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/ADT/DenseMap.h"
32 #include "llvm/ADT/SmallSet.h"
33 #include "llvm/Assembly/Writer.h"
34 #include "llvm/Support/CallSite.h"
35 #include "llvm/Support/CommandLine.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"
41 using namespace llvm::PatternMatch;
43 static cl::opt<bool> FactorCommonPreds("split-critical-paths-tweak",
44 cl::init(false), cl::Hidden);
47 class CodeGenPrepare : public FunctionPass {
48 /// TLI - Keep a pointer of a TargetLowering to consult for determining
49 /// transformation profitability.
50 const TargetLowering *TLI;
53 /// BackEdges - Keep a set of all the loop back edges.
55 SmallSet<std::pair<const BasicBlock*, const BasicBlock*>, 8> BackEdges;
57 static char ID; // Pass identification, replacement for typeid
58 explicit CodeGenPrepare(const TargetLowering *tli = 0)
59 : FunctionPass(&ID), TLI(tli) {}
60 bool runOnFunction(Function &F);
62 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
63 AU.addPreserved<ProfileInfo>();
66 virtual void releaseMemory() {
71 bool EliminateMostlyEmptyBlocks(Function &F);
72 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
73 void EliminateMostlyEmptyBlock(BasicBlock *BB);
74 bool OptimizeBlock(BasicBlock &BB);
75 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy,
76 DenseMap<Value*,Value*> &SunkAddrs);
77 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS,
78 DenseMap<Value*,Value*> &SunkAddrs);
79 bool MoveExtToFormExtLoad(Instruction *I);
80 bool OptimizeExtUses(Instruction *I);
81 void findLoopBackEdges(const Function &F);
85 char CodeGenPrepare::ID = 0;
86 static RegisterPass<CodeGenPrepare> X("codegenprepare",
87 "Optimize for code generation");
89 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
90 return new CodeGenPrepare(TLI);
93 /// findLoopBackEdges - Do a DFS walk to find loop back edges.
95 void CodeGenPrepare::findLoopBackEdges(const Function &F) {
96 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
97 FindFunctionBackedges(F, Edges);
99 BackEdges.insert(Edges.begin(), Edges.end());
103 bool CodeGenPrepare::runOnFunction(Function &F) {
104 bool EverMadeChange = false;
106 PFI = getAnalysisIfAvailable<ProfileInfo>();
107 // First pass, eliminate blocks that contain only PHI nodes and an
108 // unconditional branch.
109 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
111 // Now find loop back edges.
112 findLoopBackEdges(F);
114 bool MadeChange = true;
117 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
118 MadeChange |= OptimizeBlock(*BB);
119 EverMadeChange |= MadeChange;
121 return EverMadeChange;
124 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
125 /// debug info directives, and an unconditional branch. Passes before isel
126 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
127 /// isel. Start by eliminating these blocks so we can split them the way we
129 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
130 bool MadeChange = false;
131 // Note that this intentionally skips the entry block.
132 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
133 BasicBlock *BB = I++;
135 // If this block doesn't end with an uncond branch, ignore it.
136 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
137 if (!BI || !BI->isUnconditional())
140 // If the instruction before the branch (skipping debug info) isn't a phi
141 // node, then other stuff is happening here.
142 BasicBlock::iterator BBI = BI;
143 if (BBI != BB->begin()) {
145 while (isa<DbgInfoIntrinsic>(BBI)) {
146 if (BBI == BB->begin())
150 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
154 // Do not break infinite loops.
155 BasicBlock *DestBB = BI->getSuccessor(0);
159 if (!CanMergeBlocks(BB, DestBB))
162 EliminateMostlyEmptyBlock(BB);
168 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
169 /// single uncond branch between them, and BB contains no other non-phi
171 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
172 const BasicBlock *DestBB) const {
173 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
174 // the successor. If there are more complex condition (e.g. preheaders),
175 // don't mess around with them.
176 BasicBlock::const_iterator BBI = BB->begin();
177 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
178 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end();
180 const Instruction *User = cast<Instruction>(*UI);
181 if (User->getParent() != DestBB || !isa<PHINode>(User))
183 // If User is inside DestBB block and it is a PHINode then check
184 // incoming value. If incoming value is not from BB then this is
185 // a complex condition (e.g. preheaders) we want to avoid here.
186 if (User->getParent() == DestBB) {
187 if (const PHINode *UPN = dyn_cast<PHINode>(User))
188 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
189 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
190 if (Insn && Insn->getParent() == BB &&
191 Insn->getParent() != UPN->getIncomingBlock(I))
198 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
199 // and DestBB may have conflicting incoming values for the block. If so, we
200 // can't merge the block.
201 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
202 if (!DestBBPN) return true; // no conflict.
204 // Collect the preds of BB.
205 SmallPtrSet<const BasicBlock*, 16> BBPreds;
206 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
207 // It is faster to get preds from a PHI than with pred_iterator.
208 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
209 BBPreds.insert(BBPN->getIncomingBlock(i));
211 BBPreds.insert(pred_begin(BB), pred_end(BB));
214 // Walk the preds of DestBB.
215 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
216 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
217 if (BBPreds.count(Pred)) { // Common predecessor?
218 BBI = DestBB->begin();
219 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
220 const Value *V1 = PN->getIncomingValueForBlock(Pred);
221 const Value *V2 = PN->getIncomingValueForBlock(BB);
223 // If V2 is a phi node in BB, look up what the mapped value will be.
224 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
225 if (V2PN->getParent() == BB)
226 V2 = V2PN->getIncomingValueForBlock(Pred);
228 // If there is a conflict, bail out.
229 if (V1 != V2) return false;
238 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
239 /// an unconditional branch in it.
240 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
241 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
242 BasicBlock *DestBB = BI->getSuccessor(0);
244 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
246 // If the destination block has a single pred, then this is a trivial edge,
248 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
249 if (SinglePred != DestBB) {
250 // Remember if SinglePred was the entry block of the function. If so, we
251 // will need to move BB back to the entry position.
252 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
253 MergeBasicBlockIntoOnlyPred(DestBB, this);
255 if (isEntry && BB != &BB->getParent()->getEntryBlock())
256 BB->moveBefore(&BB->getParent()->getEntryBlock());
258 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
263 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
264 // to handle the new incoming edges it is about to have.
266 for (BasicBlock::iterator BBI = DestBB->begin();
267 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
268 // Remove the incoming value for BB, and remember it.
269 Value *InVal = PN->removeIncomingValue(BB, false);
271 // Two options: either the InVal is a phi node defined in BB or it is some
272 // value that dominates BB.
273 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
274 if (InValPhi && InValPhi->getParent() == BB) {
275 // Add all of the input values of the input PHI as inputs of this phi.
276 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
277 PN->addIncoming(InValPhi->getIncomingValue(i),
278 InValPhi->getIncomingBlock(i));
280 // Otherwise, add one instance of the dominating value for each edge that
281 // we will be adding.
282 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
283 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
284 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
286 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
287 PN->addIncoming(InVal, *PI);
292 // The PHIs are now updated, change everything that refers to BB to use
293 // DestBB and remove BB.
294 BB->replaceAllUsesWith(DestBB);
296 PFI->replaceAllUses(BB, DestBB);
297 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
299 BB->eraseFromParent();
301 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
304 /// FindReusablePredBB - Check all of the predecessors of the block DestPHI
305 /// lives in to see if there is a block that we can reuse as a critical edge
307 static BasicBlock *FindReusablePredBB(PHINode *DestPHI, BasicBlock *TIBB) {
308 BasicBlock *Dest = DestPHI->getParent();
310 /// TIPHIValues - This array is lazily computed to determine the values of
311 /// PHIs in Dest that TI would provide.
312 SmallVector<Value*, 32> TIPHIValues;
314 /// TIBBEntryNo - This is a cache to speed up pred queries for TIBB.
315 unsigned TIBBEntryNo = 0;
317 // Check to see if Dest has any blocks that can be used as a split edge for
319 for (unsigned pi = 0, e = DestPHI->getNumIncomingValues(); pi != e; ++pi) {
320 BasicBlock *Pred = DestPHI->getIncomingBlock(pi);
321 // To be usable, the pred has to end with an uncond branch to the dest.
322 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator());
323 if (!PredBr || !PredBr->isUnconditional())
325 // Must be empty other than the branch and debug info.
326 BasicBlock::iterator I = Pred->begin();
327 while (isa<DbgInfoIntrinsic>(I))
331 // Cannot be the entry block; its label does not get emitted.
332 if (Pred == &Dest->getParent()->getEntryBlock())
335 // Finally, since we know that Dest has phi nodes in it, we have to make
336 // sure that jumping to Pred will have the same effect as going to Dest in
337 // terms of PHI values.
340 unsigned PredEntryNo = pi;
342 bool FoundMatch = true;
343 for (BasicBlock::iterator I = Dest->begin();
344 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) {
345 if (PHINo == TIPHIValues.size()) {
346 if (PN->getIncomingBlock(TIBBEntryNo) != TIBB)
347 TIBBEntryNo = PN->getBasicBlockIndex(TIBB);
348 TIPHIValues.push_back(PN->getIncomingValue(TIBBEntryNo));
351 // If the PHI entry doesn't work, we can't use this pred.
352 if (PN->getIncomingBlock(PredEntryNo) != Pred)
353 PredEntryNo = PN->getBasicBlockIndex(Pred);
355 if (TIPHIValues[PHINo] != PN->getIncomingValue(PredEntryNo)) {
361 // If we found a workable predecessor, change TI to branch to Succ.
369 /// SplitEdgeNicely - Split the critical edge from TI to its specified
370 /// successor if it will improve codegen. We only do this if the successor has
371 /// phi nodes (otherwise critical edges are ok). If there is already another
372 /// predecessor of the succ that is empty (and thus has no phi nodes), use it
373 /// instead of introducing a new block.
374 static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum,
375 SmallSet<std::pair<const BasicBlock*,
376 const BasicBlock*>, 8> &BackEdges,
378 BasicBlock *TIBB = TI->getParent();
379 BasicBlock *Dest = TI->getSuccessor(SuccNum);
380 assert(isa<PHINode>(Dest->begin()) &&
381 "This should only be called if Dest has a PHI!");
382 PHINode *DestPHI = cast<PHINode>(Dest->begin());
384 // Do not split edges to EH landing pads.
385 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI))
386 if (Invoke->getSuccessor(1) == Dest)
389 // As a hack, never split backedges of loops. Even though the copy for any
390 // PHIs inserted on the backedge would be dead for exits from the loop, we
391 // assume that the cost of *splitting* the backedge would be too high.
392 if (BackEdges.count(std::make_pair(TIBB, Dest)))
395 if (!FactorCommonPreds) {
396 if (BasicBlock *ReuseBB = FindReusablePredBB(DestPHI, TIBB)) {
397 ProfileInfo *PFI = P->getAnalysisIfAvailable<ProfileInfo>();
399 PFI->splitEdge(TIBB, Dest, ReuseBB);
400 Dest->removePredecessor(TIBB);
401 TI->setSuccessor(SuccNum, ReuseBB);
405 SplitCriticalEdge(TI, SuccNum, P, true);
410 SmallVector<Value*, 8> TIPHIValues;
411 for (BasicBlock::iterator I = Dest->begin();
412 (PN = dyn_cast<PHINode>(I)); ++I)
413 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB));
415 SmallVector<BasicBlock*, 8> IdenticalPreds;
417 for (unsigned pi = 0, e = DestPHI->getNumIncomingValues(); pi != e; ++pi) {
418 BasicBlock *Pred = DestPHI->getIncomingBlock(pi);
419 if (BackEdges.count(std::make_pair(Pred, Dest)))
422 IdenticalPreds.push_back(Pred);
425 bool Identical = true;
427 for (BasicBlock::iterator I = Dest->begin();
428 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo)
429 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) {
434 IdenticalPreds.push_back(Pred);
437 assert(!IdenticalPreds.empty());
438 SplitBlockPredecessors(Dest, &IdenticalPreds[0], IdenticalPreds.size(),
443 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
444 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
445 /// sink it into user blocks to reduce the number of virtual
446 /// registers that must be created and coalesced.
448 /// Return true if any changes are made.
450 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
451 // If this is a noop copy,
452 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
453 EVT DstVT = TLI.getValueType(CI->getType());
455 // This is an fp<->int conversion?
456 if (SrcVT.isInteger() != DstVT.isInteger())
459 // If this is an extension, it will be a zero or sign extension, which
461 if (SrcVT.bitsLT(DstVT)) return false;
463 // If these values will be promoted, find out what they will be promoted
464 // to. This helps us consider truncates on PPC as noop copies when they
466 if (TLI.getTypeAction(CI->getContext(), SrcVT) == TargetLowering::Promote)
467 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
468 if (TLI.getTypeAction(CI->getContext(), DstVT) == TargetLowering::Promote)
469 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
471 // If, after promotion, these are the same types, this is a noop copy.
475 BasicBlock *DefBB = CI->getParent();
477 /// InsertedCasts - Only insert a cast in each block once.
478 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
480 bool MadeChange = false;
481 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
483 Use &TheUse = UI.getUse();
484 Instruction *User = cast<Instruction>(*UI);
486 // Figure out which BB this cast is used in. For PHI's this is the
487 // appropriate predecessor block.
488 BasicBlock *UserBB = User->getParent();
489 if (PHINode *PN = dyn_cast<PHINode>(User)) {
490 UserBB = PN->getIncomingBlock(UI);
493 // Preincrement use iterator so we don't invalidate it.
496 // If this user is in the same block as the cast, don't change the cast.
497 if (UserBB == DefBB) continue;
499 // If we have already inserted a cast into this block, use it.
500 CastInst *&InsertedCast = InsertedCasts[UserBB];
503 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
506 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
511 // Replace a use of the cast with a use of the new cast.
512 TheUse = InsertedCast;
515 // If we removed all uses, nuke the cast.
516 if (CI->use_empty()) {
517 CI->eraseFromParent();
524 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
525 /// the number of virtual registers that must be created and coalesced. This is
526 /// a clear win except on targets with multiple condition code registers
527 /// (PowerPC), where it might lose; some adjustment may be wanted there.
529 /// Return true if any changes are made.
530 static bool OptimizeCmpExpression(CmpInst *CI) {
531 BasicBlock *DefBB = CI->getParent();
533 /// InsertedCmp - Only insert a cmp in each block once.
534 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
536 bool MadeChange = false;
537 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
539 Use &TheUse = UI.getUse();
540 Instruction *User = cast<Instruction>(*UI);
542 // Preincrement use iterator so we don't invalidate it.
545 // Don't bother for PHI nodes.
546 if (isa<PHINode>(User))
549 // Figure out which BB this cmp is used in.
550 BasicBlock *UserBB = User->getParent();
552 // If this user is in the same block as the cmp, don't change the cmp.
553 if (UserBB == DefBB) continue;
555 // If we have already inserted a cmp into this block, use it.
556 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
559 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
562 CmpInst::Create(CI->getOpcode(),
563 CI->getPredicate(), CI->getOperand(0),
564 CI->getOperand(1), "", InsertPt);
568 // Replace a use of the cmp with a use of the new cmp.
569 TheUse = InsertedCmp;
572 // If we removed all uses, nuke the cmp.
574 CI->eraseFromParent();
579 //===----------------------------------------------------------------------===//
580 // Memory Optimization
581 //===----------------------------------------------------------------------===//
583 /// IsNonLocalValue - Return true if the specified values are defined in a
584 /// different basic block than BB.
585 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
586 if (Instruction *I = dyn_cast<Instruction>(V))
587 return I->getParent() != BB;
591 /// OptimizeMemoryInst - Load and Store Instructions often have
592 /// addressing modes that can do significant amounts of computation. As such,
593 /// instruction selection will try to get the load or store to do as much
594 /// computation as possible for the program. The problem is that isel can only
595 /// see within a single block. As such, we sink as much legal addressing mode
596 /// stuff into the block as possible.
598 /// This method is used to optimize both load/store and inline asms with memory
600 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
601 const Type *AccessTy,
602 DenseMap<Value*,Value*> &SunkAddrs) {
603 // Figure out what addressing mode will be built up for this operation.
604 SmallVector<Instruction*, 16> AddrModeInsts;
605 ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst,
606 AddrModeInsts, *TLI);
608 // Check to see if any of the instructions supersumed by this addr mode are
609 // non-local to I's BB.
610 bool AnyNonLocal = false;
611 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
612 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
618 // If all the instructions matched are already in this BB, don't do anything.
620 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
624 // Insert this computation right after this user. Since our caller is
625 // scanning from the top of the BB to the bottom, reuse of the expr are
626 // guaranteed to happen later.
627 BasicBlock::iterator InsertPt = MemoryInst;
629 // Now that we determined the addressing expression we want to use and know
630 // that we have to sink it into this block. Check to see if we have already
631 // done this for some other load/store instr in this block. If so, reuse the
633 Value *&SunkAddr = SunkAddrs[Addr];
635 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
637 if (SunkAddr->getType() != Addr->getType())
638 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
640 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
642 const Type *IntPtrTy =
643 TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
647 // Start with the base register. Do this first so that subsequent address
648 // matching finds it last, which will prevent it from trying to match it
649 // as the scaled value in case it happens to be a mul. That would be
650 // problematic if we've sunk a different mul for the scale, because then
651 // we'd end up sinking both muls.
652 if (AddrMode.BaseReg) {
653 Value *V = AddrMode.BaseReg;
654 if (isa<PointerType>(V->getType()))
655 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
656 if (V->getType() != IntPtrTy)
657 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true,
658 "sunkaddr", InsertPt);
662 // Add the scale value.
663 if (AddrMode.Scale) {
664 Value *V = AddrMode.ScaledReg;
665 if (V->getType() == IntPtrTy) {
667 } else if (isa<PointerType>(V->getType())) {
668 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
669 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
670 cast<IntegerType>(V->getType())->getBitWidth()) {
671 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
673 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
675 if (AddrMode.Scale != 1)
676 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
678 "sunkaddr", InsertPt);
680 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
685 // Add in the BaseGV if present.
686 if (AddrMode.BaseGV) {
687 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
690 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
695 // Add in the Base Offset if present.
696 if (AddrMode.BaseOffs) {
697 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
699 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
705 SunkAddr = Constant::getNullValue(Addr->getType());
707 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
710 MemoryInst->replaceUsesOfWith(Addr, SunkAddr);
712 if (Addr->use_empty())
713 RecursivelyDeleteTriviallyDeadInstructions(Addr);
717 /// OptimizeInlineAsmInst - If there are any memory operands, use
718 /// OptimizeMemoryInst to sink their address computing into the block when
719 /// possible / profitable.
720 bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS,
721 DenseMap<Value*,Value*> &SunkAddrs) {
722 bool MadeChange = false;
723 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
725 // Do a prepass over the constraints, canonicalizing them, and building up the
726 // ConstraintOperands list.
727 std::vector<InlineAsm::ConstraintInfo>
728 ConstraintInfos = IA->ParseConstraints();
730 /// ConstraintOperands - Information about all of the constraints.
731 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands;
732 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
733 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
735 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i]));
736 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back();
738 // Compute the value type for each operand.
739 switch (OpInfo.Type) {
740 case InlineAsm::isOutput:
741 if (OpInfo.isIndirect)
742 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
744 case InlineAsm::isInput:
745 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
747 case InlineAsm::isClobber:
752 // Compute the constraint code and ConstraintType to use.
753 TLI->ComputeConstraintToUse(OpInfo, SDValue(),
754 OpInfo.ConstraintType == TargetLowering::C_Memory);
756 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
758 Value *OpVal = OpInfo.CallOperandVal;
759 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs);
766 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
767 /// basic block as the load, unless conditions are unfavorable. This allows
768 /// SelectionDAG to fold the extend into the load.
770 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
771 // Look for a load being extended.
772 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
773 if (!LI) return false;
775 // If they're already in the same block, there's nothing to do.
776 if (LI->getParent() == I->getParent())
779 // If the load has other users and the truncate is not free, this probably
781 if (!LI->hasOneUse() &&
782 TLI && !TLI->isTruncateFree(I->getType(), LI->getType()))
785 // Check whether the target supports casts folded into loads.
787 if (isa<ZExtInst>(I))
788 LType = ISD::ZEXTLOAD;
790 assert(isa<SExtInst>(I) && "Unexpected ext type!");
791 LType = ISD::SEXTLOAD;
793 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
796 // Move the extend into the same block as the load, so that SelectionDAG
798 I->removeFromParent();
803 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
804 BasicBlock *DefBB = I->getParent();
806 // If both result of the {s|z}xt and its source are live out, rewrite all
807 // other uses of the source with result of extension.
808 Value *Src = I->getOperand(0);
809 if (Src->hasOneUse())
812 // Only do this xform if truncating is free.
813 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
816 // Only safe to perform the optimization if the source is also defined in
818 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
821 bool DefIsLiveOut = false;
822 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
824 Instruction *User = cast<Instruction>(*UI);
826 // Figure out which BB this ext is used in.
827 BasicBlock *UserBB = User->getParent();
828 if (UserBB == DefBB) continue;
835 // Make sure non of the uses are PHI nodes.
836 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
838 Instruction *User = cast<Instruction>(*UI);
839 BasicBlock *UserBB = User->getParent();
840 if (UserBB == DefBB) continue;
841 // Be conservative. We don't want this xform to end up introducing
842 // reloads just before load / store instructions.
843 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
847 // InsertedTruncs - Only insert one trunc in each block once.
848 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
850 bool MadeChange = false;
851 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
853 Use &TheUse = UI.getUse();
854 Instruction *User = cast<Instruction>(*UI);
856 // Figure out which BB this ext is used in.
857 BasicBlock *UserBB = User->getParent();
858 if (UserBB == DefBB) continue;
860 // Both src and def are live in this block. Rewrite the use.
861 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
863 if (!InsertedTrunc) {
864 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
866 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
869 // Replace a use of the {s|z}ext source with a use of the result.
870 TheUse = InsertedTrunc;
878 // In this pass we look for GEP and cast instructions that are used
879 // across basic blocks and rewrite them to improve basic-block-at-a-time
881 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
882 bool MadeChange = false;
884 // Split all critical edges where the dest block has a PHI.
885 TerminatorInst *BBTI = BB.getTerminator();
886 if (BBTI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(BBTI)) {
887 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) {
888 BasicBlock *SuccBB = BBTI->getSuccessor(i);
889 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true))
890 SplitEdgeNicely(BBTI, i, BackEdges, this);
894 // Keep track of non-local addresses that have been sunk into this block.
895 // This allows us to avoid inserting duplicate code for blocks with multiple
896 // load/stores of the same address.
897 DenseMap<Value*, Value*> SunkAddrs;
899 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) {
900 Instruction *I = BBI++;
902 if (CastInst *CI = dyn_cast<CastInst>(I)) {
903 // If the source of the cast is a constant, then this should have
904 // already been constant folded. The only reason NOT to constant fold
905 // it is if something (e.g. LSR) was careful to place the constant
906 // evaluation in a block other than then one that uses it (e.g. to hoist
907 // the address of globals out of a loop). If this is the case, we don't
908 // want to forward-subst the cast.
909 if (isa<Constant>(CI->getOperand(0)))
914 Change = OptimizeNoopCopyExpression(CI, *TLI);
915 MadeChange |= Change;
918 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) {
919 MadeChange |= MoveExtToFormExtLoad(I);
920 MadeChange |= OptimizeExtUses(I);
922 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
923 MadeChange |= OptimizeCmpExpression(CI);
924 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
926 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(),
928 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
930 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1),
931 SI->getOperand(0)->getType(),
933 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
934 if (GEPI->hasAllZeroIndices()) {
935 /// The GEP operand must be a pointer, so must its result -> BitCast
936 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
937 GEPI->getName(), GEPI);
938 GEPI->replaceAllUsesWith(NC);
939 GEPI->eraseFromParent();
943 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
944 // If we found an inline asm expession, and if the target knows how to
945 // lower it to normal LLVM code, do so now.
946 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
947 if (TLI->ExpandInlineAsm(CI)) {
949 // Avoid processing instructions out of order, which could cause
950 // reuse before a value is defined.
953 // Sink address computing for memory operands into the block.
954 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs);