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/Analysis/TargetLibraryInfo.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/LLVMContext.h"
34 #include "llvm/IR/PredIteratorCache.h"
35 #include "llvm/Support/Debug.h"
38 #define DEBUG_TYPE "memdep"
40 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
41 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
42 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
44 STATISTIC(NumCacheNonLocalPtr,
45 "Number of fully cached non-local ptr responses");
46 STATISTIC(NumCacheDirtyNonLocalPtr,
47 "Number of cached, but dirty, non-local ptr responses");
48 STATISTIC(NumUncacheNonLocalPtr,
49 "Number of uncached non-local ptr responses");
50 STATISTIC(NumCacheCompleteNonLocalPtr,
51 "Number of block queries that were completely cached");
53 // Limit for the number of instructions to scan in a block.
55 static cl::opt<unsigned> BlockScanLimit(
56 "memdep-block-scan-limit", cl::Hidden, cl::init(100),
57 cl::desc("The number of instructions to scan in a block in memory "
58 "dependency analysis (default = 100)"));
60 // Limit on the number of memdep results to process.
61 static const unsigned int NumResultsLimit = 100;
63 char MemoryDependenceAnalysis::ID = 0;
65 // Register this pass...
66 INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep",
67 "Memory Dependence Analysis", false, true)
68 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
69 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
70 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
71 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep",
72 "Memory Dependence Analysis", false, true)
74 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
76 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry());
78 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
81 /// Clean up memory in between runs
82 void MemoryDependenceAnalysis::releaseMemory() {
85 NonLocalPointerDeps.clear();
86 ReverseLocalDeps.clear();
87 ReverseNonLocalDeps.clear();
88 ReverseNonLocalPtrDeps.clear();
92 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
94 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
96 AU.addRequired<AssumptionCacheTracker>();
97 AU.addRequiredTransitive<AAResultsWrapperPass>();
98 AU.addRequiredTransitive<TargetLibraryInfoWrapperPass>();
101 bool MemoryDependenceAnalysis::runOnFunction(Function &F) {
102 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
103 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
104 DominatorTreeWrapperPass *DTWP =
105 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
106 DT = DTWP ? &DTWP->getDomTree() : nullptr;
107 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
111 /// RemoveFromReverseMap - This is a helper function that removes Val from
112 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
113 template <typename KeyTy>
114 static void RemoveFromReverseMap(DenseMap<Instruction*,
115 SmallPtrSet<KeyTy, 4> > &ReverseMap,
116 Instruction *Inst, KeyTy Val) {
117 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
118 InstIt = ReverseMap.find(Inst);
119 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
120 bool Found = InstIt->second.erase(Val);
121 assert(Found && "Invalid reverse map!"); (void)Found;
122 if (InstIt->second.empty())
123 ReverseMap.erase(InstIt);
126 /// GetLocation - If the given instruction references a specific memory
127 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
128 /// Return a ModRefInfo value describing the general behavior of the
130 static ModRefInfo GetLocation(const Instruction *Inst, MemoryLocation &Loc,
131 const TargetLibraryInfo &TLI) {
132 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
133 if (LI->isUnordered()) {
134 Loc = MemoryLocation::get(LI);
137 if (LI->getOrdering() == Monotonic) {
138 Loc = MemoryLocation::get(LI);
141 Loc = MemoryLocation();
145 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
146 if (SI->isUnordered()) {
147 Loc = MemoryLocation::get(SI);
150 if (SI->getOrdering() == Monotonic) {
151 Loc = MemoryLocation::get(SI);
154 Loc = MemoryLocation();
158 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
159 Loc = MemoryLocation::get(V);
163 if (const CallInst *CI = isFreeCall(Inst, &TLI)) {
164 // calls to free() deallocate the entire structure
165 Loc = MemoryLocation(CI->getArgOperand(0));
169 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
172 switch (II->getIntrinsicID()) {
173 case Intrinsic::lifetime_start:
174 case Intrinsic::lifetime_end:
175 case Intrinsic::invariant_start:
176 II->getAAMetadata(AAInfo);
177 Loc = MemoryLocation(
178 II->getArgOperand(1),
179 cast<ConstantInt>(II->getArgOperand(0))->getZExtValue(), AAInfo);
180 // These intrinsics don't really modify the memory, but returning Mod
181 // will allow them to be handled conservatively.
183 case Intrinsic::invariant_end:
184 II->getAAMetadata(AAInfo);
185 Loc = MemoryLocation(
186 II->getArgOperand(2),
187 cast<ConstantInt>(II->getArgOperand(1))->getZExtValue(), AAInfo);
188 // These intrinsics don't really modify the memory, but returning Mod
189 // will allow them to be handled conservatively.
196 // Otherwise, just do the coarse-grained thing that always works.
197 if (Inst->mayWriteToMemory())
199 if (Inst->mayReadFromMemory())
204 /// getCallSiteDependencyFrom - Private helper for finding the local
205 /// dependencies of a call site.
206 MemDepResult MemoryDependenceAnalysis::
207 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
208 BasicBlock::iterator ScanIt, BasicBlock *BB) {
209 unsigned Limit = BlockScanLimit;
211 // Walk backwards through the block, looking for dependencies
212 while (ScanIt != BB->begin()) {
213 // Limit the amount of scanning we do so we don't end up with quadratic
214 // running time on extreme testcases.
217 return MemDepResult::getUnknown();
219 Instruction *Inst = &*--ScanIt;
221 // If this inst is a memory op, get the pointer it accessed
223 ModRefInfo MR = GetLocation(Inst, Loc, *TLI);
225 // A simple instruction.
226 if (AA->getModRefInfo(CS, Loc) != MRI_NoModRef)
227 return MemDepResult::getClobber(Inst);
231 if (auto InstCS = CallSite(Inst)) {
232 // Debug intrinsics don't cause dependences.
233 if (isa<DbgInfoIntrinsic>(Inst)) continue;
234 // If these two calls do not interfere, look past it.
235 switch (AA->getModRefInfo(CS, InstCS)) {
237 // If the two calls are the same, return InstCS as a Def, so that
238 // CS can be found redundant and eliminated.
239 if (isReadOnlyCall && !(MR & MRI_Mod) &&
240 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
241 return MemDepResult::getDef(Inst);
243 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
247 return MemDepResult::getClobber(Inst);
251 // If we could not obtain a pointer for the instruction and the instruction
252 // touches memory then assume that this is a dependency.
253 if (MR != MRI_NoModRef)
254 return MemDepResult::getClobber(Inst);
257 // No dependence found. If this is the entry block of the function, it is
258 // unknown, otherwise it is non-local.
259 if (BB != &BB->getParent()->getEntryBlock())
260 return MemDepResult::getNonLocal();
261 return MemDepResult::getNonFuncLocal();
264 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
265 /// would fully overlap MemLoc if done as a wider legal integer load.
267 /// MemLocBase, MemLocOffset are lazily computed here the first time the
268 /// base/offs of memloc is needed.
269 static bool isLoadLoadClobberIfExtendedToFullWidth(const MemoryLocation &MemLoc,
270 const Value *&MemLocBase,
272 const LoadInst *LI) {
273 const DataLayout &DL = LI->getModule()->getDataLayout();
275 // If we haven't already computed the base/offset of MemLoc, do so now.
277 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, DL);
279 unsigned Size = MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
280 MemLocBase, MemLocOffs, MemLoc.Size, LI);
284 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
285 /// looks at a memory location for a load (specified by MemLocBase, Offs,
286 /// and Size) and compares it against a load. If the specified load could
287 /// be safely widened to a larger integer load that is 1) still efficient,
288 /// 2) safe for the target, and 3) would provide the specified memory
289 /// location value, then this function returns the size in bytes of the
290 /// load width to use. If not, this returns zero.
291 unsigned MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
292 const Value *MemLocBase, int64_t MemLocOffs, unsigned MemLocSize,
293 const LoadInst *LI) {
294 // We can only extend simple integer loads.
295 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
297 // Load widening is hostile to ThreadSanitizer: it may cause false positives
298 // or make the reports more cryptic (access sizes are wrong).
299 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
302 const DataLayout &DL = LI->getModule()->getDataLayout();
304 // Get the base of this load.
306 const Value *LIBase =
307 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, DL);
309 // If the two pointers are not based on the same pointer, we can't tell that
311 if (LIBase != MemLocBase) return 0;
313 // Okay, the two values are based on the same pointer, but returned as
314 // no-alias. This happens when we have things like two byte loads at "P+1"
315 // and "P+3". Check to see if increasing the size of the "LI" load up to its
316 // alignment (or the largest native integer type) will allow us to load all
317 // the bits required by MemLoc.
319 // If MemLoc is before LI, then no widening of LI will help us out.
320 if (MemLocOffs < LIOffs) return 0;
322 // Get the alignment of the load in bytes. We assume that it is safe to load
323 // any legal integer up to this size without a problem. For example, if we're
324 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
325 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
327 unsigned LoadAlign = LI->getAlignment();
329 int64_t MemLocEnd = MemLocOffs+MemLocSize;
331 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
332 if (LIOffs+LoadAlign < MemLocEnd) return 0;
334 // This is the size of the load to try. Start with the next larger power of
336 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
337 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
340 // If this load size is bigger than our known alignment or would not fit
341 // into a native integer register, then we fail.
342 if (NewLoadByteSize > LoadAlign ||
343 !DL.fitsInLegalInteger(NewLoadByteSize*8))
346 if (LIOffs + NewLoadByteSize > MemLocEnd &&
347 LI->getParent()->getParent()->hasFnAttribute(
348 Attribute::SanitizeAddress))
349 // We will be reading past the location accessed by the original program.
350 // While this is safe in a regular build, Address Safety analysis tools
351 // may start reporting false warnings. So, don't do widening.
354 // If a load of this width would include all of MemLoc, then we succeed.
355 if (LIOffs+NewLoadByteSize >= MemLocEnd)
356 return NewLoadByteSize;
358 NewLoadByteSize <<= 1;
362 static bool isVolatile(Instruction *Inst) {
363 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
364 return LI->isVolatile();
365 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
366 return SI->isVolatile();
367 else if (AtomicCmpXchgInst *AI = dyn_cast<AtomicCmpXchgInst>(Inst))
368 return AI->isVolatile();
373 /// getPointerDependencyFrom - Return the instruction on which a memory
374 /// location depends. If isLoad is true, this routine ignores may-aliases with
375 /// read-only operations. If isLoad is false, this routine ignores may-aliases
376 /// with reads from read-only locations. If possible, pass the query
377 /// instruction as well; this function may take advantage of the metadata
378 /// annotated to the query instruction to refine the result.
379 MemDepResult MemoryDependenceAnalysis::getPointerDependencyFrom(
380 const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
381 BasicBlock *BB, Instruction *QueryInst) {
383 if (QueryInst != nullptr) {
384 if (auto *LI = dyn_cast<LoadInst>(QueryInst)) {
385 MemDepResult invariantGroupDependency =
386 getInvariantGroupPointerDependency(LI, BB);
388 if (invariantGroupDependency.isDef())
389 return invariantGroupDependency;
392 return getSimplePointerDependencyFrom(MemLoc, isLoad, ScanIt, BB, QueryInst);
396 MemoryDependenceAnalysis::getInvariantGroupPointerDependency(LoadInst *LI,
398 Value *LoadOperand = LI->getPointerOperand();
399 // It's is not safe to walk the use list of global value, because function
400 // passes aren't allowed to look outside their functions.
401 if (isa<GlobalValue>(LoadOperand))
402 return MemDepResult::getUnknown();
404 auto *InvariantGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group);
405 if (!InvariantGroupMD)
406 return MemDepResult::getUnknown();
408 MemDepResult Result = MemDepResult::getUnknown();
409 llvm::SmallSet<Value *, 14> Seen;
410 // Queue to process all pointers that are equivalent to load operand.
411 llvm::SmallVector<Value *, 8> LoadOperandsQueue;
412 LoadOperandsQueue.push_back(LoadOperand);
413 while (!LoadOperandsQueue.empty()) {
414 Value *Ptr = LoadOperandsQueue.pop_back_val();
415 if (isa<GlobalValue>(Ptr))
418 if (auto *BCI = dyn_cast<BitCastInst>(Ptr)) {
419 if (!Seen.count(BCI->getOperand(0))) {
420 LoadOperandsQueue.push_back(BCI->getOperand(0));
421 Seen.insert(BCI->getOperand(0));
425 for (Use &Us : Ptr->uses()) {
426 auto *U = dyn_cast<Instruction>(Us.getUser());
427 if (!U || U == LI || !DT->dominates(U, LI))
430 if (auto *BCI = dyn_cast<BitCastInst>(U)) {
431 if (!Seen.count(BCI)) {
432 LoadOperandsQueue.push_back(BCI);
437 // If we hit load/store with the same invariant.group metadata (and the
438 // same pointer operand) we can assume that value pointed by pointer
439 // operand didn't change.
440 if ((isa<LoadInst>(U) || isa<StoreInst>(U)) && U->getParent() == BB &&
441 U->getMetadata(LLVMContext::MD_invariant_group) == InvariantGroupMD)
442 return MemDepResult::getDef(U);
448 MemDepResult MemoryDependenceAnalysis::getSimplePointerDependencyFrom(
449 const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
450 BasicBlock *BB, Instruction *QueryInst) {
452 const Value *MemLocBase = nullptr;
453 int64_t MemLocOffset = 0;
454 unsigned Limit = BlockScanLimit;
455 bool isInvariantLoad = false;
457 // We must be careful with atomic accesses, as they may allow another thread
458 // to touch this location, cloberring it. We are conservative: if the
459 // QueryInst is not a simple (non-atomic) memory access, we automatically
460 // return getClobber.
461 // If it is simple, we know based on the results of
462 // "Compiler testing via a theory of sound optimisations in the C11/C++11
463 // memory model" in PLDI 2013, that a non-atomic location can only be
464 // clobbered between a pair of a release and an acquire action, with no
465 // access to the location in between.
466 // Here is an example for giving the general intuition behind this rule.
467 // In the following code:
469 // release action; [1]
470 // acquire action; [4]
472 // It is unsafe to replace %val by 0 because another thread may be running:
473 // acquire action; [2]
475 // release action; [3]
476 // with synchronization from 1 to 2 and from 3 to 4, resulting in %val
477 // being 42. A key property of this program however is that if either
478 // 1 or 4 were missing, there would be a race between the store of 42
479 // either the store of 0 or the load (making the whole progam racy).
480 // The paper mentioned above shows that the same property is respected
481 // by every program that can detect any optimisation of that kind: either
482 // it is racy (undefined) or there is a release followed by an acquire
483 // between the pair of accesses under consideration.
485 // If the load is invariant, we "know" that it doesn't alias *any* write. We
486 // do want to respect mustalias results since defs are useful for value
487 // forwarding, but any mayalias write can be assumed to be noalias.
488 // Arguably, this logic should be pushed inside AliasAnalysis itself.
489 if (isLoad && QueryInst) {
490 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
491 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
492 isInvariantLoad = true;
495 const DataLayout &DL = BB->getModule()->getDataLayout();
497 // Create a numbered basic block to lazily compute and cache instruction
498 // positions inside a BB. This is used to provide fast queries for relative
499 // position between two instructions in a BB and can be used by
500 // AliasAnalysis::callCapturesBefore.
501 OrderedBasicBlock OBB(BB);
503 // Walk backwards through the basic block, looking for dependencies.
504 while (ScanIt != BB->begin()) {
505 Instruction *Inst = &*--ScanIt;
507 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
508 // Debug intrinsics don't (and can't) cause dependencies.
509 if (isa<DbgInfoIntrinsic>(II)) continue;
511 // Limit the amount of scanning we do so we don't end up with quadratic
512 // running time on extreme testcases.
515 return MemDepResult::getUnknown();
517 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
518 // If we reach a lifetime begin or end marker, then the query ends here
519 // because the value is undefined.
520 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
521 // FIXME: This only considers queries directly on the invariant-tagged
522 // pointer, not on query pointers that are indexed off of them. It'd
523 // be nice to handle that at some point (the right approach is to use
524 // GetPointerBaseWithConstantOffset).
525 if (AA->isMustAlias(MemoryLocation(II->getArgOperand(1)), MemLoc))
526 return MemDepResult::getDef(II);
531 // Values depend on loads if the pointers are must aliased. This means that
532 // a load depends on another must aliased load from the same value.
533 // One exception is atomic loads: a value can depend on an atomic load that it
534 // does not alias with when this atomic load indicates that another thread may
535 // be accessing the location.
536 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
538 // While volatile access cannot be eliminated, they do not have to clobber
539 // non-aliasing locations, as normal accesses, for example, can be safely
540 // reordered with volatile accesses.
541 if (LI->isVolatile()) {
543 // Original QueryInst *may* be volatile
544 return MemDepResult::getClobber(LI);
545 if (isVolatile(QueryInst))
546 // Ordering required if QueryInst is itself volatile
547 return MemDepResult::getClobber(LI);
548 // Otherwise, volatile doesn't imply any special ordering
551 // Atomic loads have complications involved.
552 // A Monotonic (or higher) load is OK if the query inst is itself not atomic.
553 // FIXME: This is overly conservative.
554 if (LI->isAtomic() && LI->getOrdering() > Unordered) {
556 return MemDepResult::getClobber(LI);
557 if (LI->getOrdering() != Monotonic)
558 return MemDepResult::getClobber(LI);
559 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
560 if (!QueryLI->isSimple())
561 return MemDepResult::getClobber(LI);
562 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
563 if (!QuerySI->isSimple())
564 return MemDepResult::getClobber(LI);
565 } else if (QueryInst->mayReadOrWriteMemory()) {
566 return MemDepResult::getClobber(LI);
570 MemoryLocation LoadLoc = MemoryLocation::get(LI);
572 // If we found a pointer, check if it could be the same as our pointer.
573 AliasResult R = AA->alias(LoadLoc, MemLoc);
577 // If this is an over-aligned integer load (for example,
578 // "load i8* %P, align 4") see if it would obviously overlap with the
579 // queried location if widened to a larger load (e.g. if the queried
580 // location is 1 byte at P+1). If so, return it as a load/load
581 // clobber result, allowing the client to decide to widen the load if
583 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
584 if (LI->getAlignment() * 8 > ITy->getPrimitiveSizeInBits() &&
585 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
587 return MemDepResult::getClobber(Inst);
592 // Must aliased loads are defs of each other.
594 return MemDepResult::getDef(Inst);
596 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
597 // in terms of clobbering loads, but since it does this by looking
598 // at the clobbering load directly, it doesn't know about any
599 // phi translation that may have happened along the way.
601 // If we have a partial alias, then return this as a clobber for the
603 if (R == PartialAlias)
604 return MemDepResult::getClobber(Inst);
607 // Random may-alias loads don't depend on each other without a
612 // Stores don't depend on other no-aliased accesses.
616 // Stores don't alias loads from read-only memory.
617 if (AA->pointsToConstantMemory(LoadLoc))
620 // Stores depend on may/must aliased loads.
621 return MemDepResult::getDef(Inst);
624 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
625 // Atomic stores have complications involved.
626 // A Monotonic store is OK if the query inst is itself not atomic.
627 // FIXME: This is overly conservative.
628 if (!SI->isUnordered()) {
630 return MemDepResult::getClobber(SI);
631 if (SI->getOrdering() != Monotonic)
632 return MemDepResult::getClobber(SI);
633 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
634 if (!QueryLI->isSimple())
635 return MemDepResult::getClobber(SI);
636 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
637 if (!QuerySI->isSimple())
638 return MemDepResult::getClobber(SI);
639 } else if (QueryInst->mayReadOrWriteMemory()) {
640 return MemDepResult::getClobber(SI);
644 // FIXME: this is overly conservative.
645 // While volatile access cannot be eliminated, they do not have to clobber
646 // non-aliasing locations, as normal accesses can for example be reordered
647 // with volatile accesses.
648 if (SI->isVolatile())
649 return MemDepResult::getClobber(SI);
651 // If alias analysis can tell that this store is guaranteed to not modify
652 // the query pointer, ignore it. Use getModRefInfo to handle cases where
653 // the query pointer points to constant memory etc.
654 if (AA->getModRefInfo(SI, MemLoc) == MRI_NoModRef)
657 // Ok, this store might clobber the query pointer. Check to see if it is
658 // a must alias: in this case, we want to return this as a def.
659 MemoryLocation StoreLoc = MemoryLocation::get(SI);
661 // If we found a pointer, check if it could be the same as our pointer.
662 AliasResult R = AA->alias(StoreLoc, MemLoc);
667 return MemDepResult::getDef(Inst);
670 return MemDepResult::getClobber(Inst);
673 // If this is an allocation, and if we know that the accessed pointer is to
674 // the allocation, return Def. This means that there is no dependence and
675 // the access can be optimized based on that. For example, a load could
677 // Note: Only determine this to be a malloc if Inst is the malloc call, not
678 // a subsequent bitcast of the malloc call result. There can be stores to
679 // the malloced memory between the malloc call and its bitcast uses, and we
680 // need to continue scanning until the malloc call.
681 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
682 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
684 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
685 return MemDepResult::getDef(Inst);
688 // Be conservative if the accessed pointer may alias the allocation.
689 if (AA->alias(Inst, AccessPtr) != NoAlias)
690 return MemDepResult::getClobber(Inst);
691 // If the allocation is not aliased and does not read memory (like
692 // strdup), it is safe to ignore.
693 if (isa<AllocaInst>(Inst) ||
694 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
701 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
702 ModRefInfo MR = AA->getModRefInfo(Inst, MemLoc);
703 // If necessary, perform additional analysis.
704 if (MR == MRI_ModRef)
705 MR = AA->callCapturesBefore(Inst, MemLoc, DT, &OBB);
708 // If the call has no effect on the queried pointer, just ignore it.
711 return MemDepResult::getClobber(Inst);
713 // If the call is known to never store to the pointer, and if this is a
714 // load query, we can safely ignore it (scan past it).
718 // Otherwise, there is a potential dependence. Return a clobber.
719 return MemDepResult::getClobber(Inst);
723 // No dependence found. If this is the entry block of the function, it is
724 // unknown, otherwise it is non-local.
725 if (BB != &BB->getParent()->getEntryBlock())
726 return MemDepResult::getNonLocal();
727 return MemDepResult::getNonFuncLocal();
730 /// getDependency - Return the instruction on which a memory operation
732 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
733 Instruction *ScanPos = QueryInst;
735 // Check for a cached result
736 MemDepResult &LocalCache = LocalDeps[QueryInst];
738 // If the cached entry is non-dirty, just return it. Note that this depends
739 // on MemDepResult's default constructing to 'dirty'.
740 if (!LocalCache.isDirty())
743 // Otherwise, if we have a dirty entry, we know we can start the scan at that
744 // instruction, which may save us some work.
745 if (Instruction *Inst = LocalCache.getInst()) {
748 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
751 BasicBlock *QueryParent = QueryInst->getParent();
754 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
755 // No dependence found. If this is the entry block of the function, it is
756 // unknown, otherwise it is non-local.
757 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
758 LocalCache = MemDepResult::getNonLocal();
760 LocalCache = MemDepResult::getNonFuncLocal();
762 MemoryLocation MemLoc;
763 ModRefInfo MR = GetLocation(QueryInst, MemLoc, *TLI);
765 // If we can do a pointer scan, make it happen.
766 bool isLoad = !(MR & MRI_Mod);
767 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
768 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
770 LocalCache = getPointerDependencyFrom(
771 MemLoc, isLoad, ScanPos->getIterator(), QueryParent, QueryInst);
772 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
773 CallSite QueryCS(QueryInst);
774 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
775 LocalCache = getCallSiteDependencyFrom(
776 QueryCS, isReadOnly, ScanPos->getIterator(), QueryParent);
778 // Non-memory instruction.
779 LocalCache = MemDepResult::getUnknown();
782 // Remember the result!
783 if (Instruction *I = LocalCache.getInst())
784 ReverseLocalDeps[I].insert(QueryInst);
790 /// AssertSorted - This method is used when -debug is specified to verify that
791 /// cache arrays are properly kept sorted.
792 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
794 if (Count == -1) Count = Cache.size();
795 assert(std::is_sorted(Cache.begin(), Cache.begin() + Count) &&
796 "Cache isn't sorted!");
800 /// getNonLocalCallDependency - Perform a full dependency query for the
801 /// specified call, returning the set of blocks that the value is
802 /// potentially live across. The returned set of results will include a
803 /// "NonLocal" result for all blocks where the value is live across.
805 /// This method assumes the instruction returns a "NonLocal" dependency
806 /// within its own block.
808 /// This returns a reference to an internal data structure that may be
809 /// invalidated on the next non-local query or when an instruction is
810 /// removed. Clients must copy this data if they want it around longer than
812 const MemoryDependenceAnalysis::NonLocalDepInfo &
813 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
814 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
815 "getNonLocalCallDependency should only be used on calls with non-local deps!");
816 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
817 NonLocalDepInfo &Cache = CacheP.first;
819 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
820 /// the cached case, this can happen due to instructions being deleted etc. In
821 /// the uncached case, this starts out as the set of predecessors we care
823 SmallVector<BasicBlock*, 32> DirtyBlocks;
825 if (!Cache.empty()) {
826 // Okay, we have a cache entry. If we know it is not dirty, just return it
827 // with no computation.
828 if (!CacheP.second) {
833 // If we already have a partially computed set of results, scan them to
834 // determine what is dirty, seeding our initial DirtyBlocks worklist.
835 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
837 if (I->getResult().isDirty())
838 DirtyBlocks.push_back(I->getBB());
840 // Sort the cache so that we can do fast binary search lookups below.
841 std::sort(Cache.begin(), Cache.end());
843 ++NumCacheDirtyNonLocal;
844 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
845 // << Cache.size() << " cached: " << *QueryInst;
847 // Seed DirtyBlocks with each of the preds of QueryInst's block.
848 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
849 for (BasicBlock *Pred : PredCache.get(QueryBB))
850 DirtyBlocks.push_back(Pred);
851 ++NumUncacheNonLocal;
854 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
855 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
857 SmallPtrSet<BasicBlock*, 64> Visited;
859 unsigned NumSortedEntries = Cache.size();
860 DEBUG(AssertSorted(Cache));
862 // Iterate while we still have blocks to update.
863 while (!DirtyBlocks.empty()) {
864 BasicBlock *DirtyBB = DirtyBlocks.back();
865 DirtyBlocks.pop_back();
867 // Already processed this block?
868 if (!Visited.insert(DirtyBB).second)
871 // Do a binary search to see if we already have an entry for this block in
872 // the cache set. If so, find it.
873 DEBUG(AssertSorted(Cache, NumSortedEntries));
874 NonLocalDepInfo::iterator Entry =
875 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
876 NonLocalDepEntry(DirtyBB));
877 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
880 NonLocalDepEntry *ExistingResult = nullptr;
881 if (Entry != Cache.begin()+NumSortedEntries &&
882 Entry->getBB() == DirtyBB) {
883 // If we already have an entry, and if it isn't already dirty, the block
885 if (!Entry->getResult().isDirty())
888 // Otherwise, remember this slot so we can update the value.
889 ExistingResult = &*Entry;
892 // If the dirty entry has a pointer, start scanning from it so we don't have
893 // to rescan the entire block.
894 BasicBlock::iterator ScanPos = DirtyBB->end();
895 if (ExistingResult) {
896 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
897 ScanPos = Inst->getIterator();
898 // We're removing QueryInst's use of Inst.
899 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
900 QueryCS.getInstruction());
904 // Find out if this block has a local dependency for QueryInst.
907 if (ScanPos != DirtyBB->begin()) {
908 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
909 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
910 // No dependence found. If this is the entry block of the function, it is
911 // a clobber, otherwise it is unknown.
912 Dep = MemDepResult::getNonLocal();
914 Dep = MemDepResult::getNonFuncLocal();
917 // If we had a dirty entry for the block, update it. Otherwise, just add
920 ExistingResult->setResult(Dep);
922 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
924 // If the block has a dependency (i.e. it isn't completely transparent to
925 // the value), remember the association!
926 if (!Dep.isNonLocal()) {
927 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
928 // update this when we remove instructions.
929 if (Instruction *Inst = Dep.getInst())
930 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
933 // If the block *is* completely transparent to the load, we need to check
934 // the predecessors of this block. Add them to our worklist.
935 for (BasicBlock *Pred : PredCache.get(DirtyBB))
936 DirtyBlocks.push_back(Pred);
943 /// getNonLocalPointerDependency - Perform a full dependency query for an
944 /// access to the specified (non-volatile) memory location, returning the
945 /// set of instructions that either define or clobber the value.
947 /// This method assumes the pointer has a "NonLocal" dependency within its
950 void MemoryDependenceAnalysis::
951 getNonLocalPointerDependency(Instruction *QueryInst,
952 SmallVectorImpl<NonLocalDepResult> &Result) {
953 const MemoryLocation Loc = MemoryLocation::get(QueryInst);
954 bool isLoad = isa<LoadInst>(QueryInst);
955 BasicBlock *FromBB = QueryInst->getParent();
958 assert(Loc.Ptr->getType()->isPointerTy() &&
959 "Can't get pointer deps of a non-pointer!");
962 // This routine does not expect to deal with volatile instructions.
963 // Doing so would require piping through the QueryInst all the way through.
964 // TODO: volatiles can't be elided, but they can be reordered with other
965 // non-volatile accesses.
967 // We currently give up on any instruction which is ordered, but we do handle
968 // atomic instructions which are unordered.
969 // TODO: Handle ordered instructions
970 auto isOrdered = [](Instruction *Inst) {
971 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
972 return !LI->isUnordered();
973 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
974 return !SI->isUnordered();
978 if (isVolatile(QueryInst) || isOrdered(QueryInst)) {
979 Result.push_back(NonLocalDepResult(FromBB,
980 MemDepResult::getUnknown(),
981 const_cast<Value *>(Loc.Ptr)));
984 const DataLayout &DL = FromBB->getModule()->getDataLayout();
985 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, AC);
987 // This is the set of blocks we've inspected, and the pointer we consider in
988 // each block. Because of critical edges, we currently bail out if querying
989 // a block with multiple different pointers. This can happen during PHI
991 DenseMap<BasicBlock*, Value*> Visited;
992 if (!getNonLocalPointerDepFromBB(QueryInst, Address, Loc, isLoad, FromBB,
993 Result, Visited, true))
996 Result.push_back(NonLocalDepResult(FromBB,
997 MemDepResult::getUnknown(),
998 const_cast<Value *>(Loc.Ptr)));
1001 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
1002 /// Pointer/PointeeSize using either cached information in Cache or by doing a
1003 /// lookup (which may use dirty cache info if available). If we do a lookup,
1004 /// add the result to the cache.
1005 MemDepResult MemoryDependenceAnalysis::GetNonLocalInfoForBlock(
1006 Instruction *QueryInst, const MemoryLocation &Loc, bool isLoad,
1007 BasicBlock *BB, NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
1009 // Do a binary search to see if we already have an entry for this block in
1010 // the cache set. If so, find it.
1011 NonLocalDepInfo::iterator Entry =
1012 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
1013 NonLocalDepEntry(BB));
1014 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
1017 NonLocalDepEntry *ExistingResult = nullptr;
1018 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
1019 ExistingResult = &*Entry;
1021 // If we have a cached entry, and it is non-dirty, use it as the value for
1023 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
1024 ++NumCacheNonLocalPtr;
1025 return ExistingResult->getResult();
1028 // Otherwise, we have to scan for the value. If we have a dirty cache
1029 // entry, start scanning from its position, otherwise we scan from the end
1031 BasicBlock::iterator ScanPos = BB->end();
1032 if (ExistingResult && ExistingResult->getResult().getInst()) {
1033 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
1034 "Instruction invalidated?");
1035 ++NumCacheDirtyNonLocalPtr;
1036 ScanPos = ExistingResult->getResult().getInst()->getIterator();
1038 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1039 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
1040 RemoveFromReverseMap(ReverseNonLocalPtrDeps, &*ScanPos, CacheKey);
1042 ++NumUncacheNonLocalPtr;
1045 // Scan the block for the dependency.
1046 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB,
1049 // If we had a dirty entry for the block, update it. Otherwise, just add
1052 ExistingResult->setResult(Dep);
1054 Cache->push_back(NonLocalDepEntry(BB, Dep));
1056 // If the block has a dependency (i.e. it isn't completely transparent to
1057 // the value), remember the reverse association because we just added it
1059 if (!Dep.isDef() && !Dep.isClobber())
1062 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
1063 // update MemDep when we remove instructions.
1064 Instruction *Inst = Dep.getInst();
1065 assert(Inst && "Didn't depend on anything?");
1066 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
1067 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
1071 /// SortNonLocalDepInfoCache - Sort the NonLocalDepInfo cache, given a certain
1072 /// number of elements in the array that are already properly ordered. This is
1073 /// optimized for the case when only a few entries are added.
1075 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
1076 unsigned NumSortedEntries) {
1077 switch (Cache.size() - NumSortedEntries) {
1079 // done, no new entries.
1082 // Two new entries, insert the last one into place.
1083 NonLocalDepEntry Val = Cache.back();
1085 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
1086 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
1087 Cache.insert(Entry, Val);
1091 // One new entry, Just insert the new value at the appropriate position.
1092 if (Cache.size() != 1) {
1093 NonLocalDepEntry Val = Cache.back();
1095 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
1096 std::upper_bound(Cache.begin(), Cache.end(), Val);
1097 Cache.insert(Entry, Val);
1101 // Added many values, do a full scale sort.
1102 std::sort(Cache.begin(), Cache.end());
1107 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
1108 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
1109 /// results to the results vector and keep track of which blocks are visited in
1112 /// This has special behavior for the first block queries (when SkipFirstBlock
1113 /// is true). In this special case, it ignores the contents of the specified
1114 /// block and starts returning dependence info for its predecessors.
1116 /// This function returns false on success, or true to indicate that it could
1117 /// not compute dependence information for some reason. This should be treated
1118 /// as a clobber dependence on the first instruction in the predecessor block.
1119 bool MemoryDependenceAnalysis::getNonLocalPointerDepFromBB(
1120 Instruction *QueryInst, const PHITransAddr &Pointer,
1121 const MemoryLocation &Loc, bool isLoad, BasicBlock *StartBB,
1122 SmallVectorImpl<NonLocalDepResult> &Result,
1123 DenseMap<BasicBlock *, Value *> &Visited, bool SkipFirstBlock) {
1124 // Look up the cached info for Pointer.
1125 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
1127 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
1128 // CacheKey, this value will be inserted as the associated value. Otherwise,
1129 // it'll be ignored, and we'll have to check to see if the cached size and
1130 // aa tags are consistent with the current query.
1131 NonLocalPointerInfo InitialNLPI;
1132 InitialNLPI.Size = Loc.Size;
1133 InitialNLPI.AATags = Loc.AATags;
1135 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
1136 // already have one.
1137 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
1138 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
1139 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
1141 // If we already have a cache entry for this CacheKey, we may need to do some
1142 // work to reconcile the cache entry and the current query.
1144 if (CacheInfo->Size < Loc.Size) {
1145 // The query's Size is greater than the cached one. Throw out the
1146 // cached data and proceed with the query at the greater size.
1147 CacheInfo->Pair = BBSkipFirstBlockPair();
1148 CacheInfo->Size = Loc.Size;
1149 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1150 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1151 if (Instruction *Inst = DI->getResult().getInst())
1152 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1153 CacheInfo->NonLocalDeps.clear();
1154 } else if (CacheInfo->Size > Loc.Size) {
1155 // This query's Size is less than the cached one. Conservatively restart
1156 // the query using the greater size.
1157 return getNonLocalPointerDepFromBB(QueryInst, Pointer,
1158 Loc.getWithNewSize(CacheInfo->Size),
1159 isLoad, StartBB, Result, Visited,
1163 // If the query's AATags are inconsistent with the cached one,
1164 // conservatively throw out the cached data and restart the query with
1165 // no tag if needed.
1166 if (CacheInfo->AATags != Loc.AATags) {
1167 if (CacheInfo->AATags) {
1168 CacheInfo->Pair = BBSkipFirstBlockPair();
1169 CacheInfo->AATags = AAMDNodes();
1170 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1171 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1172 if (Instruction *Inst = DI->getResult().getInst())
1173 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1174 CacheInfo->NonLocalDeps.clear();
1177 return getNonLocalPointerDepFromBB(QueryInst,
1178 Pointer, Loc.getWithoutAATags(),
1179 isLoad, StartBB, Result, Visited,
1184 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1186 // If we have valid cached information for exactly the block we are
1187 // investigating, just return it with no recomputation.
1188 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1189 // We have a fully cached result for this query then we can just return the
1190 // cached results and populate the visited set. However, we have to verify
1191 // that we don't already have conflicting results for these blocks. Check
1192 // to ensure that if a block in the results set is in the visited set that
1193 // it was for the same pointer query.
1194 if (!Visited.empty()) {
1195 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1197 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
1198 if (VI == Visited.end() || VI->second == Pointer.getAddr())
1201 // We have a pointer mismatch in a block. Just return clobber, saying
1202 // that something was clobbered in this result. We could also do a
1203 // non-fully cached query, but there is little point in doing this.
1208 Value *Addr = Pointer.getAddr();
1209 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1211 Visited.insert(std::make_pair(I->getBB(), Addr));
1212 if (I->getResult().isNonLocal()) {
1217 Result.push_back(NonLocalDepResult(I->getBB(),
1218 MemDepResult::getUnknown(),
1220 } else if (DT->isReachableFromEntry(I->getBB())) {
1221 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1224 ++NumCacheCompleteNonLocalPtr;
1228 // Otherwise, either this is a new block, a block with an invalid cache
1229 // pointer or one that we're about to invalidate by putting more info into it
1230 // than its valid cache info. If empty, the result will be valid cache info,
1231 // otherwise it isn't.
1233 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1235 CacheInfo->Pair = BBSkipFirstBlockPair();
1237 SmallVector<BasicBlock*, 32> Worklist;
1238 Worklist.push_back(StartBB);
1240 // PredList used inside loop.
1241 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1243 // Keep track of the entries that we know are sorted. Previously cached
1244 // entries will all be sorted. The entries we add we only sort on demand (we
1245 // don't insert every element into its sorted position). We know that we
1246 // won't get any reuse from currently inserted values, because we don't
1247 // revisit blocks after we insert info for them.
1248 unsigned NumSortedEntries = Cache->size();
1249 DEBUG(AssertSorted(*Cache));
1251 while (!Worklist.empty()) {
1252 BasicBlock *BB = Worklist.pop_back_val();
1254 // If we do process a large number of blocks it becomes very expensive and
1255 // likely it isn't worth worrying about
1256 if (Result.size() > NumResultsLimit) {
1258 // Sort it now (if needed) so that recursive invocations of
1259 // getNonLocalPointerDepFromBB and other routines that could reuse the
1260 // cache value will only see properly sorted cache arrays.
1261 if (Cache && NumSortedEntries != Cache->size()) {
1262 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1264 // Since we bail out, the "Cache" set won't contain all of the
1265 // results for the query. This is ok (we can still use it to accelerate
1266 // specific block queries) but we can't do the fastpath "return all
1267 // results from the set". Clear out the indicator for this.
1268 CacheInfo->Pair = BBSkipFirstBlockPair();
1272 // Skip the first block if we have it.
1273 if (!SkipFirstBlock) {
1274 // Analyze the dependency of *Pointer in FromBB. See if we already have
1276 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1278 // Get the dependency info for Pointer in BB. If we have cached
1279 // information, we will use it, otherwise we compute it.
1280 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1281 MemDepResult Dep = GetNonLocalInfoForBlock(QueryInst,
1282 Loc, isLoad, BB, Cache,
1285 // If we got a Def or Clobber, add this to the list of results.
1286 if (!Dep.isNonLocal()) {
1288 Result.push_back(NonLocalDepResult(BB,
1289 MemDepResult::getUnknown(),
1290 Pointer.getAddr()));
1292 } else if (DT->isReachableFromEntry(BB)) {
1293 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1299 // If 'Pointer' is an instruction defined in this block, then we need to do
1300 // phi translation to change it into a value live in the predecessor block.
1301 // If not, we just add the predecessors to the worklist and scan them with
1302 // the same Pointer.
1303 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1304 SkipFirstBlock = false;
1305 SmallVector<BasicBlock*, 16> NewBlocks;
1306 for (BasicBlock *Pred : PredCache.get(BB)) {
1307 // Verify that we haven't looked at this block yet.
1308 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1309 InsertRes = Visited.insert(std::make_pair(Pred, Pointer.getAddr()));
1310 if (InsertRes.second) {
1311 // First time we've looked at *PI.
1312 NewBlocks.push_back(Pred);
1316 // If we have seen this block before, but it was with a different
1317 // pointer then we have a phi translation failure and we have to treat
1318 // this as a clobber.
1319 if (InsertRes.first->second != Pointer.getAddr()) {
1320 // Make sure to clean up the Visited map before continuing on to
1321 // PredTranslationFailure.
1322 for (unsigned i = 0; i < NewBlocks.size(); i++)
1323 Visited.erase(NewBlocks[i]);
1324 goto PredTranslationFailure;
1327 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1331 // We do need to do phi translation, if we know ahead of time we can't phi
1332 // translate this value, don't even try.
1333 if (!Pointer.IsPotentiallyPHITranslatable())
1334 goto PredTranslationFailure;
1336 // We may have added values to the cache list before this PHI translation.
1337 // If so, we haven't done anything to ensure that the cache remains sorted.
1338 // Sort it now (if needed) so that recursive invocations of
1339 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1340 // value will only see properly sorted cache arrays.
1341 if (Cache && NumSortedEntries != Cache->size()) {
1342 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1343 NumSortedEntries = Cache->size();
1348 for (BasicBlock *Pred : PredCache.get(BB)) {
1349 PredList.push_back(std::make_pair(Pred, Pointer));
1351 // Get the PHI translated pointer in this predecessor. This can fail if
1352 // not translatable, in which case the getAddr() returns null.
1353 PHITransAddr &PredPointer = PredList.back().second;
1354 PredPointer.PHITranslateValue(BB, Pred, DT, /*MustDominate=*/false);
1355 Value *PredPtrVal = PredPointer.getAddr();
1357 // Check to see if we have already visited this pred block with another
1358 // pointer. If so, we can't do this lookup. This failure can occur
1359 // with PHI translation when a critical edge exists and the PHI node in
1360 // the successor translates to a pointer value different than the
1361 // pointer the block was first analyzed with.
1362 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1363 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1365 if (!InsertRes.second) {
1366 // We found the pred; take it off the list of preds to visit.
1367 PredList.pop_back();
1369 // If the predecessor was visited with PredPtr, then we already did
1370 // the analysis and can ignore it.
1371 if (InsertRes.first->second == PredPtrVal)
1374 // Otherwise, the block was previously analyzed with a different
1375 // pointer. We can't represent the result of this case, so we just
1376 // treat this as a phi translation failure.
1378 // Make sure to clean up the Visited map before continuing on to
1379 // PredTranslationFailure.
1380 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1381 Visited.erase(PredList[i].first);
1383 goto PredTranslationFailure;
1387 // Actually process results here; this need to be a separate loop to avoid
1388 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1389 // any results for. (getNonLocalPointerDepFromBB will modify our
1390 // datastructures in ways the code after the PredTranslationFailure label
1392 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1393 BasicBlock *Pred = PredList[i].first;
1394 PHITransAddr &PredPointer = PredList[i].second;
1395 Value *PredPtrVal = PredPointer.getAddr();
1397 bool CanTranslate = true;
1398 // If PHI translation was unable to find an available pointer in this
1399 // predecessor, then we have to assume that the pointer is clobbered in
1400 // that predecessor. We can still do PRE of the load, which would insert
1401 // a computation of the pointer in this predecessor.
1403 CanTranslate = false;
1405 // FIXME: it is entirely possible that PHI translating will end up with
1406 // the same value. Consider PHI translating something like:
1407 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1408 // to recurse here, pedantically speaking.
1410 // If getNonLocalPointerDepFromBB fails here, that means the cached
1411 // result conflicted with the Visited list; we have to conservatively
1412 // assume it is unknown, but this also does not block PRE of the load.
1413 if (!CanTranslate ||
1414 getNonLocalPointerDepFromBB(QueryInst, PredPointer,
1415 Loc.getWithNewPtr(PredPtrVal),
1418 // Add the entry to the Result list.
1419 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1420 Result.push_back(Entry);
1422 // Since we had a phi translation failure, the cache for CacheKey won't
1423 // include all of the entries that we need to immediately satisfy future
1424 // queries. Mark this in NonLocalPointerDeps by setting the
1425 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1426 // cached value to do more work but not miss the phi trans failure.
1427 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1428 NLPI.Pair = BBSkipFirstBlockPair();
1433 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1434 CacheInfo = &NonLocalPointerDeps[CacheKey];
1435 Cache = &CacheInfo->NonLocalDeps;
1436 NumSortedEntries = Cache->size();
1438 // Since we did phi translation, the "Cache" set won't contain all of the
1439 // results for the query. This is ok (we can still use it to accelerate
1440 // specific block queries) but we can't do the fastpath "return all
1441 // results from the set" Clear out the indicator for this.
1442 CacheInfo->Pair = BBSkipFirstBlockPair();
1443 SkipFirstBlock = false;
1446 PredTranslationFailure:
1447 // The following code is "failure"; we can't produce a sane translation
1448 // for the given block. It assumes that we haven't modified any of
1449 // our datastructures while processing the current block.
1452 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1453 CacheInfo = &NonLocalPointerDeps[CacheKey];
1454 Cache = &CacheInfo->NonLocalDeps;
1455 NumSortedEntries = Cache->size();
1458 // Since we failed phi translation, the "Cache" set won't contain all of the
1459 // results for the query. This is ok (we can still use it to accelerate
1460 // specific block queries) but we can't do the fastpath "return all
1461 // results from the set". Clear out the indicator for this.
1462 CacheInfo->Pair = BBSkipFirstBlockPair();
1464 // If *nothing* works, mark the pointer as unknown.
1466 // If this is the magic first block, return this as a clobber of the whole
1467 // incoming value. Since we can't phi translate to one of the predecessors,
1468 // we have to bail out.
1472 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1473 assert(I != Cache->rend() && "Didn't find current block??");
1474 if (I->getBB() != BB)
1477 assert((I->getResult().isNonLocal() || !DT->isReachableFromEntry(BB)) &&
1478 "Should only be here with transparent block");
1479 I->setResult(MemDepResult::getUnknown());
1480 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1481 Pointer.getAddr()));
1486 // Okay, we're done now. If we added new values to the cache, re-sort it.
1487 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1488 DEBUG(AssertSorted(*Cache));
1492 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1493 /// CachedNonLocalPointerInfo, remove it.
1494 void MemoryDependenceAnalysis::
1495 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1496 CachedNonLocalPointerInfo::iterator It =
1497 NonLocalPointerDeps.find(P);
1498 if (It == NonLocalPointerDeps.end()) return;
1500 // Remove all of the entries in the BB->val map. This involves removing
1501 // instructions from the reverse map.
1502 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1504 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1505 Instruction *Target = PInfo[i].getResult().getInst();
1506 if (!Target) continue; // Ignore non-local dep results.
1507 assert(Target->getParent() == PInfo[i].getBB());
1509 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1510 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1513 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1514 NonLocalPointerDeps.erase(It);
1518 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1519 /// information about the specified pointer, because it may be too
1520 /// conservative in memdep. This is an optional call that can be used when
1521 /// the client detects an equivalence between the pointer and some other
1522 /// value and replaces the other value with ptr. This can make Ptr available
1523 /// in more places that cached info does not necessarily keep.
1524 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1525 // If Ptr isn't really a pointer, just ignore it.
1526 if (!Ptr->getType()->isPointerTy()) return;
1527 // Flush store info for the pointer.
1528 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1529 // Flush load info for the pointer.
1530 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1533 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1534 /// This needs to be done when the CFG changes, e.g., due to splitting
1536 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1540 /// removeInstruction - Remove an instruction from the dependence analysis,
1541 /// updating the dependence of instructions that previously depended on it.
1542 /// This method attempts to keep the cache coherent using the reverse map.
1543 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1544 // Walk through the Non-local dependencies, removing this one as the value
1545 // for any cached queries.
1546 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1547 if (NLDI != NonLocalDeps.end()) {
1548 NonLocalDepInfo &BlockMap = NLDI->second.first;
1549 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1551 if (Instruction *Inst = DI->getResult().getInst())
1552 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1553 NonLocalDeps.erase(NLDI);
1556 // If we have a cached local dependence query for this instruction, remove it.
1558 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1559 if (LocalDepEntry != LocalDeps.end()) {
1560 // Remove us from DepInst's reverse set now that the local dep info is gone.
1561 if (Instruction *Inst = LocalDepEntry->second.getInst())
1562 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1564 // Remove this local dependency info.
1565 LocalDeps.erase(LocalDepEntry);
1568 // If we have any cached pointer dependencies on this instruction, remove
1569 // them. If the instruction has non-pointer type, then it can't be a pointer
1572 // Remove it from both the load info and the store info. The instruction
1573 // can't be in either of these maps if it is non-pointer.
1574 if (RemInst->getType()->isPointerTy()) {
1575 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1576 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1579 // Loop over all of the things that depend on the instruction we're removing.
1581 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1583 // If we find RemInst as a clobber or Def in any of the maps for other values,
1584 // we need to replace its entry with a dirty version of the instruction after
1585 // it. If RemInst is a terminator, we use a null dirty value.
1587 // Using a dirty version of the instruction after RemInst saves having to scan
1588 // the entire block to get to this point.
1589 MemDepResult NewDirtyVal;
1590 if (!RemInst->isTerminator())
1591 NewDirtyVal = MemDepResult::getDirty(&*++RemInst->getIterator());
1593 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1594 if (ReverseDepIt != ReverseLocalDeps.end()) {
1595 // RemInst can't be the terminator if it has local stuff depending on it.
1596 assert(!ReverseDepIt->second.empty() && !isa<TerminatorInst>(RemInst) &&
1597 "Nothing can locally depend on a terminator");
1599 for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) {
1600 assert(InstDependingOnRemInst != RemInst &&
1601 "Already removed our local dep info");
1603 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1605 // Make sure to remember that new things depend on NewDepInst.
1606 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1607 "a local dep on this if it is a terminator!");
1608 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1609 InstDependingOnRemInst));
1612 ReverseLocalDeps.erase(ReverseDepIt);
1614 // Add new reverse deps after scanning the set, to avoid invalidating the
1615 // 'ReverseDeps' reference.
1616 while (!ReverseDepsToAdd.empty()) {
1617 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1618 .insert(ReverseDepsToAdd.back().second);
1619 ReverseDepsToAdd.pop_back();
1623 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1624 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1625 for (Instruction *I : ReverseDepIt->second) {
1626 assert(I != RemInst && "Already removed NonLocalDep info for RemInst");
1628 PerInstNLInfo &INLD = NonLocalDeps[I];
1629 // The information is now dirty!
1632 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1633 DE = INLD.first.end(); DI != DE; ++DI) {
1634 if (DI->getResult().getInst() != RemInst) continue;
1636 // Convert to a dirty entry for the subsequent instruction.
1637 DI->setResult(NewDirtyVal);
1639 if (Instruction *NextI = NewDirtyVal.getInst())
1640 ReverseDepsToAdd.push_back(std::make_pair(NextI, I));
1644 ReverseNonLocalDeps.erase(ReverseDepIt);
1646 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1647 while (!ReverseDepsToAdd.empty()) {
1648 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1649 .insert(ReverseDepsToAdd.back().second);
1650 ReverseDepsToAdd.pop_back();
1654 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1655 // value in the NonLocalPointerDeps info.
1656 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1657 ReverseNonLocalPtrDeps.find(RemInst);
1658 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1659 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1661 for (ValueIsLoadPair P : ReversePtrDepIt->second) {
1662 assert(P.getPointer() != RemInst &&
1663 "Already removed NonLocalPointerDeps info for RemInst");
1665 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1667 // The cache is not valid for any specific block anymore.
1668 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1670 // Update any entries for RemInst to use the instruction after it.
1671 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1673 if (DI->getResult().getInst() != RemInst) continue;
1675 // Convert to a dirty entry for the subsequent instruction.
1676 DI->setResult(NewDirtyVal);
1678 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1679 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1682 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1683 // subsequent value may invalidate the sortedness.
1684 std::sort(NLPDI.begin(), NLPDI.end());
1687 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1689 while (!ReversePtrDepsToAdd.empty()) {
1690 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1691 .insert(ReversePtrDepsToAdd.back().second);
1692 ReversePtrDepsToAdd.pop_back();
1697 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1698 DEBUG(verifyRemoved(RemInst));
1700 /// verifyRemoved - Verify that the specified instruction does not occur
1701 /// in our internal data structures. This function verifies by asserting in
1703 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1705 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1706 E = LocalDeps.end(); I != E; ++I) {
1707 assert(I->first != D && "Inst occurs in data structures");
1708 assert(I->second.getInst() != D &&
1709 "Inst occurs in data structures");
1712 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1713 E = NonLocalPointerDeps.end(); I != E; ++I) {
1714 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1715 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1716 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1718 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1721 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1722 E = NonLocalDeps.end(); I != E; ++I) {
1723 assert(I->first != D && "Inst occurs in data structures");
1724 const PerInstNLInfo &INLD = I->second;
1725 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1726 EE = INLD.first.end(); II != EE; ++II)
1727 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1730 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1731 E = ReverseLocalDeps.end(); I != E; ++I) {
1732 assert(I->first != D && "Inst occurs in data structures");
1733 for (Instruction *Inst : I->second)
1734 assert(Inst != D && "Inst occurs in data structures");
1737 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1738 E = ReverseNonLocalDeps.end();
1740 assert(I->first != D && "Inst occurs in data structures");
1741 for (Instruction *Inst : I->second)
1742 assert(Inst != D && "Inst occurs in data structures");
1745 for (ReverseNonLocalPtrDepTy::const_iterator
1746 I = ReverseNonLocalPtrDeps.begin(),
1747 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1748 assert(I->first != D && "Inst occurs in rev NLPD map");
1750 for (ValueIsLoadPair P : I->second)
1751 assert(P != ValueIsLoadPair(D, false) &&
1752 P != ValueIsLoadPair(D, true) &&
1753 "Inst occurs in ReverseNonLocalPtrDeps map");