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/SmallSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/AssumptionCache.h"
24 #include "llvm/Analysis/MemoryBuiltins.h"
25 #include "llvm/Analysis/PHITransAddr.h"
26 #include "llvm/Analysis/OrderedBasicBlock.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/Analysis/TargetLibraryInfo.h"
29 #include "llvm/IR/CallSite.h"
30 #include "llvm/IR/Constants.h"
31 #include "llvm/IR/DataLayout.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/Dominators.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/IR/Instruction.h"
36 #include "llvm/IR/Instructions.h"
37 #include "llvm/IR/IntrinsicInst.h"
38 #include "llvm/IR/LLVMContext.h"
39 #include "llvm/IR/PredIteratorCache.h"
40 #include "llvm/Support/AtomicOrdering.h"
41 #include "llvm/Support/Casting.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Compiler.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/MathExtras.h"
52 #define DEBUG_TYPE "memdep"
54 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
55 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
56 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
58 STATISTIC(NumCacheNonLocalPtr,
59 "Number of fully cached non-local ptr responses");
60 STATISTIC(NumCacheDirtyNonLocalPtr,
61 "Number of cached, but dirty, non-local ptr responses");
62 STATISTIC(NumUncacheNonLocalPtr, "Number of uncached non-local ptr responses");
63 STATISTIC(NumCacheCompleteNonLocalPtr,
64 "Number of block queries that were completely cached");
66 // Limit for the number of instructions to scan in a block.
68 static cl::opt<unsigned> BlockScanLimit(
69 "memdep-block-scan-limit", cl::Hidden, cl::init(100),
70 cl::desc("The number of instructions to scan in a block in memory "
71 "dependency analysis (default = 100)"));
73 static cl::opt<unsigned>
74 BlockNumberLimit("memdep-block-number-limit", cl::Hidden, cl::init(1000),
75 cl::desc("The number of blocks to scan during memory "
76 "dependency analysis (default = 1000)"));
78 // Limit on the number of memdep results to process.
79 static const unsigned int NumResultsLimit = 100;
81 /// This is a helper function that removes Val from 'Inst's set in ReverseMap.
83 /// If the set becomes empty, remove Inst's entry.
84 template <typename KeyTy>
86 RemoveFromReverseMap(DenseMap<Instruction *, SmallPtrSet<KeyTy, 4>> &ReverseMap,
87 Instruction *Inst, KeyTy Val) {
88 typename DenseMap<Instruction *, SmallPtrSet<KeyTy, 4>>::iterator InstIt =
89 ReverseMap.find(Inst);
90 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
91 bool Found = InstIt->second.erase(Val);
92 assert(Found && "Invalid reverse map!");
94 if (InstIt->second.empty())
95 ReverseMap.erase(InstIt);
98 /// If the given instruction references a specific memory location, fill in Loc
99 /// with the details, otherwise set Loc.Ptr to null.
101 /// Returns a ModRefInfo value describing the general behavior of the
103 static ModRefInfo GetLocation(const Instruction *Inst, MemoryLocation &Loc,
104 const TargetLibraryInfo &TLI) {
105 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
106 if (LI->isUnordered()) {
107 Loc = MemoryLocation::get(LI);
110 if (LI->getOrdering() == AtomicOrdering::Monotonic) {
111 Loc = MemoryLocation::get(LI);
114 Loc = MemoryLocation();
118 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
119 if (SI->isUnordered()) {
120 Loc = MemoryLocation::get(SI);
123 if (SI->getOrdering() == AtomicOrdering::Monotonic) {
124 Loc = MemoryLocation::get(SI);
127 Loc = MemoryLocation();
131 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
132 Loc = MemoryLocation::get(V);
136 if (const CallInst *CI = isFreeCall(Inst, &TLI)) {
137 // calls to free() deallocate the entire structure
138 Loc = MemoryLocation(CI->getArgOperand(0));
142 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
145 switch (II->getIntrinsicID()) {
146 case Intrinsic::lifetime_start:
147 case Intrinsic::lifetime_end:
148 case Intrinsic::invariant_start:
149 II->getAAMetadata(AAInfo);
150 Loc = MemoryLocation(
151 II->getArgOperand(1),
152 cast<ConstantInt>(II->getArgOperand(0))->getZExtValue(), AAInfo);
153 // These intrinsics don't really modify the memory, but returning Mod
154 // will allow them to be handled conservatively.
156 case Intrinsic::invariant_end:
157 II->getAAMetadata(AAInfo);
158 Loc = MemoryLocation(
159 II->getArgOperand(2),
160 cast<ConstantInt>(II->getArgOperand(1))->getZExtValue(), AAInfo);
161 // These intrinsics don't really modify the memory, but returning Mod
162 // will allow them to be handled conservatively.
169 // Otherwise, just do the coarse-grained thing that always works.
170 if (Inst->mayWriteToMemory())
172 if (Inst->mayReadFromMemory())
177 /// Private helper for finding the local dependencies of a call site.
178 MemDepResult MemoryDependenceResults::getCallSiteDependencyFrom(
179 CallSite CS, bool isReadOnlyCall, BasicBlock::iterator ScanIt,
181 unsigned Limit = BlockScanLimit;
183 // Walk backwards through the block, looking for dependencies.
184 while (ScanIt != BB->begin()) {
185 // Limit the amount of scanning we do so we don't end up with quadratic
186 // running time on extreme testcases.
189 return MemDepResult::getUnknown();
191 Instruction *Inst = &*--ScanIt;
193 // If this inst is a memory op, get the pointer it accessed
195 ModRefInfo MR = GetLocation(Inst, Loc, TLI);
197 // A simple instruction.
198 if (AA.getModRefInfo(CS, Loc) != MRI_NoModRef)
199 return MemDepResult::getClobber(Inst);
203 if (auto InstCS = CallSite(Inst)) {
204 // Debug intrinsics don't cause dependences.
205 if (isa<DbgInfoIntrinsic>(Inst))
207 // If these two calls do not interfere, look past it.
208 switch (AA.getModRefInfo(CS, InstCS)) {
210 // If the two calls are the same, return InstCS as a Def, so that
211 // CS can be found redundant and eliminated.
212 if (isReadOnlyCall && !(MR & MRI_Mod) &&
213 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
214 return MemDepResult::getDef(Inst);
216 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
220 return MemDepResult::getClobber(Inst);
224 // If we could not obtain a pointer for the instruction and the instruction
225 // touches memory then assume that this is a dependency.
226 if (MR != MRI_NoModRef)
227 return MemDepResult::getClobber(Inst);
230 // No dependence found. If this is the entry block of the function, it is
231 // unknown, otherwise it is non-local.
232 if (BB != &BB->getParent()->getEntryBlock())
233 return MemDepResult::getNonLocal();
234 return MemDepResult::getNonFuncLocal();
237 unsigned MemoryDependenceResults::getLoadLoadClobberFullWidthSize(
238 const Value *MemLocBase, int64_t MemLocOffs, unsigned MemLocSize,
239 const LoadInst *LI) {
240 // We can only extend simple integer loads.
241 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple())
244 // Load widening is hostile to ThreadSanitizer: it may cause false positives
245 // or make the reports more cryptic (access sizes are wrong).
246 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
249 const DataLayout &DL = LI->getModule()->getDataLayout();
251 // Get the base of this load.
253 const Value *LIBase =
254 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, DL);
256 // If the two pointers are not based on the same pointer, we can't tell that
258 if (LIBase != MemLocBase)
261 // Okay, the two values are based on the same pointer, but returned as
262 // no-alias. This happens when we have things like two byte loads at "P+1"
263 // and "P+3". Check to see if increasing the size of the "LI" load up to its
264 // alignment (or the largest native integer type) will allow us to load all
265 // the bits required by MemLoc.
267 // If MemLoc is before LI, then no widening of LI will help us out.
268 if (MemLocOffs < LIOffs)
271 // Get the alignment of the load in bytes. We assume that it is safe to load
272 // any legal integer up to this size without a problem. For example, if we're
273 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
274 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
276 unsigned LoadAlign = LI->getAlignment();
278 int64_t MemLocEnd = MemLocOffs + MemLocSize;
280 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
281 if (LIOffs + LoadAlign < MemLocEnd)
284 // This is the size of the load to try. Start with the next larger power of
286 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits() / 8U;
287 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
290 // If this load size is bigger than our known alignment or would not fit
291 // into a native integer register, then we fail.
292 if (NewLoadByteSize > LoadAlign ||
293 !DL.fitsInLegalInteger(NewLoadByteSize * 8))
296 if (LIOffs + NewLoadByteSize > MemLocEnd &&
297 LI->getParent()->getParent()->hasFnAttribute(
298 Attribute::SanitizeAddress))
299 // We will be reading past the location accessed by the original program.
300 // While this is safe in a regular build, Address Safety analysis tools
301 // may start reporting false warnings. So, don't do widening.
304 // If a load of this width would include all of MemLoc, then we succeed.
305 if (LIOffs + NewLoadByteSize >= MemLocEnd)
306 return NewLoadByteSize;
308 NewLoadByteSize <<= 1;
312 static bool isVolatile(Instruction *Inst) {
313 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
314 return LI->isVolatile();
315 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
316 return SI->isVolatile();
317 else if (AtomicCmpXchgInst *AI = dyn_cast<AtomicCmpXchgInst>(Inst))
318 return AI->isVolatile();
322 MemDepResult MemoryDependenceResults::getPointerDependencyFrom(
323 const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
324 BasicBlock *BB, Instruction *QueryInst, unsigned *Limit) {
326 MemDepResult InvariantGroupDependency = MemDepResult::getUnknown();
327 if (QueryInst != nullptr) {
328 if (auto *LI = dyn_cast<LoadInst>(QueryInst)) {
329 InvariantGroupDependency = getInvariantGroupPointerDependency(LI, BB);
331 if (InvariantGroupDependency.isDef())
332 return InvariantGroupDependency;
335 MemDepResult SimpleDep = getSimplePointerDependencyFrom(
336 MemLoc, isLoad, ScanIt, BB, QueryInst, Limit);
337 if (SimpleDep.isDef())
339 // Non-local invariant group dependency indicates there is non local Def
340 // (it only returns nonLocal if it finds nonLocal def), which is better than
341 // local clobber and everything else.
342 if (InvariantGroupDependency.isNonLocal())
343 return InvariantGroupDependency;
345 assert(InvariantGroupDependency.isUnknown() &&
346 "InvariantGroupDependency should be only unknown at this point");
351 MemoryDependenceResults::getInvariantGroupPointerDependency(LoadInst *LI,
354 auto *InvariantGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group);
355 if (!InvariantGroupMD)
356 return MemDepResult::getUnknown();
358 // Take the ptr operand after all casts and geps 0. This way we can search
359 // cast graph down only.
360 Value *LoadOperand = LI->getPointerOperand()->stripPointerCasts();
362 // It's is not safe to walk the use list of global value, because function
363 // passes aren't allowed to look outside their functions.
364 // FIXME: this could be fixed by filtering instructions from outside
365 // of current function.
366 if (isa<GlobalValue>(LoadOperand))
367 return MemDepResult::getUnknown();
369 // Queue to process all pointers that are equivalent to load operand.
370 SmallVector<const Value *, 8> LoadOperandsQueue;
371 LoadOperandsQueue.push_back(LoadOperand);
373 Instruction *ClosestDependency = nullptr;
374 // Order of instructions in uses list is unpredictible. In order to always
375 // get the same result, we will look for the closest dominance.
376 auto GetClosestDependency = [this](Instruction *Best, Instruction *Other) {
377 assert(Other && "Must call it with not null instruction");
378 if (Best == nullptr || DT.dominates(Best, Other))
384 // FIXME: This loop is O(N^2) because dominates can be O(n) and in worst case
385 // we will see all the instructions. This should be fixed in MSSA.
386 while (!LoadOperandsQueue.empty()) {
387 const Value *Ptr = LoadOperandsQueue.pop_back_val();
388 assert(Ptr && !isa<GlobalValue>(Ptr) &&
389 "Null or GlobalValue should not be inserted");
391 for (const Use &Us : Ptr->uses()) {
392 auto *U = dyn_cast<Instruction>(Us.getUser());
393 if (!U || U == LI || !DT.dominates(U, LI))
396 // Bitcast or gep with zeros are using Ptr. Add to queue to check it's
397 // users. U = bitcast Ptr
398 if (isa<BitCastInst>(U)) {
399 LoadOperandsQueue.push_back(U);
402 // Gep with zeros is equivalent to bitcast.
403 // FIXME: we are not sure if some bitcast should be canonicalized to gep 0
404 // or gep 0 to bitcast because of SROA, so there are 2 forms. When
405 // typeless pointers will be ready then both cases will be gone
406 // (and this BFS also won't be needed).
407 if (auto *GEP = dyn_cast<GetElementPtrInst>(U))
408 if (GEP->hasAllZeroIndices()) {
409 LoadOperandsQueue.push_back(U);
413 // If we hit load/store with the same invariant.group metadata (and the
414 // same pointer operand) we can assume that value pointed by pointer
415 // operand didn't change.
416 if ((isa<LoadInst>(U) || isa<StoreInst>(U)) &&
417 U->getMetadata(LLVMContext::MD_invariant_group) == InvariantGroupMD)
418 ClosestDependency = GetClosestDependency(ClosestDependency, U);
422 if (!ClosestDependency)
423 return MemDepResult::getUnknown();
424 if (ClosestDependency->getParent() == BB)
425 return MemDepResult::getDef(ClosestDependency);
426 // Def(U) can't be returned here because it is non-local. If local
427 // dependency won't be found then return nonLocal counting that the
428 // user will call getNonLocalPointerDependency, which will return cached
430 NonLocalDefsCache.try_emplace(
431 LI, NonLocalDepResult(ClosestDependency->getParent(),
432 MemDepResult::getDef(ClosestDependency), nullptr));
433 return MemDepResult::getNonLocal();
436 MemDepResult MemoryDependenceResults::getSimplePointerDependencyFrom(
437 const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
438 BasicBlock *BB, Instruction *QueryInst, unsigned *Limit) {
439 bool isInvariantLoad = false;
442 unsigned DefaultLimit = BlockScanLimit;
443 return getSimplePointerDependencyFrom(MemLoc, isLoad, ScanIt, BB, QueryInst,
447 // We must be careful with atomic accesses, as they may allow another thread
448 // to touch this location, clobbering it. We are conservative: if the
449 // QueryInst is not a simple (non-atomic) memory access, we automatically
450 // return getClobber.
451 // If it is simple, we know based on the results of
452 // "Compiler testing via a theory of sound optimisations in the C11/C++11
453 // memory model" in PLDI 2013, that a non-atomic location can only be
454 // clobbered between a pair of a release and an acquire action, with no
455 // access to the location in between.
456 // Here is an example for giving the general intuition behind this rule.
457 // In the following code:
459 // release action; [1]
460 // acquire action; [4]
462 // It is unsafe to replace %val by 0 because another thread may be running:
463 // acquire action; [2]
465 // release action; [3]
466 // with synchronization from 1 to 2 and from 3 to 4, resulting in %val
467 // being 42. A key property of this program however is that if either
468 // 1 or 4 were missing, there would be a race between the store of 42
469 // either the store of 0 or the load (making the whole program racy).
470 // The paper mentioned above shows that the same property is respected
471 // by every program that can detect any optimization of that kind: either
472 // it is racy (undefined) or there is a release followed by an acquire
473 // between the pair of accesses under consideration.
475 // If the load is invariant, we "know" that it doesn't alias *any* write. We
476 // do want to respect mustalias results since defs are useful for value
477 // forwarding, but any mayalias write can be assumed to be noalias.
478 // Arguably, this logic should be pushed inside AliasAnalysis itself.
479 if (isLoad && QueryInst) {
480 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
481 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
482 isInvariantLoad = true;
485 const DataLayout &DL = BB->getModule()->getDataLayout();
487 // Create a numbered basic block to lazily compute and cache instruction
488 // positions inside a BB. This is used to provide fast queries for relative
489 // position between two instructions in a BB and can be used by
490 // AliasAnalysis::callCapturesBefore.
491 OrderedBasicBlock OBB(BB);
493 // Return "true" if and only if the instruction I is either a non-simple
494 // load or a non-simple store.
495 auto isNonSimpleLoadOrStore = [](Instruction *I) -> bool {
496 if (auto *LI = dyn_cast<LoadInst>(I))
497 return !LI->isSimple();
498 if (auto *SI = dyn_cast<StoreInst>(I))
499 return !SI->isSimple();
503 // Return "true" if I is not a load and not a store, but it does access
505 auto isOtherMemAccess = [](Instruction *I) -> bool {
506 return !isa<LoadInst>(I) && !isa<StoreInst>(I) && I->mayReadOrWriteMemory();
509 // Walk backwards through the basic block, looking for dependencies.
510 while (ScanIt != BB->begin()) {
511 Instruction *Inst = &*--ScanIt;
513 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
514 // Debug intrinsics don't (and can't) cause dependencies.
515 if (isa<DbgInfoIntrinsic>(II))
518 // Limit the amount of scanning we do so we don't end up with quadratic
519 // running time on extreme testcases.
522 return MemDepResult::getUnknown();
524 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
525 // If we reach a lifetime begin or end marker, then the query ends here
526 // because the value is undefined.
527 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
528 // FIXME: This only considers queries directly on the invariant-tagged
529 // pointer, not on query pointers that are indexed off of them. It'd
530 // be nice to handle that at some point (the right approach is to use
531 // GetPointerBaseWithConstantOffset).
532 if (AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), MemLoc))
533 return MemDepResult::getDef(II);
538 // Values depend on loads if the pointers are must aliased. This means
539 // that a load depends on another must aliased load from the same value.
540 // One exception is atomic loads: a value can depend on an atomic load that
541 // it does not alias with when this atomic load indicates that another
542 // thread may be accessing the location.
543 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
545 // While volatile access cannot be eliminated, they do not have to clobber
546 // non-aliasing locations, as normal accesses, for example, can be safely
547 // reordered with volatile accesses.
548 if (LI->isVolatile()) {
550 // Original QueryInst *may* be volatile
551 return MemDepResult::getClobber(LI);
552 if (isVolatile(QueryInst))
553 // Ordering required if QueryInst is itself volatile
554 return MemDepResult::getClobber(LI);
555 // Otherwise, volatile doesn't imply any special ordering
558 // Atomic loads have complications involved.
559 // A Monotonic (or higher) load is OK if the query inst is itself not
561 // FIXME: This is overly conservative.
562 if (LI->isAtomic() && isStrongerThanUnordered(LI->getOrdering())) {
563 if (!QueryInst || isNonSimpleLoadOrStore(QueryInst) ||
564 isOtherMemAccess(QueryInst))
565 return MemDepResult::getClobber(LI);
566 if (LI->getOrdering() != AtomicOrdering::Monotonic)
567 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);
579 // Must aliased loads are defs of each other.
581 return MemDepResult::getDef(Inst);
583 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
584 // in terms of clobbering loads, but since it does this by looking
585 // at the clobbering load directly, it doesn't know about any
586 // phi translation that may have happened along the way.
588 // If we have a partial alias, then return this as a clobber for the
590 if (R == PartialAlias)
591 return MemDepResult::getClobber(Inst);
594 // Random may-alias loads don't depend on each other without a
599 // Stores don't depend on other no-aliased accesses.
603 // Stores don't alias loads from read-only memory.
604 if (AA.pointsToConstantMemory(LoadLoc))
607 // Stores depend on may/must aliased loads.
608 return MemDepResult::getDef(Inst);
611 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
612 // Atomic stores have complications involved.
613 // A Monotonic store is OK if the query inst is itself not atomic.
614 // FIXME: This is overly conservative.
615 if (!SI->isUnordered() && SI->isAtomic()) {
616 if (!QueryInst || isNonSimpleLoadOrStore(QueryInst) ||
617 isOtherMemAccess(QueryInst))
618 return MemDepResult::getClobber(SI);
619 if (SI->getOrdering() != AtomicOrdering::Monotonic)
620 return MemDepResult::getClobber(SI);
623 // FIXME: this is overly conservative.
624 // While volatile access cannot be eliminated, they do not have to clobber
625 // non-aliasing locations, as normal accesses can for example be reordered
626 // with volatile accesses.
627 if (SI->isVolatile())
628 if (!QueryInst || isNonSimpleLoadOrStore(QueryInst) ||
629 isOtherMemAccess(QueryInst))
630 return MemDepResult::getClobber(SI);
632 // If alias analysis can tell that this store is guaranteed to not modify
633 // the query pointer, ignore it. Use getModRefInfo to handle cases where
634 // the query pointer points to constant memory etc.
635 if (AA.getModRefInfo(SI, MemLoc) == MRI_NoModRef)
638 // Ok, this store might clobber the query pointer. Check to see if it is
639 // a must alias: in this case, we want to return this as a def.
640 MemoryLocation StoreLoc = MemoryLocation::get(SI);
642 // If we found a pointer, check if it could be the same as our pointer.
643 AliasResult R = AA.alias(StoreLoc, MemLoc);
648 return MemDepResult::getDef(Inst);
651 return MemDepResult::getClobber(Inst);
654 // If this is an allocation, and if we know that the accessed pointer is to
655 // the allocation, return Def. This means that there is no dependence and
656 // the access can be optimized based on that. For example, a load could
657 // turn into undef. Note that we can bypass the allocation itself when
658 // looking for a clobber in many cases; that's an alias property and is
659 // handled by BasicAA.
660 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, &TLI)) {
661 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
662 if (AccessPtr == Inst || AA.isMustAlias(Inst, AccessPtr))
663 return MemDepResult::getDef(Inst);
669 // A release fence requires that all stores complete before it, but does
670 // not prevent the reordering of following loads or stores 'before' the
671 // fence. As a result, we look past it when finding a dependency for
672 // loads. DSE uses this to find preceeding stores to delete and thus we
673 // can't bypass the fence if the query instruction is a store.
674 if (FenceInst *FI = dyn_cast<FenceInst>(Inst))
675 if (isLoad && FI->getOrdering() == AtomicOrdering::Release)
678 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
679 ModRefInfo MR = AA.getModRefInfo(Inst, MemLoc);
680 // If necessary, perform additional analysis.
681 if (MR == MRI_ModRef)
682 MR = AA.callCapturesBefore(Inst, MemLoc, &DT, &OBB);
685 // If the call has no effect on the queried pointer, just ignore it.
688 return MemDepResult::getClobber(Inst);
690 // If the call is known to never store to the pointer, and if this is a
691 // load query, we can safely ignore it (scan past it).
695 // Otherwise, there is a potential dependence. Return a clobber.
696 return MemDepResult::getClobber(Inst);
700 // No dependence found. If this is the entry block of the function, it is
701 // unknown, otherwise it is non-local.
702 if (BB != &BB->getParent()->getEntryBlock())
703 return MemDepResult::getNonLocal();
704 return MemDepResult::getNonFuncLocal();
707 MemDepResult MemoryDependenceResults::getDependency(Instruction *QueryInst) {
708 Instruction *ScanPos = QueryInst;
710 // Check for a cached result
711 MemDepResult &LocalCache = LocalDeps[QueryInst];
713 // If the cached entry is non-dirty, just return it. Note that this depends
714 // on MemDepResult's default constructing to 'dirty'.
715 if (!LocalCache.isDirty())
718 // Otherwise, if we have a dirty entry, we know we can start the scan at that
719 // instruction, which may save us some work.
720 if (Instruction *Inst = LocalCache.getInst()) {
723 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
726 BasicBlock *QueryParent = QueryInst->getParent();
729 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
730 // No dependence found. If this is the entry block of the function, it is
731 // unknown, otherwise it is non-local.
732 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
733 LocalCache = MemDepResult::getNonLocal();
735 LocalCache = MemDepResult::getNonFuncLocal();
737 MemoryLocation MemLoc;
738 ModRefInfo MR = GetLocation(QueryInst, MemLoc, TLI);
740 // If we can do a pointer scan, make it happen.
741 bool isLoad = !(MR & MRI_Mod);
742 if (auto *II = dyn_cast<IntrinsicInst>(QueryInst))
743 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
745 LocalCache = getPointerDependencyFrom(
746 MemLoc, isLoad, ScanPos->getIterator(), QueryParent, QueryInst);
747 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
748 CallSite QueryCS(QueryInst);
749 bool isReadOnly = AA.onlyReadsMemory(QueryCS);
750 LocalCache = getCallSiteDependencyFrom(
751 QueryCS, isReadOnly, ScanPos->getIterator(), QueryParent);
753 // Non-memory instruction.
754 LocalCache = MemDepResult::getUnknown();
757 // Remember the result!
758 if (Instruction *I = LocalCache.getInst())
759 ReverseLocalDeps[I].insert(QueryInst);
765 /// This method is used when -debug is specified to verify that cache arrays
766 /// are properly kept sorted.
767 static void AssertSorted(MemoryDependenceResults::NonLocalDepInfo &Cache,
770 Count = Cache.size();
771 assert(std::is_sorted(Cache.begin(), Cache.begin() + Count) &&
772 "Cache isn't sorted!");
776 const MemoryDependenceResults::NonLocalDepInfo &
777 MemoryDependenceResults::getNonLocalCallDependency(CallSite QueryCS) {
778 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
779 "getNonLocalCallDependency should only be used on calls with "
781 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
782 NonLocalDepInfo &Cache = CacheP.first;
784 // This is the set of blocks that need to be recomputed. In the cached case,
785 // this can happen due to instructions being deleted etc. In the uncached
786 // case, this starts out as the set of predecessors we care about.
787 SmallVector<BasicBlock *, 32> DirtyBlocks;
789 if (!Cache.empty()) {
790 // Okay, we have a cache entry. If we know it is not dirty, just return it
791 // with no computation.
792 if (!CacheP.second) {
797 // If we already have a partially computed set of results, scan them to
798 // determine what is dirty, seeding our initial DirtyBlocks worklist.
799 for (auto &Entry : Cache)
800 if (Entry.getResult().isDirty())
801 DirtyBlocks.push_back(Entry.getBB());
803 // Sort the cache so that we can do fast binary search lookups below.
804 std::sort(Cache.begin(), Cache.end());
806 ++NumCacheDirtyNonLocal;
807 // cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
808 // << Cache.size() << " cached: " << *QueryInst;
810 // Seed DirtyBlocks with each of the preds of QueryInst's block.
811 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
812 for (BasicBlock *Pred : PredCache.get(QueryBB))
813 DirtyBlocks.push_back(Pred);
814 ++NumUncacheNonLocal;
817 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
818 bool isReadonlyCall = AA.onlyReadsMemory(QueryCS);
820 SmallPtrSet<BasicBlock *, 32> Visited;
822 unsigned NumSortedEntries = Cache.size();
823 DEBUG(AssertSorted(Cache));
825 // Iterate while we still have blocks to update.
826 while (!DirtyBlocks.empty()) {
827 BasicBlock *DirtyBB = DirtyBlocks.back();
828 DirtyBlocks.pop_back();
830 // Already processed this block?
831 if (!Visited.insert(DirtyBB).second)
834 // Do a binary search to see if we already have an entry for this block in
835 // the cache set. If so, find it.
836 DEBUG(AssertSorted(Cache, NumSortedEntries));
837 NonLocalDepInfo::iterator Entry =
838 std::upper_bound(Cache.begin(), Cache.begin() + NumSortedEntries,
839 NonLocalDepEntry(DirtyBB));
840 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
843 NonLocalDepEntry *ExistingResult = nullptr;
844 if (Entry != Cache.begin() + NumSortedEntries &&
845 Entry->getBB() == DirtyBB) {
846 // If we already have an entry, and if it isn't already dirty, the block
848 if (!Entry->getResult().isDirty())
851 // Otherwise, remember this slot so we can update the value.
852 ExistingResult = &*Entry;
855 // If the dirty entry has a pointer, start scanning from it so we don't have
856 // to rescan the entire block.
857 BasicBlock::iterator ScanPos = DirtyBB->end();
858 if (ExistingResult) {
859 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
860 ScanPos = Inst->getIterator();
861 // We're removing QueryInst's use of Inst.
862 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
863 QueryCS.getInstruction());
867 // Find out if this block has a local dependency for QueryInst.
870 if (ScanPos != DirtyBB->begin()) {
872 getCallSiteDependencyFrom(QueryCS, isReadonlyCall, ScanPos, DirtyBB);
873 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
874 // No dependence found. If this is the entry block of the function, it is
875 // a clobber, otherwise it is unknown.
876 Dep = MemDepResult::getNonLocal();
878 Dep = MemDepResult::getNonFuncLocal();
881 // If we had a dirty entry for the block, update it. Otherwise, just add
884 ExistingResult->setResult(Dep);
886 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
888 // If the block has a dependency (i.e. it isn't completely transparent to
889 // the value), remember the association!
890 if (!Dep.isNonLocal()) {
891 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
892 // update this when we remove instructions.
893 if (Instruction *Inst = Dep.getInst())
894 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
897 // If the block *is* completely transparent to the load, we need to check
898 // the predecessors of this block. Add them to our worklist.
899 for (BasicBlock *Pred : PredCache.get(DirtyBB))
900 DirtyBlocks.push_back(Pred);
907 void MemoryDependenceResults::getNonLocalPointerDependency(
908 Instruction *QueryInst, SmallVectorImpl<NonLocalDepResult> &Result) {
909 const MemoryLocation Loc = MemoryLocation::get(QueryInst);
910 bool isLoad = isa<LoadInst>(QueryInst);
911 BasicBlock *FromBB = QueryInst->getParent();
914 assert(Loc.Ptr->getType()->isPointerTy() &&
915 "Can't get pointer deps of a non-pointer!");
918 // Check if there is cached Def with invariant.group. FIXME: cache might be
919 // invalid if cached instruction would be removed between call to
920 // getPointerDependencyFrom and this function.
921 auto NonLocalDefIt = NonLocalDefsCache.find(QueryInst);
922 if (NonLocalDefIt != NonLocalDefsCache.end()) {
923 Result.push_back(std::move(NonLocalDefIt->second));
924 NonLocalDefsCache.erase(NonLocalDefIt);
928 // This routine does not expect to deal with volatile instructions.
929 // Doing so would require piping through the QueryInst all the way through.
930 // TODO: volatiles can't be elided, but they can be reordered with other
931 // non-volatile accesses.
933 // We currently give up on any instruction which is ordered, but we do handle
934 // atomic instructions which are unordered.
935 // TODO: Handle ordered instructions
936 auto isOrdered = [](Instruction *Inst) {
937 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
938 return !LI->isUnordered();
939 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
940 return !SI->isUnordered();
944 if (isVolatile(QueryInst) || isOrdered(QueryInst)) {
945 Result.push_back(NonLocalDepResult(FromBB, MemDepResult::getUnknown(),
946 const_cast<Value *>(Loc.Ptr)));
949 const DataLayout &DL = FromBB->getModule()->getDataLayout();
950 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, &AC);
952 // This is the set of blocks we've inspected, and the pointer we consider in
953 // each block. Because of critical edges, we currently bail out if querying
954 // a block with multiple different pointers. This can happen during PHI
956 DenseMap<BasicBlock *, Value *> Visited;
957 if (getNonLocalPointerDepFromBB(QueryInst, Address, Loc, isLoad, FromBB,
958 Result, Visited, true))
961 Result.push_back(NonLocalDepResult(FromBB, MemDepResult::getUnknown(),
962 const_cast<Value *>(Loc.Ptr)));
965 /// Compute the memdep value for BB with Pointer/PointeeSize using either
966 /// cached information in Cache or by doing a lookup (which may use dirty cache
967 /// info if available).
969 /// If we do a lookup, add the result to the cache.
970 MemDepResult MemoryDependenceResults::GetNonLocalInfoForBlock(
971 Instruction *QueryInst, const MemoryLocation &Loc, bool isLoad,
972 BasicBlock *BB, NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
974 // Do a binary search to see if we already have an entry for this block in
975 // the cache set. If so, find it.
976 NonLocalDepInfo::iterator Entry = std::upper_bound(
977 Cache->begin(), Cache->begin() + NumSortedEntries, NonLocalDepEntry(BB));
978 if (Entry != Cache->begin() && (Entry - 1)->getBB() == BB)
981 NonLocalDepEntry *ExistingResult = nullptr;
982 if (Entry != Cache->begin() + NumSortedEntries && Entry->getBB() == BB)
983 ExistingResult = &*Entry;
985 // If we have a cached entry, and it is non-dirty, use it as the value for
987 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
988 ++NumCacheNonLocalPtr;
989 return ExistingResult->getResult();
992 // Otherwise, we have to scan for the value. If we have a dirty cache
993 // entry, start scanning from its position, otherwise we scan from the end
995 BasicBlock::iterator ScanPos = BB->end();
996 if (ExistingResult && ExistingResult->getResult().getInst()) {
997 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
998 "Instruction invalidated?");
999 ++NumCacheDirtyNonLocalPtr;
1000 ScanPos = ExistingResult->getResult().getInst()->getIterator();
1002 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1003 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
1004 RemoveFromReverseMap(ReverseNonLocalPtrDeps, &*ScanPos, CacheKey);
1006 ++NumUncacheNonLocalPtr;
1009 // Scan the block for the dependency.
1011 getPointerDependencyFrom(Loc, isLoad, ScanPos, BB, QueryInst);
1013 // If we had a dirty entry for the block, update it. Otherwise, just add
1016 ExistingResult->setResult(Dep);
1018 Cache->push_back(NonLocalDepEntry(BB, Dep));
1020 // If the block has a dependency (i.e. it isn't completely transparent to
1021 // the value), remember the reverse association because we just added it
1023 if (!Dep.isDef() && !Dep.isClobber())
1026 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
1027 // update MemDep when we remove instructions.
1028 Instruction *Inst = Dep.getInst();
1029 assert(Inst && "Didn't depend on anything?");
1030 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
1031 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
1035 /// Sort the NonLocalDepInfo cache, given a certain number of elements in the
1036 /// array that are already properly ordered.
1038 /// This is optimized for the case when only a few entries are added.
1040 SortNonLocalDepInfoCache(MemoryDependenceResults::NonLocalDepInfo &Cache,
1041 unsigned NumSortedEntries) {
1042 switch (Cache.size() - NumSortedEntries) {
1044 // done, no new entries.
1047 // Two new entries, insert the last one into place.
1048 NonLocalDepEntry Val = Cache.back();
1050 MemoryDependenceResults::NonLocalDepInfo::iterator Entry =
1051 std::upper_bound(Cache.begin(), Cache.end() - 1, Val);
1052 Cache.insert(Entry, Val);
1056 // One new entry, Just insert the new value at the appropriate position.
1057 if (Cache.size() != 1) {
1058 NonLocalDepEntry Val = Cache.back();
1060 MemoryDependenceResults::NonLocalDepInfo::iterator Entry =
1061 std::upper_bound(Cache.begin(), Cache.end(), Val);
1062 Cache.insert(Entry, Val);
1066 // Added many values, do a full scale sort.
1067 std::sort(Cache.begin(), Cache.end());
1072 /// Perform a dependency query based on pointer/pointeesize starting at the end
1075 /// Add any clobber/def results to the results vector and keep track of which
1076 /// blocks are visited in 'Visited'.
1078 /// This has special behavior for the first block queries (when SkipFirstBlock
1079 /// is true). In this special case, it ignores the contents of the specified
1080 /// block and starts returning dependence info for its predecessors.
1082 /// This function returns true on success, or false to indicate that it could
1083 /// not compute dependence information for some reason. This should be treated
1084 /// as a clobber dependence on the first instruction in the predecessor block.
1085 bool MemoryDependenceResults::getNonLocalPointerDepFromBB(
1086 Instruction *QueryInst, const PHITransAddr &Pointer,
1087 const MemoryLocation &Loc, bool isLoad, BasicBlock *StartBB,
1088 SmallVectorImpl<NonLocalDepResult> &Result,
1089 DenseMap<BasicBlock *, Value *> &Visited, bool SkipFirstBlock) {
1090 // Look up the cached info for Pointer.
1091 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
1093 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
1094 // CacheKey, this value will be inserted as the associated value. Otherwise,
1095 // it'll be ignored, and we'll have to check to see if the cached size and
1096 // aa tags are consistent with the current query.
1097 NonLocalPointerInfo InitialNLPI;
1098 InitialNLPI.Size = Loc.Size;
1099 InitialNLPI.AATags = Loc.AATags;
1101 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
1102 // already have one.
1103 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
1104 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
1105 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
1107 // If we already have a cache entry for this CacheKey, we may need to do some
1108 // work to reconcile the cache entry and the current query.
1110 if (CacheInfo->Size < Loc.Size) {
1111 // The query's Size is greater than the cached one. Throw out the
1112 // cached data and proceed with the query at the greater size.
1113 CacheInfo->Pair = BBSkipFirstBlockPair();
1114 CacheInfo->Size = Loc.Size;
1115 for (auto &Entry : CacheInfo->NonLocalDeps)
1116 if (Instruction *Inst = Entry.getResult().getInst())
1117 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1118 CacheInfo->NonLocalDeps.clear();
1119 } else if (CacheInfo->Size > Loc.Size) {
1120 // This query's Size is less than the cached one. Conservatively restart
1121 // the query using the greater size.
1122 return getNonLocalPointerDepFromBB(
1123 QueryInst, Pointer, Loc.getWithNewSize(CacheInfo->Size), isLoad,
1124 StartBB, Result, Visited, SkipFirstBlock);
1127 // If the query's AATags are inconsistent with the cached one,
1128 // conservatively throw out the cached data and restart the query with
1129 // no tag if needed.
1130 if (CacheInfo->AATags != Loc.AATags) {
1131 if (CacheInfo->AATags) {
1132 CacheInfo->Pair = BBSkipFirstBlockPair();
1133 CacheInfo->AATags = AAMDNodes();
1134 for (auto &Entry : CacheInfo->NonLocalDeps)
1135 if (Instruction *Inst = Entry.getResult().getInst())
1136 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1137 CacheInfo->NonLocalDeps.clear();
1140 return getNonLocalPointerDepFromBB(
1141 QueryInst, Pointer, Loc.getWithoutAATags(), isLoad, StartBB, Result,
1142 Visited, SkipFirstBlock);
1146 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1148 // If we have valid cached information for exactly the block we are
1149 // investigating, just return it with no recomputation.
1150 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1151 // We have a fully cached result for this query then we can just return the
1152 // cached results and populate the visited set. However, we have to verify
1153 // that we don't already have conflicting results for these blocks. Check
1154 // to ensure that if a block in the results set is in the visited set that
1155 // it was for the same pointer query.
1156 if (!Visited.empty()) {
1157 for (auto &Entry : *Cache) {
1158 DenseMap<BasicBlock *, Value *>::iterator VI =
1159 Visited.find(Entry.getBB());
1160 if (VI == Visited.end() || VI->second == Pointer.getAddr())
1163 // We have a pointer mismatch in a block. Just return false, saying
1164 // that something was clobbered in this result. We could also do a
1165 // non-fully cached query, but there is little point in doing this.
1170 Value *Addr = Pointer.getAddr();
1171 for (auto &Entry : *Cache) {
1172 Visited.insert(std::make_pair(Entry.getBB(), Addr));
1173 if (Entry.getResult().isNonLocal()) {
1177 if (DT.isReachableFromEntry(Entry.getBB())) {
1179 NonLocalDepResult(Entry.getBB(), Entry.getResult(), Addr));
1182 ++NumCacheCompleteNonLocalPtr;
1186 // Otherwise, either this is a new block, a block with an invalid cache
1187 // pointer or one that we're about to invalidate by putting more info into it
1188 // than its valid cache info. If empty, the result will be valid cache info,
1189 // otherwise it isn't.
1191 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1193 CacheInfo->Pair = BBSkipFirstBlockPair();
1195 SmallVector<BasicBlock *, 32> Worklist;
1196 Worklist.push_back(StartBB);
1198 // PredList used inside loop.
1199 SmallVector<std::pair<BasicBlock *, PHITransAddr>, 16> PredList;
1201 // Keep track of the entries that we know are sorted. Previously cached
1202 // entries will all be sorted. The entries we add we only sort on demand (we
1203 // don't insert every element into its sorted position). We know that we
1204 // won't get any reuse from currently inserted values, because we don't
1205 // revisit blocks after we insert info for them.
1206 unsigned NumSortedEntries = Cache->size();
1207 unsigned WorklistEntries = BlockNumberLimit;
1208 bool GotWorklistLimit = false;
1209 DEBUG(AssertSorted(*Cache));
1211 while (!Worklist.empty()) {
1212 BasicBlock *BB = Worklist.pop_back_val();
1214 // If we do process a large number of blocks it becomes very expensive and
1215 // likely it isn't worth worrying about
1216 if (Result.size() > NumResultsLimit) {
1218 // Sort it now (if needed) so that recursive invocations of
1219 // getNonLocalPointerDepFromBB and other routines that could reuse the
1220 // cache value will only see properly sorted cache arrays.
1221 if (Cache && NumSortedEntries != Cache->size()) {
1222 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1224 // Since we bail out, the "Cache" set won't contain all of the
1225 // results for the query. This is ok (we can still use it to accelerate
1226 // specific block queries) but we can't do the fastpath "return all
1227 // results from the set". Clear out the indicator for this.
1228 CacheInfo->Pair = BBSkipFirstBlockPair();
1232 // Skip the first block if we have it.
1233 if (!SkipFirstBlock) {
1234 // Analyze the dependency of *Pointer in FromBB. See if we already have
1236 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1238 // Get the dependency info for Pointer in BB. If we have cached
1239 // information, we will use it, otherwise we compute it.
1240 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1241 MemDepResult Dep = GetNonLocalInfoForBlock(QueryInst, Loc, isLoad, BB,
1242 Cache, NumSortedEntries);
1244 // If we got a Def or Clobber, add this to the list of results.
1245 if (!Dep.isNonLocal()) {
1246 if (DT.isReachableFromEntry(BB)) {
1247 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1253 // If 'Pointer' is an instruction defined in this block, then we need to do
1254 // phi translation to change it into a value live in the predecessor block.
1255 // If not, we just add the predecessors to the worklist and scan them with
1256 // the same Pointer.
1257 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1258 SkipFirstBlock = false;
1259 SmallVector<BasicBlock *, 16> NewBlocks;
1260 for (BasicBlock *Pred : PredCache.get(BB)) {
1261 // Verify that we haven't looked at this block yet.
1262 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> InsertRes =
1263 Visited.insert(std::make_pair(Pred, Pointer.getAddr()));
1264 if (InsertRes.second) {
1265 // First time we've looked at *PI.
1266 NewBlocks.push_back(Pred);
1270 // If we have seen this block before, but it was with a different
1271 // pointer then we have a phi translation failure and we have to treat
1272 // this as a clobber.
1273 if (InsertRes.first->second != Pointer.getAddr()) {
1274 // Make sure to clean up the Visited map before continuing on to
1275 // PredTranslationFailure.
1276 for (unsigned i = 0; i < NewBlocks.size(); i++)
1277 Visited.erase(NewBlocks[i]);
1278 goto PredTranslationFailure;
1281 if (NewBlocks.size() > WorklistEntries) {
1282 // Make sure to clean up the Visited map before continuing on to
1283 // PredTranslationFailure.
1284 for (unsigned i = 0; i < NewBlocks.size(); i++)
1285 Visited.erase(NewBlocks[i]);
1286 GotWorklistLimit = true;
1287 goto PredTranslationFailure;
1289 WorklistEntries -= NewBlocks.size();
1290 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1294 // We do need to do phi translation, if we know ahead of time we can't phi
1295 // translate this value, don't even try.
1296 if (!Pointer.IsPotentiallyPHITranslatable())
1297 goto PredTranslationFailure;
1299 // We may have added values to the cache list before this PHI translation.
1300 // If so, we haven't done anything to ensure that the cache remains sorted.
1301 // Sort it now (if needed) so that recursive invocations of
1302 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1303 // value will only see properly sorted cache arrays.
1304 if (Cache && NumSortedEntries != Cache->size()) {
1305 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1306 NumSortedEntries = Cache->size();
1311 for (BasicBlock *Pred : PredCache.get(BB)) {
1312 PredList.push_back(std::make_pair(Pred, Pointer));
1314 // Get the PHI translated pointer in this predecessor. This can fail if
1315 // not translatable, in which case the getAddr() returns null.
1316 PHITransAddr &PredPointer = PredList.back().second;
1317 PredPointer.PHITranslateValue(BB, Pred, &DT, /*MustDominate=*/false);
1318 Value *PredPtrVal = PredPointer.getAddr();
1320 // Check to see if we have already visited this pred block with another
1321 // pointer. If so, we can't do this lookup. This failure can occur
1322 // with PHI translation when a critical edge exists and the PHI node in
1323 // the successor translates to a pointer value different than the
1324 // pointer the block was first analyzed with.
1325 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> InsertRes =
1326 Visited.insert(std::make_pair(Pred, PredPtrVal));
1328 if (!InsertRes.second) {
1329 // We found the pred; take it off the list of preds to visit.
1330 PredList.pop_back();
1332 // If the predecessor was visited with PredPtr, then we already did
1333 // the analysis and can ignore it.
1334 if (InsertRes.first->second == PredPtrVal)
1337 // Otherwise, the block was previously analyzed with a different
1338 // pointer. We can't represent the result of this case, so we just
1339 // treat this as a phi translation failure.
1341 // Make sure to clean up the Visited map before continuing on to
1342 // PredTranslationFailure.
1343 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1344 Visited.erase(PredList[i].first);
1346 goto PredTranslationFailure;
1350 // Actually process results here; this need to be a separate loop to avoid
1351 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1352 // any results for. (getNonLocalPointerDepFromBB will modify our
1353 // datastructures in ways the code after the PredTranslationFailure label
1355 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1356 BasicBlock *Pred = PredList[i].first;
1357 PHITransAddr &PredPointer = PredList[i].second;
1358 Value *PredPtrVal = PredPointer.getAddr();
1360 bool CanTranslate = true;
1361 // If PHI translation was unable to find an available pointer in this
1362 // predecessor, then we have to assume that the pointer is clobbered in
1363 // that predecessor. We can still do PRE of the load, which would insert
1364 // a computation of the pointer in this predecessor.
1366 CanTranslate = false;
1368 // FIXME: it is entirely possible that PHI translating will end up with
1369 // the same value. Consider PHI translating something like:
1370 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1371 // to recurse here, pedantically speaking.
1373 // If getNonLocalPointerDepFromBB fails here, that means the cached
1374 // result conflicted with the Visited list; we have to conservatively
1375 // assume it is unknown, but this also does not block PRE of the load.
1376 if (!CanTranslate ||
1377 !getNonLocalPointerDepFromBB(QueryInst, PredPointer,
1378 Loc.getWithNewPtr(PredPtrVal), isLoad,
1379 Pred, Result, Visited)) {
1380 // Add the entry to the Result list.
1381 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1382 Result.push_back(Entry);
1384 // Since we had a phi translation failure, the cache for CacheKey won't
1385 // include all of the entries that we need to immediately satisfy future
1386 // queries. Mark this in NonLocalPointerDeps by setting the
1387 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1388 // cached value to do more work but not miss the phi trans failure.
1389 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1390 NLPI.Pair = BBSkipFirstBlockPair();
1395 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1396 CacheInfo = &NonLocalPointerDeps[CacheKey];
1397 Cache = &CacheInfo->NonLocalDeps;
1398 NumSortedEntries = Cache->size();
1400 // Since we did phi translation, the "Cache" set won't contain all of the
1401 // results for the query. This is ok (we can still use it to accelerate
1402 // specific block queries) but we can't do the fastpath "return all
1403 // results from the set" Clear out the indicator for this.
1404 CacheInfo->Pair = BBSkipFirstBlockPair();
1405 SkipFirstBlock = false;
1408 PredTranslationFailure:
1409 // The following code is "failure"; we can't produce a sane translation
1410 // for the given block. It assumes that we haven't modified any of
1411 // our datastructures while processing the current block.
1414 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1415 CacheInfo = &NonLocalPointerDeps[CacheKey];
1416 Cache = &CacheInfo->NonLocalDeps;
1417 NumSortedEntries = Cache->size();
1420 // Since we failed phi translation, the "Cache" set won't contain all of the
1421 // results for the query. This is ok (we can still use it to accelerate
1422 // specific block queries) but we can't do the fastpath "return all
1423 // results from the set". Clear out the indicator for this.
1424 CacheInfo->Pair = BBSkipFirstBlockPair();
1426 // If *nothing* works, mark the pointer as unknown.
1428 // If this is the magic first block, return this as a clobber of the whole
1429 // incoming value. Since we can't phi translate to one of the predecessors,
1430 // we have to bail out.
1434 bool foundBlock = false;
1435 for (NonLocalDepEntry &I : llvm::reverse(*Cache)) {
1436 if (I.getBB() != BB)
1439 assert((GotWorklistLimit || I.getResult().isNonLocal() ||
1440 !DT.isReachableFromEntry(BB)) &&
1441 "Should only be here with transparent block");
1443 I.setResult(MemDepResult::getUnknown());
1445 NonLocalDepResult(I.getBB(), I.getResult(), Pointer.getAddr()));
1448 (void)foundBlock; (void)GotWorklistLimit;
1449 assert((foundBlock || GotWorklistLimit) && "Current block not in cache?");
1452 // Okay, we're done now. If we added new values to the cache, re-sort it.
1453 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1454 DEBUG(AssertSorted(*Cache));
1458 /// If P exists in CachedNonLocalPointerInfo, remove it.
1459 void MemoryDependenceResults::RemoveCachedNonLocalPointerDependencies(
1460 ValueIsLoadPair P) {
1461 CachedNonLocalPointerInfo::iterator It = NonLocalPointerDeps.find(P);
1462 if (It == NonLocalPointerDeps.end())
1465 // Remove all of the entries in the BB->val map. This involves removing
1466 // instructions from the reverse map.
1467 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1469 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1470 Instruction *Target = PInfo[i].getResult().getInst();
1472 continue; // Ignore non-local dep results.
1473 assert(Target->getParent() == PInfo[i].getBB());
1475 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1476 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1479 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1480 NonLocalPointerDeps.erase(It);
1483 void MemoryDependenceResults::invalidateCachedPointerInfo(Value *Ptr) {
1484 // If Ptr isn't really a pointer, just ignore it.
1485 if (!Ptr->getType()->isPointerTy())
1487 // Flush store info for the pointer.
1488 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1489 // Flush load info for the pointer.
1490 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1493 void MemoryDependenceResults::invalidateCachedPredecessors() {
1497 void MemoryDependenceResults::removeInstruction(Instruction *RemInst) {
1498 // Walk through the Non-local dependencies, removing this one as the value
1499 // for any cached queries.
1500 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1501 if (NLDI != NonLocalDeps.end()) {
1502 NonLocalDepInfo &BlockMap = NLDI->second.first;
1503 for (auto &Entry : BlockMap)
1504 if (Instruction *Inst = Entry.getResult().getInst())
1505 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1506 NonLocalDeps.erase(NLDI);
1509 // If we have a cached local dependence query for this instruction, remove it.
1511 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1512 if (LocalDepEntry != LocalDeps.end()) {
1513 // Remove us from DepInst's reverse set now that the local dep info is gone.
1514 if (Instruction *Inst = LocalDepEntry->second.getInst())
1515 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1517 // Remove this local dependency info.
1518 LocalDeps.erase(LocalDepEntry);
1521 // If we have any cached pointer dependencies on this instruction, remove
1522 // them. If the instruction has non-pointer type, then it can't be a pointer
1525 // Remove it from both the load info and the store info. The instruction
1526 // can't be in either of these maps if it is non-pointer.
1527 if (RemInst->getType()->isPointerTy()) {
1528 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1529 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1532 // Loop over all of the things that depend on the instruction we're removing.
1534 SmallVector<std::pair<Instruction *, Instruction *>, 8> ReverseDepsToAdd;
1536 // If we find RemInst as a clobber or Def in any of the maps for other values,
1537 // we need to replace its entry with a dirty version of the instruction after
1538 // it. If RemInst is a terminator, we use a null dirty value.
1540 // Using a dirty version of the instruction after RemInst saves having to scan
1541 // the entire block to get to this point.
1542 MemDepResult NewDirtyVal;
1543 if (!RemInst->isTerminator())
1544 NewDirtyVal = MemDepResult::getDirty(&*++RemInst->getIterator());
1546 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1547 if (ReverseDepIt != ReverseLocalDeps.end()) {
1548 // RemInst can't be the terminator if it has local stuff depending on it.
1549 assert(!ReverseDepIt->second.empty() && !isa<TerminatorInst>(RemInst) &&
1550 "Nothing can locally depend on a terminator");
1552 for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) {
1553 assert(InstDependingOnRemInst != RemInst &&
1554 "Already removed our local dep info");
1556 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1558 // Make sure to remember that new things depend on NewDepInst.
1559 assert(NewDirtyVal.getInst() &&
1560 "There is no way something else can have "
1561 "a local dep on this if it is a terminator!");
1562 ReverseDepsToAdd.push_back(
1563 std::make_pair(NewDirtyVal.getInst(), InstDependingOnRemInst));
1566 ReverseLocalDeps.erase(ReverseDepIt);
1568 // Add new reverse deps after scanning the set, to avoid invalidating the
1569 // 'ReverseDeps' reference.
1570 while (!ReverseDepsToAdd.empty()) {
1571 ReverseLocalDeps[ReverseDepsToAdd.back().first].insert(
1572 ReverseDepsToAdd.back().second);
1573 ReverseDepsToAdd.pop_back();
1577 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1578 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1579 for (Instruction *I : ReverseDepIt->second) {
1580 assert(I != RemInst && "Already removed NonLocalDep info for RemInst");
1582 PerInstNLInfo &INLD = NonLocalDeps[I];
1583 // The information is now dirty!
1586 for (auto &Entry : INLD.first) {
1587 if (Entry.getResult().getInst() != RemInst)
1590 // Convert to a dirty entry for the subsequent instruction.
1591 Entry.setResult(NewDirtyVal);
1593 if (Instruction *NextI = NewDirtyVal.getInst())
1594 ReverseDepsToAdd.push_back(std::make_pair(NextI, I));
1598 ReverseNonLocalDeps.erase(ReverseDepIt);
1600 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1601 while (!ReverseDepsToAdd.empty()) {
1602 ReverseNonLocalDeps[ReverseDepsToAdd.back().first].insert(
1603 ReverseDepsToAdd.back().second);
1604 ReverseDepsToAdd.pop_back();
1608 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1609 // value in the NonLocalPointerDeps info.
1610 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1611 ReverseNonLocalPtrDeps.find(RemInst);
1612 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1613 SmallVector<std::pair<Instruction *, ValueIsLoadPair>, 8>
1614 ReversePtrDepsToAdd;
1616 for (ValueIsLoadPair P : ReversePtrDepIt->second) {
1617 assert(P.getPointer() != RemInst &&
1618 "Already removed NonLocalPointerDeps info for RemInst");
1620 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1622 // The cache is not valid for any specific block anymore.
1623 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1625 // Update any entries for RemInst to use the instruction after it.
1626 for (auto &Entry : NLPDI) {
1627 if (Entry.getResult().getInst() != RemInst)
1630 // Convert to a dirty entry for the subsequent instruction.
1631 Entry.setResult(NewDirtyVal);
1633 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1634 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1637 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1638 // subsequent value may invalidate the sortedness.
1639 std::sort(NLPDI.begin(), NLPDI.end());
1642 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1644 while (!ReversePtrDepsToAdd.empty()) {
1645 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first].insert(
1646 ReversePtrDepsToAdd.back().second);
1647 ReversePtrDepsToAdd.pop_back();
1651 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1652 DEBUG(verifyRemoved(RemInst));
1655 /// Verify that the specified instruction does not occur in our internal data
1658 /// This function verifies by asserting in debug builds.
1659 void MemoryDependenceResults::verifyRemoved(Instruction *D) const {
1661 for (const auto &DepKV : LocalDeps) {
1662 assert(DepKV.first != D && "Inst occurs in data structures");
1663 assert(DepKV.second.getInst() != D && "Inst occurs in data structures");
1666 for (const auto &DepKV : NonLocalPointerDeps) {
1667 assert(DepKV.first.getPointer() != D && "Inst occurs in NLPD map key");
1668 for (const auto &Entry : DepKV.second.NonLocalDeps)
1669 assert(Entry.getResult().getInst() != D && "Inst occurs as NLPD value");
1672 for (const auto &DepKV : NonLocalDeps) {
1673 assert(DepKV.first != D && "Inst occurs in data structures");
1674 const PerInstNLInfo &INLD = DepKV.second;
1675 for (const auto &Entry : INLD.first)
1676 assert(Entry.getResult().getInst() != D &&
1677 "Inst occurs in data structures");
1680 for (const auto &DepKV : ReverseLocalDeps) {
1681 assert(DepKV.first != D && "Inst occurs in data structures");
1682 for (Instruction *Inst : DepKV.second)
1683 assert(Inst != D && "Inst occurs in data structures");
1686 for (const auto &DepKV : ReverseNonLocalDeps) {
1687 assert(DepKV.first != D && "Inst occurs in data structures");
1688 for (Instruction *Inst : DepKV.second)
1689 assert(Inst != D && "Inst occurs in data structures");
1692 for (const auto &DepKV : ReverseNonLocalPtrDeps) {
1693 assert(DepKV.first != D && "Inst occurs in rev NLPD map");
1695 for (ValueIsLoadPair P : DepKV.second)
1696 assert(P != ValueIsLoadPair(D, false) && P != ValueIsLoadPair(D, true) &&
1697 "Inst occurs in ReverseNonLocalPtrDeps map");
1702 AnalysisKey MemoryDependenceAnalysis::Key;
1704 MemoryDependenceResults
1705 MemoryDependenceAnalysis::run(Function &F, FunctionAnalysisManager &AM) {
1706 auto &AA = AM.getResult<AAManager>(F);
1707 auto &AC = AM.getResult<AssumptionAnalysis>(F);
1708 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1709 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1710 return MemoryDependenceResults(AA, AC, TLI, DT);
1713 char MemoryDependenceWrapperPass::ID = 0;
1715 INITIALIZE_PASS_BEGIN(MemoryDependenceWrapperPass, "memdep",
1716 "Memory Dependence Analysis", false, true)
1717 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1718 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1719 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1720 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1721 INITIALIZE_PASS_END(MemoryDependenceWrapperPass, "memdep",
1722 "Memory Dependence Analysis", false, true)
1724 MemoryDependenceWrapperPass::MemoryDependenceWrapperPass() : FunctionPass(ID) {
1725 initializeMemoryDependenceWrapperPassPass(*PassRegistry::getPassRegistry());
1728 MemoryDependenceWrapperPass::~MemoryDependenceWrapperPass() {}
1730 void MemoryDependenceWrapperPass::releaseMemory() {
1734 void MemoryDependenceWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1735 AU.setPreservesAll();
1736 AU.addRequired<AssumptionCacheTracker>();
1737 AU.addRequired<DominatorTreeWrapperPass>();
1738 AU.addRequiredTransitive<AAResultsWrapperPass>();
1739 AU.addRequiredTransitive<TargetLibraryInfoWrapperPass>();
1742 bool MemoryDependenceResults::invalidate(Function &F, const PreservedAnalyses &PA,
1743 FunctionAnalysisManager::Invalidator &Inv) {
1744 // Check whether our analysis is preserved.
1745 auto PAC = PA.getChecker<MemoryDependenceAnalysis>();
1746 if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Function>>())
1747 // If not, give up now.
1750 // Check whether the analyses we depend on became invalid for any reason.
1751 if (Inv.invalidate<AAManager>(F, PA) ||
1752 Inv.invalidate<AssumptionAnalysis>(F, PA) ||
1753 Inv.invalidate<DominatorTreeAnalysis>(F, PA))
1756 // Otherwise this analysis result remains valid.
1760 unsigned MemoryDependenceResults::getDefaultBlockScanLimit() const {
1761 return BlockScanLimit;
1764 bool MemoryDependenceWrapperPass::runOnFunction(Function &F) {
1765 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
1766 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1767 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1768 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1769 MemDep.emplace(AA, AC, TLI, DT);