1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass munges the code in the input function to better prepare it for
11 // SelectionDAG-based code generation. This works around limitations in it's
12 // basic-block-at-a-time approach. It should eventually be removed.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "codegenprepare"
17 #include "llvm/Transforms/Scalar.h"
18 #include "llvm/Constants.h"
19 #include "llvm/DerivedTypes.h"
20 #include "llvm/Function.h"
21 #include "llvm/InlineAsm.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/Pass.h"
24 #include "llvm/Target/TargetAsmInfo.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/Target/TargetLowering.h"
27 #include "llvm/Target/TargetMachine.h"
28 #include "llvm/Transforms/Utils/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/Compiler.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/GetElementPtrTypeIterator.h"
39 #include "llvm/Support/PatternMatch.h"
41 using namespace llvm::PatternMatch;
43 static cl::opt<bool> FactorCommonPreds("split-critical-paths-tweak",
44 cl::init(false), cl::Hidden);
47 class VISIBILITY_HIDDEN CodeGenPrepare : public FunctionPass {
48 /// TLI - Keep a pointer of a TargetLowering to consult for determining
49 /// transformation profitability.
50 const TargetLowering *TLI;
52 /// BackEdges - Keep a set of all the loop back edges.
54 SmallSet<std::pair<BasicBlock*,BasicBlock*>, 8> BackEdges;
56 static char ID; // Pass identification, replacement for typeid
57 explicit CodeGenPrepare(const TargetLowering *tli = 0)
58 : FunctionPass(&ID), TLI(tli) {}
59 bool runOnFunction(Function &F);
62 bool EliminateMostlyEmptyBlocks(Function &F);
63 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
64 void EliminateMostlyEmptyBlock(BasicBlock *BB);
65 bool OptimizeBlock(BasicBlock &BB);
66 bool OptimizeMemoryInst(Instruction *I, Value *Addr, const Type *AccessTy,
67 DenseMap<Value*,Value*> &SunkAddrs);
68 bool OptimizeInlineAsmInst(Instruction *I, CallSite CS,
69 DenseMap<Value*,Value*> &SunkAddrs);
70 bool OptimizeExtUses(Instruction *I);
71 void findLoopBackEdges(Function &F);
75 char CodeGenPrepare::ID = 0;
76 static RegisterPass<CodeGenPrepare> X("codegenprepare",
77 "Optimize for code generation");
79 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
80 return new CodeGenPrepare(TLI);
83 /// findLoopBackEdges - Do a DFS walk to find loop back edges.
85 void CodeGenPrepare::findLoopBackEdges(Function &F) {
86 SmallPtrSet<BasicBlock*, 8> Visited;
87 SmallVector<std::pair<BasicBlock*, succ_iterator>, 8> VisitStack;
88 SmallPtrSet<BasicBlock*, 8> InStack;
90 BasicBlock *BB = &F.getEntryBlock();
91 if (succ_begin(BB) == succ_end(BB))
94 VisitStack.push_back(std::make_pair(BB, succ_begin(BB)));
97 std::pair<BasicBlock*, succ_iterator> &Top = VisitStack.back();
98 BasicBlock *ParentBB = Top.first;
99 succ_iterator &I = Top.second;
101 bool FoundNew = false;
102 while (I != succ_end(ParentBB)) {
104 if (Visited.insert(BB)) {
108 // Successor is in VisitStack, it's a back edge.
109 if (InStack.count(BB))
110 BackEdges.insert(std::make_pair(ParentBB, BB));
114 // Go down one level if there is a unvisited successor.
116 VisitStack.push_back(std::make_pair(BB, succ_begin(BB)));
119 std::pair<BasicBlock*, succ_iterator> &Pop = VisitStack.back();
120 InStack.erase(Pop.first);
121 VisitStack.pop_back();
123 } while (!VisitStack.empty());
127 bool CodeGenPrepare::runOnFunction(Function &F) {
128 bool EverMadeChange = false;
130 // First pass, eliminate blocks that contain only PHI nodes and an
131 // unconditional branch.
132 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
134 // Now find loop back edges.
135 findLoopBackEdges(F);
137 bool MadeChange = true;
140 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
141 MadeChange |= OptimizeBlock(*BB);
142 EverMadeChange |= MadeChange;
144 return EverMadeChange;
147 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes
148 /// and an unconditional branch. Passes before isel (e.g. LSR/loopsimplify)
149 /// often split edges in ways that are non-optimal for isel. Start by
150 /// eliminating these blocks so we can split them the way we want them.
151 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
152 bool MadeChange = false;
153 // Note that this intentionally skips the entry block.
154 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
155 BasicBlock *BB = I++;
157 // If this block doesn't end with an uncond branch, ignore it.
158 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
159 if (!BI || !BI->isUnconditional())
162 // If the instruction before the branch isn't a phi node, then other stuff
163 // is happening here.
164 BasicBlock::iterator BBI = BI;
165 if (BBI != BB->begin()) {
167 if (!isa<PHINode>(BBI)) continue;
170 // Do not break infinite loops.
171 BasicBlock *DestBB = BI->getSuccessor(0);
175 if (!CanMergeBlocks(BB, DestBB))
178 EliminateMostlyEmptyBlock(BB);
184 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
185 /// single uncond branch between them, and BB contains no other non-phi
187 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
188 const BasicBlock *DestBB) const {
189 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
190 // the successor. If there are more complex condition (e.g. preheaders),
191 // don't mess around with them.
192 BasicBlock::const_iterator BBI = BB->begin();
193 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
194 for (Value::use_const_iterator UI = PN->use_begin(), E = PN->use_end();
196 const Instruction *User = cast<Instruction>(*UI);
197 if (User->getParent() != DestBB || !isa<PHINode>(User))
199 // If User is inside DestBB block and it is a PHINode then check
200 // incoming value. If incoming value is not from BB then this is
201 // a complex condition (e.g. preheaders) we want to avoid here.
202 if (User->getParent() == DestBB) {
203 if (const PHINode *UPN = dyn_cast<PHINode>(User))
204 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
205 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
206 if (Insn && Insn->getParent() == BB &&
207 Insn->getParent() != UPN->getIncomingBlock(I))
214 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
215 // and DestBB may have conflicting incoming values for the block. If so, we
216 // can't merge the block.
217 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
218 if (!DestBBPN) return true; // no conflict.
220 // Collect the preds of BB.
221 SmallPtrSet<const BasicBlock*, 16> BBPreds;
222 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
223 // It is faster to get preds from a PHI than with pred_iterator.
224 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
225 BBPreds.insert(BBPN->getIncomingBlock(i));
227 BBPreds.insert(pred_begin(BB), pred_end(BB));
230 // Walk the preds of DestBB.
231 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
232 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
233 if (BBPreds.count(Pred)) { // Common predecessor?
234 BBI = DestBB->begin();
235 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
236 const Value *V1 = PN->getIncomingValueForBlock(Pred);
237 const Value *V2 = PN->getIncomingValueForBlock(BB);
239 // If V2 is a phi node in BB, look up what the mapped value will be.
240 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
241 if (V2PN->getParent() == BB)
242 V2 = V2PN->getIncomingValueForBlock(Pred);
244 // If there is a conflict, bail out.
245 if (V1 != V2) return false;
254 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
255 /// an unconditional branch in it.
256 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
257 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
258 BasicBlock *DestBB = BI->getSuccessor(0);
260 DOUT << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB;
262 // If the destination block has a single pred, then this is a trivial edge,
264 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
265 if (SinglePred != DestBB) {
266 // Remember if SinglePred was the entry block of the function. If so, we
267 // will need to move BB back to the entry position.
268 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
269 MergeBasicBlockIntoOnlyPred(DestBB);
271 if (isEntry && BB != &BB->getParent()->getEntryBlock())
272 BB->moveBefore(&BB->getParent()->getEntryBlock());
274 DOUT << "AFTER:\n" << *DestBB << "\n\n\n";
279 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
280 // to handle the new incoming edges it is about to have.
282 for (BasicBlock::iterator BBI = DestBB->begin();
283 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
284 // Remove the incoming value for BB, and remember it.
285 Value *InVal = PN->removeIncomingValue(BB, false);
287 // Two options: either the InVal is a phi node defined in BB or it is some
288 // value that dominates BB.
289 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
290 if (InValPhi && InValPhi->getParent() == BB) {
291 // Add all of the input values of the input PHI as inputs of this phi.
292 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
293 PN->addIncoming(InValPhi->getIncomingValue(i),
294 InValPhi->getIncomingBlock(i));
296 // Otherwise, add one instance of the dominating value for each edge that
297 // we will be adding.
298 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
299 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
300 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
302 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
303 PN->addIncoming(InVal, *PI);
308 // The PHIs are now updated, change everything that refers to BB to use
309 // DestBB and remove BB.
310 BB->replaceAllUsesWith(DestBB);
311 BB->eraseFromParent();
313 DOUT << "AFTER:\n" << *DestBB << "\n\n\n";
317 /// SplitEdgeNicely - Split the critical edge from TI to its specified
318 /// successor if it will improve codegen. We only do this if the successor has
319 /// phi nodes (otherwise critical edges are ok). If there is already another
320 /// predecessor of the succ that is empty (and thus has no phi nodes), use it
321 /// instead of introducing a new block.
322 static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum,
323 SmallSet<std::pair<BasicBlock*,BasicBlock*>, 8> &BackEdges,
325 BasicBlock *TIBB = TI->getParent();
326 BasicBlock *Dest = TI->getSuccessor(SuccNum);
327 assert(isa<PHINode>(Dest->begin()) &&
328 "This should only be called if Dest has a PHI!");
330 // As a hack, never split backedges of loops. Even though the copy for any
331 // PHIs inserted on the backedge would be dead for exits from the loop, we
332 // assume that the cost of *splitting* the backedge would be too high.
333 if (BackEdges.count(std::make_pair(TIBB, Dest)))
336 if (!FactorCommonPreds) {
337 /// TIPHIValues - This array is lazily computed to determine the values of
338 /// PHIs in Dest that TI would provide.
339 SmallVector<Value*, 32> TIPHIValues;
341 // Check to see if Dest has any blocks that can be used as a split edge for
343 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) {
344 BasicBlock *Pred = *PI;
345 // To be usable, the pred has to end with an uncond branch to the dest.
346 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator());
347 if (!PredBr || !PredBr->isUnconditional() ||
348 // Must be empty other than the branch.
349 &Pred->front() != PredBr ||
350 // Cannot be the entry block; its label does not get emitted.
351 Pred == &(Dest->getParent()->getEntryBlock()))
354 // Finally, since we know that Dest has phi nodes in it, we have to make
355 // sure that jumping to Pred will have the same affect as going to Dest in
356 // terms of PHI values.
359 bool FoundMatch = true;
360 for (BasicBlock::iterator I = Dest->begin();
361 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) {
362 if (PHINo == TIPHIValues.size())
363 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB));
365 // If the PHI entry doesn't work, we can't use this pred.
366 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) {
372 // If we found a workable predecessor, change TI to branch to Succ.
374 Dest->removePredecessor(TIBB);
375 TI->setSuccessor(SuccNum, Pred);
380 SplitCriticalEdge(TI, SuccNum, P, true);
385 SmallVector<Value*, 8> TIPHIValues;
386 for (BasicBlock::iterator I = Dest->begin();
387 (PN = dyn_cast<PHINode>(I)); ++I)
388 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB));
390 SmallVector<BasicBlock*, 8> IdenticalPreds;
391 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) {
392 BasicBlock *Pred = *PI;
393 if (BackEdges.count(std::make_pair(Pred, Dest)))
396 IdenticalPreds.push_back(Pred);
398 bool Identical = true;
400 for (BasicBlock::iterator I = Dest->begin();
401 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo)
402 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) {
407 IdenticalPreds.push_back(Pred);
411 assert(!IdenticalPreds.empty());
412 SplitBlockPredecessors(Dest, &IdenticalPreds[0], IdenticalPreds.size(),
417 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
418 /// copy (e.g. it's casting from one pointer type to another, int->uint, or
419 /// int->sbyte on PPC), sink it into user blocks to reduce the number of virtual
420 /// registers that must be created and coalesced.
422 /// Return true if any changes are made.
424 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
425 // If this is a noop copy,
426 MVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
427 MVT DstVT = TLI.getValueType(CI->getType());
429 // This is an fp<->int conversion?
430 if (SrcVT.isInteger() != DstVT.isInteger())
433 // If this is an extension, it will be a zero or sign extension, which
435 if (SrcVT.bitsLT(DstVT)) return false;
437 // If these values will be promoted, find out what they will be promoted
438 // to. This helps us consider truncates on PPC as noop copies when they
440 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote)
441 SrcVT = TLI.getTypeToTransformTo(SrcVT);
442 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote)
443 DstVT = TLI.getTypeToTransformTo(DstVT);
445 // If, after promotion, these are the same types, this is a noop copy.
449 BasicBlock *DefBB = CI->getParent();
451 /// InsertedCasts - Only insert a cast in each block once.
452 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
454 bool MadeChange = false;
455 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
457 Use &TheUse = UI.getUse();
458 Instruction *User = cast<Instruction>(*UI);
460 // Figure out which BB this cast is used in. For PHI's this is the
461 // appropriate predecessor block.
462 BasicBlock *UserBB = User->getParent();
463 if (PHINode *PN = dyn_cast<PHINode>(User)) {
464 UserBB = PN->getIncomingBlock(UI);
467 // Preincrement use iterator so we don't invalidate it.
470 // If this user is in the same block as the cast, don't change the cast.
471 if (UserBB == DefBB) continue;
473 // If we have already inserted a cast into this block, use it.
474 CastInst *&InsertedCast = InsertedCasts[UserBB];
477 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
480 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
485 // Replace a use of the cast with a use of the new cast.
486 TheUse = InsertedCast;
489 // If we removed all uses, nuke the cast.
490 if (CI->use_empty()) {
491 CI->eraseFromParent();
498 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
499 /// the number of virtual registers that must be created and coalesced. This is
500 /// a clear win except on targets with multiple condition code registers
501 /// (PowerPC), where it might lose; some adjustment may be wanted there.
503 /// Return true if any changes are made.
504 static bool OptimizeCmpExpression(CmpInst *CI) {
505 BasicBlock *DefBB = CI->getParent();
507 /// InsertedCmp - Only insert a cmp in each block once.
508 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
510 bool MadeChange = false;
511 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
513 Use &TheUse = UI.getUse();
514 Instruction *User = cast<Instruction>(*UI);
516 // Preincrement use iterator so we don't invalidate it.
519 // Don't bother for PHI nodes.
520 if (isa<PHINode>(User))
523 // Figure out which BB this cmp is used in.
524 BasicBlock *UserBB = User->getParent();
526 // If this user is in the same block as the cmp, don't change the cmp.
527 if (UserBB == DefBB) continue;
529 // If we have already inserted a cmp into this block, use it.
530 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
533 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
536 CmpInst::Create(CI->getOpcode(), CI->getPredicate(), CI->getOperand(0),
537 CI->getOperand(1), "", InsertPt);
541 // Replace a use of the cmp with a use of the new cmp.
542 TheUse = InsertedCmp;
545 // If we removed all uses, nuke the cmp.
547 CI->eraseFromParent();
552 //===----------------------------------------------------------------------===//
553 // Addressing Mode Analysis and Optimization
554 //===----------------------------------------------------------------------===//
556 //===----------------------------------------------------------------------===//
557 // Memory Optimization
558 //===----------------------------------------------------------------------===//
560 /// IsNonLocalValue - Return true if the specified values are defined in a
561 /// different basic block than BB.
562 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
563 if (Instruction *I = dyn_cast<Instruction>(V))
564 return I->getParent() != BB;
568 /// OptimizeMemoryInst - Load and Store Instructions have often have
569 /// addressing modes that can do significant amounts of computation. As such,
570 /// instruction selection will try to get the load or store to do as much
571 /// computation as possible for the program. The problem is that isel can only
572 /// see within a single block. As such, we sink as much legal addressing mode
573 /// stuff into the block as possible.
575 /// This method is used to optimize both load/store and inline asms with memory
577 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
578 const Type *AccessTy,
579 DenseMap<Value*,Value*> &SunkAddrs) {
580 // Figure out what addressing mode will be built up for this operation.
581 SmallVector<Instruction*, 16> AddrModeInsts;
582 ExtAddrMode AddrMode = AddressingModeMatcher::Match(Addr, AccessTy,MemoryInst,
583 AddrModeInsts, *TLI);
585 // Check to see if any of the instructions supersumed by this addr mode are
586 // non-local to I's BB.
587 bool AnyNonLocal = false;
588 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
589 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
595 // If all the instructions matched are already in this BB, don't do anything.
597 DEBUG(cerr << "CGP: Found local addrmode: " << AddrMode << "\n");
601 // Insert this computation right after this user. Since our caller is
602 // scanning from the top of the BB to the bottom, reuse of the expr are
603 // guaranteed to happen later.
604 BasicBlock::iterator InsertPt = MemoryInst;
606 // Now that we determined the addressing expression we want to use and know
607 // that we have to sink it into this block. Check to see if we have already
608 // done this for some other load/store instr in this block. If so, reuse the
610 Value *&SunkAddr = SunkAddrs[Addr];
612 DEBUG(cerr << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
614 if (SunkAddr->getType() != Addr->getType())
615 SunkAddr = new BitCastInst(SunkAddr, Addr->getType(), "tmp", InsertPt);
617 DEBUG(cerr << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
619 const Type *IntPtrTy = TLI->getTargetData()->getIntPtrType();
622 // Start with the scale value.
623 if (AddrMode.Scale) {
624 Value *V = AddrMode.ScaledReg;
625 if (V->getType() == IntPtrTy) {
627 } else if (isa<PointerType>(V->getType())) {
628 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
629 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
630 cast<IntegerType>(V->getType())->getBitWidth()) {
631 V = new TruncInst(V, IntPtrTy, "sunkaddr", InsertPt);
633 V = new SExtInst(V, IntPtrTy, "sunkaddr", InsertPt);
635 if (AddrMode.Scale != 1)
636 V = BinaryOperator::CreateMul(V, ConstantInt::get(IntPtrTy,
638 "sunkaddr", InsertPt);
642 // Add in the base register.
643 if (AddrMode.BaseReg) {
644 Value *V = AddrMode.BaseReg;
645 if (V->getType() != IntPtrTy)
646 V = new PtrToIntInst(V, IntPtrTy, "sunkaddr", InsertPt);
648 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
653 // Add in the BaseGV if present.
654 if (AddrMode.BaseGV) {
655 Value *V = new PtrToIntInst(AddrMode.BaseGV, IntPtrTy, "sunkaddr",
658 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
663 // Add in the Base Offset if present.
664 if (AddrMode.BaseOffs) {
665 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
667 Result = BinaryOperator::CreateAdd(Result, V, "sunkaddr", InsertPt);
673 SunkAddr = Constant::getNullValue(Addr->getType());
675 SunkAddr = new IntToPtrInst(Result, Addr->getType(), "sunkaddr",InsertPt);
678 MemoryInst->replaceUsesOfWith(Addr, SunkAddr);
680 if (Addr->use_empty())
681 RecursivelyDeleteTriviallyDeadInstructions(Addr);
685 /// OptimizeInlineAsmInst - If there are any memory operands, use
686 /// OptimizeMemoryInst to sink their address computing into the block when
687 /// possible / profitable.
688 bool CodeGenPrepare::OptimizeInlineAsmInst(Instruction *I, CallSite CS,
689 DenseMap<Value*,Value*> &SunkAddrs) {
690 bool MadeChange = false;
691 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
693 // Do a prepass over the constraints, canonicalizing them, and building up the
694 // ConstraintOperands list.
695 std::vector<InlineAsm::ConstraintInfo>
696 ConstraintInfos = IA->ParseConstraints();
698 /// ConstraintOperands - Information about all of the constraints.
699 std::vector<TargetLowering::AsmOperandInfo> ConstraintOperands;
700 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
701 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
703 push_back(TargetLowering::AsmOperandInfo(ConstraintInfos[i]));
704 TargetLowering::AsmOperandInfo &OpInfo = ConstraintOperands.back();
706 // Compute the value type for each operand.
707 switch (OpInfo.Type) {
708 case InlineAsm::isOutput:
709 if (OpInfo.isIndirect)
710 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
712 case InlineAsm::isInput:
713 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
715 case InlineAsm::isClobber:
720 // Compute the constraint code and ConstraintType to use.
721 TLI->ComputeConstraintToUse(OpInfo, SDValue(),
722 OpInfo.ConstraintType == TargetLowering::C_Memory);
724 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
726 Value *OpVal = OpInfo.CallOperandVal;
727 MadeChange |= OptimizeMemoryInst(I, OpVal, OpVal->getType(), SunkAddrs);
734 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
735 BasicBlock *DefBB = I->getParent();
737 // If both result of the {s|z}xt and its source are live out, rewrite all
738 // other uses of the source with result of extension.
739 Value *Src = I->getOperand(0);
740 if (Src->hasOneUse())
743 // Only do this xform if truncating is free.
744 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
747 // Only safe to perform the optimization if the source is also defined in
749 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
752 bool DefIsLiveOut = false;
753 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
755 Instruction *User = cast<Instruction>(*UI);
757 // Figure out which BB this ext is used in.
758 BasicBlock *UserBB = User->getParent();
759 if (UserBB == DefBB) continue;
766 // Make sure non of the uses are PHI nodes.
767 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
769 Instruction *User = cast<Instruction>(*UI);
770 BasicBlock *UserBB = User->getParent();
771 if (UserBB == DefBB) continue;
772 // Be conservative. We don't want this xform to end up introducing
773 // reloads just before load / store instructions.
774 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
778 // InsertedTruncs - Only insert one trunc in each block once.
779 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
781 bool MadeChange = false;
782 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
784 Use &TheUse = UI.getUse();
785 Instruction *User = cast<Instruction>(*UI);
787 // Figure out which BB this ext is used in.
788 BasicBlock *UserBB = User->getParent();
789 if (UserBB == DefBB) continue;
791 // Both src and def are live in this block. Rewrite the use.
792 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
794 if (!InsertedTrunc) {
795 BasicBlock::iterator InsertPt = UserBB->getFirstNonPHI();
797 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
800 // Replace a use of the {s|z}ext source with a use of the result.
801 TheUse = InsertedTrunc;
809 // In this pass we look for GEP and cast instructions that are used
810 // across basic blocks and rewrite them to improve basic-block-at-a-time
812 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
813 bool MadeChange = false;
815 // Split all critical edges where the dest block has a PHI.
816 TerminatorInst *BBTI = BB.getTerminator();
817 if (BBTI->getNumSuccessors() > 1) {
818 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) {
819 BasicBlock *SuccBB = BBTI->getSuccessor(i);
820 if (isa<PHINode>(SuccBB->begin()) && isCriticalEdge(BBTI, i, true))
821 SplitEdgeNicely(BBTI, i, BackEdges, this);
825 // Keep track of non-local addresses that have been sunk into this block.
826 // This allows us to avoid inserting duplicate code for blocks with multiple
827 // load/stores of the same address.
828 DenseMap<Value*, Value*> SunkAddrs;
830 for (BasicBlock::iterator BBI = BB.begin(), E = BB.end(); BBI != E; ) {
831 Instruction *I = BBI++;
833 if (CastInst *CI = dyn_cast<CastInst>(I)) {
834 // If the source of the cast is a constant, then this should have
835 // already been constant folded. The only reason NOT to constant fold
836 // it is if something (e.g. LSR) was careful to place the constant
837 // evaluation in a block other than then one that uses it (e.g. to hoist
838 // the address of globals out of a loop). If this is the case, we don't
839 // want to forward-subst the cast.
840 if (isa<Constant>(CI->getOperand(0)))
845 Change = OptimizeNoopCopyExpression(CI, *TLI);
846 MadeChange |= Change;
849 if (!Change && (isa<ZExtInst>(I) || isa<SExtInst>(I)))
850 MadeChange |= OptimizeExtUses(I);
851 } else if (CmpInst *CI = dyn_cast<CmpInst>(I)) {
852 MadeChange |= OptimizeCmpExpression(CI);
853 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
855 MadeChange |= OptimizeMemoryInst(I, I->getOperand(0), LI->getType(),
857 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
859 MadeChange |= OptimizeMemoryInst(I, SI->getOperand(1),
860 SI->getOperand(0)->getType(),
862 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
863 if (GEPI->hasAllZeroIndices()) {
864 /// The GEP operand must be a pointer, so must its result -> BitCast
865 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
866 GEPI->getName(), GEPI);
867 GEPI->replaceAllUsesWith(NC);
868 GEPI->eraseFromParent();
872 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
873 // If we found an inline asm expession, and if the target knows how to
874 // lower it to normal LLVM code, do so now.
875 if (TLI && isa<InlineAsm>(CI->getCalledValue()))
876 if (const TargetAsmInfo *TAI =
877 TLI->getTargetMachine().getTargetAsmInfo()) {
878 if (TAI->ExpandInlineAsm(CI)) {
880 // Avoid processing instructions out of order, which could cause
881 // reuse before a value is defined.
884 // Sink address computing for memory operands into the block.
885 MadeChange |= OptimizeInlineAsmInst(I, &(*CI), SunkAddrs);