it (e.g. to hoist
1410 // the address of globals out of a loop). If this is the case, we don't
1411 // want to forward-subst the cast.
1412 if (isa<Constant>(CI->getOperand(0)))
1415 bool Change = false;
1417 Change = OptimizeNoopCopyExpression(CI, *TLI);
1418 MadeChange |= Change;
1421 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I)))
1422 MadeChange |= OptimizeExtUses(I);
1423 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
1424 MadeChange |= OptimizeCmpExpression(CI);
1425 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1427 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(),
1429 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1431 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1),
1432 SI->getOperand(0)->getType(),
1434 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1435 if (GEPI->hasAllZeroIndices()) {
1436 /// The GEP operand must be a pointer, so must its result -> BitCast
1437 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1438 GEPI->getName(), GEPI);
1439 GEPI->replaceAllUsesWith(NC);
1440 GEPI->eraseFromParent();
1444 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
1445 // If we found an inline asm expession, and if the target knows how to
1446 // lower it to normal LLVM code, do so now.
1447 if (TLI && isa<InlineAsm>(CI->getCalledValue()))
1448 if (const TargetAsmInfo *TAI =
1449 TLI->getTargetMachine().getTargetAsmInfo()) {
1450 if (TAI->ExpandInlineAsm(CI))
1453 // Sink address computing for memory operands into the block.
1454 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs);