1 //===- RewriteStatepointsForGC.cpp - Make GC relocations explicit ---------===//
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 // Rewrite an existing set of gc.statepoints such that they make potential
11 // relocations performed by the garbage collector explicit in the IR.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Pass.h"
16 #include "llvm/Analysis/CFG.h"
17 #include "llvm/Analysis/TargetTransformInfo.h"
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/ADT/DenseSet.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/StringRef.h"
23 #include "llvm/IR/BasicBlock.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/IRBuilder.h"
28 #include "llvm/IR/InstIterator.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/IR/MDBuilder.h"
34 #include "llvm/IR/Statepoint.h"
35 #include "llvm/IR/Value.h"
36 #include "llvm/IR/Verifier.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Transforms/Scalar.h"
40 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
41 #include "llvm/Transforms/Utils/Cloning.h"
42 #include "llvm/Transforms/Utils/Local.h"
43 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
45 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
49 // Print tracing output
50 static cl::opt<bool> TraceLSP("trace-rewrite-statepoints", cl::Hidden,
53 // Print the liveset found at the insert location
54 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
56 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden,
58 // Print out the base pointers for debugging
59 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden,
62 // Cost threshold measuring when it is profitable to rematerialize value instead
64 static cl::opt<unsigned>
65 RematerializationThreshold("spp-rematerialization-threshold", cl::Hidden,
69 static bool ClobberNonLive = true;
71 static bool ClobberNonLive = false;
73 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
74 cl::location(ClobberNonLive),
78 struct RewriteStatepointsForGC : public ModulePass {
79 static char ID; // Pass identification, replacement for typeid
81 RewriteStatepointsForGC() : ModulePass(ID) {
82 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
84 bool runOnFunction(Function &F);
85 bool runOnModule(Module &M) override {
88 Changed |= runOnFunction(F);
91 // stripDereferenceabilityInfo asserts that shouldRewriteStatepointsIn
92 // returns true for at least one function in the module. Since at least
93 // one function changed, we know that the precondition is satisfied.
94 stripDereferenceabilityInfo(M);
100 void getAnalysisUsage(AnalysisUsage &AU) const override {
101 // We add and rewrite a bunch of instructions, but don't really do much
102 // else. We could in theory preserve a lot more analyses here.
103 AU.addRequired<DominatorTreeWrapperPass>();
104 AU.addRequired<TargetTransformInfoWrapperPass>();
107 /// The IR fed into RewriteStatepointsForGC may have had attributes implying
108 /// dereferenceability that are no longer valid/correct after
109 /// RewriteStatepointsForGC has run. This is because semantically, after
110 /// RewriteStatepointsForGC runs, all calls to gc.statepoint "free" the entire
111 /// heap. stripDereferenceabilityInfo (conservatively) restores correctness
112 /// by erasing all attributes in the module that externally imply
113 /// dereferenceability.
115 void stripDereferenceabilityInfo(Module &M);
117 // Helpers for stripDereferenceabilityInfo
118 void stripDereferenceabilityInfoFromBody(Function &F);
119 void stripDereferenceabilityInfoFromPrototype(Function &F);
123 char RewriteStatepointsForGC::ID = 0;
125 ModulePass *llvm::createRewriteStatepointsForGCPass() {
126 return new RewriteStatepointsForGC();
129 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
130 "Make relocations explicit at statepoints", false, false)
131 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
132 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
133 "Make relocations explicit at statepoints", false, false)
136 struct GCPtrLivenessData {
137 /// Values defined in this block.
138 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet;
139 /// Values used in this block (and thus live); does not included values
140 /// killed within this block.
141 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet;
143 /// Values live into this basic block (i.e. used by any
144 /// instruction in this basic block or ones reachable from here)
145 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn;
147 /// Values live out of this basic block (i.e. live into
148 /// any successor block)
149 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut;
152 // The type of the internal cache used inside the findBasePointers family
153 // of functions. From the callers perspective, this is an opaque type and
154 // should not be inspected.
156 // In the actual implementation this caches two relations:
157 // - The base relation itself (i.e. this pointer is based on that one)
158 // - The base defining value relation (i.e. before base_phi insertion)
159 // Generally, after the execution of a full findBasePointer call, only the
160 // base relation will remain. Internally, we add a mixture of the two
161 // types, then update all the second type to the first type
162 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
163 typedef DenseSet<llvm::Value *> StatepointLiveSetTy;
164 typedef DenseMap<Instruction *, Value *> RematerializedValueMapTy;
166 struct PartiallyConstructedSafepointRecord {
167 /// The set of values known to be live accross this safepoint
168 StatepointLiveSetTy liveset;
170 /// Mapping from live pointers to a base-defining-value
171 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
173 /// The *new* gc.statepoint instruction itself. This produces the token
174 /// that normal path gc.relocates and the gc.result are tied to.
175 Instruction *StatepointToken;
177 /// Instruction to which exceptional gc relocates are attached
178 /// Makes it easier to iterate through them during relocationViaAlloca.
179 Instruction *UnwindToken;
181 /// Record live values we are rematerialized instead of relocating.
182 /// They are not included into 'liveset' field.
183 /// Maps rematerialized copy to it's original value.
184 RematerializedValueMapTy RematerializedValues;
188 /// Compute the live-in set for every basic block in the function
189 static void computeLiveInValues(DominatorTree &DT, Function &F,
190 GCPtrLivenessData &Data);
192 /// Given results from the dataflow liveness computation, find the set of live
193 /// Values at a particular instruction.
194 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
195 StatepointLiveSetTy &out);
197 // TODO: Once we can get to the GCStrategy, this becomes
198 // Optional<bool> isGCManagedPointer(const Value *V) const override {
200 static bool isGCPointerType(const Type *T) {
201 if (const PointerType *PT = dyn_cast<PointerType>(T))
202 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
203 // GC managed heap. We know that a pointer into this heap needs to be
204 // updated and that no other pointer does.
205 return (1 == PT->getAddressSpace());
209 // Return true if this type is one which a) is a gc pointer or contains a GC
210 // pointer and b) is of a type this code expects to encounter as a live value.
211 // (The insertion code will assert that a type which matches (a) and not (b)
212 // is not encountered.)
213 static bool isHandledGCPointerType(Type *T) {
214 // We fully support gc pointers
215 if (isGCPointerType(T))
217 // We partially support vectors of gc pointers. The code will assert if it
218 // can't handle something.
219 if (auto VT = dyn_cast<VectorType>(T))
220 if (isGCPointerType(VT->getElementType()))
226 /// Returns true if this type contains a gc pointer whether we know how to
227 /// handle that type or not.
228 static bool containsGCPtrType(Type *Ty) {
229 if (isGCPointerType(Ty))
231 if (VectorType *VT = dyn_cast<VectorType>(Ty))
232 return isGCPointerType(VT->getScalarType());
233 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
234 return containsGCPtrType(AT->getElementType());
235 if (StructType *ST = dyn_cast<StructType>(Ty))
237 ST->subtypes().begin(), ST->subtypes().end(),
238 [](Type *SubType) { return containsGCPtrType(SubType); });
242 // Returns true if this is a type which a) is a gc pointer or contains a GC
243 // pointer and b) is of a type which the code doesn't expect (i.e. first class
244 // aggregates). Used to trip assertions.
245 static bool isUnhandledGCPointerType(Type *Ty) {
246 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
250 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
251 if (a->hasName() && b->hasName()) {
252 return -1 == a->getName().compare(b->getName());
253 } else if (a->hasName() && !b->hasName()) {
255 } else if (!a->hasName() && b->hasName()) {
258 // Better than nothing, but not stable
263 // Conservatively identifies any definitions which might be live at the
264 // given instruction. The analysis is performed immediately before the
265 // given instruction. Values defined by that instruction are not considered
266 // live. Values used by that instruction are considered live.
267 static void analyzeParsePointLiveness(
268 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
269 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
270 Instruction *inst = CS.getInstruction();
272 StatepointLiveSetTy liveset;
273 findLiveSetAtInst(inst, OriginalLivenessData, liveset);
276 // Note: This output is used by several of the test cases
277 // The order of elemtns in a set is not stable, put them in a vec and sort
279 SmallVector<Value *, 64> temp;
280 temp.insert(temp.end(), liveset.begin(), liveset.end());
281 std::sort(temp.begin(), temp.end(), order_by_name);
282 errs() << "Live Variables:\n";
283 for (Value *V : temp) {
284 errs() << " " << V->getName(); // no newline
288 if (PrintLiveSetSize) {
289 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
290 errs() << "Number live values: " << liveset.size() << "\n";
292 result.liveset = liveset;
295 static Value *findBaseDefiningValue(Value *I);
297 /// Return a base defining value for the 'Index' element of the given vector
298 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
299 /// 'I'. As an optimization, this method will try to determine when the
300 /// element is known to already be a base pointer. If this can be established,
301 /// the second value in the returned pair will be true. Note that either a
302 /// vector or a pointer typed value can be returned. For the former, the
303 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
304 /// If the later, the return pointer is a BDV (or possibly a base) for the
305 /// particular element in 'I'.
306 static std::pair<Value *, bool>
307 findBaseDefiningValueOfVector(Value *I, Value *Index = nullptr) {
308 assert(I->getType()->isVectorTy() &&
309 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
310 "Illegal to ask for the base pointer of a non-pointer type");
312 // Each case parallels findBaseDefiningValue below, see that code for
313 // detailed motivation.
315 if (isa<Argument>(I))
316 // An incoming argument to the function is a base pointer
317 return std::make_pair(I, true);
319 // We shouldn't see the address of a global as a vector value?
320 assert(!isa<GlobalVariable>(I) &&
321 "unexpected global variable found in base of vector");
323 // inlining could possibly introduce phi node that contains
324 // undef if callee has multiple returns
325 if (isa<UndefValue>(I))
326 // utterly meaningless, but useful for dealing with partially optimized
328 return std::make_pair(I, true);
330 // Due to inheritance, this must be _after_ the global variable and undef
332 if (Constant *Con = dyn_cast<Constant>(I)) {
333 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
334 "order of checks wrong!");
335 assert(Con->isNullValue() && "null is the only case which makes sense");
336 return std::make_pair(Con, true);
339 if (isa<LoadInst>(I))
340 return std::make_pair(I, true);
342 // For an insert element, we might be able to look through it if we know
343 // something about the indexes.
344 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(I)) {
346 Value *InsertIndex = IEI->getOperand(2);
347 // This index is inserting the value, look for its BDV
348 if (InsertIndex == Index)
349 return std::make_pair(findBaseDefiningValue(IEI->getOperand(1)), false);
350 // Both constant, and can't be equal per above. This insert is definitely
351 // not relevant, look back at the rest of the vector and keep trying.
352 if (isa<ConstantInt>(Index) && isa<ConstantInt>(InsertIndex))
353 return findBaseDefiningValueOfVector(IEI->getOperand(0), Index);
356 // We don't know whether this vector contains entirely base pointers or
357 // not. To be conservatively correct, we treat it as a BDV and will
358 // duplicate code as needed to construct a parallel vector of bases.
359 return std::make_pair(IEI, false);
362 if (isa<ShuffleVectorInst>(I))
363 // We don't know whether this vector contains entirely base pointers or
364 // not. To be conservatively correct, we treat it as a BDV and will
365 // duplicate code as needed to construct a parallel vector of bases.
366 // TODO: There a number of local optimizations which could be applied here
367 // for particular sufflevector patterns.
368 return std::make_pair(I, false);
370 // A PHI or Select is a base defining value. The outer findBasePointer
371 // algorithm is responsible for constructing a base value for this BDV.
372 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
373 "unknown vector instruction - no base found for vector element");
374 return std::make_pair(I, false);
377 static bool isKnownBaseResult(Value *V);
379 /// Helper function for findBasePointer - Will return a value which either a)
380 /// defines the base pointer for the input or b) blocks the simple search
381 /// (i.e. a PHI or Select of two derived pointers)
382 static Value *findBaseDefiningValue(Value *I) {
383 if (I->getType()->isVectorTy())
384 return findBaseDefiningValueOfVector(I).first;
386 assert(I->getType()->isPointerTy() &&
387 "Illegal to ask for the base pointer of a non-pointer type");
389 // This case is a bit of a hack - it only handles extracts from vectors which
390 // trivially contain only base pointers or cases where we can directly match
391 // the index of the original extract element to an insertion into the vector.
392 // See note inside the function for how to improve this.
393 if (auto *EEI = dyn_cast<ExtractElementInst>(I)) {
394 Value *VectorOperand = EEI->getVectorOperand();
395 Value *Index = EEI->getIndexOperand();
396 std::pair<Value *, bool> pair =
397 findBaseDefiningValueOfVector(VectorOperand, Index);
398 Value *VectorBase = pair.first;
399 if (VectorBase->getType()->isPointerTy())
400 // We found a BDV for this specific element with the vector. This is an
401 // optimization, but in practice it covers most of the useful cases
402 // created via scalarization.
405 assert(VectorBase->getType()->isVectorTy());
407 // If the entire vector returned is known to be entirely base pointers,
408 // then the extractelement is valid base for this value.
411 // Otherwise, we have an instruction which potentially produces a
412 // derived pointer and we need findBasePointers to clone code for us
413 // such that we can create an instruction which produces the
414 // accompanying base pointer.
415 // Note: This code is currently rather incomplete. We don't currently
416 // support the general form of shufflevector of insertelement.
417 // Conceptually, these are just 'base defining values' of the same
418 // variety as phi or select instructions. We need to update the
419 // findBasePointers algorithm to insert new 'base-only' versions of the
420 // original instructions. This is relative straight forward to do, but
421 // the case which would motivate the work hasn't shown up in real
423 assert((isa<PHINode>(VectorBase) || isa<SelectInst>(VectorBase)) &&
424 "need to extend findBasePointers for generic vector"
425 "instruction cases");
431 if (isa<Argument>(I))
432 // An incoming argument to the function is a base pointer
433 // We should have never reached here if this argument isn't an gc value
436 if (isa<GlobalVariable>(I))
440 // inlining could possibly introduce phi node that contains
441 // undef if callee has multiple returns
442 if (isa<UndefValue>(I))
443 // utterly meaningless, but useful for dealing with
444 // partially optimized code.
447 // Due to inheritance, this must be _after_ the global variable and undef
449 if (Constant *Con = dyn_cast<Constant>(I)) {
450 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
451 "order of checks wrong!");
452 // Note: Finding a constant base for something marked for relocation
453 // doesn't really make sense. The most likely case is either a) some
454 // screwed up the address space usage or b) your validating against
455 // compiled C++ code w/o the proper separation. The only real exception
456 // is a null pointer. You could have generic code written to index of
457 // off a potentially null value and have proven it null. We also use
458 // null pointers in dead paths of relocation phis (which we might later
459 // want to find a base pointer for).
460 assert(isa<ConstantPointerNull>(Con) &&
461 "null is the only case which makes sense");
465 if (CastInst *CI = dyn_cast<CastInst>(I)) {
466 Value *Def = CI->stripPointerCasts();
467 // If we find a cast instruction here, it means we've found a cast which is
468 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
469 // handle int->ptr conversion.
470 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
471 return findBaseDefiningValue(Def);
474 if (isa<LoadInst>(I))
475 return I; // The value loaded is an gc base itself
477 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
478 // The base of this GEP is the base
479 return findBaseDefiningValue(GEP->getPointerOperand());
481 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
482 switch (II->getIntrinsicID()) {
483 case Intrinsic::experimental_gc_result_ptr:
485 // fall through to general call handling
487 case Intrinsic::experimental_gc_statepoint:
488 case Intrinsic::experimental_gc_result_float:
489 case Intrinsic::experimental_gc_result_int:
490 llvm_unreachable("these don't produce pointers");
491 case Intrinsic::experimental_gc_relocate: {
492 // Rerunning safepoint insertion after safepoints are already
493 // inserted is not supported. It could probably be made to work,
494 // but why are you doing this? There's no good reason.
495 llvm_unreachable("repeat safepoint insertion is not supported");
497 case Intrinsic::gcroot:
498 // Currently, this mechanism hasn't been extended to work with gcroot.
499 // There's no reason it couldn't be, but I haven't thought about the
500 // implications much.
502 "interaction with the gcroot mechanism is not supported");
505 // We assume that functions in the source language only return base
506 // pointers. This should probably be generalized via attributes to support
507 // both source language and internal functions.
508 if (isa<CallInst>(I) || isa<InvokeInst>(I))
511 // I have absolutely no idea how to implement this part yet. It's not
512 // neccessarily hard, I just haven't really looked at it yet.
513 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
515 if (isa<AtomicCmpXchgInst>(I))
516 // A CAS is effectively a atomic store and load combined under a
517 // predicate. From the perspective of base pointers, we just treat it
521 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
522 "binary ops which don't apply to pointers");
524 // The aggregate ops. Aggregates can either be in the heap or on the
525 // stack, but in either case, this is simply a field load. As a result,
526 // this is a defining definition of the base just like a load is.
527 if (isa<ExtractValueInst>(I))
530 // We should never see an insert vector since that would require we be
531 // tracing back a struct value not a pointer value.
532 assert(!isa<InsertValueInst>(I) &&
533 "Base pointer for a struct is meaningless");
535 // The last two cases here don't return a base pointer. Instead, they
536 // return a value which dynamically selects from amoung several base
537 // derived pointers (each with it's own base potentially). It's the job of
538 // the caller to resolve these.
539 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
540 "missing instruction case in findBaseDefiningValing");
544 /// Returns the base defining value for this value.
545 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
546 Value *&Cached = Cache[I];
548 Cached = findBaseDefiningValue(I);
549 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
550 << Cached->getName() << "\n");
552 assert(Cache[I] != nullptr);
556 /// Return a base pointer for this value if known. Otherwise, return it's
557 /// base defining value.
558 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
559 Value *Def = findBaseDefiningValueCached(I, Cache);
560 auto Found = Cache.find(Def);
561 if (Found != Cache.end()) {
562 // Either a base-of relation, or a self reference. Caller must check.
563 return Found->second;
565 // Only a BDV available
569 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
570 /// is it known to be a base pointer? Or do we need to continue searching.
571 static bool isKnownBaseResult(Value *V) {
572 if (!isa<PHINode>(V) && !isa<SelectInst>(V)) {
573 // no recursion possible
576 if (isa<Instruction>(V) &&
577 cast<Instruction>(V)->getMetadata("is_base_value")) {
578 // This is a previously inserted base phi or select. We know
579 // that this is a base value.
583 // We need to keep searching
587 // TODO: find a better name for this
591 enum Status { Unknown, Base, Conflict };
593 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
594 assert(status != Base || b);
596 PhiState(Value *b) : status(Base), base(b) {}
597 PhiState() : status(Unknown), base(nullptr) {}
599 Status getStatus() const { return status; }
600 Value *getBase() const { return base; }
602 bool isBase() const { return getStatus() == Base; }
603 bool isUnknown() const { return getStatus() == Unknown; }
604 bool isConflict() const { return getStatus() == Conflict; }
606 bool operator==(const PhiState &other) const {
607 return base == other.base && status == other.status;
610 bool operator!=(const PhiState &other) const { return !(*this == other); }
613 void dump() const { print(dbgs()); dbgs() << '\n'; }
615 void print(raw_ostream &OS) const {
616 OS << status << " (" << base << " - "
617 << (base ? base->getName() : "nullptr") << "): ";
622 Value *base; // non null only if status == base
625 inline raw_ostream &operator<<(raw_ostream &OS, const PhiState &State) {
631 typedef DenseMap<Value *, PhiState> ConflictStateMapTy;
632 // Values of type PhiState form a lattice, and this is a helper
633 // class that implementes the meet operation. The meat of the meet
634 // operation is implemented in MeetPhiStates::pureMeet
635 class MeetPhiStates {
637 /// Initializes the currentResult to the TOP state so that if can be met with
638 /// any other state to produce that state.
641 // Destructively meet the current result with the given PhiState
642 void meetWith(PhiState otherState) {
643 currentResult = meet(otherState, currentResult);
646 PhiState getResult() const { return currentResult; }
649 PhiState currentResult;
651 /// Perform a meet operation on two elements of the PhiState lattice.
652 static PhiState meet(PhiState LHS, PhiState RHS) {
653 assert((pureMeet(LHS, RHS) == pureMeet(RHS, LHS)) &&
654 "math is wrong: meet does not commute!");
655 PhiState Result = pureMeet(LHS, RHS);
656 DEBUG(dbgs() << "meet of " << LHS << " with " << RHS
657 << " produced " << Result << "\n");
661 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
662 switch (stateA.getStatus()) {
663 case PhiState::Unknown:
667 assert(stateA.getBase() && "can't be null");
668 if (stateB.isUnknown())
671 if (stateB.isBase()) {
672 if (stateA.getBase() == stateB.getBase()) {
673 assert(stateA == stateB && "equality broken!");
676 return PhiState(PhiState::Conflict);
678 assert(stateB.isConflict() && "only three states!");
679 return PhiState(PhiState::Conflict);
681 case PhiState::Conflict:
684 llvm_unreachable("only three states!");
688 /// For a given value or instruction, figure out what base ptr it's derived
689 /// from. For gc objects, this is simply itself. On success, returns a value
690 /// which is the base pointer. (This is reliable and can be used for
691 /// relocation.) On failure, returns nullptr.
692 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
693 Value *def = findBaseOrBDV(I, cache);
695 if (isKnownBaseResult(def)) {
699 // Here's the rough algorithm:
700 // - For every SSA value, construct a mapping to either an actual base
701 // pointer or a PHI which obscures the base pointer.
702 // - Construct a mapping from PHI to unknown TOP state. Use an
703 // optimistic algorithm to propagate base pointer information. Lattice
708 // When algorithm terminates, all PHIs will either have a single concrete
709 // base or be in a conflict state.
710 // - For every conflict, insert a dummy PHI node without arguments. Add
711 // these to the base[Instruction] = BasePtr mapping. For every
712 // non-conflict, add the actual base.
713 // - For every conflict, add arguments for the base[a] of each input
716 // Note: A simpler form of this would be to add the conflict form of all
717 // PHIs without running the optimistic algorithm. This would be
718 // analougous to pessimistic data flow and would likely lead to an
719 // overall worse solution.
721 ConflictStateMapTy states;
722 states[def] = PhiState();
723 // Recursively fill in all phis & selects reachable from the initial one
724 // for which we don't already know a definite base value for
725 // TODO: This should be rewritten with a worklist
729 // Since we're adding elements to 'states' as we run, we can't keep
730 // iterators into the set.
731 SmallVector<Value *, 16> Keys;
732 Keys.reserve(states.size());
733 for (auto Pair : states) {
734 Value *V = Pair.first;
737 for (Value *v : Keys) {
738 assert(!isKnownBaseResult(v) && "why did it get added?");
739 if (PHINode *phi = dyn_cast<PHINode>(v)) {
740 assert(phi->getNumIncomingValues() > 0 &&
741 "zero input phis are illegal");
742 for (Value *InVal : phi->incoming_values()) {
743 Value *local = findBaseOrBDV(InVal, cache);
744 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
745 states[local] = PhiState();
749 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
750 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
751 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
752 states[local] = PhiState();
755 local = findBaseOrBDV(sel->getFalseValue(), cache);
756 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
757 states[local] = PhiState();
765 errs() << "States after initialization:\n";
766 for (auto Pair : states)
767 dbgs() << " " << Pair.second << " for " << Pair.first << "\n";
770 // TODO: come back and revisit the state transitions around inputs which
771 // have reached conflict state. The current version seems too conservative.
773 // Return a phi state for a base defining value. We'll generate a new
774 // base state for known bases and expect to find a cached state otherwise.
775 auto getStateForBDV = [&](Value *baseValue) {
776 if (isKnownBaseResult(baseValue))
777 return PhiState(baseValue);
778 auto I = states.find(baseValue);
779 assert(I != states.end() && "lookup failed!");
783 bool progress = true;
786 size_t oldSize = states.size();
789 // We're only changing keys in this loop, thus safe to keep iterators
790 for (auto Pair : states) {
791 Value *v = Pair.first;
792 assert(!isKnownBaseResult(v) && "why did it get added?");
794 // Given an input value for the current instruction, return a PhiState
795 // instance which represents the BDV of that value.
796 auto getStateForInput = [&](Value *V) mutable {
797 Value *BDV = findBaseOrBDV(V, cache);
798 return getStateForBDV(BDV);
801 MeetPhiStates calculateMeet;
802 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
803 calculateMeet.meetWith(getStateForInput(select->getTrueValue()));
804 calculateMeet.meetWith(getStateForInput(select->getFalseValue()));
806 for (Value *Val : cast<PHINode>(v)->incoming_values())
807 calculateMeet.meetWith(getStateForInput(Val));
809 PhiState oldState = states[v];
810 PhiState newState = calculateMeet.getResult();
811 if (oldState != newState) {
813 states[v] = newState;
817 assert(oldSize <= states.size());
818 assert(oldSize == states.size() || progress);
822 errs() << "States after meet iteration:\n";
823 for (auto Pair : states)
824 dbgs() << " " << Pair.second << " for " << Pair.first << "\n";
827 // Insert Phis for all conflicts
828 // We want to keep naming deterministic in the loop that follows, so
829 // sort the keys before iteration. This is useful in allowing us to
830 // write stable tests. Note that there is no invalidation issue here.
831 SmallVector<Value *, 16> Keys;
832 Keys.reserve(states.size());
833 for (auto Pair : states) {
834 Value *V = Pair.first;
837 std::sort(Keys.begin(), Keys.end(), order_by_name);
838 // TODO: adjust naming patterns to avoid this order of iteration dependency
839 for (Value *V : Keys) {
840 Instruction *I = cast<Instruction>(V);
841 PhiState State = states[I];
842 assert(!isKnownBaseResult(I) && "why did it get added?");
843 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
844 if (!State.isConflict())
847 /// Create and insert a new instruction which will represent the base of
848 /// the given instruction 'I'.
849 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
850 if (isa<PHINode>(I)) {
851 BasicBlock *BB = I->getParent();
852 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
853 assert(NumPreds > 0 && "how did we reach here");
854 return PHINode::Create(I->getType(), NumPreds, "base_phi", I);
856 SelectInst *Sel = cast<SelectInst>(I);
857 // The undef will be replaced later
858 UndefValue *Undef = UndefValue::get(Sel->getType());
859 return SelectInst::Create(Sel->getCondition(), Undef,
860 Undef, "base_select", Sel);
862 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
863 // Add metadata marking this as a base value
864 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
865 states[I] = PhiState(PhiState::Conflict, BaseInst);
868 // Fixup all the inputs of the new PHIs
869 for (auto Pair : states) {
870 Instruction *v = cast<Instruction>(Pair.first);
871 PhiState state = Pair.second;
873 assert(!isKnownBaseResult(v) && "why did it get added?");
874 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
875 if (!state.isConflict())
878 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
879 PHINode *phi = cast<PHINode>(v);
880 unsigned NumPHIValues = phi->getNumIncomingValues();
881 for (unsigned i = 0; i < NumPHIValues; i++) {
882 Value *InVal = phi->getIncomingValue(i);
883 BasicBlock *InBB = phi->getIncomingBlock(i);
885 // If we've already seen InBB, add the same incoming value
886 // we added for it earlier. The IR verifier requires phi
887 // nodes with multiple entries from the same basic block
888 // to have the same incoming value for each of those
889 // entries. If we don't do this check here and basephi
890 // has a different type than base, we'll end up adding two
891 // bitcasts (and hence two distinct values) as incoming
892 // values for the same basic block.
894 int blockIndex = basephi->getBasicBlockIndex(InBB);
895 if (blockIndex != -1) {
896 Value *oldBase = basephi->getIncomingValue(blockIndex);
897 basephi->addIncoming(oldBase, InBB);
899 Value *base = findBaseOrBDV(InVal, cache);
900 if (!isKnownBaseResult(base)) {
901 // Either conflict or base.
902 assert(states.count(base));
903 base = states[base].getBase();
904 assert(base != nullptr && "unknown PhiState!");
907 // In essense this assert states: the only way two
908 // values incoming from the same basic block may be
909 // different is by being different bitcasts of the same
910 // value. A cleanup that remains TODO is changing
911 // findBaseOrBDV to return an llvm::Value of the correct
912 // type (and still remain pure). This will remove the
913 // need to add bitcasts.
914 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
915 "sanity -- findBaseOrBDV should be pure!");
920 // Find either the defining value for the PHI or the normal base for
922 Value *base = findBaseOrBDV(InVal, cache);
923 if (!isKnownBaseResult(base)) {
924 // Either conflict or base.
925 assert(states.count(base));
926 base = states[base].getBase();
927 assert(base != nullptr && "unknown PhiState!");
929 assert(base && "can't be null");
930 // Must use original input BB since base may not be Instruction
931 // The cast is needed since base traversal may strip away bitcasts
932 if (base->getType() != basephi->getType()) {
933 base = new BitCastInst(base, basephi->getType(), "cast",
934 InBB->getTerminator());
936 basephi->addIncoming(base, InBB);
938 assert(basephi->getNumIncomingValues() == NumPHIValues);
940 SelectInst *basesel = cast<SelectInst>(state.getBase());
941 SelectInst *sel = cast<SelectInst>(v);
942 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
943 // something more safe and less hacky.
944 for (int i = 1; i <= 2; i++) {
945 Value *InVal = sel->getOperand(i);
946 // Find either the defining value for the PHI or the normal base for
948 Value *base = findBaseOrBDV(InVal, cache);
949 if (!isKnownBaseResult(base)) {
950 // Either conflict or base.
951 assert(states.count(base));
952 base = states[base].getBase();
953 assert(base != nullptr && "unknown PhiState!");
955 assert(base && "can't be null");
956 // Must use original input BB since base may not be Instruction
957 // The cast is needed since base traversal may strip away bitcasts
958 if (base->getType() != basesel->getType()) {
959 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
961 basesel->setOperand(i, base);
966 // Cache all of our results so we can cheaply reuse them
967 // NOTE: This is actually two caches: one of the base defining value
968 // relation and one of the base pointer relation! FIXME
969 for (auto item : states) {
970 Value *v = item.first;
971 Value *base = item.second.getBase();
973 assert(!isKnownBaseResult(v) && "why did it get added?");
976 std::string fromstr =
977 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
979 errs() << "Updating base value cache"
980 << " for: " << (v->hasName() ? v->getName() : "")
981 << " from: " << fromstr
982 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
985 assert(isKnownBaseResult(base) &&
986 "must be something we 'know' is a base pointer");
987 if (cache.count(v)) {
988 // Once we transition from the BDV relation being store in the cache to
989 // the base relation being stored, it must be stable
990 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
991 "base relation should be stable");
995 assert(cache.find(def) != cache.end());
999 // For a set of live pointers (base and/or derived), identify the base
1000 // pointer of the object which they are derived from. This routine will
1001 // mutate the IR graph as needed to make the 'base' pointer live at the
1002 // definition site of 'derived'. This ensures that any use of 'derived' can
1003 // also use 'base'. This may involve the insertion of a number of
1004 // additional PHI nodes.
1006 // preconditions: live is a set of pointer type Values
1008 // side effects: may insert PHI nodes into the existing CFG, will preserve
1009 // CFG, will not remove or mutate any existing nodes
1011 // post condition: PointerToBase contains one (derived, base) pair for every
1012 // pointer in live. Note that derived can be equal to base if the original
1013 // pointer was a base pointer.
1015 findBasePointers(const StatepointLiveSetTy &live,
1016 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
1017 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1018 // For the naming of values inserted to be deterministic - which makes for
1019 // much cleaner and more stable tests - we need to assign an order to the
1020 // live values. DenseSets do not provide a deterministic order across runs.
1021 SmallVector<Value *, 64> Temp;
1022 Temp.insert(Temp.end(), live.begin(), live.end());
1023 std::sort(Temp.begin(), Temp.end(), order_by_name);
1024 for (Value *ptr : Temp) {
1025 Value *base = findBasePointer(ptr, DVCache);
1026 assert(base && "failed to find base pointer");
1027 PointerToBase[ptr] = base;
1028 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1029 DT->dominates(cast<Instruction>(base)->getParent(),
1030 cast<Instruction>(ptr)->getParent())) &&
1031 "The base we found better dominate the derived pointer");
1033 // If you see this trip and like to live really dangerously, the code should
1034 // be correct, just with idioms the verifier can't handle. You can try
1035 // disabling the verifier at your own substaintial risk.
1036 assert(!isa<ConstantPointerNull>(base) &&
1037 "the relocation code needs adjustment to handle the relocation of "
1038 "a null pointer constant without causing false positives in the "
1039 "safepoint ir verifier.");
1043 /// Find the required based pointers (and adjust the live set) for the given
1045 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1047 PartiallyConstructedSafepointRecord &result) {
1048 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
1049 findBasePointers(result.liveset, PointerToBase, &DT, DVCache);
1051 if (PrintBasePointers) {
1052 // Note: Need to print these in a stable order since this is checked in
1054 errs() << "Base Pairs (w/o Relocation):\n";
1055 SmallVector<Value *, 64> Temp;
1056 Temp.reserve(PointerToBase.size());
1057 for (auto Pair : PointerToBase) {
1058 Temp.push_back(Pair.first);
1060 std::sort(Temp.begin(), Temp.end(), order_by_name);
1061 for (Value *Ptr : Temp) {
1062 Value *Base = PointerToBase[Ptr];
1063 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
1068 result.PointerToBase = PointerToBase;
1071 /// Given an updated version of the dataflow liveness results, update the
1072 /// liveset and base pointer maps for the call site CS.
1073 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1075 PartiallyConstructedSafepointRecord &result);
1077 static void recomputeLiveInValues(
1078 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1079 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1080 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1081 // again. The old values are still live and will help it stablize quickly.
1082 GCPtrLivenessData RevisedLivenessData;
1083 computeLiveInValues(DT, F, RevisedLivenessData);
1084 for (size_t i = 0; i < records.size(); i++) {
1085 struct PartiallyConstructedSafepointRecord &info = records[i];
1086 const CallSite &CS = toUpdate[i];
1087 recomputeLiveInValues(RevisedLivenessData, CS, info);
1091 // When inserting gc.relocate calls, we need to ensure there are no uses
1092 // of the original value between the gc.statepoint and the gc.relocate call.
1093 // One case which can arise is a phi node starting one of the successor blocks.
1094 // We also need to be able to insert the gc.relocates only on the path which
1095 // goes through the statepoint. We might need to split an edge to make this
1098 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1099 DominatorTree &DT) {
1100 BasicBlock *Ret = BB;
1101 if (!BB->getUniquePredecessor()) {
1102 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1105 // Now that 'ret' has unique predecessor we can safely remove all phi nodes
1107 FoldSingleEntryPHINodes(Ret);
1108 assert(!isa<PHINode>(Ret->begin()));
1110 // At this point, we can safely insert a gc.relocate as the first instruction
1111 // in Ret if needed.
1115 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1116 auto itr = std::find(livevec.begin(), livevec.end(), val);
1117 assert(livevec.end() != itr);
1118 size_t index = std::distance(livevec.begin(), itr);
1119 assert(index < livevec.size());
1123 // Create new attribute set containing only attributes which can be transfered
1124 // from original call to the safepoint.
1125 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1128 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1129 unsigned index = AS.getSlotIndex(Slot);
1131 if (index == AttributeSet::ReturnIndex ||
1132 index == AttributeSet::FunctionIndex) {
1134 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1136 Attribute attr = *it;
1138 // Do not allow certain attributes - just skip them
1139 // Safepoint can not be read only or read none.
1140 if (attr.hasAttribute(Attribute::ReadNone) ||
1141 attr.hasAttribute(Attribute::ReadOnly))
1144 ret = ret.addAttributes(
1145 AS.getContext(), index,
1146 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1150 // Just skip parameter attributes for now
1156 /// Helper function to place all gc relocates necessary for the given
1159 /// liveVariables - list of variables to be relocated.
1160 /// liveStart - index of the first live variable.
1161 /// basePtrs - base pointers.
1162 /// statepointToken - statepoint instruction to which relocates should be
1164 /// Builder - Llvm IR builder to be used to construct new calls.
1165 static void CreateGCRelocates(ArrayRef<llvm::Value *> LiveVariables,
1166 const int LiveStart,
1167 ArrayRef<llvm::Value *> BasePtrs,
1168 Instruction *StatepointToken,
1169 IRBuilder<> Builder) {
1170 if (LiveVariables.empty())
1173 // All gc_relocate are set to i8 addrspace(1)* type. We originally generated
1174 // unique declarations for each pointer type, but this proved problematic
1175 // because the intrinsic mangling code is incomplete and fragile. Since
1176 // we're moving towards a single unified pointer type anyways, we can just
1177 // cast everything to an i8* of the right address space. A bitcast is added
1178 // later to convert gc_relocate to the actual value's type.
1179 Module *M = StatepointToken->getModule();
1180 auto AS = cast<PointerType>(LiveVariables[0]->getType())->getAddressSpace();
1181 Type *Types[] = {Type::getInt8PtrTy(M->getContext(), AS)};
1182 Value *GCRelocateDecl =
1183 Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, Types);
1185 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1186 // Generate the gc.relocate call and save the result
1188 Builder.getInt32(LiveStart + find_index(LiveVariables, BasePtrs[i]));
1190 Builder.getInt32(LiveStart + find_index(LiveVariables, LiveVariables[i]));
1192 // only specify a debug name if we can give a useful one
1193 CallInst *Reloc = Builder.CreateCall(
1194 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1195 LiveVariables[i]->hasName() ? LiveVariables[i]->getName() + ".relocated"
1197 // Trick CodeGen into thinking there are lots of free registers at this
1199 Reloc->setCallingConv(CallingConv::Cold);
1204 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1205 const SmallVectorImpl<llvm::Value *> &basePtrs,
1206 const SmallVectorImpl<llvm::Value *> &liveVariables,
1208 PartiallyConstructedSafepointRecord &result) {
1209 assert(basePtrs.size() == liveVariables.size());
1210 assert(isStatepoint(CS) &&
1211 "This method expects to be rewriting a statepoint");
1213 BasicBlock *BB = CS.getInstruction()->getParent();
1215 Function *F = BB->getParent();
1216 assert(F && "must be set");
1217 Module *M = F->getParent();
1219 assert(M && "must be set");
1221 // We're not changing the function signature of the statepoint since the gc
1222 // arguments go into the var args section.
1223 Function *gc_statepoint_decl = CS.getCalledFunction();
1225 // Then go ahead and use the builder do actually do the inserts. We insert
1226 // immediately before the previous instruction under the assumption that all
1227 // arguments will be available here. We can't insert afterwards since we may
1228 // be replacing a terminator.
1229 Instruction *insertBefore = CS.getInstruction();
1230 IRBuilder<> Builder(insertBefore);
1231 // Copy all of the arguments from the original statepoint - this includes the
1232 // target, call args, and deopt args
1233 SmallVector<llvm::Value *, 64> args;
1234 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1235 // TODO: Clear the 'needs rewrite' flag
1237 // add all the pointers to be relocated (gc arguments)
1238 // Capture the start of the live variable list for use in the gc_relocates
1239 const int live_start = args.size();
1240 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1242 // Create the statepoint given all the arguments
1243 Instruction *token = nullptr;
1244 AttributeSet return_attributes;
1246 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1248 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1249 call->setTailCall(toReplace->isTailCall());
1250 call->setCallingConv(toReplace->getCallingConv());
1252 // Currently we will fail on parameter attributes and on certain
1253 // function attributes.
1254 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1255 // In case if we can handle this set of sttributes - set up function attrs
1256 // directly on statepoint and return attrs later for gc_result intrinsic.
1257 call->setAttributes(new_attrs.getFnAttributes());
1258 return_attributes = new_attrs.getRetAttributes();
1262 // Put the following gc_result and gc_relocate calls immediately after the
1263 // the old call (which we're about to delete)
1264 BasicBlock::iterator next(toReplace);
1265 assert(BB->end() != next && "not a terminator, must have next");
1267 Instruction *IP = &*(next);
1268 Builder.SetInsertPoint(IP);
1269 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1272 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1274 // Insert the new invoke into the old block. We'll remove the old one in a
1275 // moment at which point this will become the new terminator for the
1277 InvokeInst *invoke = InvokeInst::Create(
1278 gc_statepoint_decl, toReplace->getNormalDest(),
1279 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1280 invoke->setCallingConv(toReplace->getCallingConv());
1282 // Currently we will fail on parameter attributes and on certain
1283 // function attributes.
1284 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1285 // In case if we can handle this set of sttributes - set up function attrs
1286 // directly on statepoint and return attrs later for gc_result intrinsic.
1287 invoke->setAttributes(new_attrs.getFnAttributes());
1288 return_attributes = new_attrs.getRetAttributes();
1292 // Generate gc relocates in exceptional path
1293 BasicBlock *unwindBlock = toReplace->getUnwindDest();
1294 assert(!isa<PHINode>(unwindBlock->begin()) &&
1295 unwindBlock->getUniquePredecessor() &&
1296 "can't safely insert in this block!");
1298 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1299 Builder.SetInsertPoint(IP);
1300 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1302 // Extract second element from landingpad return value. We will attach
1303 // exceptional gc relocates to it.
1304 const unsigned idx = 1;
1305 Instruction *exceptional_token =
1306 cast<Instruction>(Builder.CreateExtractValue(
1307 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1308 result.UnwindToken = exceptional_token;
1310 CreateGCRelocates(liveVariables, live_start, basePtrs,
1311 exceptional_token, Builder);
1313 // Generate gc relocates and returns for normal block
1314 BasicBlock *normalDest = toReplace->getNormalDest();
1315 assert(!isa<PHINode>(normalDest->begin()) &&
1316 normalDest->getUniquePredecessor() &&
1317 "can't safely insert in this block!");
1319 IP = &*(normalDest->getFirstInsertionPt());
1320 Builder.SetInsertPoint(IP);
1322 // gc relocates will be generated later as if it were regular call
1327 // Take the name of the original value call if it had one.
1328 token->takeName(CS.getInstruction());
1330 // The GCResult is already inserted, we just need to find it
1332 Instruction *toReplace = CS.getInstruction();
1333 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1334 "only valid use before rewrite is gc.result");
1335 assert(!toReplace->hasOneUse() ||
1336 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1339 // Update the gc.result of the original statepoint (if any) to use the newly
1340 // inserted statepoint. This is safe to do here since the token can't be
1341 // considered a live reference.
1342 CS.getInstruction()->replaceAllUsesWith(token);
1344 result.StatepointToken = token;
1346 // Second, create a gc.relocate for every live variable
1347 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1351 struct name_ordering {
1354 bool operator()(name_ordering const &a, name_ordering const &b) {
1355 return -1 == a.derived->getName().compare(b.derived->getName());
1359 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1360 SmallVectorImpl<Value *> &livevec) {
1361 assert(basevec.size() == livevec.size());
1363 SmallVector<name_ordering, 64> temp;
1364 for (size_t i = 0; i < basevec.size(); i++) {
1366 v.base = basevec[i];
1367 v.derived = livevec[i];
1370 std::sort(temp.begin(), temp.end(), name_ordering());
1371 for (size_t i = 0; i < basevec.size(); i++) {
1372 basevec[i] = temp[i].base;
1373 livevec[i] = temp[i].derived;
1377 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1378 // which make the relocations happening at this safepoint explicit.
1380 // WARNING: Does not do any fixup to adjust users of the original live
1381 // values. That's the callers responsibility.
1383 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1384 PartiallyConstructedSafepointRecord &result) {
1385 auto liveset = result.liveset;
1386 auto PointerToBase = result.PointerToBase;
1388 // Convert to vector for efficient cross referencing.
1389 SmallVector<Value *, 64> basevec, livevec;
1390 livevec.reserve(liveset.size());
1391 basevec.reserve(liveset.size());
1392 for (Value *L : liveset) {
1393 livevec.push_back(L);
1394 assert(PointerToBase.count(L));
1395 Value *base = PointerToBase[L];
1396 basevec.push_back(base);
1398 assert(livevec.size() == basevec.size());
1400 // To make the output IR slightly more stable (for use in diffs), ensure a
1401 // fixed order of the values in the safepoint (by sorting the value name).
1402 // The order is otherwise meaningless.
1403 stablize_order(basevec, livevec);
1405 // Do the actual rewriting and delete the old statepoint
1406 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1407 CS.getInstruction()->eraseFromParent();
1410 // Helper function for the relocationViaAlloca.
1411 // It receives iterator to the statepoint gc relocates and emits store to the
1413 // location (via allocaMap) for the each one of them.
1414 // Add visited values into the visitedLiveValues set we will later use them
1415 // for sanity check.
1417 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1418 DenseMap<Value *, Value *> &AllocaMap,
1419 DenseSet<Value *> &VisitedLiveValues) {
1421 for (User *U : GCRelocs) {
1422 if (!isa<IntrinsicInst>(U))
1425 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1427 // We only care about relocates
1428 if (RelocatedValue->getIntrinsicID() !=
1429 Intrinsic::experimental_gc_relocate) {
1433 GCRelocateOperands RelocateOperands(RelocatedValue);
1434 Value *OriginalValue =
1435 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1436 assert(AllocaMap.count(OriginalValue));
1437 Value *Alloca = AllocaMap[OriginalValue];
1439 // Emit store into the related alloca
1440 // All gc_relocate are i8 addrspace(1)* typed, and it must be bitcasted to
1441 // the correct type according to alloca.
1442 assert(RelocatedValue->getNextNode() && "Should always have one since it's not a terminator");
1443 IRBuilder<> Builder(RelocatedValue->getNextNode());
1444 Value *CastedRelocatedValue =
1445 Builder.CreateBitCast(RelocatedValue, cast<AllocaInst>(Alloca)->getAllocatedType(),
1446 RelocatedValue->hasName() ? RelocatedValue->getName() + ".casted" : "");
1448 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1449 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1452 VisitedLiveValues.insert(OriginalValue);
1457 // Helper function for the "relocationViaAlloca". Similar to the
1458 // "insertRelocationStores" but works for rematerialized values.
1460 insertRematerializationStores(
1461 RematerializedValueMapTy RematerializedValues,
1462 DenseMap<Value *, Value *> &AllocaMap,
1463 DenseSet<Value *> &VisitedLiveValues) {
1465 for (auto RematerializedValuePair: RematerializedValues) {
1466 Instruction *RematerializedValue = RematerializedValuePair.first;
1467 Value *OriginalValue = RematerializedValuePair.second;
1469 assert(AllocaMap.count(OriginalValue) &&
1470 "Can not find alloca for rematerialized value");
1471 Value *Alloca = AllocaMap[OriginalValue];
1473 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1474 Store->insertAfter(RematerializedValue);
1477 VisitedLiveValues.insert(OriginalValue);
1482 /// do all the relocation update via allocas and mem2reg
1483 static void relocationViaAlloca(
1484 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1485 ArrayRef<struct PartiallyConstructedSafepointRecord> Records) {
1487 // record initial number of (static) allocas; we'll check we have the same
1488 // number when we get done.
1489 int InitialAllocaNum = 0;
1490 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1492 if (isa<AllocaInst>(*I))
1496 // TODO-PERF: change data structures, reserve
1497 DenseMap<Value *, Value *> AllocaMap;
1498 SmallVector<AllocaInst *, 200> PromotableAllocas;
1499 // Used later to chack that we have enough allocas to store all values
1500 std::size_t NumRematerializedValues = 0;
1501 PromotableAllocas.reserve(Live.size());
1503 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1504 // "PromotableAllocas"
1505 auto emitAllocaFor = [&](Value *LiveValue) {
1506 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1507 F.getEntryBlock().getFirstNonPHI());
1508 AllocaMap[LiveValue] = Alloca;
1509 PromotableAllocas.push_back(Alloca);
1512 // emit alloca for each live gc pointer
1513 for (unsigned i = 0; i < Live.size(); i++) {
1514 emitAllocaFor(Live[i]);
1517 // emit allocas for rematerialized values
1518 for (size_t i = 0; i < Records.size(); i++) {
1519 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1521 for (auto RematerializedValuePair : Info.RematerializedValues) {
1522 Value *OriginalValue = RematerializedValuePair.second;
1523 if (AllocaMap.count(OriginalValue) != 0)
1526 emitAllocaFor(OriginalValue);
1527 ++NumRematerializedValues;
1531 // The next two loops are part of the same conceptual operation. We need to
1532 // insert a store to the alloca after the original def and at each
1533 // redefinition. We need to insert a load before each use. These are split
1534 // into distinct loops for performance reasons.
1536 // update gc pointer after each statepoint
1537 // either store a relocated value or null (if no relocated value found for
1538 // this gc pointer and it is not a gc_result)
1539 // this must happen before we update the statepoint with load of alloca
1540 // otherwise we lose the link between statepoint and old def
1541 for (size_t i = 0; i < Records.size(); i++) {
1542 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1543 Value *Statepoint = Info.StatepointToken;
1545 // This will be used for consistency check
1546 DenseSet<Value *> VisitedLiveValues;
1548 // Insert stores for normal statepoint gc relocates
1549 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1551 // In case if it was invoke statepoint
1552 // we will insert stores for exceptional path gc relocates.
1553 if (isa<InvokeInst>(Statepoint)) {
1554 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1558 // Do similar thing with rematerialized values
1559 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1562 if (ClobberNonLive) {
1563 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1564 // the gc.statepoint. This will turn some subtle GC problems into
1565 // slightly easier to debug SEGVs. Note that on large IR files with
1566 // lots of gc.statepoints this is extremely costly both memory and time
1568 SmallVector<AllocaInst *, 64> ToClobber;
1569 for (auto Pair : AllocaMap) {
1570 Value *Def = Pair.first;
1571 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1573 // This value was relocated
1574 if (VisitedLiveValues.count(Def)) {
1577 ToClobber.push_back(Alloca);
1580 auto InsertClobbersAt = [&](Instruction *IP) {
1581 for (auto *AI : ToClobber) {
1582 auto AIType = cast<PointerType>(AI->getType());
1583 auto PT = cast<PointerType>(AIType->getElementType());
1584 Constant *CPN = ConstantPointerNull::get(PT);
1585 StoreInst *Store = new StoreInst(CPN, AI);
1586 Store->insertBefore(IP);
1590 // Insert the clobbering stores. These may get intermixed with the
1591 // gc.results and gc.relocates, but that's fine.
1592 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1593 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1594 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1596 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1598 InsertClobbersAt(Next);
1602 // update use with load allocas and add store for gc_relocated
1603 for (auto Pair : AllocaMap) {
1604 Value *Def = Pair.first;
1605 Value *Alloca = Pair.second;
1607 // we pre-record the uses of allocas so that we dont have to worry about
1609 // that change the user information.
1610 SmallVector<Instruction *, 20> Uses;
1611 // PERF: trade a linear scan for repeated reallocation
1612 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1613 for (User *U : Def->users()) {
1614 if (!isa<ConstantExpr>(U)) {
1615 // If the def has a ConstantExpr use, then the def is either a
1616 // ConstantExpr use itself or null. In either case
1617 // (recursively in the first, directly in the second), the oop
1618 // it is ultimately dependent on is null and this particular
1619 // use does not need to be fixed up.
1620 Uses.push_back(cast<Instruction>(U));
1624 std::sort(Uses.begin(), Uses.end());
1625 auto Last = std::unique(Uses.begin(), Uses.end());
1626 Uses.erase(Last, Uses.end());
1628 for (Instruction *Use : Uses) {
1629 if (isa<PHINode>(Use)) {
1630 PHINode *Phi = cast<PHINode>(Use);
1631 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1632 if (Def == Phi->getIncomingValue(i)) {
1633 LoadInst *Load = new LoadInst(
1634 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1635 Phi->setIncomingValue(i, Load);
1639 LoadInst *Load = new LoadInst(Alloca, "", Use);
1640 Use->replaceUsesOfWith(Def, Load);
1644 // emit store for the initial gc value
1645 // store must be inserted after load, otherwise store will be in alloca's
1646 // use list and an extra load will be inserted before it
1647 StoreInst *Store = new StoreInst(Def, Alloca);
1648 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1649 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1650 // InvokeInst is a TerminatorInst so the store need to be inserted
1651 // into its normal destination block.
1652 BasicBlock *NormalDest = Invoke->getNormalDest();
1653 Store->insertBefore(NormalDest->getFirstNonPHI());
1655 assert(!Inst->isTerminator() &&
1656 "The only TerminatorInst that can produce a value is "
1657 "InvokeInst which is handled above.");
1658 Store->insertAfter(Inst);
1661 assert(isa<Argument>(Def));
1662 Store->insertAfter(cast<Instruction>(Alloca));
1666 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1667 "we must have the same allocas with lives");
1668 if (!PromotableAllocas.empty()) {
1669 // apply mem2reg to promote alloca to SSA
1670 PromoteMemToReg(PromotableAllocas, DT);
1674 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1676 if (isa<AllocaInst>(*I))
1678 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1682 /// Implement a unique function which doesn't require we sort the input
1683 /// vector. Doing so has the effect of changing the output of a couple of
1684 /// tests in ways which make them less useful in testing fused safepoints.
1685 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1686 SmallSet<T, 8> Seen;
1687 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1688 return !Seen.insert(V).second;
1692 /// Insert holders so that each Value is obviously live through the entire
1693 /// lifetime of the call.
1694 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1695 SmallVectorImpl<CallInst *> &Holders) {
1697 // No values to hold live, might as well not insert the empty holder
1700 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1701 // Use a dummy vararg function to actually hold the values live
1702 Function *Func = cast<Function>(M->getOrInsertFunction(
1703 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1705 // For call safepoints insert dummy calls right after safepoint
1706 BasicBlock::iterator Next(CS.getInstruction());
1708 Holders.push_back(CallInst::Create(Func, Values, "", Next));
1711 // For invoke safepooints insert dummy calls both in normal and
1712 // exceptional destination blocks
1713 auto *II = cast<InvokeInst>(CS.getInstruction());
1714 Holders.push_back(CallInst::Create(
1715 Func, Values, "", II->getNormalDest()->getFirstInsertionPt()));
1716 Holders.push_back(CallInst::Create(
1717 Func, Values, "", II->getUnwindDest()->getFirstInsertionPt()));
1720 static void findLiveReferences(
1721 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1722 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1723 GCPtrLivenessData OriginalLivenessData;
1724 computeLiveInValues(DT, F, OriginalLivenessData);
1725 for (size_t i = 0; i < records.size(); i++) {
1726 struct PartiallyConstructedSafepointRecord &info = records[i];
1727 const CallSite &CS = toUpdate[i];
1728 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1732 /// Remove any vector of pointers from the liveset by scalarizing them over the
1733 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1734 /// would be preferrable to include the vector in the statepoint itself, but
1735 /// the lowering code currently does not handle that. Extending it would be
1736 /// slightly non-trivial since it requires a format change. Given how rare
1737 /// such cases are (for the moment?) scalarizing is an acceptable comprimise.
1738 static void splitVectorValues(Instruction *StatepointInst,
1739 StatepointLiveSetTy &LiveSet,
1740 DenseMap<Value *, Value *>& PointerToBase,
1741 DominatorTree &DT) {
1742 SmallVector<Value *, 16> ToSplit;
1743 for (Value *V : LiveSet)
1744 if (isa<VectorType>(V->getType()))
1745 ToSplit.push_back(V);
1747 if (ToSplit.empty())
1750 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1752 Function &F = *(StatepointInst->getParent()->getParent());
1754 DenseMap<Value *, AllocaInst *> AllocaMap;
1755 // First is normal return, second is exceptional return (invoke only)
1756 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1757 for (Value *V : ToSplit) {
1758 AllocaInst *Alloca =
1759 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1760 AllocaMap[V] = Alloca;
1762 VectorType *VT = cast<VectorType>(V->getType());
1763 IRBuilder<> Builder(StatepointInst);
1764 SmallVector<Value *, 16> Elements;
1765 for (unsigned i = 0; i < VT->getNumElements(); i++)
1766 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1767 ElementMapping[V] = Elements;
1769 auto InsertVectorReform = [&](Instruction *IP) {
1770 Builder.SetInsertPoint(IP);
1771 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1772 Value *ResultVec = UndefValue::get(VT);
1773 for (unsigned i = 0; i < VT->getNumElements(); i++)
1774 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1775 Builder.getInt32(i));
1779 if (isa<CallInst>(StatepointInst)) {
1780 BasicBlock::iterator Next(StatepointInst);
1782 Instruction *IP = &*(Next);
1783 Replacements[V].first = InsertVectorReform(IP);
1784 Replacements[V].second = nullptr;
1786 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1787 // We've already normalized - check that we don't have shared destination
1789 BasicBlock *NormalDest = Invoke->getNormalDest();
1790 assert(!isa<PHINode>(NormalDest->begin()));
1791 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1792 assert(!isa<PHINode>(UnwindDest->begin()));
1793 // Insert insert element sequences in both successors
1794 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1795 Replacements[V].first = InsertVectorReform(IP);
1796 IP = &*(UnwindDest->getFirstInsertionPt());
1797 Replacements[V].second = InsertVectorReform(IP);
1801 for (Value *V : ToSplit) {
1802 AllocaInst *Alloca = AllocaMap[V];
1804 // Capture all users before we start mutating use lists
1805 SmallVector<Instruction *, 16> Users;
1806 for (User *U : V->users())
1807 Users.push_back(cast<Instruction>(U));
1809 for (Instruction *I : Users) {
1810 if (auto Phi = dyn_cast<PHINode>(I)) {
1811 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1812 if (V == Phi->getIncomingValue(i)) {
1813 LoadInst *Load = new LoadInst(
1814 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1815 Phi->setIncomingValue(i, Load);
1818 LoadInst *Load = new LoadInst(Alloca, "", I);
1819 I->replaceUsesOfWith(V, Load);
1823 // Store the original value and the replacement value into the alloca
1824 StoreInst *Store = new StoreInst(V, Alloca);
1825 if (auto I = dyn_cast<Instruction>(V))
1826 Store->insertAfter(I);
1828 Store->insertAfter(Alloca);
1830 // Normal return for invoke, or call return
1831 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1832 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1833 // Unwind return for invoke only
1834 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1836 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1839 // apply mem2reg to promote alloca to SSA
1840 SmallVector<AllocaInst *, 16> Allocas;
1841 for (Value *V : ToSplit)
1842 Allocas.push_back(AllocaMap[V]);
1843 PromoteMemToReg(Allocas, DT);
1845 // Update our tracking of live pointers and base mappings to account for the
1846 // changes we just made.
1847 for (Value *V : ToSplit) {
1848 auto &Elements = ElementMapping[V];
1851 LiveSet.insert(Elements.begin(), Elements.end());
1852 // We need to update the base mapping as well.
1853 assert(PointerToBase.count(V));
1854 Value *OldBase = PointerToBase[V];
1855 auto &BaseElements = ElementMapping[OldBase];
1856 PointerToBase.erase(V);
1857 assert(Elements.size() == BaseElements.size());
1858 for (unsigned i = 0; i < Elements.size(); i++) {
1859 Value *Elem = Elements[i];
1860 PointerToBase[Elem] = BaseElements[i];
1865 // Helper function for the "rematerializeLiveValues". It walks use chain
1866 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
1867 // values are visited (currently it is GEP's and casts). Returns true if it
1868 // sucessfully reached "BaseValue" and false otherwise.
1869 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
1871 static bool findRematerializableChainToBasePointer(
1872 SmallVectorImpl<Instruction*> &ChainToBase,
1873 Value *CurrentValue, Value *BaseValue) {
1875 // We have found a base value
1876 if (CurrentValue == BaseValue) {
1880 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
1881 ChainToBase.push_back(GEP);
1882 return findRematerializableChainToBasePointer(ChainToBase,
1883 GEP->getPointerOperand(),
1887 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
1888 Value *Def = CI->stripPointerCasts();
1890 // This two checks are basically similar. First one is here for the
1891 // consistency with findBasePointers logic.
1892 assert(!isa<CastInst>(Def) && "not a pointer cast found");
1893 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
1896 ChainToBase.push_back(CI);
1897 return findRematerializableChainToBasePointer(ChainToBase, Def, BaseValue);
1900 // Not supported instruction in the chain
1904 // Helper function for the "rematerializeLiveValues". Compute cost of the use
1905 // chain we are going to rematerialize.
1907 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
1908 TargetTransformInfo &TTI) {
1911 for (Instruction *Instr : Chain) {
1912 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
1913 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
1914 "non noop cast is found during rematerialization");
1916 Type *SrcTy = CI->getOperand(0)->getType();
1917 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
1919 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
1920 // Cost of the address calculation
1921 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
1922 Cost += TTI.getAddressComputationCost(ValTy);
1924 // And cost of the GEP itself
1925 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
1926 // allowed for the external usage)
1927 if (!GEP->hasAllConstantIndices())
1931 llvm_unreachable("unsupported instruciton type during rematerialization");
1938 // From the statepoint liveset pick values that are cheaper to recompute then to
1939 // relocate. Remove this values from the liveset, rematerialize them after
1940 // statepoint and record them in "Info" structure. Note that similar to
1941 // relocated values we don't do any user adjustments here.
1942 static void rematerializeLiveValues(CallSite CS,
1943 PartiallyConstructedSafepointRecord &Info,
1944 TargetTransformInfo &TTI) {
1945 const unsigned int ChainLengthThreshold = 10;
1947 // Record values we are going to delete from this statepoint live set.
1948 // We can not di this in following loop due to iterator invalidation.
1949 SmallVector<Value *, 32> LiveValuesToBeDeleted;
1951 for (Value *LiveValue: Info.liveset) {
1952 // For each live pointer find it's defining chain
1953 SmallVector<Instruction *, 3> ChainToBase;
1954 assert(Info.PointerToBase.count(LiveValue));
1956 findRematerializableChainToBasePointer(ChainToBase,
1958 Info.PointerToBase[LiveValue]);
1959 // Nothing to do, or chain is too long
1961 ChainToBase.size() == 0 ||
1962 ChainToBase.size() > ChainLengthThreshold)
1965 // Compute cost of this chain
1966 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
1967 // TODO: We can also account for cases when we will be able to remove some
1968 // of the rematerialized values by later optimization passes. I.e if
1969 // we rematerialized several intersecting chains. Or if original values
1970 // don't have any uses besides this statepoint.
1972 // For invokes we need to rematerialize each chain twice - for normal and
1973 // for unwind basic blocks. Model this by multiplying cost by two.
1974 if (CS.isInvoke()) {
1977 // If it's too expensive - skip it
1978 if (Cost >= RematerializationThreshold)
1981 // Remove value from the live set
1982 LiveValuesToBeDeleted.push_back(LiveValue);
1984 // Clone instructions and record them inside "Info" structure
1986 // Walk backwards to visit top-most instructions first
1987 std::reverse(ChainToBase.begin(), ChainToBase.end());
1989 // Utility function which clones all instructions from "ChainToBase"
1990 // and inserts them before "InsertBefore". Returns rematerialized value
1991 // which should be used after statepoint.
1992 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
1993 Instruction *LastClonedValue = nullptr;
1994 Instruction *LastValue = nullptr;
1995 for (Instruction *Instr: ChainToBase) {
1996 // Only GEP's and casts are suported as we need to be careful to not
1997 // introduce any new uses of pointers not in the liveset.
1998 // Note that it's fine to introduce new uses of pointers which were
1999 // otherwise not used after this statepoint.
2000 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2002 Instruction *ClonedValue = Instr->clone();
2003 ClonedValue->insertBefore(InsertBefore);
2004 ClonedValue->setName(Instr->getName() + ".remat");
2006 // If it is not first instruction in the chain then it uses previously
2007 // cloned value. We should update it to use cloned value.
2008 if (LastClonedValue) {
2010 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2012 // Assert that cloned instruction does not use any instructions from
2013 // this chain other than LastClonedValue
2014 for (auto OpValue : ClonedValue->operand_values()) {
2015 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2016 ChainToBase.end() &&
2017 "incorrect use in rematerialization chain");
2022 LastClonedValue = ClonedValue;
2025 assert(LastClonedValue);
2026 return LastClonedValue;
2029 // Different cases for calls and invokes. For invokes we need to clone
2030 // instructions both on normal and unwind path.
2032 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2033 assert(InsertBefore);
2034 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2035 Info.RematerializedValues[RematerializedValue] = LiveValue;
2037 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2039 Instruction *NormalInsertBefore =
2040 Invoke->getNormalDest()->getFirstInsertionPt();
2041 Instruction *UnwindInsertBefore =
2042 Invoke->getUnwindDest()->getFirstInsertionPt();
2044 Instruction *NormalRematerializedValue =
2045 rematerializeChain(NormalInsertBefore);
2046 Instruction *UnwindRematerializedValue =
2047 rematerializeChain(UnwindInsertBefore);
2049 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2050 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2054 // Remove rematerializaed values from the live set
2055 for (auto LiveValue: LiveValuesToBeDeleted) {
2056 Info.liveset.erase(LiveValue);
2060 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
2061 SmallVectorImpl<CallSite> &toUpdate) {
2063 // sanity check the input
2064 std::set<CallSite> uniqued;
2065 uniqued.insert(toUpdate.begin(), toUpdate.end());
2066 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
2068 for (size_t i = 0; i < toUpdate.size(); i++) {
2069 CallSite &CS = toUpdate[i];
2070 assert(CS.getInstruction()->getParent()->getParent() == &F);
2071 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
2075 // When inserting gc.relocates for invokes, we need to be able to insert at
2076 // the top of the successor blocks. See the comment on
2077 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2078 // may restructure the CFG.
2079 for (CallSite CS : toUpdate) {
2082 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
2083 normalizeForInvokeSafepoint(invoke->getNormalDest(), invoke->getParent(),
2085 normalizeForInvokeSafepoint(invoke->getUnwindDest(), invoke->getParent(),
2089 // A list of dummy calls added to the IR to keep various values obviously
2090 // live in the IR. We'll remove all of these when done.
2091 SmallVector<CallInst *, 64> holders;
2093 // Insert a dummy call with all of the arguments to the vm_state we'll need
2094 // for the actual safepoint insertion. This ensures reference arguments in
2095 // the deopt argument list are considered live through the safepoint (and
2096 // thus makes sure they get relocated.)
2097 for (size_t i = 0; i < toUpdate.size(); i++) {
2098 CallSite &CS = toUpdate[i];
2099 Statepoint StatepointCS(CS);
2101 SmallVector<Value *, 64> DeoptValues;
2102 for (Use &U : StatepointCS.vm_state_args()) {
2103 Value *Arg = cast<Value>(&U);
2104 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2105 "support for FCA unimplemented");
2106 if (isHandledGCPointerType(Arg->getType()))
2107 DeoptValues.push_back(Arg);
2109 insertUseHolderAfter(CS, DeoptValues, holders);
2112 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
2113 records.reserve(toUpdate.size());
2114 for (size_t i = 0; i < toUpdate.size(); i++) {
2115 struct PartiallyConstructedSafepointRecord info;
2116 records.push_back(info);
2118 assert(records.size() == toUpdate.size());
2120 // A) Identify all gc pointers which are staticly live at the given call
2122 findLiveReferences(F, DT, P, toUpdate, records);
2124 // B) Find the base pointers for each live pointer
2125 /* scope for caching */ {
2126 // Cache the 'defining value' relation used in the computation and
2127 // insertion of base phis and selects. This ensures that we don't insert
2128 // large numbers of duplicate base_phis.
2129 DefiningValueMapTy DVCache;
2131 for (size_t i = 0; i < records.size(); i++) {
2132 struct PartiallyConstructedSafepointRecord &info = records[i];
2133 CallSite &CS = toUpdate[i];
2134 findBasePointers(DT, DVCache, CS, info);
2136 } // end of cache scope
2138 // The base phi insertion logic (for any safepoint) may have inserted new
2139 // instructions which are now live at some safepoint. The simplest such
2142 // phi a <-- will be a new base_phi here
2143 // safepoint 1 <-- that needs to be live here
2147 // We insert some dummy calls after each safepoint to definitely hold live
2148 // the base pointers which were identified for that safepoint. We'll then
2149 // ask liveness for _every_ base inserted to see what is now live. Then we
2150 // remove the dummy calls.
2151 holders.reserve(holders.size() + records.size());
2152 for (size_t i = 0; i < records.size(); i++) {
2153 struct PartiallyConstructedSafepointRecord &info = records[i];
2154 CallSite &CS = toUpdate[i];
2156 SmallVector<Value *, 128> Bases;
2157 for (auto Pair : info.PointerToBase) {
2158 Bases.push_back(Pair.second);
2160 insertUseHolderAfter(CS, Bases, holders);
2163 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2164 // need to rerun liveness. We may *also* have inserted new defs, but that's
2165 // not the key issue.
2166 recomputeLiveInValues(F, DT, P, toUpdate, records);
2168 if (PrintBasePointers) {
2169 for (size_t i = 0; i < records.size(); i++) {
2170 struct PartiallyConstructedSafepointRecord &info = records[i];
2171 errs() << "Base Pairs: (w/Relocation)\n";
2172 for (auto Pair : info.PointerToBase) {
2173 errs() << " derived %" << Pair.first->getName() << " base %"
2174 << Pair.second->getName() << "\n";
2178 for (size_t i = 0; i < holders.size(); i++) {
2179 holders[i]->eraseFromParent();
2180 holders[i] = nullptr;
2184 // Do a limited scalarization of any live at safepoint vector values which
2185 // contain pointers. This enables this pass to run after vectorization at
2186 // the cost of some possible performance loss. TODO: it would be nice to
2187 // natively support vectors all the way through the backend so we don't need
2188 // to scalarize here.
2189 for (size_t i = 0; i < records.size(); i++) {
2190 struct PartiallyConstructedSafepointRecord &info = records[i];
2191 Instruction *statepoint = toUpdate[i].getInstruction();
2192 splitVectorValues(cast<Instruction>(statepoint), info.liveset,
2193 info.PointerToBase, DT);
2196 // In order to reduce live set of statepoint we might choose to rematerialize
2197 // some values instead of relocating them. This is purelly an optimization and
2198 // does not influence correctness.
2199 TargetTransformInfo &TTI =
2200 P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2202 for (size_t i = 0; i < records.size(); i++) {
2203 struct PartiallyConstructedSafepointRecord &info = records[i];
2204 CallSite &CS = toUpdate[i];
2206 rematerializeLiveValues(CS, info, TTI);
2209 // Now run through and replace the existing statepoints with new ones with
2210 // the live variables listed. We do not yet update uses of the values being
2211 // relocated. We have references to live variables that need to
2212 // survive to the last iteration of this loop. (By construction, the
2213 // previous statepoint can not be a live variable, thus we can and remove
2214 // the old statepoint calls as we go.)
2215 for (size_t i = 0; i < records.size(); i++) {
2216 struct PartiallyConstructedSafepointRecord &info = records[i];
2217 CallSite &CS = toUpdate[i];
2218 makeStatepointExplicit(DT, CS, P, info);
2220 toUpdate.clear(); // prevent accident use of invalid CallSites
2222 // Do all the fixups of the original live variables to their relocated selves
2223 SmallVector<Value *, 128> live;
2224 for (size_t i = 0; i < records.size(); i++) {
2225 struct PartiallyConstructedSafepointRecord &info = records[i];
2226 // We can't simply save the live set from the original insertion. One of
2227 // the live values might be the result of a call which needs a safepoint.
2228 // That Value* no longer exists and we need to use the new gc_result.
2229 // Thankfully, the liveset is embedded in the statepoint (and updated), so
2230 // we just grab that.
2231 Statepoint statepoint(info.StatepointToken);
2232 live.insert(live.end(), statepoint.gc_args_begin(),
2233 statepoint.gc_args_end());
2235 // Do some basic sanity checks on our liveness results before performing
2236 // relocation. Relocation can and will turn mistakes in liveness results
2237 // into non-sensical code which is must harder to debug.
2238 // TODO: It would be nice to test consistency as well
2239 assert(DT.isReachableFromEntry(info.StatepointToken->getParent()) &&
2240 "statepoint must be reachable or liveness is meaningless");
2241 for (Value *V : statepoint.gc_args()) {
2242 if (!isa<Instruction>(V))
2243 // Non-instruction values trivial dominate all possible uses
2245 auto LiveInst = cast<Instruction>(V);
2246 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2247 "unreachable values should never be live");
2248 assert(DT.dominates(LiveInst, info.StatepointToken) &&
2249 "basic SSA liveness expectation violated by liveness analysis");
2253 unique_unsorted(live);
2257 for (auto ptr : live) {
2258 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
2262 relocationViaAlloca(F, DT, live, records);
2263 return !records.empty();
2266 // Handles both return values and arguments for Functions and CallSites.
2267 template <typename AttrHolder>
2268 static void RemoveDerefAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2271 if (AH.getDereferenceableBytes(Index))
2272 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2273 AH.getDereferenceableBytes(Index)));
2274 if (AH.getDereferenceableOrNullBytes(Index))
2275 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2276 AH.getDereferenceableOrNullBytes(Index)));
2279 AH.setAttributes(AH.getAttributes().removeAttributes(
2280 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2284 RewriteStatepointsForGC::stripDereferenceabilityInfoFromPrototype(Function &F) {
2285 LLVMContext &Ctx = F.getContext();
2287 for (Argument &A : F.args())
2288 if (isa<PointerType>(A.getType()))
2289 RemoveDerefAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2291 if (isa<PointerType>(F.getReturnType()))
2292 RemoveDerefAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2295 void RewriteStatepointsForGC::stripDereferenceabilityInfoFromBody(Function &F) {
2299 LLVMContext &Ctx = F.getContext();
2300 MDBuilder Builder(Ctx);
2302 for (Instruction &I : inst_range(F)) {
2303 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2304 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2305 bool IsImmutableTBAA =
2306 MD->getNumOperands() == 4 &&
2307 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2309 if (!IsImmutableTBAA)
2310 continue; // no work to do, MD_tbaa is already marked mutable
2312 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2313 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2315 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2317 MDNode *MutableTBAA =
2318 Builder.createTBAAStructTagNode(Base, Access, Offset);
2319 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2322 if (CallSite CS = CallSite(&I)) {
2323 for (int i = 0, e = CS.arg_size(); i != e; i++)
2324 if (isa<PointerType>(CS.getArgument(i)->getType()))
2325 RemoveDerefAttrAtIndex(Ctx, CS, i + 1);
2326 if (isa<PointerType>(CS.getType()))
2327 RemoveDerefAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2332 /// Returns true if this function should be rewritten by this pass. The main
2333 /// point of this function is as an extension point for custom logic.
2334 static bool shouldRewriteStatepointsIn(Function &F) {
2335 // TODO: This should check the GCStrategy
2337 const char *FunctionGCName = F.getGC();
2338 const StringRef StatepointExampleName("statepoint-example");
2339 const StringRef CoreCLRName("coreclr");
2340 return (StatepointExampleName == FunctionGCName) ||
2341 (CoreCLRName == FunctionGCName);
2346 void RewriteStatepointsForGC::stripDereferenceabilityInfo(Module &M) {
2348 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2352 for (Function &F : M)
2353 stripDereferenceabilityInfoFromPrototype(F);
2355 for (Function &F : M)
2356 stripDereferenceabilityInfoFromBody(F);
2359 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2360 // Nothing to do for declarations.
2361 if (F.isDeclaration() || F.empty())
2364 // Policy choice says not to rewrite - the most common reason is that we're
2365 // compiling code without a GCStrategy.
2366 if (!shouldRewriteStatepointsIn(F))
2369 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2371 // Gather all the statepoints which need rewritten. Be careful to only
2372 // consider those in reachable code since we need to ask dominance queries
2373 // when rewriting. We'll delete the unreachable ones in a moment.
2374 SmallVector<CallSite, 64> ParsePointNeeded;
2375 bool HasUnreachableStatepoint = false;
2376 for (Instruction &I : inst_range(F)) {
2377 // TODO: only the ones with the flag set!
2378 if (isStatepoint(I)) {
2379 if (DT.isReachableFromEntry(I.getParent()))
2380 ParsePointNeeded.push_back(CallSite(&I));
2382 HasUnreachableStatepoint = true;
2386 bool MadeChange = false;
2388 // Delete any unreachable statepoints so that we don't have unrewritten
2389 // statepoints surviving this pass. This makes testing easier and the
2390 // resulting IR less confusing to human readers. Rather than be fancy, we
2391 // just reuse a utility function which removes the unreachable blocks.
2392 if (HasUnreachableStatepoint)
2393 MadeChange |= removeUnreachableBlocks(F);
2395 // Return early if no work to do.
2396 if (ParsePointNeeded.empty())
2399 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2400 // These are created by LCSSA. They have the effect of increasing the size
2401 // of liveness sets for no good reason. It may be harder to do this post
2402 // insertion since relocations and base phis can confuse things.
2403 for (BasicBlock &BB : F)
2404 if (BB.getUniquePredecessor()) {
2406 FoldSingleEntryPHINodes(&BB);
2409 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
2413 // liveness computation via standard dataflow
2414 // -------------------------------------------------------------------
2416 // TODO: Consider using bitvectors for liveness, the set of potentially
2417 // interesting values should be small and easy to pre-compute.
2419 /// Compute the live-in set for the location rbegin starting from
2420 /// the live-out set of the basic block
2421 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2422 BasicBlock::reverse_iterator rend,
2423 DenseSet<Value *> &LiveTmp) {
2425 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2426 Instruction *I = &*ritr;
2428 // KILL/Def - Remove this definition from LiveIn
2431 // Don't consider *uses* in PHI nodes, we handle their contribution to
2432 // predecessor blocks when we seed the LiveOut sets
2433 if (isa<PHINode>(I))
2436 // USE - Add to the LiveIn set for this instruction
2437 for (Value *V : I->operands()) {
2438 assert(!isUnhandledGCPointerType(V->getType()) &&
2439 "support for FCA unimplemented");
2440 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2441 // The choice to exclude all things constant here is slightly subtle.
2442 // There are two idependent reasons:
2443 // - We assume that things which are constant (from LLVM's definition)
2444 // do not move at runtime. For example, the address of a global
2445 // variable is fixed, even though it's contents may not be.
2446 // - Second, we can't disallow arbitrary inttoptr constants even
2447 // if the language frontend does. Optimization passes are free to
2448 // locally exploit facts without respect to global reachability. This
2449 // can create sections of code which are dynamically unreachable and
2450 // contain just about anything. (see constants.ll in tests)
2457 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2459 for (BasicBlock *Succ : successors(BB)) {
2460 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2461 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2462 PHINode *Phi = cast<PHINode>(&*I);
2463 Value *V = Phi->getIncomingValueForBlock(BB);
2464 assert(!isUnhandledGCPointerType(V->getType()) &&
2465 "support for FCA unimplemented");
2466 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2473 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2474 DenseSet<Value *> KillSet;
2475 for (Instruction &I : *BB)
2476 if (isHandledGCPointerType(I.getType()))
2482 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2483 /// sanity check for the liveness computation.
2484 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2485 TerminatorInst *TI, bool TermOkay = false) {
2486 for (Value *V : Live) {
2487 if (auto *I = dyn_cast<Instruction>(V)) {
2488 // The terminator can be a member of the LiveOut set. LLVM's definition
2489 // of instruction dominance states that V does not dominate itself. As
2490 // such, we need to special case this to allow it.
2491 if (TermOkay && TI == I)
2493 assert(DT.dominates(I, TI) &&
2494 "basic SSA liveness expectation violated by liveness analysis");
2499 /// Check that all the liveness sets used during the computation of liveness
2500 /// obey basic SSA properties. This is useful for finding cases where we miss
2502 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2504 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2505 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2506 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2510 static void computeLiveInValues(DominatorTree &DT, Function &F,
2511 GCPtrLivenessData &Data) {
2513 SmallSetVector<BasicBlock *, 200> Worklist;
2514 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2515 // We use a SetVector so that we don't have duplicates in the worklist.
2516 Worklist.insert(pred_begin(BB), pred_end(BB));
2518 auto NextItem = [&]() {
2519 BasicBlock *BB = Worklist.back();
2520 Worklist.pop_back();
2524 // Seed the liveness for each individual block
2525 for (BasicBlock &BB : F) {
2526 Data.KillSet[&BB] = computeKillSet(&BB);
2527 Data.LiveSet[&BB].clear();
2528 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2531 for (Value *Kill : Data.KillSet[&BB])
2532 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2535 Data.LiveOut[&BB] = DenseSet<Value *>();
2536 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2537 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2538 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2539 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2540 if (!Data.LiveIn[&BB].empty())
2541 AddPredsToWorklist(&BB);
2544 // Propagate that liveness until stable
2545 while (!Worklist.empty()) {
2546 BasicBlock *BB = NextItem();
2548 // Compute our new liveout set, then exit early if it hasn't changed
2549 // despite the contribution of our successor.
2550 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2551 const auto OldLiveOutSize = LiveOut.size();
2552 for (BasicBlock *Succ : successors(BB)) {
2553 assert(Data.LiveIn.count(Succ));
2554 set_union(LiveOut, Data.LiveIn[Succ]);
2556 // assert OutLiveOut is a subset of LiveOut
2557 if (OldLiveOutSize == LiveOut.size()) {
2558 // If the sets are the same size, then we didn't actually add anything
2559 // when unioning our successors LiveIn Thus, the LiveIn of this block
2563 Data.LiveOut[BB] = LiveOut;
2565 // Apply the effects of this basic block
2566 DenseSet<Value *> LiveTmp = LiveOut;
2567 set_union(LiveTmp, Data.LiveSet[BB]);
2568 set_subtract(LiveTmp, Data.KillSet[BB]);
2570 assert(Data.LiveIn.count(BB));
2571 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2572 // assert: OldLiveIn is a subset of LiveTmp
2573 if (OldLiveIn.size() != LiveTmp.size()) {
2574 Data.LiveIn[BB] = LiveTmp;
2575 AddPredsToWorklist(BB);
2577 } // while( !worklist.empty() )
2580 // Sanity check our ouput against SSA properties. This helps catch any
2581 // missing kills during the above iteration.
2582 for (BasicBlock &BB : F) {
2583 checkBasicSSA(DT, Data, BB);
2588 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2589 StatepointLiveSetTy &Out) {
2591 BasicBlock *BB = Inst->getParent();
2593 // Note: The copy is intentional and required
2594 assert(Data.LiveOut.count(BB));
2595 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2597 // We want to handle the statepoint itself oddly. It's
2598 // call result is not live (normal), nor are it's arguments
2599 // (unless they're used again later). This adjustment is
2600 // specifically what we need to relocate
2601 BasicBlock::reverse_iterator rend(Inst);
2602 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2603 LiveOut.erase(Inst);
2604 Out.insert(LiveOut.begin(), LiveOut.end());
2607 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2609 PartiallyConstructedSafepointRecord &Info) {
2610 Instruction *Inst = CS.getInstruction();
2611 StatepointLiveSetTy Updated;
2612 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2615 DenseSet<Value *> Bases;
2616 for (auto KVPair : Info.PointerToBase) {
2617 Bases.insert(KVPair.second);
2620 // We may have base pointers which are now live that weren't before. We need
2621 // to update the PointerToBase structure to reflect this.
2622 for (auto V : Updated)
2623 if (!Info.PointerToBase.count(V)) {
2624 assert(Bases.count(V) && "can't find base for unexpected live value");
2625 Info.PointerToBase[V] = V;
2630 for (auto V : Updated) {
2631 assert(Info.PointerToBase.count(V) &&
2632 "must be able to find base for live value");
2636 // Remove any stale base mappings - this can happen since our liveness is
2637 // more precise then the one inherent in the base pointer analysis
2638 DenseSet<Value *> ToErase;
2639 for (auto KVPair : Info.PointerToBase)
2640 if (!Updated.count(KVPair.first))
2641 ToErase.insert(KVPair.first);
2642 for (auto V : ToErase)
2643 Info.PointerToBase.erase(V);
2646 for (auto KVPair : Info.PointerToBase)
2647 assert(Updated.count(KVPair.first) && "record for non-live value");
2650 Info.liveset = Updated;