1 //===- LoadValueNumbering.cpp - Load Value #'ing Implementation -*- C++ -*-===//
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 implements a value numbering pass that value numbers load and call
11 // instructions. To do this, it finds lexically identical load instructions,
12 // and uses alias analysis to determine which loads are guaranteed to produce
13 // the same value. To value number call instructions, it looks for calls to
14 // functions that do not write to memory which do not have intervening
15 // instructions that clobber the memory that is read from.
17 // This pass builds off of another value numbering pass to implement value
18 // numbering for non-load and non-call instructions. It uses Alias Analysis so
19 // that it can disambiguate the load instructions. The more powerful these base
20 // analyses are, the more powerful the resultant value numbering will be.
22 //===----------------------------------------------------------------------===//
24 #include "llvm/Analysis/LoadValueNumbering.h"
25 #include "llvm/Constants.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Pass.h"
29 #include "llvm/Type.h"
30 #include "llvm/Analysis/ValueNumbering.h"
31 #include "llvm/Analysis/AliasAnalysis.h"
32 #include "llvm/Analysis/Dominators.h"
33 #include "llvm/Support/CFG.h"
34 #include "llvm/Target/TargetData.h"
40 // FIXME: This should not be a FunctionPass.
41 struct LoadVN : public FunctionPass, public ValueNumbering {
43 /// Pass Implementation stuff. This doesn't do any analysis.
45 bool runOnFunction(Function &) { return false; }
47 /// getAnalysisUsage - Does not modify anything. It uses Value Numbering
48 /// and Alias Analysis.
50 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
52 /// getEqualNumberNodes - Return nodes with the same value number as the
53 /// specified Value. This fills in the argument vector with any equal
56 virtual void getEqualNumberNodes(Value *V1,
57 std::vector<Value*> &RetVals) const;
59 /// deleteValue - This method should be called whenever an LLVM Value is
60 /// deleted from the program, for example when an instruction is found to be
61 /// redundant and is eliminated.
63 virtual void deleteValue(Value *V) {
64 getAnalysis<AliasAnalysis>().deleteValue(V);
67 /// copyValue - This method should be used whenever a preexisting value in
68 /// the program is copied or cloned, introducing a new value. Note that
69 /// analysis implementations should tolerate clients that use this method to
70 /// introduce the same value multiple times: if the analysis already knows
71 /// about a value, it should ignore the request.
73 virtual void copyValue(Value *From, Value *To) {
74 getAnalysis<AliasAnalysis>().copyValue(From, To);
77 /// getCallEqualNumberNodes - Given a call instruction, find other calls
78 /// that have the same value number.
79 void getCallEqualNumberNodes(CallInst *CI,
80 std::vector<Value*> &RetVals) const;
83 // Register this pass...
84 RegisterOpt<LoadVN> X("load-vn", "Load Value Numbering");
86 // Declare that we implement the ValueNumbering interface
87 RegisterAnalysisGroup<ValueNumbering, LoadVN> Y;
90 FunctionPass *llvm::createLoadValueNumberingPass() { return new LoadVN(); }
93 /// getAnalysisUsage - Does not modify anything. It uses Value Numbering and
96 void LoadVN::getAnalysisUsage(AnalysisUsage &AU) const {
98 AU.addRequired<AliasAnalysis>();
99 AU.addRequired<ValueNumbering>();
100 AU.addRequired<DominatorSet>();
101 AU.addRequired<TargetData>();
104 static bool isPathTransparentTo(BasicBlock *CurBlock, BasicBlock *Dom,
105 Value *Ptr, unsigned Size, AliasAnalysis &AA,
106 std::set<BasicBlock*> &Visited,
107 std::map<BasicBlock*, bool> &TransparentBlocks){
108 // If we have already checked out this path, or if we reached our destination,
109 // stop searching, returning success.
110 if (CurBlock == Dom || !Visited.insert(CurBlock).second)
113 // Check whether this block is known transparent or not.
114 std::map<BasicBlock*, bool>::iterator TBI =
115 TransparentBlocks.lower_bound(CurBlock);
117 if (TBI == TransparentBlocks.end() || TBI->first != CurBlock) {
118 // If this basic block can modify the memory location, then the path is not
120 if (AA.canBasicBlockModify(*CurBlock, Ptr, Size)) {
121 TransparentBlocks.insert(TBI, std::make_pair(CurBlock, false));
124 TransparentBlocks.insert(TBI, std::make_pair(CurBlock, true));
125 } else if (!TBI->second)
126 // This block is known non-transparent, so that path can't be either.
129 // The current block is known to be transparent. The entire path is
130 // transparent if all of the predecessors paths to the parent is also
131 // transparent to the memory location.
132 for (pred_iterator PI = pred_begin(CurBlock), E = pred_end(CurBlock);
134 if (!isPathTransparentTo(*PI, Dom, Ptr, Size, AA, Visited,
140 /// getCallEqualNumberNodes - Given a call instruction, find other calls that
141 /// have the same value number.
142 void LoadVN::getCallEqualNumberNodes(CallInst *CI,
143 std::vector<Value*> &RetVals) const {
144 Function *CF = CI->getCalledFunction();
145 if (CF == 0) return; // Indirect call.
146 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
147 AliasAnalysis::ModRefBehavior MRB = AA.getModRefBehavior(CF, CI);
148 if (MRB != AliasAnalysis::DoesNotAccessMemory &&
149 MRB != AliasAnalysis::OnlyReadsMemory)
150 return; // Nothing we can do for now.
152 // Scan all of the arguments of the function, looking for one that is not
153 // global. In particular, we would prefer to have an argument or instruction
154 // operand to chase the def-use chains of.
156 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
157 if (isa<Argument>(CI->getOperand(i)) ||
158 isa<Instruction>(CI->getOperand(i))) {
159 Op = CI->getOperand(i);
163 // Identify all lexically identical calls in this function.
164 std::vector<CallInst*> IdenticalCalls;
166 Function *CIFunc = CI->getParent()->getParent();
167 for (Value::use_iterator UI = Op->use_begin(), E = Op->use_end(); UI != E;
169 if (CallInst *C = dyn_cast<CallInst>(*UI))
170 if (C->getNumOperands() == CI->getNumOperands() &&
171 C->getOperand(0) == CI->getOperand(0) &&
172 C->getParent()->getParent() == CIFunc && C != CI) {
173 bool AllOperandsEqual = true;
174 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
175 if (C->getOperand(i) != CI->getOperand(i)) {
176 AllOperandsEqual = false;
180 if (AllOperandsEqual)
181 IdenticalCalls.push_back(C);
184 if (IdenticalCalls.empty()) return;
186 // Eliminate duplicates, which could occur if we chose a value that is passed
187 // into a call site multiple times.
188 std::sort(IdenticalCalls.begin(), IdenticalCalls.end());
189 IdenticalCalls.erase(std::unique(IdenticalCalls.begin(),IdenticalCalls.end()),
190 IdenticalCalls.end());
192 // If the call reads memory, we must make sure that there are no stores
193 // between the calls in question.
195 // FIXME: This should use mod/ref information. What we really care about it
196 // whether an intervening instruction could modify memory that is read, not
199 if (MRB == AliasAnalysis::OnlyReadsMemory) {
200 DominatorSet &DomSetInfo = getAnalysis<DominatorSet>();
201 BasicBlock *CIBB = CI->getParent();
202 for (unsigned i = 0; i != IdenticalCalls.size(); ++i) {
203 CallInst *C = IdenticalCalls[i];
204 bool CantEqual = false;
206 if (DomSetInfo.dominates(CIBB, C->getParent())) {
207 // FIXME: we currently only handle the case where both calls are in the
209 if (CIBB != C->getParent()) {
212 Instruction *First = CI, *Second = C;
213 if (!DomSetInfo.dominates(CI, C))
214 std::swap(First, Second);
216 // Scan the instructions between the calls, checking for stores or
217 // calls to dangerous functions.
218 BasicBlock::iterator I = First;
219 for (++First; I != BasicBlock::iterator(Second); ++I) {
220 if (isa<StoreInst>(I)) {
221 // FIXME: We could use mod/ref information to make this much
225 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
226 if (CI->getCalledFunction() == 0 ||
227 !AA.onlyReadsMemory(CI->getCalledFunction())) {
231 } else if (I->mayWriteToMemory()) {
238 } else if (DomSetInfo.dominates(C->getParent(), CIBB)) {
239 // FIXME: We could implement this, but we don't for now.
242 // FIXME: if one doesn't dominate the other, we can't tell yet.
248 // This call does not produce the same value as the one in the query.
249 std::swap(IdenticalCalls[i--], IdenticalCalls.back());
250 IdenticalCalls.pop_back();
255 // Any calls that are identical and not destroyed will produce equal values!
256 for (unsigned i = 0, e = IdenticalCalls.size(); i != e; ++i)
257 RetVals.push_back(IdenticalCalls[i]);
260 // getEqualNumberNodes - Return nodes with the same value number as the
261 // specified Value. This fills in the argument vector with any equal values.
263 void LoadVN::getEqualNumberNodes(Value *V,
264 std::vector<Value*> &RetVals) const {
265 // If the alias analysis has any must alias information to share with us, we
266 // can definitely use it.
267 if (isa<PointerType>(V->getType()))
268 getAnalysis<AliasAnalysis>().getMustAliases(V, RetVals);
270 if (!isa<LoadInst>(V)) {
271 if (CallInst *CI = dyn_cast<CallInst>(V))
272 getCallEqualNumberNodes(CI, RetVals);
274 // Not a load instruction? Just chain to the base value numbering
275 // implementation to satisfy the request...
276 assert(&getAnalysis<ValueNumbering>() != (ValueNumbering*)this &&
277 "getAnalysis() returned this!");
279 return getAnalysis<ValueNumbering>().getEqualNumberNodes(V, RetVals);
282 // Volatile loads cannot be replaced with the value of other loads.
283 LoadInst *LI = cast<LoadInst>(V);
284 if (LI->isVolatile())
285 return getAnalysis<ValueNumbering>().getEqualNumberNodes(V, RetVals);
287 Value *LoadPtr = LI->getOperand(0);
288 BasicBlock *LoadBB = LI->getParent();
289 Function *F = LoadBB->getParent();
291 // Find out how many bytes of memory are loaded by the load instruction...
292 unsigned LoadSize = getAnalysis<TargetData>().getTypeSize(LI->getType());
293 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
295 // Figure out if the load is invalidated from the entry of the block it is in
296 // until the actual instruction. This scans the block backwards from LI. If
297 // we see any candidate load or store instructions, then we know that the
298 // candidates have the same value # as LI.
299 bool LoadInvalidatedInBBBefore = false;
300 for (BasicBlock::iterator I = LI; I != LoadBB->begin(); ) {
303 // If we run into an allocation of the value being loaded, then the
304 // contents are not initialized.
305 if (isa<AllocationInst>(I))
306 RetVals.push_back(UndefValue::get(LI->getType()));
308 // Otherwise, since this is the definition of what we are loading, this
309 // loaded value cannot occur before this block.
310 LoadInvalidatedInBBBefore = true;
312 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
313 // If this instruction is a candidate load before LI, we know there are no
314 // invalidating instructions between it and LI, so they have the same
316 if (LI->getOperand(0) == LoadPtr && !LI->isVolatile())
317 RetVals.push_back(I);
320 if (AA.getModRefInfo(I, LoadPtr, LoadSize) & AliasAnalysis::Mod) {
321 // If the invalidating instruction is a store, and its in our candidate
322 // set, then we can do store-load forwarding: the load has the same value
323 // # as the stored value.
324 if (StoreInst *SI = dyn_cast<StoreInst>(I))
325 if (SI->getOperand(1) == LoadPtr)
326 RetVals.push_back(I->getOperand(0));
328 LoadInvalidatedInBBBefore = true;
331 if (dyn_cast<VAArgInst>(I)) {
332 LoadInvalidatedInBBBefore = true;
337 // Figure out if the load is invalidated between the load and the exit of the
338 // block it is defined in. While we are scanning the current basic block, if
339 // we see any candidate loads, then we know they have the same value # as LI.
341 bool LoadInvalidatedInBBAfter = false;
342 for (BasicBlock::iterator I = LI->getNext(); I != LoadBB->end(); ++I) {
343 // If this instruction is a load, then this instruction returns the same
345 if (isa<LoadInst>(I) && cast<LoadInst>(I)->getOperand(0) == LoadPtr)
346 RetVals.push_back(I);
348 if (AA.getModRefInfo(I, LoadPtr, LoadSize) & AliasAnalysis::Mod) {
349 LoadInvalidatedInBBAfter = true;
354 // If the pointer is clobbered on entry and on exit to the function, there is
355 // no need to do any global analysis at all.
356 if (LoadInvalidatedInBBBefore && LoadInvalidatedInBBAfter)
359 // Now that we know the value is not neccesarily killed on entry or exit to
360 // the BB, find out how many load and store instructions (to this location)
361 // live in each BB in the function.
363 std::map<BasicBlock*, unsigned> CandidateLoads;
364 std::set<BasicBlock*> CandidateStores;
366 for (Value::use_iterator UI = LoadPtr->use_begin(), UE = LoadPtr->use_end();
368 if (LoadInst *Cand = dyn_cast<LoadInst>(*UI)) {// Is a load of source?
369 if (Cand->getParent()->getParent() == F && // In the same function?
370 // Not in LI's block?
371 Cand->getParent() != LoadBB && !Cand->isVolatile())
372 ++CandidateLoads[Cand->getParent()]; // Got one.
373 } else if (StoreInst *Cand = dyn_cast<StoreInst>(*UI)) {
374 if (Cand->getParent()->getParent() == F && !Cand->isVolatile() &&
375 Cand->getOperand(1) == LoadPtr) // It's a store THROUGH the ptr.
376 CandidateStores.insert(Cand->getParent());
380 DominatorSet &DomSetInfo = getAnalysis<DominatorSet>();
382 // TransparentBlocks - For each basic block the load/store is alive across,
383 // figure out if the pointer is invalidated or not. If it is invalidated, the
384 // boolean is set to false, if it's not it is set to true. If we don't know
385 // yet, the entry is not in the map.
386 std::map<BasicBlock*, bool> TransparentBlocks;
388 // Loop over all of the basic blocks that also load the value. If the value
389 // is live across the CFG from the source to destination blocks, and if the
390 // value is not invalidated in either the source or destination blocks, add it
391 // to the equivalence sets.
392 for (std::map<BasicBlock*, unsigned>::iterator
393 I = CandidateLoads.begin(), E = CandidateLoads.end(); I != E; ++I) {
394 bool CantEqual = false;
396 // Right now we only can handle cases where one load dominates the other.
397 // FIXME: generalize this!
398 BasicBlock *BB1 = I->first, *BB2 = LoadBB;
399 if (DomSetInfo.dominates(BB1, BB2)) {
400 // The other load dominates LI. If the loaded value is killed entering
401 // the LoadBB block, we know the load is not live.
402 if (LoadInvalidatedInBBBefore)
404 } else if (DomSetInfo.dominates(BB2, BB1)) {
405 std::swap(BB1, BB2); // Canonicalize
406 // LI dominates the other load. If the loaded value is killed exiting
407 // the LoadBB block, we know the load is not live.
408 if (LoadInvalidatedInBBAfter)
411 // None of these loads can VN the same.
416 // Ok, at this point, we know that BB1 dominates BB2, and that there is
417 // nothing in the LI block that kills the loaded value. Check to see if
418 // the value is live across the CFG.
419 std::set<BasicBlock*> Visited;
420 for (pred_iterator PI = pred_begin(BB2), E = pred_end(BB2); PI!=E; ++PI)
421 if (!isPathTransparentTo(*PI, BB1, LoadPtr, LoadSize, AA,
422 Visited, TransparentBlocks)) {
423 // None of these loads can VN the same.
429 // If the loads can equal so far, scan the basic block that contains the
430 // loads under consideration to see if they are invalidated in the block.
431 // For any loads that are not invalidated, add them to the equivalence
434 unsigned NumLoads = I->second;
436 // If LI dominates the block in question, check to see if any of the
437 // loads in this block are invalidated before they are reached.
438 for (BasicBlock::iterator BBI = I->first->begin(); ; ++BBI) {
439 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
440 if (LI->getOperand(0) == LoadPtr && !LI->isVolatile()) {
441 // The load is in the set!
442 RetVals.push_back(BBI);
443 if (--NumLoads == 0) break; // Found last load to check.
445 } else if (AA.getModRefInfo(BBI, LoadPtr, LoadSize)
446 & AliasAnalysis::Mod) {
447 // If there is a modifying instruction, nothing below it will value
453 // If the block dominates LI, make sure that the loads in the block are
454 // not invalidated before the block ends.
455 BasicBlock::iterator BBI = I->first->end();
458 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
459 if (LI->getOperand(0) == LoadPtr && !LI->isVolatile()) {
460 // The load is the same as this load!
461 RetVals.push_back(BBI);
462 if (--NumLoads == 0) break; // Found all of the laods.
464 } else if (AA.getModRefInfo(BBI, LoadPtr, LoadSize)
465 & AliasAnalysis::Mod) {
466 // If there is a modifying instruction, nothing above it will value
475 // Handle candidate stores. If the loaded location is clobbered on entrance
476 // to the LoadBB, no store outside of the LoadBB can value number equal, so
478 if (LoadInvalidatedInBBBefore)
481 // Stores in the load-bb are handled above.
482 CandidateStores.erase(LoadBB);
484 for (std::set<BasicBlock*>::iterator I = CandidateStores.begin(),
485 E = CandidateStores.end(); I != E; ++I)
486 if (DomSetInfo.dominates(*I, LoadBB)) {
487 BasicBlock *StoreBB = *I;
489 // Check to see if the path from the store to the load is transparent
490 // w.r.t. the memory location.
491 bool CantEqual = false;
492 std::set<BasicBlock*> Visited;
493 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
495 if (!isPathTransparentTo(*PI, StoreBB, LoadPtr, LoadSize, AA,
496 Visited, TransparentBlocks)) {
497 // None of these stores can VN the same.
503 // Okay, the path from the store block to the load block is clear, and
504 // we know that there are no invalidating instructions from the start
505 // of the load block to the load itself. Now we just scan the store
508 BasicBlock::iterator BBI = StoreBB->end();
510 assert(BBI != StoreBB->begin() &&
511 "There is a store in this block of the pointer, but the store"
512 " doesn't mod the address being stored to?? Must be a bug in"
513 " the alias analysis implementation!");
515 if (AA.getModRefInfo(BBI, LoadPtr, LoadSize) & AliasAnalysis::Mod) {
516 // If the invalidating instruction is one of the candidates,
517 // then it provides the value the load loads.
518 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
519 if (SI->getOperand(1) == LoadPtr)
520 RetVals.push_back(SI->getOperand(0));