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/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/GetElementPtrTypeIterator.h"
39 #include "llvm/Support/PatternMatch.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Support/IRBuilder.h"
43 using namespace llvm::PatternMatch;
46 CriticalEdgeSplit("cgp-critical-edge-splitting",
47 cl::desc("Split critical edges during codegen prepare"),
48 cl::init(true), cl::Hidden);
51 class CodeGenPrepare : public FunctionPass {
52 /// TLI - Keep a pointer of a TargetLowering to consult for determining
53 /// transformation profitability.
54 const TargetLowering *TLI;
57 /// BackEdges - Keep a set of all the loop back edges.
59 SmallSet<std::pair<const BasicBlock*, const BasicBlock*>, 8> BackEdges;
61 static char ID; // Pass identification, replacement for typeid
62 explicit CodeGenPrepare(const TargetLowering *tli = 0)
63 : FunctionPass(ID), TLI(tli) {}
64 bool runOnFunction(Function &F);
66 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
67 AU.addPreserved<ProfileInfo>();
70 virtual void releaseMemory() {
75 bool EliminateMostlyEmptyBlocks(Function &F);
76 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
77 void EliminateMostlyEmptyBlock(BasicBlock *BB);
78 bool OptimizeBlock(BasicBlock &BB);
79 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy,
80 DenseMap<Value*,Value*> &SunkAddrs);
81 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS,
82 DenseMap<Value*,Value*> &SunkAddrs);
83 bool OptimizeCallInst(CallInst *CI);
84 bool MoveExtToFormExtLoad(Instruction *I);
85 bool OptimizeExtUses(Instruction *I);
86 void findLoopBackEdges(const Function &F);
90 char CodeGenPrepare::ID = 0;
91 INITIALIZE_PASS(CodeGenPrepare, "codegenprepare",
92 "Optimize for code generation", false, false);
94 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
95 return new CodeGenPrepare(TLI);
98 /// findLoopBackEdges - Do a DFS walk to find loop back edges.
100 void CodeGenPrepare::findLoopBackEdges(const Function &F) {
101 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
102 FindFunctionBackedges(F, Edges);
104 BackEdges.insert(Edges.begin(), Edges.end());
108 bool CodeGenPrepare::runOnFunction(Function &F) {
109 bool EverMadeChange = false;
111 PFI = getAnalysisIfAvailable<ProfileInfo>();
112 // First pass, eliminate blocks that contain only PHI nodes and an
113 // unconditional branch.
114 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
116 // Now find loop back edges.
117 findLoopBackEdges(F);
119 bool MadeChange = true;
122 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
123 MadeChange |= OptimizeBlock(*BB);
124 EverMadeChange |= MadeChange;
126 return EverMadeChange;
129 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
130 /// debug info directives, and an unconditional branch. Passes before isel
131 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
132 /// isel. Start by eliminating these blocks so we can split them the way we
134 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
135 bool MadeChange = false;
136 // Note that this intentionally skips the entry block.
137 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
138 BasicBlock *BB = I++;
140 // If this block doesn't end with an uncond branch, ignore it.
141 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
142 if (!BI || !BI->isUnconditional())
145 // If the instruction before the branch (skipping debug info) isn't a phi
146 // node, then other stuff is happening here.
147 BasicBlock::iterator BBI = BI;
148 if (BBI != BB->begin()) {
150 while (isa<DbgInfoIntrinsic>(BBI)) {
151 if (BBI == BB->begin())
155 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
159 // Do not break infinite loops.
160 BasicBlock *DestBB = BI->getSuccessor(0);
164 if (!CanMergeBlocks(BB, DestBB))
167 EliminateMostlyEmptyBlock(BB);
173 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
174 /// single uncond branch between them, and BB contains no other non-phi
176 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
177 const BasicBlock *DestBB) const {
178 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
179 // the successor. If there are more complex condition (e.g. preheaders),
180 // don't mess around with them.
181 BasicBlock::const_iterator BBI = BB->begin();
182 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
183 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
185 const Instruction *User = cast<Instruction>(*UI);
186 if (User->getParent() != DestBB || !isa<PHINode>(User))
188 // If User is inside DestBB block and it is a PHINode then check
189 // incoming value. If incoming value is not from BB then this is
190 // a complex condition (e.g. preheaders) we want to avoid here.
191 if (User->getParent() == DestBB) {
192 if (const PHINode *UPN = dyn_cast<PHINode>(User))
193 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
194 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
195 if (Insn && Insn->getParent() == BB &&
196 Insn->getParent() != UPN->getIncomingBlock(I))
203 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
204 // and DestBB may have conflicting incoming values for the block. If so, we
205 // can't merge the block.
206 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
207 if (!DestBBPN) return true; // no conflict.
209 // Collect the preds of BB.
210 SmallPtrSet<const BasicBlock*, 16> BBPreds;
211 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
212 // It is faster to get preds from a PHI than with pred_iterator.
213 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
214 BBPreds.insert(BBPN->getIncomingBlock(i));
216 BBPreds.insert(pred_begin(BB), pred_end(BB));
219 // Walk the preds of DestBB.
220 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
221 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
222 if (BBPreds.count(Pred)) { // Common predecessor?
223 BBI = DestBB->begin();
224 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
225 const Value *V1 = PN->getIncomingValueForBlock(Pred);
226 const Value *V2 = PN->getIncomingValueForBlock(BB);
228 // If V2 is a phi node in BB, look up what the mapped value will be.
229 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
230 if (V2PN->getParent() == BB)
231 V2 = V2PN->getIncomingValueForBlock(Pred);
233 // If there is a conflict, bail out.
234 if (V1 != V2) return false;
243 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
244 /// an unconditional branch in it.
245 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
246 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
247 BasicBlock *DestBB = BI->getSuccessor(0);
249 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
251 // If the destination block has a single pred, then this is a trivial edge,
253 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
254 if (SinglePred != DestBB) {
255 // Remember if SinglePred was the entry block of the function. If so, we
256 // will need to move BB back to the entry position.
257 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
258 MergeBasicBlockIntoOnlyPred(DestBB, this);
260 if (isEntry && BB != &BB->getParent()->getEntryBlock())
261 BB->moveBefore(&BB->getParent()->getEntryBlock());
263 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
268 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
269 // to handle the new incoming edges it is about to have.
271 for (BasicBlock::iterator BBI = DestBB->begin();
272 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
273 // Remove the incoming value for BB, and remember it.
274 Value *InVal = PN->removeIncomingValue(BB, false);
276 // Two options: either the InVal is a phi node defined in BB or it is some
277 // value that dominates BB.
278 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
279 if (InValPhi && InValPhi->getParent() == BB) {
280 // Add all of the input values of the input PHI as inputs of this phi.
281 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
282 PN->addIncoming(InValPhi->getIncomingValue(i),
283 InValPhi->getIncomingBlock(i));
285 // Otherwise, add one instance of the dominating value for each edge that
286 // we will be adding.
287 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
288 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
289 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
291 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
292 PN->addIncoming(InVal, *PI);
297 // The PHIs are now updated, change everything that refers to BB to use
298 // DestBB and remove BB.
299 BB->replaceAllUsesWith(DestBB);
301 PFI->replaceAllUses(BB, DestBB);
302 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
304 BB->eraseFromParent();
306 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
309 /// FindReusablePredBB - Check all of the predecessors of the block DestPHI
310 /// lives in to see if there is a block that we can reuse as a critical edge
312 static BasicBlock *FindReusablePredBB(PHINode *DestPHI, BasicBlock *TIBB) {
313 BasicBlock *Dest = DestPHI->getParent();
315 /// TIPHIValues - This array is lazily computed to determine the values of
316 /// PHIs in Dest that TI would provide.
317 SmallVector<Value*, 32> TIPHIValues;
319 /// TIBBEntryNo - This is a cache to speed up pred queries for TIBB.
320 unsigned TIBBEntryNo = 0;
322 // Check to see if Dest has any blocks that can be used as a split edge for
324 for (unsigned pi = 0, e = DestPHI->getNumIncomingValues(); pi != e; ++pi) {
325 BasicBlock *Pred = DestPHI->getIncomingBlock(pi);
326 // To be usable, the pred has to end with an uncond branch to the dest.
327 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator());
328 if (!PredBr || !PredBr->isUnconditional())
330 // Must be empty other than the branch and debug info.
331 BasicBlock::iterator I = Pred->begin();
332 while (isa<DbgInfoIntrinsic>(I))
336 // Cannot be the entry block; its label does not get emitted.
337 if (Pred == &Dest->getParent()->getEntryBlock())
340 // Finally, since we know that Dest has phi nodes in it, we have to make
341 // sure that jumping to Pred will have the same effect as going to Dest in
342 // terms of PHI values.
345 unsigned PredEntryNo = pi;
347 bool FoundMatch = true;
348 for (BasicBlock::iterator I = Dest->begin();
349 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) {
350 if (PHINo == TIPHIValues.size()) {
351 if (PN->getIncomingBlock(TIBBEntryNo) != TIBB)
352 TIBBEntryNo = PN->getBasicBlockIndex(TIBB);
353 TIPHIValues.push_back(PN->getIncomingValue(TIBBEntryNo));
356 // If the PHI entry doesn't work, we can't use this pred.
357 if (PN->getIncomingBlock(PredEntryNo) != Pred)
358 PredEntryNo = PN->getBasicBlockIndex(Pred);
360 if (TIPHIValues[PHINo] != PN->getIncomingValue(PredEntryNo)) {
366 // If we found a workable predecessor, change TI to branch to Succ.
374 /// SplitEdgeNicely - Split the critical edge from TI to its specified
375 /// successor if it will improve codegen. We only do this if the successor has
376 /// phi nodes (otherwise critical edges are ok). If there is already another
377 /// predecessor of the succ that is empty (and thus has no phi nodes), use it
378 /// instead of introducing a new block.
379 static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum,
380 SmallSet<std::pair<const BasicBlock*,
381 const BasicBlock*>, 8> &BackEdges,
383 BasicBlock *TIBB = TI->getParent();
384 BasicBlock *Dest = TI->getSuccessor(SuccNum);
385 assert(isa<PHINode>(Dest->begin()) &&
386 "This should only be called if Dest has a PHI!");
387 PHINode *DestPHI = cast<PHINode>(Dest->begin());
389 // Do not split edges to EH landing pads.
390 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(TI))
391 if (Invoke->getSuccessor(1) == Dest)
394 // As a hack, never split backedges of loops. Even though the copy for any
395 // PHIs inserted on the backedge would be dead for exits from the loop, we
396 // assume that the cost of *splitting* the backedge would be too high.
397 if (BackEdges.count(std::make_pair(TIBB, Dest)))
400 if (BasicBlock *ReuseBB = FindReusablePredBB(DestPHI, TIBB)) {
401 ProfileInfo *PFI = P->getAnalysisIfAvailable<ProfileInfo>();
403 PFI->splitEdge(TIBB, Dest, ReuseBB);
404 Dest->removePredecessor(TIBB);
405 TI->setSuccessor(SuccNum, ReuseBB);
409 SplitCriticalEdge(TI, SuccNum, P, true);
413 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
414 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
415 /// sink it into user blocks to reduce the number of virtual
416 /// registers that must be created and coalesced.
418 /// Return true if any changes are made.
420 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
421 // If this is a noop copy,
422 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
423 EVT DstVT = TLI.getValueType(CI->getType());
425 // This is an fp<->int conversion?
426 if (SrcVT.isInteger() != DstVT.isInteger())
429 // If this is an extension, it will be a zero or sign extension, which
431 if (SrcVT.bitsLT(DstVT)) return false;
433 // If these values will be promoted, find out what they will be promoted
434 // to. This helps us consider truncates on PPC as noop copies when they
436 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote)
437 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
438 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote)
439 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
441 // If, after promotion, these are the same types, this is a noop copy.
445 BasicBlock *DefBB = CI->getParent();
447 /// InsertedCasts - Only insert a cast in each block once.
448 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
450 bool MadeChange = false;
451 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
453 Use &TheUse = UI.getUse();
454 Instruction *User = cast<Instruction>(*UI);
456 // Figure out which BB this cast is used in. For PHI's this is the
457 // appropriate predecessor block.
458 BasicBlock *UserBB = User->getParent();
459 if (PHINode *PN = dyn_cast<PHINode>(User)) {
460 UserBB = PN->getIncomingBlock(UI);
463 // Preincrement use iterator so we don't invalidate it.
466 // If this user is in the same block as the cast, don't change the cast.
467 if (UserBB == DefBB) continue;
469 // If we have already inserted a cast into this block, use it.
470 CastInst *&InsertedCast = InsertedCasts[UserBB];
473 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
476 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
481 // Replace a use of the cast with a use of the new cast.
482 TheUse = InsertedCast;
485 // If we removed all uses, nuke the cast.
486 if (CI->use_empty()) {
487 CI->eraseFromParent();
494 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
495 /// the number of virtual registers that must be created and coalesced. This is
496 /// a clear win except on targets with multiple condition code registers
497 /// (PowerPC), where it might lose; some adjustment may be wanted there.
499 /// Return true if any changes are made.
500 static bool OptimizeCmpExpression(CmpInst *CI) {
501 BasicBlock *DefBB = CI->getParent();
503 /// InsertedCmp - Only insert a cmp in each block once.
504 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
506 bool MadeChange = false;
507 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
509 Use &TheUse = UI.getUse();
510 Instruction *User = cast<Instruction>(*UI);
512 // Preincrement use iterator so we don't invalidate it.
515 // Don't bother for PHI nodes.
516 if (isa<PHINode>(User))
519 // Figure out which BB this cmp is used in.
520 BasicBlock *UserBB = User->getParent();
522 // If this user is in the same block as the cmp, don't change the cmp.
523 if (UserBB == DefBB) continue;
525 // If we have already inserted a cmp into this block, use it.
526 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
529 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
532 CmpInst::Create(CI->getOpcode(),
533 CI->getPredicate(), CI->getOperand(0),
534 CI->getOperand(1), "", InsertPt);
538 // Replace a use of the cmp with a use of the new cmp.
539 TheUse = InsertedCmp;
542 // If we removed all uses, nuke the cmp.
544 CI->eraseFromParent();
550 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
552 void replaceCall(Value *With) {
553 CI->replaceAllUsesWith(With);
554 CI->eraseFromParent();
556 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
557 if (ConstantInt *SizeCI =
558 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
559 return SizeCI->isAllOnesValue();
563 } // end anonymous namespace
565 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
566 // Lower all uses of llvm.objectsize.*
567 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
568 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
569 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
570 const Type *ReturnTy = CI->getType();
571 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
572 CI->replaceAllUsesWith(RetVal);
573 CI->eraseFromParent();
577 // From here on out we're working with named functions.
578 if (CI->getCalledFunction() == 0) return false;
580 // We'll need TargetData from here on out.
581 const TargetData *TD = TLI ? TLI->getTargetData() : 0;
582 if (!TD) return false;
584 // Lower all default uses of _chk calls. This is very similar
585 // to what InstCombineCalls does, but here we are only lowering calls
586 // that have the default "don't know" as the objectsize. Anything else
587 // should be left alone.
588 CodeGenPrepareFortifiedLibCalls Simplifier;
589 return Simplifier.fold(CI, TD);
591 //===----------------------------------------------------------------------===//
592 // Memory Optimization
593 //===----------------------------------------------------------------------===//
595 /// IsNonLocalValue - Return true if the specified values are defined in a
596 /// different basic block than BB.
597 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
598 if (Instruction *I = dyn_cast<Instruction>(V))
599 return I->getParent() != BB;
603 /// OptimizeMemoryInst - Load and Store Instructions often have
604 /// addressing modes that can do significant amounts of computation. As such,
605 /// instruction selection will try to get the load or store to do as much
606 /// computation as possible for the program. The problem is that isel can only
607 /// see within a single block. As such, we sink as much legal addressing mode
608 /// stuff into the block as possible.
610 /// This method is used to optimize both load/store and inline asms with memory
612 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
613 const Type *AccessTy,
614 DenseMap<Value*,Value*> &SunkAddrs) {
615 // Figure out what addressing mode will be built up for this operation.
616 SmallVector<Instruction*, 16> AddrModeInsts;
617 ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst,
618 AddrModeInsts, *TLI);
620 // Check to see if any of the instructions supersumed by this addr mode are
621 // non-local to I's BB.
622 bool AnyNonLocal = false;
623 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
624 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
630 // If all the instructions matched are already in this BB, don't do anything.
632 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
636 // Insert this computation right after this user. Since our caller is
637 // scanning from the top of the BB to the bottom, reuse of the expr are
638 // guaranteed to happen later.
639 BasicBlock::iterator InsertPt = MemoryInst;
641 // Now that we determined the addressing expression we want to use and know
642 // that we have to sink it into this block. Check to see if we have already
643 // done this for some other load/store instr in this block. If so, reuse the
645 Value *&SunkAddr = SunkAddrs[Addr];
647 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
649 if (SunkAddr->getType() != Addr->getType())
650 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
652 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
654 const Type *IntPtrTy =
655 TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
659 // Start with the base register. Do this first so that subsequent address
660 // matching finds it last, which will prevent it from trying to match it
661 // as the scaled value in case it happens to be a mul. That would be
662 // problematic if we've sunk a different mul for the scale, because then
663 // we'd end up sinking both muls.
664 if (AddrMode.BaseReg) {
665 Value *V = AddrMode.BaseReg;
666 if (V->getType()->isPointerTy())
667 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
668 if (V->getType() != IntPtrTy)
669 V = CastInst::CreateIntegerCast(V, IntPtrTy, /*isSigned=*/true,
670 "sunkaddr", InsertPt);
674 // Add the scale value.
675 if (AddrMode.Scale) {
676 Value *V = AddrMode.ScaledReg;
677 if (V->getType() == IntPtrTy) {
679 } else if (V->getType()->isPointerTy()) {
680 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
681 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
682 cast<IntegerType>(V->getType())->getBitWidth()) {
683 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
685 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
687 if (AddrMode.Scale != 1)
688 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
690 "sunkaddr", InsertPt);
692 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
697 // Add in the BaseGV if present.
698 if (AddrMode.BaseGV) {
699 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
702 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
707 // Add in the Base Offset if present.
708 if (AddrMode.BaseOffs) {
709 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
711 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
717 SunkAddr = Constant::getNullValue(Addr->getType());
719 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
722 MemoryInst->replaceUsesOfWith(Addr, SunkAddr);
724 if (Addr->use_empty()) {
725 RecursivelyDeleteTriviallyDeadInstructions(Addr);
726 // This address is now available for reassignment, so erase the table entry;
727 // we don't want to match some completely different instruction.
733 /// OptimizeInlineAsmInst - If there are any memory operands, use
734 /// OptimizeMemoryInst to sink their address computing into the block when
735 /// possible / profitable.
736 bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS,
737 DenseMap<Value*,Value*> &SunkAddrs) {
738 bool MadeChange = false;
740 std::vector<TargetLowering::AsmOperandInfo> TargetConstraints = TLI->ParseConstraints(CS);
741 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
742 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
744 // Compute the constraint code and ConstraintType to use.
745 TLI->ComputeConstraintToUse(OpInfo, SDValue());
747 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
749 Value *OpVal = OpInfo.CallOperandVal;
750 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs);
757 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
758 /// basic block as the load, unless conditions are unfavorable. This allows
759 /// SelectionDAG to fold the extend into the load.
761 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
762 // Look for a load being extended.
763 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
764 if (!LI) return false;
766 // If they're already in the same block, there's nothing to do.
767 if (LI->getParent() == I->getParent())
770 // If the load has other users and the truncate is not free, this probably
772 if (!LI->hasOneUse() &&
773 TLI && !TLI->isTruncateFree(I->getType(), LI->getType()))
776 // Check whether the target supports casts folded into loads.
778 if (isa<ZExtInst>(I))
779 LType = ISD::ZEXTLOAD;
781 assert(isa<SExtInst>(I) && "Unexpected ext type!");
782 LType = ISD::SEXTLOAD;
784 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
787 // Move the extend into the same block as the load, so that SelectionDAG
789 I->removeFromParent();
794 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
795 BasicBlock *DefBB = I->getParent();
797 // If both result of the {s|z}xt and its source are live out, rewrite all
798 // other uses of the source with result of extension.
799 Value *Src = I->getOperand(0);
800 if (Src->hasOneUse())
803 // Only do this xform if truncating is free.
804 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
807 // Only safe to perform the optimization if the source is also defined in
809 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
812 bool DefIsLiveOut = false;
813 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
815 Instruction *User = cast<Instruction>(*UI);
817 // Figure out which BB this ext is used in.
818 BasicBlock *UserBB = User->getParent();
819 if (UserBB == DefBB) continue;
826 // Make sure non of the uses are PHI nodes.
827 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
829 Instruction *User = cast<Instruction>(*UI);
830 BasicBlock *UserBB = User->getParent();
831 if (UserBB == DefBB) continue;
832 // Be conservative. We don't want this xform to end up introducing
833 // reloads just before load / store instructions.
834 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
838 // InsertedTruncs - Only insert one trunc in each block once.
839 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
841 bool MadeChange = false;
842 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
844 Use &TheUse = UI.getUse();
845 Instruction *User = cast<Instruction>(*UI);
847 // Figure out which BB this ext is used in.
848 BasicBlock *UserBB = User->getParent();
849 if (UserBB == DefBB) continue;
851 // Both src and def are live in this block. Rewrite the use.
852 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
854 if (!InsertedTrunc) {
855 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
857 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
860 // Replace a use of the {s|z}ext source with a use of the result.
861 TheUse = InsertedTrunc;
869 // In this pass we look for GEP and cast instructions that are used
870 // across basic blocks and rewrite them to improve basic-block-at-a-time
872 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
873 bool MadeChange = false;
875 // Split all critical edges where the dest block has a PHI.
876 if (CriticalEdgeSplit) {
877 TerminatorInst *BBTI = BB.getTerminator();
878 if (BBTI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(BBTI)) {
879 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) {
880 BasicBlock *SuccBB = BBTI->getSuccessor(i);
881 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true))
882 SplitEdgeNicely(BBTI, i, BackEdges, this);
887 // Keep track of non-local addresses that have been sunk into this block.
888 // This allows us to avoid inserting duplicate code for blocks with multiple
889 // load/stores of the same address.
890 DenseMap<Value*, Value*> SunkAddrs;
892 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) {
893 Instruction *I = BBI++;
895 if (CastInst *CI = dyn_cast<CastInst>(I)) {
896 // If the source of the cast is a constant, then this should have
897 // already been constant folded. The only reason NOT to constant fold
898 // it is if something (e.g. LSR) was careful to place the constant
899 // evaluation in a block other than then one that uses it (e.g. to hoist
900 // the address of globals out of a loop). If this is the case, we don't
901 // want to forward-subst the cast.
902 if (isa<Constant>(CI->getOperand(0)))
907 Change = OptimizeNoopCopyExpression(CI, *TLI);
908 MadeChange |= Change;
911 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I))) {
912 MadeChange |= MoveExtToFormExtLoad(I);
913 MadeChange |= OptimizeExtUses(I);
915 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
916 MadeChange |= OptimizeCmpExpression(CI);
917 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
919 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(),
921 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
923 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1),
924 SI->getOperand(0)->getType(),
926 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
927 if (GEPI->hasAllZeroIndices()) {
928 /// The GEP operand must be a pointer, so must its result -> BitCast
929 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
930 GEPI->getName(), GEPI);
931 GEPI->replaceAllUsesWith(NC);
932 GEPI->eraseFromParent();
936 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
937 // If we found an inline asm expession, and if the target knows how to
938 // lower it to normal LLVM code, do so now.
939 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
940 if (TLI->ExpandInlineAsm(CI)) {
942 // Avoid processing instructions out of order, which could cause
943 // reuse before a value is defined.
946 // Sink address computing for memory operands into the block.
947 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs);
949 // Other CallInst optimizations that don't need to muck with the
950 // enclosing iterator here.
951 MadeChange |= OptimizeCallInst(CI);