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);
550 assert(Cache[I] != nullptr);
553 dbgs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
559 /// Return a base pointer for this value if known. Otherwise, return it's
560 /// base defining value.
561 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
562 Value *Def = findBaseDefiningValueCached(I, Cache);
563 auto Found = Cache.find(Def);
564 if (Found != Cache.end()) {
565 // Either a base-of relation, or a self reference. Caller must check.
566 return Found->second;
568 // Only a BDV available
572 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
573 /// is it known to be a base pointer? Or do we need to continue searching.
574 static bool isKnownBaseResult(Value *V) {
575 if (!isa<PHINode>(V) && !isa<SelectInst>(V)) {
576 // no recursion possible
579 if (isa<Instruction>(V) &&
580 cast<Instruction>(V)->getMetadata("is_base_value")) {
581 // This is a previously inserted base phi or select. We know
582 // that this is a base value.
586 // We need to keep searching
590 // TODO: find a better name for this
594 enum Status { Unknown, Base, Conflict };
596 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
597 assert(status != Base || b);
599 PhiState(Value *b) : status(Base), base(b) {}
600 PhiState() : status(Unknown), base(nullptr) {}
602 Status getStatus() const { return status; }
603 Value *getBase() const { return base; }
605 bool isBase() const { return getStatus() == Base; }
606 bool isUnknown() const { return getStatus() == Unknown; }
607 bool isConflict() const { return getStatus() == Conflict; }
609 bool operator==(const PhiState &other) const {
610 return base == other.base && status == other.status;
613 bool operator!=(const PhiState &other) const { return !(*this == other); }
616 errs() << status << " (" << base << " - "
617 << (base ? base->getName() : "nullptr") << "): ";
622 Value *base; // non null only if status == base
625 typedef DenseMap<Value *, PhiState> ConflictStateMapTy;
626 // Values of type PhiState form a lattice, and this is a helper
627 // class that implementes the meet operation. The meat of the meet
628 // operation is implemented in MeetPhiStates::pureMeet
629 class MeetPhiStates {
631 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
632 explicit MeetPhiStates(const ConflictStateMapTy &phiStates)
633 : phiStates(phiStates) {}
635 // Destructively meet the current result with the base V. V can
636 // either be a merge instruction (SelectInst / PHINode), in which
637 // case its status is looked up in the phiStates map; or a regular
638 // SSA value, in which case it is assumed to be a base.
639 void meetWith(Value *V) {
640 PhiState otherState = getStateForBDV(V);
641 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
642 MeetPhiStates::pureMeet(currentResult, otherState)) &&
643 "math is wrong: meet does not commute!");
644 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
647 PhiState getResult() const { return currentResult; }
650 const ConflictStateMapTy &phiStates;
651 PhiState currentResult;
653 /// Return a phi state for a base defining value. We'll generate a new
654 /// base state for known bases and expect to find a cached state otherwise
655 PhiState getStateForBDV(Value *baseValue) {
656 if (isKnownBaseResult(baseValue)) {
657 return PhiState(baseValue);
659 return lookupFromMap(baseValue);
663 PhiState lookupFromMap(Value *V) {
664 auto I = phiStates.find(V);
665 assert(I != phiStates.end() && "lookup failed!");
669 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
670 switch (stateA.getStatus()) {
671 case PhiState::Unknown:
675 assert(stateA.getBase() && "can't be null");
676 if (stateB.isUnknown())
679 if (stateB.isBase()) {
680 if (stateA.getBase() == stateB.getBase()) {
681 assert(stateA == stateB && "equality broken!");
684 return PhiState(PhiState::Conflict);
686 assert(stateB.isConflict() && "only three states!");
687 return PhiState(PhiState::Conflict);
689 case PhiState::Conflict:
692 llvm_unreachable("only three states!");
696 /// For a given value or instruction, figure out what base ptr it's derived
697 /// from. For gc objects, this is simply itself. On success, returns a value
698 /// which is the base pointer. (This is reliable and can be used for
699 /// relocation.) On failure, returns nullptr.
700 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
701 Value *def = findBaseOrBDV(I, cache);
703 if (isKnownBaseResult(def)) {
707 // Here's the rough algorithm:
708 // - For every SSA value, construct a mapping to either an actual base
709 // pointer or a PHI which obscures the base pointer.
710 // - Construct a mapping from PHI to unknown TOP state. Use an
711 // optimistic algorithm to propagate base pointer information. Lattice
716 // When algorithm terminates, all PHIs will either have a single concrete
717 // base or be in a conflict state.
718 // - For every conflict, insert a dummy PHI node without arguments. Add
719 // these to the base[Instruction] = BasePtr mapping. For every
720 // non-conflict, add the actual base.
721 // - For every conflict, add arguments for the base[a] of each input
724 // Note: A simpler form of this would be to add the conflict form of all
725 // PHIs without running the optimistic algorithm. This would be
726 // analougous to pessimistic data flow and would likely lead to an
727 // overall worse solution.
729 ConflictStateMapTy states;
730 states[def] = PhiState();
731 // Recursively fill in all phis & selects reachable from the initial one
732 // for which we don't already know a definite base value for
733 // TODO: This should be rewritten with a worklist
737 // Since we're adding elements to 'states' as we run, we can't keep
738 // iterators into the set.
739 SmallVector<Value *, 16> Keys;
740 Keys.reserve(states.size());
741 for (auto Pair : states) {
742 Value *V = Pair.first;
745 for (Value *v : Keys) {
746 assert(!isKnownBaseResult(v) && "why did it get added?");
747 if (PHINode *phi = dyn_cast<PHINode>(v)) {
748 assert(phi->getNumIncomingValues() > 0 &&
749 "zero input phis are illegal");
750 for (Value *InVal : phi->incoming_values()) {
751 Value *local = findBaseOrBDV(InVal, cache);
752 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
753 states[local] = PhiState();
757 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
758 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
759 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
760 states[local] = PhiState();
763 local = findBaseOrBDV(sel->getFalseValue(), cache);
764 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
765 states[local] = PhiState();
773 errs() << "States after initialization:\n";
774 for (auto Pair : states) {
775 Instruction *v = cast<Instruction>(Pair.first);
776 PhiState state = Pair.second;
782 // TODO: come back and revisit the state transitions around inputs which
783 // have reached conflict state. The current version seems too conservative.
785 bool progress = true;
788 size_t oldSize = states.size();
791 // We're only changing keys in this loop, thus safe to keep iterators
792 for (auto Pair : states) {
793 MeetPhiStates calculateMeet(states);
794 Value *v = Pair.first;
795 assert(!isKnownBaseResult(v) && "why did it get added?");
796 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
797 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
798 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
800 for (Value *Val : cast<PHINode>(v)->incoming_values())
801 calculateMeet.meetWith(findBaseOrBDV(Val, cache));
803 PhiState oldState = states[v];
804 PhiState newState = calculateMeet.getResult();
805 if (oldState != newState) {
807 states[v] = newState;
811 assert(oldSize <= states.size());
812 assert(oldSize == states.size() || progress);
816 errs() << "States after meet iteration:\n";
817 for (auto Pair : states) {
818 Instruction *v = cast<Instruction>(Pair.first);
819 PhiState state = Pair.second;
825 // Insert Phis for all conflicts
826 // We want to keep naming deterministic in the loop that follows, so
827 // sort the keys before iteration. This is useful in allowing us to
828 // write stable tests. Note that there is no invalidation issue here.
829 SmallVector<Value *, 16> Keys;
830 Keys.reserve(states.size());
831 for (auto Pair : states) {
832 Value *V = Pair.first;
835 std::sort(Keys.begin(), Keys.end(), order_by_name);
836 // TODO: adjust naming patterns to avoid this order of iteration dependency
837 for (Value *V : Keys) {
838 Instruction *v = cast<Instruction>(V);
839 PhiState state = states[V];
840 assert(!isKnownBaseResult(v) && "why did it get added?");
841 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
842 if (!state.isConflict())
845 if (isa<PHINode>(v)) {
847 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
848 assert(num_preds > 0 && "how did we reach here");
849 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
850 // Add metadata marking this as a base value
851 phi->setMetadata("is_base_value", MDNode::get(v->getContext(), {}));
852 states[v] = PhiState(PhiState::Conflict, phi);
854 SelectInst *sel = cast<SelectInst>(v);
855 // The undef will be replaced later
856 UndefValue *undef = UndefValue::get(sel->getType());
857 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
858 undef, "base_select", sel);
859 // Add metadata marking this as a base value
860 basesel->setMetadata("is_base_value", MDNode::get(v->getContext(), {}));
861 states[v] = PhiState(PhiState::Conflict, basesel);
865 // Fixup all the inputs of the new PHIs
866 for (auto Pair : states) {
867 Instruction *v = cast<Instruction>(Pair.first);
868 PhiState state = Pair.second;
870 assert(!isKnownBaseResult(v) && "why did it get added?");
871 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
872 if (!state.isConflict())
875 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
876 PHINode *phi = cast<PHINode>(v);
877 unsigned NumPHIValues = phi->getNumIncomingValues();
878 for (unsigned i = 0; i < NumPHIValues; i++) {
879 Value *InVal = phi->getIncomingValue(i);
880 BasicBlock *InBB = phi->getIncomingBlock(i);
882 // If we've already seen InBB, add the same incoming value
883 // we added for it earlier. The IR verifier requires phi
884 // nodes with multiple entries from the same basic block
885 // to have the same incoming value for each of those
886 // entries. If we don't do this check here and basephi
887 // has a different type than base, we'll end up adding two
888 // bitcasts (and hence two distinct values) as incoming
889 // values for the same basic block.
891 int blockIndex = basephi->getBasicBlockIndex(InBB);
892 if (blockIndex != -1) {
893 Value *oldBase = basephi->getIncomingValue(blockIndex);
894 basephi->addIncoming(oldBase, InBB);
896 Value *base = findBaseOrBDV(InVal, cache);
897 if (!isKnownBaseResult(base)) {
898 // Either conflict or base.
899 assert(states.count(base));
900 base = states[base].getBase();
901 assert(base != nullptr && "unknown PhiState!");
904 // In essense this assert states: the only way two
905 // values incoming from the same basic block may be
906 // different is by being different bitcasts of the same
907 // value. A cleanup that remains TODO is changing
908 // findBaseOrBDV to return an llvm::Value of the correct
909 // type (and still remain pure). This will remove the
910 // need to add bitcasts.
911 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
912 "sanity -- findBaseOrBDV should be pure!");
917 // Find either the defining value for the PHI or the normal base for
919 Value *base = findBaseOrBDV(InVal, cache);
920 if (!isKnownBaseResult(base)) {
921 // Either conflict or base.
922 assert(states.count(base));
923 base = states[base].getBase();
924 assert(base != nullptr && "unknown PhiState!");
926 assert(base && "can't be null");
927 // Must use original input BB since base may not be Instruction
928 // The cast is needed since base traversal may strip away bitcasts
929 if (base->getType() != basephi->getType()) {
930 base = new BitCastInst(base, basephi->getType(), "cast",
931 InBB->getTerminator());
933 basephi->addIncoming(base, InBB);
935 assert(basephi->getNumIncomingValues() == NumPHIValues);
937 SelectInst *basesel = cast<SelectInst>(state.getBase());
938 SelectInst *sel = cast<SelectInst>(v);
939 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
940 // something more safe and less hacky.
941 for (int i = 1; i <= 2; i++) {
942 Value *InVal = sel->getOperand(i);
943 // Find either the defining value for the PHI or the normal base for
945 Value *base = findBaseOrBDV(InVal, cache);
946 if (!isKnownBaseResult(base)) {
947 // Either conflict or base.
948 assert(states.count(base));
949 base = states[base].getBase();
950 assert(base != nullptr && "unknown PhiState!");
952 assert(base && "can't be null");
953 // Must use original input BB since base may not be Instruction
954 // The cast is needed since base traversal may strip away bitcasts
955 if (base->getType() != basesel->getType()) {
956 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
958 basesel->setOperand(i, base);
963 // Cache all of our results so we can cheaply reuse them
964 // NOTE: This is actually two caches: one of the base defining value
965 // relation and one of the base pointer relation! FIXME
966 for (auto item : states) {
967 Value *v = item.first;
968 Value *base = item.second.getBase();
970 assert(!isKnownBaseResult(v) && "why did it get added?");
973 std::string fromstr =
974 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
976 errs() << "Updating base value cache"
977 << " for: " << (v->hasName() ? v->getName() : "")
978 << " from: " << fromstr
979 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
982 assert(isKnownBaseResult(base) &&
983 "must be something we 'know' is a base pointer");
984 if (cache.count(v)) {
985 // Once we transition from the BDV relation being store in the cache to
986 // the base relation being stored, it must be stable
987 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
988 "base relation should be stable");
992 assert(cache.find(def) != cache.end());
996 // For a set of live pointers (base and/or derived), identify the base
997 // pointer of the object which they are derived from. This routine will
998 // mutate the IR graph as needed to make the 'base' pointer live at the
999 // definition site of 'derived'. This ensures that any use of 'derived' can
1000 // also use 'base'. This may involve the insertion of a number of
1001 // additional PHI nodes.
1003 // preconditions: live is a set of pointer type Values
1005 // side effects: may insert PHI nodes into the existing CFG, will preserve
1006 // CFG, will not remove or mutate any existing nodes
1008 // post condition: PointerToBase contains one (derived, base) pair for every
1009 // pointer in live. Note that derived can be equal to base if the original
1010 // pointer was a base pointer.
1012 findBasePointers(const StatepointLiveSetTy &live,
1013 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
1014 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1015 // For the naming of values inserted to be deterministic - which makes for
1016 // much cleaner and more stable tests - we need to assign an order to the
1017 // live values. DenseSets do not provide a deterministic order across runs.
1018 SmallVector<Value *, 64> Temp;
1019 Temp.insert(Temp.end(), live.begin(), live.end());
1020 std::sort(Temp.begin(), Temp.end(), order_by_name);
1021 for (Value *ptr : Temp) {
1022 Value *base = findBasePointer(ptr, DVCache);
1023 assert(base && "failed to find base pointer");
1024 PointerToBase[ptr] = base;
1025 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1026 DT->dominates(cast<Instruction>(base)->getParent(),
1027 cast<Instruction>(ptr)->getParent())) &&
1028 "The base we found better dominate the derived pointer");
1030 // If you see this trip and like to live really dangerously, the code should
1031 // be correct, just with idioms the verifier can't handle. You can try
1032 // disabling the verifier at your own substaintial risk.
1033 assert(!isa<ConstantPointerNull>(base) &&
1034 "the relocation code needs adjustment to handle the relocation of "
1035 "a null pointer constant without causing false positives in the "
1036 "safepoint ir verifier.");
1040 /// Find the required based pointers (and adjust the live set) for the given
1042 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1044 PartiallyConstructedSafepointRecord &result) {
1045 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
1046 findBasePointers(result.liveset, PointerToBase, &DT, DVCache);
1048 if (PrintBasePointers) {
1049 // Note: Need to print these in a stable order since this is checked in
1051 errs() << "Base Pairs (w/o Relocation):\n";
1052 SmallVector<Value *, 64> Temp;
1053 Temp.reserve(PointerToBase.size());
1054 for (auto Pair : PointerToBase) {
1055 Temp.push_back(Pair.first);
1057 std::sort(Temp.begin(), Temp.end(), order_by_name);
1058 for (Value *Ptr : Temp) {
1059 Value *Base = PointerToBase[Ptr];
1060 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
1065 result.PointerToBase = PointerToBase;
1068 /// Given an updated version of the dataflow liveness results, update the
1069 /// liveset and base pointer maps for the call site CS.
1070 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1072 PartiallyConstructedSafepointRecord &result);
1074 static void recomputeLiveInValues(
1075 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1076 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1077 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1078 // again. The old values are still live and will help it stablize quickly.
1079 GCPtrLivenessData RevisedLivenessData;
1080 computeLiveInValues(DT, F, RevisedLivenessData);
1081 for (size_t i = 0; i < records.size(); i++) {
1082 struct PartiallyConstructedSafepointRecord &info = records[i];
1083 const CallSite &CS = toUpdate[i];
1084 recomputeLiveInValues(RevisedLivenessData, CS, info);
1088 // When inserting gc.relocate calls, we need to ensure there are no uses
1089 // of the original value between the gc.statepoint and the gc.relocate call.
1090 // One case which can arise is a phi node starting one of the successor blocks.
1091 // We also need to be able to insert the gc.relocates only on the path which
1092 // goes through the statepoint. We might need to split an edge to make this
1095 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1096 DominatorTree &DT) {
1097 BasicBlock *Ret = BB;
1098 if (!BB->getUniquePredecessor()) {
1099 Ret = SplitBlockPredecessors(BB, InvokeParent, "", nullptr, &DT);
1102 // Now that 'ret' has unique predecessor we can safely remove all phi nodes
1104 FoldSingleEntryPHINodes(Ret);
1105 assert(!isa<PHINode>(Ret->begin()));
1107 // At this point, we can safely insert a gc.relocate as the first instruction
1108 // in Ret if needed.
1112 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1113 auto itr = std::find(livevec.begin(), livevec.end(), val);
1114 assert(livevec.end() != itr);
1115 size_t index = std::distance(livevec.begin(), itr);
1116 assert(index < livevec.size());
1120 // Create new attribute set containing only attributes which can be transfered
1121 // from original call to the safepoint.
1122 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1125 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1126 unsigned index = AS.getSlotIndex(Slot);
1128 if (index == AttributeSet::ReturnIndex ||
1129 index == AttributeSet::FunctionIndex) {
1131 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1133 Attribute attr = *it;
1135 // Do not allow certain attributes - just skip them
1136 // Safepoint can not be read only or read none.
1137 if (attr.hasAttribute(Attribute::ReadNone) ||
1138 attr.hasAttribute(Attribute::ReadOnly))
1141 ret = ret.addAttributes(
1142 AS.getContext(), index,
1143 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1147 // Just skip parameter attributes for now
1153 /// Helper function to place all gc relocates necessary for the given
1156 /// liveVariables - list of variables to be relocated.
1157 /// liveStart - index of the first live variable.
1158 /// basePtrs - base pointers.
1159 /// statepointToken - statepoint instruction to which relocates should be
1161 /// Builder - Llvm IR builder to be used to construct new calls.
1162 static void CreateGCRelocates(ArrayRef<llvm::Value *> LiveVariables,
1163 const int LiveStart,
1164 ArrayRef<llvm::Value *> BasePtrs,
1165 Instruction *StatepointToken,
1166 IRBuilder<> Builder) {
1167 if (LiveVariables.empty())
1170 // All gc_relocate are set to i8 addrspace(1)* type. We originally generated
1171 // unique declarations for each pointer type, but this proved problematic
1172 // because the intrinsic mangling code is incomplete and fragile. Since
1173 // we're moving towards a single unified pointer type anyways, we can just
1174 // cast everything to an i8* of the right address space. A bitcast is added
1175 // later to convert gc_relocate to the actual value's type.
1176 Module *M = StatepointToken->getModule();
1177 auto AS = cast<PointerType>(LiveVariables[0]->getType())->getAddressSpace();
1178 Type *Types[] = {Type::getInt8PtrTy(M->getContext(), AS)};
1179 Value *GCRelocateDecl =
1180 Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, Types);
1182 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1183 // Generate the gc.relocate call and save the result
1185 Builder.getInt32(LiveStart + find_index(LiveVariables, BasePtrs[i]));
1187 Builder.getInt32(LiveStart + find_index(LiveVariables, LiveVariables[i]));
1189 // only specify a debug name if we can give a useful one
1190 CallInst *Reloc = Builder.CreateCall(
1191 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1192 LiveVariables[i]->hasName() ? LiveVariables[i]->getName() + ".relocated"
1194 // Trick CodeGen into thinking there are lots of free registers at this
1196 Reloc->setCallingConv(CallingConv::Cold);
1201 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1202 const SmallVectorImpl<llvm::Value *> &basePtrs,
1203 const SmallVectorImpl<llvm::Value *> &liveVariables,
1205 PartiallyConstructedSafepointRecord &result) {
1206 assert(basePtrs.size() == liveVariables.size());
1207 assert(isStatepoint(CS) &&
1208 "This method expects to be rewriting a statepoint");
1210 BasicBlock *BB = CS.getInstruction()->getParent();
1212 Function *F = BB->getParent();
1213 assert(F && "must be set");
1214 Module *M = F->getParent();
1216 assert(M && "must be set");
1218 // We're not changing the function signature of the statepoint since the gc
1219 // arguments go into the var args section.
1220 Function *gc_statepoint_decl = CS.getCalledFunction();
1222 // Then go ahead and use the builder do actually do the inserts. We insert
1223 // immediately before the previous instruction under the assumption that all
1224 // arguments will be available here. We can't insert afterwards since we may
1225 // be replacing a terminator.
1226 Instruction *insertBefore = CS.getInstruction();
1227 IRBuilder<> Builder(insertBefore);
1228 // Copy all of the arguments from the original statepoint - this includes the
1229 // target, call args, and deopt args
1230 SmallVector<llvm::Value *, 64> args;
1231 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1232 // TODO: Clear the 'needs rewrite' flag
1234 // add all the pointers to be relocated (gc arguments)
1235 // Capture the start of the live variable list for use in the gc_relocates
1236 const int live_start = args.size();
1237 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1239 // Create the statepoint given all the arguments
1240 Instruction *token = nullptr;
1241 AttributeSet return_attributes;
1243 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1245 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1246 call->setTailCall(toReplace->isTailCall());
1247 call->setCallingConv(toReplace->getCallingConv());
1249 // Currently we will fail on parameter attributes and on certain
1250 // function attributes.
1251 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1252 // In case if we can handle this set of sttributes - set up function attrs
1253 // directly on statepoint and return attrs later for gc_result intrinsic.
1254 call->setAttributes(new_attrs.getFnAttributes());
1255 return_attributes = new_attrs.getRetAttributes();
1259 // Put the following gc_result and gc_relocate calls immediately after the
1260 // the old call (which we're about to delete)
1261 BasicBlock::iterator next(toReplace);
1262 assert(BB->end() != next && "not a terminator, must have next");
1264 Instruction *IP = &*(next);
1265 Builder.SetInsertPoint(IP);
1266 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1269 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1271 // Insert the new invoke into the old block. We'll remove the old one in a
1272 // moment at which point this will become the new terminator for the
1274 InvokeInst *invoke = InvokeInst::Create(
1275 gc_statepoint_decl, toReplace->getNormalDest(),
1276 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1277 invoke->setCallingConv(toReplace->getCallingConv());
1279 // Currently we will fail on parameter attributes and on certain
1280 // function attributes.
1281 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1282 // In case if we can handle this set of sttributes - set up function attrs
1283 // directly on statepoint and return attrs later for gc_result intrinsic.
1284 invoke->setAttributes(new_attrs.getFnAttributes());
1285 return_attributes = new_attrs.getRetAttributes();
1289 // Generate gc relocates in exceptional path
1290 BasicBlock *unwindBlock = toReplace->getUnwindDest();
1291 assert(!isa<PHINode>(unwindBlock->begin()) &&
1292 unwindBlock->getUniquePredecessor() &&
1293 "can't safely insert in this block!");
1295 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1296 Builder.SetInsertPoint(IP);
1297 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1299 // Extract second element from landingpad return value. We will attach
1300 // exceptional gc relocates to it.
1301 const unsigned idx = 1;
1302 Instruction *exceptional_token =
1303 cast<Instruction>(Builder.CreateExtractValue(
1304 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1305 result.UnwindToken = exceptional_token;
1307 // Just throw away return value. We will use the one we got for normal
1309 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1310 exceptional_token, Builder);
1312 // Generate gc relocates and returns for normal block
1313 BasicBlock *normalDest = toReplace->getNormalDest();
1314 assert(!isa<PHINode>(normalDest->begin()) &&
1315 normalDest->getUniquePredecessor() &&
1316 "can't safely insert in this block!");
1318 IP = &*(normalDest->getFirstInsertionPt());
1319 Builder.SetInsertPoint(IP);
1321 // gc relocates will be generated later as if it were regular call
1326 // Take the name of the original value call if it had one.
1327 token->takeName(CS.getInstruction());
1329 // The GCResult is already inserted, we just need to find it
1331 Instruction *toReplace = CS.getInstruction();
1332 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1333 "only valid use before rewrite is gc.result");
1334 assert(!toReplace->hasOneUse() ||
1335 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1338 // Update the gc.result of the original statepoint (if any) to use the newly
1339 // inserted statepoint. This is safe to do here since the token can't be
1340 // considered a live reference.
1341 CS.getInstruction()->replaceAllUsesWith(token);
1343 result.StatepointToken = token;
1345 // Second, create a gc.relocate for every live variable
1346 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1350 struct name_ordering {
1353 bool operator()(name_ordering const &a, name_ordering const &b) {
1354 return -1 == a.derived->getName().compare(b.derived->getName());
1358 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1359 SmallVectorImpl<Value *> &livevec) {
1360 assert(basevec.size() == livevec.size());
1362 SmallVector<name_ordering, 64> temp;
1363 for (size_t i = 0; i < basevec.size(); i++) {
1365 v.base = basevec[i];
1366 v.derived = livevec[i];
1369 std::sort(temp.begin(), temp.end(), name_ordering());
1370 for (size_t i = 0; i < basevec.size(); i++) {
1371 basevec[i] = temp[i].base;
1372 livevec[i] = temp[i].derived;
1376 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1377 // which make the relocations happening at this safepoint explicit.
1379 // WARNING: Does not do any fixup to adjust users of the original live
1380 // values. That's the callers responsibility.
1382 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1383 PartiallyConstructedSafepointRecord &result) {
1384 auto liveset = result.liveset;
1385 auto PointerToBase = result.PointerToBase;
1387 // Convert to vector for efficient cross referencing.
1388 SmallVector<Value *, 64> basevec, livevec;
1389 livevec.reserve(liveset.size());
1390 basevec.reserve(liveset.size());
1391 for (Value *L : liveset) {
1392 livevec.push_back(L);
1393 assert(PointerToBase.count(L));
1394 Value *base = PointerToBase[L];
1395 basevec.push_back(base);
1397 assert(livevec.size() == basevec.size());
1399 // To make the output IR slightly more stable (for use in diffs), ensure a
1400 // fixed order of the values in the safepoint (by sorting the value name).
1401 // The order is otherwise meaningless.
1402 stablize_order(basevec, livevec);
1404 // Do the actual rewriting and delete the old statepoint
1405 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1406 CS.getInstruction()->eraseFromParent();
1409 // Helper function for the relocationViaAlloca.
1410 // It receives iterator to the statepoint gc relocates and emits store to the
1412 // location (via allocaMap) for the each one of them.
1413 // Add visited values into the visitedLiveValues set we will later use them
1414 // for sanity check.
1416 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1417 DenseMap<Value *, Value *> &AllocaMap,
1418 DenseSet<Value *> &VisitedLiveValues) {
1420 for (User *U : GCRelocs) {
1421 if (!isa<IntrinsicInst>(U))
1424 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1426 // We only care about relocates
1427 if (RelocatedValue->getIntrinsicID() !=
1428 Intrinsic::experimental_gc_relocate) {
1432 GCRelocateOperands RelocateOperands(RelocatedValue);
1433 Value *OriginalValue =
1434 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1435 assert(AllocaMap.count(OriginalValue));
1436 Value *Alloca = AllocaMap[OriginalValue];
1438 // Emit store into the related alloca
1439 // All gc_relocate are i8 addrspace(1)* typed, and it must be bitcasted to
1440 // the correct type according to alloca.
1441 assert(RelocatedValue->getNextNode() && "Should always have one since it's not a terminator");
1442 IRBuilder<> Builder(RelocatedValue->getNextNode());
1443 Value *CastedRelocatedValue =
1444 Builder.CreateBitCast(RelocatedValue, cast<AllocaInst>(Alloca)->getAllocatedType(),
1445 RelocatedValue->hasName() ? RelocatedValue->getName() + ".casted" : "");
1447 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1448 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1451 VisitedLiveValues.insert(OriginalValue);
1456 // Helper function for the "relocationViaAlloca". Similar to the
1457 // "insertRelocationStores" but works for rematerialized values.
1459 insertRematerializationStores(
1460 RematerializedValueMapTy RematerializedValues,
1461 DenseMap<Value *, Value *> &AllocaMap,
1462 DenseSet<Value *> &VisitedLiveValues) {
1464 for (auto RematerializedValuePair: RematerializedValues) {
1465 Instruction *RematerializedValue = RematerializedValuePair.first;
1466 Value *OriginalValue = RematerializedValuePair.second;
1468 assert(AllocaMap.count(OriginalValue) &&
1469 "Can not find alloca for rematerialized value");
1470 Value *Alloca = AllocaMap[OriginalValue];
1472 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1473 Store->insertAfter(RematerializedValue);
1476 VisitedLiveValues.insert(OriginalValue);
1481 /// do all the relocation update via allocas and mem2reg
1482 static void relocationViaAlloca(
1483 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1484 ArrayRef<struct PartiallyConstructedSafepointRecord> Records) {
1486 // record initial number of (static) allocas; we'll check we have the same
1487 // number when we get done.
1488 int InitialAllocaNum = 0;
1489 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1491 if (isa<AllocaInst>(*I))
1495 // TODO-PERF: change data structures, reserve
1496 DenseMap<Value *, Value *> AllocaMap;
1497 SmallVector<AllocaInst *, 200> PromotableAllocas;
1498 // Used later to chack that we have enough allocas to store all values
1499 std::size_t NumRematerializedValues = 0;
1500 PromotableAllocas.reserve(Live.size());
1502 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1503 // "PromotableAllocas"
1504 auto emitAllocaFor = [&](Value *LiveValue) {
1505 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1506 F.getEntryBlock().getFirstNonPHI());
1507 AllocaMap[LiveValue] = Alloca;
1508 PromotableAllocas.push_back(Alloca);
1511 // emit alloca for each live gc pointer
1512 for (unsigned i = 0; i < Live.size(); i++) {
1513 emitAllocaFor(Live[i]);
1516 // emit allocas for rematerialized values
1517 for (size_t i = 0; i < Records.size(); i++) {
1518 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1520 for (auto RematerializedValuePair : Info.RematerializedValues) {
1521 Value *OriginalValue = RematerializedValuePair.second;
1522 if (AllocaMap.count(OriginalValue) != 0)
1525 emitAllocaFor(OriginalValue);
1526 ++NumRematerializedValues;
1530 // The next two loops are part of the same conceptual operation. We need to
1531 // insert a store to the alloca after the original def and at each
1532 // redefinition. We need to insert a load before each use. These are split
1533 // into distinct loops for performance reasons.
1535 // update gc pointer after each statepoint
1536 // either store a relocated value or null (if no relocated value found for
1537 // this gc pointer and it is not a gc_result)
1538 // this must happen before we update the statepoint with load of alloca
1539 // otherwise we lose the link between statepoint and old def
1540 for (size_t i = 0; i < Records.size(); i++) {
1541 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1542 Value *Statepoint = Info.StatepointToken;
1544 // This will be used for consistency check
1545 DenseSet<Value *> VisitedLiveValues;
1547 // Insert stores for normal statepoint gc relocates
1548 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1550 // In case if it was invoke statepoint
1551 // we will insert stores for exceptional path gc relocates.
1552 if (isa<InvokeInst>(Statepoint)) {
1553 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1557 // Do similar thing with rematerialized values
1558 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1561 if (ClobberNonLive) {
1562 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1563 // the gc.statepoint. This will turn some subtle GC problems into
1564 // slightly easier to debug SEGVs. Note that on large IR files with
1565 // lots of gc.statepoints this is extremely costly both memory and time
1567 SmallVector<AllocaInst *, 64> ToClobber;
1568 for (auto Pair : AllocaMap) {
1569 Value *Def = Pair.first;
1570 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1572 // This value was relocated
1573 if (VisitedLiveValues.count(Def)) {
1576 ToClobber.push_back(Alloca);
1579 auto InsertClobbersAt = [&](Instruction *IP) {
1580 for (auto *AI : ToClobber) {
1581 auto AIType = cast<PointerType>(AI->getType());
1582 auto PT = cast<PointerType>(AIType->getElementType());
1583 Constant *CPN = ConstantPointerNull::get(PT);
1584 StoreInst *Store = new StoreInst(CPN, AI);
1585 Store->insertBefore(IP);
1589 // Insert the clobbering stores. These may get intermixed with the
1590 // gc.results and gc.relocates, but that's fine.
1591 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1592 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1593 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1595 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1597 InsertClobbersAt(Next);
1601 // update use with load allocas and add store for gc_relocated
1602 for (auto Pair : AllocaMap) {
1603 Value *Def = Pair.first;
1604 Value *Alloca = Pair.second;
1606 // we pre-record the uses of allocas so that we dont have to worry about
1608 // that change the user information.
1609 SmallVector<Instruction *, 20> Uses;
1610 // PERF: trade a linear scan for repeated reallocation
1611 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1612 for (User *U : Def->users()) {
1613 if (!isa<ConstantExpr>(U)) {
1614 // If the def has a ConstantExpr use, then the def is either a
1615 // ConstantExpr use itself or null. In either case
1616 // (recursively in the first, directly in the second), the oop
1617 // it is ultimately dependent on is null and this particular
1618 // use does not need to be fixed up.
1619 Uses.push_back(cast<Instruction>(U));
1623 std::sort(Uses.begin(), Uses.end());
1624 auto Last = std::unique(Uses.begin(), Uses.end());
1625 Uses.erase(Last, Uses.end());
1627 for (Instruction *Use : Uses) {
1628 if (isa<PHINode>(Use)) {
1629 PHINode *Phi = cast<PHINode>(Use);
1630 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1631 if (Def == Phi->getIncomingValue(i)) {
1632 LoadInst *Load = new LoadInst(
1633 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1634 Phi->setIncomingValue(i, Load);
1638 LoadInst *Load = new LoadInst(Alloca, "", Use);
1639 Use->replaceUsesOfWith(Def, Load);
1643 // emit store for the initial gc value
1644 // store must be inserted after load, otherwise store will be in alloca's
1645 // use list and an extra load will be inserted before it
1646 StoreInst *Store = new StoreInst(Def, Alloca);
1647 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1648 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1649 // InvokeInst is a TerminatorInst so the store need to be inserted
1650 // into its normal destination block.
1651 BasicBlock *NormalDest = Invoke->getNormalDest();
1652 Store->insertBefore(NormalDest->getFirstNonPHI());
1654 assert(!Inst->isTerminator() &&
1655 "The only TerminatorInst that can produce a value is "
1656 "InvokeInst which is handled above.");
1657 Store->insertAfter(Inst);
1660 assert(isa<Argument>(Def));
1661 Store->insertAfter(cast<Instruction>(Alloca));
1665 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1666 "we must have the same allocas with lives");
1667 if (!PromotableAllocas.empty()) {
1668 // apply mem2reg to promote alloca to SSA
1669 PromoteMemToReg(PromotableAllocas, DT);
1673 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1675 if (isa<AllocaInst>(*I))
1677 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1681 /// Implement a unique function which doesn't require we sort the input
1682 /// vector. Doing so has the effect of changing the output of a couple of
1683 /// tests in ways which make them less useful in testing fused safepoints.
1684 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1685 SmallSet<T, 8> Seen;
1686 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1687 return !Seen.insert(V).second;
1691 /// Insert holders so that each Value is obviously live through the entire
1692 /// lifetime of the call.
1693 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1694 SmallVectorImpl<CallInst *> &Holders) {
1696 // No values to hold live, might as well not insert the empty holder
1699 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1700 // Use a dummy vararg function to actually hold the values live
1701 Function *Func = cast<Function>(M->getOrInsertFunction(
1702 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1704 // For call safepoints insert dummy calls right after safepoint
1705 BasicBlock::iterator Next(CS.getInstruction());
1707 Holders.push_back(CallInst::Create(Func, Values, "", Next));
1710 // For invoke safepooints insert dummy calls both in normal and
1711 // exceptional destination blocks
1712 auto *II = cast<InvokeInst>(CS.getInstruction());
1713 Holders.push_back(CallInst::Create(
1714 Func, Values, "", II->getNormalDest()->getFirstInsertionPt()));
1715 Holders.push_back(CallInst::Create(
1716 Func, Values, "", II->getUnwindDest()->getFirstInsertionPt()));
1719 static void findLiveReferences(
1720 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1721 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1722 GCPtrLivenessData OriginalLivenessData;
1723 computeLiveInValues(DT, F, OriginalLivenessData);
1724 for (size_t i = 0; i < records.size(); i++) {
1725 struct PartiallyConstructedSafepointRecord &info = records[i];
1726 const CallSite &CS = toUpdate[i];
1727 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1731 /// Remove any vector of pointers from the liveset by scalarizing them over the
1732 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1733 /// would be preferrable to include the vector in the statepoint itself, but
1734 /// the lowering code currently does not handle that. Extending it would be
1735 /// slightly non-trivial since it requires a format change. Given how rare
1736 /// such cases are (for the moment?) scalarizing is an acceptable comprimise.
1737 static void splitVectorValues(Instruction *StatepointInst,
1738 StatepointLiveSetTy &LiveSet,
1739 DenseMap<Value *, Value *>& PointerToBase,
1740 DominatorTree &DT) {
1741 SmallVector<Value *, 16> ToSplit;
1742 for (Value *V : LiveSet)
1743 if (isa<VectorType>(V->getType()))
1744 ToSplit.push_back(V);
1746 if (ToSplit.empty())
1749 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1751 Function &F = *(StatepointInst->getParent()->getParent());
1753 DenseMap<Value *, AllocaInst *> AllocaMap;
1754 // First is normal return, second is exceptional return (invoke only)
1755 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1756 for (Value *V : ToSplit) {
1757 AllocaInst *Alloca =
1758 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1759 AllocaMap[V] = Alloca;
1761 VectorType *VT = cast<VectorType>(V->getType());
1762 IRBuilder<> Builder(StatepointInst);
1763 SmallVector<Value *, 16> Elements;
1764 for (unsigned i = 0; i < VT->getNumElements(); i++)
1765 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1766 ElementMapping[V] = Elements;
1768 auto InsertVectorReform = [&](Instruction *IP) {
1769 Builder.SetInsertPoint(IP);
1770 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1771 Value *ResultVec = UndefValue::get(VT);
1772 for (unsigned i = 0; i < VT->getNumElements(); i++)
1773 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1774 Builder.getInt32(i));
1778 if (isa<CallInst>(StatepointInst)) {
1779 BasicBlock::iterator Next(StatepointInst);
1781 Instruction *IP = &*(Next);
1782 Replacements[V].first = InsertVectorReform(IP);
1783 Replacements[V].second = nullptr;
1785 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1786 // We've already normalized - check that we don't have shared destination
1788 BasicBlock *NormalDest = Invoke->getNormalDest();
1789 assert(!isa<PHINode>(NormalDest->begin()));
1790 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1791 assert(!isa<PHINode>(UnwindDest->begin()));
1792 // Insert insert element sequences in both successors
1793 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1794 Replacements[V].first = InsertVectorReform(IP);
1795 IP = &*(UnwindDest->getFirstInsertionPt());
1796 Replacements[V].second = InsertVectorReform(IP);
1800 for (Value *V : ToSplit) {
1801 AllocaInst *Alloca = AllocaMap[V];
1803 // Capture all users before we start mutating use lists
1804 SmallVector<Instruction *, 16> Users;
1805 for (User *U : V->users())
1806 Users.push_back(cast<Instruction>(U));
1808 for (Instruction *I : Users) {
1809 if (auto Phi = dyn_cast<PHINode>(I)) {
1810 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1811 if (V == Phi->getIncomingValue(i)) {
1812 LoadInst *Load = new LoadInst(
1813 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1814 Phi->setIncomingValue(i, Load);
1817 LoadInst *Load = new LoadInst(Alloca, "", I);
1818 I->replaceUsesOfWith(V, Load);
1822 // Store the original value and the replacement value into the alloca
1823 StoreInst *Store = new StoreInst(V, Alloca);
1824 if (auto I = dyn_cast<Instruction>(V))
1825 Store->insertAfter(I);
1827 Store->insertAfter(Alloca);
1829 // Normal return for invoke, or call return
1830 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1831 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1832 // Unwind return for invoke only
1833 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1835 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1838 // apply mem2reg to promote alloca to SSA
1839 SmallVector<AllocaInst *, 16> Allocas;
1840 for (Value *V : ToSplit)
1841 Allocas.push_back(AllocaMap[V]);
1842 PromoteMemToReg(Allocas, DT);
1844 // Update our tracking of live pointers and base mappings to account for the
1845 // changes we just made.
1846 for (Value *V : ToSplit) {
1847 auto &Elements = ElementMapping[V];
1850 LiveSet.insert(Elements.begin(), Elements.end());
1851 // We need to update the base mapping as well.
1852 assert(PointerToBase.count(V));
1853 Value *OldBase = PointerToBase[V];
1854 auto &BaseElements = ElementMapping[OldBase];
1855 PointerToBase.erase(V);
1856 assert(Elements.size() == BaseElements.size());
1857 for (unsigned i = 0; i < Elements.size(); i++) {
1858 Value *Elem = Elements[i];
1859 PointerToBase[Elem] = BaseElements[i];
1864 // Helper function for the "rematerializeLiveValues". It walks use chain
1865 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
1866 // values are visited (currently it is GEP's and casts). Returns true if it
1867 // sucessfully reached "BaseValue" and false otherwise.
1868 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
1870 static bool findRematerializableChainToBasePointer(
1871 SmallVectorImpl<Instruction*> &ChainToBase,
1872 Value *CurrentValue, Value *BaseValue) {
1874 // We have found a base value
1875 if (CurrentValue == BaseValue) {
1879 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
1880 ChainToBase.push_back(GEP);
1881 return findRematerializableChainToBasePointer(ChainToBase,
1882 GEP->getPointerOperand(),
1886 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
1887 Value *Def = CI->stripPointerCasts();
1889 // This two checks are basically similar. First one is here for the
1890 // consistency with findBasePointers logic.
1891 assert(!isa<CastInst>(Def) && "not a pointer cast found");
1892 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
1895 ChainToBase.push_back(CI);
1896 return findRematerializableChainToBasePointer(ChainToBase, Def, BaseValue);
1899 // Not supported instruction in the chain
1903 // Helper function for the "rematerializeLiveValues". Compute cost of the use
1904 // chain we are going to rematerialize.
1906 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
1907 TargetTransformInfo &TTI) {
1910 for (Instruction *Instr : Chain) {
1911 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
1912 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
1913 "non noop cast is found during rematerialization");
1915 Type *SrcTy = CI->getOperand(0)->getType();
1916 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
1918 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
1919 // Cost of the address calculation
1920 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
1921 Cost += TTI.getAddressComputationCost(ValTy);
1923 // And cost of the GEP itself
1924 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
1925 // allowed for the external usage)
1926 if (!GEP->hasAllConstantIndices())
1930 llvm_unreachable("unsupported instruciton type during rematerialization");
1937 // From the statepoint liveset pick values that are cheaper to recompute then to
1938 // relocate. Remove this values from the liveset, rematerialize them after
1939 // statepoint and record them in "Info" structure. Note that similar to
1940 // relocated values we don't do any user adjustments here.
1941 static void rematerializeLiveValues(CallSite CS,
1942 PartiallyConstructedSafepointRecord &Info,
1943 TargetTransformInfo &TTI) {
1944 const unsigned int ChainLengthThreshold = 10;
1946 // Record values we are going to delete from this statepoint live set.
1947 // We can not di this in following loop due to iterator invalidation.
1948 SmallVector<Value *, 32> LiveValuesToBeDeleted;
1950 for (Value *LiveValue: Info.liveset) {
1951 // For each live pointer find it's defining chain
1952 SmallVector<Instruction *, 3> ChainToBase;
1953 assert(Info.PointerToBase.count(LiveValue));
1955 findRematerializableChainToBasePointer(ChainToBase,
1957 Info.PointerToBase[LiveValue]);
1958 // Nothing to do, or chain is too long
1960 ChainToBase.size() == 0 ||
1961 ChainToBase.size() > ChainLengthThreshold)
1964 // Compute cost of this chain
1965 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
1966 // TODO: We can also account for cases when we will be able to remove some
1967 // of the rematerialized values by later optimization passes. I.e if
1968 // we rematerialized several intersecting chains. Or if original values
1969 // don't have any uses besides this statepoint.
1971 // For invokes we need to rematerialize each chain twice - for normal and
1972 // for unwind basic blocks. Model this by multiplying cost by two.
1973 if (CS.isInvoke()) {
1976 // If it's too expensive - skip it
1977 if (Cost >= RematerializationThreshold)
1980 // Remove value from the live set
1981 LiveValuesToBeDeleted.push_back(LiveValue);
1983 // Clone instructions and record them inside "Info" structure
1985 // Walk backwards to visit top-most instructions first
1986 std::reverse(ChainToBase.begin(), ChainToBase.end());
1988 // Utility function which clones all instructions from "ChainToBase"
1989 // and inserts them before "InsertBefore". Returns rematerialized value
1990 // which should be used after statepoint.
1991 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
1992 Instruction *LastClonedValue = nullptr;
1993 Instruction *LastValue = nullptr;
1994 for (Instruction *Instr: ChainToBase) {
1995 // Only GEP's and casts are suported as we need to be careful to not
1996 // introduce any new uses of pointers not in the liveset.
1997 // Note that it's fine to introduce new uses of pointers which were
1998 // otherwise not used after this statepoint.
1999 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2001 Instruction *ClonedValue = Instr->clone();
2002 ClonedValue->insertBefore(InsertBefore);
2003 ClonedValue->setName(Instr->getName() + ".remat");
2005 // If it is not first instruction in the chain then it uses previously
2006 // cloned value. We should update it to use cloned value.
2007 if (LastClonedValue) {
2009 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2011 // Assert that cloned instruction does not use any instructions from
2012 // this chain other than LastClonedValue
2013 for (auto OpValue : ClonedValue->operand_values()) {
2014 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2015 ChainToBase.end() &&
2016 "incorrect use in rematerialization chain");
2021 LastClonedValue = ClonedValue;
2024 assert(LastClonedValue);
2025 return LastClonedValue;
2028 // Different cases for calls and invokes. For invokes we need to clone
2029 // instructions both on normal and unwind path.
2031 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2032 assert(InsertBefore);
2033 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2034 Info.RematerializedValues[RematerializedValue] = LiveValue;
2036 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2038 Instruction *NormalInsertBefore =
2039 Invoke->getNormalDest()->getFirstInsertionPt();
2040 Instruction *UnwindInsertBefore =
2041 Invoke->getUnwindDest()->getFirstInsertionPt();
2043 Instruction *NormalRematerializedValue =
2044 rematerializeChain(NormalInsertBefore);
2045 Instruction *UnwindRematerializedValue =
2046 rematerializeChain(UnwindInsertBefore);
2048 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2049 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2053 // Remove rematerializaed values from the live set
2054 for (auto LiveValue: LiveValuesToBeDeleted) {
2055 Info.liveset.erase(LiveValue);
2059 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
2060 SmallVectorImpl<CallSite> &toUpdate) {
2062 // sanity check the input
2063 std::set<CallSite> uniqued;
2064 uniqued.insert(toUpdate.begin(), toUpdate.end());
2065 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
2067 for (size_t i = 0; i < toUpdate.size(); i++) {
2068 CallSite &CS = toUpdate[i];
2069 assert(CS.getInstruction()->getParent()->getParent() == &F);
2070 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
2074 // When inserting gc.relocates for invokes, we need to be able to insert at
2075 // the top of the successor blocks. See the comment on
2076 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2077 // may restructure the CFG.
2078 for (CallSite CS : toUpdate) {
2081 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
2082 normalizeForInvokeSafepoint(invoke->getNormalDest(), invoke->getParent(),
2084 normalizeForInvokeSafepoint(invoke->getUnwindDest(), invoke->getParent(),
2088 // A list of dummy calls added to the IR to keep various values obviously
2089 // live in the IR. We'll remove all of these when done.
2090 SmallVector<CallInst *, 64> holders;
2092 // Insert a dummy call with all of the arguments to the vm_state we'll need
2093 // for the actual safepoint insertion. This ensures reference arguments in
2094 // the deopt argument list are considered live through the safepoint (and
2095 // thus makes sure they get relocated.)
2096 for (size_t i = 0; i < toUpdate.size(); i++) {
2097 CallSite &CS = toUpdate[i];
2098 Statepoint StatepointCS(CS);
2100 SmallVector<Value *, 64> DeoptValues;
2101 for (Use &U : StatepointCS.vm_state_args()) {
2102 Value *Arg = cast<Value>(&U);
2103 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2104 "support for FCA unimplemented");
2105 if (isHandledGCPointerType(Arg->getType()))
2106 DeoptValues.push_back(Arg);
2108 insertUseHolderAfter(CS, DeoptValues, holders);
2111 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
2112 records.reserve(toUpdate.size());
2113 for (size_t i = 0; i < toUpdate.size(); i++) {
2114 struct PartiallyConstructedSafepointRecord info;
2115 records.push_back(info);
2117 assert(records.size() == toUpdate.size());
2119 // A) Identify all gc pointers which are staticly live at the given call
2121 findLiveReferences(F, DT, P, toUpdate, records);
2123 // B) Find the base pointers for each live pointer
2124 /* scope for caching */ {
2125 // Cache the 'defining value' relation used in the computation and
2126 // insertion of base phis and selects. This ensures that we don't insert
2127 // large numbers of duplicate base_phis.
2128 DefiningValueMapTy DVCache;
2130 for (size_t i = 0; i < records.size(); i++) {
2131 struct PartiallyConstructedSafepointRecord &info = records[i];
2132 CallSite &CS = toUpdate[i];
2133 findBasePointers(DT, DVCache, CS, info);
2135 } // end of cache scope
2137 // The base phi insertion logic (for any safepoint) may have inserted new
2138 // instructions which are now live at some safepoint. The simplest such
2141 // phi a <-- will be a new base_phi here
2142 // safepoint 1 <-- that needs to be live here
2146 // We insert some dummy calls after each safepoint to definitely hold live
2147 // the base pointers which were identified for that safepoint. We'll then
2148 // ask liveness for _every_ base inserted to see what is now live. Then we
2149 // remove the dummy calls.
2150 holders.reserve(holders.size() + records.size());
2151 for (size_t i = 0; i < records.size(); i++) {
2152 struct PartiallyConstructedSafepointRecord &info = records[i];
2153 CallSite &CS = toUpdate[i];
2155 SmallVector<Value *, 128> Bases;
2156 for (auto Pair : info.PointerToBase) {
2157 Bases.push_back(Pair.second);
2159 insertUseHolderAfter(CS, Bases, holders);
2162 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2163 // need to rerun liveness. We may *also* have inserted new defs, but that's
2164 // not the key issue.
2165 recomputeLiveInValues(F, DT, P, toUpdate, records);
2167 if (PrintBasePointers) {
2168 for (size_t i = 0; i < records.size(); i++) {
2169 struct PartiallyConstructedSafepointRecord &info = records[i];
2170 errs() << "Base Pairs: (w/Relocation)\n";
2171 for (auto Pair : info.PointerToBase) {
2172 errs() << " derived %" << Pair.first->getName() << " base %"
2173 << Pair.second->getName() << "\n";
2177 for (size_t i = 0; i < holders.size(); i++) {
2178 holders[i]->eraseFromParent();
2179 holders[i] = nullptr;
2183 // Do a limited scalarization of any live at safepoint vector values which
2184 // contain pointers. This enables this pass to run after vectorization at
2185 // the cost of some possible performance loss. TODO: it would be nice to
2186 // natively support vectors all the way through the backend so we don't need
2187 // to scalarize here.
2188 for (size_t i = 0; i < records.size(); i++) {
2189 struct PartiallyConstructedSafepointRecord &info = records[i];
2190 Instruction *statepoint = toUpdate[i].getInstruction();
2191 splitVectorValues(cast<Instruction>(statepoint), info.liveset,
2192 info.PointerToBase, DT);
2195 // In order to reduce live set of statepoint we might choose to rematerialize
2196 // some values instead of relocating them. This is purelly an optimization and
2197 // does not influence correctness.
2198 TargetTransformInfo &TTI =
2199 P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2201 for (size_t i = 0; i < records.size(); i++) {
2202 struct PartiallyConstructedSafepointRecord &info = records[i];
2203 CallSite &CS = toUpdate[i];
2205 rematerializeLiveValues(CS, info, TTI);
2208 // Now run through and replace the existing statepoints with new ones with
2209 // the live variables listed. We do not yet update uses of the values being
2210 // relocated. We have references to live variables that need to
2211 // survive to the last iteration of this loop. (By construction, the
2212 // previous statepoint can not be a live variable, thus we can and remove
2213 // the old statepoint calls as we go.)
2214 for (size_t i = 0; i < records.size(); i++) {
2215 struct PartiallyConstructedSafepointRecord &info = records[i];
2216 CallSite &CS = toUpdate[i];
2217 makeStatepointExplicit(DT, CS, P, info);
2219 toUpdate.clear(); // prevent accident use of invalid CallSites
2221 // Do all the fixups of the original live variables to their relocated selves
2222 SmallVector<Value *, 128> live;
2223 for (size_t i = 0; i < records.size(); i++) {
2224 struct PartiallyConstructedSafepointRecord &info = records[i];
2225 // We can't simply save the live set from the original insertion. One of
2226 // the live values might be the result of a call which needs a safepoint.
2227 // That Value* no longer exists and we need to use the new gc_result.
2228 // Thankfully, the liveset is embedded in the statepoint (and updated), so
2229 // we just grab that.
2230 Statepoint statepoint(info.StatepointToken);
2231 live.insert(live.end(), statepoint.gc_args_begin(),
2232 statepoint.gc_args_end());
2234 // Do some basic sanity checks on our liveness results before performing
2235 // relocation. Relocation can and will turn mistakes in liveness results
2236 // into non-sensical code which is must harder to debug.
2237 // TODO: It would be nice to test consistency as well
2238 assert(DT.isReachableFromEntry(info.StatepointToken->getParent()) &&
2239 "statepoint must be reachable or liveness is meaningless");
2240 for (Value *V : statepoint.gc_args()) {
2241 if (!isa<Instruction>(V))
2242 // Non-instruction values trivial dominate all possible uses
2244 auto LiveInst = cast<Instruction>(V);
2245 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2246 "unreachable values should never be live");
2247 assert(DT.dominates(LiveInst, info.StatepointToken) &&
2248 "basic SSA liveness expectation violated by liveness analysis");
2252 unique_unsorted(live);
2256 for (auto ptr : live) {
2257 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
2261 relocationViaAlloca(F, DT, live, records);
2262 return !records.empty();
2265 // Handles both return values and arguments for Functions and CallSites.
2266 template <typename AttrHolder>
2267 static void RemoveDerefAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2270 if (AH.getDereferenceableBytes(Index))
2271 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2272 AH.getDereferenceableBytes(Index)));
2273 if (AH.getDereferenceableOrNullBytes(Index))
2274 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2275 AH.getDereferenceableOrNullBytes(Index)));
2278 AH.setAttributes(AH.getAttributes().removeAttributes(
2279 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2283 RewriteStatepointsForGC::stripDereferenceabilityInfoFromPrototype(Function &F) {
2284 LLVMContext &Ctx = F.getContext();
2286 for (Argument &A : F.args())
2287 if (isa<PointerType>(A.getType()))
2288 RemoveDerefAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2290 if (isa<PointerType>(F.getReturnType()))
2291 RemoveDerefAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2294 void RewriteStatepointsForGC::stripDereferenceabilityInfoFromBody(Function &F) {
2298 LLVMContext &Ctx = F.getContext();
2299 MDBuilder Builder(Ctx);
2301 for (Instruction &I : inst_range(F)) {
2302 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2303 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2304 bool IsImmutableTBAA =
2305 MD->getNumOperands() == 4 &&
2306 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2308 if (!IsImmutableTBAA)
2309 continue; // no work to do, MD_tbaa is already marked mutable
2311 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2312 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2314 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2316 MDNode *MutableTBAA =
2317 Builder.createTBAAStructTagNode(Base, Access, Offset);
2318 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2321 if (CallSite CS = CallSite(&I)) {
2322 for (int i = 0, e = CS.arg_size(); i != e; i++)
2323 if (isa<PointerType>(CS.getArgument(i)->getType()))
2324 RemoveDerefAttrAtIndex(Ctx, CS, i + 1);
2325 if (isa<PointerType>(CS.getType()))
2326 RemoveDerefAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2331 /// Returns true if this function should be rewritten by this pass. The main
2332 /// point of this function is as an extension point for custom logic.
2333 static bool shouldRewriteStatepointsIn(Function &F) {
2334 // TODO: This should check the GCStrategy
2336 const char *FunctionGCName = F.getGC();
2337 const StringRef StatepointExampleName("statepoint-example");
2338 const StringRef CoreCLRName("coreclr");
2339 return (StatepointExampleName == FunctionGCName) ||
2340 (CoreCLRName == FunctionGCName);
2345 void RewriteStatepointsForGC::stripDereferenceabilityInfo(Module &M) {
2347 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2351 for (Function &F : M)
2352 stripDereferenceabilityInfoFromPrototype(F);
2354 for (Function &F : M)
2355 stripDereferenceabilityInfoFromBody(F);
2358 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2359 // Nothing to do for declarations.
2360 if (F.isDeclaration() || F.empty())
2363 // Policy choice says not to rewrite - the most common reason is that we're
2364 // compiling code without a GCStrategy.
2365 if (!shouldRewriteStatepointsIn(F))
2368 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2370 // Gather all the statepoints which need rewritten. Be careful to only
2371 // consider those in reachable code since we need to ask dominance queries
2372 // when rewriting. We'll delete the unreachable ones in a moment.
2373 SmallVector<CallSite, 64> ParsePointNeeded;
2374 bool HasUnreachableStatepoint = false;
2375 for (Instruction &I : inst_range(F)) {
2376 // TODO: only the ones with the flag set!
2377 if (isStatepoint(I)) {
2378 if (DT.isReachableFromEntry(I.getParent()))
2379 ParsePointNeeded.push_back(CallSite(&I));
2381 HasUnreachableStatepoint = true;
2385 bool MadeChange = false;
2387 // Delete any unreachable statepoints so that we don't have unrewritten
2388 // statepoints surviving this pass. This makes testing easier and the
2389 // resulting IR less confusing to human readers. Rather than be fancy, we
2390 // just reuse a utility function which removes the unreachable blocks.
2391 if (HasUnreachableStatepoint)
2392 MadeChange |= removeUnreachableBlocks(F);
2394 // Return early if no work to do.
2395 if (ParsePointNeeded.empty())
2398 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2399 // These are created by LCSSA. They have the effect of increasing the size
2400 // of liveness sets for no good reason. It may be harder to do this post
2401 // insertion since relocations and base phis can confuse things.
2402 for (BasicBlock &BB : F)
2403 if (BB.getUniquePredecessor()) {
2405 FoldSingleEntryPHINodes(&BB);
2408 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
2412 // liveness computation via standard dataflow
2413 // -------------------------------------------------------------------
2415 // TODO: Consider using bitvectors for liveness, the set of potentially
2416 // interesting values should be small and easy to pre-compute.
2418 /// Compute the live-in set for the location rbegin starting from
2419 /// the live-out set of the basic block
2420 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2421 BasicBlock::reverse_iterator rend,
2422 DenseSet<Value *> &LiveTmp) {
2424 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2425 Instruction *I = &*ritr;
2427 // KILL/Def - Remove this definition from LiveIn
2430 // Don't consider *uses* in PHI nodes, we handle their contribution to
2431 // predecessor blocks when we seed the LiveOut sets
2432 if (isa<PHINode>(I))
2435 // USE - Add to the LiveIn set for this instruction
2436 for (Value *V : I->operands()) {
2437 assert(!isUnhandledGCPointerType(V->getType()) &&
2438 "support for FCA unimplemented");
2439 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2440 // The choice to exclude all things constant here is slightly subtle.
2441 // There are two idependent reasons:
2442 // - We assume that things which are constant (from LLVM's definition)
2443 // do not move at runtime. For example, the address of a global
2444 // variable is fixed, even though it's contents may not be.
2445 // - Second, we can't disallow arbitrary inttoptr constants even
2446 // if the language frontend does. Optimization passes are free to
2447 // locally exploit facts without respect to global reachability. This
2448 // can create sections of code which are dynamically unreachable and
2449 // contain just about anything. (see constants.ll in tests)
2456 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2458 for (BasicBlock *Succ : successors(BB)) {
2459 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2460 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2461 PHINode *Phi = cast<PHINode>(&*I);
2462 Value *V = Phi->getIncomingValueForBlock(BB);
2463 assert(!isUnhandledGCPointerType(V->getType()) &&
2464 "support for FCA unimplemented");
2465 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2472 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2473 DenseSet<Value *> KillSet;
2474 for (Instruction &I : *BB)
2475 if (isHandledGCPointerType(I.getType()))
2481 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2482 /// sanity check for the liveness computation.
2483 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2484 TerminatorInst *TI, bool TermOkay = false) {
2485 for (Value *V : Live) {
2486 if (auto *I = dyn_cast<Instruction>(V)) {
2487 // The terminator can be a member of the LiveOut set. LLVM's definition
2488 // of instruction dominance states that V does not dominate itself. As
2489 // such, we need to special case this to allow it.
2490 if (TermOkay && TI == I)
2492 assert(DT.dominates(I, TI) &&
2493 "basic SSA liveness expectation violated by liveness analysis");
2498 /// Check that all the liveness sets used during the computation of liveness
2499 /// obey basic SSA properties. This is useful for finding cases where we miss
2501 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2503 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2504 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2505 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2509 static void computeLiveInValues(DominatorTree &DT, Function &F,
2510 GCPtrLivenessData &Data) {
2512 SmallSetVector<BasicBlock *, 200> Worklist;
2513 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2514 // We use a SetVector so that we don't have duplicates in the worklist.
2515 Worklist.insert(pred_begin(BB), pred_end(BB));
2517 auto NextItem = [&]() {
2518 BasicBlock *BB = Worklist.back();
2519 Worklist.pop_back();
2523 // Seed the liveness for each individual block
2524 for (BasicBlock &BB : F) {
2525 Data.KillSet[&BB] = computeKillSet(&BB);
2526 Data.LiveSet[&BB].clear();
2527 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2530 for (Value *Kill : Data.KillSet[&BB])
2531 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2534 Data.LiveOut[&BB] = DenseSet<Value *>();
2535 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2536 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2537 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2538 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2539 if (!Data.LiveIn[&BB].empty())
2540 AddPredsToWorklist(&BB);
2543 // Propagate that liveness until stable
2544 while (!Worklist.empty()) {
2545 BasicBlock *BB = NextItem();
2547 // Compute our new liveout set, then exit early if it hasn't changed
2548 // despite the contribution of our successor.
2549 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2550 const auto OldLiveOutSize = LiveOut.size();
2551 for (BasicBlock *Succ : successors(BB)) {
2552 assert(Data.LiveIn.count(Succ));
2553 set_union(LiveOut, Data.LiveIn[Succ]);
2555 // assert OutLiveOut is a subset of LiveOut
2556 if (OldLiveOutSize == LiveOut.size()) {
2557 // If the sets are the same size, then we didn't actually add anything
2558 // when unioning our successors LiveIn Thus, the LiveIn of this block
2562 Data.LiveOut[BB] = LiveOut;
2564 // Apply the effects of this basic block
2565 DenseSet<Value *> LiveTmp = LiveOut;
2566 set_union(LiveTmp, Data.LiveSet[BB]);
2567 set_subtract(LiveTmp, Data.KillSet[BB]);
2569 assert(Data.LiveIn.count(BB));
2570 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2571 // assert: OldLiveIn is a subset of LiveTmp
2572 if (OldLiveIn.size() != LiveTmp.size()) {
2573 Data.LiveIn[BB] = LiveTmp;
2574 AddPredsToWorklist(BB);
2576 } // while( !worklist.empty() )
2579 // Sanity check our ouput against SSA properties. This helps catch any
2580 // missing kills during the above iteration.
2581 for (BasicBlock &BB : F) {
2582 checkBasicSSA(DT, Data, BB);
2587 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2588 StatepointLiveSetTy &Out) {
2590 BasicBlock *BB = Inst->getParent();
2592 // Note: The copy is intentional and required
2593 assert(Data.LiveOut.count(BB));
2594 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2596 // We want to handle the statepoint itself oddly. It's
2597 // call result is not live (normal), nor are it's arguments
2598 // (unless they're used again later). This adjustment is
2599 // specifically what we need to relocate
2600 BasicBlock::reverse_iterator rend(Inst);
2601 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2602 LiveOut.erase(Inst);
2603 Out.insert(LiveOut.begin(), LiveOut.end());
2606 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2608 PartiallyConstructedSafepointRecord &Info) {
2609 Instruction *Inst = CS.getInstruction();
2610 StatepointLiveSetTy Updated;
2611 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2614 DenseSet<Value *> Bases;
2615 for (auto KVPair : Info.PointerToBase) {
2616 Bases.insert(KVPair.second);
2619 // We may have base pointers which are now live that weren't before. We need
2620 // to update the PointerToBase structure to reflect this.
2621 for (auto V : Updated)
2622 if (!Info.PointerToBase.count(V)) {
2623 assert(Bases.count(V) && "can't find base for unexpected live value");
2624 Info.PointerToBase[V] = V;
2629 for (auto V : Updated) {
2630 assert(Info.PointerToBase.count(V) &&
2631 "must be able to find base for live value");
2635 // Remove any stale base mappings - this can happen since our liveness is
2636 // more precise then the one inherent in the base pointer analysis
2637 DenseSet<Value *> ToErase;
2638 for (auto KVPair : Info.PointerToBase)
2639 if (!Updated.count(KVPair.first))
2640 ToErase.insert(KVPair.first);
2641 for (auto V : ToErase)
2642 Info.PointerToBase.erase(V);
2645 for (auto KVPair : Info.PointerToBase)
2646 assert(Updated.count(KVPair.first) && "record for non-live value");
2649 Info.liveset = Updated;