1 //===- PromoteMemoryToRegister.cpp - Convert allocas to registers ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file promote memory references to be register references. It promotes
11 // alloca instructions which only have loads and stores as uses (or that have
12 // PHI nodes which are only loaded from). An alloca is transformed by using
13 // dominator frontiers to place PHI nodes, then traversing the function in
14 // depth-first order to rewrite loads and stores as appropriate. This is just
15 // the standard SSA construction algorithm to construct "pruned" SSA form.
17 //===----------------------------------------------------------------------===//
19 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
20 #include "llvm/Analysis/Dominators.h"
21 #include "llvm/iMemory.h"
22 #include "llvm/iPHINode.h"
23 #include "llvm/iOther.h"
24 #include "llvm/Function.h"
25 #include "llvm/Constant.h"
26 #include "llvm/Support/CFG.h"
27 #include "Support/StringExtras.h"
30 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
31 /// This is true if there are only loads and stores to the alloca... of if there
32 /// is a PHI node using the address which can be trivially transformed.
34 bool llvm::isAllocaPromotable(const AllocaInst *AI, const TargetData &TD) {
35 // FIXME: If the memory unit is of pointer or integer type, we can permit
36 // assignments to subsections of the memory unit.
38 // Only allow direct loads and stores...
39 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
40 UI != UE; ++UI) // Loop over all of the uses of the alloca
41 if (isa<LoadInst>(*UI)) {
43 } else if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
44 if (SI->getOperand(0) == AI)
45 return false; // Don't allow a store OF the AI, only INTO the AI.
46 } else if (const PHINode *PN = dyn_cast<PHINode>(*UI)) {
47 // We only support PHI nodes in a few simple cases. The PHI node is only
48 // allowed to have one use, which must be a load instruction, and can only
49 // use alloca instructions (no random pointers). Also, there cannot be
50 // any accesses to AI between the PHI node and the use of the PHI.
51 if (!PN->hasOneUse()) return false;
53 // Our transformation causes the unconditional loading of all pointer
54 // operands to the PHI node. Because this could cause a fault if there is
55 // a critical edge in the CFG and if one of the pointers is illegal, we
56 // refuse to promote PHI nodes unless they are obviously safe. For now,
57 // obviously safe means that all of the operands are allocas.
59 // If we wanted to extend this code to break critical edges, this
60 // restriction could be relaxed, and we could even handle uses of the PHI
61 // node that are volatile loads or stores.
63 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
64 if (!isa<AllocaInst>(PN->getIncomingValue(i)))
67 // Now make sure the one user instruction is in the same basic block as
68 // the PHI, and that there are no loads or stores between the PHI node and
70 BasicBlock::const_iterator UI = cast<Instruction>(PN->use_back());
71 if (!isa<LoadInst>(UI) || cast<LoadInst>(UI)->isVolatile()) return false;
73 // Scan looking for memory accesses.
74 // FIXME: this should REALLY use alias analysis.
75 for (--UI; !isa<PHINode>(UI); --UI)
76 if (isa<LoadInst>(UI) || isa<StoreInst>(UI) || isa<CallInst>(UI))
79 // If we got this far, we can promote the PHI use.
80 } else if (const SelectInst *SI = dyn_cast<SelectInst>(*UI)) {
81 // We only support selects in a few simple cases. The select is only
82 // allowed to have one use, which must be a load instruction, and can only
83 // use alloca instructions (no random pointers). Also, there cannot be
84 // any accesses to AI between the PHI node and the use of the PHI.
85 if (!SI->hasOneUse()) return false;
87 // Our transformation causes the unconditional loading of all pointer
88 // operands of the select. Because this could cause a fault if there is a
89 // critical edge in the CFG and if one of the pointers is illegal, we
90 // refuse to promote the select unless it is obviously safe. For now,
91 // obviously safe means that all of the operands are allocas.
93 if (!isa<AllocaInst>(SI->getOperand(1)) ||
94 !isa<AllocaInst>(SI->getOperand(2)))
97 // Now make sure the one user instruction is in the same basic block as
98 // the PHI, and that there are no loads or stores between the PHI node and
100 BasicBlock::const_iterator UI = cast<Instruction>(SI->use_back());
101 if (!isa<LoadInst>(UI) || cast<LoadInst>(UI)->isVolatile()) return false;
103 // Scan looking for memory accesses.
104 // FIXME: this should REALLY use alias analysis.
105 for (--UI; &*UI != SI; --UI)
106 if (isa<LoadInst>(UI) || isa<StoreInst>(UI) || isa<CallInst>(UI))
109 // If we got this far, we can promote the select use.
111 return false; // Not a load, store, or promotable PHI?
118 struct PromoteMem2Reg {
119 // Allocas - The alloca instructions being promoted
120 std::vector<AllocaInst*> Allocas;
122 DominanceFrontier &DF;
123 const TargetData &TD;
125 // AllocaLookup - Reverse mapping of Allocas
126 std::map<AllocaInst*, unsigned> AllocaLookup;
128 // NewPhiNodes - The PhiNodes we're adding.
129 std::map<BasicBlock*, std::vector<PHINode*> > NewPhiNodes;
131 // Visited - The set of basic blocks the renamer has already visited.
132 std::set<BasicBlock*> Visited;
134 // BasicBlockNumbering - Holds a numbering of the basic blocks in the
135 // function in a stable order that does not depend on their address.
136 std::map<BasicBlock*, unsigned> BasicBlockNumbering;
138 // NumberedBasicBlock - Holds the inverse mapping of BasicBlockNumbering.
139 std::vector<BasicBlock*> NumberedBasicBlock;
142 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
143 DominanceFrontier &df, const TargetData &td)
144 : Allocas(A), DT(dt), DF(df), TD(td) {}
149 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
150 std::set<PHINode*> &DeadPHINodes);
151 void PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI);
152 void PromoteLocallyUsedAllocas(BasicBlock *BB,
153 const std::vector<AllocaInst*> &AIs);
155 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
156 std::vector<Value*> &IncVals);
157 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
158 std::set<PHINode*> &InsertedPHINodes);
160 } // end of anonymous namespace
162 void PromoteMem2Reg::run() {
163 Function &F = *DF.getRoot()->getParent();
165 // LocallyUsedAllocas - Keep track of all of the alloca instructions which are
166 // only used in a single basic block. These instructions can be efficiently
167 // promoted by performing a single linear scan over that one block. Since
168 // individual basic blocks are sometimes large, we group together all allocas
169 // that are live in a single basic block by the basic block they are live in.
170 std::map<BasicBlock*, std::vector<AllocaInst*> > LocallyUsedAllocas;
173 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
174 AllocaInst *AI = Allocas[AllocaNum];
176 assert(isAllocaPromotable(AI, TD) &&
177 "Cannot promote non-promotable alloca!");
178 assert(AI->getParent()->getParent() == &F &&
179 "All allocas should be in the same function, which is same as DF!");
181 if (AI->use_empty()) {
182 // If there are no uses of the alloca, just delete it now.
183 AI->getParent()->getInstList().erase(AI);
185 // Remove the alloca from the Allocas list, since it has been processed
186 Allocas[AllocaNum] = Allocas.back();
192 // Calculate the set of read and write-locations for each alloca. This is
193 // analogous to finding the 'uses' and 'definitions' of each variable.
194 std::vector<BasicBlock*> DefiningBlocks;
195 std::vector<BasicBlock*> UsingBlocks;
197 BasicBlock *OnlyBlock = 0;
198 bool OnlyUsedInOneBlock = true;
200 // As we scan the uses of the alloca instruction, keep track of stores, and
201 // decide whether all of the loads and stores to the alloca are within the
204 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E;++U){
205 Instruction *User = cast<Instruction>(*U);
206 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
207 // Remember the basic blocks which define new values for the alloca
208 DefiningBlocks.push_back(SI->getParent());
209 } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
210 // Otherwise it must be a load instruction, keep track of variable reads
211 UsingBlocks.push_back(LI->getParent());
212 } else if (SelectInst *SI = dyn_cast<SelectInst>(User)) {
213 // Because of the restrictions we placed on Select instruction uses
214 // above things are very simple. Transform the PHI of addresses into a
215 // select of loaded values.
216 LoadInst *Load = cast<LoadInst>(SI->use_back());
217 std::string LoadName = Load->getName(); Load->setName("");
219 Value *TrueVal = new LoadInst(SI->getOperand(1),
220 SI->getOperand(1)->getName()+".val", SI);
221 Value *FalseVal = new LoadInst(SI->getOperand(2),
222 SI->getOperand(2)->getName()+".val", SI);
224 Value *NewSI = new SelectInst(SI->getOperand(0), TrueVal,
225 FalseVal, Load->getName(), SI);
226 Load->replaceAllUsesWith(NewSI);
227 Load->getParent()->getInstList().erase(Load);
228 SI->getParent()->getInstList().erase(SI);
230 // Restart our scan of uses...
231 DefiningBlocks.clear();
235 // Because of the restrictions we placed on PHI node uses above, the PHI
236 // node reads the block in any using predecessors. Transform the PHI of
237 // addresses into a PHI of loaded values.
238 PHINode *PN = cast<PHINode>(User);
239 assert(PN->hasOneUse() && "Cannot handle PHI Node with != 1 use!");
240 LoadInst *PNUser = cast<LoadInst>(PN->use_back());
241 std::string PNUserName = PNUser->getName(); PNUser->setName("");
243 // Create the new PHI node and insert load instructions as appropriate.
244 PHINode *NewPN = new PHINode(AI->getAllocatedType(), PNUserName, PN);
245 std::map<BasicBlock*, LoadInst*> NewLoads;
246 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
247 BasicBlock *Pred = PN->getIncomingBlock(i);
248 LoadInst *&NewLoad = NewLoads[Pred];
249 if (NewLoad == 0) // Insert the new load in the predecessor
250 NewLoad = new LoadInst(PN->getIncomingValue(i),
251 PN->getIncomingValue(i)->getName()+".val",
252 Pred->getTerminator());
253 NewPN->addIncoming(NewLoad, Pred);
256 // Remove the old load.
257 PNUser->replaceAllUsesWith(NewPN);
258 PNUser->getParent()->getInstList().erase(PNUser);
260 // Remove the old PHI node.
261 PN->getParent()->getInstList().erase(PN);
263 // Restart our scan of uses...
264 DefiningBlocks.clear();
269 if (OnlyUsedInOneBlock) {
271 OnlyBlock = User->getParent();
272 else if (OnlyBlock != User->getParent())
273 OnlyUsedInOneBlock = false;
277 // If the alloca is only read and written in one basic block, just perform a
278 // linear sweep over the block to eliminate it.
279 if (OnlyUsedInOneBlock) {
280 LocallyUsedAllocas[OnlyBlock].push_back(AI);
282 // Remove the alloca from the Allocas list, since it will be processed.
283 Allocas[AllocaNum] = Allocas.back();
289 // If we haven't computed a numbering for the BB's in the function, do so
291 if (NumberedBasicBlock.empty()) {
293 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I, ++n) {
294 NumberedBasicBlock.push_back(I);
295 BasicBlockNumbering[I] = n;
299 // Compute the locations where PhiNodes need to be inserted. Look at the
300 // dominance frontier of EACH basic-block we have a write in.
302 unsigned CurrentVersion = 0;
303 std::set<PHINode*> InsertedPHINodes;
304 std::vector<unsigned> DFBlocks;
305 while (!DefiningBlocks.empty()) {
306 BasicBlock *BB = DefiningBlocks.back();
307 DefiningBlocks.pop_back();
309 // Look up the DF for this write, add it to PhiNodes
310 DominanceFrontier::const_iterator it = DF.find(BB);
311 if (it != DF.end()) {
312 const DominanceFrontier::DomSetType &S = it->second;
314 // In theory we don't need the indirection through the DFBlocks vector.
315 // In practice, the order of calling QueuePhiNode would depend on the
316 // (unspecified) ordering of basic blocks in the dominance frontier,
317 // which would give PHI nodes non-determinstic subscripts. Fix this by
318 // processing blocks in order of the occurance in the function.
319 for (DominanceFrontier::DomSetType::iterator P = S.begin(),PE = S.end();
321 DFBlocks.push_back(BasicBlockNumbering[*P]);
323 // Sort by which the block ordering in the function.
324 std::sort(DFBlocks.begin(), DFBlocks.end());
326 for (unsigned i = 0, e = DFBlocks.size(); i != e; ++i) {
327 BasicBlock *BB = NumberedBasicBlock[DFBlocks[i]];
328 if (QueuePhiNode(BB, AllocaNum, CurrentVersion, InsertedPHINodes))
329 DefiningBlocks.push_back(BB);
335 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier
336 // of the writes to the variable, scan through the reads of the variable,
337 // marking PHI nodes which are actually necessary as alive (by removing them
338 // from the InsertedPHINodes set). This is not perfect: there may PHI
339 // marked alive because of loads which are dominated by stores, but there
340 // will be no unmarked PHI nodes which are actually used.
342 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
343 MarkDominatingPHILive(UsingBlocks[i], AllocaNum, InsertedPHINodes);
346 // If there are any PHI nodes which are now known to be dead, remove them!
347 for (std::set<PHINode*>::iterator I = InsertedPHINodes.begin(),
348 E = InsertedPHINodes.end(); I != E; ++I) {
350 std::vector<PHINode*> &BBPNs = NewPhiNodes[PN->getParent()];
351 BBPNs[AllocaNum] = 0;
353 // Check to see if we just removed the last inserted PHI node from this
354 // basic block. If so, remove the entry for the basic block.
355 bool HasOtherPHIs = false;
356 for (unsigned i = 0, e = BBPNs.size(); i != e; ++i)
362 NewPhiNodes.erase(PN->getParent());
364 PN->getParent()->getInstList().erase(PN);
367 // Keep the reverse mapping of the 'Allocas' array.
368 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
371 // Process all allocas which are only used in a single basic block.
372 for (std::map<BasicBlock*, std::vector<AllocaInst*> >::iterator I =
373 LocallyUsedAllocas.begin(), E = LocallyUsedAllocas.end(); I != E; ++I){
374 const std::vector<AllocaInst*> &Allocas = I->second;
375 assert(!Allocas.empty() && "empty alloca list??");
377 // It's common for there to only be one alloca in the list. Handle it
379 if (Allocas.size() == 1)
380 PromoteLocallyUsedAlloca(I->first, Allocas[0]);
382 PromoteLocallyUsedAllocas(I->first, Allocas);
386 return; // All of the allocas must have been trivial!
388 // Set the incoming values for the basic block to be null values for all of
389 // the alloca's. We do this in case there is a load of a value that has not
390 // been stored yet. In this case, it will get this null value.
392 std::vector<Value *> Values(Allocas.size());
393 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
394 Values[i] = Constant::getNullValue(Allocas[i]->getAllocatedType());
396 // Walks all basic blocks in the function performing the SSA rename algorithm
397 // and inserting the phi nodes we marked as necessary
399 RenamePass(F.begin(), 0, Values);
401 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
404 // Remove the allocas themselves from the function...
405 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
406 Instruction *A = Allocas[i];
408 // If there are any uses of the alloca instructions left, they must be in
409 // sections of dead code that were not processed on the dominance frontier.
410 // Just delete the users now.
413 A->replaceAllUsesWith(Constant::getNullValue(A->getType()));
414 A->getParent()->getInstList().erase(A);
417 // At this point, the renamer has added entries to PHI nodes for all reachable
418 // code. Unfortunately, there may be blocks which are not reachable, which
419 // the renamer hasn't traversed. If this is the case, the PHI nodes may not
420 // have incoming values for all predecessors. Loop over all PHI nodes we have
421 // created, inserting null constants if they are missing any incoming values.
423 for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I =
424 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
426 std::vector<BasicBlock*> Preds(pred_begin(I->first), pred_end(I->first));
427 std::vector<PHINode*> &PNs = I->second;
428 assert(!PNs.empty() && "Empty PHI node list??");
430 // Only do work here if there the PHI nodes are missing incoming values. We
431 // know that all PHI nodes that were inserted in a block will have the same
432 // number of incoming values, so we can just check any PHI node.
434 for (unsigned i = 0; (FirstPHI = PNs[i]) == 0; ++i)
437 if (Preds.size() != FirstPHI->getNumIncomingValues()) {
438 // Ok, now we know that all of the PHI nodes are missing entries for some
439 // basic blocks. Start by sorting the incoming predecessors for efficient
441 std::sort(Preds.begin(), Preds.end());
443 // Now we loop through all BB's which have entries in FirstPHI and remove
444 // them from the Preds list.
445 for (unsigned i = 0, e = FirstPHI->getNumIncomingValues(); i != e; ++i) {
446 // Do a log(n) search of the Preds list for the entry we want.
447 std::vector<BasicBlock*>::iterator EntIt =
448 std::lower_bound(Preds.begin(), Preds.end(),
449 FirstPHI->getIncomingBlock(i));
450 assert(EntIt != Preds.end() && *EntIt == FirstPHI->getIncomingBlock(i)&&
451 "PHI node has entry for a block which is not a predecessor!");
457 // At this point, the blocks left in the preds list must have dummy
458 // entries inserted into every PHI nodes for the block.
459 for (unsigned i = 0, e = PNs.size(); i != e; ++i)
460 if (PHINode *PN = PNs[i]) {
461 Value *NullVal = Constant::getNullValue(PN->getType());
462 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
463 PN->addIncoming(NullVal, Preds[pred]);
469 // MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not
470 // "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF
471 // as usual (inserting the PHI nodes in the DeadPHINodes set), then processes
472 // each read of the variable. For each block that reads the variable, this
473 // function is called, which removes used PHI nodes from the DeadPHINodes set.
474 // After all of the reads have been processed, any PHI nodes left in the
475 // DeadPHINodes set are removed.
477 void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
478 std::set<PHINode*> &DeadPHINodes) {
479 // Scan the immediate dominators of this block looking for a block which has a
480 // PHI node for Alloca num. If we find it, mark the PHI node as being alive!
481 for (DominatorTree::Node *N = DT[BB]; N; N = N->getIDom()) {
482 BasicBlock *DomBB = N->getBlock();
483 std::map<BasicBlock*, std::vector<PHINode*> >::iterator
484 I = NewPhiNodes.find(DomBB);
485 if (I != NewPhiNodes.end() && I->second[AllocaNum]) {
486 // Ok, we found an inserted PHI node which dominates this value.
487 PHINode *DominatingPHI = I->second[AllocaNum];
489 // Find out if we previously thought it was dead.
490 std::set<PHINode*>::iterator DPNI = DeadPHINodes.find(DominatingPHI);
491 if (DPNI != DeadPHINodes.end()) {
492 // Ok, until now, we thought this PHI node was dead. Mark it as being
494 DeadPHINodes.erase(DPNI);
496 // Now that we have marked the PHI node alive, also mark any PHI nodes
497 // which it might use as being alive as well.
498 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB);
500 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes);
506 /// PromoteLocallyUsedAlloca - Many allocas are only used within a single basic
507 /// block. If this is the case, avoid traversing the CFG and inserting a lot of
508 /// potentially useless PHI nodes by just performing a single linear pass over
509 /// the basic block using the Alloca.
511 void PromoteMem2Reg::PromoteLocallyUsedAlloca(BasicBlock *BB, AllocaInst *AI) {
512 assert(!AI->use_empty() && "There are no uses of the alloca!");
514 // Handle degenerate cases quickly.
515 if (AI->hasOneUse()) {
516 Instruction *U = cast<Instruction>(AI->use_back());
517 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
518 // Must be a load of uninitialized value.
519 LI->replaceAllUsesWith(Constant::getNullValue(AI->getAllocatedType()));
521 // Otherwise it must be a store which is never read.
522 assert(isa<StoreInst>(U));
524 BB->getInstList().erase(U);
526 // Uses of the uninitialized memory location shall get zero...
527 Value *CurVal = Constant::getNullValue(AI->getAllocatedType());
529 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
530 Instruction *Inst = I++;
531 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
532 if (LI->getOperand(0) == AI) {
533 // Loads just returns the "current value"...
534 LI->replaceAllUsesWith(CurVal);
535 BB->getInstList().erase(LI);
537 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
538 if (SI->getOperand(1) == AI) {
539 // Store updates the "current value"...
540 CurVal = SI->getOperand(0);
541 BB->getInstList().erase(SI);
547 // After traversing the basic block, there should be no more uses of the
548 // alloca, remove it now.
549 assert(AI->use_empty() && "Uses of alloca from more than one BB??");
550 AI->getParent()->getInstList().erase(AI);
553 /// PromoteLocallyUsedAllocas - This method is just like
554 /// PromoteLocallyUsedAlloca, except that it processes multiple alloca
555 /// instructions in parallel. This is important in cases where we have large
556 /// basic blocks, as we don't want to rescan the entire basic block for each
557 /// alloca which is locally used in it (which might be a lot).
558 void PromoteMem2Reg::
559 PromoteLocallyUsedAllocas(BasicBlock *BB, const std::vector<AllocaInst*> &AIs) {
560 std::map<AllocaInst*, Value*> CurValues;
561 for (unsigned i = 0, e = AIs.size(); i != e; ++i)
562 CurValues[AIs[i]] = 0; // Insert with null value
564 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
565 Instruction *Inst = I++;
566 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
567 // Is this a load of an alloca we are tracking?
568 if (AllocaInst *AI = dyn_cast<AllocaInst>(LI->getOperand(0))) {
569 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
570 if (AIt != CurValues.end()) {
571 // Loads just returns the "current value"...
572 if (AIt->second == 0) // Uninitialized value??
573 AIt->second =Constant::getNullValue(AIt->first->getAllocatedType());
574 LI->replaceAllUsesWith(AIt->second);
575 BB->getInstList().erase(LI);
578 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
579 if (AllocaInst *AI = dyn_cast<AllocaInst>(SI->getOperand(1))) {
580 std::map<AllocaInst*, Value*>::iterator AIt = CurValues.find(AI);
581 if (AIt != CurValues.end()) {
582 // Store updates the "current value"...
583 AIt->second = SI->getOperand(0);
584 BB->getInstList().erase(SI);
593 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
594 // Alloca returns true if there wasn't already a phi-node for that variable
596 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
598 std::set<PHINode*> &InsertedPHINodes) {
599 // Look up the basic-block in question
600 std::vector<PHINode*> &BBPNs = NewPhiNodes[BB];
601 if (BBPNs.empty()) BBPNs.resize(Allocas.size());
603 // If the BB already has a phi node added for the i'th alloca then we're done!
604 if (BBPNs[AllocaNo]) return false;
606 // Create a PhiNode using the dereferenced type... and add the phi-node to the
608 BBPNs[AllocaNo] = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
609 Allocas[AllocaNo]->getName() + "." +
610 utostr(Version++), BB->begin());
611 InsertedPHINodes.insert(BBPNs[AllocaNo]);
616 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
617 // stores to the allocas which we are promoting. IncomingVals indicates what
618 // value each Alloca contains on exit from the predecessor block Pred.
620 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
621 std::vector<Value*> &IncomingVals) {
623 // If this BB needs a PHI node, update the PHI node for each variable we need
625 std::map<BasicBlock*, std::vector<PHINode *> >::iterator
626 BBPNI = NewPhiNodes.find(BB);
627 if (BBPNI != NewPhiNodes.end()) {
628 std::vector<PHINode *> &BBPNs = BBPNI->second;
629 for (unsigned k = 0; k != BBPNs.size(); ++k)
630 if (PHINode *PN = BBPNs[k]) {
631 // Add this incoming value to the PHI node.
632 PN->addIncoming(IncomingVals[k], Pred);
634 // The currently active variable for this block is now the PHI.
635 IncomingVals[k] = PN;
639 // don't revisit nodes
640 if (Visited.count(BB)) return;
645 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
646 Instruction *I = II++; // get the instruction, increment iterator
648 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
649 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) {
650 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
651 if (AI != AllocaLookup.end()) {
652 Value *V = IncomingVals[AI->second];
654 // walk the use list of this load and replace all uses with r
655 LI->replaceAllUsesWith(V);
656 BB->getInstList().erase(LI);
659 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
660 // Delete this instruction and mark the name as the current holder of the
662 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) {
663 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
664 if (ai != AllocaLookup.end()) {
665 // what value were we writing?
666 IncomingVals[ai->second] = SI->getOperand(0);
667 BB->getInstList().erase(SI);
673 // Recurse to our successors.
674 TerminatorInst *TI = BB->getTerminator();
675 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) {
676 std::vector<Value*> OutgoingVals(IncomingVals);
677 RenamePass(TI->getSuccessor(i), BB, OutgoingVals);
681 /// PromoteMemToReg - Promote the specified list of alloca instructions into
682 /// scalar registers, inserting PHI nodes as appropriate. This function makes
683 /// use of DominanceFrontier information. This function does not modify the CFG
684 /// of the function at all. All allocas must be from the same function.
686 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
687 DominatorTree &DT, DominanceFrontier &DF,
688 const TargetData &TD) {
689 // If there is nothing to do, bail out...
690 if (Allocas.empty()) return;
691 PromoteMem2Reg(Allocas, DT, DF, TD).run();