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 iterated dominator frontiers to place PHI nodes, then
13 // traversing the function in depth-first order to rewrite loads and stores as
16 // The algorithm used here is based on:
18 // Sreedhar and Gao. A linear time algorithm for placing phi-nodes.
19 // In Proceedings of the 22nd ACM SIGPLAN-SIGACT Symposium on Principles of
20 // Programming Languages
21 // POPL '95. ACM, New York, NY, 62-73.
23 // It has been modified to not explicitly use the DJ graph data structure and to
24 // directly compute pruned SSA using per-variable liveness information.
26 //===----------------------------------------------------------------------===//
28 #define DEBUG_TYPE "mem2reg"
29 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
30 #include "llvm/Constants.h"
31 #include "llvm/DerivedTypes.h"
32 #include "llvm/Function.h"
33 #include "llvm/Instructions.h"
34 #include "llvm/IntrinsicInst.h"
35 #include "llvm/Metadata.h"
36 #include "llvm/Analysis/AliasSetTracker.h"
37 #include "llvm/Analysis/DebugInfo.h"
38 #include "llvm/Analysis/DIBuilder.h"
39 #include "llvm/Analysis/Dominators.h"
40 #include "llvm/Analysis/InstructionSimplify.h"
41 #include "llvm/Analysis/ValueTracking.h"
42 #include "llvm/Transforms/Utils/Local.h"
43 #include "llvm/ADT/DenseMap.h"
44 #include "llvm/ADT/SmallPtrSet.h"
45 #include "llvm/ADT/SmallVector.h"
46 #include "llvm/ADT/Statistic.h"
47 #include "llvm/ADT/STLExtras.h"
48 #include "llvm/Support/CFG.h"
53 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
54 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
55 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
56 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
60 struct DenseMapInfo<std::pair<BasicBlock*, unsigned> > {
61 typedef std::pair<BasicBlock*, unsigned> EltTy;
62 static inline EltTy getEmptyKey() {
63 return EltTy(reinterpret_cast<BasicBlock*>(-1), ~0U);
65 static inline EltTy getTombstoneKey() {
66 return EltTy(reinterpret_cast<BasicBlock*>(-2), 0U);
68 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
69 return DenseMapInfo<void*>::getHashValue(Val.first) + Val.second*2;
71 static bool isEqual(const EltTy &LHS, const EltTy &RHS) {
77 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
78 /// This is true if there are only loads and stores to the alloca.
80 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
81 // FIXME: If the memory unit is of pointer or integer type, we can permit
82 // assignments to subsections of the memory unit.
84 // Only allow direct and non-volatile loads and stores...
85 for (Value::const_use_iterator UI = AI->use_begin(), UE = AI->use_end();
86 UI != UE; ++UI) { // Loop over all of the uses of the alloca
88 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
89 // Note that atomic loads can be transformed; atomic semantics do
90 // not have any meaning for a local alloca.
93 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
94 if (SI->getOperand(0) == AI)
95 return false; // Don't allow a store OF the AI, only INTO the AI.
96 // Note that atomic stores can be transformed; atomic semantics do
97 // not have any meaning for a local alloca.
100 } else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
101 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
102 II->getIntrinsicID() != Intrinsic::lifetime_end)
104 } else if (const BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
105 if (BCI->getType() != Type::getInt8PtrTy(U->getContext()))
107 if (!onlyUsedByLifetimeMarkers(BCI))
109 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
110 if (GEPI->getType() != Type::getInt8PtrTy(U->getContext()))
112 if (!GEPI->hasAllZeroIndices())
114 if (!onlyUsedByLifetimeMarkers(GEPI))
127 // Data package used by RenamePass()
128 class RenamePassData {
130 typedef std::vector<Value *> ValVector;
132 RenamePassData() : BB(NULL), Pred(NULL), Values() {}
133 RenamePassData(BasicBlock *B, BasicBlock *P,
134 const ValVector &V) : BB(B), Pred(P), Values(V) {}
139 void swap(RenamePassData &RHS) {
140 std::swap(BB, RHS.BB);
141 std::swap(Pred, RHS.Pred);
142 Values.swap(RHS.Values);
146 /// LargeBlockInfo - This assigns and keeps a per-bb relative ordering of
147 /// load/store instructions in the block that directly load or store an alloca.
149 /// This functionality is important because it avoids scanning large basic
150 /// blocks multiple times when promoting many allocas in the same block.
151 class LargeBlockInfo {
152 /// InstNumbers - For each instruction that we track, keep the index of the
153 /// instruction. The index starts out as the number of the instruction from
154 /// the start of the block.
155 DenseMap<const Instruction *, unsigned> InstNumbers;
158 /// isInterestingInstruction - This code only looks at accesses to allocas.
159 static bool isInterestingInstruction(const Instruction *I) {
160 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
161 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
164 /// getInstructionIndex - Get or calculate the index of the specified
166 unsigned getInstructionIndex(const Instruction *I) {
167 assert(isInterestingInstruction(I) &&
168 "Not a load/store to/from an alloca?");
170 // If we already have this instruction number, return it.
171 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
172 if (It != InstNumbers.end()) return It->second;
174 // Scan the whole block to get the instruction. This accumulates
175 // information for every interesting instruction in the block, in order to
176 // avoid gratuitus rescans.
177 const BasicBlock *BB = I->getParent();
179 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end();
181 if (isInterestingInstruction(BBI))
182 InstNumbers[BBI] = InstNo++;
183 It = InstNumbers.find(I);
185 assert(It != InstNumbers.end() && "Didn't insert instruction?");
189 void deleteValue(const Instruction *I) {
190 InstNumbers.erase(I);
198 struct PromoteMem2Reg {
199 /// Allocas - The alloca instructions being promoted.
201 std::vector<AllocaInst*> Allocas;
205 /// AST - An AliasSetTracker object to update. If null, don't update it.
207 AliasSetTracker *AST;
209 /// AllocaLookup - Reverse mapping of Allocas.
211 DenseMap<AllocaInst*, unsigned> AllocaLookup;
213 /// NewPhiNodes - The PhiNodes we're adding.
215 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
217 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
218 /// it corresponds to.
219 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
221 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
222 /// each alloca that is of pointer type, we keep track of what to copyValue
223 /// to the inserted PHI nodes here.
225 std::vector<Value*> PointerAllocaValues;
227 /// AllocaDbgDeclares - For each alloca, we keep track of the dbg.declare
228 /// intrinsic that describes it, if any, so that we can convert it to a
229 /// dbg.value intrinsic if the alloca gets promoted.
230 SmallVector<DbgDeclareInst*, 8> AllocaDbgDeclares;
232 /// Visited - The set of basic blocks the renamer has already visited.
234 SmallPtrSet<BasicBlock*, 16> Visited;
236 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
237 /// non-determinstic behavior.
238 DenseMap<BasicBlock*, unsigned> BBNumbers;
240 /// DomLevels - Maps DomTreeNodes to their level in the dominator tree.
241 DenseMap<DomTreeNode*, unsigned> DomLevels;
243 /// BBNumPreds - Lazily compute the number of predecessors a block has.
244 DenseMap<const BasicBlock*, unsigned> BBNumPreds;
246 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
247 AliasSetTracker *ast)
248 : Allocas(A), DT(dt), DIB(0), AST(ast) {}
255 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
257 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
258 return DT.dominates(BB1, BB2);
262 void RemoveFromAllocasList(unsigned &AllocaIdx) {
263 Allocas[AllocaIdx] = Allocas.back();
268 unsigned getNumPreds(const BasicBlock *BB) {
269 unsigned &NP = BBNumPreds[BB];
271 NP = std::distance(pred_begin(BB), pred_end(BB))+1;
275 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
277 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
278 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
279 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks);
281 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
282 LargeBlockInfo &LBI);
283 void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
284 LargeBlockInfo &LBI);
286 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
287 RenamePassData::ValVector &IncVals,
288 std::vector<RenamePassData> &Worklist);
289 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version);
293 SmallVector<BasicBlock*, 32> DefiningBlocks;
294 SmallVector<BasicBlock*, 32> UsingBlocks;
296 StoreInst *OnlyStore;
297 BasicBlock *OnlyBlock;
298 bool OnlyUsedInOneBlock;
300 Value *AllocaPointerVal;
301 DbgDeclareInst *DbgDeclare;
304 DefiningBlocks.clear();
308 OnlyUsedInOneBlock = true;
309 AllocaPointerVal = 0;
313 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
315 void AnalyzeAlloca(AllocaInst *AI) {
318 // As we scan the uses of the alloca instruction, keep track of stores,
319 // and decide whether all of the loads and stores to the alloca are within
320 // the same basic block.
321 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
323 Instruction *User = cast<Instruction>(*UI++);
325 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
326 // Remember the basic blocks which define new values for the alloca
327 DefiningBlocks.push_back(SI->getParent());
328 AllocaPointerVal = SI->getOperand(0);
331 LoadInst *LI = cast<LoadInst>(User);
332 // Otherwise it must be a load instruction, keep track of variable
334 UsingBlocks.push_back(LI->getParent());
335 AllocaPointerVal = LI;
338 if (OnlyUsedInOneBlock) {
340 OnlyBlock = User->getParent();
341 else if (OnlyBlock != User->getParent())
342 OnlyUsedInOneBlock = false;
346 DbgDeclare = FindAllocaDbgDeclare(AI);
350 typedef std::pair<DomTreeNode*, unsigned> DomTreeNodePair;
352 struct DomTreeNodeCompare {
353 bool operator()(const DomTreeNodePair &LHS, const DomTreeNodePair &RHS) {
354 return LHS.second < RHS.second;
357 } // end of anonymous namespace
359 static void removeLifetimeIntrinsicUsers(AllocaInst *AI) {
360 // Knowing that this alloca is promotable, we know that it's safe to kill all
361 // instructions except for load and store.
363 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
365 Instruction *I = cast<Instruction>(*UI);
367 if (isa<LoadInst>(I) || isa<StoreInst>(I))
370 if (!I->getType()->isVoidTy()) {
371 // The only users of this bitcast/GEP instruction are lifetime intrinsics.
372 // Follow the use/def chain to erase them now instead of leaving it for
373 // dead code elimination later.
374 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
376 Instruction *Inst = cast<Instruction>(*UI);
378 Inst->eraseFromParent();
381 I->eraseFromParent();
385 void PromoteMem2Reg::run() {
386 Function &F = *DT.getRoot()->getParent();
388 if (AST) PointerAllocaValues.resize(Allocas.size());
389 AllocaDbgDeclares.resize(Allocas.size());
394 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
395 AllocaInst *AI = Allocas[AllocaNum];
397 assert(isAllocaPromotable(AI) &&
398 "Cannot promote non-promotable alloca!");
399 assert(AI->getParent()->getParent() == &F &&
400 "All allocas should be in the same function, which is same as DF!");
402 removeLifetimeIntrinsicUsers(AI);
404 if (AI->use_empty()) {
405 // If there are no uses of the alloca, just delete it now.
406 if (AST) AST->deleteValue(AI);
407 AI->eraseFromParent();
409 // Remove the alloca from the Allocas list, since it has been processed
410 RemoveFromAllocasList(AllocaNum);
415 // Calculate the set of read and write-locations for each alloca. This is
416 // analogous to finding the 'uses' and 'definitions' of each variable.
417 Info.AnalyzeAlloca(AI);
419 // If there is only a single store to this value, replace any loads of
420 // it that are directly dominated by the definition with the value stored.
421 if (Info.DefiningBlocks.size() == 1) {
422 RewriteSingleStoreAlloca(AI, Info, LBI);
424 // Finally, after the scan, check to see if the store is all that is left.
425 if (Info.UsingBlocks.empty()) {
426 // Record debuginfo for the store and remove the declaration's debuginfo.
427 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
429 DIB = new DIBuilder(*DDI->getParent()->getParent()->getParent());
430 ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore, *DIB);
431 DDI->eraseFromParent();
433 // Remove the (now dead) store and alloca.
434 Info.OnlyStore->eraseFromParent();
435 LBI.deleteValue(Info.OnlyStore);
437 if (AST) AST->deleteValue(AI);
438 AI->eraseFromParent();
441 // The alloca has been processed, move on.
442 RemoveFromAllocasList(AllocaNum);
449 // If the alloca is only read and written in one basic block, just perform a
450 // linear sweep over the block to eliminate it.
451 if (Info.OnlyUsedInOneBlock) {
452 PromoteSingleBlockAlloca(AI, Info, LBI);
454 // Finally, after the scan, check to see if the stores are all that is
456 if (Info.UsingBlocks.empty()) {
458 // Remove the (now dead) stores and alloca.
459 while (!AI->use_empty()) {
460 StoreInst *SI = cast<StoreInst>(AI->use_back());
461 // Record debuginfo for the store before removing it.
462 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
464 DIB = new DIBuilder(*SI->getParent()->getParent()->getParent());
465 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB);
467 SI->eraseFromParent();
471 if (AST) AST->deleteValue(AI);
472 AI->eraseFromParent();
475 // The alloca has been processed, move on.
476 RemoveFromAllocasList(AllocaNum);
478 // The alloca's debuginfo can be removed as well.
479 if (DbgDeclareInst *DDI = Info.DbgDeclare)
480 DDI->eraseFromParent();
487 // If we haven't computed dominator tree levels, do so now.
488 if (DomLevels.empty()) {
489 SmallVector<DomTreeNode*, 32> Worklist;
491 DomTreeNode *Root = DT.getRootNode();
493 Worklist.push_back(Root);
495 while (!Worklist.empty()) {
496 DomTreeNode *Node = Worklist.pop_back_val();
497 unsigned ChildLevel = DomLevels[Node] + 1;
498 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end();
500 DomLevels[*CI] = ChildLevel;
501 Worklist.push_back(*CI);
506 // If we haven't computed a numbering for the BB's in the function, do so
508 if (BBNumbers.empty()) {
510 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
514 // If we have an AST to keep updated, remember some pointer value that is
515 // stored into the alloca.
517 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
519 // Remember the dbg.declare intrinsic describing this alloca, if any.
520 if (Info.DbgDeclare) AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare;
522 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
523 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
525 // At this point, we're committed to promoting the alloca using IDF's, and
526 // the standard SSA construction algorithm. Determine which blocks need PHI
527 // nodes and see if we can optimize out some work by avoiding insertion of
529 DetermineInsertionPoint(AI, AllocaNum, Info);
533 return; // All of the allocas must have been trivial!
538 // Set the incoming values for the basic block to be null values for all of
539 // the alloca's. We do this in case there is a load of a value that has not
540 // been stored yet. In this case, it will get this null value.
542 RenamePassData::ValVector Values(Allocas.size());
543 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
544 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
546 // Walks all basic blocks in the function performing the SSA rename algorithm
547 // and inserting the phi nodes we marked as necessary
549 std::vector<RenamePassData> RenamePassWorkList;
550 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
553 RPD.swap(RenamePassWorkList.back());
554 RenamePassWorkList.pop_back();
555 // RenamePass may add new worklist entries.
556 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
557 } while (!RenamePassWorkList.empty());
559 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
562 // Remove the allocas themselves from the function.
563 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
564 Instruction *A = Allocas[i];
566 // If there are any uses of the alloca instructions left, they must be in
567 // unreachable basic blocks that were not processed by walking the dominator
568 // tree. Just delete the users now.
570 A->replaceAllUsesWith(UndefValue::get(A->getType()));
571 if (AST) AST->deleteValue(A);
572 A->eraseFromParent();
575 // Remove alloca's dbg.declare instrinsics from the function.
576 for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i)
577 if (DbgDeclareInst *DDI = AllocaDbgDeclares[i])
578 DDI->eraseFromParent();
580 // Loop over all of the PHI nodes and see if there are any that we can get
581 // rid of because they merge all of the same incoming values. This can
582 // happen due to undef values coming into the PHI nodes. This process is
583 // iterative, because eliminating one PHI node can cause others to be removed.
584 bool EliminatedAPHI = true;
585 while (EliminatedAPHI) {
586 EliminatedAPHI = false;
588 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
589 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
590 PHINode *PN = I->second;
592 // If this PHI node merges one value and/or undefs, get the value.
593 if (Value *V = SimplifyInstruction(PN, 0, 0, &DT)) {
594 if (AST && PN->getType()->isPointerTy())
595 AST->deleteValue(PN);
596 PN->replaceAllUsesWith(V);
597 PN->eraseFromParent();
598 NewPhiNodes.erase(I++);
599 EliminatedAPHI = true;
606 // At this point, the renamer has added entries to PHI nodes for all reachable
607 // code. Unfortunately, there may be unreachable blocks which the renamer
608 // hasn't traversed. If this is the case, the PHI nodes may not
609 // have incoming values for all predecessors. Loop over all PHI nodes we have
610 // created, inserting undef values if they are missing any incoming values.
612 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
613 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
614 // We want to do this once per basic block. As such, only process a block
615 // when we find the PHI that is the first entry in the block.
616 PHINode *SomePHI = I->second;
617 BasicBlock *BB = SomePHI->getParent();
618 if (&BB->front() != SomePHI)
621 // Only do work here if there the PHI nodes are missing incoming values. We
622 // know that all PHI nodes that were inserted in a block will have the same
623 // number of incoming values, so we can just check any of them.
624 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
627 // Get the preds for BB.
628 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
630 // Ok, now we know that all of the PHI nodes are missing entries for some
631 // basic blocks. Start by sorting the incoming predecessors for efficient
633 std::sort(Preds.begin(), Preds.end());
635 // Now we loop through all BB's which have entries in SomePHI and remove
636 // them from the Preds list.
637 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
638 // Do a log(n) search of the Preds list for the entry we want.
639 SmallVector<BasicBlock*, 16>::iterator EntIt =
640 std::lower_bound(Preds.begin(), Preds.end(),
641 SomePHI->getIncomingBlock(i));
642 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
643 "PHI node has entry for a block which is not a predecessor!");
649 // At this point, the blocks left in the preds list must have dummy
650 // entries inserted into every PHI nodes for the block. Update all the phi
651 // nodes in this block that we are inserting (there could be phis before
653 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
654 BasicBlock::iterator BBI = BB->begin();
655 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
656 SomePHI->getNumIncomingValues() == NumBadPreds) {
657 Value *UndefVal = UndefValue::get(SomePHI->getType());
658 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
659 SomePHI->addIncoming(UndefVal, Preds[pred]);
667 /// ComputeLiveInBlocks - Determine which blocks the value is live in. These
668 /// are blocks which lead to uses. Knowing this allows us to avoid inserting
669 /// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
671 void PromoteMem2Reg::
672 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
673 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
674 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
676 // To determine liveness, we must iterate through the predecessors of blocks
677 // where the def is live. Blocks are added to the worklist if we need to
678 // check their predecessors. Start with all the using blocks.
679 SmallVector<BasicBlock*, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
680 Info.UsingBlocks.end());
682 // If any of the using blocks is also a definition block, check to see if the
683 // definition occurs before or after the use. If it happens before the use,
684 // the value isn't really live-in.
685 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
686 BasicBlock *BB = LiveInBlockWorklist[i];
687 if (!DefBlocks.count(BB)) continue;
689 // Okay, this is a block that both uses and defines the value. If the first
690 // reference to the alloca is a def (store), then we know it isn't live-in.
691 for (BasicBlock::iterator I = BB->begin(); ; ++I) {
692 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
693 if (SI->getOperand(1) != AI) continue;
695 // We found a store to the alloca before a load. The alloca is not
696 // actually live-in here.
697 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
698 LiveInBlockWorklist.pop_back();
703 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
704 if (LI->getOperand(0) != AI) continue;
706 // Okay, we found a load before a store to the alloca. It is actually
707 // live into this block.
713 // Now that we have a set of blocks where the phi is live-in, recursively add
714 // their predecessors until we find the full region the value is live.
715 while (!LiveInBlockWorklist.empty()) {
716 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
718 // The block really is live in here, insert it into the set. If already in
719 // the set, then it has already been processed.
720 if (!LiveInBlocks.insert(BB))
723 // Since the value is live into BB, it is either defined in a predecessor or
724 // live into it to. Add the preds to the worklist unless they are a
726 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
729 // The value is not live into a predecessor if it defines the value.
730 if (DefBlocks.count(P))
733 // Otherwise it is, add to the worklist.
734 LiveInBlockWorklist.push_back(P);
739 /// DetermineInsertionPoint - At this point, we're committed to promoting the
740 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
741 /// which blocks need phi nodes and see if we can optimize out some work by
742 /// avoiding insertion of dead phi nodes.
743 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
745 // Unique the set of defining blocks for efficient lookup.
746 SmallPtrSet<BasicBlock*, 32> DefBlocks;
747 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
749 // Determine which blocks the value is live in. These are blocks which lead
751 SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
752 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
754 // Use a priority queue keyed on dominator tree level so that inserted nodes
755 // are handled from the bottom of the dominator tree upwards.
756 typedef std::priority_queue<DomTreeNodePair, SmallVector<DomTreeNodePair, 32>,
757 DomTreeNodeCompare> IDFPriorityQueue;
760 for (SmallPtrSet<BasicBlock*, 32>::const_iterator I = DefBlocks.begin(),
761 E = DefBlocks.end(); I != E; ++I) {
762 if (DomTreeNode *Node = DT.getNode(*I))
763 PQ.push(std::make_pair(Node, DomLevels[Node]));
766 SmallVector<std::pair<unsigned, BasicBlock*>, 32> DFBlocks;
767 SmallPtrSet<DomTreeNode*, 32> Visited;
768 SmallVector<DomTreeNode*, 32> Worklist;
769 while (!PQ.empty()) {
770 DomTreeNodePair RootPair = PQ.top();
772 DomTreeNode *Root = RootPair.first;
773 unsigned RootLevel = RootPair.second;
775 // Walk all dominator tree children of Root, inspecting their CFG edges with
776 // targets elsewhere on the dominator tree. Only targets whose level is at
777 // most Root's level are added to the iterated dominance frontier of the
781 Worklist.push_back(Root);
783 while (!Worklist.empty()) {
784 DomTreeNode *Node = Worklist.pop_back_val();
785 BasicBlock *BB = Node->getBlock();
787 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE;
789 DomTreeNode *SuccNode = DT.getNode(*SI);
791 // Quickly skip all CFG edges that are also dominator tree edges instead
792 // of catching them below.
793 if (SuccNode->getIDom() == Node)
796 unsigned SuccLevel = DomLevels[SuccNode];
797 if (SuccLevel > RootLevel)
800 if (!Visited.insert(SuccNode))
803 BasicBlock *SuccBB = SuccNode->getBlock();
804 if (!LiveInBlocks.count(SuccBB))
807 DFBlocks.push_back(std::make_pair(BBNumbers[SuccBB], SuccBB));
808 if (!DefBlocks.count(SuccBB))
809 PQ.push(std::make_pair(SuccNode, SuccLevel));
812 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end(); CI != CE;
814 if (!Visited.count(*CI))
815 Worklist.push_back(*CI);
820 if (DFBlocks.size() > 1)
821 std::sort(DFBlocks.begin(), DFBlocks.end());
823 unsigned CurrentVersion = 0;
824 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i)
825 QueuePhiNode(DFBlocks[i].second, AllocaNum, CurrentVersion);
828 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
829 /// replace any loads of it that are directly dominated by the definition with
830 /// the value stored.
831 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
833 LargeBlockInfo &LBI) {
834 StoreInst *OnlyStore = Info.OnlyStore;
835 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
836 BasicBlock *StoreBB = OnlyStore->getParent();
839 // Clear out UsingBlocks. We will reconstruct it here if needed.
840 Info.UsingBlocks.clear();
842 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
843 Instruction *UserInst = cast<Instruction>(*UI++);
844 if (!isa<LoadInst>(UserInst)) {
845 assert(UserInst == OnlyStore && "Should only have load/stores");
848 LoadInst *LI = cast<LoadInst>(UserInst);
850 // Okay, if we have a load from the alloca, we want to replace it with the
851 // only value stored to the alloca. We can do this if the value is
852 // dominated by the store. If not, we use the rest of the mem2reg machinery
853 // to insert the phi nodes as needed.
854 if (!StoringGlobalVal) { // Non-instructions are always dominated.
855 if (LI->getParent() == StoreBB) {
856 // If we have a use that is in the same block as the store, compare the
857 // indices of the two instructions to see which one came first. If the
858 // load came before the store, we can't handle it.
859 if (StoreIndex == -1)
860 StoreIndex = LBI.getInstructionIndex(OnlyStore);
862 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
863 // Can't handle this load, bail out.
864 Info.UsingBlocks.push_back(StoreBB);
868 } else if (LI->getParent() != StoreBB &&
869 !dominates(StoreBB, LI->getParent())) {
870 // If the load and store are in different blocks, use BB dominance to
871 // check their relationships. If the store doesn't dom the use, bail
873 Info.UsingBlocks.push_back(LI->getParent());
878 // Otherwise, we *can* safely rewrite this load.
879 Value *ReplVal = OnlyStore->getOperand(0);
880 // If the replacement value is the load, this must occur in unreachable
883 ReplVal = UndefValue::get(LI->getType());
884 LI->replaceAllUsesWith(ReplVal);
885 if (AST && LI->getType()->isPointerTy())
886 AST->deleteValue(LI);
887 LI->eraseFromParent();
894 /// StoreIndexSearchPredicate - This is a helper predicate used to search by the
895 /// first element of a pair.
896 struct StoreIndexSearchPredicate {
897 bool operator()(const std::pair<unsigned, StoreInst*> &LHS,
898 const std::pair<unsigned, StoreInst*> &RHS) {
899 return LHS.first < RHS.first;
905 /// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
906 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
907 /// potentially useless PHI nodes by just performing a single linear pass over
908 /// the basic block using the Alloca.
910 /// If we cannot promote this alloca (because it is read before it is written),
911 /// return true. This is necessary in cases where, due to control flow, the
912 /// alloca is potentially undefined on some control flow paths. e.g. code like
913 /// this is potentially correct:
915 /// for (...) { if (c) { A = undef; undef = B; } }
917 /// ... so long as A is not used before undef is set.
919 void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
920 LargeBlockInfo &LBI) {
921 // The trickiest case to handle is when we have large blocks. Because of this,
922 // this code is optimized assuming that large blocks happen. This does not
923 // significantly pessimize the small block case. This uses LargeBlockInfo to
924 // make it efficient to get the index of various operations in the block.
926 // Clear out UsingBlocks. We will reconstruct it here if needed.
927 Info.UsingBlocks.clear();
929 // Walk the use-def list of the alloca, getting the locations of all stores.
930 typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
931 StoresByIndexTy StoresByIndex;
933 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
935 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
936 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
938 // If there are no stores to the alloca, just replace any loads with undef.
939 if (StoresByIndex.empty()) {
940 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
941 if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
942 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
943 if (AST && LI->getType()->isPointerTy())
944 AST->deleteValue(LI);
946 LI->eraseFromParent();
951 // Sort the stores by their index, making it efficient to do a lookup with a
953 std::sort(StoresByIndex.begin(), StoresByIndex.end());
955 // Walk all of the loads from this alloca, replacing them with the nearest
956 // store above them, if any.
957 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
958 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
961 unsigned LoadIdx = LBI.getInstructionIndex(LI);
963 // Find the nearest store that has a lower than this load.
964 StoresByIndexTy::iterator I =
965 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
966 std::pair<unsigned, StoreInst*>(LoadIdx, static_cast<StoreInst*>(0)),
967 StoreIndexSearchPredicate());
969 // If there is no store before this load, then we can't promote this load.
970 if (I == StoresByIndex.begin()) {
971 // Can't handle this load, bail out.
972 Info.UsingBlocks.push_back(LI->getParent());
976 // Otherwise, there was a store before this load, the load takes its value.
978 LI->replaceAllUsesWith(I->second->getOperand(0));
979 if (AST && LI->getType()->isPointerTy())
980 AST->deleteValue(LI);
981 LI->eraseFromParent();
986 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
987 // Alloca returns true if there wasn't already a phi-node for that variable
989 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
991 // Look up the basic-block in question.
992 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
994 // If the BB already has a phi node added for the i'th alloca then we're done!
995 if (PN) return false;
997 // Create a PhiNode using the dereferenced type... and add the phi-node to the
999 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB),
1000 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
1003 PhiToAllocaMap[PN] = AllocaNo;
1005 if (AST && PN->getType()->isPointerTy())
1006 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
1011 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
1012 // stores to the allocas which we are promoting. IncomingVals indicates what
1013 // value each Alloca contains on exit from the predecessor block Pred.
1015 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
1016 RenamePassData::ValVector &IncomingVals,
1017 std::vector<RenamePassData> &Worklist) {
1019 // If we are inserting any phi nodes into this BB, they will already be in the
1021 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
1022 // If we have PHI nodes to update, compute the number of edges from Pred to
1024 if (PhiToAllocaMap.count(APN)) {
1025 // We want to be able to distinguish between PHI nodes being inserted by
1026 // this invocation of mem2reg from those phi nodes that already existed in
1027 // the IR before mem2reg was run. We determine that APN is being inserted
1028 // because it is missing incoming edges. All other PHI nodes being
1029 // inserted by this pass of mem2reg will have the same number of incoming
1030 // operands so far. Remember this count.
1031 unsigned NewPHINumOperands = APN->getNumOperands();
1033 unsigned NumEdges = 0;
1034 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
1037 assert(NumEdges && "Must be at least one edge from Pred to BB!");
1039 // Add entries for all the phis.
1040 BasicBlock::iterator PNI = BB->begin();
1042 unsigned AllocaNo = PhiToAllocaMap[APN];
1044 // Add N incoming values to the PHI node.
1045 for (unsigned i = 0; i != NumEdges; ++i)
1046 APN->addIncoming(IncomingVals[AllocaNo], Pred);
1048 // The currently active variable for this block is now the PHI.
1049 IncomingVals[AllocaNo] = APN;
1051 // Get the next phi node.
1053 APN = dyn_cast<PHINode>(PNI);
1054 if (APN == 0) break;
1056 // Verify that it is missing entries. If not, it is not being inserted
1057 // by this mem2reg invocation so we want to ignore it.
1058 } while (APN->getNumOperands() == NewPHINumOperands);
1062 // Don't revisit blocks.
1063 if (!Visited.insert(BB)) return;
1065 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
1066 Instruction *I = II++; // get the instruction, increment iterator
1068 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1069 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
1072 DenseMap<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
1073 if (AI == AllocaLookup.end()) continue;
1075 Value *V = IncomingVals[AI->second];
1077 // Anything using the load now uses the current value.
1078 LI->replaceAllUsesWith(V);
1079 if (AST && LI->getType()->isPointerTy())
1080 AST->deleteValue(LI);
1081 BB->getInstList().erase(LI);
1082 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1083 // Delete this instruction and mark the name as the current holder of the
1085 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
1086 if (!Dest) continue;
1088 DenseMap<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
1089 if (ai == AllocaLookup.end())
1092 // what value were we writing?
1093 IncomingVals[ai->second] = SI->getOperand(0);
1094 // Record debuginfo for the store before removing it.
1095 if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second]) {
1097 DIB = new DIBuilder(*SI->getParent()->getParent()->getParent());
1098 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB);
1100 BB->getInstList().erase(SI);
1104 // 'Recurse' to our successors.
1105 succ_iterator I = succ_begin(BB), E = succ_end(BB);
1108 // Keep track of the successors so we don't visit the same successor twice
1109 SmallPtrSet<BasicBlock*, 8> VisitedSuccs;
1111 // Handle the first successor without using the worklist.
1112 VisitedSuccs.insert(*I);
1118 if (VisitedSuccs.insert(*I))
1119 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
1124 /// PromoteMemToReg - Promote the specified list of alloca instructions into
1125 /// scalar registers, inserting PHI nodes as appropriate. This function does
1126 /// not modify the CFG of the function at all. All allocas must be from the
1129 /// If AST is specified, the specified tracker is updated to reflect changes
1132 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
1133 DominatorTree &DT, AliasSetTracker *AST) {
1134 // If there is nothing to do, bail out...
1135 if (Allocas.empty()) return;
1137 PromoteMem2Reg(Allocas, DT, AST).run();