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. An alloca is
12 // transformed by using dominator frontiers to place PHI nodes, then traversing
13 // the function in depth-first order to rewrite loads and stores as appropriate.
14 // This is just the standard SSA construction algorithm to construct "pruned"
17 //===----------------------------------------------------------------------===//
19 #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/Function.h"
24 #include "llvm/Constant.h"
25 #include "llvm/Support/CFG.h"
26 #include "Support/StringExtras.h"
29 /// isAllocaPromotable - Return true if this alloca is legal for promotion.
30 /// This is true if there are only loads and stores to the alloca...
32 bool llvm::isAllocaPromotable(const AllocaInst *AI, const TargetData &TD) {
33 // FIXME: If the memory unit is of pointer or integer type, we can permit
34 // assignments to subsections of the memory unit.
36 // Only allow direct loads and stores...
37 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
38 UI != UE; ++UI) // Loop over all of the uses of the alloca
39 if (!isa<LoadInst>(*UI))
40 if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
41 if (SI->getOperand(0) == AI)
42 return false; // Don't allow a store of the AI, only INTO the AI.
44 return false; // Not a load or store?
51 struct PromoteMem2Reg {
52 // Allocas - The alloca instructions being promoted
53 std::vector<AllocaInst*> Allocas;
55 DominanceFrontier &DF;
58 // AllocaLookup - Reverse mapping of Allocas
59 std::map<AllocaInst*, unsigned> AllocaLookup;
61 // NewPhiNodes - The PhiNodes we're adding.
62 std::map<BasicBlock*, std::vector<PHINode*> > NewPhiNodes;
64 // Visited - The set of basic blocks the renamer has already visited.
65 std::set<BasicBlock*> Visited;
68 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
69 DominanceFrontier &df, const TargetData &td)
70 : Allocas(A), DT(dt), DF(df), TD(td) {}
75 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
76 std::set<PHINode*> &DeadPHINodes);
77 void PromoteLocallyUsedAlloca(AllocaInst *AI);
79 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
80 std::vector<Value*> &IncVals);
81 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
82 std::set<PHINode*> &InsertedPHINodes);
84 } // end of anonymous namespace
86 void PromoteMem2Reg::run() {
87 Function &F = *DF.getRoot()->getParent();
89 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
90 AllocaInst *AI = Allocas[AllocaNum];
92 assert(isAllocaPromotable(AI, TD) &&
93 "Cannot promote non-promotable alloca!");
94 assert(AI->getParent()->getParent() == &F &&
95 "All allocas should be in the same function, which is same as DF!");
97 if (AI->use_empty()) {
98 // If there are no uses of the alloca, just delete it now.
99 AI->getParent()->getInstList().erase(AI);
101 // Remove the alloca from the Allocas list, since it has been processed
102 Allocas[AllocaNum] = Allocas.back();
108 // Calculate the set of read and write-locations for each alloca. This is
109 // analogous to counting the number of 'uses' and 'definitions' of each
111 std::vector<BasicBlock*> DefiningBlocks;
112 std::vector<BasicBlock*> UsingBlocks;
114 BasicBlock *OnlyBlock = 0;
115 bool OnlyUsedInOneBlock = true;
117 // As we scan the uses of the alloca instruction, keep track of stores, and
118 // decide whether all of the loads and stores to the alloca are within the
120 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E;++U){
121 Instruction *User = cast<Instruction>(*U);
122 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
123 // Remember the basic blocks which define new values for the alloca
124 DefiningBlocks.push_back(SI->getParent());
126 // Otherwise it must be a load instruction, keep track of variable reads
127 UsingBlocks.push_back(cast<LoadInst>(User)->getParent());
130 if (OnlyUsedInOneBlock) {
132 OnlyBlock = User->getParent();
133 else if (OnlyBlock != User->getParent())
134 OnlyUsedInOneBlock = false;
138 // If the alloca is only read and written in one basic block, just perform a
139 // linear sweep over the block to eliminate it.
140 if (OnlyUsedInOneBlock) {
141 PromoteLocallyUsedAlloca(AI);
143 // Remove the alloca from the Allocas list, since it has been processed
144 Allocas[AllocaNum] = Allocas.back();
150 // Compute the locations where PhiNodes need to be inserted. Look at the
151 // dominance frontier of EACH basic-block we have a write in.
153 unsigned CurrentVersion = 0;
154 std::set<PHINode*> InsertedPHINodes;
155 while (!DefiningBlocks.empty()) {
156 BasicBlock *BB = DefiningBlocks.back();
157 DefiningBlocks.pop_back();
159 // Look up the DF for this write, add it to PhiNodes
160 DominanceFrontier::const_iterator it = DF.find(BB);
161 if (it != DF.end()) {
162 const DominanceFrontier::DomSetType &S = it->second;
163 for (DominanceFrontier::DomSetType::iterator P = S.begin(),PE = S.end();
165 if (QueuePhiNode(*P, AllocaNum, CurrentVersion, InsertedPHINodes))
166 DefiningBlocks.push_back(*P);
170 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier
171 // of the writes to the variable, scan through the reads of the variable,
172 // marking PHI nodes which are actually necessary as alive (by removing them
173 // from the InsertedPHINodes set). This is not perfect: there may PHI
174 // marked alive because of loads which are dominated by stores, but there
175 // will be no unmarked PHI nodes which are actually used.
177 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
178 MarkDominatingPHILive(UsingBlocks[i], AllocaNum, InsertedPHINodes);
181 // If there are any PHI nodes which are now known to be dead, remove them!
182 for (std::set<PHINode*>::iterator I = InsertedPHINodes.begin(),
183 E = InsertedPHINodes.end(); I != E; ++I) {
185 std::vector<PHINode*> &BBPNs = NewPhiNodes[PN->getParent()];
186 BBPNs[AllocaNum] = 0;
188 // Check to see if we just removed the last inserted PHI node from this
189 // basic block. If so, remove the entry for the basic block.
190 bool HasOtherPHIs = false;
191 for (unsigned i = 0, e = BBPNs.size(); i != e; ++i)
197 NewPhiNodes.erase(PN->getParent());
199 PN->getParent()->getInstList().erase(PN);
202 // Keep the reverse mapping of the 'Allocas' array.
203 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
207 return; // All of the allocas must have been trivial!
209 // Set the incoming values for the basic block to be null values for all of
210 // the alloca's. We do this in case there is a load of a value that has not
211 // been stored yet. In this case, it will get this null value.
213 std::vector<Value *> Values(Allocas.size());
214 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
215 Values[i] = Constant::getNullValue(Allocas[i]->getAllocatedType());
217 // Walks all basic blocks in the function performing the SSA rename algorithm
218 // and inserting the phi nodes we marked as necessary
220 RenamePass(F.begin(), 0, Values);
222 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
225 // Remove the allocas themselves from the function...
226 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
227 Instruction *A = Allocas[i];
229 // If there are any uses of the alloca instructions left, they must be in
230 // sections of dead code that were not processed on the dominance frontier.
231 // Just delete the users now.
234 A->replaceAllUsesWith(Constant::getNullValue(A->getType()));
235 A->getParent()->getInstList().erase(A);
238 // At this point, the renamer has added entries to PHI nodes for all reachable
239 // code. Unfortunately, there may be blocks which are not reachable, which
240 // the renamer hasn't traversed. If this is the case, the PHI nodes may not
241 // have incoming values for all predecessors. Loop over all PHI nodes we have
242 // created, inserting null constants if they are missing any incoming values.
244 for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I =
245 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
247 std::vector<BasicBlock*> Preds(pred_begin(I->first), pred_end(I->first));
248 std::vector<PHINode*> &PNs = I->second;
249 assert(!PNs.empty() && "Empty PHI node list??");
251 // Only do work here if there the PHI nodes are missing incoming values. We
252 // know that all PHI nodes that were inserted in a block will have the same
253 // number of incoming values, so we can just check any PHI node.
255 for (unsigned i = 0; (FirstPHI = PNs[i]) == 0; ++i)
258 if (Preds.size() != FirstPHI->getNumIncomingValues()) {
259 // Ok, now we know that all of the PHI nodes are missing entries for some
260 // basic blocks. Start by sorting the incoming predecessors for efficient
262 std::sort(Preds.begin(), Preds.end());
264 // Now we loop through all BB's which have entries in FirstPHI and remove
265 // them from the Preds list.
266 for (unsigned i = 0, e = FirstPHI->getNumIncomingValues(); i != e; ++i) {
267 // Do a log(n) search of the Preds list for the entry we want.
268 std::vector<BasicBlock*>::iterator EntIt =
269 std::lower_bound(Preds.begin(), Preds.end(),
270 FirstPHI->getIncomingBlock(i));
271 assert(EntIt != Preds.end() && *EntIt == FirstPHI->getIncomingBlock(i)&&
272 "PHI node has entry for a block which is not a predecessor!");
278 // At this point, the blocks left in the preds list must have dummy
279 // entries inserted into every PHI nodes for the block.
280 for (unsigned i = 0, e = PNs.size(); i != e; ++i)
281 if (PHINode *PN = PNs[i]) {
282 Value *NullVal = Constant::getNullValue(PN->getType());
283 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
284 PN->addIncoming(NullVal, Preds[pred]);
290 // MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not
291 // "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF
292 // as usual (inserting the PHI nodes in the DeadPHINodes set), then processes
293 // each read of the variable. For each block that reads the variable, this
294 // function is called, which removes used PHI nodes from the DeadPHINodes set.
295 // After all of the reads have been processed, any PHI nodes left in the
296 // DeadPHINodes set are removed.
298 void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
299 std::set<PHINode*> &DeadPHINodes) {
300 // Scan the immediate dominators of this block looking for a block which has a
301 // PHI node for Alloca num. If we find it, mark the PHI node as being alive!
302 for (DominatorTree::Node *N = DT[BB]; N; N = N->getIDom()) {
303 BasicBlock *DomBB = N->getBlock();
304 std::map<BasicBlock*, std::vector<PHINode*> >::iterator
305 I = NewPhiNodes.find(DomBB);
306 if (I != NewPhiNodes.end() && I->second[AllocaNum]) {
307 // Ok, we found an inserted PHI node which dominates this value.
308 PHINode *DominatingPHI = I->second[AllocaNum];
310 // Find out if we previously thought it was dead.
311 std::set<PHINode*>::iterator DPNI = DeadPHINodes.find(DominatingPHI);
312 if (DPNI != DeadPHINodes.end()) {
313 // Ok, until now, we thought this PHI node was dead. Mark it as being
315 DeadPHINodes.erase(DPNI);
317 // Now that we have marked the PHI node alive, also mark any PHI nodes
318 // which it might use as being alive as well.
319 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB);
321 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes);
327 // PromoteLocallyUsedAlloca - Many allocas are only used within a single basic
328 // block. If this is the case, avoid traversing the CFG and inserting a lot of
329 // potentially useless PHI nodes by just performing a single linear pass over
330 // the basic block using the Alloca.
332 void PromoteMem2Reg::PromoteLocallyUsedAlloca(AllocaInst *AI) {
333 assert(!AI->use_empty() && "There are no uses of the alloca!");
335 // Uses of the uninitialized memory location shall get zero...
336 Value *CurVal = Constant::getNullValue(AI->getAllocatedType());
338 BasicBlock *BB = cast<Instruction>(AI->use_back())->getParent();
340 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
341 Instruction *Inst = I++;
342 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
343 if (LI->getOperand(0) == AI) {
344 // Loads just return the "current value"...
345 LI->replaceAllUsesWith(CurVal);
346 BB->getInstList().erase(LI);
348 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
349 if (SI->getOperand(1) == AI) {
350 // Loads just update the "current value"...
351 CurVal = SI->getOperand(0);
352 BB->getInstList().erase(SI);
357 // After traversing the basic block, there should be no more uses of the
358 // alloca, remove it now.
359 assert(AI->use_empty() && "Uses of alloca from more than one BB??");
360 AI->getParent()->getInstList().erase(AI);
363 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
364 // Alloca returns true if there wasn't already a phi-node for that variable
366 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
368 std::set<PHINode*> &InsertedPHINodes) {
369 // Look up the basic-block in question
370 std::vector<PHINode*> &BBPNs = NewPhiNodes[BB];
371 if (BBPNs.empty()) BBPNs.resize(Allocas.size());
373 // If the BB already has a phi node added for the i'th alloca then we're done!
374 if (BBPNs[AllocaNo]) return false;
376 // Create a PhiNode using the dereferenced type... and add the phi-node to the
378 BBPNs[AllocaNo] = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
379 Allocas[AllocaNo]->getName() + "." +
380 utostr(Version++), BB->begin());
381 InsertedPHINodes.insert(BBPNs[AllocaNo]);
386 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
387 // stores to the allocas which we are promoting. IncomingVals indicates what
388 // value each Alloca contains on exit from the predecessor block Pred.
390 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
391 std::vector<Value*> &IncomingVals) {
393 // If this BB needs a PHI node, update the PHI node for each variable we need
395 std::map<BasicBlock*, std::vector<PHINode *> >::iterator
396 BBPNI = NewPhiNodes.find(BB);
397 if (BBPNI != NewPhiNodes.end()) {
398 std::vector<PHINode *> &BBPNs = BBPNI->second;
399 for (unsigned k = 0; k != BBPNs.size(); ++k)
400 if (PHINode *PN = BBPNs[k]) {
401 // Add this incoming value to the PHI node.
402 PN->addIncoming(IncomingVals[k], Pred);
404 // The currently active variable for this block is now the PHI.
405 IncomingVals[k] = PN;
409 // don't revisit nodes
410 if (Visited.count(BB)) return;
415 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
416 Instruction *I = II++; // get the instruction, increment iterator
418 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
419 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) {
420 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
421 if (AI != AllocaLookup.end()) {
422 Value *V = IncomingVals[AI->second];
424 // walk the use list of this load and replace all uses with r
425 LI->replaceAllUsesWith(V);
426 BB->getInstList().erase(LI);
429 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
430 // Delete this instruction and mark the name as the current holder of the
432 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) {
433 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
434 if (ai != AllocaLookup.end()) {
435 // what value were we writing?
436 IncomingVals[ai->second] = SI->getOperand(0);
437 BB->getInstList().erase(SI);
443 // Recurse to our successors.
444 TerminatorInst *TI = BB->getTerminator();
445 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) {
446 std::vector<Value*> OutgoingVals(IncomingVals);
447 RenamePass(TI->getSuccessor(i), BB, OutgoingVals);
451 /// PromoteMemToReg - Promote the specified list of alloca instructions into
452 /// scalar registers, inserting PHI nodes as appropriate. This function makes
453 /// use of DominanceFrontier information. This function does not modify the CFG
454 /// of the function at all. All allocas must be from the same function.
456 void llvm::PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
457 DominatorTree &DT, DominanceFrontier &DF,
458 const TargetData &TD) {
459 // If there is nothing to do, bail out...
460 if (Allocas.empty()) return;
461 PromoteMem2Reg(Allocas, DT, DF, TD).run();