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/Support/Compiler.h"
35 #include "llvm/Target/TargetData.h"
41 // FIXME: This should not be a FunctionPass.
42 struct VISIBILITY_HIDDEN LoadVN : public FunctionPass, public ValueNumbering {
44 /// Pass Implementation stuff. This doesn't do any analysis.
46 bool runOnFunction(Function &) { return false; }
48 /// getAnalysisUsage - Does not modify anything. It uses Value Numbering
49 /// and Alias Analysis.
51 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
53 /// getEqualNumberNodes - Return nodes with the same value number as the
54 /// specified Value. This fills in the argument vector with any equal
57 virtual void getEqualNumberNodes(Value *V1,
58 std::vector<Value*> &RetVals) const;
60 /// deleteValue - This method should be called whenever an LLVM Value is
61 /// deleted from the program, for example when an instruction is found to be
62 /// redundant and is eliminated.
64 virtual void deleteValue(Value *V) {
65 getAnalysis<AliasAnalysis>().deleteValue(V);
68 /// copyValue - This method should be used whenever a preexisting value in
69 /// the program is copied or cloned, introducing a new value. Note that
70 /// analysis implementations should tolerate clients that use this method to
71 /// introduce the same value multiple times: if the analysis already knows
72 /// about a value, it should ignore the request.
74 virtual void copyValue(Value *From, Value *To) {
75 getAnalysis<AliasAnalysis>().copyValue(From, To);
78 /// getCallEqualNumberNodes - Given a call instruction, find other calls
79 /// that have the same value number.
80 void getCallEqualNumberNodes(CallInst *CI,
81 std::vector<Value*> &RetVals) const;
84 // Register this pass...
85 RegisterPass<LoadVN> X("load-vn", "Load Value Numbering");
87 // Declare that we implement the ValueNumbering interface
88 RegisterAnalysisGroup<ValueNumbering> Y(X);
91 FunctionPass *llvm::createLoadValueNumberingPass() { return new LoadVN(); }
94 /// getAnalysisUsage - Does not modify anything. It uses Value Numbering and
97 void LoadVN::getAnalysisUsage(AnalysisUsage &AU) const {
99 AU.addRequiredTransitive<AliasAnalysis>();
100 AU.addRequired<ValueNumbering>();
101 AU.addRequiredTransitive<DominatorSet>();
102 AU.addRequiredTransitive<TargetData>();
105 static bool isPathTransparentTo(BasicBlock *CurBlock, BasicBlock *Dom,
106 Value *Ptr, unsigned Size, AliasAnalysis &AA,
107 std::set<BasicBlock*> &Visited,
108 std::map<BasicBlock*, bool> &TransparentBlocks){
109 // If we have already checked out this path, or if we reached our destination,
110 // stop searching, returning success.
111 if (CurBlock == Dom || !Visited.insert(CurBlock).second)
114 // Check whether this block is known transparent or not.
115 std::map<BasicBlock*, bool>::iterator TBI =
116 TransparentBlocks.lower_bound(CurBlock);
118 if (TBI == TransparentBlocks.end() || TBI->first != CurBlock) {
119 // If this basic block can modify the memory location, then the path is not
121 if (AA.canBasicBlockModify(*CurBlock, Ptr, Size)) {
122 TransparentBlocks.insert(TBI, std::make_pair(CurBlock, false));
125 TransparentBlocks.insert(TBI, std::make_pair(CurBlock, true));
126 } else if (!TBI->second)
127 // This block is known non-transparent, so that path can't be either.
130 // The current block is known to be transparent. The entire path is
131 // transparent if all of the predecessors paths to the parent is also
132 // transparent to the memory location.
133 for (pred_iterator PI = pred_begin(CurBlock), E = pred_end(CurBlock);
135 if (!isPathTransparentTo(*PI, Dom, Ptr, Size, AA, Visited,
141 /// getCallEqualNumberNodes - Given a call instruction, find other calls that
142 /// have the same value number.
143 void LoadVN::getCallEqualNumberNodes(CallInst *CI,
144 std::vector<Value*> &RetVals) const {
145 Function *CF = CI->getCalledFunction();
146 if (CF == 0) return; // Indirect call.
147 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
148 AliasAnalysis::ModRefBehavior MRB = AA.getModRefBehavior(CF, CI);
149 if (MRB != AliasAnalysis::DoesNotAccessMemory &&
150 MRB != AliasAnalysis::OnlyReadsMemory)
151 return; // Nothing we can do for now.
153 // Scan all of the arguments of the function, looking for one that is not
154 // global. In particular, we would prefer to have an argument or instruction
155 // operand to chase the def-use chains of.
157 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
158 if (isa<Argument>(CI->getOperand(i)) ||
159 isa<Instruction>(CI->getOperand(i))) {
160 Op = CI->getOperand(i);
164 // Identify all lexically identical calls in this function.
165 std::vector<CallInst*> IdenticalCalls;
167 Function *CIFunc = CI->getParent()->getParent();
168 for (Value::use_iterator UI = Op->use_begin(), E = Op->use_end(); UI != E;
170 if (CallInst *C = dyn_cast<CallInst>(*UI))
171 if (C->getNumOperands() == CI->getNumOperands() &&
172 C->getOperand(0) == CI->getOperand(0) &&
173 C->getParent()->getParent() == CIFunc && C != CI) {
174 bool AllOperandsEqual = true;
175 for (unsigned i = 1, e = CI->getNumOperands(); i != e; ++i)
176 if (C->getOperand(i) != CI->getOperand(i)) {
177 AllOperandsEqual = false;
181 if (AllOperandsEqual)
182 IdenticalCalls.push_back(C);
185 if (IdenticalCalls.empty()) return;
187 // Eliminate duplicates, which could occur if we chose a value that is passed
188 // into a call site multiple times.
189 std::sort(IdenticalCalls.begin(), IdenticalCalls.end());
190 IdenticalCalls.erase(std::unique(IdenticalCalls.begin(),IdenticalCalls.end()),
191 IdenticalCalls.end());
193 // If the call reads memory, we must make sure that there are no stores
194 // between the calls in question.
196 // FIXME: This should use mod/ref information. What we really care about it
197 // whether an intervening instruction could modify memory that is read, not
200 if (MRB == AliasAnalysis::OnlyReadsMemory) {
201 DominatorSet &DomSetInfo = getAnalysis<DominatorSet>();
202 BasicBlock *CIBB = CI->getParent();
203 for (unsigned i = 0; i != IdenticalCalls.size(); ++i) {
204 CallInst *C = IdenticalCalls[i];
205 bool CantEqual = false;
207 if (DomSetInfo.dominates(CIBB, C->getParent())) {
208 // FIXME: we currently only handle the case where both calls are in the
210 if (CIBB != C->getParent()) {
213 Instruction *First = CI, *Second = C;
214 if (!DomSetInfo.dominates(CI, C))
215 std::swap(First, Second);
217 // Scan the instructions between the calls, checking for stores or
218 // calls to dangerous functions.
219 BasicBlock::iterator I = First;
220 for (++First; I != BasicBlock::iterator(Second); ++I) {
221 if (isa<StoreInst>(I)) {
222 // FIXME: We could use mod/ref information to make this much
226 } else if (CallInst *CI = dyn_cast<CallInst>(I)) {
227 if (CI->getCalledFunction() == 0 ||
228 !AA.onlyReadsMemory(CI->getCalledFunction())) {
232 } else if (I->mayWriteToMemory()) {
239 } else if (DomSetInfo.dominates(C->getParent(), CIBB)) {
240 // FIXME: We could implement this, but we don't for now.
243 // FIXME: if one doesn't dominate the other, we can't tell yet.
249 // This call does not produce the same value as the one in the query.
250 std::swap(IdenticalCalls[i--], IdenticalCalls.back());
251 IdenticalCalls.pop_back();
256 // Any calls that are identical and not destroyed will produce equal values!
257 for (unsigned i = 0, e = IdenticalCalls.size(); i != e; ++i)
258 RetVals.push_back(IdenticalCalls[i]);
261 // getEqualNumberNodes - Return nodes with the same value number as the
262 // specified Value. This fills in the argument vector with any equal values.
264 void LoadVN::getEqualNumberNodes(Value *V,
265 std::vector<Value*> &RetVals) const {
266 // If the alias analysis has any must alias information to share with us, we
267 // can definitely use it.
268 if (isa<PointerType>(V->getType()))
269 getAnalysis<AliasAnalysis>().getMustAliases(V, RetVals);
271 if (!isa<LoadInst>(V)) {
272 if (CallInst *CI = dyn_cast<CallInst>(V))
273 getCallEqualNumberNodes(CI, RetVals);
275 // Not a load instruction? Just chain to the base value numbering
276 // implementation to satisfy the request...
277 assert(&getAnalysis<ValueNumbering>() != (ValueNumbering*)this &&
278 "getAnalysis() returned this!");
280 return getAnalysis<ValueNumbering>().getEqualNumberNodes(V, RetVals);
283 // Volatile loads cannot be replaced with the value of other loads.
284 LoadInst *LI = cast<LoadInst>(V);
285 if (LI->isVolatile())
286 return getAnalysis<ValueNumbering>().getEqualNumberNodes(V, RetVals);
288 Value *LoadPtr = LI->getOperand(0);
289 BasicBlock *LoadBB = LI->getParent();
290 Function *F = LoadBB->getParent();
292 // Find out how many bytes of memory are loaded by the load instruction...
293 unsigned LoadSize = getAnalysis<TargetData>().getTypeSize(LI->getType());
294 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
296 // Figure out if the load is invalidated from the entry of the block it is in
297 // until the actual instruction. This scans the block backwards from LI. If
298 // we see any candidate load or store instructions, then we know that the
299 // candidates have the same value # as LI.
300 bool LoadInvalidatedInBBBefore = false;
301 for (BasicBlock::iterator I = LI; I != LoadBB->begin(); ) {
304 // If we run into an allocation of the value being loaded, then the
305 // contents are not initialized.
306 if (isa<AllocationInst>(I))
307 RetVals.push_back(UndefValue::get(LI->getType()));
309 // Otherwise, since this is the definition of what we are loading, this
310 // loaded value cannot occur before this block.
311 LoadInvalidatedInBBBefore = true;
313 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
314 // If this instruction is a candidate load before LI, we know there are no
315 // invalidating instructions between it and LI, so they have the same
317 if (LI->getOperand(0) == LoadPtr && !LI->isVolatile())
318 RetVals.push_back(I);
321 if (AA.getModRefInfo(I, LoadPtr, LoadSize) & AliasAnalysis::Mod) {
322 // If the invalidating instruction is a store, and its in our candidate
323 // set, then we can do store-load forwarding: the load has the same value
324 // # as the stored value.
325 if (StoreInst *SI = dyn_cast<StoreInst>(I))
326 if (SI->getOperand(1) == LoadPtr)
327 RetVals.push_back(I->getOperand(0));
329 LoadInvalidatedInBBBefore = true;
334 // Figure out if the load is invalidated between the load and the exit of the
335 // block it is defined in. While we are scanning the current basic block, if
336 // we see any candidate loads, then we know they have the same value # as LI.
338 bool LoadInvalidatedInBBAfter = false;
339 for (BasicBlock::iterator I = LI->getNext(); I != LoadBB->end(); ++I) {
340 // If this instruction is a load, then this instruction returns the same
342 if (isa<LoadInst>(I) && cast<LoadInst>(I)->getOperand(0) == LoadPtr)
343 RetVals.push_back(I);
345 if (AA.getModRefInfo(I, LoadPtr, LoadSize) & AliasAnalysis::Mod) {
346 LoadInvalidatedInBBAfter = true;
351 // If the pointer is clobbered on entry and on exit to the function, there is
352 // no need to do any global analysis at all.
353 if (LoadInvalidatedInBBBefore && LoadInvalidatedInBBAfter)
356 // Now that we know the value is not neccesarily killed on entry or exit to
357 // the BB, find out how many load and store instructions (to this location)
358 // live in each BB in the function.
360 std::map<BasicBlock*, unsigned> CandidateLoads;
361 std::set<BasicBlock*> CandidateStores;
363 for (Value::use_iterator UI = LoadPtr->use_begin(), UE = LoadPtr->use_end();
365 if (LoadInst *Cand = dyn_cast<LoadInst>(*UI)) {// Is a load of source?
366 if (Cand->getParent()->getParent() == F && // In the same function?
367 // Not in LI's block?
368 Cand->getParent() != LoadBB && !Cand->isVolatile())
369 ++CandidateLoads[Cand->getParent()]; // Got one.
370 } else if (StoreInst *Cand = dyn_cast<StoreInst>(*UI)) {
371 if (Cand->getParent()->getParent() == F && !Cand->isVolatile() &&
372 Cand->getOperand(1) == LoadPtr) // It's a store THROUGH the ptr.
373 CandidateStores.insert(Cand->getParent());
377 DominatorSet &DomSetInfo = getAnalysis<DominatorSet>();
379 // TransparentBlocks - For each basic block the load/store is alive across,
380 // figure out if the pointer is invalidated or not. If it is invalidated, the
381 // boolean is set to false, if it's not it is set to true. If we don't know
382 // yet, the entry is not in the map.
383 std::map<BasicBlock*, bool> TransparentBlocks;
385 // Loop over all of the basic blocks that also load the value. If the value
386 // is live across the CFG from the source to destination blocks, and if the
387 // value is not invalidated in either the source or destination blocks, add it
388 // to the equivalence sets.
389 for (std::map<BasicBlock*, unsigned>::iterator
390 I = CandidateLoads.begin(), E = CandidateLoads.end(); I != E; ++I) {
391 bool CantEqual = false;
393 // Right now we only can handle cases where one load dominates the other.
394 // FIXME: generalize this!
395 BasicBlock *BB1 = I->first, *BB2 = LoadBB;
396 if (DomSetInfo.dominates(BB1, BB2)) {
397 // The other load dominates LI. If the loaded value is killed entering
398 // the LoadBB block, we know the load is not live.
399 if (LoadInvalidatedInBBBefore)
401 } else if (DomSetInfo.dominates(BB2, BB1)) {
402 std::swap(BB1, BB2); // Canonicalize
403 // LI dominates the other load. If the loaded value is killed exiting
404 // the LoadBB block, we know the load is not live.
405 if (LoadInvalidatedInBBAfter)
408 // None of these loads can VN the same.
413 // Ok, at this point, we know that BB1 dominates BB2, and that there is
414 // nothing in the LI block that kills the loaded value. Check to see if
415 // the value is live across the CFG.
416 std::set<BasicBlock*> Visited;
417 for (pred_iterator PI = pred_begin(BB2), E = pred_end(BB2); PI!=E; ++PI)
418 if (!isPathTransparentTo(*PI, BB1, LoadPtr, LoadSize, AA,
419 Visited, TransparentBlocks)) {
420 // None of these loads can VN the same.
426 // If the loads can equal so far, scan the basic block that contains the
427 // loads under consideration to see if they are invalidated in the block.
428 // For any loads that are not invalidated, add them to the equivalence
431 unsigned NumLoads = I->second;
433 // If LI dominates the block in question, check to see if any of the
434 // loads in this block are invalidated before they are reached.
435 for (BasicBlock::iterator BBI = I->first->begin(); ; ++BBI) {
436 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
437 if (LI->getOperand(0) == LoadPtr && !LI->isVolatile()) {
438 // The load is in the set!
439 RetVals.push_back(BBI);
440 if (--NumLoads == 0) break; // Found last load to check.
442 } else if (AA.getModRefInfo(BBI, LoadPtr, LoadSize)
443 & AliasAnalysis::Mod) {
444 // If there is a modifying instruction, nothing below it will value
450 // If the block dominates LI, make sure that the loads in the block are
451 // not invalidated before the block ends.
452 BasicBlock::iterator BBI = I->first->end();
455 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
456 if (LI->getOperand(0) == LoadPtr && !LI->isVolatile()) {
457 // The load is the same as this load!
458 RetVals.push_back(BBI);
459 if (--NumLoads == 0) break; // Found all of the laods.
461 } else if (AA.getModRefInfo(BBI, LoadPtr, LoadSize)
462 & AliasAnalysis::Mod) {
463 // If there is a modifying instruction, nothing above it will value
472 // Handle candidate stores. If the loaded location is clobbered on entrance
473 // to the LoadBB, no store outside of the LoadBB can value number equal, so
475 if (LoadInvalidatedInBBBefore)
478 // Stores in the load-bb are handled above.
479 CandidateStores.erase(LoadBB);
481 for (std::set<BasicBlock*>::iterator I = CandidateStores.begin(),
482 E = CandidateStores.end(); I != E; ++I)
483 if (DomSetInfo.dominates(*I, LoadBB)) {
484 BasicBlock *StoreBB = *I;
486 // Check to see if the path from the store to the load is transparent
487 // w.r.t. the memory location.
488 bool CantEqual = false;
489 std::set<BasicBlock*> Visited;
490 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
492 if (!isPathTransparentTo(*PI, StoreBB, LoadPtr, LoadSize, AA,
493 Visited, TransparentBlocks)) {
494 // None of these stores can VN the same.
500 // Okay, the path from the store block to the load block is clear, and
501 // we know that there are no invalidating instructions from the start
502 // of the load block to the load itself. Now we just scan the store
505 BasicBlock::iterator BBI = StoreBB->end();
507 assert(BBI != StoreBB->begin() &&
508 "There is a store in this block of the pointer, but the store"
509 " doesn't mod the address being stored to?? Must be a bug in"
510 " the alias analysis implementation!");
512 if (AA.getModRefInfo(BBI, LoadPtr, LoadSize) & AliasAnalysis::Mod) {
513 // If the invalidating instruction is one of the candidates,
514 // then it provides the value the load loads.
515 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
516 if (SI->getOperand(1) == LoadPtr)
517 RetVals.push_back(SI->getOperand(0));