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/Transforms/Utils/Local.h"
42 #include "llvm/ADT/DenseMap.h"
43 #include "llvm/ADT/SmallPtrSet.h"
44 #include "llvm/ADT/SmallVector.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/ADT/STLExtras.h"
47 #include "llvm/Support/CFG.h"
52 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
53 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
54 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
55 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
59 struct DenseMapInfo<std::pair<BasicBlock*, unsigned> > {
60 typedef std::pair<BasicBlock*, unsigned> EltTy;
61 static inline EltTy getEmptyKey() {
62 return EltTy(reinterpret_cast<BasicBlock*>(-1), ~0U);
64 static inline EltTy getTombstoneKey() {
65 return EltTy(reinterpret_cast<BasicBlock*>(-2), 0U);
67 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
68 return DenseMapInfo<void*>::getHashValue(Val.first) + Val.second*2;
70 static bool isEqual(const EltTy &LHS, const EltTy &RHS) {
76 /// onlyUsedByLifetimeMarkers - Return true if the only users of this pointer
77 /// are lifetime markers.
79 static bool onlyUsedByLifetimeMarkers(const Value *V) {
80 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end();
82 const IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI);
83 if (!II) return false;
85 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
86 II->getIntrinsicID() != Intrinsic::lifetime_end)
92 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
93 /// This is true if there are only loads and stores to the alloca.
95 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
96 // FIXME: If the memory unit is of pointer or integer type, we can permit
97 // assignments to subsections of the memory unit.
99 // Only allow direct and non-volatile loads and stores...
100 for (Value::const_use_iterator UI = AI->use_begin(), UE = AI->use_end();
101 UI != UE; ++UI) { // Loop over all of the uses of the alloca
103 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
104 if (LI->isVolatile())
106 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
107 if (SI->getOperand(0) == AI)
108 return false; // Don't allow a store OF the AI, only INTO the AI.
109 if (SI->isVolatile())
111 } else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
112 if (II->getIntrinsicID() != Intrinsic::lifetime_start &&
113 II->getIntrinsicID() != Intrinsic::lifetime_end)
115 } else if (const BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
116 if (BCI->getType() != Type::getInt8PtrTy(U->getContext()))
118 if (!onlyUsedByLifetimeMarkers(BCI))
120 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
121 if (GEPI->getType() != Type::getInt8PtrTy(U->getContext()))
123 if (!GEPI->hasAllZeroIndices())
125 if (!onlyUsedByLifetimeMarkers(GEPI))
138 // Data package used by RenamePass()
139 class RenamePassData {
141 typedef std::vector<Value *> ValVector;
143 RenamePassData() : BB(NULL), Pred(NULL), Values() {}
144 RenamePassData(BasicBlock *B, BasicBlock *P,
145 const ValVector &V) : BB(B), Pred(P), Values(V) {}
150 void swap(RenamePassData &RHS) {
151 std::swap(BB, RHS.BB);
152 std::swap(Pred, RHS.Pred);
153 Values.swap(RHS.Values);
157 /// LargeBlockInfo - This assigns and keeps a per-bb relative ordering of
158 /// load/store instructions in the block that directly load or store an alloca.
160 /// This functionality is important because it avoids scanning large basic
161 /// blocks multiple times when promoting many allocas in the same block.
162 class LargeBlockInfo {
163 /// InstNumbers - For each instruction that we track, keep the index of the
164 /// instruction. The index starts out as the number of the instruction from
165 /// the start of the block.
166 DenseMap<const Instruction *, unsigned> InstNumbers;
169 /// isInterestingInstruction - This code only looks at accesses to allocas.
170 static bool isInterestingInstruction(const Instruction *I) {
171 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
172 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
175 /// getInstructionIndex - Get or calculate the index of the specified
177 unsigned getInstructionIndex(const Instruction *I) {
178 assert(isInterestingInstruction(I) &&
179 "Not a load/store to/from an alloca?");
181 // If we already have this instruction number, return it.
182 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
183 if (It != InstNumbers.end()) return It->second;
185 // Scan the whole block to get the instruction. This accumulates
186 // information for every interesting instruction in the block, in order to
187 // avoid gratuitus rescans.
188 const BasicBlock *BB = I->getParent();
190 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end();
192 if (isInterestingInstruction(BBI))
193 InstNumbers[BBI] = InstNo++;
194 It = InstNumbers.find(I);
196 assert(It != InstNumbers.end() && "Didn't insert instruction?");
200 void deleteValue(const Instruction *I) {
201 InstNumbers.erase(I);
209 struct PromoteMem2Reg {
210 /// Allocas - The alloca instructions being promoted.
212 std::vector<AllocaInst*> Allocas;
216 /// AST - An AliasSetTracker object to update. If null, don't update it.
218 AliasSetTracker *AST;
220 /// AllocaLookup - Reverse mapping of Allocas.
222 DenseMap<AllocaInst*, unsigned> AllocaLookup;
224 /// NewPhiNodes - The PhiNodes we're adding.
226 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
228 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
229 /// it corresponds to.
230 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
232 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
233 /// each alloca that is of pointer type, we keep track of what to copyValue
234 /// to the inserted PHI nodes here.
236 std::vector<Value*> PointerAllocaValues;
238 /// AllocaDbgDeclares - For each alloca, we keep track of the dbg.declare
239 /// intrinsic that describes it, if any, so that we can convert it to a
240 /// dbg.value intrinsic if the alloca gets promoted.
241 SmallVector<DbgDeclareInst*, 8> AllocaDbgDeclares;
243 /// Visited - The set of basic blocks the renamer has already visited.
245 SmallPtrSet<BasicBlock*, 16> Visited;
247 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
248 /// non-determinstic behavior.
249 DenseMap<BasicBlock*, unsigned> BBNumbers;
251 /// DomLevels - Maps DomTreeNodes to their level in the dominator tree.
252 DenseMap<DomTreeNode*, unsigned> DomLevels;
254 /// BBNumPreds - Lazily compute the number of predecessors a block has.
255 DenseMap<const BasicBlock*, unsigned> BBNumPreds;
257 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
258 AliasSetTracker *ast)
259 : Allocas(A), DT(dt), DIB(0), AST(ast) {}
266 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
268 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
269 return DT.dominates(BB1, BB2);
273 void RemoveFromAllocasList(unsigned &AllocaIdx) {
274 Allocas[AllocaIdx] = Allocas.back();
279 unsigned getNumPreds(const BasicBlock *BB) {
280 unsigned &NP = BBNumPreds[BB];
282 NP = std::distance(pred_begin(BB), pred_end(BB))+1;
286 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
288 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
289 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
290 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks);
292 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
293 LargeBlockInfo &LBI);
294 void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
295 LargeBlockInfo &LBI);
297 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
298 RenamePassData::ValVector &IncVals,
299 std::vector<RenamePassData> &Worklist);
300 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version);
304 SmallVector<BasicBlock*, 32> DefiningBlocks;
305 SmallVector<BasicBlock*, 32> UsingBlocks;
307 StoreInst *OnlyStore;
308 BasicBlock *OnlyBlock;
309 bool OnlyUsedInOneBlock;
311 Value *AllocaPointerVal;
312 DbgDeclareInst *DbgDeclare;
315 DefiningBlocks.clear();
319 OnlyUsedInOneBlock = true;
320 AllocaPointerVal = 0;
324 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
326 void AnalyzeAlloca(AllocaInst *AI) {
329 // As we scan the uses of the alloca instruction, keep track of stores,
330 // and decide whether all of the loads and stores to the alloca are within
331 // the same basic block.
332 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
334 Instruction *User = cast<Instruction>(*UI++);
336 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
337 // Remember the basic blocks which define new values for the alloca
338 DefiningBlocks.push_back(SI->getParent());
339 AllocaPointerVal = SI->getOperand(0);
342 LoadInst *LI = cast<LoadInst>(User);
343 // Otherwise it must be a load instruction, keep track of variable
345 UsingBlocks.push_back(LI->getParent());
346 AllocaPointerVal = LI;
349 if (OnlyUsedInOneBlock) {
351 OnlyBlock = User->getParent();
352 else if (OnlyBlock != User->getParent())
353 OnlyUsedInOneBlock = false;
357 DbgDeclare = FindAllocaDbgDeclare(AI);
361 typedef std::pair<DomTreeNode*, unsigned> DomTreeNodePair;
363 struct DomTreeNodeCompare {
364 bool operator()(const DomTreeNodePair &LHS, const DomTreeNodePair &RHS) {
365 return LHS.second < RHS.second;
368 } // end of anonymous namespace
370 static void removeLifetimeIntrinsicUsers(AllocaInst *AI) {
371 // Knowing that this alloca is promotable, we know that it's safe to kill all
372 // instructions except for load and store.
374 for (Value::use_iterator UI = AI->use_begin(), UE = AI->use_end();
376 Instruction *I = cast<Instruction>(*UI);
378 if (isa<LoadInst>(I) || isa<StoreInst>(I))
381 if (!I->getType()->isVoidTy()) {
382 // The only users of this bitcast/GEP instruction are lifetime intrinsics.
383 // Follow the use/def chain to erase them now instead of leaving it for
384 // dead code elimination later.
385 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
387 Instruction *Inst = cast<Instruction>(*UI);
389 Inst->eraseFromParent();
392 I->eraseFromParent();
396 void PromoteMem2Reg::run() {
397 Function &F = *DT.getRoot()->getParent();
399 if (AST) PointerAllocaValues.resize(Allocas.size());
400 AllocaDbgDeclares.resize(Allocas.size());
405 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
406 AllocaInst *AI = Allocas[AllocaNum];
408 assert(isAllocaPromotable(AI) &&
409 "Cannot promote non-promotable alloca!");
410 assert(AI->getParent()->getParent() == &F &&
411 "All allocas should be in the same function, which is same as DF!");
413 removeLifetimeIntrinsicUsers(AI);
415 if (AI->use_empty()) {
416 // If there are no uses of the alloca, just delete it now.
417 if (AST) AST->deleteValue(AI);
418 AI->eraseFromParent();
420 // Remove the alloca from the Allocas list, since it has been processed
421 RemoveFromAllocasList(AllocaNum);
426 // Calculate the set of read and write-locations for each alloca. This is
427 // analogous to finding the 'uses' and 'definitions' of each variable.
428 Info.AnalyzeAlloca(AI);
430 // If there is only a single store to this value, replace any loads of
431 // it that are directly dominated by the definition with the value stored.
432 if (Info.DefiningBlocks.size() == 1) {
433 RewriteSingleStoreAlloca(AI, Info, LBI);
435 // Finally, after the scan, check to see if the store is all that is left.
436 if (Info.UsingBlocks.empty()) {
437 // Record debuginfo for the store and remove the declaration's debuginfo.
438 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
440 DIB = new DIBuilder(*DDI->getParent()->getParent()->getParent());
441 ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore, *DIB);
442 DDI->eraseFromParent();
444 // Remove the (now dead) store and alloca.
445 Info.OnlyStore->eraseFromParent();
446 LBI.deleteValue(Info.OnlyStore);
448 if (AST) AST->deleteValue(AI);
449 AI->eraseFromParent();
452 // The alloca has been processed, move on.
453 RemoveFromAllocasList(AllocaNum);
460 // If the alloca is only read and written in one basic block, just perform a
461 // linear sweep over the block to eliminate it.
462 if (Info.OnlyUsedInOneBlock) {
463 PromoteSingleBlockAlloca(AI, Info, LBI);
465 // Finally, after the scan, check to see if the stores are all that is
467 if (Info.UsingBlocks.empty()) {
469 // Remove the (now dead) stores and alloca.
470 while (!AI->use_empty()) {
471 StoreInst *SI = cast<StoreInst>(AI->use_back());
472 // Record debuginfo for the store before removing it.
473 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
475 DIB = new DIBuilder(*SI->getParent()->getParent()->getParent());
476 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB);
478 SI->eraseFromParent();
482 if (AST) AST->deleteValue(AI);
483 AI->eraseFromParent();
486 // The alloca has been processed, move on.
487 RemoveFromAllocasList(AllocaNum);
489 // The alloca's debuginfo can be removed as well.
490 if (DbgDeclareInst *DDI = Info.DbgDeclare)
491 DDI->eraseFromParent();
498 // If we haven't computed dominator tree levels, do so now.
499 if (DomLevels.empty()) {
500 SmallVector<DomTreeNode*, 32> Worklist;
502 DomTreeNode *Root = DT.getRootNode();
504 Worklist.push_back(Root);
506 while (!Worklist.empty()) {
507 DomTreeNode *Node = Worklist.pop_back_val();
508 unsigned ChildLevel = DomLevels[Node] + 1;
509 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end();
511 DomLevels[*CI] = ChildLevel;
512 Worklist.push_back(*CI);
517 // If we haven't computed a numbering for the BB's in the function, do so
519 if (BBNumbers.empty()) {
521 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
525 // If we have an AST to keep updated, remember some pointer value that is
526 // stored into the alloca.
528 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
530 // Remember the dbg.declare intrinsic describing this alloca, if any.
531 if (Info.DbgDeclare) AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare;
533 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
534 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
536 // At this point, we're committed to promoting the alloca using IDF's, and
537 // the standard SSA construction algorithm. Determine which blocks need PHI
538 // nodes and see if we can optimize out some work by avoiding insertion of
540 DetermineInsertionPoint(AI, AllocaNum, Info);
544 return; // All of the allocas must have been trivial!
549 // Set the incoming values for the basic block to be null values for all of
550 // the alloca's. We do this in case there is a load of a value that has not
551 // been stored yet. In this case, it will get this null value.
553 RenamePassData::ValVector Values(Allocas.size());
554 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
555 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
557 // Walks all basic blocks in the function performing the SSA rename algorithm
558 // and inserting the phi nodes we marked as necessary
560 std::vector<RenamePassData> RenamePassWorkList;
561 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
564 RPD.swap(RenamePassWorkList.back());
565 RenamePassWorkList.pop_back();
566 // RenamePass may add new worklist entries.
567 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
568 } while (!RenamePassWorkList.empty());
570 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
573 // Remove the allocas themselves from the function.
574 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
575 Instruction *A = Allocas[i];
577 // If there are any uses of the alloca instructions left, they must be in
578 // unreachable basic blocks that were not processed by walking the dominator
579 // tree. Just delete the users now.
581 A->replaceAllUsesWith(UndefValue::get(A->getType()));
582 if (AST) AST->deleteValue(A);
583 A->eraseFromParent();
586 // Remove alloca's dbg.declare instrinsics from the function.
587 for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i)
588 if (DbgDeclareInst *DDI = AllocaDbgDeclares[i])
589 DDI->eraseFromParent();
591 // Loop over all of the PHI nodes and see if there are any that we can get
592 // rid of because they merge all of the same incoming values. This can
593 // happen due to undef values coming into the PHI nodes. This process is
594 // iterative, because eliminating one PHI node can cause others to be removed.
595 bool EliminatedAPHI = true;
596 while (EliminatedAPHI) {
597 EliminatedAPHI = false;
599 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
600 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
601 PHINode *PN = I->second;
603 // If this PHI node merges one value and/or undefs, get the value.
604 if (Value *V = SimplifyInstruction(PN, 0, &DT)) {
605 if (AST && PN->getType()->isPointerTy())
606 AST->deleteValue(PN);
607 PN->replaceAllUsesWith(V);
608 PN->eraseFromParent();
609 NewPhiNodes.erase(I++);
610 EliminatedAPHI = true;
617 // At this point, the renamer has added entries to PHI nodes for all reachable
618 // code. Unfortunately, there may be unreachable blocks which the renamer
619 // hasn't traversed. If this is the case, the PHI nodes may not
620 // have incoming values for all predecessors. Loop over all PHI nodes we have
621 // created, inserting undef values if they are missing any incoming values.
623 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
624 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
625 // We want to do this once per basic block. As such, only process a block
626 // when we find the PHI that is the first entry in the block.
627 PHINode *SomePHI = I->second;
628 BasicBlock *BB = SomePHI->getParent();
629 if (&BB->front() != SomePHI)
632 // Only do work here if there the PHI nodes are missing incoming values. We
633 // know that all PHI nodes that were inserted in a block will have the same
634 // number of incoming values, so we can just check any of them.
635 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
638 // Get the preds for BB.
639 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
641 // Ok, now we know that all of the PHI nodes are missing entries for some
642 // basic blocks. Start by sorting the incoming predecessors for efficient
644 std::sort(Preds.begin(), Preds.end());
646 // Now we loop through all BB's which have entries in SomePHI and remove
647 // them from the Preds list.
648 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
649 // Do a log(n) search of the Preds list for the entry we want.
650 SmallVector<BasicBlock*, 16>::iterator EntIt =
651 std::lower_bound(Preds.begin(), Preds.end(),
652 SomePHI->getIncomingBlock(i));
653 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
654 "PHI node has entry for a block which is not a predecessor!");
660 // At this point, the blocks left in the preds list must have dummy
661 // entries inserted into every PHI nodes for the block. Update all the phi
662 // nodes in this block that we are inserting (there could be phis before
664 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
665 BasicBlock::iterator BBI = BB->begin();
666 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
667 SomePHI->getNumIncomingValues() == NumBadPreds) {
668 Value *UndefVal = UndefValue::get(SomePHI->getType());
669 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
670 SomePHI->addIncoming(UndefVal, Preds[pred]);
678 /// ComputeLiveInBlocks - Determine which blocks the value is live in. These
679 /// are blocks which lead to uses. Knowing this allows us to avoid inserting
680 /// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
682 void PromoteMem2Reg::
683 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
684 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
685 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
687 // To determine liveness, we must iterate through the predecessors of blocks
688 // where the def is live. Blocks are added to the worklist if we need to
689 // check their predecessors. Start with all the using blocks.
690 SmallVector<BasicBlock*, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
691 Info.UsingBlocks.end());
693 // If any of the using blocks is also a definition block, check to see if the
694 // definition occurs before or after the use. If it happens before the use,
695 // the value isn't really live-in.
696 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
697 BasicBlock *BB = LiveInBlockWorklist[i];
698 if (!DefBlocks.count(BB)) continue;
700 // Okay, this is a block that both uses and defines the value. If the first
701 // reference to the alloca is a def (store), then we know it isn't live-in.
702 for (BasicBlock::iterator I = BB->begin(); ; ++I) {
703 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
704 if (SI->getOperand(1) != AI) continue;
706 // We found a store to the alloca before a load. The alloca is not
707 // actually live-in here.
708 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
709 LiveInBlockWorklist.pop_back();
714 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
715 if (LI->getOperand(0) != AI) continue;
717 // Okay, we found a load before a store to the alloca. It is actually
718 // live into this block.
724 // Now that we have a set of blocks where the phi is live-in, recursively add
725 // their predecessors until we find the full region the value is live.
726 while (!LiveInBlockWorklist.empty()) {
727 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
729 // The block really is live in here, insert it into the set. If already in
730 // the set, then it has already been processed.
731 if (!LiveInBlocks.insert(BB))
734 // Since the value is live into BB, it is either defined in a predecessor or
735 // live into it to. Add the preds to the worklist unless they are a
737 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
740 // The value is not live into a predecessor if it defines the value.
741 if (DefBlocks.count(P))
744 // Otherwise it is, add to the worklist.
745 LiveInBlockWorklist.push_back(P);
750 /// DetermineInsertionPoint - At this point, we're committed to promoting the
751 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
752 /// which blocks need phi nodes and see if we can optimize out some work by
753 /// avoiding insertion of dead phi nodes.
754 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
756 // Unique the set of defining blocks for efficient lookup.
757 SmallPtrSet<BasicBlock*, 32> DefBlocks;
758 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
760 // Determine which blocks the value is live in. These are blocks which lead
762 SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
763 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
765 // Use a priority queue keyed on dominator tree level so that inserted nodes
766 // are handled from the bottom of the dominator tree upwards.
767 typedef std::priority_queue<DomTreeNodePair, SmallVector<DomTreeNodePair, 32>,
768 DomTreeNodeCompare> IDFPriorityQueue;
771 for (SmallPtrSet<BasicBlock*, 32>::const_iterator I = DefBlocks.begin(),
772 E = DefBlocks.end(); I != E; ++I) {
773 if (DomTreeNode *Node = DT.getNode(*I))
774 PQ.push(std::make_pair(Node, DomLevels[Node]));
777 SmallVector<std::pair<unsigned, BasicBlock*>, 32> DFBlocks;
778 SmallPtrSet<DomTreeNode*, 32> Visited;
779 SmallVector<DomTreeNode*, 32> Worklist;
780 while (!PQ.empty()) {
781 DomTreeNodePair RootPair = PQ.top();
783 DomTreeNode *Root = RootPair.first;
784 unsigned RootLevel = RootPair.second;
786 // Walk all dominator tree children of Root, inspecting their CFG edges with
787 // targets elsewhere on the dominator tree. Only targets whose level is at
788 // most Root's level are added to the iterated dominance frontier of the
792 Worklist.push_back(Root);
794 while (!Worklist.empty()) {
795 DomTreeNode *Node = Worklist.pop_back_val();
796 BasicBlock *BB = Node->getBlock();
798 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE;
800 DomTreeNode *SuccNode = DT.getNode(*SI);
802 // Quickly skip all CFG edges that are also dominator tree edges instead
803 // of catching them below.
804 if (SuccNode->getIDom() == Node)
807 unsigned SuccLevel = DomLevels[SuccNode];
808 if (SuccLevel > RootLevel)
811 if (!Visited.insert(SuccNode))
814 BasicBlock *SuccBB = SuccNode->getBlock();
815 if (!LiveInBlocks.count(SuccBB))
818 DFBlocks.push_back(std::make_pair(BBNumbers[SuccBB], SuccBB));
819 if (!DefBlocks.count(SuccBB))
820 PQ.push(std::make_pair(SuccNode, SuccLevel));
823 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end(); CI != CE;
825 if (!Visited.count(*CI))
826 Worklist.push_back(*CI);
831 if (DFBlocks.size() > 1)
832 std::sort(DFBlocks.begin(), DFBlocks.end());
834 unsigned CurrentVersion = 0;
835 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i)
836 QueuePhiNode(DFBlocks[i].second, AllocaNum, CurrentVersion);
839 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
840 /// replace any loads of it that are directly dominated by the definition with
841 /// the value stored.
842 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
844 LargeBlockInfo &LBI) {
845 StoreInst *OnlyStore = Info.OnlyStore;
846 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
847 BasicBlock *StoreBB = OnlyStore->getParent();
850 // Clear out UsingBlocks. We will reconstruct it here if needed.
851 Info.UsingBlocks.clear();
853 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
854 Instruction *UserInst = cast<Instruction>(*UI++);
855 if (!isa<LoadInst>(UserInst)) {
856 assert(UserInst == OnlyStore && "Should only have load/stores");
859 LoadInst *LI = cast<LoadInst>(UserInst);
861 // Okay, if we have a load from the alloca, we want to replace it with the
862 // only value stored to the alloca. We can do this if the value is
863 // dominated by the store. If not, we use the rest of the mem2reg machinery
864 // to insert the phi nodes as needed.
865 if (!StoringGlobalVal) { // Non-instructions are always dominated.
866 if (LI->getParent() == StoreBB) {
867 // If we have a use that is in the same block as the store, compare the
868 // indices of the two instructions to see which one came first. If the
869 // load came before the store, we can't handle it.
870 if (StoreIndex == -1)
871 StoreIndex = LBI.getInstructionIndex(OnlyStore);
873 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
874 // Can't handle this load, bail out.
875 Info.UsingBlocks.push_back(StoreBB);
879 } else if (LI->getParent() != StoreBB &&
880 !dominates(StoreBB, LI->getParent())) {
881 // If the load and store are in different blocks, use BB dominance to
882 // check their relationships. If the store doesn't dom the use, bail
884 Info.UsingBlocks.push_back(LI->getParent());
889 // Otherwise, we *can* safely rewrite this load.
890 Value *ReplVal = OnlyStore->getOperand(0);
891 // If the replacement value is the load, this must occur in unreachable
894 ReplVal = UndefValue::get(LI->getType());
895 LI->replaceAllUsesWith(ReplVal);
896 if (AST && LI->getType()->isPointerTy())
897 AST->deleteValue(LI);
898 LI->eraseFromParent();
905 /// StoreIndexSearchPredicate - This is a helper predicate used to search by the
906 /// first element of a pair.
907 struct StoreIndexSearchPredicate {
908 bool operator()(const std::pair<unsigned, StoreInst*> &LHS,
909 const std::pair<unsigned, StoreInst*> &RHS) {
910 return LHS.first < RHS.first;
916 /// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
917 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
918 /// potentially useless PHI nodes by just performing a single linear pass over
919 /// the basic block using the Alloca.
921 /// If we cannot promote this alloca (because it is read before it is written),
922 /// return true. This is necessary in cases where, due to control flow, the
923 /// alloca is potentially undefined on some control flow paths. e.g. code like
924 /// this is potentially correct:
926 /// for (...) { if (c) { A = undef; undef = B; } }
928 /// ... so long as A is not used before undef is set.
930 void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
931 LargeBlockInfo &LBI) {
932 // The trickiest case to handle is when we have large blocks. Because of this,
933 // this code is optimized assuming that large blocks happen. This does not
934 // significantly pessimize the small block case. This uses LargeBlockInfo to
935 // make it efficient to get the index of various operations in the block.
937 // Clear out UsingBlocks. We will reconstruct it here if needed.
938 Info.UsingBlocks.clear();
940 // Walk the use-def list of the alloca, getting the locations of all stores.
941 typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
942 StoresByIndexTy StoresByIndex;
944 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
946 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
947 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
949 // If there are no stores to the alloca, just replace any loads with undef.
950 if (StoresByIndex.empty()) {
951 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
952 if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
953 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
954 if (AST && LI->getType()->isPointerTy())
955 AST->deleteValue(LI);
957 LI->eraseFromParent();
962 // Sort the stores by their index, making it efficient to do a lookup with a
964 std::sort(StoresByIndex.begin(), StoresByIndex.end());
966 // Walk all of the loads from this alloca, replacing them with the nearest
967 // store above them, if any.
968 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
969 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
972 unsigned LoadIdx = LBI.getInstructionIndex(LI);
974 // Find the nearest store that has a lower than this load.
975 StoresByIndexTy::iterator I =
976 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
977 std::pair<unsigned, StoreInst*>(LoadIdx, static_cast<StoreInst*>(0)),
978 StoreIndexSearchPredicate());
980 // If there is no store before this load, then we can't promote this load.
981 if (I == StoresByIndex.begin()) {
982 // Can't handle this load, bail out.
983 Info.UsingBlocks.push_back(LI->getParent());
987 // Otherwise, there was a store before this load, the load takes its value.
989 LI->replaceAllUsesWith(I->second->getOperand(0));
990 if (AST && LI->getType()->isPointerTy())
991 AST->deleteValue(LI);
992 LI->eraseFromParent();
997 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
998 // Alloca returns true if there wasn't already a phi-node for that variable
1000 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
1001 unsigned &Version) {
1002 // Look up the basic-block in question.
1003 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
1005 // If the BB already has a phi node added for the i'th alloca then we're done!
1006 if (PN) return false;
1008 // Create a PhiNode using the dereferenced type... and add the phi-node to the
1010 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(), getNumPreds(BB),
1011 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
1014 PhiToAllocaMap[PN] = AllocaNo;
1016 if (AST && PN->getType()->isPointerTy())
1017 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
1022 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
1023 // stores to the allocas which we are promoting. IncomingVals indicates what
1024 // value each Alloca contains on exit from the predecessor block Pred.
1026 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
1027 RenamePassData::ValVector &IncomingVals,
1028 std::vector<RenamePassData> &Worklist) {
1030 // If we are inserting any phi nodes into this BB, they will already be in the
1032 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
1033 // If we have PHI nodes to update, compute the number of edges from Pred to
1035 if (PhiToAllocaMap.count(APN)) {
1036 // We want to be able to distinguish between PHI nodes being inserted by
1037 // this invocation of mem2reg from those phi nodes that already existed in
1038 // the IR before mem2reg was run. We determine that APN is being inserted
1039 // because it is missing incoming edges. All other PHI nodes being
1040 // inserted by this pass of mem2reg will have the same number of incoming
1041 // operands so far. Remember this count.
1042 unsigned NewPHINumOperands = APN->getNumOperands();
1044 unsigned NumEdges = 0;
1045 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
1048 assert(NumEdges && "Must be at least one edge from Pred to BB!");
1050 // Add entries for all the phis.
1051 BasicBlock::iterator PNI = BB->begin();
1053 unsigned AllocaNo = PhiToAllocaMap[APN];
1055 // Add N incoming values to the PHI node.
1056 for (unsigned i = 0; i != NumEdges; ++i)
1057 APN->addIncoming(IncomingVals[AllocaNo], Pred);
1059 // The currently active variable for this block is now the PHI.
1060 IncomingVals[AllocaNo] = APN;
1062 // Get the next phi node.
1064 APN = dyn_cast<PHINode>(PNI);
1065 if (APN == 0) break;
1067 // Verify that it is missing entries. If not, it is not being inserted
1068 // by this mem2reg invocation so we want to ignore it.
1069 } while (APN->getNumOperands() == NewPHINumOperands);
1073 // Don't revisit blocks.
1074 if (!Visited.insert(BB)) return;
1076 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
1077 Instruction *I = II++; // get the instruction, increment iterator
1079 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1080 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
1083 DenseMap<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
1084 if (AI == AllocaLookup.end()) continue;
1086 Value *V = IncomingVals[AI->second];
1088 // Anything using the load now uses the current value.
1089 LI->replaceAllUsesWith(V);
1090 if (AST && LI->getType()->isPointerTy())
1091 AST->deleteValue(LI);
1092 BB->getInstList().erase(LI);
1093 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1094 // Delete this instruction and mark the name as the current holder of the
1096 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
1097 if (!Dest) continue;
1099 DenseMap<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
1100 if (ai == AllocaLookup.end())
1103 // what value were we writing?
1104 IncomingVals[ai->second] = SI->getOperand(0);
1105 // Record debuginfo for the store before removing it.
1106 if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second]) {
1108 DIB = new DIBuilder(*SI->getParent()->getParent()->getParent());
1109 ConvertDebugDeclareToDebugValue(DDI, SI, *DIB);
1111 BB->getInstList().erase(SI);
1115 // 'Recurse' to our successors.
1116 succ_iterator I = succ_begin(BB), E = succ_end(BB);
1119 // Keep track of the successors so we don't visit the same successor twice
1120 SmallPtrSet<BasicBlock*, 8> VisitedSuccs;
1122 // Handle the first successor without using the worklist.
1123 VisitedSuccs.insert(*I);
1129 if (VisitedSuccs.insert(*I))
1130 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
1135 /// PromoteMemToReg - Promote the specified list of alloca instructions into
1136 /// scalar registers, inserting PHI nodes as appropriate. This function does
1137 /// not modify the CFG of the function at all. All allocas must be from the
1140 /// If AST is specified, the specified tracker is updated to reflect changes
1143 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
1144 DominatorTree &DT, AliasSetTracker *AST) {
1145 // If there is nothing to do, bail out...
1146 if (Allocas.empty()) return;
1148 PromoteMem2Reg(Allocas, DT, AST).run();