1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "codegenprepare"
17 #include "llvm/Transforms/Scalar.h"
18 #include "llvm/Constants.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Function.h"
21 #include "llvm/InlineAsm.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/IntrinsicInst.h"
24 #include "llvm/Pass.h"
25 #include "llvm/Analysis/ProfileInfo.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Target/TargetLowering.h"
28 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
29 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
30 #include "llvm/Transforms/Utils/Local.h"
31 #include "llvm/Transforms/Utils/BuildLibCalls.h"
32 #include "llvm/ADT/DenseMap.h"
33 #include "llvm/ADT/SmallSet.h"
34 #include "llvm/Assembly/Writer.h"
35 #include "llvm/Support/CallSite.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/GetElementPtrTypeIterator.h"
38 #include "llvm/Support/PatternMatch.h"
39 #include "llvm/Support/raw_ostream.h"
40 #include "llvm/Support/IRBuilder.h"
42 using namespace llvm::PatternMatch;
45 class CodeGenPrepare : public FunctionPass {
46 /// TLI - Keep a pointer of a TargetLowering to consult for determining
47 /// transformation profitability.
48 const TargetLowering *TLI;
51 /// BackEdges - Keep a set of all the loop back edges.
53 SmallSet<std::pair<const BasicBlock*, const BasicBlock*>, 8> BackEdges;
55 static char ID; // Pass identification, replacement for typeid
56 explicit CodeGenPrepare(const TargetLowering *tli = 0)
57 : FunctionPass(ID), TLI(tli) {}
58 bool runOnFunction(Function &F);
60 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
61 AU.addPreserved<ProfileInfo>();
64 virtual void releaseMemory() {
69 bool EliminateMostlyEmptyBlocks(Function &F);
70 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
71 void EliminateMostlyEmptyBlock(BasicBlock *BB);
72 bool OptimizeBlock(BasicBlock &BB);
73 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy,
74 DenseMap<Value*,Value*> &SunkAddrs);
75 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS,
76 DenseMap<Value*,Value*> &SunkAddrs);
77 bool OptimizeCallInst(CallInst *CI);
78 bool MoveExtToFormExtLoad(Instruction *I);
79 bool OptimizeExtUses(Instruction *I);
80 void findLoopBackEdges(const Function &F);
84 char CodeGenPrepare::ID = 0;
85 INITIALIZE_PASS(CodeGenPrepare, "codegenprepare",
86 "Optimize for code generation", false, false);
88 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
89 return new CodeGenPrepare(TLI);
92 /// findLoopBackEdges - Do a DFS walk to find loop back edges.
94 void CodeGenPrepare::findLoopBackEdges(const Function &F) {
95 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
96 FindFunctionBackedges(F, Edges);
98 BackEdges.insert(Edges.begin(), Edges.end());
102 bool CodeGenPrepare::runOnFunction(Function &F) {
103 bool EverMadeChange = false;
105 PFI = getAnalysisIfAvailable<ProfileInfo>();
106 // First pass, eliminate blocks that contain only PHI nodes and an
107 // unconditional branch.
108 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
110 // Now find loop back edges.
111 findLoopBackEdges(F);
113 bool MadeChange = true;
116 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
117 MadeChange |= OptimizeBlock(*BB);
118 EverMadeChange |= MadeChange;
120 return EverMadeChange;
123 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
124 /// debug info directives, and an unconditional branch. Passes before isel
125 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
126 /// isel. Start by eliminating these blocks so we can split them the way we
128 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
129 bool MadeChange = false;
130 // Note that this intentionally skips the entry block.
131 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
132 BasicBlock *BB = I++;
134 // If this block doesn't end with an uncond branch, ignore it.
135 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
136 if (!BI || !BI->isUnconditional())
139 // If the instruction before the branch (skipping debug info) isn't a phi
140 // node, then other stuff is happening here.
141 BasicBlock::iterator BBI = BI;
142 if (BBI != BB->begin()) {
144 while (isa<DbgInfoIntrinsic>(BBI)) {
145 if (BBI == BB->begin())
149 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
153 // Do not break infinite loops.
154 BasicBlock *DestBB = BI->getSuccessor(0);
158 if (!CanMergeBlocks(BB, DestBB))
161 EliminateMostlyEmptyBlock(BB);
167 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
168 /// single uncond branch between them, and BB contains no other non-phi
170 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
171 const BasicBlock *DestBB) const {
172 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
173 // the successor. If there are more complex condition (e.g. preheaders),
174 // don't mess around with them.
175 BasicBlock::const_iterator BBI = BB->begin();
176 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
177 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
179 const Instruction *User = cast<Instruction>(*UI);
180 if (User->getParent() != DestBB || !isa<PHINode>(User))
182 // If User is inside DestBB block and it is a PHINode then check
183 // incoming value. If incoming value is not from BB then this is
184 // a complex condition (e.g. preheaders) we want to avoid here.
185 if (User->getParent() == DestBB) {
186 if (const PHINode *UPN = dyn_cast<PHINode>(User))
187 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
188 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
189 if (Insn && Insn->getParent() == BB &&
190 Insn->getParent() != UPN->getIncomingBlock(I))
197 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
198 // and DestBB may have conflicting incoming values for the block. If so, we
199 // can't merge the block.
200 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
201 if (!DestBBPN) return true; // no conflict.
203 // Collect the preds of BB.
204 SmallPtrSet<const BasicBlock*, 16> BBPreds;
205 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
206 // It is faster to get preds from a PHI than with pred_iterator.
207 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
208 BBPreds.insert(BBPN->getIncomingBlock(i));
210 BBPreds.insert(pred_begin(BB), pred_end(BB));
213 // Walk the preds of DestBB.
214 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
215 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
216 if (BBPreds.count(Pred)) { // Common predecessor?
217 BBI = DestBB->begin();
218 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
219 const Value *V1 = PN->getIncomingValueForBlock(Pred);
220 const Value *V2 = PN->getIncomingValueForBlock(BB);
222 // If V2 is a phi node in BB, look up what the mapped value will be.
223 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
224 if (V2PN->getParent() == BB)
225 V2 = V2PN->getIncomingValueForBlock(Pred);
227 // If there is a conflict, bail out.
228 if (V1 != V2) return false;
237 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
238 /// an unconditional branch in it.
239 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
240 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
241 BasicBlock *DestBB = BI->getSuccessor(0);
243 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
245 // If the destination block has a single pred, then this is a trivial edge,
247 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
248 if (SinglePred != DestBB) {
249 // Remember if SinglePred was the entry block of the function. If so, we
250 // will need to move BB back to the entry position.
251 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
252 MergeBasicBlockIntoOnlyPred(DestBB, this);
254 if (isEntry && BB != &BB->getParent()->getEntryBlock())
255 BB->moveBefore(&BB->getParent()->getEntryBlock());
257 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
262 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
263 // to handle the new incoming edges it is about to have.
265 for (BasicBlock::iterator BBI = DestBB->begin();
266 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
267 // Remove the incoming value for BB, and remember it.
268 Value *InVal = PN->removeIncomingValue(BB, false);
270 // Two options: either the InVal is a phi node defined in BB or it is some
271 // value that dominates BB.
272 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
273 if (InValPhi && InValPhi->getParent() == BB) {
274 // Add all of the input values of the input PHI as inputs of this phi.
275 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
276 PN->addIncoming(InValPhi->getIncomingValue(i),
277 InValPhi->getIncomingBlock(i));
279 // Otherwise, add one instance of the dominating value for each edge that
280 // we will be adding.
281 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
282 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
283 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
285 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
286 PN->addIncoming(InVal, *PI);
291 // The PHIs are now updated, change everything that refers to BB to use
292 // DestBB and remove BB.
293 BB->replaceAllUsesWith(DestBB);
295 PFI->replaceAllUses(BB, DestBB);
296 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
298 BB->eraseFromParent();
300 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
303 /// FindReusablePredBB - Check all of the predecessors of the block DestPHI
304 /// lives in to see if there is a block that we can reuse as a critical edge
306 static BasicBlock *FindReusablePredBB(PHINode *DestPHI, BasicBlock *TIBB) {
307 BasicBlock *Dest = DestPHI->getParent();
309 /// TIPHIValues - This array is lazily computed to determine the values of
310 /// PHIs in Dest that TI would provide.
311 SmallVector<Value*, 32> TIPHIValues;
313 /// TIBBEntryNo - This is a cache to speed up pred queries for TIBB.
314 unsigned TIBBEntryNo = 0;
316 // Check to see if Dest has any blocks that can be used as a split edge for
318 for (unsigned pi = 0, e = DestPHI->getNumIncomingValues(); pi != e; ++pi) {
319 BasicBlock *Pred = DestPHI->getIncomingBlock(pi);
320 // To be usable, the pred has to end with an uncond branch to the dest.
321 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator());
322 if (!PredBr || !PredBr->isUnconditional())
324 // Must be empty other than the branch and debug info.
325 BasicBlock::iterator I = Pred->begin();
326 while (isa<DbgInfoIntrinsic>(I))
330 // Cannot be the entry block; its label does not get emitted.
331 if (Pred == &Dest->getParent()->getEntryBlock())
334 // Finally, since we know that Dest has phi nodes in it, we have to make
335 // sure that jumping to Pred will have the same effect as going to Dest in
336 // terms of PHI values.
339 unsigned PredEntryNo = pi;
341 bool FoundMatch = true;
342 for (BasicBlock::iterator I = Dest->begin();
343 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) {
344 if (PHINo == TIPHIValues.size()) {
345 if (PN->getIncomingBlock(TIBBEntryNo) != TIBB)
346 TIBBEntryNo = PN->getBasicBlockIndex(TIBB);
347 TIPHIValues.push_back(PN->getIncomingValue(TIBBEntryNo));
350 // If the PHI entry doesn't work, we can't use this pred.
351 if (PN->getIncomingBlock(PredEntryNo) != Pred)
352 PredEntryNo = PN->getBasicBlockIndex(Pred);
354 if (TIPHIValues[PHINo] != PN->getIncomingValue(PredEntryNo)) {
360 // If we found a workable predecessor, change TI to branch to Succ.
368 /// SplitEdgeNicely - Split the critical edge from TI to its specified
369 /// successor if it will improve codegen. We only do this if the successor has
370 /// phi nodes (otherwise critical edges are ok). If there is already another
371 /// predecessor of the succ that is empty (and thus has no phi nodes), use it
372 /// instead of introducing a new block.
373 static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum,
374 SmallSet<std::pair<const BasicBlock*,
375 const BasicBlock*>, 8> &BackEdges,
377 BasicBlock *TIBB = TI->getParent();
378 BasicBlock *Dest = TI->getSuccessor(SuccNum);
379 assert(isa<PHINode>(Dest->begin()) &&
380 "This should only be called if Dest has a PHI!");
381 PHINode *DestPHI = cast<PHINode>(Dest->begin());
383 // Do not split edges to EH landing pads.
384 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI))
385 if (Invoke->getSuccessor(1) == Dest)
388 // As a hack, never split backedges of loops. Even though the copy for any
389 // PHIs inserted on the backedge would be dead for exits from the loop, we
390 // assume that the cost of *splitting* the backedge would be too high.
391 if (BackEdges.count(std::make_pair(TIBB, Dest)))
394 if (BasicBlock *ReuseBB = FindReusablePredBB(DestPHI, TIBB)) {
395 ProfileInfo *PFI = P->getAnalysisIfAvailable<ProfileInfo>();
397 PFI->splitEdge(TIBB, Dest, ReuseBB);
398 Dest->removePredecessor(TIBB);
399 TI->setSuccessor(SuccNum, ReuseBB);
403 SplitCriticalEdge(TI, SuccNum, P, true);
407 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
408 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
409 /// sink it into user blocks to reduce the number of virtual
410 /// registers that must be created and coalesced.
412 /// Return true if any changes are made.
414 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
415 // If this is a noop copy,
416 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
417 EVT DstVT = TLI.getValueType(CI->getType());
419 // This is an fp<->int conversion?
420 if (SrcVT.isInteger() != DstVT.isInteger())
423 // If this is an extension, it will be a zero or sign extension, which
425 if (SrcVT.bitsLT(DstVT)) return false;
427 // If these values will be promoted, find out what they will be promoted
428 // to. This helps us consider truncates on PPC as noop copies when they
430 if (TLI.getTypeAction(CI->getContext(), SrcVT) == TargetLowering::Promote)
431 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
432 if (TLI.getTypeAction(CI->getContext(), DstVT) == TargetLowering::Promote)
433 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
435 // If, after promotion, these are the same types, this is a noop copy.
439 BasicBlock *DefBB = CI->getParent();
441 /// InsertedCasts - Only insert a cast in each block once.
442 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
444 bool MadeChange = false;
445 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
447 Use &TheUse = UI.getUse();
448 Instruction *User = cast<Instruction>(*UI);
450 // Figure out which BB this cast is used in. For PHI's this is the
451 // appropriate predecessor block.
452 BasicBlock *UserBB = User->getParent();
453 if (PHINode *PN = dyn_cast<PHINode>(User)) {
454 UserBB = PN->getIncomingBlock(UI);
457 // Preincrement use iterator so we don't invalidate it.
460 // If this user is in the same block as the cast, don't change the cast.
461 if (UserBB == DefBB) continue;
463 // If we have already inserted a cast into this block, use it.
464 CastInst *&InsertedCast = InsertedCasts[UserBB];
467 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
470 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
475 // Replace a use of the cast with a use of the new cast.
476 TheUse = InsertedCast;
479 // If we removed all uses, nuke the cast.
480 if (CI->use_empty()) {
481 CI->eraseFromParent();
488 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
489 /// the number of virtual registers that must be created and coalesced. This is
490 /// a clear win except on targets with multiple condition code registers
491 /// (PowerPC), where it might lose; some adjustment may be wanted there.
493 /// Return true if any changes are made.
494 static bool OptimizeCmpExpression(CmpInst *CI) {
495 BasicBlock *DefBB = CI->getParent();
497 /// InsertedCmp - Only insert a cmp in each block once.
498 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
500 bool MadeChange = false;
501 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
503 Use &TheUse = UI.getUse();
504 Instruction *User = cast<Instruction>(*UI);
506 // Preincrement use iterator so we don't invalidate it.
509 // Don't bother for PHI nodes.
510 if (isa<PHINode>(User))
513 // Figure out which BB this cmp is used in.
514 BasicBlock *UserBB = User->getParent();
516 // If this user is in the same block as the cmp, don't change the cmp.
517 if (UserBB == DefBB) continue;
519 // If we have already inserted a cmp into this block, use it.
520 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
523 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
526 CmpInst::Create(CI->getOpcode(),
527 CI->getPredicate(), CI->getOperand(0),
528 CI->getOperand(1), "", InsertPt);
532 // Replace a use of the cmp with a use of the new cmp.
533 TheUse = InsertedCmp;
536 // If we removed all uses, nuke the cmp.
538 CI->eraseFromParent();
544 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
546 void replaceCall(Value *With) {
547 CI->replaceAllUsesWith(With);
548 CI->eraseFromParent();
550 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
551 if (ConstantInt *SizeCI =
552 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
553 return SizeCI->isAllOnesValue();
557 } // end anonymous namespace
559 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
560 // Lower all uses of llvm.objectsize.*
561 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
562 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
563 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
564 const Type *ReturnTy = CI->getType();
565 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
566 CI->replaceAllUsesWith(RetVal);
567 CI->eraseFromParent();
571 // From here on out we're working with named functions.
572 if (CI->getCalledFunction() == 0) return false;
574 // We'll need TargetData from here on out.
575 const TargetData *TD = TLI ? TLI->getTargetData() : 0;
576 if (!TD) return false;
578 // Lower all default uses of _chk calls. This is very similar
579 // to what InstCombineCalls does, but here we are only lowering calls
580 // that have the default "don't know" as the objectsize. Anything else
581 // should be left alone.
582 CodeGenPrepareFortifiedLibCalls Simplifier;
583 return Simplifier.fold(CI, TD);
585 //===----------------------------------------------------------------------===//
586 // Memory Optimization
587 //===----------------------------------------------------------------------===//
589 /// IsNonLocalValue - Return true if the specified values are defined in a
590 /// different basic block than BB.
591 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
592 if (Instruction *I = dyn_cast<Instruction>(V))
593 return I->getParent() != BB;
597 /// OptimizeMemoryInst - Load and Store Instructions often have
598 /// addressing modes that can do significant amounts of computation. As such,
599 /// instruction selection will try to get the load or store to do as much
600 /// computation as possible for the program. The problem is that isel can only
601 /// see within a single block. As such, we sink as much legal addressing mode
602 /// stuff into the block as possible.
604 /// This method is used to optimize both load/store and inline asms with memory
606 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
607 const Type *AccessTy,
608 DenseMap<Value*,Value*> &SunkAddrs) {
609 // Figure out what addressing mode will be built up for this operation.
610 SmallVector<Instruction*, 16> AddrModeInsts;
611 ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst,
612 AddrModeInsts, *TLI);
614 // Check to see if any of the instructions supersumed by this addr mode are
615 // non-local to I's BB.
616 bool AnyNonLocal = false;
617 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
618 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
624 // If all the instructions matched are already in this BB, don't do anything.
626 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
630 // Insert this computation right after this user. Since our caller is
631 // scanning from the top of the BB to the bottom, reuse of the expr are
632 // guaranteed to happen later.
633 BasicBlock::iterator InsertPt = MemoryInst;
635 // Now that we determined the addressing expression we want to use and know
636 // that we have to sink it into this block. Check to see if we have already
637 // done this for some other load/store instr in this block. If so, reuse the
639 Value *&SunkAddr = SunkAddrs[Addr];
641 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
643 if (SunkAddr->getType() != Addr->getType())
644 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
646 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
648 const Type *IntPtrTy =
649 TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
653 // Start with the base register. Do this first so that subsequent address
654 // matching finds it last, which will prevent it from trying to match it
655 // as the scaled value in case it happens to be a mul. That would be
656 // problematic if we've sunk a different mul for the scale, because then
657 // we'd end up sinking both muls.
658 if (AddrMode.BaseReg) {
659 Value *V = AddrMode.BaseReg;
660 if (V->getType()->isPointerTy())
661 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
662 if (V->getType() != IntPtrTy)
663 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true,
664 "sunkaddr", InsertPt);
668 // Add the scale value.
669 if (AddrMode.Scale) {
670 Value *V = AddrMode.ScaledReg;
671 if (V->getType() == IntPtrTy) {
673 } else if (V->getType()->isPointerTy()) {
674 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
675 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
676 cast<IntegerType>(V->getType())->getBitWidth()) {
677 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
679 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
681 if (AddrMode.Scale != 1)
682 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
684 "sunkaddr", InsertPt);
686 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
691 // Add in the BaseGV if present.
692 if (AddrMode.BaseGV) {
693 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
696 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
701 // Add in the Base Offset if present.
702 if (AddrMode.BaseOffs) {
703 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
705 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
711 SunkAddr = Constant::getNullValue(Addr->getType());
713 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
716 MemoryInst->replaceUsesOfWith(Addr, SunkAddr);
718 if (Addr->use_empty()) {
719 RecursivelyDeleteTriviallyDeadInstructions(Addr);
720 // This address is now available for reassignment, so erase the table entry;
721 // we don't want to match some completely different instruction.
727 /// OptimizeInlineAsmInst - If there are any memory operands, use
728 /// OptimizeMemoryInst to sink their address computing into the block when
729 /// possible / profitable.
730 bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS,
731 DenseMap<Value*,Value*> &SunkAddrs) {
732 bool MadeChange = false;
733 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
735 // Do a prepass over the constraints, canonicalizing them, and building up the
736 // ConstraintOperands list.
737 std::vector<InlineAsm::ConstraintInfo>
738 ConstraintInfos = IA->ParseConstraints();
740 /// ConstraintOperands - Information about all of the constraints.
741 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands;
742 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
743 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
745 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i]));
746 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back();
748 // Compute the value type for each operand.
749 switch (OpInfo.Type) {
750 case InlineAsm::isOutput:
751 if (OpInfo.isIndirect)
752 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
754 case InlineAsm::isInput:
755 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
757 case InlineAsm::isClobber:
762 // Compute the constraint code and ConstraintType to use.
763 TLI->ComputeConstraintToUse(OpInfo, SDValue());
765 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
767 Value *OpVal = OpInfo.CallOperandVal;
768 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs);
775 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
776 /// basic block as the load, unless conditions are unfavorable. This allows
777 /// SelectionDAG to fold the extend into the load.
779 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
780 // Look for a load being extended.
781 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
782 if (!LI) return false;
784 // If they're already in the same block, there's nothing to do.
785 if (LI->getParent() == I->getParent())
788 // If the load has other users and the truncate is not free, this probably
790 if (!LI->hasOneUse() &&
791 TLI && !TLI->isTruncateFree(I->getType(), LI->getType()))
794 // Check whether the target supports casts folded into loads.
796 if (isa<ZExtInst>(I))
797 LType = ISD::ZEXTLOAD;
799 assert(isa<SExtInst>(I) && "Unexpected ext type!");
800 LType = ISD::SEXTLOAD;
802 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
805 // Move the extend into the same block as the load, so that SelectionDAG
807 I->removeFromParent();
812 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
813 BasicBlock *DefBB = I->getParent();
815 // If both result of the {s|z}xt and its source are live out, rewrite all
816 // other uses of the source with result of extension.
817 Value *Src = I->getOperand(0);
818 if (Src->hasOneUse())
821 // Only do this xform if truncating is free.
822 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
825 // Only safe to perform the optimization if the source is also defined in
827 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
830 bool DefIsLiveOut = false;
831 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
833 Instruction *User = cast<Instruction>(*UI);
835 // Figure out which BB this ext is used in.
836 BasicBlock *UserBB = User->getParent();
837 if (UserBB == DefBB) continue;
844 // Make sure non of the uses are PHI nodes.
845 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
847 Instruction *User = cast<Instruction>(*UI);
848 BasicBlock *UserBB = User->getParent();
849 if (UserBB == DefBB) continue;
850 // Be conservative. We don't want this xform to end up introducing
851 // reloads just before load / store instructions.
852 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
856 // InsertedTruncs - Only insert one trunc in each block once.
857 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
859 bool MadeChange = false;
860 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
862 Use &TheUse = UI.getUse();
863 Instruction *User = cast<Instruction>(*UI);
865 // Figure out which BB this ext is used in.
866 BasicBlock *UserBB = User->getParent();
867 if (UserBB == DefBB) continue;
869 // Both src and def are live in this block. Rewrite the use.
870 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
872 if (!InsertedTrunc) {
873 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
875 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
878 // Replace a use of the {s|z}ext source with a use of the result.
879 TheUse = InsertedTrunc;
887 // In this pass we look for GEP and cast instructions that are used
888 // across basic blocks and rewrite them to improve basic-block-at-a-time
890 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
891 bool MadeChange = false;
893 // Split all critical edges where the dest block has a PHI.
894 TerminatorInst *BBTI = BB.getTerminator();
895 if (BBTI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(BBTI)) {
896 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) {
897 BasicBlock *SuccBB = BBTI->getSuccessor(i);
898 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true))
899 SplitEdgeNicely(BBTI, i, BackEdges, this);
903 // Keep track of non-local addresses that have been sunk into this block.
904 // This allows us to avoid inserting duplicate code for blocks with multiple
905 // load/stores of the same address.
906 DenseMap<Value*, Value*> SunkAddrs;
908 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) {
909 Instruction *I = BBI++;
911 if (CastInst *CI = dyn_cast<CastInst>(I)) {
912 // If the source of the cast is a constant, then this should have
913 // already been constant folded. The only reason NOT to constant fold
914 // it is if something (e.g. LSR) was careful to place the constant
915 // evaluation in a block other than then one that uses it (e.g. to hoist
916 // the address of globals out of a loop). If this is the case, we don't
917 // want to forward-subst the cast.
918 if (isa<Constant>(CI->getOperand(0)))
923 Change = OptimizeNoopCopyExpression(CI, *TLI);
924 MadeChange |= Change;
927 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) {
928 MadeChange |= MoveExtToFormExtLoad(I);
929 MadeChange |= OptimizeExtUses(I);
931 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
932 MadeChange |= OptimizeCmpExpression(CI);
933 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
935 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(),
937 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
939 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1),
940 SI->getOperand(0)->getType(),
942 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
943 if (GEPI->hasAllZeroIndices()) {
944 /// The GEP operand must be a pointer, so must its result -> BitCast
945 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
946 GEPI->getName(), GEPI);
947 GEPI->replaceAllUsesWith(NC);
948 GEPI->eraseFromParent();
952 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
953 // If we found an inline asm expession, and if the target knows how to
954 // lower it to normal LLVM code, do so now.
955 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
956 if (TLI->ExpandInlineAsm(CI)) {
958 // Avoid processing instructions out of order, which could cause
959 // reuse before a value is defined.
962 // Sink address computing for memory operands into the block.
963 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs);
965 // Other CallInst optimizations that don't need to muck with the
966 // enclosing iterator here.
967 MadeChange |= OptimizeCallInst(CI);