1 //===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===//
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 file promotes memory references to be register references. It promotes
11 // alloca instructions which only have loads and stores as uses. An alloca is
12 // transformed by using dominator frontiers to place PHI nodes, then traversing
13 // the function in depth-first order to rewrite loads and stores as appropriate.
14 // This is just the standard SSA construction algorithm to construct "pruned"
17 //===----------------------------------------------------------------------===//
19 #define DEBUG_TYPE "mem2reg"
20 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/Function.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/IntrinsicInst.h"
26 #include "llvm/Metadata.h"
27 #include "llvm/Analysis/AliasSetTracker.h"
28 #include "llvm/Analysis/DebugInfo.h"
29 #include "llvm/Analysis/DominanceFrontier.h"
30 #include "llvm/Analysis/InstructionSimplify.h"
31 #include "llvm/ADT/DenseMap.h"
32 #include "llvm/ADT/SmallPtrSet.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/ADT/STLExtras.h"
36 #include "llvm/Support/CFG.h"
40 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
41 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
42 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
43 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
47 struct DenseMapInfo<std::pair<BasicBlock*, unsigned> > {
48 typedef std::pair<BasicBlock*, unsigned> EltTy;
49 static inline EltTy getEmptyKey() {
50 return EltTy(reinterpret_cast<BasicBlock*>(-1), ~0U);
52 static inline EltTy getTombstoneKey() {
53 return EltTy(reinterpret_cast<BasicBlock*>(-2), 0U);
55 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
56 return DenseMapInfo<void*>::getHashValue(Val.first) + Val.second*2;
58 static bool isEqual(const EltTy &LHS, const EltTy &RHS) {
64 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
65 /// This is true if there are only loads and stores to the alloca.
67 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
68 // FIXME: If the memory unit is of pointer or integer type, we can permit
69 // assignments to subsections of the memory unit.
71 // Only allow direct and non-volatile loads and stores...
72 for (Value::const_use_iterator UI = AI->use_begin(), UE = AI->use_end();
73 UI != UE; ++UI) { // Loop over all of the uses of the alloca
75 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
78 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
79 if (SI->getOperand(0) == AI)
80 return false; // Don't allow a store OF the AI, only INTO the AI.
91 /// FindAllocaDbgDeclare - Finds the llvm.dbg.declare intrinsic describing the
92 /// alloca 'V', if any.
93 static DbgDeclareInst *FindAllocaDbgDeclare(Value *V) {
94 if (MDNode *DebugNode = MDNode::getIfExists(V->getContext(), &V, 1))
95 for (Value::use_iterator UI = DebugNode->use_begin(),
96 E = DebugNode->use_end(); UI != E; ++UI)
97 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
106 // Data package used by RenamePass()
107 class RenamePassData {
109 typedef std::vector<Value *> ValVector;
111 RenamePassData() : BB(NULL), Pred(NULL), Values() {}
112 RenamePassData(BasicBlock *B, BasicBlock *P,
113 const ValVector &V) : BB(B), Pred(P), Values(V) {}
118 void swap(RenamePassData &RHS) {
119 std::swap(BB, RHS.BB);
120 std::swap(Pred, RHS.Pred);
121 Values.swap(RHS.Values);
125 /// LargeBlockInfo - This assigns and keeps a per-bb relative ordering of
126 /// load/store instructions in the block that directly load or store an alloca.
128 /// This functionality is important because it avoids scanning large basic
129 /// blocks multiple times when promoting many allocas in the same block.
130 class LargeBlockInfo {
131 /// InstNumbers - For each instruction that we track, keep the index of the
132 /// instruction. The index starts out as the number of the instruction from
133 /// the start of the block.
134 DenseMap<const Instruction *, unsigned> InstNumbers;
137 /// isInterestingInstruction - This code only looks at accesses to allocas.
138 static bool isInterestingInstruction(const Instruction *I) {
139 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
140 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
143 /// getInstructionIndex - Get or calculate the index of the specified
145 unsigned getInstructionIndex(const Instruction *I) {
146 assert(isInterestingInstruction(I) &&
147 "Not a load/store to/from an alloca?");
149 // If we already have this instruction number, return it.
150 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
151 if (It != InstNumbers.end()) return It->second;
153 // Scan the whole block to get the instruction. This accumulates
154 // information for every interesting instruction in the block, in order to
155 // avoid gratuitus rescans.
156 const BasicBlock *BB = I->getParent();
158 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end();
160 if (isInterestingInstruction(BBI))
161 InstNumbers[BBI] = InstNo++;
162 It = InstNumbers.find(I);
164 assert(It != InstNumbers.end() && "Didn't insert instruction?");
168 void deleteValue(const Instruction *I) {
169 InstNumbers.erase(I);
177 struct PromoteMem2Reg {
178 /// Allocas - The alloca instructions being promoted.
180 std::vector<AllocaInst*> Allocas;
182 DominanceFrontier &DF;
185 /// AST - An AliasSetTracker object to update. If null, don't update it.
187 AliasSetTracker *AST;
189 /// AllocaLookup - Reverse mapping of Allocas.
191 std::map<AllocaInst*, unsigned> AllocaLookup;
193 /// NewPhiNodes - The PhiNodes we're adding.
195 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
197 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
198 /// it corresponds to.
199 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
201 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
202 /// each alloca that is of pointer type, we keep track of what to copyValue
203 /// to the inserted PHI nodes here.
205 std::vector<Value*> PointerAllocaValues;
207 /// AllocaDbgDeclares - For each alloca, we keep track of the dbg.declare
208 /// intrinsic that describes it, if any, so that we can convert it to a
209 /// dbg.value intrinsic if the alloca gets promoted.
210 SmallVector<DbgDeclareInst*, 8> AllocaDbgDeclares;
212 /// Visited - The set of basic blocks the renamer has already visited.
214 SmallPtrSet<BasicBlock*, 16> Visited;
216 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
217 /// non-determinstic behavior.
218 DenseMap<BasicBlock*, unsigned> BBNumbers;
220 /// BBNumPreds - Lazily compute the number of predecessors a block has.
221 DenseMap<const BasicBlock*, unsigned> BBNumPreds;
223 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
224 DominanceFrontier &df, AliasSetTracker *ast)
225 : Allocas(A), DT(dt), DF(df), DIF(0), AST(ast) {}
232 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
234 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
235 return DT.dominates(BB1, BB2);
239 void RemoveFromAllocasList(unsigned &AllocaIdx) {
240 Allocas[AllocaIdx] = Allocas.back();
245 unsigned getNumPreds(const BasicBlock *BB) {
246 unsigned &NP = BBNumPreds[BB];
248 NP = std::distance(pred_begin(BB), pred_end(BB))+1;
252 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
254 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
255 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
256 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks);
258 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
259 LargeBlockInfo &LBI);
260 void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
261 LargeBlockInfo &LBI);
262 void ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI, StoreInst *SI);
265 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
266 RenamePassData::ValVector &IncVals,
267 std::vector<RenamePassData> &Worklist);
268 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version);
272 std::vector<BasicBlock*> DefiningBlocks;
273 std::vector<BasicBlock*> UsingBlocks;
275 StoreInst *OnlyStore;
276 BasicBlock *OnlyBlock;
277 bool OnlyUsedInOneBlock;
279 Value *AllocaPointerVal;
280 DbgDeclareInst *DbgDeclare;
283 DefiningBlocks.clear();
287 OnlyUsedInOneBlock = true;
288 AllocaPointerVal = 0;
292 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
294 void AnalyzeAlloca(AllocaInst *AI) {
297 // As we scan the uses of the alloca instruction, keep track of stores,
298 // and decide whether all of the loads and stores to the alloca are within
299 // the same basic block.
300 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
302 Instruction *User = cast<Instruction>(*UI++);
304 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
305 // Remember the basic blocks which define new values for the alloca
306 DefiningBlocks.push_back(SI->getParent());
307 AllocaPointerVal = SI->getOperand(0);
310 LoadInst *LI = cast<LoadInst>(User);
311 // Otherwise it must be a load instruction, keep track of variable
313 UsingBlocks.push_back(LI->getParent());
314 AllocaPointerVal = LI;
317 if (OnlyUsedInOneBlock) {
319 OnlyBlock = User->getParent();
320 else if (OnlyBlock != User->getParent())
321 OnlyUsedInOneBlock = false;
325 DbgDeclare = FindAllocaDbgDeclare(AI);
328 } // end of anonymous namespace
331 void PromoteMem2Reg::run() {
332 Function &F = *DF.getRoot()->getParent();
334 if (AST) PointerAllocaValues.resize(Allocas.size());
335 AllocaDbgDeclares.resize(Allocas.size());
340 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
341 AllocaInst *AI = Allocas[AllocaNum];
343 assert(isAllocaPromotable(AI) &&
344 "Cannot promote non-promotable alloca!");
345 assert(AI->getParent()->getParent() == &F &&
346 "All allocas should be in the same function, which is same as DF!");
348 if (AI->use_empty()) {
349 // If there are no uses of the alloca, just delete it now.
350 if (AST) AST->deleteValue(AI);
351 AI->eraseFromParent();
353 // Remove the alloca from the Allocas list, since it has been processed
354 RemoveFromAllocasList(AllocaNum);
359 // Calculate the set of read and write-locations for each alloca. This is
360 // analogous to finding the 'uses' and 'definitions' of each variable.
361 Info.AnalyzeAlloca(AI);
363 // If there is only a single store to this value, replace any loads of
364 // it that are directly dominated by the definition with the value stored.
365 if (Info.DefiningBlocks.size() == 1) {
366 RewriteSingleStoreAlloca(AI, Info, LBI);
368 // Finally, after the scan, check to see if the store is all that is left.
369 if (Info.UsingBlocks.empty()) {
370 // Record debuginfo for the store and remove the declaration's debuginfo.
371 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
372 ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore);
373 DDI->eraseFromParent();
375 // Remove the (now dead) store and alloca.
376 Info.OnlyStore->eraseFromParent();
377 LBI.deleteValue(Info.OnlyStore);
379 if (AST) AST->deleteValue(AI);
380 AI->eraseFromParent();
383 // The alloca has been processed, move on.
384 RemoveFromAllocasList(AllocaNum);
391 // If the alloca is only read and written in one basic block, just perform a
392 // linear sweep over the block to eliminate it.
393 if (Info.OnlyUsedInOneBlock) {
394 PromoteSingleBlockAlloca(AI, Info, LBI);
396 // Finally, after the scan, check to see if the stores are all that is
398 if (Info.UsingBlocks.empty()) {
400 // Remove the (now dead) stores and alloca.
401 while (!AI->use_empty()) {
402 StoreInst *SI = cast<StoreInst>(AI->use_back());
403 // Record debuginfo for the store before removing it.
404 if (DbgDeclareInst *DDI = Info.DbgDeclare)
405 ConvertDebugDeclareToDebugValue(DDI, SI);
406 SI->eraseFromParent();
410 if (AST) AST->deleteValue(AI);
411 AI->eraseFromParent();
414 // The alloca has been processed, move on.
415 RemoveFromAllocasList(AllocaNum);
417 // The alloca's debuginfo can be removed as well.
418 if (DbgDeclareInst *DDI = Info.DbgDeclare)
419 DDI->eraseFromParent();
426 // If we haven't computed a numbering for the BB's in the function, do so
428 if (BBNumbers.empty()) {
430 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
434 // If we have an AST to keep updated, remember some pointer value that is
435 // stored into the alloca.
437 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
439 // Remember the dbg.declare intrinsic describing this alloca, if any.
440 if (Info.DbgDeclare) AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare;
442 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
443 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
445 // At this point, we're committed to promoting the alloca using IDF's, and
446 // the standard SSA construction algorithm. Determine which blocks need PHI
447 // nodes and see if we can optimize out some work by avoiding insertion of
449 DetermineInsertionPoint(AI, AllocaNum, Info);
453 return; // All of the allocas must have been trivial!
458 // Set the incoming values for the basic block to be null values for all of
459 // the alloca's. We do this in case there is a load of a value that has not
460 // been stored yet. In this case, it will get this null value.
462 RenamePassData::ValVector Values(Allocas.size());
463 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
464 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
466 // Walks all basic blocks in the function performing the SSA rename algorithm
467 // and inserting the phi nodes we marked as necessary
469 std::vector<RenamePassData> RenamePassWorkList;
470 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
473 RPD.swap(RenamePassWorkList.back());
474 RenamePassWorkList.pop_back();
475 // RenamePass may add new worklist entries.
476 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
477 } while (!RenamePassWorkList.empty());
479 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
482 // Remove the allocas themselves from the function.
483 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
484 Instruction *A = Allocas[i];
486 // If there are any uses of the alloca instructions left, they must be in
487 // sections of dead code that were not processed on the dominance frontier.
488 // Just delete the users now.
491 A->replaceAllUsesWith(UndefValue::get(A->getType()));
492 if (AST) AST->deleteValue(A);
493 A->eraseFromParent();
496 // Remove alloca's dbg.declare instrinsics from the function.
497 for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i)
498 if (DbgDeclareInst *DDI = AllocaDbgDeclares[i])
499 DDI->eraseFromParent();
501 // Loop over all of the PHI nodes and see if there are any that we can get
502 // rid of because they merge all of the same incoming values. This can
503 // happen due to undef values coming into the PHI nodes. This process is
504 // iterative, because eliminating one PHI node can cause others to be removed.
505 bool EliminatedAPHI = true;
506 while (EliminatedAPHI) {
507 EliminatedAPHI = false;
509 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
510 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
511 PHINode *PN = I->second;
513 // If this PHI node merges one value and/or undefs, get the value.
514 if (Value *V = SimplifyInstruction(PN, 0, &DT)) {
515 if (AST && PN->getType()->isPointerTy())
516 AST->deleteValue(PN);
517 PN->replaceAllUsesWith(V);
518 PN->eraseFromParent();
519 NewPhiNodes.erase(I++);
520 EliminatedAPHI = true;
527 // At this point, the renamer has added entries to PHI nodes for all reachable
528 // code. Unfortunately, there may be unreachable blocks which the renamer
529 // hasn't traversed. If this is the case, the PHI nodes may not
530 // have incoming values for all predecessors. Loop over all PHI nodes we have
531 // created, inserting undef values if they are missing any incoming values.
533 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
534 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
535 // We want to do this once per basic block. As such, only process a block
536 // when we find the PHI that is the first entry in the block.
537 PHINode *SomePHI = I->second;
538 BasicBlock *BB = SomePHI->getParent();
539 if (&BB->front() != SomePHI)
542 // Only do work here if there the PHI nodes are missing incoming values. We
543 // know that all PHI nodes that were inserted in a block will have the same
544 // number of incoming values, so we can just check any of them.
545 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
548 // Get the preds for BB.
549 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
551 // Ok, now we know that all of the PHI nodes are missing entries for some
552 // basic blocks. Start by sorting the incoming predecessors for efficient
554 std::sort(Preds.begin(), Preds.end());
556 // Now we loop through all BB's which have entries in SomePHI and remove
557 // them from the Preds list.
558 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
559 // Do a log(n) search of the Preds list for the entry we want.
560 SmallVector<BasicBlock*, 16>::iterator EntIt =
561 std::lower_bound(Preds.begin(), Preds.end(),
562 SomePHI->getIncomingBlock(i));
563 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
564 "PHI node has entry for a block which is not a predecessor!");
570 // At this point, the blocks left in the preds list must have dummy
571 // entries inserted into every PHI nodes for the block. Update all the phi
572 // nodes in this block that we are inserting (there could be phis before
574 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
575 BasicBlock::iterator BBI = BB->begin();
576 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
577 SomePHI->getNumIncomingValues() == NumBadPreds) {
578 Value *UndefVal = UndefValue::get(SomePHI->getType());
579 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
580 SomePHI->addIncoming(UndefVal, Preds[pred]);
588 /// ComputeLiveInBlocks - Determine which blocks the value is live in. These
589 /// are blocks which lead to uses. Knowing this allows us to avoid inserting
590 /// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
592 void PromoteMem2Reg::
593 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
594 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
595 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
597 // To determine liveness, we must iterate through the predecessors of blocks
598 // where the def is live. Blocks are added to the worklist if we need to
599 // check their predecessors. Start with all the using blocks.
600 SmallVector<BasicBlock*, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
601 Info.UsingBlocks.end());
603 // If any of the using blocks is also a definition block, check to see if the
604 // definition occurs before or after the use. If it happens before the use,
605 // the value isn't really live-in.
606 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
607 BasicBlock *BB = LiveInBlockWorklist[i];
608 if (!DefBlocks.count(BB)) continue;
610 // Okay, this is a block that both uses and defines the value. If the first
611 // reference to the alloca is a def (store), then we know it isn't live-in.
612 for (BasicBlock::iterator I = BB->begin(); ; ++I) {
613 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
614 if (SI->getOperand(1) != AI) continue;
616 // We found a store to the alloca before a load. The alloca is not
617 // actually live-in here.
618 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
619 LiveInBlockWorklist.pop_back();
624 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
625 if (LI->getOperand(0) != AI) continue;
627 // Okay, we found a load before a store to the alloca. It is actually
628 // live into this block.
634 // Now that we have a set of blocks where the phi is live-in, recursively add
635 // their predecessors until we find the full region the value is live.
636 while (!LiveInBlockWorklist.empty()) {
637 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
639 // The block really is live in here, insert it into the set. If already in
640 // the set, then it has already been processed.
641 if (!LiveInBlocks.insert(BB))
644 // Since the value is live into BB, it is either defined in a predecessor or
645 // live into it to. Add the preds to the worklist unless they are a
647 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
650 // The value is not live into a predecessor if it defines the value.
651 if (DefBlocks.count(P))
654 // Otherwise it is, add to the worklist.
655 LiveInBlockWorklist.push_back(P);
660 /// DetermineInsertionPoint - At this point, we're committed to promoting the
661 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
662 /// which blocks need phi nodes and see if we can optimize out some work by
663 /// avoiding insertion of dead phi nodes.
664 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
667 // Unique the set of defining blocks for efficient lookup.
668 SmallPtrSet<BasicBlock*, 32> DefBlocks;
669 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
671 // Determine which blocks the value is live in. These are blocks which lead
673 SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
674 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
676 // Compute the locations where PhiNodes need to be inserted. Look at the
677 // dominance frontier of EACH basic-block we have a write in.
678 unsigned CurrentVersion = 0;
679 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
680 while (!Info.DefiningBlocks.empty()) {
681 BasicBlock *BB = Info.DefiningBlocks.back();
682 Info.DefiningBlocks.pop_back();
684 // Look up the DF for this write, add it to defining blocks.
685 DominanceFrontier::const_iterator it = DF.find(BB);
686 if (it == DF.end()) continue;
688 const DominanceFrontier::DomSetType &S = it->second;
690 // In theory we don't need the indirection through the DFBlocks vector.
691 // In practice, the order of calling QueuePhiNode would depend on the
692 // (unspecified) ordering of basic blocks in the dominance frontier,
693 // which would give PHI nodes non-determinstic subscripts. Fix this by
694 // processing blocks in order of the occurance in the function.
695 for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
696 PE = S.end(); P != PE; ++P) {
697 // If the frontier block is not in the live-in set for the alloca, don't
698 // bother processing it.
699 if (!LiveInBlocks.count(*P))
702 DFBlocks.push_back(std::make_pair(BBNumbers[*P], *P));
705 // Sort by which the block ordering in the function.
706 if (DFBlocks.size() > 1)
707 std::sort(DFBlocks.begin(), DFBlocks.end());
709 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
710 BasicBlock *BB = DFBlocks[i].second;
711 if (QueuePhiNode(BB, AllocaNum, CurrentVersion))
712 Info.DefiningBlocks.push_back(BB);
718 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
719 /// replace any loads of it that are directly dominated by the definition with
720 /// the value stored.
721 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
723 LargeBlockInfo &LBI) {
724 StoreInst *OnlyStore = Info.OnlyStore;
725 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
726 BasicBlock *StoreBB = OnlyStore->getParent();
729 // Clear out UsingBlocks. We will reconstruct it here if needed.
730 Info.UsingBlocks.clear();
732 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
733 Instruction *UserInst = cast<Instruction>(*UI++);
734 if (!isa<LoadInst>(UserInst)) {
735 assert(UserInst == OnlyStore && "Should only have load/stores");
738 LoadInst *LI = cast<LoadInst>(UserInst);
740 // Okay, if we have a load from the alloca, we want to replace it with the
741 // only value stored to the alloca. We can do this if the value is
742 // dominated by the store. If not, we use the rest of the mem2reg machinery
743 // to insert the phi nodes as needed.
744 if (!StoringGlobalVal) { // Non-instructions are always dominated.
745 if (LI->getParent() == StoreBB) {
746 // If we have a use that is in the same block as the store, compare the
747 // indices of the two instructions to see which one came first. If the
748 // load came before the store, we can't handle it.
749 if (StoreIndex == -1)
750 StoreIndex = LBI.getInstructionIndex(OnlyStore);
752 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
753 // Can't handle this load, bail out.
754 Info.UsingBlocks.push_back(StoreBB);
758 } else if (LI->getParent() != StoreBB &&
759 !dominates(StoreBB, LI->getParent())) {
760 // If the load and store are in different blocks, use BB dominance to
761 // check their relationships. If the store doesn't dom the use, bail
763 Info.UsingBlocks.push_back(LI->getParent());
768 // Otherwise, we *can* safely rewrite this load.
769 Value *ReplVal = OnlyStore->getOperand(0);
770 // If the replacement value is the load, this must occur in unreachable
773 ReplVal = UndefValue::get(LI->getType());
774 LI->replaceAllUsesWith(ReplVal);
775 if (AST && LI->getType()->isPointerTy())
776 AST->deleteValue(LI);
777 LI->eraseFromParent();
784 /// StoreIndexSearchPredicate - This is a helper predicate used to search by the
785 /// first element of a pair.
786 struct StoreIndexSearchPredicate {
787 bool operator()(const std::pair<unsigned, StoreInst*> &LHS,
788 const std::pair<unsigned, StoreInst*> &RHS) {
789 return LHS.first < RHS.first;
795 /// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
796 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
797 /// potentially useless PHI nodes by just performing a single linear pass over
798 /// the basic block using the Alloca.
800 /// If we cannot promote this alloca (because it is read before it is written),
801 /// return true. This is necessary in cases where, due to control flow, the
802 /// alloca is potentially undefined on some control flow paths. e.g. code like
803 /// this is potentially correct:
805 /// for (...) { if (c) { A = undef; undef = B; } }
807 /// ... so long as A is not used before undef is set.
809 void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
810 LargeBlockInfo &LBI) {
811 // The trickiest case to handle is when we have large blocks. Because of this,
812 // this code is optimized assuming that large blocks happen. This does not
813 // significantly pessimize the small block case. This uses LargeBlockInfo to
814 // make it efficient to get the index of various operations in the block.
816 // Clear out UsingBlocks. We will reconstruct it here if needed.
817 Info.UsingBlocks.clear();
819 // Walk the use-def list of the alloca, getting the locations of all stores.
820 typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
821 StoresByIndexTy StoresByIndex;
823 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
825 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
826 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
828 // If there are no stores to the alloca, just replace any loads with undef.
829 if (StoresByIndex.empty()) {
830 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
831 if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
832 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
833 if (AST && LI->getType()->isPointerTy())
834 AST->deleteValue(LI);
836 LI->eraseFromParent();
841 // Sort the stores by their index, making it efficient to do a lookup with a
843 std::sort(StoresByIndex.begin(), StoresByIndex.end());
845 // Walk all of the loads from this alloca, replacing them with the nearest
846 // store above them, if any.
847 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
848 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
851 unsigned LoadIdx = LBI.getInstructionIndex(LI);
853 // Find the nearest store that has a lower than this load.
854 StoresByIndexTy::iterator I =
855 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
856 std::pair<unsigned, StoreInst*>(LoadIdx, static_cast<StoreInst*>(0)),
857 StoreIndexSearchPredicate());
859 // If there is no store before this load, then we can't promote this load.
860 if (I == StoresByIndex.begin()) {
861 // Can't handle this load, bail out.
862 Info.UsingBlocks.push_back(LI->getParent());
866 // Otherwise, there was a store before this load, the load takes its value.
868 LI->replaceAllUsesWith(I->second->getOperand(0));
869 if (AST && LI->getType()->isPointerTy())
870 AST->deleteValue(LI);
871 LI->eraseFromParent();
876 // Inserts a llvm.dbg.value instrinsic before the stores to an alloca'd value
877 // that has an associated llvm.dbg.decl intrinsic.
878 void PromoteMem2Reg::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
880 DIVariable DIVar(DDI->getVariable());
885 DIF = new DIFactory(*SI->getParent()->getParent()->getParent());
886 Instruction *DbgVal = DIF->InsertDbgValueIntrinsic(SI->getOperand(0), 0,
889 // Propagate any debug metadata from the store onto the dbg.value.
890 DebugLoc SIDL = SI->getDebugLoc();
891 if (!SIDL.isUnknown())
892 DbgVal->setDebugLoc(SIDL);
893 // Otherwise propagate debug metadata from dbg.declare.
895 DbgVal->setDebugLoc(DDI->getDebugLoc());
898 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
899 // Alloca returns true if there wasn't already a phi-node for that variable
901 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
903 // Look up the basic-block in question.
904 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
906 // If the BB already has a phi node added for the i'th alloca then we're done!
907 if (PN) return false;
909 // Create a PhiNode using the dereferenced type... and add the phi-node to the
911 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(),
912 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
915 PhiToAllocaMap[PN] = AllocaNo;
916 PN->reserveOperandSpace(getNumPreds(BB));
918 if (AST && PN->getType()->isPointerTy())
919 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
924 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
925 // stores to the allocas which we are promoting. IncomingVals indicates what
926 // value each Alloca contains on exit from the predecessor block Pred.
928 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
929 RenamePassData::ValVector &IncomingVals,
930 std::vector<RenamePassData> &Worklist) {
932 // If we are inserting any phi nodes into this BB, they will already be in the
934 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
935 // If we have PHI nodes to update, compute the number of edges from Pred to
937 if (PhiToAllocaMap.count(APN)) {
938 // We want to be able to distinguish between PHI nodes being inserted by
939 // this invocation of mem2reg from those phi nodes that already existed in
940 // the IR before mem2reg was run. We determine that APN is being inserted
941 // because it is missing incoming edges. All other PHI nodes being
942 // inserted by this pass of mem2reg will have the same number of incoming
943 // operands so far. Remember this count.
944 unsigned NewPHINumOperands = APN->getNumOperands();
946 unsigned NumEdges = 0;
947 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
950 assert(NumEdges && "Must be at least one edge from Pred to BB!");
952 // Add entries for all the phis.
953 BasicBlock::iterator PNI = BB->begin();
955 unsigned AllocaNo = PhiToAllocaMap[APN];
957 // Add N incoming values to the PHI node.
958 for (unsigned i = 0; i != NumEdges; ++i)
959 APN->addIncoming(IncomingVals[AllocaNo], Pred);
961 // The currently active variable for this block is now the PHI.
962 IncomingVals[AllocaNo] = APN;
964 // Get the next phi node.
966 APN = dyn_cast<PHINode>(PNI);
969 // Verify that it is missing entries. If not, it is not being inserted
970 // by this mem2reg invocation so we want to ignore it.
971 } while (APN->getNumOperands() == NewPHINumOperands);
975 // Don't revisit blocks.
976 if (!Visited.insert(BB)) return;
978 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
979 Instruction *I = II++; // get the instruction, increment iterator
981 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
982 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
985 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
986 if (AI == AllocaLookup.end()) continue;
988 Value *V = IncomingVals[AI->second];
990 // Anything using the load now uses the current value.
991 LI->replaceAllUsesWith(V);
992 if (AST && LI->getType()->isPointerTy())
993 AST->deleteValue(LI);
994 BB->getInstList().erase(LI);
995 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
996 // Delete this instruction and mark the name as the current holder of the
998 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
1001 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
1002 if (ai == AllocaLookup.end())
1005 // what value were we writing?
1006 IncomingVals[ai->second] = SI->getOperand(0);
1007 // Record debuginfo for the store before removing it.
1008 if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second])
1009 ConvertDebugDeclareToDebugValue(DDI, SI);
1010 BB->getInstList().erase(SI);
1014 // 'Recurse' to our successors.
1015 succ_iterator I = succ_begin(BB), E = succ_end(BB);
1018 // Keep track of the successors so we don't visit the same successor twice
1019 SmallPtrSet<BasicBlock*, 8> VisitedSuccs;
1021 // Handle the first successor without using the worklist.
1022 VisitedSuccs.insert(*I);
1028 if (VisitedSuccs.insert(*I))
1029 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
1034 /// PromoteMemToReg - Promote the specified list of alloca instructions into
1035 /// scalar registers, inserting PHI nodes as appropriate. This function makes
1036 /// use of DominanceFrontier information. This function does not modify the CFG
1037 /// of the function at all. All allocas must be from the same function.
1039 /// If AST is specified, the specified tracker is updated to reflect changes
1042 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
1043 DominatorTree &DT, DominanceFrontier &DF,
1044 AliasSetTracker *AST) {
1045 // If there is nothing to do, bail out...
1046 if (Allocas.empty()) return;
1048 PromoteMem2Reg(Allocas, DT, DF, AST).run();