1 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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.
13 //
14 //===----------------------------------------------------------------------===//
15
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/Dominators.h"
26 #include "llvm/Analysis/InstructionSimplify.h"
27 #include "llvm/Analysis/ProfileInfo.h"
28 #include "llvm/Target/TargetData.h"
29 #include "llvm/Target/TargetLowering.h"
30 #include "llvm/Transforms/Utils/AddrModeMatcher.h"
31 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
32 #include "llvm/Transforms/Utils/Local.h"
33 #include "llvm/Transforms/Utils/BuildLibCalls.h"
34 #include "llvm/ADT/DenseMap.h"
35 #include "llvm/ADT/SmallSet.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/Assembly/Writer.h"
38 #include "llvm/Support/CallSite.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/GetElementPtrTypeIterator.h"
42 #include "llvm/Support/PatternMatch.h"
43 #include "llvm/Support/raw_ostream.h"
44 #include "llvm/Support/IRBuilder.h"
45 #include "llvm/Support/ValueHandle.h"
46 using namespace llvm;
47 using namespace llvm::PatternMatch;
48
49 STATISTIC(NumBlocksElim, "Number of blocks eliminated");
50 STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated");
51 STATISTIC(NumGEPsElim, "Number of GEPs converted to casts");
52 STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "
53 "sunken Cmps");
54 STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "
55 "of sunken Casts");
56 STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "
57 "computations were sunk");
58 STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads");
59 STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized");
60 STATISTIC(NumRetsDup, "Number of return instructions duplicated");
61 STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved");
62
63 static cl::opt<bool> DisableBranchOpts(
64 "disable-cgp-branch-opts", cl::Hidden, cl::init(false),
65 cl::desc("Disable branch optimizations in CodeGenPrepare"));
66
67 namespace {
68 class CodeGenPrepare : public FunctionPass {
69 /// TLI - Keep a pointer of a TargetLowering to consult for determining
70 /// transformation profitability.
71 const TargetLowering *TLI;
72 DominatorTree *DT;
73 ProfileInfo *PFI;
74
75 /// CurInstIterator - As we scan instructions optimizing them, this is the
76 /// next instruction to optimize. Xforms that can invalidate this should
77 /// update it.
78 BasicBlock::iterator CurInstIterator;
79
80 /// Keeps track of non-local addresses that have been sunk into a block.
81 /// This allows us to avoid inserting duplicate code for blocks with
82 /// multiple load/stores of the same address.
83 DenseMap<Value*, Value*> SunkAddrs;
84
85 /// ModifiedDT - If CFG is modified in anyway, dominator tree may need to
86 /// be updated.
87 bool ModifiedDT;
88
89 public:
90 static char ID; // Pass identification, replacement for typeid
CodeGenPrepare(const TargetLowering * tli=0)91 explicit CodeGenPrepare(const TargetLowering *tli = 0)
92 : FunctionPass(ID), TLI(tli) {
93 initializeCodeGenPreparePass(*PassRegistry::getPassRegistry());
94 }
95 bool runOnFunction(Function &F);
96
getAnalysisUsage(AnalysisUsage & AU) const97 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
98 AU.addPreserved<DominatorTree>();
99 AU.addPreserved<ProfileInfo>();
100 }
101
102 private:
103 bool EliminateMostlyEmptyBlocks(Function &F);
104 bool CanMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
105 void EliminateMostlyEmptyBlock(BasicBlock *BB);
106 bool OptimizeBlock(BasicBlock &BB);
107 bool OptimizeInst(Instruction *I);
108 bool OptimizeMemoryInst(Instruction *I, Value *Addr, Type *AccessTy);
109 bool OptimizeInlineAsmInst(CallInst *CS);
110 bool OptimizeCallInst(CallInst *CI);
111 bool MoveExtToFormExtLoad(Instruction *I);
112 bool OptimizeExtUses(Instruction *I);
113 bool DupRetToEnableTailCallOpts(ReturnInst *RI);
114 bool PlaceDbgValues(Function &F);
115 };
116 }
117
118 char CodeGenPrepare::ID = 0;
119 INITIALIZE_PASS(CodeGenPrepare, "codegenprepare",
120 "Optimize for code generation", false, false)
121
createCodeGenPreparePass(const TargetLowering * TLI)122 FunctionPass *llvm::createCodeGenPreparePass(const TargetLowering *TLI) {
123 return new CodeGenPrepare(TLI);
124 }
125
runOnFunction(Function & F)126 bool CodeGenPrepare::runOnFunction(Function &F) {
127 bool EverMadeChange = false;
128
129 ModifiedDT = false;
130 DT = getAnalysisIfAvailable<DominatorTree>();
131 PFI = getAnalysisIfAvailable<ProfileInfo>();
132
133 // First pass, eliminate blocks that contain only PHI nodes and an
134 // unconditional branch.
135 EverMadeChange |= EliminateMostlyEmptyBlocks(F);
136
137 // llvm.dbg.value is far away from the value then iSel may not be able
138 // handle it properly. iSel will drop llvm.dbg.value if it can not
139 // find a node corresponding to the value.
140 EverMadeChange |= PlaceDbgValues(F);
141
142 bool MadeChange = true;
143 while (MadeChange) {
144 MadeChange = false;
145 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
146 BasicBlock *BB = I++;
147 MadeChange |= OptimizeBlock(*BB);
148 }
149 EverMadeChange |= MadeChange;
150 }
151
152 SunkAddrs.clear();
153
154 if (!DisableBranchOpts) {
155 MadeChange = false;
156 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
157 MadeChange |= ConstantFoldTerminator(BB, true);
158
159 if (MadeChange)
160 ModifiedDT = true;
161 EverMadeChange |= MadeChange;
162 }
163
164 if (ModifiedDT && DT)
165 DT->DT->recalculate(F);
166
167 return EverMadeChange;
168 }
169
170 /// EliminateMostlyEmptyBlocks - eliminate blocks that contain only PHI nodes,
171 /// debug info directives, and an unconditional branch. Passes before isel
172 /// (e.g. LSR/loopsimplify) often split edges in ways that are non-optimal for
173 /// isel. Start by eliminating these blocks so we can split them the way we
174 /// want them.
EliminateMostlyEmptyBlocks(Function & F)175 bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
176 bool MadeChange = false;
177 // Note that this intentionally skips the entry block.
178 for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
179 BasicBlock *BB = I++;
180
181 // If this block doesn't end with an uncond branch, ignore it.
182 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
183 if (!BI || !BI->isUnconditional())
184 continue;
185
186 // If the instruction before the branch (skipping debug info) isn't a phi
187 // node, then other stuff is happening here.
188 BasicBlock::iterator BBI = BI;
189 if (BBI != BB->begin()) {
190 --BBI;
191 while (isa<DbgInfoIntrinsic>(BBI)) {
192 if (BBI == BB->begin())
193 break;
194 --BBI;
195 }
196 if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI))
197 continue;
198 }
199
200 // Do not break infinite loops.
201 BasicBlock *DestBB = BI->getSuccessor(0);
202 if (DestBB == BB)
203 continue;
204
205 if (!CanMergeBlocks(BB, DestBB))
206 continue;
207
208 EliminateMostlyEmptyBlock(BB);
209 MadeChange = true;
210 }
211 return MadeChange;
212 }
213
214 /// CanMergeBlocks - Return true if we can merge BB into DestBB if there is a
215 /// single uncond branch between them, and BB contains no other non-phi
216 /// instructions.
CanMergeBlocks(const BasicBlock * BB,const BasicBlock * DestBB) const217 bool CodeGenPrepare::CanMergeBlocks(const BasicBlock *BB,
218 const BasicBlock *DestBB) const {
219 // We only want to eliminate blocks whose phi nodes are used by phi nodes in
220 // the successor. If there are more complex condition (e.g. preheaders),
221 // don't mess around with them.
222 BasicBlock::const_iterator BBI = BB->begin();
223 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
224 for (Value::const_use_iterator UI = PN->use_begin(), E = PN->use_end();
225 UI != E; ++UI) {
226 const Instruction *User = cast<Instruction>(*UI);
227 if (User->getParent() != DestBB || !isa<PHINode>(User))
228 return false;
229 // If User is inside DestBB block and it is a PHINode then check
230 // incoming value. If incoming value is not from BB then this is
231 // a complex condition (e.g. preheaders) we want to avoid here.
232 if (User->getParent() == DestBB) {
233 if (const PHINode *UPN = dyn_cast<PHINode>(User))
234 for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) {
235 Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I));
236 if (Insn && Insn->getParent() == BB &&
237 Insn->getParent() != UPN->getIncomingBlock(I))
238 return false;
239 }
240 }
241 }
242 }
243
244 // If BB and DestBB contain any common predecessors, then the phi nodes in BB
245 // and DestBB may have conflicting incoming values for the block. If so, we
246 // can't merge the block.
247 const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin());
248 if (!DestBBPN) return true; // no conflict.
249
250 // Collect the preds of BB.
251 SmallPtrSet<const BasicBlock*, 16> BBPreds;
252 if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
253 // It is faster to get preds from a PHI than with pred_iterator.
254 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
255 BBPreds.insert(BBPN->getIncomingBlock(i));
256 } else {
257 BBPreds.insert(pred_begin(BB), pred_end(BB));
258 }
259
260 // Walk the preds of DestBB.
261 for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) {
262 BasicBlock *Pred = DestBBPN->getIncomingBlock(i);
263 if (BBPreds.count(Pred)) { // Common predecessor?
264 BBI = DestBB->begin();
265 while (const PHINode *PN = dyn_cast<PHINode>(BBI++)) {
266 const Value *V1 = PN->getIncomingValueForBlock(Pred);
267 const Value *V2 = PN->getIncomingValueForBlock(BB);
268
269 // If V2 is a phi node in BB, look up what the mapped value will be.
270 if (const PHINode *V2PN = dyn_cast<PHINode>(V2))
271 if (V2PN->getParent() == BB)
272 V2 = V2PN->getIncomingValueForBlock(Pred);
273
274 // If there is a conflict, bail out.
275 if (V1 != V2) return false;
276 }
277 }
278 }
279
280 return true;
281 }
282
283
284 /// EliminateMostlyEmptyBlock - Eliminate a basic block that have only phi's and
285 /// an unconditional branch in it.
EliminateMostlyEmptyBlock(BasicBlock * BB)286 void CodeGenPrepare::EliminateMostlyEmptyBlock(BasicBlock *BB) {
287 BranchInst *BI = cast<BranchInst>(BB->getTerminator());
288 BasicBlock *DestBB = BI->getSuccessor(0);
289
290 DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n" << *BB << *DestBB);
291
292 // If the destination block has a single pred, then this is a trivial edge,
293 // just collapse it.
294 if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) {
295 if (SinglePred != DestBB) {
296 // Remember if SinglePred was the entry block of the function. If so, we
297 // will need to move BB back to the entry position.
298 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock();
299 MergeBasicBlockIntoOnlyPred(DestBB, this);
300
301 if (isEntry && BB != &BB->getParent()->getEntryBlock())
302 BB->moveBefore(&BB->getParent()->getEntryBlock());
303
304 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
305 return;
306 }
307 }
308
309 // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB
310 // to handle the new incoming edges it is about to have.
311 PHINode *PN;
312 for (BasicBlock::iterator BBI = DestBB->begin();
313 (PN = dyn_cast<PHINode>(BBI)); ++BBI) {
314 // Remove the incoming value for BB, and remember it.
315 Value *InVal = PN->removeIncomingValue(BB, false);
316
317 // Two options: either the InVal is a phi node defined in BB or it is some
318 // value that dominates BB.
319 PHINode *InValPhi = dyn_cast<PHINode>(InVal);
320 if (InValPhi && InValPhi->getParent() == BB) {
321 // Add all of the input values of the input PHI as inputs of this phi.
322 for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i)
323 PN->addIncoming(InValPhi->getIncomingValue(i),
324 InValPhi->getIncomingBlock(i));
325 } else {
326 // Otherwise, add one instance of the dominating value for each edge that
327 // we will be adding.
328 if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) {
329 for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i)
330 PN->addIncoming(InVal, BBPN->getIncomingBlock(i));
331 } else {
332 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
333 PN->addIncoming(InVal, *PI);
334 }
335 }
336 }
337
338 // The PHIs are now updated, change everything that refers to BB to use
339 // DestBB and remove BB.
340 BB->replaceAllUsesWith(DestBB);
341 if (DT && !ModifiedDT) {
342 BasicBlock *BBIDom = DT->getNode(BB)->getIDom()->getBlock();
343 BasicBlock *DestBBIDom = DT->getNode(DestBB)->getIDom()->getBlock();
344 BasicBlock *NewIDom = DT->findNearestCommonDominator(BBIDom, DestBBIDom);
345 DT->changeImmediateDominator(DestBB, NewIDom);
346 DT->eraseNode(BB);
347 }
348 if (PFI) {
349 PFI->replaceAllUses(BB, DestBB);
350 PFI->removeEdge(ProfileInfo::getEdge(BB, DestBB));
351 }
352 BB->eraseFromParent();
353 ++NumBlocksElim;
354
355 DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n");
356 }
357
358 /// OptimizeNoopCopyExpression - If the specified cast instruction is a noop
359 /// copy (e.g. it's casting from one pointer type to another, i32->i8 on PPC),
360 /// sink it into user blocks to reduce the number of virtual
361 /// registers that must be created and coalesced.
362 ///
363 /// Return true if any changes are made.
364 ///
OptimizeNoopCopyExpression(CastInst * CI,const TargetLowering & TLI)365 static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI){
366 // If this is a noop copy,
367 EVT SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
368 EVT DstVT = TLI.getValueType(CI->getType());
369
370 // This is an fp<->int conversion?
371 if (SrcVT.isInteger() != DstVT.isInteger())
372 return false;
373
374 // If this is an extension, it will be a zero or sign extension, which
375 // isn't a noop.
376 if (SrcVT.bitsLT(DstVT)) return false;
377
378 // If these values will be promoted, find out what they will be promoted
379 // to. This helps us consider truncates on PPC as noop copies when they
380 // are.
381 if (TLI.getTypeAction(CI->getContext(), SrcVT) ==
382 TargetLowering::TypePromoteInteger)
383 SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT);
384 if (TLI.getTypeAction(CI->getContext(), DstVT) ==
385 TargetLowering::TypePromoteInteger)
386 DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT);
387
388 // If, after promotion, these are the same types, this is a noop copy.
389 if (SrcVT != DstVT)
390 return false;
391
392 BasicBlock *DefBB = CI->getParent();
393
394 /// InsertedCasts - Only insert a cast in each block once.
395 DenseMap<BasicBlock*, CastInst*> InsertedCasts;
396
397 bool MadeChange = false;
398 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
399 UI != E; ) {
400 Use &TheUse = UI.getUse();
401 Instruction *User = cast<Instruction>(*UI);
402
403 // Figure out which BB this cast is used in. For PHI's this is the
404 // appropriate predecessor block.
405 BasicBlock *UserBB = User->getParent();
406 if (PHINode *PN = dyn_cast<PHINode>(User)) {
407 UserBB = PN->getIncomingBlock(UI);
408 }
409
410 // Preincrement use iterator so we don't invalidate it.
411 ++UI;
412
413 // If this user is in the same block as the cast, don't change the cast.
414 if (UserBB == DefBB) continue;
415
416 // If we have already inserted a cast into this block, use it.
417 CastInst *&InsertedCast = InsertedCasts[UserBB];
418
419 if (!InsertedCast) {
420 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
421 InsertedCast =
422 CastInst::Create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "",
423 InsertPt);
424 MadeChange = true;
425 }
426
427 // Replace a use of the cast with a use of the new cast.
428 TheUse = InsertedCast;
429 ++NumCastUses;
430 }
431
432 // If we removed all uses, nuke the cast.
433 if (CI->use_empty()) {
434 CI->eraseFromParent();
435 MadeChange = true;
436 }
437
438 return MadeChange;
439 }
440
441 /// OptimizeCmpExpression - sink the given CmpInst into user blocks to reduce
442 /// the number of virtual registers that must be created and coalesced. This is
443 /// a clear win except on targets with multiple condition code registers
444 /// (PowerPC), where it might lose; some adjustment may be wanted there.
445 ///
446 /// Return true if any changes are made.
OptimizeCmpExpression(CmpInst * CI)447 static bool OptimizeCmpExpression(CmpInst *CI) {
448 BasicBlock *DefBB = CI->getParent();
449
450 /// InsertedCmp - Only insert a cmp in each block once.
451 DenseMap<BasicBlock*, CmpInst*> InsertedCmps;
452
453 bool MadeChange = false;
454 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
455 UI != E; ) {
456 Use &TheUse = UI.getUse();
457 Instruction *User = cast<Instruction>(*UI);
458
459 // Preincrement use iterator so we don't invalidate it.
460 ++UI;
461
462 // Don't bother for PHI nodes.
463 if (isa<PHINode>(User))
464 continue;
465
466 // Figure out which BB this cmp is used in.
467 BasicBlock *UserBB = User->getParent();
468
469 // If this user is in the same block as the cmp, don't change the cmp.
470 if (UserBB == DefBB) continue;
471
472 // If we have already inserted a cmp into this block, use it.
473 CmpInst *&InsertedCmp = InsertedCmps[UserBB];
474
475 if (!InsertedCmp) {
476 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
477 InsertedCmp =
478 CmpInst::Create(CI->getOpcode(),
479 CI->getPredicate(), CI->getOperand(0),
480 CI->getOperand(1), "", InsertPt);
481 MadeChange = true;
482 }
483
484 // Replace a use of the cmp with a use of the new cmp.
485 TheUse = InsertedCmp;
486 ++NumCmpUses;
487 }
488
489 // If we removed all uses, nuke the cmp.
490 if (CI->use_empty())
491 CI->eraseFromParent();
492
493 return MadeChange;
494 }
495
496 namespace {
497 class CodeGenPrepareFortifiedLibCalls : public SimplifyFortifiedLibCalls {
498 protected:
replaceCall(Value * With)499 void replaceCall(Value *With) {
500 CI->replaceAllUsesWith(With);
501 CI->eraseFromParent();
502 }
isFoldable(unsigned SizeCIOp,unsigned,bool) const503 bool isFoldable(unsigned SizeCIOp, unsigned, bool) const {
504 if (ConstantInt *SizeCI =
505 dyn_cast<ConstantInt>(CI->getArgOperand(SizeCIOp)))
506 return SizeCI->isAllOnesValue();
507 return false;
508 }
509 };
510 } // end anonymous namespace
511
OptimizeCallInst(CallInst * CI)512 bool CodeGenPrepare::OptimizeCallInst(CallInst *CI) {
513 BasicBlock *BB = CI->getParent();
514
515 // Lower inline assembly if we can.
516 // If we found an inline asm expession, and if the target knows how to
517 // lower it to normal LLVM code, do so now.
518 if (TLI && isa<InlineAsm>(CI->getCalledValue())) {
519 if (TLI->ExpandInlineAsm(CI)) {
520 // Avoid invalidating the iterator.
521 CurInstIterator = BB->begin();
522 // Avoid processing instructions out of order, which could cause
523 // reuse before a value is defined.
524 SunkAddrs.clear();
525 return true;
526 }
527 // Sink address computing for memory operands into the block.
528 if (OptimizeInlineAsmInst(CI))
529 return true;
530 }
531
532 // Lower all uses of llvm.objectsize.*
533 IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
534 if (II && II->getIntrinsicID() == Intrinsic::objectsize) {
535 bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
536 Type *ReturnTy = CI->getType();
537 Constant *RetVal = ConstantInt::get(ReturnTy, Min ? 0 : -1ULL);
538
539 // Substituting this can cause recursive simplifications, which can
540 // invalidate our iterator. Use a WeakVH to hold onto it in case this
541 // happens.
542 WeakVH IterHandle(CurInstIterator);
543
544 ReplaceAndSimplifyAllUses(CI, RetVal, TLI ? TLI->getTargetData() : 0,
545 ModifiedDT ? 0 : DT);
546
547 // If the iterator instruction was recursively deleted, start over at the
548 // start of the block.
549 if (IterHandle != CurInstIterator) {
550 CurInstIterator = BB->begin();
551 SunkAddrs.clear();
552 }
553 return true;
554 }
555
556 // From here on out we're working with named functions.
557 if (CI->getCalledFunction() == 0) return false;
558
559 // We'll need TargetData from here on out.
560 const TargetData *TD = TLI ? TLI->getTargetData() : 0;
561 if (!TD) return false;
562
563 // Lower all default uses of _chk calls. This is very similar
564 // to what InstCombineCalls does, but here we are only lowering calls
565 // that have the default "don't know" as the objectsize. Anything else
566 // should be left alone.
567 CodeGenPrepareFortifiedLibCalls Simplifier;
568 return Simplifier.fold(CI, TD);
569 }
570
571 /// DupRetToEnableTailCallOpts - Look for opportunities to duplicate return
572 /// instructions to the predecessor to enable tail call optimizations. The
573 /// case it is currently looking for is:
574 /// bb0:
575 /// %tmp0 = tail call i32 @f0()
576 /// br label %return
577 /// bb1:
578 /// %tmp1 = tail call i32 @f1()
579 /// br label %return
580 /// bb2:
581 /// %tmp2 = tail call i32 @f2()
582 /// br label %return
583 /// return:
584 /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ]
585 /// ret i32 %retval
586 ///
587 /// =>
588 ///
589 /// bb0:
590 /// %tmp0 = tail call i32 @f0()
591 /// ret i32 %tmp0
592 /// bb1:
593 /// %tmp1 = tail call i32 @f1()
594 /// ret i32 %tmp1
595 /// bb2:
596 /// %tmp2 = tail call i32 @f2()
597 /// ret i32 %tmp2
598 ///
DupRetToEnableTailCallOpts(ReturnInst * RI)599 bool CodeGenPrepare::DupRetToEnableTailCallOpts(ReturnInst *RI) {
600 if (!TLI)
601 return false;
602
603 Value *V = RI->getReturnValue();
604 PHINode *PN = V ? dyn_cast<PHINode>(V) : NULL;
605 if (V && !PN)
606 return false;
607
608 BasicBlock *BB = RI->getParent();
609 if (PN && PN->getParent() != BB)
610 return false;
611
612 // It's not safe to eliminate the sign / zero extension of the return value.
613 // See llvm::isInTailCallPosition().
614 const Function *F = BB->getParent();
615 unsigned CallerRetAttr = F->getAttributes().getRetAttributes();
616 if ((CallerRetAttr & Attribute::ZExt) || (CallerRetAttr & Attribute::SExt))
617 return false;
618
619 // Make sure there are no instructions between the PHI and return, or that the
620 // return is the first instruction in the block.
621 if (PN) {
622 BasicBlock::iterator BI = BB->begin();
623 do { ++BI; } while (isa<DbgInfoIntrinsic>(BI));
624 if (&*BI != RI)
625 return false;
626 } else {
627 BasicBlock::iterator BI = BB->begin();
628 while (isa<DbgInfoIntrinsic>(BI)) ++BI;
629 if (&*BI != RI)
630 return false;
631 }
632
633 /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail
634 /// call.
635 SmallVector<CallInst*, 4> TailCalls;
636 if (PN) {
637 for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) {
638 CallInst *CI = dyn_cast<CallInst>(PN->getIncomingValue(I));
639 // Make sure the phi value is indeed produced by the tail call.
640 if (CI && CI->hasOneUse() && CI->getParent() == PN->getIncomingBlock(I) &&
641 TLI->mayBeEmittedAsTailCall(CI))
642 TailCalls.push_back(CI);
643 }
644 } else {
645 SmallPtrSet<BasicBlock*, 4> VisitedBBs;
646 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); PI != PE; ++PI) {
647 if (!VisitedBBs.insert(*PI))
648 continue;
649
650 BasicBlock::InstListType &InstList = (*PI)->getInstList();
651 BasicBlock::InstListType::reverse_iterator RI = InstList.rbegin();
652 BasicBlock::InstListType::reverse_iterator RE = InstList.rend();
653 do { ++RI; } while (RI != RE && isa<DbgInfoIntrinsic>(&*RI));
654 if (RI == RE)
655 continue;
656
657 CallInst *CI = dyn_cast<CallInst>(&*RI);
658 if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI))
659 TailCalls.push_back(CI);
660 }
661 }
662
663 bool Changed = false;
664 for (unsigned i = 0, e = TailCalls.size(); i != e; ++i) {
665 CallInst *CI = TailCalls[i];
666 CallSite CS(CI);
667
668 // Conservatively require the attributes of the call to match those of the
669 // return. Ignore noalias because it doesn't affect the call sequence.
670 unsigned CalleeRetAttr = CS.getAttributes().getRetAttributes();
671 if ((CalleeRetAttr ^ CallerRetAttr) & ~Attribute::NoAlias)
672 continue;
673
674 // Make sure the call instruction is followed by an unconditional branch to
675 // the return block.
676 BasicBlock *CallBB = CI->getParent();
677 BranchInst *BI = dyn_cast<BranchInst>(CallBB->getTerminator());
678 if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB)
679 continue;
680
681 // Duplicate the return into CallBB.
682 (void)FoldReturnIntoUncondBranch(RI, BB, CallBB);
683 ModifiedDT = Changed = true;
684 ++NumRetsDup;
685 }
686
687 // If we eliminated all predecessors of the block, delete the block now.
688 if (Changed && pred_begin(BB) == pred_end(BB))
689 BB->eraseFromParent();
690
691 return Changed;
692 }
693
694 //===----------------------------------------------------------------------===//
695 // Memory Optimization
696 //===----------------------------------------------------------------------===//
697
698 /// IsNonLocalValue - Return true if the specified values are defined in a
699 /// different basic block than BB.
IsNonLocalValue(Value * V,BasicBlock * BB)700 static bool IsNonLocalValue(Value *V, BasicBlock *BB) {
701 if (Instruction *I = dyn_cast<Instruction>(V))
702 return I->getParent() != BB;
703 return false;
704 }
705
706 /// OptimizeMemoryInst - Load and Store Instructions often have
707 /// addressing modes that can do significant amounts of computation. As such,
708 /// instruction selection will try to get the load or store to do as much
709 /// computation as possible for the program. The problem is that isel can only
710 /// see within a single block. As such, we sink as much legal addressing mode
711 /// stuff into the block as possible.
712 ///
713 /// This method is used to optimize both load/store and inline asms with memory
714 /// operands.
OptimizeMemoryInst(Instruction * MemoryInst,Value * Addr,Type * AccessTy)715 bool CodeGenPrepare::OptimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
716 Type *AccessTy) {
717 Value *Repl = Addr;
718
719 // Try to collapse single-value PHI nodes. This is necessary to undo
720 // unprofitable PRE transformations.
721 SmallVector<Value*, 8> worklist;
722 SmallPtrSet<Value*, 16> Visited;
723 worklist.push_back(Addr);
724
725 // Use a worklist to iteratively look through PHI nodes, and ensure that
726 // the addressing mode obtained from the non-PHI roots of the graph
727 // are equivalent.
728 Value *Consensus = 0;
729 unsigned NumUsesConsensus = 0;
730 bool IsNumUsesConsensusValid = false;
731 SmallVector<Instruction*, 16> AddrModeInsts;
732 ExtAddrMode AddrMode;
733 while (!worklist.empty()) {
734 Value *V = worklist.back();
735 worklist.pop_back();
736
737 // Break use-def graph loops.
738 if (!Visited.insert(V)) {
739 Consensus = 0;
740 break;
741 }
742
743 // For a PHI node, push all of its incoming values.
744 if (PHINode *P = dyn_cast<PHINode>(V)) {
745 for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i)
746 worklist.push_back(P->getIncomingValue(i));
747 continue;
748 }
749
750 // For non-PHIs, determine the addressing mode being computed.
751 SmallVector<Instruction*, 16> NewAddrModeInsts;
752 ExtAddrMode NewAddrMode =
753 AddressingModeMatcher::Match(V, AccessTy, MemoryInst,
754 NewAddrModeInsts, *TLI);
755
756 // This check is broken into two cases with very similar code to avoid using
757 // getNumUses() as much as possible. Some values have a lot of uses, so
758 // calling getNumUses() unconditionally caused a significant compile-time
759 // regression.
760 if (!Consensus) {
761 Consensus = V;
762 AddrMode = NewAddrMode;
763 AddrModeInsts = NewAddrModeInsts;
764 continue;
765 } else if (NewAddrMode == AddrMode) {
766 if (!IsNumUsesConsensusValid) {
767 NumUsesConsensus = Consensus->getNumUses();
768 IsNumUsesConsensusValid = true;
769 }
770
771 // Ensure that the obtained addressing mode is equivalent to that obtained
772 // for all other roots of the PHI traversal. Also, when choosing one
773 // such root as representative, select the one with the most uses in order
774 // to keep the cost modeling heuristics in AddressingModeMatcher
775 // applicable.
776 unsigned NumUses = V->getNumUses();
777 if (NumUses > NumUsesConsensus) {
778 Consensus = V;
779 NumUsesConsensus = NumUses;
780 AddrModeInsts = NewAddrModeInsts;
781 }
782 continue;
783 }
784
785 Consensus = 0;
786 break;
787 }
788
789 // If the addressing mode couldn't be determined, or if multiple different
790 // ones were determined, bail out now.
791 if (!Consensus) return false;
792
793 // Check to see if any of the instructions supersumed by this addr mode are
794 // non-local to I's BB.
795 bool AnyNonLocal = false;
796 for (unsigned i = 0, e = AddrModeInsts.size(); i != e; ++i) {
797 if (IsNonLocalValue(AddrModeInsts[i], MemoryInst->getParent())) {
798 AnyNonLocal = true;
799 break;
800 }
801 }
802
803 // If all the instructions matched are already in this BB, don't do anything.
804 if (!AnyNonLocal) {
805 DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrMode << "\n");
806 return false;
807 }
808
809 // Insert this computation right after this user. Since our caller is
810 // scanning from the top of the BB to the bottom, reuse of the expr are
811 // guaranteed to happen later.
812 IRBuilder<> Builder(MemoryInst);
813
814 // Now that we determined the addressing expression we want to use and know
815 // that we have to sink it into this block. Check to see if we have already
816 // done this for some other load/store instr in this block. If so, reuse the
817 // computation.
818 Value *&SunkAddr = SunkAddrs[Addr];
819 if (SunkAddr) {
820 DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrMode << " for "
821 << *MemoryInst);
822 if (SunkAddr->getType() != Addr->getType())
823 SunkAddr = Builder.CreateBitCast(SunkAddr, Addr->getType());
824 } else {
825 DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrMode << " for "
826 << *MemoryInst);
827 Type *IntPtrTy =
828 TLI->getTargetData()->getIntPtrType(AccessTy->getContext());
829
830 Value *Result = 0;
831
832 // Start with the base register. Do this first so that subsequent address
833 // matching finds it last, which will prevent it from trying to match it
834 // as the scaled value in case it happens to be a mul. That would be
835 // problematic if we've sunk a different mul for the scale, because then
836 // we'd end up sinking both muls.
837 if (AddrMode.BaseReg) {
838 Value *V = AddrMode.BaseReg;
839 if (V->getType()->isPointerTy())
840 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
841 if (V->getType() != IntPtrTy)
842 V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr");
843 Result = V;
844 }
845
846 // Add the scale value.
847 if (AddrMode.Scale) {
848 Value *V = AddrMode.ScaledReg;
849 if (V->getType() == IntPtrTy) {
850 // done.
851 } else if (V->getType()->isPointerTy()) {
852 V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr");
853 } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() <
854 cast<IntegerType>(V->getType())->getBitWidth()) {
855 V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr");
856 } else {
857 V = Builder.CreateSExt(V, IntPtrTy, "sunkaddr");
858 }
859 if (AddrMode.Scale != 1)
860 V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale),
861 "sunkaddr");
862 if (Result)
863 Result = Builder.CreateAdd(Result, V, "sunkaddr");
864 else
865 Result = V;
866 }
867
868 // Add in the BaseGV if present.
869 if (AddrMode.BaseGV) {
870 Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr");
871 if (Result)
872 Result = Builder.CreateAdd(Result, V, "sunkaddr");
873 else
874 Result = V;
875 }
876
877 // Add in the Base Offset if present.
878 if (AddrMode.BaseOffs) {
879 Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs);
880 if (Result)
881 Result = Builder.CreateAdd(Result, V, "sunkaddr");
882 else
883 Result = V;
884 }
885
886 if (Result == 0)
887 SunkAddr = Constant::getNullValue(Addr->getType());
888 else
889 SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr");
890 }
891
892 MemoryInst->replaceUsesOfWith(Repl, SunkAddr);
893
894 // If we have no uses, recursively delete the value and all dead instructions
895 // using it.
896 if (Repl->use_empty()) {
897 // This can cause recursive deletion, which can invalidate our iterator.
898 // Use a WeakVH to hold onto it in case this happens.
899 WeakVH IterHandle(CurInstIterator);
900 BasicBlock *BB = CurInstIterator->getParent();
901
902 RecursivelyDeleteTriviallyDeadInstructions(Repl);
903
904 if (IterHandle != CurInstIterator) {
905 // If the iterator instruction was recursively deleted, start over at the
906 // start of the block.
907 CurInstIterator = BB->begin();
908 SunkAddrs.clear();
909 } else {
910 // This address is now available for reassignment, so erase the table
911 // entry; we don't want to match some completely different instruction.
912 SunkAddrs[Addr] = 0;
913 }
914 }
915 ++NumMemoryInsts;
916 return true;
917 }
918
919 /// OptimizeInlineAsmInst - If there are any memory operands, use
920 /// OptimizeMemoryInst to sink their address computing into the block when
921 /// possible / profitable.
OptimizeInlineAsmInst(CallInst * CS)922 bool CodeGenPrepare::OptimizeInlineAsmInst(CallInst *CS) {
923 bool MadeChange = false;
924
925 TargetLowering::AsmOperandInfoVector
926 TargetConstraints = TLI->ParseConstraints(CS);
927 unsigned ArgNo = 0;
928 for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
929 TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i];
930
931 // Compute the constraint code and ConstraintType to use.
932 TLI->ComputeConstraintToUse(OpInfo, SDValue());
933
934 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
935 OpInfo.isIndirect) {
936 Value *OpVal = CS->getArgOperand(ArgNo++);
937 MadeChange |= OptimizeMemoryInst(CS, OpVal, OpVal->getType());
938 } else if (OpInfo.Type == InlineAsm::isInput)
939 ArgNo++;
940 }
941
942 return MadeChange;
943 }
944
945 /// MoveExtToFormExtLoad - Move a zext or sext fed by a load into the same
946 /// basic block as the load, unless conditions are unfavorable. This allows
947 /// SelectionDAG to fold the extend into the load.
948 ///
MoveExtToFormExtLoad(Instruction * I)949 bool CodeGenPrepare::MoveExtToFormExtLoad(Instruction *I) {
950 // Look for a load being extended.
951 LoadInst *LI = dyn_cast<LoadInst>(I->getOperand(0));
952 if (!LI) return false;
953
954 // If they're already in the same block, there's nothing to do.
955 if (LI->getParent() == I->getParent())
956 return false;
957
958 // If the load has other users and the truncate is not free, this probably
959 // isn't worthwhile.
960 if (!LI->hasOneUse() &&
961 TLI && (TLI->isTypeLegal(TLI->getValueType(LI->getType())) ||
962 !TLI->isTypeLegal(TLI->getValueType(I->getType()))) &&
963 !TLI->isTruncateFree(I->getType(), LI->getType()))
964 return false;
965
966 // Check whether the target supports casts folded into loads.
967 unsigned LType;
968 if (isa<ZExtInst>(I))
969 LType = ISD::ZEXTLOAD;
970 else {
971 assert(isa<SExtInst>(I) && "Unexpected ext type!");
972 LType = ISD::SEXTLOAD;
973 }
974 if (TLI && !TLI->isLoadExtLegal(LType, TLI->getValueType(LI->getType())))
975 return false;
976
977 // Move the extend into the same block as the load, so that SelectionDAG
978 // can fold it.
979 I->removeFromParent();
980 I->insertAfter(LI);
981 ++NumExtsMoved;
982 return true;
983 }
984
OptimizeExtUses(Instruction * I)985 bool CodeGenPrepare::OptimizeExtUses(Instruction *I) {
986 BasicBlock *DefBB = I->getParent();
987
988 // If the result of a {s|z}ext and its source are both live out, rewrite all
989 // other uses of the source with result of extension.
990 Value *Src = I->getOperand(0);
991 if (Src->hasOneUse())
992 return false;
993
994 // Only do this xform if truncating is free.
995 if (TLI && !TLI->isTruncateFree(I->getType(), Src->getType()))
996 return false;
997
998 // Only safe to perform the optimization if the source is also defined in
999 // this block.
1000 if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent())
1001 return false;
1002
1003 bool DefIsLiveOut = false;
1004 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
1005 UI != E; ++UI) {
1006 Instruction *User = cast<Instruction>(*UI);
1007
1008 // Figure out which BB this ext is used in.
1009 BasicBlock *UserBB = User->getParent();
1010 if (UserBB == DefBB) continue;
1011 DefIsLiveOut = true;
1012 break;
1013 }
1014 if (!DefIsLiveOut)
1015 return false;
1016
1017 // Make sure non of the uses are PHI nodes.
1018 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1019 UI != E; ++UI) {
1020 Instruction *User = cast<Instruction>(*UI);
1021 BasicBlock *UserBB = User->getParent();
1022 if (UserBB == DefBB) continue;
1023 // Be conservative. We don't want this xform to end up introducing
1024 // reloads just before load / store instructions.
1025 if (isa<PHINode>(User) || isa<LoadInst>(User) || isa<StoreInst>(User))
1026 return false;
1027 }
1028
1029 // InsertedTruncs - Only insert one trunc in each block once.
1030 DenseMap<BasicBlock*, Instruction*> InsertedTruncs;
1031
1032 bool MadeChange = false;
1033 for (Value::use_iterator UI = Src->use_begin(), E = Src->use_end();
1034 UI != E; ++UI) {
1035 Use &TheUse = UI.getUse();
1036 Instruction *User = cast<Instruction>(*UI);
1037
1038 // Figure out which BB this ext is used in.
1039 BasicBlock *UserBB = User->getParent();
1040 if (UserBB == DefBB) continue;
1041
1042 // Both src and def are live in this block. Rewrite the use.
1043 Instruction *&InsertedTrunc = InsertedTruncs[UserBB];
1044
1045 if (!InsertedTrunc) {
1046 BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt();
1047 InsertedTrunc = new TruncInst(I, Src->getType(), "", InsertPt);
1048 }
1049
1050 // Replace a use of the {s|z}ext source with a use of the result.
1051 TheUse = InsertedTrunc;
1052 ++NumExtUses;
1053 MadeChange = true;
1054 }
1055
1056 return MadeChange;
1057 }
1058
OptimizeInst(Instruction * I)1059 bool CodeGenPrepare::OptimizeInst(Instruction *I) {
1060 if (PHINode *P = dyn_cast<PHINode>(I)) {
1061 // It is possible for very late stage optimizations (such as SimplifyCFG)
1062 // to introduce PHI nodes too late to be cleaned up. If we detect such a
1063 // trivial PHI, go ahead and zap it here.
1064 if (Value *V = SimplifyInstruction(P)) {
1065 P->replaceAllUsesWith(V);
1066 P->eraseFromParent();
1067 ++NumPHIsElim;
1068 return true;
1069 }
1070 return false;
1071 }
1072
1073 if (CastInst *CI = dyn_cast<CastInst>(I)) {
1074 // If the source of the cast is a constant, then this should have
1075 // already been constant folded. The only reason NOT to constant fold
1076 // it is if something (e.g. LSR) was careful to place the constant
1077 // evaluation in a block other than then one that uses it (e.g. to hoist
1078 // the address of globals out of a loop). If this is the case, we don't
1079 // want to forward-subst the cast.
1080 if (isa<Constant>(CI->getOperand(0)))
1081 return false;
1082
1083 if (TLI && OptimizeNoopCopyExpression(CI, *TLI))
1084 return true;
1085
1086 if (isa<ZExtInst>(I) || isa<SExtInst>(I)) {
1087 bool MadeChange = MoveExtToFormExtLoad(I);
1088 return MadeChange | OptimizeExtUses(I);
1089 }
1090 return false;
1091 }
1092
1093 if (CmpInst *CI = dyn_cast<CmpInst>(I))
1094 return OptimizeCmpExpression(CI);
1095
1096 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1097 if (TLI)
1098 return OptimizeMemoryInst(I, I->getOperand(0), LI->getType());
1099 return false;
1100 }
1101
1102 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1103 if (TLI)
1104 return OptimizeMemoryInst(I, SI->getOperand(1),
1105 SI->getOperand(0)->getType());
1106 return false;
1107 }
1108
1109 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
1110 if (GEPI->hasAllZeroIndices()) {
1111 /// The GEP operand must be a pointer, so must its result -> BitCast
1112 Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(),
1113 GEPI->getName(), GEPI);
1114 GEPI->replaceAllUsesWith(NC);
1115 GEPI->eraseFromParent();
1116 ++NumGEPsElim;
1117 OptimizeInst(NC);
1118 return true;
1119 }
1120 return false;
1121 }
1122
1123 if (CallInst *CI = dyn_cast<CallInst>(I))
1124 return OptimizeCallInst(CI);
1125
1126 if (ReturnInst *RI = dyn_cast<ReturnInst>(I))
1127 return DupRetToEnableTailCallOpts(RI);
1128
1129 return false;
1130 }
1131
1132 // In this pass we look for GEP and cast instructions that are used
1133 // across basic blocks and rewrite them to improve basic-block-at-a-time
1134 // selection.
OptimizeBlock(BasicBlock & BB)1135 bool CodeGenPrepare::OptimizeBlock(BasicBlock &BB) {
1136 SunkAddrs.clear();
1137 bool MadeChange = false;
1138
1139 CurInstIterator = BB.begin();
1140 for (BasicBlock::iterator E = BB.end(); CurInstIterator != E; )
1141 MadeChange |= OptimizeInst(CurInstIterator++);
1142
1143 return MadeChange;
1144 }
1145
1146 // llvm.dbg.value is far away from the value then iSel may not be able
1147 // handle it properly. iSel will drop llvm.dbg.value if it can not
1148 // find a node corresponding to the value.
PlaceDbgValues(Function & F)1149 bool CodeGenPrepare::PlaceDbgValues(Function &F) {
1150 bool MadeChange = false;
1151 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) {
1152 Instruction *PrevNonDbgInst = NULL;
1153 for (BasicBlock::iterator BI = I->begin(), BE = I->end(); BI != BE;) {
1154 Instruction *Insn = BI; ++BI;
1155 DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn);
1156 if (!DVI) {
1157 PrevNonDbgInst = Insn;
1158 continue;
1159 }
1160
1161 Instruction *VI = dyn_cast_or_null<Instruction>(DVI->getValue());
1162 if (VI && VI != PrevNonDbgInst && !VI->isTerminator()) {
1163 DEBUG(dbgs() << "Moving Debug Value before :\n" << *DVI << ' ' << *VI);
1164 DVI->removeFromParent();
1165 if (isa<PHINode>(VI))
1166 DVI->insertBefore(VI->getParent()->getFirstInsertionPt());
1167 else
1168 DVI->insertAfter(VI);
1169 MadeChange = true;
1170 ++NumDbgValueMoved;
1171 }
1172 }
1173 }
1174 return MadeChange;
1175 }
1176