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"
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...
33 bool isAllocaPromotable(const AllocaInst *AI, const TargetData &TD) {
34 // FIXME: If the memory unit is of pointer or integer type, we can permit
35 // assignments to subsections of the memory unit.
37 // Only allow direct loads and stores...
38 for (Value::use_const_iterator UI = AI->use_begin(), UE = AI->use_end();
39 UI != UE; ++UI) // Loop over all of the uses of the alloca
40 if (!isa<LoadInst>(*UI))
41 if (const StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
42 if (SI->getOperand(0) == AI)
43 return false; // Don't allow a store of the AI, only INTO the AI.
45 return false; // Not a load or store?
52 struct PromoteMem2Reg {
53 // Allocas - The alloca instructions being promoted
54 std::vector<AllocaInst*> Allocas;
56 DominanceFrontier &DF;
59 // AllocaLookup - Reverse mapping of Allocas
60 std::map<AllocaInst*, unsigned> AllocaLookup;
62 // NewPhiNodes - The PhiNodes we're adding.
63 std::map<BasicBlock*, std::vector<PHINode*> > NewPhiNodes;
65 // Visited - The set of basic blocks the renamer has already visited.
66 std::set<BasicBlock*> Visited;
69 PromoteMem2Reg(const std::vector<AllocaInst*> &A, DominatorTree &dt,
70 DominanceFrontier &df, const TargetData &td)
71 : Allocas(A), DT(dt), DF(df), TD(td) {}
76 void MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
77 std::set<PHINode*> &DeadPHINodes);
78 void PromoteLocallyUsedAlloca(AllocaInst *AI);
80 void RenamePass(BasicBlock *BB, BasicBlock *Pred,
81 std::vector<Value*> &IncVals);
82 bool QueuePhiNode(BasicBlock *BB, unsigned AllocaIdx, unsigned &Version,
83 std::set<PHINode*> &InsertedPHINodes);
85 } // end of anonymous namespace
87 void PromoteMem2Reg::run() {
88 Function &F = *DF.getRoot()->getParent();
90 for (unsigned AllocaNum = 0; AllocaNum != Allocas.size(); ++AllocaNum) {
91 AllocaInst *AI = Allocas[AllocaNum];
93 assert(isAllocaPromotable(AI, TD) &&
94 "Cannot promote non-promotable alloca!");
95 assert(AI->getParent()->getParent() == &F &&
96 "All allocas should be in the same function, which is same as DF!");
98 if (AI->use_empty()) {
99 // If there are no uses of the alloca, just delete it now.
100 AI->getParent()->getInstList().erase(AI);
102 // Remove the alloca from the Allocas list, since it has been processed
103 Allocas[AllocaNum] = Allocas.back();
109 // Calculate the set of read and write-locations for each alloca. This is
110 // analogous to counting the number of 'uses' and 'definitions' of each
112 std::vector<BasicBlock*> DefiningBlocks;
113 std::vector<BasicBlock*> UsingBlocks;
115 BasicBlock *OnlyBlock = 0;
116 bool OnlyUsedInOneBlock = true;
118 // As we scan the uses of the alloca instruction, keep track of stores, and
119 // decide whether all of the loads and stores to the alloca are within the
121 for (Value::use_iterator U =AI->use_begin(), E = AI->use_end(); U != E;++U){
122 Instruction *User = cast<Instruction>(*U);
123 if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
124 // Remember the basic blocks which define new values for the alloca
125 DefiningBlocks.push_back(SI->getParent());
127 // Otherwise it must be a load instruction, keep track of variable reads
128 UsingBlocks.push_back(cast<LoadInst>(User)->getParent());
131 if (OnlyUsedInOneBlock) {
133 OnlyBlock = User->getParent();
134 else if (OnlyBlock != User->getParent())
135 OnlyUsedInOneBlock = false;
139 // If the alloca is only read and written in one basic block, just perform a
140 // linear sweep over the block to eliminate it.
141 if (OnlyUsedInOneBlock) {
142 PromoteLocallyUsedAlloca(AI);
144 // Remove the alloca from the Allocas list, since it has been processed
145 Allocas[AllocaNum] = Allocas.back();
151 // Compute the locations where PhiNodes need to be inserted. Look at the
152 // dominance frontier of EACH basic-block we have a write in.
154 unsigned CurrentVersion = 0;
155 std::set<PHINode*> InsertedPHINodes;
156 while (!DefiningBlocks.empty()) {
157 BasicBlock *BB = DefiningBlocks.back();
158 DefiningBlocks.pop_back();
160 // Look up the DF for this write, add it to PhiNodes
161 DominanceFrontier::const_iterator it = DF.find(BB);
162 if (it != DF.end()) {
163 const DominanceFrontier::DomSetType &S = it->second;
164 for (DominanceFrontier::DomSetType::iterator P = S.begin(),PE = S.end();
166 if (QueuePhiNode(*P, AllocaNum, CurrentVersion, InsertedPHINodes))
167 DefiningBlocks.push_back(*P);
171 // Now that we have inserted PHI nodes along the Iterated Dominance Frontier
172 // of the writes to the variable, scan through the reads of the variable,
173 // marking PHI nodes which are actually necessary as alive (by removing them
174 // from the InsertedPHINodes set). This is not perfect: there may PHI
175 // marked alive because of loads which are dominated by stores, but there
176 // will be no unmarked PHI nodes which are actually used.
178 for (unsigned i = 0, e = UsingBlocks.size(); i != e; ++i)
179 MarkDominatingPHILive(UsingBlocks[i], AllocaNum, InsertedPHINodes);
182 // If there are any PHI nodes which are now known to be dead, remove them!
183 for (std::set<PHINode*>::iterator I = InsertedPHINodes.begin(),
184 E = InsertedPHINodes.end(); I != E; ++I) {
186 std::vector<PHINode*> &BBPNs = NewPhiNodes[PN->getParent()];
187 BBPNs[AllocaNum] = 0;
189 // Check to see if we just removed the last inserted PHI node from this
190 // basic block. If so, remove the entry for the basic block.
191 bool HasOtherPHIs = false;
192 for (unsigned i = 0, e = BBPNs.size(); i != e; ++i)
198 NewPhiNodes.erase(PN->getParent());
200 PN->getParent()->getInstList().erase(PN);
203 // Keep the reverse mapping of the 'Allocas' array.
204 AllocaLookup[Allocas[AllocaNum]] = AllocaNum;
208 return; // All of the allocas must have been trivial!
210 // Set the incoming values for the basic block to be null values for all of
211 // the alloca's. We do this in case there is a load of a value that has not
212 // been stored yet. In this case, it will get this null value.
214 std::vector<Value *> Values(Allocas.size());
215 for (unsigned i = 0, e = Allocas.size(); i != e; ++i)
216 Values[i] = Constant::getNullValue(Allocas[i]->getAllocatedType());
218 // Walks all basic blocks in the function performing the SSA rename algorithm
219 // and inserting the phi nodes we marked as necessary
221 RenamePass(F.begin(), 0, Values);
223 // The renamer uses the Visited set to avoid infinite loops. Clear it now.
226 // Remove the allocas themselves from the function...
227 for (unsigned i = 0, e = Allocas.size(); i != e; ++i) {
228 Instruction *A = Allocas[i];
230 // If there are any uses of the alloca instructions left, they must be in
231 // sections of dead code that were not processed on the dominance frontier.
232 // Just delete the users now.
235 A->replaceAllUsesWith(Constant::getNullValue(A->getType()));
236 A->getParent()->getInstList().erase(A);
239 // At this point, the renamer has added entries to PHI nodes for all reachable
240 // code. Unfortunately, there may be blocks which are not reachable, which
241 // the renamer hasn't traversed. If this is the case, the PHI nodes may not
242 // have incoming values for all predecessors. Loop over all PHI nodes we have
243 // created, inserting null constants if they are missing any incoming values.
245 for (std::map<BasicBlock*, std::vector<PHINode *> >::iterator I =
246 NewPhiNodes.begin(), E = NewPhiNodes.end(); I != E; ++I) {
248 std::vector<BasicBlock*> Preds(pred_begin(I->first), pred_end(I->first));
249 std::vector<PHINode*> &PNs = I->second;
250 assert(!PNs.empty() && "Empty PHI node list??");
252 // Only do work here if there the PHI nodes are missing incoming values. We
253 // know that all PHI nodes that were inserted in a block will have the same
254 // number of incoming values, so we can just check any PHI node.
256 for (unsigned i = 0; (FirstPHI = PNs[i]) == 0; ++i)
259 if (Preds.size() != FirstPHI->getNumIncomingValues()) {
260 // Ok, now we know that all of the PHI nodes are missing entries for some
261 // basic blocks. Start by sorting the incoming predecessors for efficient
263 std::sort(Preds.begin(), Preds.end());
265 // Now we loop through all BB's which have entries in FirstPHI and remove
266 // them from the Preds list.
267 for (unsigned i = 0, e = FirstPHI->getNumIncomingValues(); i != e; ++i) {
268 // Do a log(n) search of the Preds list for the entry we want.
269 std::vector<BasicBlock*>::iterator EntIt =
270 std::lower_bound(Preds.begin(), Preds.end(),
271 FirstPHI->getIncomingBlock(i));
272 assert(EntIt != Preds.end() && *EntIt == FirstPHI->getIncomingBlock(i)&&
273 "PHI node has entry for a block which is not a predecessor!");
279 // At this point, the blocks left in the preds list must have dummy
280 // entries inserted into every PHI nodes for the block.
281 for (unsigned i = 0, e = PNs.size(); i != e; ++i)
282 if (PHINode *PN = PNs[i]) {
283 Value *NullVal = Constant::getNullValue(PN->getType());
284 for (unsigned pred = 0, e = Preds.size(); pred != e; ++pred)
285 PN->addIncoming(NullVal, Preds[pred]);
291 // MarkDominatingPHILive - Mem2Reg wants to construct "pruned" SSA form, not
292 // "minimal" SSA form. To do this, it inserts all of the PHI nodes on the IDF
293 // as usual (inserting the PHI nodes in the DeadPHINodes set), then processes
294 // each read of the variable. For each block that reads the variable, this
295 // function is called, which removes used PHI nodes from the DeadPHINodes set.
296 // After all of the reads have been processed, any PHI nodes left in the
297 // DeadPHINodes set are removed.
299 void PromoteMem2Reg::MarkDominatingPHILive(BasicBlock *BB, unsigned AllocaNum,
300 std::set<PHINode*> &DeadPHINodes) {
301 // Scan the immediate dominators of this block looking for a block which has a
302 // PHI node for Alloca num. If we find it, mark the PHI node as being alive!
303 for (DominatorTree::Node *N = DT[BB]; N; N = N->getIDom()) {
304 BasicBlock *DomBB = N->getBlock();
305 std::map<BasicBlock*, std::vector<PHINode*> >::iterator
306 I = NewPhiNodes.find(DomBB);
307 if (I != NewPhiNodes.end() && I->second[AllocaNum]) {
308 // Ok, we found an inserted PHI node which dominates this value.
309 PHINode *DominatingPHI = I->second[AllocaNum];
311 // Find out if we previously thought it was dead.
312 std::set<PHINode*>::iterator DPNI = DeadPHINodes.find(DominatingPHI);
313 if (DPNI != DeadPHINodes.end()) {
314 // Ok, until now, we thought this PHI node was dead. Mark it as being
316 DeadPHINodes.erase(DPNI);
318 // Now that we have marked the PHI node alive, also mark any PHI nodes
319 // which it might use as being alive as well.
320 for (pred_iterator PI = pred_begin(DomBB), PE = pred_end(DomBB);
322 MarkDominatingPHILive(*PI, AllocaNum, DeadPHINodes);
328 // PromoteLocallyUsedAlloca - Many allocas are only used within a single basic
329 // block. If this is the case, avoid traversing the CFG and inserting a lot of
330 // potentially useless PHI nodes by just performing a single linear pass over
331 // the basic block using the Alloca.
333 void PromoteMem2Reg::PromoteLocallyUsedAlloca(AllocaInst *AI) {
334 assert(!AI->use_empty() && "There are no uses of the alloca!");
336 // Uses of the uninitialized memory location shall get zero...
337 Value *CurVal = Constant::getNullValue(AI->getAllocatedType());
339 BasicBlock *BB = cast<Instruction>(AI->use_back())->getParent();
341 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
342 Instruction *Inst = I++;
343 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
344 if (LI->getOperand(0) == AI) {
345 // Loads just return the "current value"...
346 LI->replaceAllUsesWith(CurVal);
347 BB->getInstList().erase(LI);
349 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
350 if (SI->getOperand(1) == AI) {
351 // Loads just update the "current value"...
352 CurVal = SI->getOperand(0);
353 BB->getInstList().erase(SI);
358 // After traversing the basic block, there should be no more uses of the
359 // alloca, remove it now.
360 assert(AI->use_empty() && "Uses of alloca from more than one BB??");
361 AI->getParent()->getInstList().erase(AI);
364 // QueuePhiNode - queues a phi-node to be added to a basic-block for a specific
365 // Alloca returns true if there wasn't already a phi-node for that variable
367 bool PromoteMem2Reg::QueuePhiNode(BasicBlock *BB, unsigned AllocaNo,
369 std::set<PHINode*> &InsertedPHINodes) {
370 // Look up the basic-block in question
371 std::vector<PHINode*> &BBPNs = NewPhiNodes[BB];
372 if (BBPNs.empty()) BBPNs.resize(Allocas.size());
374 // If the BB already has a phi node added for the i'th alloca then we're done!
375 if (BBPNs[AllocaNo]) return false;
377 // Create a PhiNode using the dereferenced type... and add the phi-node to the
379 BBPNs[AllocaNo] = new PHINode(Allocas[AllocaNo]->getAllocatedType(),
380 Allocas[AllocaNo]->getName() + "." +
381 utostr(Version++), BB->begin());
382 InsertedPHINodes.insert(BBPNs[AllocaNo]);
387 // RenamePass - Recursively traverse the CFG of the function, renaming loads and
388 // stores to the allocas which we are promoting. IncomingVals indicates what
389 // value each Alloca contains on exit from the predecessor block Pred.
391 void PromoteMem2Reg::RenamePass(BasicBlock *BB, BasicBlock *Pred,
392 std::vector<Value*> &IncomingVals) {
394 // If this BB needs a PHI node, update the PHI node for each variable we need
396 std::map<BasicBlock*, std::vector<PHINode *> >::iterator
397 BBPNI = NewPhiNodes.find(BB);
398 if (BBPNI != NewPhiNodes.end()) {
399 std::vector<PHINode *> &BBPNs = BBPNI->second;
400 for (unsigned k = 0; k != BBPNs.size(); ++k)
401 if (PHINode *PN = BBPNs[k]) {
402 // Add this incoming value to the PHI node.
403 PN->addIncoming(IncomingVals[k], Pred);
405 // The currently active variable for this block is now the PHI.
406 IncomingVals[k] = PN;
410 // don't revisit nodes
411 if (Visited.count(BB)) return;
416 for (BasicBlock::iterator II = BB->begin(); !isa<TerminatorInst>(II); ) {
417 Instruction *I = II++; // get the instruction, increment iterator
419 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
420 if (AllocaInst *Src = dyn_cast<AllocaInst>(LI->getPointerOperand())) {
421 std::map<AllocaInst*, unsigned>::iterator AI = AllocaLookup.find(Src);
422 if (AI != AllocaLookup.end()) {
423 Value *V = IncomingVals[AI->second];
425 // walk the use list of this load and replace all uses with r
426 LI->replaceAllUsesWith(V);
427 BB->getInstList().erase(LI);
430 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
431 // Delete this instruction and mark the name as the current holder of the
433 if (AllocaInst *Dest = dyn_cast<AllocaInst>(SI->getPointerOperand())) {
434 std::map<AllocaInst *, unsigned>::iterator ai = AllocaLookup.find(Dest);
435 if (ai != AllocaLookup.end()) {
436 // what value were we writing?
437 IncomingVals[ai->second] = SI->getOperand(0);
438 BB->getInstList().erase(SI);
444 // Recurse to our successors.
445 TerminatorInst *TI = BB->getTerminator();
446 for (unsigned i = 0; i != TI->getNumSuccessors(); i++) {
447 std::vector<Value*> OutgoingVals(IncomingVals);
448 RenamePass(TI->getSuccessor(i), BB, OutgoingVals);
452 /// PromoteMemToReg - Promote the specified list of alloca instructions into
453 /// scalar registers, inserting PHI nodes as appropriate. This function makes
454 /// use of DominanceFrontier information. This function does not modify the CFG
455 /// of the function at all. All allocas must be from the same function.
457 void PromoteMemToReg(const std::vector<AllocaInst*> &Allocas,
458 DominatorTree &DT, DominanceFrontier &DF,
459 const TargetData &TD) {
460 // If there is nothing to do, bail out...
461 if (Allocas.empty()) return;
462 PromoteMem2Reg(Allocas, DT, DF, TD).run();
465 } // End llvm namespace