1 //===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements an analysis that determines, for a given memory
11 // operation, what preceding memory operations it depends on. It builds on
12 // alias analysis information, and tries to provide a lazy, caching interface to
13 // a common kind of alias information query.
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/AssumptionCache.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/MemoryBuiltins.h"
24 #include "llvm/Analysis/PHITransAddr.h"
25 #include "llvm/Analysis/OrderedBasicBlock.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/LLVMContext.h"
33 #include "llvm/IR/PredIteratorCache.h"
34 #include "llvm/Support/Debug.h"
37 #define DEBUG_TYPE "memdep"
39 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
40 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
41 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
43 STATISTIC(NumCacheNonLocalPtr,
44 "Number of fully cached non-local ptr responses");
45 STATISTIC(NumCacheDirtyNonLocalPtr,
46 "Number of cached, but dirty, non-local ptr responses");
47 STATISTIC(NumUncacheNonLocalPtr,
48 "Number of uncached non-local ptr responses");
49 STATISTIC(NumCacheCompleteNonLocalPtr,
50 "Number of block queries that were completely cached");
52 // Limit for the number of instructions to scan in a block.
54 static cl::opt<unsigned> BlockScanLimit(
55 "memdep-block-scan-limit", cl::Hidden, cl::init(100),
56 cl::desc("The number of instructions to scan in a block in memory "
57 "dependency analysis (default = 100)"));
59 // Limit on the number of memdep results to process.
60 static const unsigned int NumResultsLimit = 100;
62 char MemoryDependenceAnalysis::ID = 0;
64 // Register this pass...
65 INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep",
66 "Memory Dependence Analysis", false, true)
67 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
68 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
69 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep",
70 "Memory Dependence Analysis", false, true)
72 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
74 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry());
76 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
79 /// Clean up memory in between runs
80 void MemoryDependenceAnalysis::releaseMemory() {
83 NonLocalPointerDeps.clear();
84 ReverseLocalDeps.clear();
85 ReverseNonLocalDeps.clear();
86 ReverseNonLocalPtrDeps.clear();
90 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
92 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
94 AU.addRequired<AssumptionCacheTracker>();
95 AU.addRequiredTransitive<AliasAnalysis>();
98 bool MemoryDependenceAnalysis::runOnFunction(Function &F) {
99 AA = &getAnalysis<AliasAnalysis>();
100 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
101 DominatorTreeWrapperPass *DTWP =
102 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
103 DT = DTWP ? &DTWP->getDomTree() : nullptr;
107 /// RemoveFromReverseMap - This is a helper function that removes Val from
108 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
109 template <typename KeyTy>
110 static void RemoveFromReverseMap(DenseMap<Instruction*,
111 SmallPtrSet<KeyTy, 4> > &ReverseMap,
112 Instruction *Inst, KeyTy Val) {
113 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
114 InstIt = ReverseMap.find(Inst);
115 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
116 bool Found = InstIt->second.erase(Val);
117 assert(Found && "Invalid reverse map!"); (void)Found;
118 if (InstIt->second.empty())
119 ReverseMap.erase(InstIt);
122 /// GetLocation - If the given instruction references a specific memory
123 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
124 /// Return a ModRefInfo value describing the general behavior of the
126 static ModRefInfo GetLocation(const Instruction *Inst, MemoryLocation &Loc,
128 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
129 if (LI->isUnordered()) {
130 Loc = MemoryLocation::get(LI);
133 if (LI->getOrdering() == Monotonic) {
134 Loc = MemoryLocation::get(LI);
137 Loc = MemoryLocation();
141 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
142 if (SI->isUnordered()) {
143 Loc = MemoryLocation::get(SI);
146 if (SI->getOrdering() == Monotonic) {
147 Loc = MemoryLocation::get(SI);
150 Loc = MemoryLocation();
154 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
155 Loc = MemoryLocation::get(V);
159 if (const CallInst *CI = isFreeCall(Inst, AA->getTargetLibraryInfo())) {
160 // calls to free() deallocate the entire structure
161 Loc = MemoryLocation(CI->getArgOperand(0));
165 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
168 switch (II->getIntrinsicID()) {
169 case Intrinsic::lifetime_start:
170 case Intrinsic::lifetime_end:
171 case Intrinsic::invariant_start:
172 II->getAAMetadata(AAInfo);
173 Loc = MemoryLocation(
174 II->getArgOperand(1),
175 cast<ConstantInt>(II->getArgOperand(0))->getZExtValue(), AAInfo);
176 // These intrinsics don't really modify the memory, but returning Mod
177 // will allow them to be handled conservatively.
179 case Intrinsic::invariant_end:
180 II->getAAMetadata(AAInfo);
181 Loc = MemoryLocation(
182 II->getArgOperand(2),
183 cast<ConstantInt>(II->getArgOperand(1))->getZExtValue(), AAInfo);
184 // These intrinsics don't really modify the memory, but returning Mod
185 // will allow them to be handled conservatively.
192 // Otherwise, just do the coarse-grained thing that always works.
193 if (Inst->mayWriteToMemory())
195 if (Inst->mayReadFromMemory())
200 /// getCallSiteDependencyFrom - Private helper for finding the local
201 /// dependencies of a call site.
202 MemDepResult MemoryDependenceAnalysis::
203 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
204 BasicBlock::iterator ScanIt, BasicBlock *BB) {
205 unsigned Limit = BlockScanLimit;
207 // Walk backwards through the block, looking for dependencies
208 while (ScanIt != BB->begin()) {
209 // Limit the amount of scanning we do so we don't end up with quadratic
210 // running time on extreme testcases.
213 return MemDepResult::getUnknown();
215 Instruction *Inst = --ScanIt;
217 // If this inst is a memory op, get the pointer it accessed
219 ModRefInfo MR = GetLocation(Inst, Loc, AA);
221 // A simple instruction.
222 if (AA->getModRefInfo(CS, Loc) != MRI_NoModRef)
223 return MemDepResult::getClobber(Inst);
227 if (auto InstCS = CallSite(Inst)) {
228 // Debug intrinsics don't cause dependences.
229 if (isa<DbgInfoIntrinsic>(Inst)) continue;
230 // If these two calls do not interfere, look past it.
231 switch (AA->getModRefInfo(CS, InstCS)) {
233 // If the two calls are the same, return InstCS as a Def, so that
234 // CS can be found redundant and eliminated.
235 if (isReadOnlyCall && !(MR & MRI_Mod) &&
236 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
237 return MemDepResult::getDef(Inst);
239 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
243 return MemDepResult::getClobber(Inst);
247 // If we could not obtain a pointer for the instruction and the instruction
248 // touches memory then assume that this is a dependency.
249 if (MR != MRI_NoModRef)
250 return MemDepResult::getClobber(Inst);
253 // No dependence found. If this is the entry block of the function, it is
254 // unknown, otherwise it is non-local.
255 if (BB != &BB->getParent()->getEntryBlock())
256 return MemDepResult::getNonLocal();
257 return MemDepResult::getNonFuncLocal();
260 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
261 /// would fully overlap MemLoc if done as a wider legal integer load.
263 /// MemLocBase, MemLocOffset are lazily computed here the first time the
264 /// base/offs of memloc is needed.
265 static bool isLoadLoadClobberIfExtendedToFullWidth(const MemoryLocation &MemLoc,
266 const Value *&MemLocBase,
268 const LoadInst *LI) {
269 const DataLayout &DL = LI->getModule()->getDataLayout();
271 // If we haven't already computed the base/offset of MemLoc, do so now.
273 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, DL);
275 unsigned Size = MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
276 MemLocBase, MemLocOffs, MemLoc.Size, LI);
280 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
281 /// looks at a memory location for a load (specified by MemLocBase, Offs,
282 /// and Size) and compares it against a load. If the specified load could
283 /// be safely widened to a larger integer load that is 1) still efficient,
284 /// 2) safe for the target, and 3) would provide the specified memory
285 /// location value, then this function returns the size in bytes of the
286 /// load width to use. If not, this returns zero.
287 unsigned MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
288 const Value *MemLocBase, int64_t MemLocOffs, unsigned MemLocSize,
289 const LoadInst *LI) {
290 // We can only extend simple integer loads.
291 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
293 // Load widening is hostile to ThreadSanitizer: it may cause false positives
294 // or make the reports more cryptic (access sizes are wrong).
295 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
298 const DataLayout &DL = LI->getModule()->getDataLayout();
300 // Get the base of this load.
302 const Value *LIBase =
303 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, DL);
305 // If the two pointers are not based on the same pointer, we can't tell that
307 if (LIBase != MemLocBase) return 0;
309 // Okay, the two values are based on the same pointer, but returned as
310 // no-alias. This happens when we have things like two byte loads at "P+1"
311 // and "P+3". Check to see if increasing the size of the "LI" load up to its
312 // alignment (or the largest native integer type) will allow us to load all
313 // the bits required by MemLoc.
315 // If MemLoc is before LI, then no widening of LI will help us out.
316 if (MemLocOffs < LIOffs) return 0;
318 // Get the alignment of the load in bytes. We assume that it is safe to load
319 // any legal integer up to this size without a problem. For example, if we're
320 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
321 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
323 unsigned LoadAlign = LI->getAlignment();
325 int64_t MemLocEnd = MemLocOffs+MemLocSize;
327 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
328 if (LIOffs+LoadAlign < MemLocEnd) return 0;
330 // This is the size of the load to try. Start with the next larger power of
332 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
333 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
336 // If this load size is bigger than our known alignment or would not fit
337 // into a native integer register, then we fail.
338 if (NewLoadByteSize > LoadAlign ||
339 !DL.fitsInLegalInteger(NewLoadByteSize*8))
342 if (LIOffs + NewLoadByteSize > MemLocEnd &&
343 LI->getParent()->getParent()->hasFnAttribute(
344 Attribute::SanitizeAddress))
345 // We will be reading past the location accessed by the original program.
346 // While this is safe in a regular build, Address Safety analysis tools
347 // may start reporting false warnings. So, don't do widening.
350 // If a load of this width would include all of MemLoc, then we succeed.
351 if (LIOffs+NewLoadByteSize >= MemLocEnd)
352 return NewLoadByteSize;
354 NewLoadByteSize <<= 1;
358 static bool isVolatile(Instruction *Inst) {
359 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
360 return LI->isVolatile();
361 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
362 return SI->isVolatile();
363 else if (AtomicCmpXchgInst *AI = dyn_cast<AtomicCmpXchgInst>(Inst))
364 return AI->isVolatile();
369 /// getPointerDependencyFrom - Return the instruction on which a memory
370 /// location depends. If isLoad is true, this routine ignores may-aliases with
371 /// read-only operations. If isLoad is false, this routine ignores may-aliases
372 /// with reads from read-only locations. If possible, pass the query
373 /// instruction as well; this function may take advantage of the metadata
374 /// annotated to the query instruction to refine the result.
375 MemDepResult MemoryDependenceAnalysis::getPointerDependencyFrom(
376 const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
377 BasicBlock *BB, Instruction *QueryInst) {
379 const Value *MemLocBase = nullptr;
380 int64_t MemLocOffset = 0;
381 unsigned Limit = BlockScanLimit;
382 bool isInvariantLoad = false;
384 // We must be careful with atomic accesses, as they may allow another thread
385 // to touch this location, cloberring it. We are conservative: if the
386 // QueryInst is not a simple (non-atomic) memory access, we automatically
387 // return getClobber.
388 // If it is simple, we know based on the results of
389 // "Compiler testing via a theory of sound optimisations in the C11/C++11
390 // memory model" in PLDI 2013, that a non-atomic location can only be
391 // clobbered between a pair of a release and an acquire action, with no
392 // access to the location in between.
393 // Here is an example for giving the general intuition behind this rule.
394 // In the following code:
396 // release action; [1]
397 // acquire action; [4]
399 // It is unsafe to replace %val by 0 because another thread may be running:
400 // acquire action; [2]
402 // release action; [3]
403 // with synchronization from 1 to 2 and from 3 to 4, resulting in %val
404 // being 42. A key property of this program however is that if either
405 // 1 or 4 were missing, there would be a race between the store of 42
406 // either the store of 0 or the load (making the whole progam racy).
407 // The paper mentionned above shows that the same property is respected
408 // by every program that can detect any optimisation of that kind: either
409 // it is racy (undefined) or there is a release followed by an acquire
410 // between the pair of accesses under consideration.
412 // If the load is invariant, we "know" that it doesn't alias *any* write. We
413 // do want to respect mustalias results since defs are useful for value
414 // forwarding, but any mayalias write can be assumed to be noalias.
415 // Arguably, this logic should be pushed inside AliasAnalysis itself.
416 if (isLoad && QueryInst) {
417 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
418 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
419 isInvariantLoad = true;
422 const DataLayout &DL = BB->getModule()->getDataLayout();
424 // Create a numbered basic block to lazily compute and cache instruction
425 // positions inside a BB. This is used to provide fast queries for relative
426 // position between two instructions in a BB and can be used by
427 // AliasAnalysis::callCapturesBefore.
428 OrderedBasicBlock OBB(BB);
430 // Walk backwards through the basic block, looking for dependencies.
431 while (ScanIt != BB->begin()) {
432 Instruction *Inst = --ScanIt;
434 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
435 // Debug intrinsics don't (and can't) cause dependencies.
436 if (isa<DbgInfoIntrinsic>(II)) continue;
438 // Limit the amount of scanning we do so we don't end up with quadratic
439 // running time on extreme testcases.
442 return MemDepResult::getUnknown();
444 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
445 // If we reach a lifetime begin or end marker, then the query ends here
446 // because the value is undefined.
447 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
448 // FIXME: This only considers queries directly on the invariant-tagged
449 // pointer, not on query pointers that are indexed off of them. It'd
450 // be nice to handle that at some point (the right approach is to use
451 // GetPointerBaseWithConstantOffset).
452 if (AA->isMustAlias(MemoryLocation(II->getArgOperand(1)), MemLoc))
453 return MemDepResult::getDef(II);
458 // Values depend on loads if the pointers are must aliased. This means that
459 // a load depends on another must aliased load from the same value.
460 // One exception is atomic loads: a value can depend on an atomic load that it
461 // does not alias with when this atomic load indicates that another thread may
462 // be accessing the location.
463 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
465 // While volatile access cannot be eliminated, they do not have to clobber
466 // non-aliasing locations, as normal accesses, for example, can be safely
467 // reordered with volatile accesses.
468 if (LI->isVolatile()) {
470 // Original QueryInst *may* be volatile
471 return MemDepResult::getClobber(LI);
472 if (isVolatile(QueryInst))
473 // Ordering required if QueryInst is itself volatile
474 return MemDepResult::getClobber(LI);
475 // Otherwise, volatile doesn't imply any special ordering
478 // Atomic loads have complications involved.
479 // A Monotonic (or higher) load is OK if the query inst is itself not atomic.
480 // FIXME: This is overly conservative.
481 if (LI->isAtomic() && LI->getOrdering() > Unordered) {
483 return MemDepResult::getClobber(LI);
484 if (LI->getOrdering() != Monotonic)
485 return MemDepResult::getClobber(LI);
486 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
487 if (!QueryLI->isSimple())
488 return MemDepResult::getClobber(LI);
489 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
490 if (!QuerySI->isSimple())
491 return MemDepResult::getClobber(LI);
492 } else if (QueryInst->mayReadOrWriteMemory()) {
493 return MemDepResult::getClobber(LI);
497 MemoryLocation LoadLoc = MemoryLocation::get(LI);
499 // If we found a pointer, check if it could be the same as our pointer.
500 AliasResult R = AA->alias(LoadLoc, MemLoc);
504 // If this is an over-aligned integer load (for example,
505 // "load i8* %P, align 4") see if it would obviously overlap with the
506 // queried location if widened to a larger load (e.g. if the queried
507 // location is 1 byte at P+1). If so, return it as a load/load
508 // clobber result, allowing the client to decide to widen the load if
510 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
511 if (LI->getAlignment() * 8 > ITy->getPrimitiveSizeInBits() &&
512 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
514 return MemDepResult::getClobber(Inst);
519 // Must aliased loads are defs of each other.
521 return MemDepResult::getDef(Inst);
523 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
524 // in terms of clobbering loads, but since it does this by looking
525 // at the clobbering load directly, it doesn't know about any
526 // phi translation that may have happened along the way.
528 // If we have a partial alias, then return this as a clobber for the
530 if (R == PartialAlias)
531 return MemDepResult::getClobber(Inst);
534 // Random may-alias loads don't depend on each other without a
539 // Stores don't depend on other no-aliased accesses.
543 // Stores don't alias loads from read-only memory.
544 if (AA->pointsToConstantMemory(LoadLoc))
547 // Stores depend on may/must aliased loads.
548 return MemDepResult::getDef(Inst);
551 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
552 // Atomic stores have complications involved.
553 // A Monotonic store is OK if the query inst is itself not atomic.
554 // FIXME: This is overly conservative.
555 if (!SI->isUnordered()) {
557 return MemDepResult::getClobber(SI);
558 if (SI->getOrdering() != Monotonic)
559 return MemDepResult::getClobber(SI);
560 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
561 if (!QueryLI->isSimple())
562 return MemDepResult::getClobber(SI);
563 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
564 if (!QuerySI->isSimple())
565 return MemDepResult::getClobber(SI);
566 } else if (QueryInst->mayReadOrWriteMemory()) {
567 return MemDepResult::getClobber(SI);
571 // FIXME: this is overly conservative.
572 // While volatile access cannot be eliminated, they do not have to clobber
573 // non-aliasing locations, as normal accesses can for example be reordered
574 // with volatile accesses.
575 if (SI->isVolatile())
576 return MemDepResult::getClobber(SI);
578 // If alias analysis can tell that this store is guaranteed to not modify
579 // the query pointer, ignore it. Use getModRefInfo to handle cases where
580 // the query pointer points to constant memory etc.
581 if (AA->getModRefInfo(SI, MemLoc) == MRI_NoModRef)
584 // Ok, this store might clobber the query pointer. Check to see if it is
585 // a must alias: in this case, we want to return this as a def.
586 MemoryLocation StoreLoc = MemoryLocation::get(SI);
588 // If we found a pointer, check if it could be the same as our pointer.
589 AliasResult R = AA->alias(StoreLoc, MemLoc);
594 return MemDepResult::getDef(Inst);
597 return MemDepResult::getClobber(Inst);
600 // If this is an allocation, and if we know that the accessed pointer is to
601 // the allocation, return Def. This means that there is no dependence and
602 // the access can be optimized based on that. For example, a load could
604 // Note: Only determine this to be a malloc if Inst is the malloc call, not
605 // a subsequent bitcast of the malloc call result. There can be stores to
606 // the malloced memory between the malloc call and its bitcast uses, and we
607 // need to continue scanning until the malloc call.
608 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
609 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
610 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
612 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
613 return MemDepResult::getDef(Inst);
616 // Be conservative if the accessed pointer may alias the allocation.
617 if (AA->alias(Inst, AccessPtr) != NoAlias)
618 return MemDepResult::getClobber(Inst);
619 // If the allocation is not aliased and does not read memory (like
620 // strdup), it is safe to ignore.
621 if (isa<AllocaInst>(Inst) ||
622 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
629 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
630 ModRefInfo MR = AA->getModRefInfo(Inst, MemLoc);
631 // If necessary, perform additional analysis.
632 if (MR == MRI_ModRef)
633 MR = AA->callCapturesBefore(Inst, MemLoc, DT, &OBB);
636 // If the call has no effect on the queried pointer, just ignore it.
639 return MemDepResult::getClobber(Inst);
641 // If the call is known to never store to the pointer, and if this is a
642 // load query, we can safely ignore it (scan past it).
646 // Otherwise, there is a potential dependence. Return a clobber.
647 return MemDepResult::getClobber(Inst);
651 // No dependence found. If this is the entry block of the function, it is
652 // unknown, otherwise it is non-local.
653 if (BB != &BB->getParent()->getEntryBlock())
654 return MemDepResult::getNonLocal();
655 return MemDepResult::getNonFuncLocal();
658 /// getDependency - Return the instruction on which a memory operation
660 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
661 Instruction *ScanPos = QueryInst;
663 // Check for a cached result
664 MemDepResult &LocalCache = LocalDeps[QueryInst];
666 // If the cached entry is non-dirty, just return it. Note that this depends
667 // on MemDepResult's default constructing to 'dirty'.
668 if (!LocalCache.isDirty())
671 // Otherwise, if we have a dirty entry, we know we can start the scan at that
672 // instruction, which may save us some work.
673 if (Instruction *Inst = LocalCache.getInst()) {
676 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
679 BasicBlock *QueryParent = QueryInst->getParent();
682 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
683 // No dependence found. If this is the entry block of the function, it is
684 // unknown, otherwise it is non-local.
685 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
686 LocalCache = MemDepResult::getNonLocal();
688 LocalCache = MemDepResult::getNonFuncLocal();
690 MemoryLocation MemLoc;
691 ModRefInfo MR = GetLocation(QueryInst, MemLoc, AA);
693 // If we can do a pointer scan, make it happen.
694 bool isLoad = !(MR & MRI_Mod);
695 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
696 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
698 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
699 QueryParent, QueryInst);
700 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
701 CallSite QueryCS(QueryInst);
702 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
703 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
706 // Non-memory instruction.
707 LocalCache = MemDepResult::getUnknown();
710 // Remember the result!
711 if (Instruction *I = LocalCache.getInst())
712 ReverseLocalDeps[I].insert(QueryInst);
718 /// AssertSorted - This method is used when -debug is specified to verify that
719 /// cache arrays are properly kept sorted.
720 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
722 if (Count == -1) Count = Cache.size();
723 if (Count == 0) return;
725 for (unsigned i = 1; i != unsigned(Count); ++i)
726 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
730 /// getNonLocalCallDependency - Perform a full dependency query for the
731 /// specified call, returning the set of blocks that the value is
732 /// potentially live across. The returned set of results will include a
733 /// "NonLocal" result for all blocks where the value is live across.
735 /// This method assumes the instruction returns a "NonLocal" dependency
736 /// within its own block.
738 /// This returns a reference to an internal data structure that may be
739 /// invalidated on the next non-local query or when an instruction is
740 /// removed. Clients must copy this data if they want it around longer than
742 const MemoryDependenceAnalysis::NonLocalDepInfo &
743 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
744 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
745 "getNonLocalCallDependency should only be used on calls with non-local deps!");
746 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
747 NonLocalDepInfo &Cache = CacheP.first;
749 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
750 /// the cached case, this can happen due to instructions being deleted etc. In
751 /// the uncached case, this starts out as the set of predecessors we care
753 SmallVector<BasicBlock*, 32> DirtyBlocks;
755 if (!Cache.empty()) {
756 // Okay, we have a cache entry. If we know it is not dirty, just return it
757 // with no computation.
758 if (!CacheP.second) {
763 // If we already have a partially computed set of results, scan them to
764 // determine what is dirty, seeding our initial DirtyBlocks worklist.
765 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
767 if (I->getResult().isDirty())
768 DirtyBlocks.push_back(I->getBB());
770 // Sort the cache so that we can do fast binary search lookups below.
771 std::sort(Cache.begin(), Cache.end());
773 ++NumCacheDirtyNonLocal;
774 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
775 // << Cache.size() << " cached: " << *QueryInst;
777 // Seed DirtyBlocks with each of the preds of QueryInst's block.
778 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
779 for (BasicBlock *Pred : PredCache.get(QueryBB))
780 DirtyBlocks.push_back(Pred);
781 ++NumUncacheNonLocal;
784 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
785 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
787 SmallPtrSet<BasicBlock*, 64> Visited;
789 unsigned NumSortedEntries = Cache.size();
790 DEBUG(AssertSorted(Cache));
792 // Iterate while we still have blocks to update.
793 while (!DirtyBlocks.empty()) {
794 BasicBlock *DirtyBB = DirtyBlocks.back();
795 DirtyBlocks.pop_back();
797 // Already processed this block?
798 if (!Visited.insert(DirtyBB).second)
801 // Do a binary search to see if we already have an entry for this block in
802 // the cache set. If so, find it.
803 DEBUG(AssertSorted(Cache, NumSortedEntries));
804 NonLocalDepInfo::iterator Entry =
805 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
806 NonLocalDepEntry(DirtyBB));
807 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
810 NonLocalDepEntry *ExistingResult = nullptr;
811 if (Entry != Cache.begin()+NumSortedEntries &&
812 Entry->getBB() == DirtyBB) {
813 // If we already have an entry, and if it isn't already dirty, the block
815 if (!Entry->getResult().isDirty())
818 // Otherwise, remember this slot so we can update the value.
819 ExistingResult = &*Entry;
822 // If the dirty entry has a pointer, start scanning from it so we don't have
823 // to rescan the entire block.
824 BasicBlock::iterator ScanPos = DirtyBB->end();
825 if (ExistingResult) {
826 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
828 // We're removing QueryInst's use of Inst.
829 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
830 QueryCS.getInstruction());
834 // Find out if this block has a local dependency for QueryInst.
837 if (ScanPos != DirtyBB->begin()) {
838 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
839 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
840 // No dependence found. If this is the entry block of the function, it is
841 // a clobber, otherwise it is unknown.
842 Dep = MemDepResult::getNonLocal();
844 Dep = MemDepResult::getNonFuncLocal();
847 // If we had a dirty entry for the block, update it. Otherwise, just add
850 ExistingResult->setResult(Dep);
852 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
854 // If the block has a dependency (i.e. it isn't completely transparent to
855 // the value), remember the association!
856 if (!Dep.isNonLocal()) {
857 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
858 // update this when we remove instructions.
859 if (Instruction *Inst = Dep.getInst())
860 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
863 // If the block *is* completely transparent to the load, we need to check
864 // the predecessors of this block. Add them to our worklist.
865 for (BasicBlock *Pred : PredCache.get(DirtyBB))
866 DirtyBlocks.push_back(Pred);
873 /// getNonLocalPointerDependency - Perform a full dependency query for an
874 /// access to the specified (non-volatile) memory location, returning the
875 /// set of instructions that either define or clobber the value.
877 /// This method assumes the pointer has a "NonLocal" dependency within its
880 void MemoryDependenceAnalysis::
881 getNonLocalPointerDependency(Instruction *QueryInst,
882 SmallVectorImpl<NonLocalDepResult> &Result) {
883 const MemoryLocation Loc = MemoryLocation::get(QueryInst);
884 bool isLoad = isa<LoadInst>(QueryInst);
885 BasicBlock *FromBB = QueryInst->getParent();
888 assert(Loc.Ptr->getType()->isPointerTy() &&
889 "Can't get pointer deps of a non-pointer!");
892 // This routine does not expect to deal with volatile instructions.
893 // Doing so would require piping through the QueryInst all the way through.
894 // TODO: volatiles can't be elided, but they can be reordered with other
895 // non-volatile accesses.
897 // We currently give up on any instruction which is ordered, but we do handle
898 // atomic instructions which are unordered.
899 // TODO: Handle ordered instructions
900 auto isOrdered = [](Instruction *Inst) {
901 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
902 return !LI->isUnordered();
903 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
904 return !SI->isUnordered();
908 if (isVolatile(QueryInst) || isOrdered(QueryInst)) {
909 Result.push_back(NonLocalDepResult(FromBB,
910 MemDepResult::getUnknown(),
911 const_cast<Value *>(Loc.Ptr)));
914 const DataLayout &DL = FromBB->getModule()->getDataLayout();
915 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, AC);
917 // This is the set of blocks we've inspected, and the pointer we consider in
918 // each block. Because of critical edges, we currently bail out if querying
919 // a block with multiple different pointers. This can happen during PHI
921 DenseMap<BasicBlock*, Value*> Visited;
922 if (!getNonLocalPointerDepFromBB(QueryInst, Address, Loc, isLoad, FromBB,
923 Result, Visited, true))
926 Result.push_back(NonLocalDepResult(FromBB,
927 MemDepResult::getUnknown(),
928 const_cast<Value *>(Loc.Ptr)));
931 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
932 /// Pointer/PointeeSize using either cached information in Cache or by doing a
933 /// lookup (which may use dirty cache info if available). If we do a lookup,
934 /// add the result to the cache.
935 MemDepResult MemoryDependenceAnalysis::GetNonLocalInfoForBlock(
936 Instruction *QueryInst, const MemoryLocation &Loc, bool isLoad,
937 BasicBlock *BB, NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
939 // Do a binary search to see if we already have an entry for this block in
940 // the cache set. If so, find it.
941 NonLocalDepInfo::iterator Entry =
942 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
943 NonLocalDepEntry(BB));
944 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
947 NonLocalDepEntry *ExistingResult = nullptr;
948 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
949 ExistingResult = &*Entry;
951 // If we have a cached entry, and it is non-dirty, use it as the value for
953 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
954 ++NumCacheNonLocalPtr;
955 return ExistingResult->getResult();
958 // Otherwise, we have to scan for the value. If we have a dirty cache
959 // entry, start scanning from its position, otherwise we scan from the end
961 BasicBlock::iterator ScanPos = BB->end();
962 if (ExistingResult && ExistingResult->getResult().getInst()) {
963 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
964 "Instruction invalidated?");
965 ++NumCacheDirtyNonLocalPtr;
966 ScanPos = ExistingResult->getResult().getInst();
968 // Eliminating the dirty entry from 'Cache', so update the reverse info.
969 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
970 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
972 ++NumUncacheNonLocalPtr;
975 // Scan the block for the dependency.
976 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB,
979 // If we had a dirty entry for the block, update it. Otherwise, just add
982 ExistingResult->setResult(Dep);
984 Cache->push_back(NonLocalDepEntry(BB, Dep));
986 // If the block has a dependency (i.e. it isn't completely transparent to
987 // the value), remember the reverse association because we just added it
989 if (!Dep.isDef() && !Dep.isClobber())
992 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
993 // update MemDep when we remove instructions.
994 Instruction *Inst = Dep.getInst();
995 assert(Inst && "Didn't depend on anything?");
996 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
997 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
1001 /// SortNonLocalDepInfoCache - Sort the NonLocalDepInfo cache, given a certain
1002 /// number of elements in the array that are already properly ordered. This is
1003 /// optimized for the case when only a few entries are added.
1005 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
1006 unsigned NumSortedEntries) {
1007 switch (Cache.size() - NumSortedEntries) {
1009 // done, no new entries.
1012 // Two new entries, insert the last one into place.
1013 NonLocalDepEntry Val = Cache.back();
1015 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
1016 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
1017 Cache.insert(Entry, Val);
1021 // One new entry, Just insert the new value at the appropriate position.
1022 if (Cache.size() != 1) {
1023 NonLocalDepEntry Val = Cache.back();
1025 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
1026 std::upper_bound(Cache.begin(), Cache.end(), Val);
1027 Cache.insert(Entry, Val);
1031 // Added many values, do a full scale sort.
1032 std::sort(Cache.begin(), Cache.end());
1037 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
1038 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
1039 /// results to the results vector and keep track of which blocks are visited in
1042 /// This has special behavior for the first block queries (when SkipFirstBlock
1043 /// is true). In this special case, it ignores the contents of the specified
1044 /// block and starts returning dependence info for its predecessors.
1046 /// This function returns false on success, or true to indicate that it could
1047 /// not compute dependence information for some reason. This should be treated
1048 /// as a clobber dependence on the first instruction in the predecessor block.
1049 bool MemoryDependenceAnalysis::getNonLocalPointerDepFromBB(
1050 Instruction *QueryInst, const PHITransAddr &Pointer,
1051 const MemoryLocation &Loc, bool isLoad, BasicBlock *StartBB,
1052 SmallVectorImpl<NonLocalDepResult> &Result,
1053 DenseMap<BasicBlock *, Value *> &Visited, bool SkipFirstBlock) {
1054 // Look up the cached info for Pointer.
1055 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
1057 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
1058 // CacheKey, this value will be inserted as the associated value. Otherwise,
1059 // it'll be ignored, and we'll have to check to see if the cached size and
1060 // aa tags are consistent with the current query.
1061 NonLocalPointerInfo InitialNLPI;
1062 InitialNLPI.Size = Loc.Size;
1063 InitialNLPI.AATags = Loc.AATags;
1065 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
1066 // already have one.
1067 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
1068 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
1069 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
1071 // If we already have a cache entry for this CacheKey, we may need to do some
1072 // work to reconcile the cache entry and the current query.
1074 if (CacheInfo->Size < Loc.Size) {
1075 // The query's Size is greater than the cached one. Throw out the
1076 // cached data and proceed with the query at the greater size.
1077 CacheInfo->Pair = BBSkipFirstBlockPair();
1078 CacheInfo->Size = Loc.Size;
1079 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1080 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1081 if (Instruction *Inst = DI->getResult().getInst())
1082 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1083 CacheInfo->NonLocalDeps.clear();
1084 } else if (CacheInfo->Size > Loc.Size) {
1085 // This query's Size is less than the cached one. Conservatively restart
1086 // the query using the greater size.
1087 return getNonLocalPointerDepFromBB(QueryInst, Pointer,
1088 Loc.getWithNewSize(CacheInfo->Size),
1089 isLoad, StartBB, Result, Visited,
1093 // If the query's AATags are inconsistent with the cached one,
1094 // conservatively throw out the cached data and restart the query with
1095 // no tag if needed.
1096 if (CacheInfo->AATags != Loc.AATags) {
1097 if (CacheInfo->AATags) {
1098 CacheInfo->Pair = BBSkipFirstBlockPair();
1099 CacheInfo->AATags = AAMDNodes();
1100 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1101 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1102 if (Instruction *Inst = DI->getResult().getInst())
1103 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1104 CacheInfo->NonLocalDeps.clear();
1107 return getNonLocalPointerDepFromBB(QueryInst,
1108 Pointer, Loc.getWithoutAATags(),
1109 isLoad, StartBB, Result, Visited,
1114 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1116 // If we have valid cached information for exactly the block we are
1117 // investigating, just return it with no recomputation.
1118 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1119 // We have a fully cached result for this query then we can just return the
1120 // cached results and populate the visited set. However, we have to verify
1121 // that we don't already have conflicting results for these blocks. Check
1122 // to ensure that if a block in the results set is in the visited set that
1123 // it was for the same pointer query.
1124 if (!Visited.empty()) {
1125 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1127 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
1128 if (VI == Visited.end() || VI->second == Pointer.getAddr())
1131 // We have a pointer mismatch in a block. Just return clobber, saying
1132 // that something was clobbered in this result. We could also do a
1133 // non-fully cached query, but there is little point in doing this.
1138 Value *Addr = Pointer.getAddr();
1139 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1141 Visited.insert(std::make_pair(I->getBB(), Addr));
1142 if (I->getResult().isNonLocal()) {
1147 Result.push_back(NonLocalDepResult(I->getBB(),
1148 MemDepResult::getUnknown(),
1150 } else if (DT->isReachableFromEntry(I->getBB())) {
1151 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1154 ++NumCacheCompleteNonLocalPtr;
1158 // Otherwise, either this is a new block, a block with an invalid cache
1159 // pointer or one that we're about to invalidate by putting more info into it
1160 // than its valid cache info. If empty, the result will be valid cache info,
1161 // otherwise it isn't.
1163 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1165 CacheInfo->Pair = BBSkipFirstBlockPair();
1167 SmallVector<BasicBlock*, 32> Worklist;
1168 Worklist.push_back(StartBB);
1170 // PredList used inside loop.
1171 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1173 // Keep track of the entries that we know are sorted. Previously cached
1174 // entries will all be sorted. The entries we add we only sort on demand (we
1175 // don't insert every element into its sorted position). We know that we
1176 // won't get any reuse from currently inserted values, because we don't
1177 // revisit blocks after we insert info for them.
1178 unsigned NumSortedEntries = Cache->size();
1179 DEBUG(AssertSorted(*Cache));
1181 while (!Worklist.empty()) {
1182 BasicBlock *BB = Worklist.pop_back_val();
1184 // If we do process a large number of blocks it becomes very expensive and
1185 // likely it isn't worth worrying about
1186 if (Result.size() > NumResultsLimit) {
1188 // Sort it now (if needed) so that recursive invocations of
1189 // getNonLocalPointerDepFromBB and other routines that could reuse the
1190 // cache value will only see properly sorted cache arrays.
1191 if (Cache && NumSortedEntries != Cache->size()) {
1192 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1194 // Since we bail out, the "Cache" set won't contain all of the
1195 // results for the query. This is ok (we can still use it to accelerate
1196 // specific block queries) but we can't do the fastpath "return all
1197 // results from the set". Clear out the indicator for this.
1198 CacheInfo->Pair = BBSkipFirstBlockPair();
1202 // Skip the first block if we have it.
1203 if (!SkipFirstBlock) {
1204 // Analyze the dependency of *Pointer in FromBB. See if we already have
1206 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1208 // Get the dependency info for Pointer in BB. If we have cached
1209 // information, we will use it, otherwise we compute it.
1210 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1211 MemDepResult Dep = GetNonLocalInfoForBlock(QueryInst,
1212 Loc, isLoad, BB, Cache,
1215 // If we got a Def or Clobber, add this to the list of results.
1216 if (!Dep.isNonLocal()) {
1218 Result.push_back(NonLocalDepResult(BB,
1219 MemDepResult::getUnknown(),
1220 Pointer.getAddr()));
1222 } else if (DT->isReachableFromEntry(BB)) {
1223 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1229 // If 'Pointer' is an instruction defined in this block, then we need to do
1230 // phi translation to change it into a value live in the predecessor block.
1231 // If not, we just add the predecessors to the worklist and scan them with
1232 // the same Pointer.
1233 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1234 SkipFirstBlock = false;
1235 SmallVector<BasicBlock*, 16> NewBlocks;
1236 for (BasicBlock *Pred : PredCache.get(BB)) {
1237 // Verify that we haven't looked at this block yet.
1238 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1239 InsertRes = Visited.insert(std::make_pair(Pred, Pointer.getAddr()));
1240 if (InsertRes.second) {
1241 // First time we've looked at *PI.
1242 NewBlocks.push_back(Pred);
1246 // If we have seen this block before, but it was with a different
1247 // pointer then we have a phi translation failure and we have to treat
1248 // this as a clobber.
1249 if (InsertRes.first->second != Pointer.getAddr()) {
1250 // Make sure to clean up the Visited map before continuing on to
1251 // PredTranslationFailure.
1252 for (unsigned i = 0; i < NewBlocks.size(); i++)
1253 Visited.erase(NewBlocks[i]);
1254 goto PredTranslationFailure;
1257 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1261 // We do need to do phi translation, if we know ahead of time we can't phi
1262 // translate this value, don't even try.
1263 if (!Pointer.IsPotentiallyPHITranslatable())
1264 goto PredTranslationFailure;
1266 // We may have added values to the cache list before this PHI translation.
1267 // If so, we haven't done anything to ensure that the cache remains sorted.
1268 // Sort it now (if needed) so that recursive invocations of
1269 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1270 // value will only see properly sorted cache arrays.
1271 if (Cache && NumSortedEntries != Cache->size()) {
1272 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1273 NumSortedEntries = Cache->size();
1278 for (BasicBlock *Pred : PredCache.get(BB)) {
1279 PredList.push_back(std::make_pair(Pred, Pointer));
1281 // Get the PHI translated pointer in this predecessor. This can fail if
1282 // not translatable, in which case the getAddr() returns null.
1283 PHITransAddr &PredPointer = PredList.back().second;
1284 PredPointer.PHITranslateValue(BB, Pred, DT, /*MustDominate=*/false);
1285 Value *PredPtrVal = PredPointer.getAddr();
1287 // Check to see if we have already visited this pred block with another
1288 // pointer. If so, we can't do this lookup. This failure can occur
1289 // with PHI translation when a critical edge exists and the PHI node in
1290 // the successor translates to a pointer value different than the
1291 // pointer the block was first analyzed with.
1292 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1293 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1295 if (!InsertRes.second) {
1296 // We found the pred; take it off the list of preds to visit.
1297 PredList.pop_back();
1299 // If the predecessor was visited with PredPtr, then we already did
1300 // the analysis and can ignore it.
1301 if (InsertRes.first->second == PredPtrVal)
1304 // Otherwise, the block was previously analyzed with a different
1305 // pointer. We can't represent the result of this case, so we just
1306 // treat this as a phi translation failure.
1308 // Make sure to clean up the Visited map before continuing on to
1309 // PredTranslationFailure.
1310 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1311 Visited.erase(PredList[i].first);
1313 goto PredTranslationFailure;
1317 // Actually process results here; this need to be a separate loop to avoid
1318 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1319 // any results for. (getNonLocalPointerDepFromBB will modify our
1320 // datastructures in ways the code after the PredTranslationFailure label
1322 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1323 BasicBlock *Pred = PredList[i].first;
1324 PHITransAddr &PredPointer = PredList[i].second;
1325 Value *PredPtrVal = PredPointer.getAddr();
1327 bool CanTranslate = true;
1328 // If PHI translation was unable to find an available pointer in this
1329 // predecessor, then we have to assume that the pointer is clobbered in
1330 // that predecessor. We can still do PRE of the load, which would insert
1331 // a computation of the pointer in this predecessor.
1333 CanTranslate = false;
1335 // FIXME: it is entirely possible that PHI translating will end up with
1336 // the same value. Consider PHI translating something like:
1337 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1338 // to recurse here, pedantically speaking.
1340 // If getNonLocalPointerDepFromBB fails here, that means the cached
1341 // result conflicted with the Visited list; we have to conservatively
1342 // assume it is unknown, but this also does not block PRE of the load.
1343 if (!CanTranslate ||
1344 getNonLocalPointerDepFromBB(QueryInst, PredPointer,
1345 Loc.getWithNewPtr(PredPtrVal),
1348 // Add the entry to the Result list.
1349 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1350 Result.push_back(Entry);
1352 // Since we had a phi translation failure, the cache for CacheKey won't
1353 // include all of the entries that we need to immediately satisfy future
1354 // queries. Mark this in NonLocalPointerDeps by setting the
1355 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1356 // cached value to do more work but not miss the phi trans failure.
1357 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1358 NLPI.Pair = BBSkipFirstBlockPair();
1363 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1364 CacheInfo = &NonLocalPointerDeps[CacheKey];
1365 Cache = &CacheInfo->NonLocalDeps;
1366 NumSortedEntries = Cache->size();
1368 // Since we did phi translation, the "Cache" set won't contain all of the
1369 // results for the query. This is ok (we can still use it to accelerate
1370 // specific block queries) but we can't do the fastpath "return all
1371 // results from the set" Clear out the indicator for this.
1372 CacheInfo->Pair = BBSkipFirstBlockPair();
1373 SkipFirstBlock = false;
1376 PredTranslationFailure:
1377 // The following code is "failure"; we can't produce a sane translation
1378 // for the given block. It assumes that we haven't modified any of
1379 // our datastructures while processing the current block.
1382 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1383 CacheInfo = &NonLocalPointerDeps[CacheKey];
1384 Cache = &CacheInfo->NonLocalDeps;
1385 NumSortedEntries = Cache->size();
1388 // Since we failed phi translation, the "Cache" set won't contain all of the
1389 // results for the query. This is ok (we can still use it to accelerate
1390 // specific block queries) but we can't do the fastpath "return all
1391 // results from the set". Clear out the indicator for this.
1392 CacheInfo->Pair = BBSkipFirstBlockPair();
1394 // If *nothing* works, mark the pointer as unknown.
1396 // If this is the magic first block, return this as a clobber of the whole
1397 // incoming value. Since we can't phi translate to one of the predecessors,
1398 // we have to bail out.
1402 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1403 assert(I != Cache->rend() && "Didn't find current block??");
1404 if (I->getBB() != BB)
1407 assert((I->getResult().isNonLocal() || !DT->isReachableFromEntry(BB)) &&
1408 "Should only be here with transparent block");
1409 I->setResult(MemDepResult::getUnknown());
1410 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1411 Pointer.getAddr()));
1416 // Okay, we're done now. If we added new values to the cache, re-sort it.
1417 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1418 DEBUG(AssertSorted(*Cache));
1422 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1423 /// CachedNonLocalPointerInfo, remove it.
1424 void MemoryDependenceAnalysis::
1425 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1426 CachedNonLocalPointerInfo::iterator It =
1427 NonLocalPointerDeps.find(P);
1428 if (It == NonLocalPointerDeps.end()) return;
1430 // Remove all of the entries in the BB->val map. This involves removing
1431 // instructions from the reverse map.
1432 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1434 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1435 Instruction *Target = PInfo[i].getResult().getInst();
1436 if (!Target) continue; // Ignore non-local dep results.
1437 assert(Target->getParent() == PInfo[i].getBB());
1439 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1440 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1443 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1444 NonLocalPointerDeps.erase(It);
1448 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1449 /// information about the specified pointer, because it may be too
1450 /// conservative in memdep. This is an optional call that can be used when
1451 /// the client detects an equivalence between the pointer and some other
1452 /// value and replaces the other value with ptr. This can make Ptr available
1453 /// in more places that cached info does not necessarily keep.
1454 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1455 // If Ptr isn't really a pointer, just ignore it.
1456 if (!Ptr->getType()->isPointerTy()) return;
1457 // Flush store info for the pointer.
1458 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1459 // Flush load info for the pointer.
1460 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1463 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1464 /// This needs to be done when the CFG changes, e.g., due to splitting
1466 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1470 /// removeInstruction - Remove an instruction from the dependence analysis,
1471 /// updating the dependence of instructions that previously depended on it.
1472 /// This method attempts to keep the cache coherent using the reverse map.
1473 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1474 // Walk through the Non-local dependencies, removing this one as the value
1475 // for any cached queries.
1476 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1477 if (NLDI != NonLocalDeps.end()) {
1478 NonLocalDepInfo &BlockMap = NLDI->second.first;
1479 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1481 if (Instruction *Inst = DI->getResult().getInst())
1482 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1483 NonLocalDeps.erase(NLDI);
1486 // If we have a cached local dependence query for this instruction, remove it.
1488 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1489 if (LocalDepEntry != LocalDeps.end()) {
1490 // Remove us from DepInst's reverse set now that the local dep info is gone.
1491 if (Instruction *Inst = LocalDepEntry->second.getInst())
1492 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1494 // Remove this local dependency info.
1495 LocalDeps.erase(LocalDepEntry);
1498 // If we have any cached pointer dependencies on this instruction, remove
1499 // them. If the instruction has non-pointer type, then it can't be a pointer
1502 // Remove it from both the load info and the store info. The instruction
1503 // can't be in either of these maps if it is non-pointer.
1504 if (RemInst->getType()->isPointerTy()) {
1505 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1506 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1509 // Loop over all of the things that depend on the instruction we're removing.
1511 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1513 // If we find RemInst as a clobber or Def in any of the maps for other values,
1514 // we need to replace its entry with a dirty version of the instruction after
1515 // it. If RemInst is a terminator, we use a null dirty value.
1517 // Using a dirty version of the instruction after RemInst saves having to scan
1518 // the entire block to get to this point.
1519 MemDepResult NewDirtyVal;
1520 if (!RemInst->isTerminator())
1521 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1523 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1524 if (ReverseDepIt != ReverseLocalDeps.end()) {
1525 // RemInst can't be the terminator if it has local stuff depending on it.
1526 assert(!ReverseDepIt->second.empty() && !isa<TerminatorInst>(RemInst) &&
1527 "Nothing can locally depend on a terminator");
1529 for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) {
1530 assert(InstDependingOnRemInst != RemInst &&
1531 "Already removed our local dep info");
1533 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1535 // Make sure to remember that new things depend on NewDepInst.
1536 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1537 "a local dep on this if it is a terminator!");
1538 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1539 InstDependingOnRemInst));
1542 ReverseLocalDeps.erase(ReverseDepIt);
1544 // Add new reverse deps after scanning the set, to avoid invalidating the
1545 // 'ReverseDeps' reference.
1546 while (!ReverseDepsToAdd.empty()) {
1547 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1548 .insert(ReverseDepsToAdd.back().second);
1549 ReverseDepsToAdd.pop_back();
1553 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1554 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1555 for (Instruction *I : ReverseDepIt->second) {
1556 assert(I != RemInst && "Already removed NonLocalDep info for RemInst");
1558 PerInstNLInfo &INLD = NonLocalDeps[I];
1559 // The information is now dirty!
1562 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1563 DE = INLD.first.end(); DI != DE; ++DI) {
1564 if (DI->getResult().getInst() != RemInst) continue;
1566 // Convert to a dirty entry for the subsequent instruction.
1567 DI->setResult(NewDirtyVal);
1569 if (Instruction *NextI = NewDirtyVal.getInst())
1570 ReverseDepsToAdd.push_back(std::make_pair(NextI, I));
1574 ReverseNonLocalDeps.erase(ReverseDepIt);
1576 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1577 while (!ReverseDepsToAdd.empty()) {
1578 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1579 .insert(ReverseDepsToAdd.back().second);
1580 ReverseDepsToAdd.pop_back();
1584 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1585 // value in the NonLocalPointerDeps info.
1586 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1587 ReverseNonLocalPtrDeps.find(RemInst);
1588 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1589 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1591 for (ValueIsLoadPair P : ReversePtrDepIt->second) {
1592 assert(P.getPointer() != RemInst &&
1593 "Already removed NonLocalPointerDeps info for RemInst");
1595 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1597 // The cache is not valid for any specific block anymore.
1598 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1600 // Update any entries for RemInst to use the instruction after it.
1601 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1603 if (DI->getResult().getInst() != RemInst) continue;
1605 // Convert to a dirty entry for the subsequent instruction.
1606 DI->setResult(NewDirtyVal);
1608 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1609 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1612 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1613 // subsequent value may invalidate the sortedness.
1614 std::sort(NLPDI.begin(), NLPDI.end());
1617 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1619 while (!ReversePtrDepsToAdd.empty()) {
1620 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1621 .insert(ReversePtrDepsToAdd.back().second);
1622 ReversePtrDepsToAdd.pop_back();
1627 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1628 DEBUG(verifyRemoved(RemInst));
1630 /// verifyRemoved - Verify that the specified instruction does not occur
1631 /// in our internal data structures. This function verifies by asserting in
1633 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1635 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1636 E = LocalDeps.end(); I != E; ++I) {
1637 assert(I->first != D && "Inst occurs in data structures");
1638 assert(I->second.getInst() != D &&
1639 "Inst occurs in data structures");
1642 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1643 E = NonLocalPointerDeps.end(); I != E; ++I) {
1644 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1645 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1646 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1648 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1651 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1652 E = NonLocalDeps.end(); I != E; ++I) {
1653 assert(I->first != D && "Inst occurs in data structures");
1654 const PerInstNLInfo &INLD = I->second;
1655 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1656 EE = INLD.first.end(); II != EE; ++II)
1657 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1660 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1661 E = ReverseLocalDeps.end(); I != E; ++I) {
1662 assert(I->first != D && "Inst occurs in data structures");
1663 for (Instruction *Inst : I->second)
1664 assert(Inst != D && "Inst occurs in data structures");
1667 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1668 E = ReverseNonLocalDeps.end();
1670 assert(I->first != D && "Inst occurs in data structures");
1671 for (Instruction *Inst : I->second)
1672 assert(Inst != D && "Inst occurs in data structures");
1675 for (ReverseNonLocalPtrDepTy::const_iterator
1676 I = ReverseNonLocalPtrDeps.begin(),
1677 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1678 assert(I->first != D && "Inst occurs in rev NLPD map");
1680 for (ValueIsLoadPair P : I->second)
1681 assert(P != ValueIsLoadPair(D, false) &&
1682 P != ValueIsLoadPair(D, true) &&
1683 "Inst occurs in ReverseNonLocalPtrDeps map");