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 auto *const_1 = ConstantInt::get(
853 v->getParent()->getParent()->getParent()->getContext()),
855 auto MDConst = ConstantAsMetadata::get(const_1);
856 MDNode *md = MDNode::get(
857 v->getParent()->getParent()->getParent()->getContext(), MDConst);
858 phi->setMetadata("is_base_value", md);
859 states[v] = PhiState(PhiState::Conflict, phi);
861 SelectInst *sel = cast<SelectInst>(v);
862 // The undef will be replaced later
863 UndefValue *undef = UndefValue::get(sel->getType());
864 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
865 undef, "base_select", sel);
866 // Add metadata marking this as a base value
867 auto *const_1 = ConstantInt::get(
869 v->getParent()->getParent()->getParent()->getContext()),
871 auto MDConst = ConstantAsMetadata::get(const_1);
872 MDNode *md = MDNode::get(
873 v->getParent()->getParent()->getParent()->getContext(), MDConst);
874 basesel->setMetadata("is_base_value", md);
875 states[v] = PhiState(PhiState::Conflict, basesel);
879 // Fixup all the inputs of the new PHIs
880 for (auto Pair : states) {
881 Instruction *v = cast<Instruction>(Pair.first);
882 PhiState state = Pair.second;
884 assert(!isKnownBaseResult(v) && "why did it get added?");
885 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
886 if (!state.isConflict())
889 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
890 PHINode *phi = cast<PHINode>(v);
891 unsigned NumPHIValues = phi->getNumIncomingValues();
892 for (unsigned i = 0; i < NumPHIValues; i++) {
893 Value *InVal = phi->getIncomingValue(i);
894 BasicBlock *InBB = phi->getIncomingBlock(i);
896 // If we've already seen InBB, add the same incoming value
897 // we added for it earlier. The IR verifier requires phi
898 // nodes with multiple entries from the same basic block
899 // to have the same incoming value for each of those
900 // entries. If we don't do this check here and basephi
901 // has a different type than base, we'll end up adding two
902 // bitcasts (and hence two distinct values) as incoming
903 // values for the same basic block.
905 int blockIndex = basephi->getBasicBlockIndex(InBB);
906 if (blockIndex != -1) {
907 Value *oldBase = basephi->getIncomingValue(blockIndex);
908 basephi->addIncoming(oldBase, InBB);
910 Value *base = findBaseOrBDV(InVal, cache);
911 if (!isKnownBaseResult(base)) {
912 // Either conflict or base.
913 assert(states.count(base));
914 base = states[base].getBase();
915 assert(base != nullptr && "unknown PhiState!");
918 // In essense this assert states: the only way two
919 // values incoming from the same basic block may be
920 // different is by being different bitcasts of the same
921 // value. A cleanup that remains TODO is changing
922 // findBaseOrBDV to return an llvm::Value of the correct
923 // type (and still remain pure). This will remove the
924 // need to add bitcasts.
925 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
926 "sanity -- findBaseOrBDV should be pure!");
931 // Find either the defining value for the PHI or the normal base for
933 Value *base = findBaseOrBDV(InVal, cache);
934 if (!isKnownBaseResult(base)) {
935 // Either conflict or base.
936 assert(states.count(base));
937 base = states[base].getBase();
938 assert(base != nullptr && "unknown PhiState!");
940 assert(base && "can't be null");
941 // Must use original input BB since base may not be Instruction
942 // The cast is needed since base traversal may strip away bitcasts
943 if (base->getType() != basephi->getType()) {
944 base = new BitCastInst(base, basephi->getType(), "cast",
945 InBB->getTerminator());
947 basephi->addIncoming(base, InBB);
949 assert(basephi->getNumIncomingValues() == NumPHIValues);
951 SelectInst *basesel = cast<SelectInst>(state.getBase());
952 SelectInst *sel = cast<SelectInst>(v);
953 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
954 // something more safe and less hacky.
955 for (int i = 1; i <= 2; i++) {
956 Value *InVal = sel->getOperand(i);
957 // Find either the defining value for the PHI or the normal base for
959 Value *base = findBaseOrBDV(InVal, cache);
960 if (!isKnownBaseResult(base)) {
961 // Either conflict or base.
962 assert(states.count(base));
963 base = states[base].getBase();
964 assert(base != nullptr && "unknown PhiState!");
966 assert(base && "can't be null");
967 // Must use original input BB since base may not be Instruction
968 // The cast is needed since base traversal may strip away bitcasts
969 if (base->getType() != basesel->getType()) {
970 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
972 basesel->setOperand(i, base);
977 // Cache all of our results so we can cheaply reuse them
978 // NOTE: This is actually two caches: one of the base defining value
979 // relation and one of the base pointer relation! FIXME
980 for (auto item : states) {
981 Value *v = item.first;
982 Value *base = item.second.getBase();
984 assert(!isKnownBaseResult(v) && "why did it get added?");
987 std::string fromstr =
988 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
990 errs() << "Updating base value cache"
991 << " for: " << (v->hasName() ? v->getName() : "")
992 << " from: " << fromstr
993 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
996 assert(isKnownBaseResult(base) &&
997 "must be something we 'know' is a base pointer");
998 if (cache.count(v)) {
999 // Once we transition from the BDV relation being store in the cache to
1000 // the base relation being stored, it must be stable
1001 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
1002 "base relation should be stable");
1006 assert(cache.find(def) != cache.end());
1010 // For a set of live pointers (base and/or derived), identify the base
1011 // pointer of the object which they are derived from. This routine will
1012 // mutate the IR graph as needed to make the 'base' pointer live at the
1013 // definition site of 'derived'. This ensures that any use of 'derived' can
1014 // also use 'base'. This may involve the insertion of a number of
1015 // additional PHI nodes.
1017 // preconditions: live is a set of pointer type Values
1019 // side effects: may insert PHI nodes into the existing CFG, will preserve
1020 // CFG, will not remove or mutate any existing nodes
1022 // post condition: PointerToBase contains one (derived, base) pair for every
1023 // pointer in live. Note that derived can be equal to base if the original
1024 // pointer was a base pointer.
1026 findBasePointers(const StatepointLiveSetTy &live,
1027 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
1028 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1029 // For the naming of values inserted to be deterministic - which makes for
1030 // much cleaner and more stable tests - we need to assign an order to the
1031 // live values. DenseSets do not provide a deterministic order across runs.
1032 SmallVector<Value *, 64> Temp;
1033 Temp.insert(Temp.end(), live.begin(), live.end());
1034 std::sort(Temp.begin(), Temp.end(), order_by_name);
1035 for (Value *ptr : Temp) {
1036 Value *base = findBasePointer(ptr, DVCache);
1037 assert(base && "failed to find base pointer");
1038 PointerToBase[ptr] = base;
1039 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1040 DT->dominates(cast<Instruction>(base)->getParent(),
1041 cast<Instruction>(ptr)->getParent())) &&
1042 "The base we found better dominate the derived pointer");
1044 // If you see this trip and like to live really dangerously, the code should
1045 // be correct, just with idioms the verifier can't handle. You can try
1046 // disabling the verifier at your own substaintial risk.
1047 assert(!isa<ConstantPointerNull>(base) &&
1048 "the relocation code needs adjustment to handle the relocation of "
1049 "a null pointer constant without causing false positives in the "
1050 "safepoint ir verifier.");
1054 /// Find the required based pointers (and adjust the live set) for the given
1056 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1058 PartiallyConstructedSafepointRecord &result) {
1059 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
1060 findBasePointers(result.liveset, PointerToBase, &DT, DVCache);
1062 if (PrintBasePointers) {
1063 // Note: Need to print these in a stable order since this is checked in
1065 errs() << "Base Pairs (w/o Relocation):\n";
1066 SmallVector<Value *, 64> Temp;
1067 Temp.reserve(PointerToBase.size());
1068 for (auto Pair : PointerToBase) {
1069 Temp.push_back(Pair.first);
1071 std::sort(Temp.begin(), Temp.end(), order_by_name);
1072 for (Value *Ptr : Temp) {
1073 Value *Base = PointerToBase[Ptr];
1074 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
1079 result.PointerToBase = PointerToBase;
1082 /// Given an updated version of the dataflow liveness results, update the
1083 /// liveset and base pointer maps for the call site CS.
1084 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1086 PartiallyConstructedSafepointRecord &result);
1088 static void recomputeLiveInValues(
1089 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1090 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1091 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1092 // again. The old values are still live and will help it stablize quickly.
1093 GCPtrLivenessData RevisedLivenessData;
1094 computeLiveInValues(DT, F, RevisedLivenessData);
1095 for (size_t i = 0; i < records.size(); i++) {
1096 struct PartiallyConstructedSafepointRecord &info = records[i];
1097 const CallSite &CS = toUpdate[i];
1098 recomputeLiveInValues(RevisedLivenessData, CS, info);
1102 // When inserting gc.relocate calls, we need to ensure there are no uses
1103 // of the original value between the gc.statepoint and the gc.relocate call.
1104 // One case which can arise is a phi node starting one of the successor blocks.
1105 // We also need to be able to insert the gc.relocates only on the path which
1106 // goes through the statepoint. We might need to split an edge to make this
1109 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1110 DominatorTree &DT) {
1111 BasicBlock *Ret = BB;
1112 if (!BB->getUniquePredecessor()) {
1113 Ret = SplitBlockPredecessors(BB, InvokeParent, "", nullptr, &DT);
1116 // Now that 'ret' has unique predecessor we can safely remove all phi nodes
1118 FoldSingleEntryPHINodes(Ret);
1119 assert(!isa<PHINode>(Ret->begin()));
1121 // At this point, we can safely insert a gc.relocate as the first instruction
1122 // in Ret if needed.
1126 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1127 auto itr = std::find(livevec.begin(), livevec.end(), val);
1128 assert(livevec.end() != itr);
1129 size_t index = std::distance(livevec.begin(), itr);
1130 assert(index < livevec.size());
1134 // Create new attribute set containing only attributes which can be transfered
1135 // from original call to the safepoint.
1136 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1139 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1140 unsigned index = AS.getSlotIndex(Slot);
1142 if (index == AttributeSet::ReturnIndex ||
1143 index == AttributeSet::FunctionIndex) {
1145 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1147 Attribute attr = *it;
1149 // Do not allow certain attributes - just skip them
1150 // Safepoint can not be read only or read none.
1151 if (attr.hasAttribute(Attribute::ReadNone) ||
1152 attr.hasAttribute(Attribute::ReadOnly))
1155 ret = ret.addAttributes(
1156 AS.getContext(), index,
1157 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1161 // Just skip parameter attributes for now
1167 /// Helper function to place all gc relocates necessary for the given
1170 /// liveVariables - list of variables to be relocated.
1171 /// liveStart - index of the first live variable.
1172 /// basePtrs - base pointers.
1173 /// statepointToken - statepoint instruction to which relocates should be
1175 /// Builder - Llvm IR builder to be used to construct new calls.
1176 static void CreateGCRelocates(ArrayRef<llvm::Value *> LiveVariables,
1177 const int LiveStart,
1178 ArrayRef<llvm::Value *> BasePtrs,
1179 Instruction *StatepointToken,
1180 IRBuilder<> Builder) {
1181 SmallVector<Instruction *, 64> NewDefs;
1182 NewDefs.reserve(LiveVariables.size());
1184 Module *M = StatepointToken->getParent()->getParent()->getParent();
1186 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1187 // We generate a (potentially) unique declaration for every pointer type
1188 // combination. This results is some blow up the function declarations in
1189 // the IR, but removes the need for argument bitcasts which shrinks the IR
1190 // greatly and makes it much more readable.
1191 SmallVector<Type *, 1> Types; // one per 'any' type
1192 // All gc_relocate are set to i8 addrspace(1)* type. This could help avoid
1193 // cases where the actual value's type mangling is not supported by llvm. A
1194 // bitcast is added later to convert gc_relocate to the actual value's type.
1195 Types.push_back(Type::getInt8PtrTy(M->getContext(), 1));
1196 Value *GCRelocateDecl = Intrinsic::getDeclaration(
1197 M, Intrinsic::experimental_gc_relocate, Types);
1199 // Generate the gc.relocate call and save the result
1201 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1202 LiveStart + find_index(LiveVariables, BasePtrs[i]));
1203 Value *LiveIdx = ConstantInt::get(
1204 Type::getInt32Ty(M->getContext()),
1205 LiveStart + find_index(LiveVariables, LiveVariables[i]));
1207 // only specify a debug name if we can give a useful one
1208 Value *Reloc = Builder.CreateCall(
1209 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1210 LiveVariables[i]->hasName() ? LiveVariables[i]->getName() + ".relocated"
1212 // Trick CodeGen into thinking there are lots of free registers at this
1214 cast<CallInst>(Reloc)->setCallingConv(CallingConv::Cold);
1216 NewDefs.push_back(cast<Instruction>(Reloc));
1218 assert(NewDefs.size() == LiveVariables.size() &&
1219 "missing or extra redefinition at safepoint");
1223 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1224 const SmallVectorImpl<llvm::Value *> &basePtrs,
1225 const SmallVectorImpl<llvm::Value *> &liveVariables,
1227 PartiallyConstructedSafepointRecord &result) {
1228 assert(basePtrs.size() == liveVariables.size());
1229 assert(isStatepoint(CS) &&
1230 "This method expects to be rewriting a statepoint");
1232 BasicBlock *BB = CS.getInstruction()->getParent();
1234 Function *F = BB->getParent();
1235 assert(F && "must be set");
1236 Module *M = F->getParent();
1238 assert(M && "must be set");
1240 // We're not changing the function signature of the statepoint since the gc
1241 // arguments go into the var args section.
1242 Function *gc_statepoint_decl = CS.getCalledFunction();
1244 // Then go ahead and use the builder do actually do the inserts. We insert
1245 // immediately before the previous instruction under the assumption that all
1246 // arguments will be available here. We can't insert afterwards since we may
1247 // be replacing a terminator.
1248 Instruction *insertBefore = CS.getInstruction();
1249 IRBuilder<> Builder(insertBefore);
1250 // Copy all of the arguments from the original statepoint - this includes the
1251 // target, call args, and deopt args
1252 SmallVector<llvm::Value *, 64> args;
1253 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1254 // TODO: Clear the 'needs rewrite' flag
1256 // add all the pointers to be relocated (gc arguments)
1257 // Capture the start of the live variable list for use in the gc_relocates
1258 const int live_start = args.size();
1259 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1261 // Create the statepoint given all the arguments
1262 Instruction *token = nullptr;
1263 AttributeSet return_attributes;
1265 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1267 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1268 call->setTailCall(toReplace->isTailCall());
1269 call->setCallingConv(toReplace->getCallingConv());
1271 // Currently we will fail on parameter attributes and on certain
1272 // function attributes.
1273 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1274 // In case if we can handle this set of sttributes - set up function attrs
1275 // directly on statepoint and return attrs later for gc_result intrinsic.
1276 call->setAttributes(new_attrs.getFnAttributes());
1277 return_attributes = new_attrs.getRetAttributes();
1281 // Put the following gc_result and gc_relocate calls immediately after the
1282 // the old call (which we're about to delete)
1283 BasicBlock::iterator next(toReplace);
1284 assert(BB->end() != next && "not a terminator, must have next");
1286 Instruction *IP = &*(next);
1287 Builder.SetInsertPoint(IP);
1288 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1291 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1293 // Insert the new invoke into the old block. We'll remove the old one in a
1294 // moment at which point this will become the new terminator for the
1296 InvokeInst *invoke = InvokeInst::Create(
1297 gc_statepoint_decl, toReplace->getNormalDest(),
1298 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1299 invoke->setCallingConv(toReplace->getCallingConv());
1301 // Currently we will fail on parameter attributes and on certain
1302 // function attributes.
1303 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1304 // In case if we can handle this set of sttributes - set up function attrs
1305 // directly on statepoint and return attrs later for gc_result intrinsic.
1306 invoke->setAttributes(new_attrs.getFnAttributes());
1307 return_attributes = new_attrs.getRetAttributes();
1311 // Generate gc relocates in exceptional path
1312 BasicBlock *unwindBlock = toReplace->getUnwindDest();
1313 assert(!isa<PHINode>(unwindBlock->begin()) &&
1314 unwindBlock->getUniquePredecessor() &&
1315 "can't safely insert in this block!");
1317 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1318 Builder.SetInsertPoint(IP);
1319 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1321 // Extract second element from landingpad return value. We will attach
1322 // exceptional gc relocates to it.
1323 const unsigned idx = 1;
1324 Instruction *exceptional_token =
1325 cast<Instruction>(Builder.CreateExtractValue(
1326 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1327 result.UnwindToken = exceptional_token;
1329 // Just throw away return value. We will use the one we got for normal
1331 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1332 exceptional_token, Builder);
1334 // Generate gc relocates and returns for normal block
1335 BasicBlock *normalDest = toReplace->getNormalDest();
1336 assert(!isa<PHINode>(normalDest->begin()) &&
1337 normalDest->getUniquePredecessor() &&
1338 "can't safely insert in this block!");
1340 IP = &*(normalDest->getFirstInsertionPt());
1341 Builder.SetInsertPoint(IP);
1343 // gc relocates will be generated later as if it were regular call
1348 // Take the name of the original value call if it had one.
1349 token->takeName(CS.getInstruction());
1351 // The GCResult is already inserted, we just need to find it
1353 Instruction *toReplace = CS.getInstruction();
1354 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1355 "only valid use before rewrite is gc.result");
1356 assert(!toReplace->hasOneUse() ||
1357 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1360 // Update the gc.result of the original statepoint (if any) to use the newly
1361 // inserted statepoint. This is safe to do here since the token can't be
1362 // considered a live reference.
1363 CS.getInstruction()->replaceAllUsesWith(token);
1365 result.StatepointToken = token;
1367 // Second, create a gc.relocate for every live variable
1368 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1372 struct name_ordering {
1375 bool operator()(name_ordering const &a, name_ordering const &b) {
1376 return -1 == a.derived->getName().compare(b.derived->getName());
1380 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1381 SmallVectorImpl<Value *> &livevec) {
1382 assert(basevec.size() == livevec.size());
1384 SmallVector<name_ordering, 64> temp;
1385 for (size_t i = 0; i < basevec.size(); i++) {
1387 v.base = basevec[i];
1388 v.derived = livevec[i];
1391 std::sort(temp.begin(), temp.end(), name_ordering());
1392 for (size_t i = 0; i < basevec.size(); i++) {
1393 basevec[i] = temp[i].base;
1394 livevec[i] = temp[i].derived;
1398 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1399 // which make the relocations happening at this safepoint explicit.
1401 // WARNING: Does not do any fixup to adjust users of the original live
1402 // values. That's the callers responsibility.
1404 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1405 PartiallyConstructedSafepointRecord &result) {
1406 auto liveset = result.liveset;
1407 auto PointerToBase = result.PointerToBase;
1409 // Convert to vector for efficient cross referencing.
1410 SmallVector<Value *, 64> basevec, livevec;
1411 livevec.reserve(liveset.size());
1412 basevec.reserve(liveset.size());
1413 for (Value *L : liveset) {
1414 livevec.push_back(L);
1416 assert(PointerToBase.find(L) != PointerToBase.end());
1417 Value *base = PointerToBase[L];
1418 basevec.push_back(base);
1420 assert(livevec.size() == basevec.size());
1422 // To make the output IR slightly more stable (for use in diffs), ensure a
1423 // fixed order of the values in the safepoint (by sorting the value name).
1424 // The order is otherwise meaningless.
1425 stablize_order(basevec, livevec);
1427 // Do the actual rewriting and delete the old statepoint
1428 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1429 CS.getInstruction()->eraseFromParent();
1432 // Helper function for the relocationViaAlloca.
1433 // It receives iterator to the statepoint gc relocates and emits store to the
1435 // location (via allocaMap) for the each one of them.
1436 // Add visited values into the visitedLiveValues set we will later use them
1437 // for sanity check.
1439 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1440 DenseMap<Value *, Value *> &AllocaMap,
1441 DenseSet<Value *> &VisitedLiveValues) {
1443 for (User *U : GCRelocs) {
1444 if (!isa<IntrinsicInst>(U))
1447 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1449 // We only care about relocates
1450 if (RelocatedValue->getIntrinsicID() !=
1451 Intrinsic::experimental_gc_relocate) {
1455 GCRelocateOperands RelocateOperands(RelocatedValue);
1456 Value *OriginalValue =
1457 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1458 assert(AllocaMap.count(OriginalValue));
1459 Value *Alloca = AllocaMap[OriginalValue];
1461 // Emit store into the related alloca
1462 // All gc_relocate are i8 addrspace(1)* typed, and it must be bitcasted to
1463 // the correct type according to alloca.
1464 assert(RelocatedValue->getNextNode() && "Should always have one since it's not a terminator");
1465 IRBuilder<> Builder(RelocatedValue->getNextNode());
1466 Value *CastedRelocatedValue =
1467 Builder.CreateBitCast(RelocatedValue, cast<AllocaInst>(Alloca)->getAllocatedType(),
1468 RelocatedValue->hasName() ? RelocatedValue->getName() + ".casted" : "");
1470 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1471 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1474 VisitedLiveValues.insert(OriginalValue);
1479 // Helper function for the "relocationViaAlloca". Similar to the
1480 // "insertRelocationStores" but works for rematerialized values.
1482 insertRematerializationStores(
1483 RematerializedValueMapTy RematerializedValues,
1484 DenseMap<Value *, Value *> &AllocaMap,
1485 DenseSet<Value *> &VisitedLiveValues) {
1487 for (auto RematerializedValuePair: RematerializedValues) {
1488 Instruction *RematerializedValue = RematerializedValuePair.first;
1489 Value *OriginalValue = RematerializedValuePair.second;
1491 assert(AllocaMap.count(OriginalValue) &&
1492 "Can not find alloca for rematerialized value");
1493 Value *Alloca = AllocaMap[OriginalValue];
1495 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1496 Store->insertAfter(RematerializedValue);
1499 VisitedLiveValues.insert(OriginalValue);
1504 /// do all the relocation update via allocas and mem2reg
1505 static void relocationViaAlloca(
1506 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1507 ArrayRef<struct PartiallyConstructedSafepointRecord> Records) {
1509 // record initial number of (static) allocas; we'll check we have the same
1510 // number when we get done.
1511 int InitialAllocaNum = 0;
1512 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1514 if (isa<AllocaInst>(*I))
1518 // TODO-PERF: change data structures, reserve
1519 DenseMap<Value *, Value *> AllocaMap;
1520 SmallVector<AllocaInst *, 200> PromotableAllocas;
1521 // Used later to chack that we have enough allocas to store all values
1522 std::size_t NumRematerializedValues = 0;
1523 PromotableAllocas.reserve(Live.size());
1525 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1526 // "PromotableAllocas"
1527 auto emitAllocaFor = [&](Value *LiveValue) {
1528 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1529 F.getEntryBlock().getFirstNonPHI());
1530 AllocaMap[LiveValue] = Alloca;
1531 PromotableAllocas.push_back(Alloca);
1534 // emit alloca for each live gc pointer
1535 for (unsigned i = 0; i < Live.size(); i++) {
1536 emitAllocaFor(Live[i]);
1539 // emit allocas for rematerialized values
1540 for (size_t i = 0; i < Records.size(); i++) {
1541 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1543 for (auto RematerializedValuePair : Info.RematerializedValues) {
1544 Value *OriginalValue = RematerializedValuePair.second;
1545 if (AllocaMap.count(OriginalValue) != 0)
1548 emitAllocaFor(OriginalValue);
1549 ++NumRematerializedValues;
1553 // The next two loops are part of the same conceptual operation. We need to
1554 // insert a store to the alloca after the original def and at each
1555 // redefinition. We need to insert a load before each use. These are split
1556 // into distinct loops for performance reasons.
1558 // update gc pointer after each statepoint
1559 // either store a relocated value or null (if no relocated value found for
1560 // this gc pointer and it is not a gc_result)
1561 // this must happen before we update the statepoint with load of alloca
1562 // otherwise we lose the link between statepoint and old def
1563 for (size_t i = 0; i < Records.size(); i++) {
1564 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1565 Value *Statepoint = Info.StatepointToken;
1567 // This will be used for consistency check
1568 DenseSet<Value *> VisitedLiveValues;
1570 // Insert stores for normal statepoint gc relocates
1571 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1573 // In case if it was invoke statepoint
1574 // we will insert stores for exceptional path gc relocates.
1575 if (isa<InvokeInst>(Statepoint)) {
1576 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1580 // Do similar thing with rematerialized values
1581 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1584 if (ClobberNonLive) {
1585 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1586 // the gc.statepoint. This will turn some subtle GC problems into
1587 // slightly easier to debug SEGVs. Note that on large IR files with
1588 // lots of gc.statepoints this is extremely costly both memory and time
1590 SmallVector<AllocaInst *, 64> ToClobber;
1591 for (auto Pair : AllocaMap) {
1592 Value *Def = Pair.first;
1593 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1595 // This value was relocated
1596 if (VisitedLiveValues.count(Def)) {
1599 ToClobber.push_back(Alloca);
1602 auto InsertClobbersAt = [&](Instruction *IP) {
1603 for (auto *AI : ToClobber) {
1604 auto AIType = cast<PointerType>(AI->getType());
1605 auto PT = cast<PointerType>(AIType->getElementType());
1606 Constant *CPN = ConstantPointerNull::get(PT);
1607 StoreInst *Store = new StoreInst(CPN, AI);
1608 Store->insertBefore(IP);
1612 // Insert the clobbering stores. These may get intermixed with the
1613 // gc.results and gc.relocates, but that's fine.
1614 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1615 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1616 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1618 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1620 InsertClobbersAt(Next);
1624 // update use with load allocas and add store for gc_relocated
1625 for (auto Pair : AllocaMap) {
1626 Value *Def = Pair.first;
1627 Value *Alloca = Pair.second;
1629 // we pre-record the uses of allocas so that we dont have to worry about
1631 // that change the user information.
1632 SmallVector<Instruction *, 20> Uses;
1633 // PERF: trade a linear scan for repeated reallocation
1634 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1635 for (User *U : Def->users()) {
1636 if (!isa<ConstantExpr>(U)) {
1637 // If the def has a ConstantExpr use, then the def is either a
1638 // ConstantExpr use itself or null. In either case
1639 // (recursively in the first, directly in the second), the oop
1640 // it is ultimately dependent on is null and this particular
1641 // use does not need to be fixed up.
1642 Uses.push_back(cast<Instruction>(U));
1646 std::sort(Uses.begin(), Uses.end());
1647 auto Last = std::unique(Uses.begin(), Uses.end());
1648 Uses.erase(Last, Uses.end());
1650 for (Instruction *Use : Uses) {
1651 if (isa<PHINode>(Use)) {
1652 PHINode *Phi = cast<PHINode>(Use);
1653 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1654 if (Def == Phi->getIncomingValue(i)) {
1655 LoadInst *Load = new LoadInst(
1656 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1657 Phi->setIncomingValue(i, Load);
1661 LoadInst *Load = new LoadInst(Alloca, "", Use);
1662 Use->replaceUsesOfWith(Def, Load);
1666 // emit store for the initial gc value
1667 // store must be inserted after load, otherwise store will be in alloca's
1668 // use list and an extra load will be inserted before it
1669 StoreInst *Store = new StoreInst(Def, Alloca);
1670 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1671 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1672 // InvokeInst is a TerminatorInst so the store need to be inserted
1673 // into its normal destination block.
1674 BasicBlock *NormalDest = Invoke->getNormalDest();
1675 Store->insertBefore(NormalDest->getFirstNonPHI());
1677 assert(!Inst->isTerminator() &&
1678 "The only TerminatorInst that can produce a value is "
1679 "InvokeInst which is handled above.");
1680 Store->insertAfter(Inst);
1683 assert(isa<Argument>(Def));
1684 Store->insertAfter(cast<Instruction>(Alloca));
1688 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1689 "we must have the same allocas with lives");
1690 if (!PromotableAllocas.empty()) {
1691 // apply mem2reg to promote alloca to SSA
1692 PromoteMemToReg(PromotableAllocas, DT);
1696 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1698 if (isa<AllocaInst>(*I))
1700 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1704 /// Implement a unique function which doesn't require we sort the input
1705 /// vector. Doing so has the effect of changing the output of a couple of
1706 /// tests in ways which make them less useful in testing fused safepoints.
1707 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1708 SmallSet<T, 8> Seen;
1709 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1710 return !Seen.insert(V).second;
1714 /// Insert holders so that each Value is obviously live through the entire
1715 /// lifetime of the call.
1716 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1717 SmallVectorImpl<CallInst *> &Holders) {
1719 // No values to hold live, might as well not insert the empty holder
1722 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1723 // Use a dummy vararg function to actually hold the values live
1724 Function *Func = cast<Function>(M->getOrInsertFunction(
1725 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1727 // For call safepoints insert dummy calls right after safepoint
1728 BasicBlock::iterator Next(CS.getInstruction());
1730 Holders.push_back(CallInst::Create(Func, Values, "", Next));
1733 // For invoke safepooints insert dummy calls both in normal and
1734 // exceptional destination blocks
1735 auto *II = cast<InvokeInst>(CS.getInstruction());
1736 Holders.push_back(CallInst::Create(
1737 Func, Values, "", II->getNormalDest()->getFirstInsertionPt()));
1738 Holders.push_back(CallInst::Create(
1739 Func, Values, "", II->getUnwindDest()->getFirstInsertionPt()));
1742 static void findLiveReferences(
1743 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1744 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1745 GCPtrLivenessData OriginalLivenessData;
1746 computeLiveInValues(DT, F, OriginalLivenessData);
1747 for (size_t i = 0; i < records.size(); i++) {
1748 struct PartiallyConstructedSafepointRecord &info = records[i];
1749 const CallSite &CS = toUpdate[i];
1750 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1754 /// Remove any vector of pointers from the liveset by scalarizing them over the
1755 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1756 /// would be preferrable to include the vector in the statepoint itself, but
1757 /// the lowering code currently does not handle that. Extending it would be
1758 /// slightly non-trivial since it requires a format change. Given how rare
1759 /// such cases are (for the moment?) scalarizing is an acceptable comprimise.
1760 static void splitVectorValues(Instruction *StatepointInst,
1761 StatepointLiveSetTy &LiveSet,
1762 DenseMap<Value *, Value *>& PointerToBase,
1763 DominatorTree &DT) {
1764 SmallVector<Value *, 16> ToSplit;
1765 for (Value *V : LiveSet)
1766 if (isa<VectorType>(V->getType()))
1767 ToSplit.push_back(V);
1769 if (ToSplit.empty())
1772 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1774 Function &F = *(StatepointInst->getParent()->getParent());
1776 DenseMap<Value *, AllocaInst *> AllocaMap;
1777 // First is normal return, second is exceptional return (invoke only)
1778 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1779 for (Value *V : ToSplit) {
1780 AllocaInst *Alloca =
1781 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1782 AllocaMap[V] = Alloca;
1784 VectorType *VT = cast<VectorType>(V->getType());
1785 IRBuilder<> Builder(StatepointInst);
1786 SmallVector<Value *, 16> Elements;
1787 for (unsigned i = 0; i < VT->getNumElements(); i++)
1788 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1789 ElementMapping[V] = Elements;
1791 auto InsertVectorReform = [&](Instruction *IP) {
1792 Builder.SetInsertPoint(IP);
1793 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1794 Value *ResultVec = UndefValue::get(VT);
1795 for (unsigned i = 0; i < VT->getNumElements(); i++)
1796 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1797 Builder.getInt32(i));
1801 if (isa<CallInst>(StatepointInst)) {
1802 BasicBlock::iterator Next(StatepointInst);
1804 Instruction *IP = &*(Next);
1805 Replacements[V].first = InsertVectorReform(IP);
1806 Replacements[V].second = nullptr;
1808 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1809 // We've already normalized - check that we don't have shared destination
1811 BasicBlock *NormalDest = Invoke->getNormalDest();
1812 assert(!isa<PHINode>(NormalDest->begin()));
1813 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1814 assert(!isa<PHINode>(UnwindDest->begin()));
1815 // Insert insert element sequences in both successors
1816 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1817 Replacements[V].first = InsertVectorReform(IP);
1818 IP = &*(UnwindDest->getFirstInsertionPt());
1819 Replacements[V].second = InsertVectorReform(IP);
1823 for (Value *V : ToSplit) {
1824 AllocaInst *Alloca = AllocaMap[V];
1826 // Capture all users before we start mutating use lists
1827 SmallVector<Instruction *, 16> Users;
1828 for (User *U : V->users())
1829 Users.push_back(cast<Instruction>(U));
1831 for (Instruction *I : Users) {
1832 if (auto Phi = dyn_cast<PHINode>(I)) {
1833 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1834 if (V == Phi->getIncomingValue(i)) {
1835 LoadInst *Load = new LoadInst(
1836 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1837 Phi->setIncomingValue(i, Load);
1840 LoadInst *Load = new LoadInst(Alloca, "", I);
1841 I->replaceUsesOfWith(V, Load);
1845 // Store the original value and the replacement value into the alloca
1846 StoreInst *Store = new StoreInst(V, Alloca);
1847 if (auto I = dyn_cast<Instruction>(V))
1848 Store->insertAfter(I);
1850 Store->insertAfter(Alloca);
1852 // Normal return for invoke, or call return
1853 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1854 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1855 // Unwind return for invoke only
1856 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1858 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1861 // apply mem2reg to promote alloca to SSA
1862 SmallVector<AllocaInst *, 16> Allocas;
1863 for (Value *V : ToSplit)
1864 Allocas.push_back(AllocaMap[V]);
1865 PromoteMemToReg(Allocas, DT);
1867 // Update our tracking of live pointers and base mappings to account for the
1868 // changes we just made.
1869 for (Value *V : ToSplit) {
1870 auto &Elements = ElementMapping[V];
1873 LiveSet.insert(Elements.begin(), Elements.end());
1874 // We need to update the base mapping as well.
1875 assert(PointerToBase.count(V));
1876 Value *OldBase = PointerToBase[V];
1877 auto &BaseElements = ElementMapping[OldBase];
1878 PointerToBase.erase(V);
1879 assert(Elements.size() == BaseElements.size());
1880 for (unsigned i = 0; i < Elements.size(); i++) {
1881 Value *Elem = Elements[i];
1882 PointerToBase[Elem] = BaseElements[i];
1887 // Helper function for the "rematerializeLiveValues". It walks use chain
1888 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
1889 // values are visited (currently it is GEP's and casts). Returns true if it
1890 // sucessfully reached "BaseValue" and false otherwise.
1891 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
1893 static bool findRematerializableChainToBasePointer(
1894 SmallVectorImpl<Instruction*> &ChainToBase,
1895 Value *CurrentValue, Value *BaseValue) {
1897 // We have found a base value
1898 if (CurrentValue == BaseValue) {
1902 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
1903 ChainToBase.push_back(GEP);
1904 return findRematerializableChainToBasePointer(ChainToBase,
1905 GEP->getPointerOperand(),
1909 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
1910 Value *Def = CI->stripPointerCasts();
1912 // This two checks are basically similar. First one is here for the
1913 // consistency with findBasePointers logic.
1914 assert(!isa<CastInst>(Def) && "not a pointer cast found");
1915 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
1918 ChainToBase.push_back(CI);
1919 return findRematerializableChainToBasePointer(ChainToBase, Def, BaseValue);
1922 // Not supported instruction in the chain
1926 // Helper function for the "rematerializeLiveValues". Compute cost of the use
1927 // chain we are going to rematerialize.
1929 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
1930 TargetTransformInfo &TTI) {
1933 for (Instruction *Instr : Chain) {
1934 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
1935 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
1936 "non noop cast is found during rematerialization");
1938 Type *SrcTy = CI->getOperand(0)->getType();
1939 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
1941 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
1942 // Cost of the address calculation
1943 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
1944 Cost += TTI.getAddressComputationCost(ValTy);
1946 // And cost of the GEP itself
1947 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
1948 // allowed for the external usage)
1949 if (!GEP->hasAllConstantIndices())
1953 llvm_unreachable("unsupported instruciton type during rematerialization");
1960 // From the statepoint liveset pick values that are cheaper to recompute then to
1961 // relocate. Remove this values from the liveset, rematerialize them after
1962 // statepoint and record them in "Info" structure. Note that similar to
1963 // relocated values we don't do any user adjustments here.
1964 static void rematerializeLiveValues(CallSite CS,
1965 PartiallyConstructedSafepointRecord &Info,
1966 TargetTransformInfo &TTI) {
1967 const unsigned int ChainLengthThreshold = 10;
1969 // Record values we are going to delete from this statepoint live set.
1970 // We can not di this in following loop due to iterator invalidation.
1971 SmallVector<Value *, 32> LiveValuesToBeDeleted;
1973 for (Value *LiveValue: Info.liveset) {
1974 // For each live pointer find it's defining chain
1975 SmallVector<Instruction *, 3> ChainToBase;
1976 assert(Info.PointerToBase.find(LiveValue) != Info.PointerToBase.end());
1978 findRematerializableChainToBasePointer(ChainToBase,
1980 Info.PointerToBase[LiveValue]);
1981 // Nothing to do, or chain is too long
1983 ChainToBase.size() == 0 ||
1984 ChainToBase.size() > ChainLengthThreshold)
1987 // Compute cost of this chain
1988 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
1989 // TODO: We can also account for cases when we will be able to remove some
1990 // of the rematerialized values by later optimization passes. I.e if
1991 // we rematerialized several intersecting chains. Or if original values
1992 // don't have any uses besides this statepoint.
1994 // For invokes we need to rematerialize each chain twice - for normal and
1995 // for unwind basic blocks. Model this by multiplying cost by two.
1996 if (CS.isInvoke()) {
1999 // If it's too expensive - skip it
2000 if (Cost >= RematerializationThreshold)
2003 // Remove value from the live set
2004 LiveValuesToBeDeleted.push_back(LiveValue);
2006 // Clone instructions and record them inside "Info" structure
2008 // Walk backwards to visit top-most instructions first
2009 std::reverse(ChainToBase.begin(), ChainToBase.end());
2011 // Utility function which clones all instructions from "ChainToBase"
2012 // and inserts them before "InsertBefore". Returns rematerialized value
2013 // which should be used after statepoint.
2014 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
2015 Instruction *LastClonedValue = nullptr;
2016 Instruction *LastValue = nullptr;
2017 for (Instruction *Instr: ChainToBase) {
2018 // Only GEP's and casts are suported as we need to be careful to not
2019 // introduce any new uses of pointers not in the liveset.
2020 // Note that it's fine to introduce new uses of pointers which were
2021 // otherwise not used after this statepoint.
2022 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2024 Instruction *ClonedValue = Instr->clone();
2025 ClonedValue->insertBefore(InsertBefore);
2026 ClonedValue->setName(Instr->getName() + ".remat");
2028 // If it is not first instruction in the chain then it uses previously
2029 // cloned value. We should update it to use cloned value.
2030 if (LastClonedValue) {
2032 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2034 // Assert that cloned instruction does not use any instructions from
2035 // this chain other than LastClonedValue
2036 for (auto OpValue : ClonedValue->operand_values()) {
2037 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2038 ChainToBase.end() &&
2039 "incorrect use in rematerialization chain");
2044 LastClonedValue = ClonedValue;
2047 assert(LastClonedValue);
2048 return LastClonedValue;
2051 // Different cases for calls and invokes. For invokes we need to clone
2052 // instructions both on normal and unwind path.
2054 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2055 assert(InsertBefore);
2056 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2057 Info.RematerializedValues[RematerializedValue] = LiveValue;
2059 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2061 Instruction *NormalInsertBefore =
2062 Invoke->getNormalDest()->getFirstInsertionPt();
2063 Instruction *UnwindInsertBefore =
2064 Invoke->getUnwindDest()->getFirstInsertionPt();
2066 Instruction *NormalRematerializedValue =
2067 rematerializeChain(NormalInsertBefore);
2068 Instruction *UnwindRematerializedValue =
2069 rematerializeChain(UnwindInsertBefore);
2071 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2072 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2076 // Remove rematerializaed values from the live set
2077 for (auto LiveValue: LiveValuesToBeDeleted) {
2078 Info.liveset.erase(LiveValue);
2082 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
2083 SmallVectorImpl<CallSite> &toUpdate) {
2085 // sanity check the input
2086 std::set<CallSite> uniqued;
2087 uniqued.insert(toUpdate.begin(), toUpdate.end());
2088 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
2090 for (size_t i = 0; i < toUpdate.size(); i++) {
2091 CallSite &CS = toUpdate[i];
2092 assert(CS.getInstruction()->getParent()->getParent() == &F);
2093 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
2097 // When inserting gc.relocates for invokes, we need to be able to insert at
2098 // the top of the successor blocks. See the comment on
2099 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2100 // may restructure the CFG.
2101 for (CallSite CS : toUpdate) {
2104 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
2105 normalizeForInvokeSafepoint(invoke->getNormalDest(), invoke->getParent(),
2107 normalizeForInvokeSafepoint(invoke->getUnwindDest(), invoke->getParent(),
2111 // A list of dummy calls added to the IR to keep various values obviously
2112 // live in the IR. We'll remove all of these when done.
2113 SmallVector<CallInst *, 64> holders;
2115 // Insert a dummy call with all of the arguments to the vm_state we'll need
2116 // for the actual safepoint insertion. This ensures reference arguments in
2117 // the deopt argument list are considered live through the safepoint (and
2118 // thus makes sure they get relocated.)
2119 for (size_t i = 0; i < toUpdate.size(); i++) {
2120 CallSite &CS = toUpdate[i];
2121 Statepoint StatepointCS(CS);
2123 SmallVector<Value *, 64> DeoptValues;
2124 for (Use &U : StatepointCS.vm_state_args()) {
2125 Value *Arg = cast<Value>(&U);
2126 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2127 "support for FCA unimplemented");
2128 if (isHandledGCPointerType(Arg->getType()))
2129 DeoptValues.push_back(Arg);
2131 insertUseHolderAfter(CS, DeoptValues, holders);
2134 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
2135 records.reserve(toUpdate.size());
2136 for (size_t i = 0; i < toUpdate.size(); i++) {
2137 struct PartiallyConstructedSafepointRecord info;
2138 records.push_back(info);
2140 assert(records.size() == toUpdate.size());
2142 // A) Identify all gc pointers which are staticly live at the given call
2144 findLiveReferences(F, DT, P, toUpdate, records);
2146 // B) Find the base pointers for each live pointer
2147 /* scope for caching */ {
2148 // Cache the 'defining value' relation used in the computation and
2149 // insertion of base phis and selects. This ensures that we don't insert
2150 // large numbers of duplicate base_phis.
2151 DefiningValueMapTy DVCache;
2153 for (size_t i = 0; i < records.size(); i++) {
2154 struct PartiallyConstructedSafepointRecord &info = records[i];
2155 CallSite &CS = toUpdate[i];
2156 findBasePointers(DT, DVCache, CS, info);
2158 } // end of cache scope
2160 // The base phi insertion logic (for any safepoint) may have inserted new
2161 // instructions which are now live at some safepoint. The simplest such
2164 // phi a <-- will be a new base_phi here
2165 // safepoint 1 <-- that needs to be live here
2169 // We insert some dummy calls after each safepoint to definitely hold live
2170 // the base pointers which were identified for that safepoint. We'll then
2171 // ask liveness for _every_ base inserted to see what is now live. Then we
2172 // remove the dummy calls.
2173 holders.reserve(holders.size() + records.size());
2174 for (size_t i = 0; i < records.size(); i++) {
2175 struct PartiallyConstructedSafepointRecord &info = records[i];
2176 CallSite &CS = toUpdate[i];
2178 SmallVector<Value *, 128> Bases;
2179 for (auto Pair : info.PointerToBase) {
2180 Bases.push_back(Pair.second);
2182 insertUseHolderAfter(CS, Bases, holders);
2185 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2186 // need to rerun liveness. We may *also* have inserted new defs, but that's
2187 // not the key issue.
2188 recomputeLiveInValues(F, DT, P, toUpdate, records);
2190 if (PrintBasePointers) {
2191 for (size_t i = 0; i < records.size(); i++) {
2192 struct PartiallyConstructedSafepointRecord &info = records[i];
2193 errs() << "Base Pairs: (w/Relocation)\n";
2194 for (auto Pair : info.PointerToBase) {
2195 errs() << " derived %" << Pair.first->getName() << " base %"
2196 << Pair.second->getName() << "\n";
2200 for (size_t i = 0; i < holders.size(); i++) {
2201 holders[i]->eraseFromParent();
2202 holders[i] = nullptr;
2206 // Do a limited scalarization of any live at safepoint vector values which
2207 // contain pointers. This enables this pass to run after vectorization at
2208 // the cost of some possible performance loss. TODO: it would be nice to
2209 // natively support vectors all the way through the backend so we don't need
2210 // to scalarize here.
2211 for (size_t i = 0; i < records.size(); i++) {
2212 struct PartiallyConstructedSafepointRecord &info = records[i];
2213 Instruction *statepoint = toUpdate[i].getInstruction();
2214 splitVectorValues(cast<Instruction>(statepoint), info.liveset,
2215 info.PointerToBase, DT);
2218 // In order to reduce live set of statepoint we might choose to rematerialize
2219 // some values instead of relocating them. This is purelly an optimization and
2220 // does not influence correctness.
2221 TargetTransformInfo &TTI =
2222 P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2224 for (size_t i = 0; i < records.size(); i++) {
2225 struct PartiallyConstructedSafepointRecord &info = records[i];
2226 CallSite &CS = toUpdate[i];
2228 rematerializeLiveValues(CS, info, TTI);
2231 // Now run through and replace the existing statepoints with new ones with
2232 // the live variables listed. We do not yet update uses of the values being
2233 // relocated. We have references to live variables that need to
2234 // survive to the last iteration of this loop. (By construction, the
2235 // previous statepoint can not be a live variable, thus we can and remove
2236 // the old statepoint calls as we go.)
2237 for (size_t i = 0; i < records.size(); i++) {
2238 struct PartiallyConstructedSafepointRecord &info = records[i];
2239 CallSite &CS = toUpdate[i];
2240 makeStatepointExplicit(DT, CS, P, info);
2242 toUpdate.clear(); // prevent accident use of invalid CallSites
2244 // Do all the fixups of the original live variables to their relocated selves
2245 SmallVector<Value *, 128> live;
2246 for (size_t i = 0; i < records.size(); i++) {
2247 struct PartiallyConstructedSafepointRecord &info = records[i];
2248 // We can't simply save the live set from the original insertion. One of
2249 // the live values might be the result of a call which needs a safepoint.
2250 // That Value* no longer exists and we need to use the new gc_result.
2251 // Thankfully, the liveset is embedded in the statepoint (and updated), so
2252 // we just grab that.
2253 Statepoint statepoint(info.StatepointToken);
2254 live.insert(live.end(), statepoint.gc_args_begin(),
2255 statepoint.gc_args_end());
2257 // Do some basic sanity checks on our liveness results before performing
2258 // relocation. Relocation can and will turn mistakes in liveness results
2259 // into non-sensical code which is must harder to debug.
2260 // TODO: It would be nice to test consistency as well
2261 assert(DT.isReachableFromEntry(info.StatepointToken->getParent()) &&
2262 "statepoint must be reachable or liveness is meaningless");
2263 for (Value *V : statepoint.gc_args()) {
2264 if (!isa<Instruction>(V))
2265 // Non-instruction values trivial dominate all possible uses
2267 auto LiveInst = cast<Instruction>(V);
2268 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2269 "unreachable values should never be live");
2270 assert(DT.dominates(LiveInst, info.StatepointToken) &&
2271 "basic SSA liveness expectation violated by liveness analysis");
2275 unique_unsorted(live);
2279 for (auto ptr : live) {
2280 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
2284 relocationViaAlloca(F, DT, live, records);
2285 return !records.empty();
2288 // Handles both return values and arguments for Functions and CallSites.
2289 template <typename AttrHolder>
2290 static void RemoveDerefAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2293 if (AH.getDereferenceableBytes(Index))
2294 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2295 AH.getDereferenceableBytes(Index)));
2296 if (AH.getDereferenceableOrNullBytes(Index))
2297 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2298 AH.getDereferenceableOrNullBytes(Index)));
2301 AH.setAttributes(AH.getAttributes().removeAttributes(
2302 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2306 RewriteStatepointsForGC::stripDereferenceabilityInfoFromPrototype(Function &F) {
2307 LLVMContext &Ctx = F.getContext();
2309 for (Argument &A : F.args())
2310 if (isa<PointerType>(A.getType()))
2311 RemoveDerefAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2313 if (isa<PointerType>(F.getReturnType()))
2314 RemoveDerefAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2317 void RewriteStatepointsForGC::stripDereferenceabilityInfoFromBody(Function &F) {
2321 LLVMContext &Ctx = F.getContext();
2322 MDBuilder Builder(Ctx);
2324 for (Instruction &I : inst_range(F)) {
2325 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2326 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2327 bool IsImmutableTBAA =
2328 MD->getNumOperands() == 4 &&
2329 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2331 if (!IsImmutableTBAA)
2332 continue; // no work to do, MD_tbaa is already marked mutable
2334 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2335 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2337 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2339 MDNode *MutableTBAA =
2340 Builder.createTBAAStructTagNode(Base, Access, Offset);
2341 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2344 if (CallSite CS = CallSite(&I)) {
2345 for (int i = 0, e = CS.arg_size(); i != e; i++)
2346 if (isa<PointerType>(CS.getArgument(i)->getType()))
2347 RemoveDerefAttrAtIndex(Ctx, CS, i + 1);
2348 if (isa<PointerType>(CS.getType()))
2349 RemoveDerefAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2354 /// Returns true if this function should be rewritten by this pass. The main
2355 /// point of this function is as an extension point for custom logic.
2356 static bool shouldRewriteStatepointsIn(Function &F) {
2357 // TODO: This should check the GCStrategy
2359 const char *FunctionGCName = F.getGC();
2360 const StringRef StatepointExampleName("statepoint-example");
2361 const StringRef CoreCLRName("coreclr");
2362 return (StatepointExampleName == FunctionGCName) ||
2363 (CoreCLRName == FunctionGCName);
2368 void RewriteStatepointsForGC::stripDereferenceabilityInfo(Module &M) {
2370 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2374 for (Function &F : M)
2375 stripDereferenceabilityInfoFromPrototype(F);
2377 for (Function &F : M)
2378 stripDereferenceabilityInfoFromBody(F);
2381 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2382 // Nothing to do for declarations.
2383 if (F.isDeclaration() || F.empty())
2386 // Policy choice says not to rewrite - the most common reason is that we're
2387 // compiling code without a GCStrategy.
2388 if (!shouldRewriteStatepointsIn(F))
2391 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2393 // Gather all the statepoints which need rewritten. Be careful to only
2394 // consider those in reachable code since we need to ask dominance queries
2395 // when rewriting. We'll delete the unreachable ones in a moment.
2396 SmallVector<CallSite, 64> ParsePointNeeded;
2397 bool HasUnreachableStatepoint = false;
2398 for (Instruction &I : inst_range(F)) {
2399 // TODO: only the ones with the flag set!
2400 if (isStatepoint(I)) {
2401 if (DT.isReachableFromEntry(I.getParent()))
2402 ParsePointNeeded.push_back(CallSite(&I));
2404 HasUnreachableStatepoint = true;
2408 bool MadeChange = false;
2410 // Delete any unreachable statepoints so that we don't have unrewritten
2411 // statepoints surviving this pass. This makes testing easier and the
2412 // resulting IR less confusing to human readers. Rather than be fancy, we
2413 // just reuse a utility function which removes the unreachable blocks.
2414 if (HasUnreachableStatepoint)
2415 MadeChange |= removeUnreachableBlocks(F);
2417 // Return early if no work to do.
2418 if (ParsePointNeeded.empty())
2421 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2422 // These are created by LCSSA. They have the effect of increasing the size
2423 // of liveness sets for no good reason. It may be harder to do this post
2424 // insertion since relocations and base phis can confuse things.
2425 for (BasicBlock &BB : F)
2426 if (BB.getUniquePredecessor()) {
2428 FoldSingleEntryPHINodes(&BB);
2431 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
2435 // liveness computation via standard dataflow
2436 // -------------------------------------------------------------------
2438 // TODO: Consider using bitvectors for liveness, the set of potentially
2439 // interesting values should be small and easy to pre-compute.
2441 /// Compute the live-in set for the location rbegin starting from
2442 /// the live-out set of the basic block
2443 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2444 BasicBlock::reverse_iterator rend,
2445 DenseSet<Value *> &LiveTmp) {
2447 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2448 Instruction *I = &*ritr;
2450 // KILL/Def - Remove this definition from LiveIn
2453 // Don't consider *uses* in PHI nodes, we handle their contribution to
2454 // predecessor blocks when we seed the LiveOut sets
2455 if (isa<PHINode>(I))
2458 // USE - Add to the LiveIn set for this instruction
2459 for (Value *V : I->operands()) {
2460 assert(!isUnhandledGCPointerType(V->getType()) &&
2461 "support for FCA unimplemented");
2462 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2463 // The choice to exclude all things constant here is slightly subtle.
2464 // There are two idependent reasons:
2465 // - We assume that things which are constant (from LLVM's definition)
2466 // do not move at runtime. For example, the address of a global
2467 // variable is fixed, even though it's contents may not be.
2468 // - Second, we can't disallow arbitrary inttoptr constants even
2469 // if the language frontend does. Optimization passes are free to
2470 // locally exploit facts without respect to global reachability. This
2471 // can create sections of code which are dynamically unreachable and
2472 // contain just about anything. (see constants.ll in tests)
2479 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2481 for (BasicBlock *Succ : successors(BB)) {
2482 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2483 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2484 PHINode *Phi = cast<PHINode>(&*I);
2485 Value *V = Phi->getIncomingValueForBlock(BB);
2486 assert(!isUnhandledGCPointerType(V->getType()) &&
2487 "support for FCA unimplemented");
2488 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2495 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2496 DenseSet<Value *> KillSet;
2497 for (Instruction &I : *BB)
2498 if (isHandledGCPointerType(I.getType()))
2504 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2505 /// sanity check for the liveness computation.
2506 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2507 TerminatorInst *TI, bool TermOkay = false) {
2508 for (Value *V : Live) {
2509 if (auto *I = dyn_cast<Instruction>(V)) {
2510 // The terminator can be a member of the LiveOut set. LLVM's definition
2511 // of instruction dominance states that V does not dominate itself. As
2512 // such, we need to special case this to allow it.
2513 if (TermOkay && TI == I)
2515 assert(DT.dominates(I, TI) &&
2516 "basic SSA liveness expectation violated by liveness analysis");
2521 /// Check that all the liveness sets used during the computation of liveness
2522 /// obey basic SSA properties. This is useful for finding cases where we miss
2524 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2526 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2527 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2528 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2532 static void computeLiveInValues(DominatorTree &DT, Function &F,
2533 GCPtrLivenessData &Data) {
2535 SmallSetVector<BasicBlock *, 200> Worklist;
2536 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2537 // We use a SetVector so that we don't have duplicates in the worklist.
2538 Worklist.insert(pred_begin(BB), pred_end(BB));
2540 auto NextItem = [&]() {
2541 BasicBlock *BB = Worklist.back();
2542 Worklist.pop_back();
2546 // Seed the liveness for each individual block
2547 for (BasicBlock &BB : F) {
2548 Data.KillSet[&BB] = computeKillSet(&BB);
2549 Data.LiveSet[&BB].clear();
2550 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2553 for (Value *Kill : Data.KillSet[&BB])
2554 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2557 Data.LiveOut[&BB] = DenseSet<Value *>();
2558 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2559 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2560 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2561 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2562 if (!Data.LiveIn[&BB].empty())
2563 AddPredsToWorklist(&BB);
2566 // Propagate that liveness until stable
2567 while (!Worklist.empty()) {
2568 BasicBlock *BB = NextItem();
2570 // Compute our new liveout set, then exit early if it hasn't changed
2571 // despite the contribution of our successor.
2572 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2573 const auto OldLiveOutSize = LiveOut.size();
2574 for (BasicBlock *Succ : successors(BB)) {
2575 assert(Data.LiveIn.count(Succ));
2576 set_union(LiveOut, Data.LiveIn[Succ]);
2578 // assert OutLiveOut is a subset of LiveOut
2579 if (OldLiveOutSize == LiveOut.size()) {
2580 // If the sets are the same size, then we didn't actually add anything
2581 // when unioning our successors LiveIn Thus, the LiveIn of this block
2585 Data.LiveOut[BB] = LiveOut;
2587 // Apply the effects of this basic block
2588 DenseSet<Value *> LiveTmp = LiveOut;
2589 set_union(LiveTmp, Data.LiveSet[BB]);
2590 set_subtract(LiveTmp, Data.KillSet[BB]);
2592 assert(Data.LiveIn.count(BB));
2593 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2594 // assert: OldLiveIn is a subset of LiveTmp
2595 if (OldLiveIn.size() != LiveTmp.size()) {
2596 Data.LiveIn[BB] = LiveTmp;
2597 AddPredsToWorklist(BB);
2599 } // while( !worklist.empty() )
2602 // Sanity check our ouput against SSA properties. This helps catch any
2603 // missing kills during the above iteration.
2604 for (BasicBlock &BB : F) {
2605 checkBasicSSA(DT, Data, BB);
2610 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2611 StatepointLiveSetTy &Out) {
2613 BasicBlock *BB = Inst->getParent();
2615 // Note: The copy is intentional and required
2616 assert(Data.LiveOut.count(BB));
2617 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2619 // We want to handle the statepoint itself oddly. It's
2620 // call result is not live (normal), nor are it's arguments
2621 // (unless they're used again later). This adjustment is
2622 // specifically what we need to relocate
2623 BasicBlock::reverse_iterator rend(Inst);
2624 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2625 LiveOut.erase(Inst);
2626 Out.insert(LiveOut.begin(), LiveOut.end());
2629 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2631 PartiallyConstructedSafepointRecord &Info) {
2632 Instruction *Inst = CS.getInstruction();
2633 StatepointLiveSetTy Updated;
2634 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2637 DenseSet<Value *> Bases;
2638 for (auto KVPair : Info.PointerToBase) {
2639 Bases.insert(KVPair.second);
2642 // We may have base pointers which are now live that weren't before. We need
2643 // to update the PointerToBase structure to reflect this.
2644 for (auto V : Updated)
2645 if (!Info.PointerToBase.count(V)) {
2646 assert(Bases.count(V) && "can't find base for unexpected live value");
2647 Info.PointerToBase[V] = V;
2652 for (auto V : Updated) {
2653 assert(Info.PointerToBase.count(V) &&
2654 "must be able to find base for live value");
2658 // Remove any stale base mappings - this can happen since our liveness is
2659 // more precise then the one inherent in the base pointer analysis
2660 DenseSet<Value *> ToErase;
2661 for (auto KVPair : Info.PointerToBase)
2662 if (!Updated.count(KVPair.first))
2663 ToErase.insert(KVPair.first);
2664 for (auto V : ToErase)
2665 Info.PointerToBase.erase(V);
2668 for (auto KVPair : Info.PointerToBase)
2669 assert(Updated.count(KVPair.first) && "record for non-live value");
2672 Info.liveset = Updated;