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/DenseMap.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/AssumptionCache.h"
25 #include "llvm/Analysis/MemoryBuiltins.h"
26 #include "llvm/Analysis/MemoryLocation.h"
27 #include "llvm/Analysis/OrderedBasicBlock.h"
28 #include "llvm/Analysis/PHITransAddr.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/Attributes.h"
32 #include "llvm/IR/BasicBlock.h"
33 #include "llvm/IR/CallSite.h"
34 #include "llvm/IR/Constants.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/DerivedTypes.h"
37 #include "llvm/IR/Dominators.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/InstrTypes.h"
40 #include "llvm/IR/Instruction.h"
41 #include "llvm/IR/Instructions.h"
42 #include "llvm/IR/IntrinsicInst.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/Metadata.h"
45 #include "llvm/IR/Module.h"
46 #include "llvm/IR/PredIteratorCache.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/IR/Use.h"
49 #include "llvm/IR/User.h"
50 #include "llvm/IR/Value.h"
51 #include "llvm/Pass.h"
52 #include "llvm/Support/AtomicOrdering.h"
53 #include "llvm/Support/Casting.h"
54 #include "llvm/Support/CommandLine.h"
55 #include "llvm/Support/Compiler.h"
56 #include "llvm/Support/Debug.h"
57 #include "llvm/Support/MathExtras.h"
66 #define DEBUG_TYPE "memdep"
68 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
69 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
70 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
72 STATISTIC(NumCacheNonLocalPtr,
73 "Number of fully cached non-local ptr responses");
74 STATISTIC(NumCacheDirtyNonLocalPtr,
75 "Number of cached, but dirty, non-local ptr responses");
76 STATISTIC(NumUncacheNonLocalPtr, "Number of uncached non-local ptr responses");
77 STATISTIC(NumCacheCompleteNonLocalPtr,
78 "Number of block queries that were completely cached");
80 // Limit for the number of instructions to scan in a block.
82 static cl::opt<unsigned> BlockScanLimit(
83 "memdep-block-scan-limit", cl::Hidden, cl::init(100),
84 cl::desc("The number of instructions to scan in a block in memory "
85 "dependency analysis (default = 100)"));
87 static cl::opt<unsigned>
88 BlockNumberLimit("memdep-block-number-limit", cl::Hidden, cl::init(1000),
89 cl::desc("The number of blocks to scan during memory "
90 "dependency analysis (default = 1000)"));
92 // Limit on the number of memdep results to process.
93 static const unsigned int NumResultsLimit = 100;
95 /// This is a helper function that removes Val from 'Inst's set in ReverseMap.
97 /// If the set becomes empty, remove Inst's entry.
98 template <typename KeyTy>
100 RemoveFromReverseMap(DenseMap<Instruction *, SmallPtrSet<KeyTy, 4>> &ReverseMap,
101 Instruction *Inst, KeyTy Val) {
102 typename DenseMap<Instruction *, SmallPtrSet<KeyTy, 4>>::iterator InstIt =
103 ReverseMap.find(Inst);
104 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
105 bool Found = InstIt->second.erase(Val);
106 assert(Found && "Invalid reverse map!");
108 if (InstIt->second.empty())
109 ReverseMap.erase(InstIt);
112 /// If the given instruction references a specific memory location, fill in Loc
113 /// with the details, otherwise set Loc.Ptr to null.
115 /// Returns a ModRefInfo value describing the general behavior of the
117 static ModRefInfo GetLocation(const Instruction *Inst, MemoryLocation &Loc,
118 const TargetLibraryInfo &TLI) {
119 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
120 if (LI->isUnordered()) {
121 Loc = MemoryLocation::get(LI);
122 return ModRefInfo::Ref;
124 if (LI->getOrdering() == AtomicOrdering::Monotonic) {
125 Loc = MemoryLocation::get(LI);
126 return ModRefInfo::ModRef;
128 Loc = MemoryLocation();
129 return ModRefInfo::ModRef;
132 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
133 if (SI->isUnordered()) {
134 Loc = MemoryLocation::get(SI);
135 return ModRefInfo::Mod;
137 if (SI->getOrdering() == AtomicOrdering::Monotonic) {
138 Loc = MemoryLocation::get(SI);
139 return ModRefInfo::ModRef;
141 Loc = MemoryLocation();
142 return ModRefInfo::ModRef;
145 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
146 Loc = MemoryLocation::get(V);
147 return ModRefInfo::ModRef;
150 if (const CallInst *CI = isFreeCall(Inst, &TLI)) {
151 // calls to free() deallocate the entire structure
152 Loc = MemoryLocation(CI->getArgOperand(0));
153 return ModRefInfo::Mod;
156 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
159 switch (II->getIntrinsicID()) {
160 case Intrinsic::lifetime_start:
161 case Intrinsic::lifetime_end:
162 case Intrinsic::invariant_start:
163 II->getAAMetadata(AAInfo);
164 Loc = MemoryLocation(
165 II->getArgOperand(1),
166 cast<ConstantInt>(II->getArgOperand(0))->getZExtValue(), AAInfo);
167 // These intrinsics don't really modify the memory, but returning Mod
168 // will allow them to be handled conservatively.
169 return ModRefInfo::Mod;
170 case Intrinsic::invariant_end:
171 II->getAAMetadata(AAInfo);
172 Loc = MemoryLocation(
173 II->getArgOperand(2),
174 cast<ConstantInt>(II->getArgOperand(1))->getZExtValue(), AAInfo);
175 // These intrinsics don't really modify the memory, but returning Mod
176 // will allow them to be handled conservatively.
177 return ModRefInfo::Mod;
183 // Otherwise, just do the coarse-grained thing that always works.
184 if (Inst->mayWriteToMemory())
185 return ModRefInfo::ModRef;
186 if (Inst->mayReadFromMemory())
187 return ModRefInfo::Ref;
188 return ModRefInfo::NoModRef;
191 /// Private helper for finding the local dependencies of a call site.
192 MemDepResult MemoryDependenceResults::getCallSiteDependencyFrom(
193 CallSite CS, bool isReadOnlyCall, BasicBlock::iterator ScanIt,
195 unsigned Limit = BlockScanLimit;
197 // Walk backwards through the block, looking for dependencies.
198 while (ScanIt != BB->begin()) {
199 Instruction *Inst = &*--ScanIt;
200 // Debug intrinsics don't cause dependences and should not affect Limit
201 if (isa<DbgInfoIntrinsic>(Inst))
204 // Limit the amount of scanning we do so we don't end up with quadratic
205 // running time on extreme testcases.
208 return MemDepResult::getUnknown();
210 // If this inst is a memory op, get the pointer it accessed
212 ModRefInfo MR = GetLocation(Inst, Loc, TLI);
214 // A simple instruction.
215 if (isModOrRefSet(AA.getModRefInfo(CS, Loc)))
216 return MemDepResult::getClobber(Inst);
220 if (auto InstCS = CallSite(Inst)) {
221 // If these two calls do not interfere, look past it.
222 if (isNoModRef(AA.getModRefInfo(CS, InstCS))) {
223 // If the two calls are the same, return InstCS as a Def, so that
224 // CS can be found redundant and eliminated.
225 if (isReadOnlyCall && !isModSet(MR) &&
226 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
227 return MemDepResult::getDef(Inst);
229 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
233 return MemDepResult::getClobber(Inst);
236 // If we could not obtain a pointer for the instruction and the instruction
237 // touches memory then assume that this is a dependency.
238 if (isModOrRefSet(MR))
239 return MemDepResult::getClobber(Inst);
242 // No dependence found. If this is the entry block of the function, it is
243 // unknown, otherwise it is non-local.
244 if (BB != &BB->getParent()->getEntryBlock())
245 return MemDepResult::getNonLocal();
246 return MemDepResult::getNonFuncLocal();
249 unsigned MemoryDependenceResults::getLoadLoadClobberFullWidthSize(
250 const Value *MemLocBase, int64_t MemLocOffs, unsigned MemLocSize,
251 const LoadInst *LI) {
252 // We can only extend simple integer loads.
253 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple())
256 // Load widening is hostile to ThreadSanitizer: it may cause false positives
257 // or make the reports more cryptic (access sizes are wrong).
258 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
261 const DataLayout &DL = LI->getModule()->getDataLayout();
263 // Get the base of this load.
265 const Value *LIBase =
266 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, DL);
268 // If the two pointers are not based on the same pointer, we can't tell that
270 if (LIBase != MemLocBase)
273 // Okay, the two values are based on the same pointer, but returned as
274 // no-alias. This happens when we have things like two byte loads at "P+1"
275 // and "P+3". Check to see if increasing the size of the "LI" load up to its
276 // alignment (or the largest native integer type) will allow us to load all
277 // the bits required by MemLoc.
279 // If MemLoc is before LI, then no widening of LI will help us out.
280 if (MemLocOffs < LIOffs)
283 // Get the alignment of the load in bytes. We assume that it is safe to load
284 // any legal integer up to this size without a problem. For example, if we're
285 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
286 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
288 unsigned LoadAlign = LI->getAlignment();
290 int64_t MemLocEnd = MemLocOffs + MemLocSize;
292 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
293 if (LIOffs + LoadAlign < MemLocEnd)
296 // This is the size of the load to try. Start with the next larger power of
298 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits() / 8U;
299 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
302 // If this load size is bigger than our known alignment or would not fit
303 // into a native integer register, then we fail.
304 if (NewLoadByteSize > LoadAlign ||
305 !DL.fitsInLegalInteger(NewLoadByteSize * 8))
308 if (LIOffs + NewLoadByteSize > MemLocEnd &&
309 (LI->getParent()->getParent()->hasFnAttribute(
310 Attribute::SanitizeAddress) ||
311 LI->getParent()->getParent()->hasFnAttribute(
312 Attribute::SanitizeHWAddress)))
313 // We will be reading past the location accessed by the original program.
314 // While this is safe in a regular build, Address Safety analysis tools
315 // may start reporting false warnings. So, don't do widening.
318 // If a load of this width would include all of MemLoc, then we succeed.
319 if (LIOffs + NewLoadByteSize >= MemLocEnd)
320 return NewLoadByteSize;
322 NewLoadByteSize <<= 1;
326 static bool isVolatile(Instruction *Inst) {
327 if (auto *LI = dyn_cast<LoadInst>(Inst))
328 return LI->isVolatile();
329 if (auto *SI = dyn_cast<StoreInst>(Inst))
330 return SI->isVolatile();
331 if (auto *AI = dyn_cast<AtomicCmpXchgInst>(Inst))
332 return AI->isVolatile();
336 MemDepResult MemoryDependenceResults::getPointerDependencyFrom(
337 const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
338 BasicBlock *BB, Instruction *QueryInst, unsigned *Limit) {
339 MemDepResult InvariantGroupDependency = MemDepResult::getUnknown();
340 if (QueryInst != nullptr) {
341 if (auto *LI = dyn_cast<LoadInst>(QueryInst)) {
342 InvariantGroupDependency = getInvariantGroupPointerDependency(LI, BB);
344 if (InvariantGroupDependency.isDef())
345 return InvariantGroupDependency;
348 MemDepResult SimpleDep = getSimplePointerDependencyFrom(
349 MemLoc, isLoad, ScanIt, BB, QueryInst, Limit);
350 if (SimpleDep.isDef())
352 // Non-local invariant group dependency indicates there is non local Def
353 // (it only returns nonLocal if it finds nonLocal def), which is better than
354 // local clobber and everything else.
355 if (InvariantGroupDependency.isNonLocal())
356 return InvariantGroupDependency;
358 assert(InvariantGroupDependency.isUnknown() &&
359 "InvariantGroupDependency should be only unknown at this point");
364 MemoryDependenceResults::getInvariantGroupPointerDependency(LoadInst *LI,
366 auto *InvariantGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group);
367 if (!InvariantGroupMD)
368 return MemDepResult::getUnknown();
370 // Take the ptr operand after all casts and geps 0. This way we can search
371 // cast graph down only.
372 Value *LoadOperand = LI->getPointerOperand()->stripPointerCasts();
374 // It's is not safe to walk the use list of global value, because function
375 // passes aren't allowed to look outside their functions.
376 // FIXME: this could be fixed by filtering instructions from outside
377 // of current function.
378 if (isa<GlobalValue>(LoadOperand))
379 return MemDepResult::getUnknown();
381 // Queue to process all pointers that are equivalent to load operand.
382 SmallVector<const Value *, 8> LoadOperandsQueue;
383 LoadOperandsQueue.push_back(LoadOperand);
385 Instruction *ClosestDependency = nullptr;
386 // Order of instructions in uses list is unpredictible. In order to always
387 // get the same result, we will look for the closest dominance.
388 auto GetClosestDependency = [this](Instruction *Best, Instruction *Other) {
389 assert(Other && "Must call it with not null instruction");
390 if (Best == nullptr || DT.dominates(Best, Other))
395 // FIXME: This loop is O(N^2) because dominates can be O(n) and in worst case
396 // we will see all the instructions. This should be fixed in MSSA.
397 while (!LoadOperandsQueue.empty()) {
398 const Value *Ptr = LoadOperandsQueue.pop_back_val();
399 assert(Ptr && !isa<GlobalValue>(Ptr) &&
400 "Null or GlobalValue should not be inserted");
402 for (const Use &Us : Ptr->uses()) {
403 auto *U = dyn_cast<Instruction>(Us.getUser());
404 if (!U || U == LI || !DT.dominates(U, LI))
407 // Bitcast or gep with zeros are using Ptr. Add to queue to check it's
408 // users. U = bitcast Ptr
409 if (isa<BitCastInst>(U)) {
410 LoadOperandsQueue.push_back(U);
413 // Gep with zeros is equivalent to bitcast.
414 // FIXME: we are not sure if some bitcast should be canonicalized to gep 0
415 // or gep 0 to bitcast because of SROA, so there are 2 forms. When
416 // typeless pointers will be ready then both cases will be gone
417 // (and this BFS also won't be needed).
418 if (auto *GEP = dyn_cast<GetElementPtrInst>(U))
419 if (GEP->hasAllZeroIndices()) {
420 LoadOperandsQueue.push_back(U);
424 // If we hit load/store with the same invariant.group metadata (and the
425 // same pointer operand) we can assume that value pointed by pointer
426 // operand didn't change.
427 if ((isa<LoadInst>(U) || isa<StoreInst>(U)) &&
428 U->getMetadata(LLVMContext::MD_invariant_group) == InvariantGroupMD)
429 ClosestDependency = GetClosestDependency(ClosestDependency, U);
433 if (!ClosestDependency)
434 return MemDepResult::getUnknown();
435 if (ClosestDependency->getParent() == BB)
436 return MemDepResult::getDef(ClosestDependency);
437 // Def(U) can't be returned here because it is non-local. If local
438 // dependency won't be found then return nonLocal counting that the
439 // user will call getNonLocalPointerDependency, which will return cached
441 NonLocalDefsCache.try_emplace(
442 LI, NonLocalDepResult(ClosestDependency->getParent(),
443 MemDepResult::getDef(ClosestDependency), nullptr));
444 return MemDepResult::getNonLocal();
447 MemDepResult MemoryDependenceResults::getSimplePointerDependencyFrom(
448 const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
449 BasicBlock *BB, Instruction *QueryInst, unsigned *Limit) {
450 bool isInvariantLoad = false;
453 unsigned DefaultLimit = BlockScanLimit;
454 return getSimplePointerDependencyFrom(MemLoc, isLoad, ScanIt, BB, QueryInst,
458 // We must be careful with atomic accesses, as they may allow another thread
459 // to touch this location, clobbering it. We are conservative: if the
460 // QueryInst is not a simple (non-atomic) memory access, we automatically
461 // return getClobber.
462 // If it is simple, we know based on the results of
463 // "Compiler testing via a theory of sound optimisations in the C11/C++11
464 // memory model" in PLDI 2013, that a non-atomic location can only be
465 // clobbered between a pair of a release and an acquire action, with no
466 // access to the location in between.
467 // Here is an example for giving the general intuition behind this rule.
468 // In the following code:
470 // release action; [1]
471 // acquire action; [4]
473 // It is unsafe to replace %val by 0 because another thread may be running:
474 // acquire action; [2]
476 // release action; [3]
477 // with synchronization from 1 to 2 and from 3 to 4, resulting in %val
478 // being 42. A key property of this program however is that if either
479 // 1 or 4 were missing, there would be a race between the store of 42
480 // either the store of 0 or the load (making the whole program racy).
481 // The paper mentioned above shows that the same property is respected
482 // by every program that can detect any optimization of that kind: either
483 // it is racy (undefined) or there is a release followed by an acquire
484 // between the pair of accesses under consideration.
486 // If the load is invariant, we "know" that it doesn't alias *any* write. We
487 // do want to respect mustalias results since defs are useful for value
488 // forwarding, but any mayalias write can be assumed to be noalias.
489 // Arguably, this logic should be pushed inside AliasAnalysis itself.
490 if (isLoad && QueryInst) {
491 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
492 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
493 isInvariantLoad = true;
496 const DataLayout &DL = BB->getModule()->getDataLayout();
498 // Create a numbered basic block to lazily compute and cache instruction
499 // positions inside a BB. This is used to provide fast queries for relative
500 // position between two instructions in a BB and can be used by
501 // AliasAnalysis::callCapturesBefore.
502 OrderedBasicBlock OBB(BB);
504 // Return "true" if and only if the instruction I is either a non-simple
505 // load or a non-simple store.
506 auto isNonSimpleLoadOrStore = [](Instruction *I) -> bool {
507 if (auto *LI = dyn_cast<LoadInst>(I))
508 return !LI->isSimple();
509 if (auto *SI = dyn_cast<StoreInst>(I))
510 return !SI->isSimple();
514 // Return "true" if I is not a load and not a store, but it does access
516 auto isOtherMemAccess = [](Instruction *I) -> bool {
517 return !isa<LoadInst>(I) && !isa<StoreInst>(I) && I->mayReadOrWriteMemory();
520 // Walk backwards through the basic block, looking for dependencies.
521 while (ScanIt != BB->begin()) {
522 Instruction *Inst = &*--ScanIt;
524 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
525 // Debug intrinsics don't (and can't) cause dependencies.
526 if (isa<DbgInfoIntrinsic>(II))
529 // Limit the amount of scanning we do so we don't end up with quadratic
530 // running time on extreme testcases.
533 return MemDepResult::getUnknown();
535 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
536 // If we reach a lifetime begin or end marker, then the query ends here
537 // because the value is undefined.
538 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
539 // FIXME: This only considers queries directly on the invariant-tagged
540 // pointer, not on query pointers that are indexed off of them. It'd
541 // be nice to handle that at some point (the right approach is to use
542 // GetPointerBaseWithConstantOffset).
543 if (AA.isMustAlias(MemoryLocation(II->getArgOperand(1)), MemLoc))
544 return MemDepResult::getDef(II);
549 // Values depend on loads if the pointers are must aliased. This means
550 // that a load depends on another must aliased load from the same value.
551 // One exception is atomic loads: a value can depend on an atomic load that
552 // it does not alias with when this atomic load indicates that another
553 // thread may be accessing the location.
554 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
555 // While volatile access cannot be eliminated, they do not have to clobber
556 // non-aliasing locations, as normal accesses, for example, can be safely
557 // reordered with volatile accesses.
558 if (LI->isVolatile()) {
560 // Original QueryInst *may* be volatile
561 return MemDepResult::getClobber(LI);
562 if (isVolatile(QueryInst))
563 // Ordering required if QueryInst is itself volatile
564 return MemDepResult::getClobber(LI);
565 // Otherwise, volatile doesn't imply any special ordering
568 // Atomic loads have complications involved.
569 // A Monotonic (or higher) load is OK if the query inst is itself not
571 // FIXME: This is overly conservative.
572 if (LI->isAtomic() && isStrongerThanUnordered(LI->getOrdering())) {
573 if (!QueryInst || isNonSimpleLoadOrStore(QueryInst) ||
574 isOtherMemAccess(QueryInst))
575 return MemDepResult::getClobber(LI);
576 if (LI->getOrdering() != AtomicOrdering::Monotonic)
577 return MemDepResult::getClobber(LI);
580 MemoryLocation LoadLoc = MemoryLocation::get(LI);
582 // If we found a pointer, check if it could be the same as our pointer.
583 AliasResult R = AA.alias(LoadLoc, MemLoc);
589 // Must aliased loads are defs of each other.
591 return MemDepResult::getDef(Inst);
593 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
594 // in terms of clobbering loads, but since it does this by looking
595 // at the clobbering load directly, it doesn't know about any
596 // phi translation that may have happened along the way.
598 // If we have a partial alias, then return this as a clobber for the
600 if (R == PartialAlias)
601 return MemDepResult::getClobber(Inst);
604 // Random may-alias loads don't depend on each other without a
609 // Stores don't depend on other no-aliased accesses.
613 // Stores don't alias loads from read-only memory.
614 if (AA.pointsToConstantMemory(LoadLoc))
617 // Stores depend on may/must aliased loads.
618 return MemDepResult::getDef(Inst);
621 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
622 // Atomic stores have complications involved.
623 // A Monotonic store is OK if the query inst is itself not atomic.
624 // FIXME: This is overly conservative.
625 if (!SI->isUnordered() && SI->isAtomic()) {
626 if (!QueryInst || isNonSimpleLoadOrStore(QueryInst) ||
627 isOtherMemAccess(QueryInst))
628 return MemDepResult::getClobber(SI);
629 if (SI->getOrdering() != AtomicOrdering::Monotonic)
630 return MemDepResult::getClobber(SI);
633 // FIXME: this is overly conservative.
634 // While volatile access cannot be eliminated, they do not have to clobber
635 // non-aliasing locations, as normal accesses can for example be reordered
636 // with volatile accesses.
637 if (SI->isVolatile())
638 if (!QueryInst || isNonSimpleLoadOrStore(QueryInst) ||
639 isOtherMemAccess(QueryInst))
640 return MemDepResult::getClobber(SI);
642 // If alias analysis can tell that this store is guaranteed to not modify
643 // the query pointer, ignore it. Use getModRefInfo to handle cases where
644 // the query pointer points to constant memory etc.
645 if (!isModOrRefSet(AA.getModRefInfo(SI, MemLoc)))
648 // Ok, this store might clobber the query pointer. Check to see if it is
649 // a must alias: in this case, we want to return this as a def.
650 // FIXME: Use ModRefInfo::Must bit from getModRefInfo call above.
651 MemoryLocation StoreLoc = MemoryLocation::get(SI);
653 // If we found a pointer, check if it could be the same as our pointer.
654 AliasResult R = AA.alias(StoreLoc, MemLoc);
659 return MemDepResult::getDef(Inst);
662 return MemDepResult::getClobber(Inst);
665 // If this is an allocation, and if we know that the accessed pointer is to
666 // the allocation, return Def. This means that there is no dependence and
667 // the access can be optimized based on that. For example, a load could
668 // turn into undef. Note that we can bypass the allocation itself when
669 // looking for a clobber in many cases; that's an alias property and is
670 // handled by BasicAA.
671 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, &TLI)) {
672 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
673 if (AccessPtr == Inst || AA.isMustAlias(Inst, AccessPtr))
674 return MemDepResult::getDef(Inst);
680 // A release fence requires that all stores complete before it, but does
681 // not prevent the reordering of following loads or stores 'before' the
682 // fence. As a result, we look past it when finding a dependency for
683 // loads. DSE uses this to find preceeding stores to delete and thus we
684 // can't bypass the fence if the query instruction is a store.
685 if (FenceInst *FI = dyn_cast<FenceInst>(Inst))
686 if (isLoad && FI->getOrdering() == AtomicOrdering::Release)
689 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
690 ModRefInfo MR = AA.getModRefInfo(Inst, MemLoc);
691 // If necessary, perform additional analysis.
692 if (isModAndRefSet(MR))
693 MR = AA.callCapturesBefore(Inst, MemLoc, &DT, &OBB);
694 switch (clearMust(MR)) {
695 case ModRefInfo::NoModRef:
696 // If the call has no effect on the queried pointer, just ignore it.
698 case ModRefInfo::Mod:
699 return MemDepResult::getClobber(Inst);
700 case ModRefInfo::Ref:
701 // If the call is known to never store to the pointer, and if this is a
702 // load query, we can safely ignore it (scan past it).
707 // Otherwise, there is a potential dependence. Return a clobber.
708 return MemDepResult::getClobber(Inst);
712 // No dependence found. If this is the entry block of the function, it is
713 // unknown, otherwise it is non-local.
714 if (BB != &BB->getParent()->getEntryBlock())
715 return MemDepResult::getNonLocal();
716 return MemDepResult::getNonFuncLocal();
719 MemDepResult MemoryDependenceResults::getDependency(Instruction *QueryInst) {
720 Instruction *ScanPos = QueryInst;
722 // Check for a cached result
723 MemDepResult &LocalCache = LocalDeps[QueryInst];
725 // If the cached entry is non-dirty, just return it. Note that this depends
726 // on MemDepResult's default constructing to 'dirty'.
727 if (!LocalCache.isDirty())
730 // Otherwise, if we have a dirty entry, we know we can start the scan at that
731 // instruction, which may save us some work.
732 if (Instruction *Inst = LocalCache.getInst()) {
735 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
738 BasicBlock *QueryParent = QueryInst->getParent();
741 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
742 // No dependence found. If this is the entry block of the function, it is
743 // unknown, otherwise it is non-local.
744 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
745 LocalCache = MemDepResult::getNonLocal();
747 LocalCache = MemDepResult::getNonFuncLocal();
749 MemoryLocation MemLoc;
750 ModRefInfo MR = GetLocation(QueryInst, MemLoc, TLI);
752 // If we can do a pointer scan, make it happen.
753 bool isLoad = !isModSet(MR);
754 if (auto *II = dyn_cast<IntrinsicInst>(QueryInst))
755 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
757 LocalCache = getPointerDependencyFrom(
758 MemLoc, isLoad, ScanPos->getIterator(), QueryParent, QueryInst);
759 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
760 CallSite QueryCS(QueryInst);
761 bool isReadOnly = AA.onlyReadsMemory(QueryCS);
762 LocalCache = getCallSiteDependencyFrom(
763 QueryCS, isReadOnly, ScanPos->getIterator(), QueryParent);
765 // Non-memory instruction.
766 LocalCache = MemDepResult::getUnknown();
769 // Remember the result!
770 if (Instruction *I = LocalCache.getInst())
771 ReverseLocalDeps[I].insert(QueryInst);
777 /// This method is used when -debug is specified to verify that cache arrays
778 /// are properly kept sorted.
779 static void AssertSorted(MemoryDependenceResults::NonLocalDepInfo &Cache,
782 Count = Cache.size();
783 assert(std::is_sorted(Cache.begin(), Cache.begin() + Count) &&
784 "Cache isn't sorted!");
788 const MemoryDependenceResults::NonLocalDepInfo &
789 MemoryDependenceResults::getNonLocalCallDependency(CallSite QueryCS) {
790 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
791 "getNonLocalCallDependency should only be used on calls with "
793 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
794 NonLocalDepInfo &Cache = CacheP.first;
796 // This is the set of blocks that need to be recomputed. In the cached case,
797 // this can happen due to instructions being deleted etc. In the uncached
798 // case, this starts out as the set of predecessors we care about.
799 SmallVector<BasicBlock *, 32> DirtyBlocks;
801 if (!Cache.empty()) {
802 // Okay, we have a cache entry. If we know it is not dirty, just return it
803 // with no computation.
804 if (!CacheP.second) {
809 // If we already have a partially computed set of results, scan them to
810 // determine what is dirty, seeding our initial DirtyBlocks worklist.
811 for (auto &Entry : Cache)
812 if (Entry.getResult().isDirty())
813 DirtyBlocks.push_back(Entry.getBB());
815 // Sort the cache so that we can do fast binary search lookups below.
816 std::sort(Cache.begin(), Cache.end());
818 ++NumCacheDirtyNonLocal;
819 // cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
820 // << Cache.size() << " cached: " << *QueryInst;
822 // Seed DirtyBlocks with each of the preds of QueryInst's block.
823 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
824 for (BasicBlock *Pred : PredCache.get(QueryBB))
825 DirtyBlocks.push_back(Pred);
826 ++NumUncacheNonLocal;
829 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
830 bool isReadonlyCall = AA.onlyReadsMemory(QueryCS);
832 SmallPtrSet<BasicBlock *, 32> Visited;
834 unsigned NumSortedEntries = Cache.size();
835 DEBUG(AssertSorted(Cache));
837 // Iterate while we still have blocks to update.
838 while (!DirtyBlocks.empty()) {
839 BasicBlock *DirtyBB = DirtyBlocks.back();
840 DirtyBlocks.pop_back();
842 // Already processed this block?
843 if (!Visited.insert(DirtyBB).second)
846 // Do a binary search to see if we already have an entry for this block in
847 // the cache set. If so, find it.
848 DEBUG(AssertSorted(Cache, NumSortedEntries));
849 NonLocalDepInfo::iterator Entry =
850 std::upper_bound(Cache.begin(), Cache.begin() + NumSortedEntries,
851 NonLocalDepEntry(DirtyBB));
852 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
855 NonLocalDepEntry *ExistingResult = nullptr;
856 if (Entry != Cache.begin() + NumSortedEntries &&
857 Entry->getBB() == DirtyBB) {
858 // If we already have an entry, and if it isn't already dirty, the block
860 if (!Entry->getResult().isDirty())
863 // Otherwise, remember this slot so we can update the value.
864 ExistingResult = &*Entry;
867 // If the dirty entry has a pointer, start scanning from it so we don't have
868 // to rescan the entire block.
869 BasicBlock::iterator ScanPos = DirtyBB->end();
870 if (ExistingResult) {
871 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
872 ScanPos = Inst->getIterator();
873 // We're removing QueryInst's use of Inst.
874 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
875 QueryCS.getInstruction());
879 // Find out if this block has a local dependency for QueryInst.
882 if (ScanPos != DirtyBB->begin()) {
884 getCallSiteDependencyFrom(QueryCS, isReadonlyCall, ScanPos, DirtyBB);
885 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
886 // No dependence found. If this is the entry block of the function, it is
887 // a clobber, otherwise it is unknown.
888 Dep = MemDepResult::getNonLocal();
890 Dep = MemDepResult::getNonFuncLocal();
893 // If we had a dirty entry for the block, update it. Otherwise, just add
896 ExistingResult->setResult(Dep);
898 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
900 // If the block has a dependency (i.e. it isn't completely transparent to
901 // the value), remember the association!
902 if (!Dep.isNonLocal()) {
903 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
904 // update this when we remove instructions.
905 if (Instruction *Inst = Dep.getInst())
906 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
909 // If the block *is* completely transparent to the load, we need to check
910 // the predecessors of this block. Add them to our worklist.
911 for (BasicBlock *Pred : PredCache.get(DirtyBB))
912 DirtyBlocks.push_back(Pred);
919 void MemoryDependenceResults::getNonLocalPointerDependency(
920 Instruction *QueryInst, SmallVectorImpl<NonLocalDepResult> &Result) {
921 const MemoryLocation Loc = MemoryLocation::get(QueryInst);
922 bool isLoad = isa<LoadInst>(QueryInst);
923 return getNonLocalPointerDependencyFrom(QueryInst, Loc, isLoad, Result);
926 void MemoryDependenceResults::getNonLocalPointerDependencyFrom(
927 Instruction *QueryInst,
928 const MemoryLocation &Loc,
930 SmallVectorImpl<NonLocalDepResult> &Result) {
931 BasicBlock *FromBB = QueryInst->getParent();
934 assert(Loc.Ptr->getType()->isPointerTy() &&
935 "Can't get pointer deps of a non-pointer!");
938 // Check if there is cached Def with invariant.group. FIXME: cache might be
939 // invalid if cached instruction would be removed between call to
940 // getPointerDependencyFrom and this function.
941 auto NonLocalDefIt = NonLocalDefsCache.find(QueryInst);
942 if (NonLocalDefIt != NonLocalDefsCache.end()) {
943 Result.push_back(std::move(NonLocalDefIt->second));
944 NonLocalDefsCache.erase(NonLocalDefIt);
948 // This routine does not expect to deal with volatile instructions.
949 // Doing so would require piping through the QueryInst all the way through.
950 // TODO: volatiles can't be elided, but they can be reordered with other
951 // non-volatile accesses.
953 // We currently give up on any instruction which is ordered, but we do handle
954 // atomic instructions which are unordered.
955 // TODO: Handle ordered instructions
956 auto isOrdered = [](Instruction *Inst) {
957 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
958 return !LI->isUnordered();
959 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
960 return !SI->isUnordered();
964 if (isVolatile(QueryInst) || isOrdered(QueryInst)) {
965 Result.push_back(NonLocalDepResult(FromBB, MemDepResult::getUnknown(),
966 const_cast<Value *>(Loc.Ptr)));
969 const DataLayout &DL = FromBB->getModule()->getDataLayout();
970 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, &AC);
972 // This is the set of blocks we've inspected, and the pointer we consider in
973 // each block. Because of critical edges, we currently bail out if querying
974 // a block with multiple different pointers. This can happen during PHI
976 DenseMap<BasicBlock *, Value *> Visited;
977 if (getNonLocalPointerDepFromBB(QueryInst, Address, Loc, isLoad, FromBB,
978 Result, Visited, true))
981 Result.push_back(NonLocalDepResult(FromBB, MemDepResult::getUnknown(),
982 const_cast<Value *>(Loc.Ptr)));
985 /// Compute the memdep value for BB with Pointer/PointeeSize using either
986 /// cached information in Cache or by doing a lookup (which may use dirty cache
987 /// info if available).
989 /// If we do a lookup, add the result to the cache.
990 MemDepResult MemoryDependenceResults::GetNonLocalInfoForBlock(
991 Instruction *QueryInst, const MemoryLocation &Loc, bool isLoad,
992 BasicBlock *BB, NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
994 // Do a binary search to see if we already have an entry for this block in
995 // the cache set. If so, find it.
996 NonLocalDepInfo::iterator Entry = std::upper_bound(
997 Cache->begin(), Cache->begin() + NumSortedEntries, NonLocalDepEntry(BB));
998 if (Entry != Cache->begin() && (Entry - 1)->getBB() == BB)
1001 NonLocalDepEntry *ExistingResult = nullptr;
1002 if (Entry != Cache->begin() + NumSortedEntries && Entry->getBB() == BB)
1003 ExistingResult = &*Entry;
1005 // If we have a cached entry, and it is non-dirty, use it as the value for
1007 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
1008 ++NumCacheNonLocalPtr;
1009 return ExistingResult->getResult();
1012 // Otherwise, we have to scan for the value. If we have a dirty cache
1013 // entry, start scanning from its position, otherwise we scan from the end
1015 BasicBlock::iterator ScanPos = BB->end();
1016 if (ExistingResult && ExistingResult->getResult().getInst()) {
1017 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
1018 "Instruction invalidated?");
1019 ++NumCacheDirtyNonLocalPtr;
1020 ScanPos = ExistingResult->getResult().getInst()->getIterator();
1022 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1023 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
1024 RemoveFromReverseMap(ReverseNonLocalPtrDeps, &*ScanPos, CacheKey);
1026 ++NumUncacheNonLocalPtr;
1029 // Scan the block for the dependency.
1031 getPointerDependencyFrom(Loc, isLoad, ScanPos, BB, QueryInst);
1033 // If we had a dirty entry for the block, update it. Otherwise, just add
1036 ExistingResult->setResult(Dep);
1038 Cache->push_back(NonLocalDepEntry(BB, Dep));
1040 // If the block has a dependency (i.e. it isn't completely transparent to
1041 // the value), remember the reverse association because we just added it
1043 if (!Dep.isDef() && !Dep.isClobber())
1046 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
1047 // update MemDep when we remove instructions.
1048 Instruction *Inst = Dep.getInst();
1049 assert(Inst && "Didn't depend on anything?");
1050 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
1051 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
1055 /// Sort the NonLocalDepInfo cache, given a certain number of elements in the
1056 /// array that are already properly ordered.
1058 /// This is optimized for the case when only a few entries are added.
1060 SortNonLocalDepInfoCache(MemoryDependenceResults::NonLocalDepInfo &Cache,
1061 unsigned NumSortedEntries) {
1062 switch (Cache.size() - NumSortedEntries) {
1064 // done, no new entries.
1067 // Two new entries, insert the last one into place.
1068 NonLocalDepEntry Val = Cache.back();
1070 MemoryDependenceResults::NonLocalDepInfo::iterator Entry =
1071 std::upper_bound(Cache.begin(), Cache.end() - 1, Val);
1072 Cache.insert(Entry, Val);
1076 // One new entry, Just insert the new value at the appropriate position.
1077 if (Cache.size() != 1) {
1078 NonLocalDepEntry Val = Cache.back();
1080 MemoryDependenceResults::NonLocalDepInfo::iterator Entry =
1081 std::upper_bound(Cache.begin(), Cache.end(), Val);
1082 Cache.insert(Entry, Val);
1086 // Added many values, do a full scale sort.
1087 std::sort(Cache.begin(), Cache.end());
1092 /// Perform a dependency query based on pointer/pointeesize starting at the end
1095 /// Add any clobber/def results to the results vector and keep track of which
1096 /// blocks are visited in 'Visited'.
1098 /// This has special behavior for the first block queries (when SkipFirstBlock
1099 /// is true). In this special case, it ignores the contents of the specified
1100 /// block and starts returning dependence info for its predecessors.
1102 /// This function returns true on success, or false to indicate that it could
1103 /// not compute dependence information for some reason. This should be treated
1104 /// as a clobber dependence on the first instruction in the predecessor block.
1105 bool MemoryDependenceResults::getNonLocalPointerDepFromBB(
1106 Instruction *QueryInst, const PHITransAddr &Pointer,
1107 const MemoryLocation &Loc, bool isLoad, BasicBlock *StartBB,
1108 SmallVectorImpl<NonLocalDepResult> &Result,
1109 DenseMap<BasicBlock *, Value *> &Visited, bool SkipFirstBlock) {
1110 // Look up the cached info for Pointer.
1111 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
1113 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
1114 // CacheKey, this value will be inserted as the associated value. Otherwise,
1115 // it'll be ignored, and we'll have to check to see if the cached size and
1116 // aa tags are consistent with the current query.
1117 NonLocalPointerInfo InitialNLPI;
1118 InitialNLPI.Size = Loc.Size;
1119 InitialNLPI.AATags = Loc.AATags;
1121 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
1122 // already have one.
1123 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
1124 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
1125 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
1127 // If we already have a cache entry for this CacheKey, we may need to do some
1128 // work to reconcile the cache entry and the current query.
1130 if (CacheInfo->Size != Loc.Size) {
1131 // The query's Size differs from the cached one. Throw out the
1132 // cached data and proceed with the query at the new size.
1133 CacheInfo->Pair = BBSkipFirstBlockPair();
1134 CacheInfo->Size = Loc.Size;
1135 for (auto &Entry : CacheInfo->NonLocalDeps)
1136 if (Instruction *Inst = Entry.getResult().getInst())
1137 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1138 CacheInfo->NonLocalDeps.clear();
1141 // If the query's AATags are inconsistent with the cached one,
1142 // conservatively throw out the cached data and restart the query with
1143 // no tag if needed.
1144 if (CacheInfo->AATags != Loc.AATags) {
1145 if (CacheInfo->AATags) {
1146 CacheInfo->Pair = BBSkipFirstBlockPair();
1147 CacheInfo->AATags = AAMDNodes();
1148 for (auto &Entry : CacheInfo->NonLocalDeps)
1149 if (Instruction *Inst = Entry.getResult().getInst())
1150 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1151 CacheInfo->NonLocalDeps.clear();
1154 return getNonLocalPointerDepFromBB(
1155 QueryInst, Pointer, Loc.getWithoutAATags(), isLoad, StartBB, Result,
1156 Visited, SkipFirstBlock);
1160 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1162 // If we have valid cached information for exactly the block we are
1163 // investigating, just return it with no recomputation.
1164 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1165 // We have a fully cached result for this query then we can just return the
1166 // cached results and populate the visited set. However, we have to verify
1167 // that we don't already have conflicting results for these blocks. Check
1168 // to ensure that if a block in the results set is in the visited set that
1169 // it was for the same pointer query.
1170 if (!Visited.empty()) {
1171 for (auto &Entry : *Cache) {
1172 DenseMap<BasicBlock *, Value *>::iterator VI =
1173 Visited.find(Entry.getBB());
1174 if (VI == Visited.end() || VI->second == Pointer.getAddr())
1177 // We have a pointer mismatch in a block. Just return false, saying
1178 // that something was clobbered in this result. We could also do a
1179 // non-fully cached query, but there is little point in doing this.
1184 Value *Addr = Pointer.getAddr();
1185 for (auto &Entry : *Cache) {
1186 Visited.insert(std::make_pair(Entry.getBB(), Addr));
1187 if (Entry.getResult().isNonLocal()) {
1191 if (DT.isReachableFromEntry(Entry.getBB())) {
1193 NonLocalDepResult(Entry.getBB(), Entry.getResult(), Addr));
1196 ++NumCacheCompleteNonLocalPtr;
1200 // Otherwise, either this is a new block, a block with an invalid cache
1201 // pointer or one that we're about to invalidate by putting more info into it
1202 // than its valid cache info. If empty, the result will be valid cache info,
1203 // otherwise it isn't.
1205 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1207 CacheInfo->Pair = BBSkipFirstBlockPair();
1209 SmallVector<BasicBlock *, 32> Worklist;
1210 Worklist.push_back(StartBB);
1212 // PredList used inside loop.
1213 SmallVector<std::pair<BasicBlock *, PHITransAddr>, 16> PredList;
1215 // Keep track of the entries that we know are sorted. Previously cached
1216 // entries will all be sorted. The entries we add we only sort on demand (we
1217 // don't insert every element into its sorted position). We know that we
1218 // won't get any reuse from currently inserted values, because we don't
1219 // revisit blocks after we insert info for them.
1220 unsigned NumSortedEntries = Cache->size();
1221 unsigned WorklistEntries = BlockNumberLimit;
1222 bool GotWorklistLimit = false;
1223 DEBUG(AssertSorted(*Cache));
1225 while (!Worklist.empty()) {
1226 BasicBlock *BB = Worklist.pop_back_val();
1228 // If we do process a large number of blocks it becomes very expensive and
1229 // likely it isn't worth worrying about
1230 if (Result.size() > NumResultsLimit) {
1232 // Sort it now (if needed) so that recursive invocations of
1233 // getNonLocalPointerDepFromBB and other routines that could reuse the
1234 // cache value will only see properly sorted cache arrays.
1235 if (Cache && NumSortedEntries != Cache->size()) {
1236 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1238 // Since we bail out, the "Cache" set won't contain all of the
1239 // results for the query. This is ok (we can still use it to accelerate
1240 // specific block queries) but we can't do the fastpath "return all
1241 // results from the set". Clear out the indicator for this.
1242 CacheInfo->Pair = BBSkipFirstBlockPair();
1246 // Skip the first block if we have it.
1247 if (!SkipFirstBlock) {
1248 // Analyze the dependency of *Pointer in FromBB. See if we already have
1250 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1252 // Get the dependency info for Pointer in BB. If we have cached
1253 // information, we will use it, otherwise we compute it.
1254 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1255 MemDepResult Dep = GetNonLocalInfoForBlock(QueryInst, Loc, isLoad, BB,
1256 Cache, NumSortedEntries);
1258 // If we got a Def or Clobber, add this to the list of results.
1259 if (!Dep.isNonLocal()) {
1260 if (DT.isReachableFromEntry(BB)) {
1261 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1267 // If 'Pointer' is an instruction defined in this block, then we need to do
1268 // phi translation to change it into a value live in the predecessor block.
1269 // If not, we just add the predecessors to the worklist and scan them with
1270 // the same Pointer.
1271 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1272 SkipFirstBlock = false;
1273 SmallVector<BasicBlock *, 16> NewBlocks;
1274 for (BasicBlock *Pred : PredCache.get(BB)) {
1275 // Verify that we haven't looked at this block yet.
1276 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> InsertRes =
1277 Visited.insert(std::make_pair(Pred, Pointer.getAddr()));
1278 if (InsertRes.second) {
1279 // First time we've looked at *PI.
1280 NewBlocks.push_back(Pred);
1284 // If we have seen this block before, but it was with a different
1285 // pointer then we have a phi translation failure and we have to treat
1286 // this as a clobber.
1287 if (InsertRes.first->second != Pointer.getAddr()) {
1288 // Make sure to clean up the Visited map before continuing on to
1289 // PredTranslationFailure.
1290 for (unsigned i = 0; i < NewBlocks.size(); i++)
1291 Visited.erase(NewBlocks[i]);
1292 goto PredTranslationFailure;
1295 if (NewBlocks.size() > WorklistEntries) {
1296 // Make sure to clean up the Visited map before continuing on to
1297 // PredTranslationFailure.
1298 for (unsigned i = 0; i < NewBlocks.size(); i++)
1299 Visited.erase(NewBlocks[i]);
1300 GotWorklistLimit = true;
1301 goto PredTranslationFailure;
1303 WorklistEntries -= NewBlocks.size();
1304 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1308 // We do need to do phi translation, if we know ahead of time we can't phi
1309 // translate this value, don't even try.
1310 if (!Pointer.IsPotentiallyPHITranslatable())
1311 goto PredTranslationFailure;
1313 // We may have added values to the cache list before this PHI translation.
1314 // If so, we haven't done anything to ensure that the cache remains sorted.
1315 // Sort it now (if needed) so that recursive invocations of
1316 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1317 // value will only see properly sorted cache arrays.
1318 if (Cache && NumSortedEntries != Cache->size()) {
1319 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1320 NumSortedEntries = Cache->size();
1325 for (BasicBlock *Pred : PredCache.get(BB)) {
1326 PredList.push_back(std::make_pair(Pred, Pointer));
1328 // Get the PHI translated pointer in this predecessor. This can fail if
1329 // not translatable, in which case the getAddr() returns null.
1330 PHITransAddr &PredPointer = PredList.back().second;
1331 PredPointer.PHITranslateValue(BB, Pred, &DT, /*MustDominate=*/false);
1332 Value *PredPtrVal = PredPointer.getAddr();
1334 // Check to see if we have already visited this pred block with another
1335 // pointer. If so, we can't do this lookup. This failure can occur
1336 // with PHI translation when a critical edge exists and the PHI node in
1337 // the successor translates to a pointer value different than the
1338 // pointer the block was first analyzed with.
1339 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> InsertRes =
1340 Visited.insert(std::make_pair(Pred, PredPtrVal));
1342 if (!InsertRes.second) {
1343 // We found the pred; take it off the list of preds to visit.
1344 PredList.pop_back();
1346 // If the predecessor was visited with PredPtr, then we already did
1347 // the analysis and can ignore it.
1348 if (InsertRes.first->second == PredPtrVal)
1351 // Otherwise, the block was previously analyzed with a different
1352 // pointer. We can't represent the result of this case, so we just
1353 // treat this as a phi translation failure.
1355 // Make sure to clean up the Visited map before continuing on to
1356 // PredTranslationFailure.
1357 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1358 Visited.erase(PredList[i].first);
1360 goto PredTranslationFailure;
1364 // Actually process results here; this need to be a separate loop to avoid
1365 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1366 // any results for. (getNonLocalPointerDepFromBB will modify our
1367 // datastructures in ways the code after the PredTranslationFailure label
1369 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1370 BasicBlock *Pred = PredList[i].first;
1371 PHITransAddr &PredPointer = PredList[i].second;
1372 Value *PredPtrVal = PredPointer.getAddr();
1374 bool CanTranslate = true;
1375 // If PHI translation was unable to find an available pointer in this
1376 // predecessor, then we have to assume that the pointer is clobbered in
1377 // that predecessor. We can still do PRE of the load, which would insert
1378 // a computation of the pointer in this predecessor.
1380 CanTranslate = false;
1382 // FIXME: it is entirely possible that PHI translating will end up with
1383 // the same value. Consider PHI translating something like:
1384 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1385 // to recurse here, pedantically speaking.
1387 // If getNonLocalPointerDepFromBB fails here, that means the cached
1388 // result conflicted with the Visited list; we have to conservatively
1389 // assume it is unknown, but this also does not block PRE of the load.
1390 if (!CanTranslate ||
1391 !getNonLocalPointerDepFromBB(QueryInst, PredPointer,
1392 Loc.getWithNewPtr(PredPtrVal), isLoad,
1393 Pred, Result, Visited)) {
1394 // Add the entry to the Result list.
1395 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1396 Result.push_back(Entry);
1398 // Since we had a phi translation failure, the cache for CacheKey won't
1399 // include all of the entries that we need to immediately satisfy future
1400 // queries. Mark this in NonLocalPointerDeps by setting the
1401 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1402 // cached value to do more work but not miss the phi trans failure.
1403 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1404 NLPI.Pair = BBSkipFirstBlockPair();
1409 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1410 CacheInfo = &NonLocalPointerDeps[CacheKey];
1411 Cache = &CacheInfo->NonLocalDeps;
1412 NumSortedEntries = Cache->size();
1414 // Since we did phi translation, the "Cache" set won't contain all of the
1415 // results for the query. This is ok (we can still use it to accelerate
1416 // specific block queries) but we can't do the fastpath "return all
1417 // results from the set" Clear out the indicator for this.
1418 CacheInfo->Pair = BBSkipFirstBlockPair();
1419 SkipFirstBlock = false;
1422 PredTranslationFailure:
1423 // The following code is "failure"; we can't produce a sane translation
1424 // for the given block. It assumes that we haven't modified any of
1425 // our datastructures while processing the current block.
1428 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1429 CacheInfo = &NonLocalPointerDeps[CacheKey];
1430 Cache = &CacheInfo->NonLocalDeps;
1431 NumSortedEntries = Cache->size();
1434 // Since we failed phi translation, the "Cache" set won't contain all of the
1435 // results for the query. This is ok (we can still use it to accelerate
1436 // specific block queries) but we can't do the fastpath "return all
1437 // results from the set". Clear out the indicator for this.
1438 CacheInfo->Pair = BBSkipFirstBlockPair();
1440 // If *nothing* works, mark the pointer as unknown.
1442 // If this is the magic first block, return this as a clobber of the whole
1443 // incoming value. Since we can't phi translate to one of the predecessors,
1444 // we have to bail out.
1448 bool foundBlock = false;
1449 for (NonLocalDepEntry &I : llvm::reverse(*Cache)) {
1450 if (I.getBB() != BB)
1453 assert((GotWorklistLimit || I.getResult().isNonLocal() ||
1454 !DT.isReachableFromEntry(BB)) &&
1455 "Should only be here with transparent block");
1457 I.setResult(MemDepResult::getUnknown());
1459 NonLocalDepResult(I.getBB(), I.getResult(), Pointer.getAddr()));
1462 (void)foundBlock; (void)GotWorklistLimit;
1463 assert((foundBlock || GotWorklistLimit) && "Current block not in cache?");
1466 // Okay, we're done now. If we added new values to the cache, re-sort it.
1467 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1468 DEBUG(AssertSorted(*Cache));
1472 /// If P exists in CachedNonLocalPointerInfo, remove it.
1473 void MemoryDependenceResults::RemoveCachedNonLocalPointerDependencies(
1474 ValueIsLoadPair P) {
1475 CachedNonLocalPointerInfo::iterator It = NonLocalPointerDeps.find(P);
1476 if (It == NonLocalPointerDeps.end())
1479 // Remove all of the entries in the BB->val map. This involves removing
1480 // instructions from the reverse map.
1481 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1483 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1484 Instruction *Target = PInfo[i].getResult().getInst();
1486 continue; // Ignore non-local dep results.
1487 assert(Target->getParent() == PInfo[i].getBB());
1489 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1490 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1493 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1494 NonLocalPointerDeps.erase(It);
1497 void MemoryDependenceResults::invalidateCachedPointerInfo(Value *Ptr) {
1498 // If Ptr isn't really a pointer, just ignore it.
1499 if (!Ptr->getType()->isPointerTy())
1501 // Flush store info for the pointer.
1502 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1503 // Flush load info for the pointer.
1504 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1507 void MemoryDependenceResults::invalidateCachedPredecessors() {
1511 void MemoryDependenceResults::removeInstruction(Instruction *RemInst) {
1512 // Walk through the Non-local dependencies, removing this one as the value
1513 // for any cached queries.
1514 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1515 if (NLDI != NonLocalDeps.end()) {
1516 NonLocalDepInfo &BlockMap = NLDI->second.first;
1517 for (auto &Entry : BlockMap)
1518 if (Instruction *Inst = Entry.getResult().getInst())
1519 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1520 NonLocalDeps.erase(NLDI);
1523 // If we have a cached local dependence query for this instruction, remove it.
1524 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1525 if (LocalDepEntry != LocalDeps.end()) {
1526 // Remove us from DepInst's reverse set now that the local dep info is gone.
1527 if (Instruction *Inst = LocalDepEntry->second.getInst())
1528 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1530 // Remove this local dependency info.
1531 LocalDeps.erase(LocalDepEntry);
1534 // If we have any cached pointer dependencies on this instruction, remove
1535 // them. If the instruction has non-pointer type, then it can't be a pointer
1538 // Remove it from both the load info and the store info. The instruction
1539 // can't be in either of these maps if it is non-pointer.
1540 if (RemInst->getType()->isPointerTy()) {
1541 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1542 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1545 // Loop over all of the things that depend on the instruction we're removing.
1546 SmallVector<std::pair<Instruction *, Instruction *>, 8> ReverseDepsToAdd;
1548 // If we find RemInst as a clobber or Def in any of the maps for other values,
1549 // we need to replace its entry with a dirty version of the instruction after
1550 // it. If RemInst is a terminator, we use a null dirty value.
1552 // Using a dirty version of the instruction after RemInst saves having to scan
1553 // the entire block to get to this point.
1554 MemDepResult NewDirtyVal;
1555 if (!RemInst->isTerminator())
1556 NewDirtyVal = MemDepResult::getDirty(&*++RemInst->getIterator());
1558 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1559 if (ReverseDepIt != ReverseLocalDeps.end()) {
1560 // RemInst can't be the terminator if it has local stuff depending on it.
1561 assert(!ReverseDepIt->second.empty() && !isa<TerminatorInst>(RemInst) &&
1562 "Nothing can locally depend on a terminator");
1564 for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) {
1565 assert(InstDependingOnRemInst != RemInst &&
1566 "Already removed our local dep info");
1568 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1570 // Make sure to remember that new things depend on NewDepInst.
1571 assert(NewDirtyVal.getInst() &&
1572 "There is no way something else can have "
1573 "a local dep on this if it is a terminator!");
1574 ReverseDepsToAdd.push_back(
1575 std::make_pair(NewDirtyVal.getInst(), InstDependingOnRemInst));
1578 ReverseLocalDeps.erase(ReverseDepIt);
1580 // Add new reverse deps after scanning the set, to avoid invalidating the
1581 // 'ReverseDeps' reference.
1582 while (!ReverseDepsToAdd.empty()) {
1583 ReverseLocalDeps[ReverseDepsToAdd.back().first].insert(
1584 ReverseDepsToAdd.back().second);
1585 ReverseDepsToAdd.pop_back();
1589 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1590 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1591 for (Instruction *I : ReverseDepIt->second) {
1592 assert(I != RemInst && "Already removed NonLocalDep info for RemInst");
1594 PerInstNLInfo &INLD = NonLocalDeps[I];
1595 // The information is now dirty!
1598 for (auto &Entry : INLD.first) {
1599 if (Entry.getResult().getInst() != RemInst)
1602 // Convert to a dirty entry for the subsequent instruction.
1603 Entry.setResult(NewDirtyVal);
1605 if (Instruction *NextI = NewDirtyVal.getInst())
1606 ReverseDepsToAdd.push_back(std::make_pair(NextI, I));
1610 ReverseNonLocalDeps.erase(ReverseDepIt);
1612 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1613 while (!ReverseDepsToAdd.empty()) {
1614 ReverseNonLocalDeps[ReverseDepsToAdd.back().first].insert(
1615 ReverseDepsToAdd.back().second);
1616 ReverseDepsToAdd.pop_back();
1620 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1621 // value in the NonLocalPointerDeps info.
1622 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1623 ReverseNonLocalPtrDeps.find(RemInst);
1624 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1625 SmallVector<std::pair<Instruction *, ValueIsLoadPair>, 8>
1626 ReversePtrDepsToAdd;
1628 for (ValueIsLoadPair P : ReversePtrDepIt->second) {
1629 assert(P.getPointer() != RemInst &&
1630 "Already removed NonLocalPointerDeps info for RemInst");
1632 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1634 // The cache is not valid for any specific block anymore.
1635 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1637 // Update any entries for RemInst to use the instruction after it.
1638 for (auto &Entry : NLPDI) {
1639 if (Entry.getResult().getInst() != RemInst)
1642 // Convert to a dirty entry for the subsequent instruction.
1643 Entry.setResult(NewDirtyVal);
1645 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1646 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1649 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1650 // subsequent value may invalidate the sortedness.
1651 std::sort(NLPDI.begin(), NLPDI.end());
1654 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1656 while (!ReversePtrDepsToAdd.empty()) {
1657 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first].insert(
1658 ReversePtrDepsToAdd.back().second);
1659 ReversePtrDepsToAdd.pop_back();
1663 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1664 DEBUG(verifyRemoved(RemInst));
1667 /// Verify that the specified instruction does not occur in our internal data
1670 /// This function verifies by asserting in debug builds.
1671 void MemoryDependenceResults::verifyRemoved(Instruction *D) const {
1673 for (const auto &DepKV : LocalDeps) {
1674 assert(DepKV.first != D && "Inst occurs in data structures");
1675 assert(DepKV.second.getInst() != D && "Inst occurs in data structures");
1678 for (const auto &DepKV : NonLocalPointerDeps) {
1679 assert(DepKV.first.getPointer() != D && "Inst occurs in NLPD map key");
1680 for (const auto &Entry : DepKV.second.NonLocalDeps)
1681 assert(Entry.getResult().getInst() != D && "Inst occurs as NLPD value");
1684 for (const auto &DepKV : NonLocalDeps) {
1685 assert(DepKV.first != D && "Inst occurs in data structures");
1686 const PerInstNLInfo &INLD = DepKV.second;
1687 for (const auto &Entry : INLD.first)
1688 assert(Entry.getResult().getInst() != D &&
1689 "Inst occurs in data structures");
1692 for (const auto &DepKV : ReverseLocalDeps) {
1693 assert(DepKV.first != D && "Inst occurs in data structures");
1694 for (Instruction *Inst : DepKV.second)
1695 assert(Inst != D && "Inst occurs in data structures");
1698 for (const auto &DepKV : ReverseNonLocalDeps) {
1699 assert(DepKV.first != D && "Inst occurs in data structures");
1700 for (Instruction *Inst : DepKV.second)
1701 assert(Inst != D && "Inst occurs in data structures");
1704 for (const auto &DepKV : ReverseNonLocalPtrDeps) {
1705 assert(DepKV.first != D && "Inst occurs in rev NLPD map");
1707 for (ValueIsLoadPair P : DepKV.second)
1708 assert(P != ValueIsLoadPair(D, false) && P != ValueIsLoadPair(D, true) &&
1709 "Inst occurs in ReverseNonLocalPtrDeps map");
1714 AnalysisKey MemoryDependenceAnalysis::Key;
1716 MemoryDependenceResults
1717 MemoryDependenceAnalysis::run(Function &F, FunctionAnalysisManager &AM) {
1718 auto &AA = AM.getResult<AAManager>(F);
1719 auto &AC = AM.getResult<AssumptionAnalysis>(F);
1720 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1721 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1722 return MemoryDependenceResults(AA, AC, TLI, DT);
1725 char MemoryDependenceWrapperPass::ID = 0;
1727 INITIALIZE_PASS_BEGIN(MemoryDependenceWrapperPass, "memdep",
1728 "Memory Dependence Analysis", false, true)
1729 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1730 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1731 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1732 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
1733 INITIALIZE_PASS_END(MemoryDependenceWrapperPass, "memdep",
1734 "Memory Dependence Analysis", false, true)
1736 MemoryDependenceWrapperPass::MemoryDependenceWrapperPass() : FunctionPass(ID) {
1737 initializeMemoryDependenceWrapperPassPass(*PassRegistry::getPassRegistry());
1740 MemoryDependenceWrapperPass::~MemoryDependenceWrapperPass() = default;
1742 void MemoryDependenceWrapperPass::releaseMemory() {
1746 void MemoryDependenceWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
1747 AU.setPreservesAll();
1748 AU.addRequired<AssumptionCacheTracker>();
1749 AU.addRequired<DominatorTreeWrapperPass>();
1750 AU.addRequiredTransitive<AAResultsWrapperPass>();
1751 AU.addRequiredTransitive<TargetLibraryInfoWrapperPass>();
1754 bool MemoryDependenceResults::invalidate(Function &F, const PreservedAnalyses &PA,
1755 FunctionAnalysisManager::Invalidator &Inv) {
1756 // Check whether our analysis is preserved.
1757 auto PAC = PA.getChecker<MemoryDependenceAnalysis>();
1758 if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Function>>())
1759 // If not, give up now.
1762 // Check whether the analyses we depend on became invalid for any reason.
1763 if (Inv.invalidate<AAManager>(F, PA) ||
1764 Inv.invalidate<AssumptionAnalysis>(F, PA) ||
1765 Inv.invalidate<DominatorTreeAnalysis>(F, PA))
1768 // Otherwise this analysis result remains valid.
1772 unsigned MemoryDependenceResults::getDefaultBlockScanLimit() const {
1773 return BlockScanLimit;
1776 bool MemoryDependenceWrapperPass::runOnFunction(Function &F) {
1777 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
1778 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1779 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1780 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1781 MemDep.emplace(AA, AC, TLI, DT);