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/Dominators.h"
39 #include "llvm/Analysis/InstructionSimplify.h"
40 #include "llvm/ADT/DenseMap.h"
41 #include "llvm/ADT/SmallPtrSet.h"
42 #include "llvm/ADT/SmallVector.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/STLExtras.h"
45 #include "llvm/Support/CFG.h"
51 STATISTIC(NumLocalPromoted, "Number of alloca's promoted within one block");
52 STATISTIC(NumSingleStore, "Number of alloca's promoted with a single store");
53 STATISTIC(NumDeadAlloca, "Number of dead alloca's removed");
54 STATISTIC(NumPHIInsert, "Number of PHI nodes inserted");
58 struct DenseMapInfo<std::pair<BasicBlock*, unsigned> > {
59 typedef std::pair<BasicBlock*, unsigned> EltTy;
60 static inline EltTy getEmptyKey() {
61 return EltTy(reinterpret_cast<BasicBlock*>(-1), ~0U);
63 static inline EltTy getTombstoneKey() {
64 return EltTy(reinterpret_cast<BasicBlock*>(-2), 0U);
66 static unsigned getHashValue(const std::pair<BasicBlock*, unsigned> &Val) {
67 return DenseMapInfo<void*>::getHashValue(Val.first) + Val.second*2;
69 static bool isEqual(const EltTy &LHS, const EltTy &RHS) {
75 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
76 /// This is true if there are only loads and stores to the alloca.
78 bool llvm::isAllocaPromotable(const AllocaInst *AI) {
79 // FIXME: If the memory unit is of pointer or integer type, we can permit
80 // assignments to subsections of the memory unit.
82 // Only allow direct and non-volatile loads and stores...
83 for (Value::const_use_iterator UI = AI->use_begin(), UE = AI->use_end();
84 UI != UE; ++UI) { // Loop over all of the uses of the alloca
86 if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
89 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
90 if (SI->getOperand(0) == AI)
91 return false; // Don't allow a store OF the AI, only INTO the AI.
102 /// FindAllocaDbgDeclare - Finds the llvm.dbg.declare intrinsic describing the
103 /// alloca 'V', if any.
104 static DbgDeclareInst *FindAllocaDbgDeclare(Value *V) {
105 if (MDNode *DebugNode = MDNode::getIfExists(V->getContext(), &V, 1))
106 for (Value::use_iterator UI = DebugNode->use_begin(),
107 E = DebugNode->use_end(); UI != E; ++UI)
108 if (DbgDeclareInst *DDI = dyn_cast<DbgDeclareInst>(*UI))
117 // Data package used by RenamePass()
118 class RenamePassData {
120 typedef std::vector<Value *> ValVector;
122 RenamePassData() : BB(NULL), Pred(NULL), Values() {}
123 RenamePassData(BasicBlock *B, BasicBlock *P,
124 const ValVector &V) : BB(B), Pred(P), Values(V) {}
129 void swap(RenamePassData &RHS) {
130 std::swap(BB, RHS.BB);
131 std::swap(Pred, RHS.Pred);
132 Values.swap(RHS.Values);
136 /// LargeBlockInfo - This assigns and keeps a per-bb relative ordering of
137 /// load/store instructions in the block that directly load or store an alloca.
139 /// This functionality is important because it avoids scanning large basic
140 /// blocks multiple times when promoting many allocas in the same block.
141 class LargeBlockInfo {
142 /// InstNumbers - For each instruction that we track, keep the index of the
143 /// instruction. The index starts out as the number of the instruction from
144 /// the start of the block.
145 DenseMap<const Instruction *, unsigned> InstNumbers;
148 /// isInterestingInstruction - This code only looks at accesses to allocas.
149 static bool isInterestingInstruction(const Instruction *I) {
150 return (isa<LoadInst>(I) && isa<AllocaInst>(I->getOperand(0))) ||
151 (isa<StoreInst>(I) && isa<AllocaInst>(I->getOperand(1)));
154 /// getInstructionIndex - Get or calculate the index of the specified
156 unsigned getInstructionIndex(const Instruction *I) {
157 assert(isInterestingInstruction(I) &&
158 "Not a load/store to/from an alloca?");
160 // If we already have this instruction number, return it.
161 DenseMap<const Instruction *, unsigned>::iterator It = InstNumbers.find(I);
162 if (It != InstNumbers.end()) return It->second;
164 // Scan the whole block to get the instruction. This accumulates
165 // information for every interesting instruction in the block, in order to
166 // avoid gratuitus rescans.
167 const BasicBlock *BB = I->getParent();
169 for (BasicBlock::const_iterator BBI = BB->begin(), E = BB->end();
171 if (isInterestingInstruction(BBI))
172 InstNumbers[BBI] = InstNo++;
173 It = InstNumbers.find(I);
175 assert(It != InstNumbers.end() && "Didn't insert instruction?");
179 void deleteValue(const Instruction *I) {
180 InstNumbers.erase(I);
188 struct PromoteMem2Reg {
189 /// Allocas - The alloca instructions being promoted.
191 std::vector<AllocaInst*> Allocas;
195 /// AST - An AliasSetTracker object to update. If null, don't update it.
197 AliasSetTracker *AST;
199 /// AllocaLookup - Reverse mapping of Allocas.
201 std::map<AllocaInst*, unsigned> AllocaLookup;
203 /// NewPhiNodes - The PhiNodes we're adding.
205 DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*> NewPhiNodes;
207 /// PhiToAllocaMap - For each PHI node, keep track of which entry in Allocas
208 /// it corresponds to.
209 DenseMap<PHINode*, unsigned> PhiToAllocaMap;
211 /// PointerAllocaValues - If we are updating an AliasSetTracker, then for
212 /// each alloca that is of pointer type, we keep track of what to copyValue
213 /// to the inserted PHI nodes here.
215 std::vector<Value*> PointerAllocaValues;
217 /// AllocaDbgDeclares - For each alloca, we keep track of the dbg.declare
218 /// intrinsic that describes it, if any, so that we can convert it to a
219 /// dbg.value intrinsic if the alloca gets promoted.
220 SmallVector<DbgDeclareInst*, 8> AllocaDbgDeclares;
222 /// Visited - The set of basic blocks the renamer has already visited.
224 SmallPtrSet<BasicBlock*, 16> Visited;
226 /// BBNumbers - Contains a stable numbering of basic blocks to avoid
227 /// non-determinstic behavior.
228 DenseMap<BasicBlock*, unsigned> BBNumbers;
230 /// DomLevels - Maps DomTreeNodes to their level in the dominator tree.
231 DenseMap<DomTreeNode*, unsigned> DomLevels;
233 /// BBNumPreds - Lazily compute the number of predecessors a block has.
234 DenseMap<const BasicBlock*, unsigned> BBNumPreds;
236 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
237 AliasSetTracker *ast)
238 : Allocas(A), DT(dt), DIF(0), AST(ast) {}
245 /// dominates - Return true if BB1 dominates BB2 using the DominatorTree.
247 bool dominates(BasicBlock *BB1, BasicBlock *BB2) const {
248 return DT.dominates(BB1, BB2);
252 void RemoveFromAllocasList(unsigned &AllocaIdx) {
253 Allocas[AllocaIdx] = Allocas.back();
258 unsigned getNumPreds(const BasicBlock *BB) {
259 unsigned &NP = BBNumPreds[BB];
261 NP = std::distance(pred_begin(BB), pred_end(BB))+1;
265 void DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
267 void ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
268 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
269 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks);
271 void RewriteSingleStoreAlloca(AllocaInst *AI, AllocaInfo &Info,
272 LargeBlockInfo &LBI);
273 void PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
274 LargeBlockInfo &LBI);
275 void ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI, StoreInst *SI);
278 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
279 RenamePassData::ValVector &IncVals,
280 std::vector<RenamePassData> &Worklist);
281 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version);
285 std::vector<BasicBlock*> DefiningBlocks;
286 std::vector<BasicBlock*> UsingBlocks;
288 StoreInst *OnlyStore;
289 BasicBlock *OnlyBlock;
290 bool OnlyUsedInOneBlock;
292 Value *AllocaPointerVal;
293 DbgDeclareInst *DbgDeclare;
296 DefiningBlocks.clear();
300 OnlyUsedInOneBlock = true;
301 AllocaPointerVal = 0;
305 /// AnalyzeAlloca - Scan the uses of the specified alloca, filling in our
307 void AnalyzeAlloca(AllocaInst *AI) {
310 // As we scan the uses of the alloca instruction, keep track of stores,
311 // and decide whether all of the loads and stores to the alloca are within
312 // the same basic block.
313 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
315 Instruction *User = cast<Instruction>(*UI++);
317 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
318 // Remember the basic blocks which define new values for the alloca
319 DefiningBlocks.push_back(SI->getParent());
320 AllocaPointerVal = SI->getOperand(0);
323 LoadInst *LI = cast<LoadInst>(User);
324 // Otherwise it must be a load instruction, keep track of variable
326 UsingBlocks.push_back(LI->getParent());
327 AllocaPointerVal = LI;
330 if (OnlyUsedInOneBlock) {
332 OnlyBlock = User->getParent();
333 else if (OnlyBlock != User->getParent())
334 OnlyUsedInOneBlock = false;
338 DbgDeclare = FindAllocaDbgDeclare(AI);
342 typedef std::pair<DomTreeNode*, unsigned> DomTreeNodePair;
344 struct DomTreeNodeCompare {
345 bool operator()(const DomTreeNodePair &LHS, const DomTreeNodePair &RHS) {
346 return LHS.second < RHS.second;
349 } // end of anonymous namespace
352 void PromoteMem2Reg::run() {
353 Function &F = *DT.getRoot()->getParent();
355 if (AST) PointerAllocaValues.resize(Allocas.size());
356 AllocaDbgDeclares.resize(Allocas.size());
361 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
362 AllocaInst *AI = Allocas[AllocaNum];
364 assert(isAllocaPromotable(AI) &&
365 "Cannot promote non-promotable alloca!");
366 assert(AI->getParent()->getParent() == &F &&
367 "All allocas should be in the same function, which is same as DF!");
369 if (AI->use_empty()) {
370 // If there are no uses of the alloca, just delete it now.
371 if (AST) AST->deleteValue(AI);
372 AI->eraseFromParent();
374 // Remove the alloca from the Allocas list, since it has been processed
375 RemoveFromAllocasList(AllocaNum);
380 // Calculate the set of read and write-locations for each alloca. This is
381 // analogous to finding the 'uses' and 'definitions' of each variable.
382 Info.AnalyzeAlloca(AI);
384 // If there is only a single store to this value, replace any loads of
385 // it that are directly dominated by the definition with the value stored.
386 if (Info.DefiningBlocks.size() == 1) {
387 RewriteSingleStoreAlloca(AI, Info, LBI);
389 // Finally, after the scan, check to see if the store is all that is left.
390 if (Info.UsingBlocks.empty()) {
391 // Record debuginfo for the store and remove the declaration's debuginfo.
392 if (DbgDeclareInst *DDI = Info.DbgDeclare) {
393 ConvertDebugDeclareToDebugValue(DDI, Info.OnlyStore);
394 DDI->eraseFromParent();
396 // Remove the (now dead) store and alloca.
397 Info.OnlyStore->eraseFromParent();
398 LBI.deleteValue(Info.OnlyStore);
400 if (AST) AST->deleteValue(AI);
401 AI->eraseFromParent();
404 // The alloca has been processed, move on.
405 RemoveFromAllocasList(AllocaNum);
412 // If the alloca is only read and written in one basic block, just perform a
413 // linear sweep over the block to eliminate it.
414 if (Info.OnlyUsedInOneBlock) {
415 PromoteSingleBlockAlloca(AI, Info, LBI);
417 // Finally, after the scan, check to see if the stores are all that is
419 if (Info.UsingBlocks.empty()) {
421 // Remove the (now dead) stores and alloca.
422 while (!AI->use_empty()) {
423 StoreInst *SI = cast<StoreInst>(AI->use_back());
424 // Record debuginfo for the store before removing it.
425 if (DbgDeclareInst *DDI = Info.DbgDeclare)
426 ConvertDebugDeclareToDebugValue(DDI, SI);
427 SI->eraseFromParent();
431 if (AST) AST->deleteValue(AI);
432 AI->eraseFromParent();
435 // The alloca has been processed, move on.
436 RemoveFromAllocasList(AllocaNum);
438 // The alloca's debuginfo can be removed as well.
439 if (DbgDeclareInst *DDI = Info.DbgDeclare)
440 DDI->eraseFromParent();
447 // If we haven't computed dominator tree levels, do so now.
448 if (DomLevels.empty()) {
449 SmallVector<DomTreeNode*, 32> Worklist;
451 DomTreeNode *Root = DT.getRootNode();
453 Worklist.push_back(Root);
455 while (!Worklist.empty()) {
456 DomTreeNode *Node = Worklist.pop_back_val();
457 unsigned ChildLevel = DomLevels[Node] + 1;
458 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end();
460 DomLevels[*CI] = ChildLevel;
461 Worklist.push_back(*CI);
466 // If we haven't computed a numbering for the BB's in the function, do so
468 if (BBNumbers.empty()) {
470 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
474 // If we have an AST to keep updated, remember some pointer value that is
475 // stored into the alloca.
477 PointerAllocaValues[AllocaNum] = Info.AllocaPointerVal;
479 // Remember the dbg.declare intrinsic describing this alloca, if any.
480 if (Info.DbgDeclare) AllocaDbgDeclares[AllocaNum] = Info.DbgDeclare;
482 // Keep the reverse mapping of the 'Allocas' array for the rename pass.
483 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
485 // At this point, we're committed to promoting the alloca using IDF's, and
486 // the standard SSA construction algorithm. Determine which blocks need PHI
487 // nodes and see if we can optimize out some work by avoiding insertion of
489 DetermineInsertionPoint(AI, AllocaNum, Info);
493 return; // All of the allocas must have been trivial!
498 // Set the incoming values for the basic block to be null values for all of
499 // the alloca's. We do this in case there is a load of a value that has not
500 // been stored yet. In this case, it will get this null value.
502 RenamePassData::ValVector Values(Allocas.size());
503 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
504 Values[i] = UndefValue::get(Allocas[i]->getAllocatedType());
506 // Walks all basic blocks in the function performing the SSA rename algorithm
507 // and inserting the phi nodes we marked as necessary
509 std::vector<RenamePassData> RenamePassWorkList;
510 RenamePassWorkList.push_back(RenamePassData(F.begin(), 0, Values));
513 RPD.swap(RenamePassWorkList.back());
514 RenamePassWorkList.pop_back();
515 // RenamePass may add new worklist entries.
516 RenamePass(RPD.BB, RPD.Pred, RPD.Values, RenamePassWorkList);
517 } while (!RenamePassWorkList.empty());
519 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
522 // Remove the allocas themselves from the function.
523 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
524 Instruction *A = Allocas[i];
526 // If there are any uses of the alloca instructions left, they must be in
527 // unreachable basic blocks that were not processed by walking the dominator
528 // tree. Just delete the users now.
530 A->replaceAllUsesWith(UndefValue::get(A->getType()));
531 if (AST) AST->deleteValue(A);
532 A->eraseFromParent();
535 // Remove alloca's dbg.declare instrinsics from the function.
536 for (unsigned i = 0, e = AllocaDbgDeclares.size(); i != e; ++i)
537 if (DbgDeclareInst *DDI = AllocaDbgDeclares[i])
538 DDI->eraseFromParent();
540 // Loop over all of the PHI nodes and see if there are any that we can get
541 // rid of because they merge all of the same incoming values. This can
542 // happen due to undef values coming into the PHI nodes. This process is
543 // iterative, because eliminating one PHI node can cause others to be removed.
544 bool EliminatedAPHI = true;
545 while (EliminatedAPHI) {
546 EliminatedAPHI = false;
548 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
549 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E;) {
550 PHINode *PN = I->second;
552 // If this PHI node merges one value and/or undefs, get the value.
553 if (Value *V = SimplifyInstruction(PN, 0, &DT)) {
554 if (AST && PN->getType()->isPointerTy())
555 AST->deleteValue(PN);
556 PN->replaceAllUsesWith(V);
557 PN->eraseFromParent();
558 NewPhiNodes.erase(I++);
559 EliminatedAPHI = true;
566 // At this point, the renamer has added entries to PHI nodes for all reachable
567 // code. Unfortunately, there may be unreachable blocks which the renamer
568 // hasn't traversed. If this is the case, the PHI nodes may not
569 // have incoming values for all predecessors. Loop over all PHI nodes we have
570 // created, inserting undef values if they are missing any incoming values.
572 for (DenseMap<std::pair<BasicBlock*, unsigned>, PHINode*>::iterator I =
573 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
574 // We want to do this once per basic block. As such, only process a block
575 // when we find the PHI that is the first entry in the block.
576 PHINode *SomePHI = I->second;
577 BasicBlock *BB = SomePHI->getParent();
578 if (&BB->front() != SomePHI)
581 // Only do work here if there the PHI nodes are missing incoming values. We
582 // know that all PHI nodes that were inserted in a block will have the same
583 // number of incoming values, so we can just check any of them.
584 if (SomePHI->getNumIncomingValues() == getNumPreds(BB))
587 // Get the preds for BB.
588 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB));
590 // Ok, now we know that all of the PHI nodes are missing entries for some
591 // basic blocks. Start by sorting the incoming predecessors for efficient
593 std::sort(Preds.begin(), Preds.end());
595 // Now we loop through all BB's which have entries in SomePHI and remove
596 // them from the Preds list.
597 for (unsigned i = 0, e = SomePHI->getNumIncomingValues(); i != e; ++i) {
598 // Do a log(n) search of the Preds list for the entry we want.
599 SmallVector<BasicBlock*, 16>::iterator EntIt =
600 std::lower_bound(Preds.begin(), Preds.end(),
601 SomePHI->getIncomingBlock(i));
602 assert(EntIt != Preds.end() && *EntIt == SomePHI->getIncomingBlock(i)&&
603 "PHI node has entry for a block which is not a predecessor!");
609 // At this point, the blocks left in the preds list must have dummy
610 // entries inserted into every PHI nodes for the block. Update all the phi
611 // nodes in this block that we are inserting (there could be phis before
613 unsigned NumBadPreds = SomePHI->getNumIncomingValues();
614 BasicBlock::iterator BBI = BB->begin();
615 while ((SomePHI = dyn_cast<PHINode>(BBI++)) &&
616 SomePHI->getNumIncomingValues() == NumBadPreds) {
617 Value *UndefVal = UndefValue::get(SomePHI->getType());
618 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
619 SomePHI->addIncoming(UndefVal, Preds[pred]);
627 /// ComputeLiveInBlocks - Determine which blocks the value is live in. These
628 /// are blocks which lead to uses. Knowing this allows us to avoid inserting
629 /// PHI nodes into blocks which don't lead to uses (thus, the inserted phi nodes
631 void PromoteMem2Reg::
632 ComputeLiveInBlocks(AllocaInst *AI, AllocaInfo &Info,
633 const SmallPtrSet<BasicBlock*, 32> &DefBlocks,
634 SmallPtrSet<BasicBlock*, 32> &LiveInBlocks) {
636 // To determine liveness, we must iterate through the predecessors of blocks
637 // where the def is live. Blocks are added to the worklist if we need to
638 // check their predecessors. Start with all the using blocks.
639 SmallVector<BasicBlock*, 64> LiveInBlockWorklist(Info.UsingBlocks.begin(),
640 Info.UsingBlocks.end());
642 // If any of the using blocks is also a definition block, check to see if the
643 // definition occurs before or after the use. If it happens before the use,
644 // the value isn't really live-in.
645 for (unsigned i = 0, e = LiveInBlockWorklist.size(); i != e; ++i) {
646 BasicBlock *BB = LiveInBlockWorklist[i];
647 if (!DefBlocks.count(BB)) continue;
649 // Okay, this is a block that both uses and defines the value. If the first
650 // reference to the alloca is a def (store), then we know it isn't live-in.
651 for (BasicBlock::iterator I = BB->begin(); ; ++I) {
652 if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
653 if (SI->getOperand(1) != AI) continue;
655 // We found a store to the alloca before a load. The alloca is not
656 // actually live-in here.
657 LiveInBlockWorklist[i] = LiveInBlockWorklist.back();
658 LiveInBlockWorklist.pop_back();
663 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
664 if (LI->getOperand(0) != AI) continue;
666 // Okay, we found a load before a store to the alloca. It is actually
667 // live into this block.
673 // Now that we have a set of blocks where the phi is live-in, recursively add
674 // their predecessors until we find the full region the value is live.
675 while (!LiveInBlockWorklist.empty()) {
676 BasicBlock *BB = LiveInBlockWorklist.pop_back_val();
678 // The block really is live in here, insert it into the set. If already in
679 // the set, then it has already been processed.
680 if (!LiveInBlocks.insert(BB))
683 // Since the value is live into BB, it is either defined in a predecessor or
684 // live into it to. Add the preds to the worklist unless they are a
686 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
689 // The value is not live into a predecessor if it defines the value.
690 if (DefBlocks.count(P))
693 // Otherwise it is, add to the worklist.
694 LiveInBlockWorklist.push_back(P);
699 /// DetermineInsertionPoint - At this point, we're committed to promoting the
700 /// alloca using IDF's, and the standard SSA construction algorithm. Determine
701 /// which blocks need phi nodes and see if we can optimize out some work by
702 /// avoiding insertion of dead phi nodes.
703 void PromoteMem2Reg::DetermineInsertionPoint(AllocaInst *AI, unsigned AllocaNum,
705 // Unique the set of defining blocks for efficient lookup.
706 SmallPtrSet<BasicBlock*, 32> DefBlocks;
707 DefBlocks.insert(Info.DefiningBlocks.begin(), Info.DefiningBlocks.end());
709 // Determine which blocks the value is live in. These are blocks which lead
711 SmallPtrSet<BasicBlock*, 32> LiveInBlocks;
712 ComputeLiveInBlocks(AI, Info, DefBlocks, LiveInBlocks);
714 // Use a priority queue keyed on dominator tree level so that inserted nodes
715 // are handled from the bottom of the dominator tree upwards.
716 typedef std::priority_queue<DomTreeNodePair, SmallVector<DomTreeNodePair, 32>,
717 DomTreeNodeCompare> IDFPriorityQueue;
720 for (SmallPtrSet<BasicBlock*, 32>::const_iterator I = DefBlocks.begin(),
721 E = DefBlocks.end(); I != E; ++I) {
722 if (DomTreeNode *Node = DT.getNode(*I))
723 PQ.push(std::make_pair(Node, DomLevels[Node]));
726 std::vector<std::pair<unsigned, BasicBlock*> > DFBlocks;
727 SmallPtrSet<DomTreeNode*, 32> Visited;
728 SmallVector<DomTreeNode*, 32> Worklist;
729 while (!PQ.empty()) {
730 DomTreeNodePair RootPair = PQ.top();
732 DomTreeNode *Root = RootPair.first;
733 unsigned RootLevel = RootPair.second;
735 // Walk all dominator tree children of Root, inspecting their CFG edges with
736 // targets elsewhere on the dominator tree. Only targets whose level is at
737 // most Root's level are added to the iterated dominance frontier of the
741 Worklist.push_back(Root);
743 while (!Worklist.empty()) {
744 DomTreeNode *Node = Worklist.pop_back_val();
745 BasicBlock *BB = Node->getBlock();
747 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE;
749 DomTreeNode *SuccNode = DT.getNode(*SI);
751 // Quickly skip all CFG edges that are also dominator tree edges instead
752 // of catching them below.
753 if (SuccNode->getIDom() == Node)
756 unsigned SuccLevel = DomLevels[SuccNode];
757 if (SuccLevel > RootLevel)
760 if (!Visited.insert(SuccNode))
763 BasicBlock *SuccBB = SuccNode->getBlock();
764 if (!LiveInBlocks.count(SuccBB))
767 DFBlocks.push_back(std::make_pair(BBNumbers[SuccBB], SuccBB));
768 if (!DefBlocks.count(SuccBB))
769 PQ.push(std::make_pair(SuccNode, SuccLevel));
772 for (DomTreeNode::iterator CI = Node->begin(), CE = Node->end(); CI != CE;
774 if (!Visited.count(*CI))
775 Worklist.push_back(*CI);
780 if (DFBlocks.size() > 1)
781 std::sort(DFBlocks.begin(), DFBlocks.end());
783 unsigned CurrentVersion = 0;
784 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i)
785 QueuePhiNode(DFBlocks[i].second, AllocaNum, CurrentVersion);
788 /// RewriteSingleStoreAlloca - If there is only a single store to this value,
789 /// replace any loads of it that are directly dominated by the definition with
790 /// the value stored.
791 void PromoteMem2Reg::RewriteSingleStoreAlloca(AllocaInst *AI,
793 LargeBlockInfo &LBI) {
794 StoreInst *OnlyStore = Info.OnlyStore;
795 bool StoringGlobalVal = !isa<Instruction>(OnlyStore->getOperand(0));
796 BasicBlock *StoreBB = OnlyStore->getParent();
799 // Clear out UsingBlocks. We will reconstruct it here if needed.
800 Info.UsingBlocks.clear();
802 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E; ) {
803 Instruction *UserInst = cast<Instruction>(*UI++);
804 if (!isa<LoadInst>(UserInst)) {
805 assert(UserInst == OnlyStore && "Should only have load/stores");
808 LoadInst *LI = cast<LoadInst>(UserInst);
810 // Okay, if we have a load from the alloca, we want to replace it with the
811 // only value stored to the alloca. We can do this if the value is
812 // dominated by the store. If not, we use the rest of the mem2reg machinery
813 // to insert the phi nodes as needed.
814 if (!StoringGlobalVal) { // Non-instructions are always dominated.
815 if (LI->getParent() == StoreBB) {
816 // If we have a use that is in the same block as the store, compare the
817 // indices of the two instructions to see which one came first. If the
818 // load came before the store, we can't handle it.
819 if (StoreIndex == -1)
820 StoreIndex = LBI.getInstructionIndex(OnlyStore);
822 if (unsigned(StoreIndex) > LBI.getInstructionIndex(LI)) {
823 // Can't handle this load, bail out.
824 Info.UsingBlocks.push_back(StoreBB);
828 } else if (LI->getParent() != StoreBB &&
829 !dominates(StoreBB, LI->getParent())) {
830 // If the load and store are in different blocks, use BB dominance to
831 // check their relationships. If the store doesn't dom the use, bail
833 Info.UsingBlocks.push_back(LI->getParent());
838 // Otherwise, we *can* safely rewrite this load.
839 Value *ReplVal = OnlyStore->getOperand(0);
840 // If the replacement value is the load, this must occur in unreachable
843 ReplVal = UndefValue::get(LI->getType());
844 LI->replaceAllUsesWith(ReplVal);
845 if (AST && LI->getType()->isPointerTy())
846 AST->deleteValue(LI);
847 LI->eraseFromParent();
854 /// StoreIndexSearchPredicate - This is a helper predicate used to search by the
855 /// first element of a pair.
856 struct StoreIndexSearchPredicate {
857 bool operator()(const std::pair<unsigned, StoreInst*> &LHS,
858 const std::pair<unsigned, StoreInst*> &RHS) {
859 return LHS.first < RHS.first;
865 /// PromoteSingleBlockAlloca - Many allocas are only used within a single basic
866 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
867 /// potentially useless PHI nodes by just performing a single linear pass over
868 /// the basic block using the Alloca.
870 /// If we cannot promote this alloca (because it is read before it is written),
871 /// return true. This is necessary in cases where, due to control flow, the
872 /// alloca is potentially undefined on some control flow paths. e.g. code like
873 /// this is potentially correct:
875 /// for (...) { if (c) { A = undef; undef = B; } }
877 /// ... so long as A is not used before undef is set.
879 void PromoteMem2Reg::PromoteSingleBlockAlloca(AllocaInst *AI, AllocaInfo &Info,
880 LargeBlockInfo &LBI) {
881 // The trickiest case to handle is when we have large blocks. Because of this,
882 // this code is optimized assuming that large blocks happen. This does not
883 // significantly pessimize the small block case. This uses LargeBlockInfo to
884 // make it efficient to get the index of various operations in the block.
886 // Clear out UsingBlocks. We will reconstruct it here if needed.
887 Info.UsingBlocks.clear();
889 // Walk the use-def list of the alloca, getting the locations of all stores.
890 typedef SmallVector<std::pair<unsigned, StoreInst*>, 64> StoresByIndexTy;
891 StoresByIndexTy StoresByIndex;
893 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
895 if (StoreInst *SI = dyn_cast<StoreInst>(*UI))
896 StoresByIndex.push_back(std::make_pair(LBI.getInstructionIndex(SI), SI));
898 // If there are no stores to the alloca, just replace any loads with undef.
899 if (StoresByIndex.empty()) {
900 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;)
901 if (LoadInst *LI = dyn_cast<LoadInst>(*UI++)) {
902 LI->replaceAllUsesWith(UndefValue::get(LI->getType()));
903 if (AST && LI->getType()->isPointerTy())
904 AST->deleteValue(LI);
906 LI->eraseFromParent();
911 // Sort the stores by their index, making it efficient to do a lookup with a
913 std::sort(StoresByIndex.begin(), StoresByIndex.end());
915 // Walk all of the loads from this alloca, replacing them with the nearest
916 // store above them, if any.
917 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end(); UI != E;) {
918 LoadInst *LI = dyn_cast<LoadInst>(*UI++);
921 unsigned LoadIdx = LBI.getInstructionIndex(LI);
923 // Find the nearest store that has a lower than this load.
924 StoresByIndexTy::iterator I =
925 std::lower_bound(StoresByIndex.begin(), StoresByIndex.end(),
926 std::pair<unsigned, StoreInst*>(LoadIdx, static_cast<StoreInst*>(0)),
927 StoreIndexSearchPredicate());
929 // If there is no store before this load, then we can't promote this load.
930 if (I == StoresByIndex.begin()) {
931 // Can't handle this load, bail out.
932 Info.UsingBlocks.push_back(LI->getParent());
936 // Otherwise, there was a store before this load, the load takes its value.
938 LI->replaceAllUsesWith(I->second->getOperand(0));
939 if (AST && LI->getType()->isPointerTy())
940 AST->deleteValue(LI);
941 LI->eraseFromParent();
946 // Inserts a llvm.dbg.value instrinsic before the stores to an alloca'd value
947 // that has an associated llvm.dbg.decl intrinsic.
948 void PromoteMem2Reg::ConvertDebugDeclareToDebugValue(DbgDeclareInst *DDI,
950 DIVariable DIVar(DDI->getVariable());
955 DIF = new DIFactory(*SI->getParent()->getParent()->getParent());
956 Instruction *DbgVal = DIF->InsertDbgValueIntrinsic(SI->getOperand(0), 0,
959 // Propagate any debug metadata from the store onto the dbg.value.
960 DebugLoc SIDL = SI->getDebugLoc();
961 if (!SIDL.isUnknown())
962 DbgVal->setDebugLoc(SIDL);
963 // Otherwise propagate debug metadata from dbg.declare.
965 DbgVal->setDebugLoc(DDI->getDebugLoc());
968 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
969 // Alloca returns true if there wasn't already a phi-node for that variable
971 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
973 // Look up the basic-block in question.
974 PHINode *&PN = NewPhiNodes[std::make_pair(BB, AllocaNo)];
976 // If the BB already has a phi node added for the i'th alloca then we're done!
977 if (PN) return false;
979 // Create a PhiNode using the dereferenced type... and add the phi-node to the
981 PN = PHINode::Create(Allocas[AllocaNo]->getAllocatedType(),
982 Allocas[AllocaNo]->getName() + "." + Twine(Version++),
985 PhiToAllocaMap[PN] = AllocaNo;
986 PN->reserveOperandSpace(getNumPreds(BB));
988 if (AST && PN->getType()->isPointerTy())
989 AST->copyValue(PointerAllocaValues[AllocaNo], PN);
994 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
995 // stores to the allocas which we are promoting. IncomingVals indicates what
996 // value each Alloca contains on exit from the predecessor block Pred.
998 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
999 RenamePassData::ValVector &IncomingVals,
1000 std::vector<RenamePassData> &Worklist) {
1002 // If we are inserting any phi nodes into this BB, they will already be in the
1004 if (PHINode *APN = dyn_cast<PHINode>(BB->begin())) {
1005 // If we have PHI nodes to update, compute the number of edges from Pred to
1007 if (PhiToAllocaMap.count(APN)) {
1008 // We want to be able to distinguish between PHI nodes being inserted by
1009 // this invocation of mem2reg from those phi nodes that already existed in
1010 // the IR before mem2reg was run. We determine that APN is being inserted
1011 // because it is missing incoming edges. All other PHI nodes being
1012 // inserted by this pass of mem2reg will have the same number of incoming
1013 // operands so far. Remember this count.
1014 unsigned NewPHINumOperands = APN->getNumOperands();
1016 unsigned NumEdges = 0;
1017 for (succ_iterator I = succ_begin(Pred), E = succ_end(Pred); I != E; ++I)
1020 assert(NumEdges && "Must be at least one edge from Pred to BB!");
1022 // Add entries for all the phis.
1023 BasicBlock::iterator PNI = BB->begin();
1025 unsigned AllocaNo = PhiToAllocaMap[APN];
1027 // Add N incoming values to the PHI node.
1028 for (unsigned i = 0; i != NumEdges; ++i)
1029 APN->addIncoming(IncomingVals[AllocaNo], Pred);
1031 // The currently active variable for this block is now the PHI.
1032 IncomingVals[AllocaNo] = APN;
1034 // Get the next phi node.
1036 APN = dyn_cast<PHINode>(PNI);
1037 if (APN == 0) break;
1039 // Verify that it is missing entries. If not, it is not being inserted
1040 // by this mem2reg invocation so we want to ignore it.
1041 } while (APN->getNumOperands() == NewPHINumOperands);
1045 // Don't revisit blocks.
1046 if (!Visited.insert(BB)) return;
1048 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
1049 Instruction *I = II++; // get the instruction, increment iterator
1051 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1052 AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand());
1055 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
1056 if (AI == AllocaLookup.end()) continue;
1058 Value *V = IncomingVals[AI->second];
1060 // Anything using the load now uses the current value.
1061 LI->replaceAllUsesWith(V);
1062 if (AST && LI->getType()->isPointerTy())
1063 AST->deleteValue(LI);
1064 BB->getInstList().erase(LI);
1065 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
1066 // Delete this instruction and mark the name as the current holder of the
1068 AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand());
1069 if (!Dest) continue;
1071 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
1072 if (ai == AllocaLookup.end())
1075 // what value were we writing?
1076 IncomingVals[ai->second] = SI->getOperand(0);
1077 // Record debuginfo for the store before removing it.
1078 if (DbgDeclareInst *DDI = AllocaDbgDeclares[ai->second])
1079 ConvertDebugDeclareToDebugValue(DDI, SI);
1080 BB->getInstList().erase(SI);
1084 // 'Recurse' to our successors.
1085 succ_iterator I = succ_begin(BB), E = succ_end(BB);
1088 // Keep track of the successors so we don't visit the same successor twice
1089 SmallPtrSet<BasicBlock*, 8> VisitedSuccs;
1091 // Handle the first successor without using the worklist.
1092 VisitedSuccs.insert(*I);
1098 if (VisitedSuccs.insert(*I))
1099 Worklist.push_back(RenamePassData(*I, Pred, IncomingVals));
1104 /// PromoteMemToReg - Promote the specified list of alloca instructions into
1105 /// scalar registers, inserting PHI nodes as appropriate. This function does
1106 /// not modify the CFG of the function at all. All allocas must be from the
1109 /// If AST is specified, the specified tracker is updated to reflect changes
1112 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
1113 DominatorTree &DT, AliasSetTracker *AST) {
1114 // If there is nothing to do, bail out...
1115 if (Allocas.empty()) return;
1117 PromoteMem2Reg(Allocas, DT, AST).run();