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/ADT/SetOperations.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/DenseSet.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/IR/BasicBlock.h"
22 #include "llvm/IR/CallSite.h"
23 #include "llvm/IR/Dominators.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/InstIterator.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/Module.h"
31 #include "llvm/IR/Statepoint.h"
32 #include "llvm/IR/Value.h"
33 #include "llvm/IR/Verifier.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Transforms/Scalar.h"
37 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
38 #include "llvm/Transforms/Utils/Cloning.h"
39 #include "llvm/Transforms/Utils/Local.h"
40 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
42 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
46 // Print tracing output
47 static cl::opt<bool> TraceLSP("trace-rewrite-statepoints", cl::Hidden,
50 // Print the liveset found at the insert location
51 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
53 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden,
55 // Print out the base pointers for debugging
56 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden,
60 static bool ClobberNonLive = true;
62 static bool ClobberNonLive = false;
64 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
65 cl::location(ClobberNonLive),
69 struct RewriteStatepointsForGC : public FunctionPass {
70 static char ID; // Pass identification, replacement for typeid
72 RewriteStatepointsForGC() : FunctionPass(ID) {
73 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
75 bool runOnFunction(Function &F) override;
77 void getAnalysisUsage(AnalysisUsage &AU) const override {
78 // We add and rewrite a bunch of instructions, but don't really do much
79 // else. We could in theory preserve a lot more analyses here.
80 AU.addRequired<DominatorTreeWrapperPass>();
85 char RewriteStatepointsForGC::ID = 0;
87 FunctionPass *llvm::createRewriteStatepointsForGCPass() {
88 return new RewriteStatepointsForGC();
91 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
92 "Make relocations explicit at statepoints", false, false)
93 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
94 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
95 "Make relocations explicit at statepoints", false, false)
98 struct GCPtrLivenessData {
99 /// Values defined in this block.
100 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet;
101 /// Values used in this block (and thus live); does not included values
102 /// killed within this block.
103 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet;
105 /// Values live into this basic block (i.e. used by any
106 /// instruction in this basic block or ones reachable from here)
107 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn;
109 /// Values live out of this basic block (i.e. live into
110 /// any successor block)
111 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut;
114 // The type of the internal cache used inside the findBasePointers family
115 // of functions. From the callers perspective, this is an opaque type and
116 // should not be inspected.
118 // In the actual implementation this caches two relations:
119 // - The base relation itself (i.e. this pointer is based on that one)
120 // - The base defining value relation (i.e. before base_phi insertion)
121 // Generally, after the execution of a full findBasePointer call, only the
122 // base relation will remain. Internally, we add a mixture of the two
123 // types, then update all the second type to the first type
124 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
125 typedef DenseSet<llvm::Value *> StatepointLiveSetTy;
127 struct PartiallyConstructedSafepointRecord {
128 /// The set of values known to be live accross this safepoint
129 StatepointLiveSetTy liveset;
131 /// Mapping from live pointers to a base-defining-value
132 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
134 /// The *new* gc.statepoint instruction itself. This produces the token
135 /// that normal path gc.relocates and the gc.result are tied to.
136 Instruction *StatepointToken;
138 /// Instruction to which exceptional gc relocates are attached
139 /// Makes it easier to iterate through them during relocationViaAlloca.
140 Instruction *UnwindToken;
144 /// Compute the live-in set for every basic block in the function
145 static void computeLiveInValues(DominatorTree &DT, Function &F,
146 GCPtrLivenessData &Data);
148 /// Given results from the dataflow liveness computation, find the set of live
149 /// Values at a particular instruction.
150 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
151 StatepointLiveSetTy &out);
153 // TODO: Once we can get to the GCStrategy, this becomes
154 // Optional<bool> isGCManagedPointer(const Value *V) const override {
156 static bool isGCPointerType(const Type *T) {
157 if (const PointerType *PT = dyn_cast<PointerType>(T))
158 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
159 // GC managed heap. We know that a pointer into this heap needs to be
160 // updated and that no other pointer does.
161 return (1 == PT->getAddressSpace());
165 // Return true if this type is one which a) is a gc pointer or contains a GC
166 // pointer and b) is of a type this code expects to encounter as a live value.
167 // (The insertion code will assert that a type which matches (a) and not (b)
168 // is not encountered.)
169 static bool isHandledGCPointerType(Type *T) {
170 // We fully support gc pointers
171 if (isGCPointerType(T))
173 // We partially support vectors of gc pointers. The code will assert if it
174 // can't handle something.
175 if (auto VT = dyn_cast<VectorType>(T))
176 if (isGCPointerType(VT->getElementType()))
182 /// Returns true if this type contains a gc pointer whether we know how to
183 /// handle that type or not.
184 static bool containsGCPtrType(Type *Ty) {
185 if (isGCPointerType(Ty))
187 if (VectorType *VT = dyn_cast<VectorType>(Ty))
188 return isGCPointerType(VT->getScalarType());
189 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
190 return containsGCPtrType(AT->getElementType());
191 if (StructType *ST = dyn_cast<StructType>(Ty))
193 ST->subtypes().begin(), ST->subtypes().end(),
194 [](Type *SubType) { return containsGCPtrType(SubType); });
198 // Returns true if this is a type which a) is a gc pointer or contains a GC
199 // pointer and b) is of a type which the code doesn't expect (i.e. first class
200 // aggregates). Used to trip assertions.
201 static bool isUnhandledGCPointerType(Type *Ty) {
202 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
206 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
207 if (a->hasName() && b->hasName()) {
208 return -1 == a->getName().compare(b->getName());
209 } else if (a->hasName() && !b->hasName()) {
211 } else if (!a->hasName() && b->hasName()) {
214 // Better than nothing, but not stable
219 // Conservatively identifies any definitions which might be live at the
220 // given instruction. The analysis is performed immediately before the
221 // given instruction. Values defined by that instruction are not considered
222 // live. Values used by that instruction are considered live.
223 static void analyzeParsePointLiveness(
224 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
225 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
226 Instruction *inst = CS.getInstruction();
228 StatepointLiveSetTy liveset;
229 findLiveSetAtInst(inst, OriginalLivenessData, liveset);
232 // Note: This output is used by several of the test cases
233 // The order of elemtns in a set is not stable, put them in a vec and sort
235 SmallVector<Value *, 64> temp;
236 temp.insert(temp.end(), liveset.begin(), liveset.end());
237 std::sort(temp.begin(), temp.end(), order_by_name);
238 errs() << "Live Variables:\n";
239 for (Value *V : temp) {
240 errs() << " " << V->getName(); // no newline
244 if (PrintLiveSetSize) {
245 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
246 errs() << "Number live values: " << liveset.size() << "\n";
248 result.liveset = liveset;
251 static Value *findBaseDefiningValue(Value *I);
253 /// If we can trivially determine that the index specified in the given vector
254 /// is a base pointer, return it. In cases where the entire vector is known to
255 /// consist of base pointers, the entire vector will be returned. This
256 /// indicates that the relevant extractelement is a valid base pointer and
257 /// should be used directly.
258 static Value *findBaseOfVector(Value *I, Value *Index) {
259 assert(I->getType()->isVectorTy() &&
260 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
261 "Illegal to ask for the base pointer of a non-pointer type");
263 // Each case parallels findBaseDefiningValue below, see that code for
264 // detailed motivation.
266 if (isa<Argument>(I))
267 // An incoming argument to the function is a base pointer
270 // We shouldn't see the address of a global as a vector value?
271 assert(!isa<GlobalVariable>(I) &&
272 "unexpected global variable found in base of vector");
274 // inlining could possibly introduce phi node that contains
275 // undef if callee has multiple returns
276 if (isa<UndefValue>(I))
277 // utterly meaningless, but useful for dealing with partially optimized
281 // Due to inheritance, this must be _after_ the global variable and undef
283 if (Constant *Con = dyn_cast<Constant>(I)) {
284 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
285 "order of checks wrong!");
286 assert(Con->isNullValue() && "null is the only case which makes sense");
290 if (isa<LoadInst>(I))
293 // For an insert element, we might be able to look through it if we know
294 // something about the indexes, but if the indices are arbitrary values, we
295 // can't without much more extensive scalarization.
296 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(I)) {
297 Value *InsertIndex = IEI->getOperand(2);
298 // This index is inserting the value, look for it's base
299 if (InsertIndex == Index)
300 return findBaseDefiningValue(IEI->getOperand(1));
301 // Both constant, and can't be equal per above. This insert is definitely
302 // not relevant, look back at the rest of the vector and keep trying.
303 if (isa<ConstantInt>(Index) && isa<ConstantInt>(InsertIndex))
304 return findBaseOfVector(IEI->getOperand(0), Index);
307 // Note: This code is currently rather incomplete. We are essentially only
308 // handling cases where the vector element is trivially a base pointer. We
309 // need to update the entire base pointer construction algorithm to know how
310 // to track vector elements and potentially scalarize, but the case which
311 // would motivate the work hasn't shown up in real workloads yet.
312 llvm_unreachable("no base found for vector element");
315 /// Helper function for findBasePointer - Will return a value which either a)
316 /// defines the base pointer for the input or b) blocks the simple search
317 /// (i.e. a PHI or Select of two derived pointers)
318 static Value *findBaseDefiningValue(Value *I) {
319 assert(I->getType()->isPointerTy() &&
320 "Illegal to ask for the base pointer of a non-pointer type");
322 // This case is a bit of a hack - it only handles extracts from vectors which
323 // trivially contain only base pointers or cases where we can directly match
324 // the index of the original extract element to an insertion into the vector.
325 // See note inside the function for how to improve this.
326 if (auto *EEI = dyn_cast<ExtractElementInst>(I)) {
327 Value *VectorOperand = EEI->getVectorOperand();
328 Value *Index = EEI->getIndexOperand();
329 Value *VectorBase = findBaseOfVector(VectorOperand, Index);
330 // If the result returned is a vector, we know the entire vector must
331 // contain base pointers. In that case, the extractelement is a valid base
333 if (VectorBase->getType()->isVectorTy())
335 // Otherwise, we needed to look through the vector to find the base for
336 // this particular element.
337 assert(VectorBase->getType()->isPointerTy());
341 if (isa<Argument>(I))
342 // An incoming argument to the function is a base pointer
343 // We should have never reached here if this argument isn't an gc value
346 if (isa<GlobalVariable>(I))
350 // inlining could possibly introduce phi node that contains
351 // undef if callee has multiple returns
352 if (isa<UndefValue>(I))
353 // utterly meaningless, but useful for dealing with
354 // partially optimized code.
357 // Due to inheritance, this must be _after_ the global variable and undef
359 if (Constant *Con = dyn_cast<Constant>(I)) {
360 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
361 "order of checks wrong!");
362 // Note: Finding a constant base for something marked for relocation
363 // doesn't really make sense. The most likely case is either a) some
364 // screwed up the address space usage or b) your validating against
365 // compiled C++ code w/o the proper separation. The only real exception
366 // is a null pointer. You could have generic code written to index of
367 // off a potentially null value and have proven it null. We also use
368 // null pointers in dead paths of relocation phis (which we might later
369 // want to find a base pointer for).
370 assert(isa<ConstantPointerNull>(Con) &&
371 "null is the only case which makes sense");
375 if (CastInst *CI = dyn_cast<CastInst>(I)) {
376 Value *Def = CI->stripPointerCasts();
377 // If we find a cast instruction here, it means we've found a cast which is
378 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
379 // handle int->ptr conversion.
380 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
381 return findBaseDefiningValue(Def);
384 if (isa<LoadInst>(I))
385 return I; // The value loaded is an gc base itself
387 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
388 // The base of this GEP is the base
389 return findBaseDefiningValue(GEP->getPointerOperand());
391 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
392 switch (II->getIntrinsicID()) {
393 case Intrinsic::experimental_gc_result_ptr:
395 // fall through to general call handling
397 case Intrinsic::experimental_gc_statepoint:
398 case Intrinsic::experimental_gc_result_float:
399 case Intrinsic::experimental_gc_result_int:
400 llvm_unreachable("these don't produce pointers");
401 case Intrinsic::experimental_gc_relocate: {
402 // Rerunning safepoint insertion after safepoints are already
403 // inserted is not supported. It could probably be made to work,
404 // but why are you doing this? There's no good reason.
405 llvm_unreachable("repeat safepoint insertion is not supported");
407 case Intrinsic::gcroot:
408 // Currently, this mechanism hasn't been extended to work with gcroot.
409 // There's no reason it couldn't be, but I haven't thought about the
410 // implications much.
412 "interaction with the gcroot mechanism is not supported");
415 // We assume that functions in the source language only return base
416 // pointers. This should probably be generalized via attributes to support
417 // both source language and internal functions.
418 if (isa<CallInst>(I) || isa<InvokeInst>(I))
421 // I have absolutely no idea how to implement this part yet. It's not
422 // neccessarily hard, I just haven't really looked at it yet.
423 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
425 if (isa<AtomicCmpXchgInst>(I))
426 // A CAS is effectively a atomic store and load combined under a
427 // predicate. From the perspective of base pointers, we just treat it
431 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
432 "binary ops which don't apply to pointers");
434 // The aggregate ops. Aggregates can either be in the heap or on the
435 // stack, but in either case, this is simply a field load. As a result,
436 // this is a defining definition of the base just like a load is.
437 if (isa<ExtractValueInst>(I))
440 // We should never see an insert vector since that would require we be
441 // tracing back a struct value not a pointer value.
442 assert(!isa<InsertValueInst>(I) &&
443 "Base pointer for a struct is meaningless");
445 // The last two cases here don't return a base pointer. Instead, they
446 // return a value which dynamically selects from amoung several base
447 // derived pointers (each with it's own base potentially). It's the job of
448 // the caller to resolve these.
449 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
450 "missing instruction case in findBaseDefiningValing");
454 /// Returns the base defining value for this value.
455 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
456 Value *&Cached = Cache[I];
458 Cached = findBaseDefiningValue(I);
460 assert(Cache[I] != nullptr);
463 dbgs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
469 /// Return a base pointer for this value if known. Otherwise, return it's
470 /// base defining value.
471 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
472 Value *Def = findBaseDefiningValueCached(I, Cache);
473 auto Found = Cache.find(Def);
474 if (Found != Cache.end()) {
475 // Either a base-of relation, or a self reference. Caller must check.
476 return Found->second;
478 // Only a BDV available
482 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
483 /// is it known to be a base pointer? Or do we need to continue searching.
484 static bool isKnownBaseResult(Value *V) {
485 if (!isa<PHINode>(V) && !isa<SelectInst>(V)) {
486 // no recursion possible
489 if (isa<Instruction>(V) &&
490 cast<Instruction>(V)->getMetadata("is_base_value")) {
491 // This is a previously inserted base phi or select. We know
492 // that this is a base value.
496 // We need to keep searching
500 // TODO: find a better name for this
504 enum Status { Unknown, Base, Conflict };
506 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
507 assert(status != Base || b);
509 PhiState(Value *b) : status(Base), base(b) {}
510 PhiState() : status(Unknown), base(nullptr) {}
512 Status getStatus() const { return status; }
513 Value *getBase() const { return base; }
515 bool isBase() const { return getStatus() == Base; }
516 bool isUnknown() const { return getStatus() == Unknown; }
517 bool isConflict() const { return getStatus() == Conflict; }
519 bool operator==(const PhiState &other) const {
520 return base == other.base && status == other.status;
523 bool operator!=(const PhiState &other) const { return !(*this == other); }
526 errs() << status << " (" << base << " - "
527 << (base ? base->getName() : "nullptr") << "): ";
532 Value *base; // non null only if status == base
535 typedef DenseMap<Value *, PhiState> ConflictStateMapTy;
536 // Values of type PhiState form a lattice, and this is a helper
537 // class that implementes the meet operation. The meat of the meet
538 // operation is implemented in MeetPhiStates::pureMeet
539 class MeetPhiStates {
541 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
542 explicit MeetPhiStates(const ConflictStateMapTy &phiStates)
543 : phiStates(phiStates) {}
545 // Destructively meet the current result with the base V. V can
546 // either be a merge instruction (SelectInst / PHINode), in which
547 // case its status is looked up in the phiStates map; or a regular
548 // SSA value, in which case it is assumed to be a base.
549 void meetWith(Value *V) {
550 PhiState otherState = getStateForBDV(V);
551 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
552 MeetPhiStates::pureMeet(currentResult, otherState)) &&
553 "math is wrong: meet does not commute!");
554 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
557 PhiState getResult() const { return currentResult; }
560 const ConflictStateMapTy &phiStates;
561 PhiState currentResult;
563 /// Return a phi state for a base defining value. We'll generate a new
564 /// base state for known bases and expect to find a cached state otherwise
565 PhiState getStateForBDV(Value *baseValue) {
566 if (isKnownBaseResult(baseValue)) {
567 return PhiState(baseValue);
569 return lookupFromMap(baseValue);
573 PhiState lookupFromMap(Value *V) {
574 auto I = phiStates.find(V);
575 assert(I != phiStates.end() && "lookup failed!");
579 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
580 switch (stateA.getStatus()) {
581 case PhiState::Unknown:
585 assert(stateA.getBase() && "can't be null");
586 if (stateB.isUnknown())
589 if (stateB.isBase()) {
590 if (stateA.getBase() == stateB.getBase()) {
591 assert(stateA == stateB && "equality broken!");
594 return PhiState(PhiState::Conflict);
596 assert(stateB.isConflict() && "only three states!");
597 return PhiState(PhiState::Conflict);
599 case PhiState::Conflict:
602 llvm_unreachable("only three states!");
606 /// For a given value or instruction, figure out what base ptr it's derived
607 /// from. For gc objects, this is simply itself. On success, returns a value
608 /// which is the base pointer. (This is reliable and can be used for
609 /// relocation.) On failure, returns nullptr.
610 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
611 Value *def = findBaseOrBDV(I, cache);
613 if (isKnownBaseResult(def)) {
617 // Here's the rough algorithm:
618 // - For every SSA value, construct a mapping to either an actual base
619 // pointer or a PHI which obscures the base pointer.
620 // - Construct a mapping from PHI to unknown TOP state. Use an
621 // optimistic algorithm to propagate base pointer information. Lattice
626 // When algorithm terminates, all PHIs will either have a single concrete
627 // base or be in a conflict state.
628 // - For every conflict, insert a dummy PHI node without arguments. Add
629 // these to the base[Instruction] = BasePtr mapping. For every
630 // non-conflict, add the actual base.
631 // - For every conflict, add arguments for the base[a] of each input
634 // Note: A simpler form of this would be to add the conflict form of all
635 // PHIs without running the optimistic algorithm. This would be
636 // analougous to pessimistic data flow and would likely lead to an
637 // overall worse solution.
639 ConflictStateMapTy states;
640 states[def] = PhiState();
641 // Recursively fill in all phis & selects reachable from the initial one
642 // for which we don't already know a definite base value for
643 // TODO: This should be rewritten with a worklist
647 // Since we're adding elements to 'states' as we run, we can't keep
648 // iterators into the set.
649 SmallVector<Value *, 16> Keys;
650 Keys.reserve(states.size());
651 for (auto Pair : states) {
652 Value *V = Pair.first;
655 for (Value *v : Keys) {
656 assert(!isKnownBaseResult(v) && "why did it get added?");
657 if (PHINode *phi = dyn_cast<PHINode>(v)) {
658 assert(phi->getNumIncomingValues() > 0 &&
659 "zero input phis are illegal");
660 for (Value *InVal : phi->incoming_values()) {
661 Value *local = findBaseOrBDV(InVal, cache);
662 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
663 states[local] = PhiState();
667 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
668 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
669 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
670 states[local] = PhiState();
673 local = findBaseOrBDV(sel->getFalseValue(), cache);
674 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
675 states[local] = PhiState();
683 errs() << "States after initialization:\n";
684 for (auto Pair : states) {
685 Instruction *v = cast<Instruction>(Pair.first);
686 PhiState state = Pair.second;
692 // TODO: come back and revisit the state transitions around inputs which
693 // have reached conflict state. The current version seems too conservative.
695 bool progress = true;
698 size_t oldSize = states.size();
701 // We're only changing keys in this loop, thus safe to keep iterators
702 for (auto Pair : states) {
703 MeetPhiStates calculateMeet(states);
704 Value *v = Pair.first;
705 assert(!isKnownBaseResult(v) && "why did it get added?");
706 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
707 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
708 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
710 for (Value *Val : cast<PHINode>(v)->incoming_values())
711 calculateMeet.meetWith(findBaseOrBDV(Val, cache));
713 PhiState oldState = states[v];
714 PhiState newState = calculateMeet.getResult();
715 if (oldState != newState) {
717 states[v] = newState;
721 assert(oldSize <= states.size());
722 assert(oldSize == states.size() || progress);
726 errs() << "States after meet iteration:\n";
727 for (auto Pair : states) {
728 Instruction *v = cast<Instruction>(Pair.first);
729 PhiState state = Pair.second;
735 // Insert Phis for all conflicts
736 // We want to keep naming deterministic in the loop that follows, so
737 // sort the keys before iteration. This is useful in allowing us to
738 // write stable tests. Note that there is no invalidation issue here.
739 SmallVector<Value *, 16> Keys;
740 Keys.reserve(states.size());
741 for (auto Pair : states) {
742 Value *V = Pair.first;
745 std::sort(Keys.begin(), Keys.end(), order_by_name);
746 // TODO: adjust naming patterns to avoid this order of iteration dependency
747 for (Value *V : Keys) {
748 Instruction *v = cast<Instruction>(V);
749 PhiState state = states[V];
750 assert(!isKnownBaseResult(v) && "why did it get added?");
751 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
752 if (!state.isConflict())
755 if (isa<PHINode>(v)) {
757 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
758 assert(num_preds > 0 && "how did we reach here");
759 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
760 // Add metadata marking this as a base value
761 auto *const_1 = ConstantInt::get(
763 v->getParent()->getParent()->getParent()->getContext()),
765 auto MDConst = ConstantAsMetadata::get(const_1);
766 MDNode *md = MDNode::get(
767 v->getParent()->getParent()->getParent()->getContext(), MDConst);
768 phi->setMetadata("is_base_value", md);
769 states[v] = PhiState(PhiState::Conflict, phi);
771 SelectInst *sel = cast<SelectInst>(v);
772 // The undef will be replaced later
773 UndefValue *undef = UndefValue::get(sel->getType());
774 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
775 undef, "base_select", sel);
776 // Add metadata marking this as a base value
777 auto *const_1 = ConstantInt::get(
779 v->getParent()->getParent()->getParent()->getContext()),
781 auto MDConst = ConstantAsMetadata::get(const_1);
782 MDNode *md = MDNode::get(
783 v->getParent()->getParent()->getParent()->getContext(), MDConst);
784 basesel->setMetadata("is_base_value", md);
785 states[v] = PhiState(PhiState::Conflict, basesel);
789 // Fixup all the inputs of the new PHIs
790 for (auto Pair : states) {
791 Instruction *v = cast<Instruction>(Pair.first);
792 PhiState state = Pair.second;
794 assert(!isKnownBaseResult(v) && "why did it get added?");
795 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
796 if (!state.isConflict())
799 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
800 PHINode *phi = cast<PHINode>(v);
801 unsigned NumPHIValues = phi->getNumIncomingValues();
802 for (unsigned i = 0; i < NumPHIValues; i++) {
803 Value *InVal = phi->getIncomingValue(i);
804 BasicBlock *InBB = phi->getIncomingBlock(i);
806 // If we've already seen InBB, add the same incoming value
807 // we added for it earlier. The IR verifier requires phi
808 // nodes with multiple entries from the same basic block
809 // to have the same incoming value for each of those
810 // entries. If we don't do this check here and basephi
811 // has a different type than base, we'll end up adding two
812 // bitcasts (and hence two distinct values) as incoming
813 // values for the same basic block.
815 int blockIndex = basephi->getBasicBlockIndex(InBB);
816 if (blockIndex != -1) {
817 Value *oldBase = basephi->getIncomingValue(blockIndex);
818 basephi->addIncoming(oldBase, InBB);
820 Value *base = findBaseOrBDV(InVal, cache);
821 if (!isKnownBaseResult(base)) {
822 // Either conflict or base.
823 assert(states.count(base));
824 base = states[base].getBase();
825 assert(base != nullptr && "unknown PhiState!");
828 // In essense this assert states: the only way two
829 // values incoming from the same basic block may be
830 // different is by being different bitcasts of the same
831 // value. A cleanup that remains TODO is changing
832 // findBaseOrBDV to return an llvm::Value of the correct
833 // type (and still remain pure). This will remove the
834 // need to add bitcasts.
835 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
836 "sanity -- findBaseOrBDV should be pure!");
841 // Find either the defining value for the PHI or the normal base for
843 Value *base = findBaseOrBDV(InVal, cache);
844 if (!isKnownBaseResult(base)) {
845 // Either conflict or base.
846 assert(states.count(base));
847 base = states[base].getBase();
848 assert(base != nullptr && "unknown PhiState!");
850 assert(base && "can't be null");
851 // Must use original input BB since base may not be Instruction
852 // The cast is needed since base traversal may strip away bitcasts
853 if (base->getType() != basephi->getType()) {
854 base = new BitCastInst(base, basephi->getType(), "cast",
855 InBB->getTerminator());
857 basephi->addIncoming(base, InBB);
859 assert(basephi->getNumIncomingValues() == NumPHIValues);
861 SelectInst *basesel = cast<SelectInst>(state.getBase());
862 SelectInst *sel = cast<SelectInst>(v);
863 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
864 // something more safe and less hacky.
865 for (int i = 1; i <= 2; i++) {
866 Value *InVal = sel->getOperand(i);
867 // Find either the defining value for the PHI or the normal base for
869 Value *base = findBaseOrBDV(InVal, cache);
870 if (!isKnownBaseResult(base)) {
871 // Either conflict or base.
872 assert(states.count(base));
873 base = states[base].getBase();
874 assert(base != nullptr && "unknown PhiState!");
876 assert(base && "can't be null");
877 // Must use original input BB since base may not be Instruction
878 // The cast is needed since base traversal may strip away bitcasts
879 if (base->getType() != basesel->getType()) {
880 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
882 basesel->setOperand(i, base);
887 // Cache all of our results so we can cheaply reuse them
888 // NOTE: This is actually two caches: one of the base defining value
889 // relation and one of the base pointer relation! FIXME
890 for (auto item : states) {
891 Value *v = item.first;
892 Value *base = item.second.getBase();
894 assert(!isKnownBaseResult(v) && "why did it get added?");
897 std::string fromstr =
898 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
900 errs() << "Updating base value cache"
901 << " for: " << (v->hasName() ? v->getName() : "")
902 << " from: " << fromstr
903 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
906 assert(isKnownBaseResult(base) &&
907 "must be something we 'know' is a base pointer");
908 if (cache.count(v)) {
909 // Once we transition from the BDV relation being store in the cache to
910 // the base relation being stored, it must be stable
911 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
912 "base relation should be stable");
916 assert(cache.find(def) != cache.end());
920 // For a set of live pointers (base and/or derived), identify the base
921 // pointer of the object which they are derived from. This routine will
922 // mutate the IR graph as needed to make the 'base' pointer live at the
923 // definition site of 'derived'. This ensures that any use of 'derived' can
924 // also use 'base'. This may involve the insertion of a number of
925 // additional PHI nodes.
927 // preconditions: live is a set of pointer type Values
929 // side effects: may insert PHI nodes into the existing CFG, will preserve
930 // CFG, will not remove or mutate any existing nodes
932 // post condition: PointerToBase contains one (derived, base) pair for every
933 // pointer in live. Note that derived can be equal to base if the original
934 // pointer was a base pointer.
936 findBasePointers(const StatepointLiveSetTy &live,
937 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
938 DominatorTree *DT, DefiningValueMapTy &DVCache) {
939 // For the naming of values inserted to be deterministic - which makes for
940 // much cleaner and more stable tests - we need to assign an order to the
941 // live values. DenseSets do not provide a deterministic order across runs.
942 SmallVector<Value *, 64> Temp;
943 Temp.insert(Temp.end(), live.begin(), live.end());
944 std::sort(Temp.begin(), Temp.end(), order_by_name);
945 for (Value *ptr : Temp) {
946 Value *base = findBasePointer(ptr, DVCache);
947 assert(base && "failed to find base pointer");
948 PointerToBase[ptr] = base;
949 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
950 DT->dominates(cast<Instruction>(base)->getParent(),
951 cast<Instruction>(ptr)->getParent())) &&
952 "The base we found better dominate the derived pointer");
954 // If you see this trip and like to live really dangerously, the code should
955 // be correct, just with idioms the verifier can't handle. You can try
956 // disabling the verifier at your own substaintial risk.
957 assert(!isa<ConstantPointerNull>(base) &&
958 "the relocation code needs adjustment to handle the relocation of "
959 "a null pointer constant without causing false positives in the "
960 "safepoint ir verifier.");
964 /// Find the required based pointers (and adjust the live set) for the given
966 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
968 PartiallyConstructedSafepointRecord &result) {
969 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
970 findBasePointers(result.liveset, PointerToBase, &DT, DVCache);
972 if (PrintBasePointers) {
973 // Note: Need to print these in a stable order since this is checked in
975 errs() << "Base Pairs (w/o Relocation):\n";
976 SmallVector<Value *, 64> Temp;
977 Temp.reserve(PointerToBase.size());
978 for (auto Pair : PointerToBase) {
979 Temp.push_back(Pair.first);
981 std::sort(Temp.begin(), Temp.end(), order_by_name);
982 for (Value *Ptr : Temp) {
983 Value *Base = PointerToBase[Ptr];
984 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
989 result.PointerToBase = PointerToBase;
992 /// Given an updated version of the dataflow liveness results, update the
993 /// liveset and base pointer maps for the call site CS.
994 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
996 PartiallyConstructedSafepointRecord &result);
998 static void recomputeLiveInValues(
999 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1000 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1001 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1002 // again. The old values are still live and will help it stablize quickly.
1003 GCPtrLivenessData RevisedLivenessData;
1004 computeLiveInValues(DT, F, RevisedLivenessData);
1005 for (size_t i = 0; i < records.size(); i++) {
1006 struct PartiallyConstructedSafepointRecord &info = records[i];
1007 const CallSite &CS = toUpdate[i];
1008 recomputeLiveInValues(RevisedLivenessData, CS, info);
1012 // When inserting gc.relocate calls, we need to ensure there are no uses
1013 // of the original value between the gc.statepoint and the gc.relocate call.
1014 // One case which can arise is a phi node starting one of the successor blocks.
1015 // We also need to be able to insert the gc.relocates only on the path which
1016 // goes through the statepoint. We might need to split an edge to make this
1019 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent, Pass *P) {
1020 DominatorTree *DT = nullptr;
1021 if (auto *DTP = P->getAnalysisIfAvailable<DominatorTreeWrapperPass>())
1022 DT = &DTP->getDomTree();
1024 BasicBlock *Ret = BB;
1025 if (!BB->getUniquePredecessor()) {
1026 Ret = SplitBlockPredecessors(BB, InvokeParent, "", nullptr, DT);
1029 // Now that 'ret' has unique predecessor we can safely remove all phi nodes
1031 FoldSingleEntryPHINodes(Ret);
1032 assert(!isa<PHINode>(Ret->begin()));
1034 // At this point, we can safely insert a gc.relocate as the first instruction
1035 // in Ret if needed.
1039 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1040 auto itr = std::find(livevec.begin(), livevec.end(), val);
1041 assert(livevec.end() != itr);
1042 size_t index = std::distance(livevec.begin(), itr);
1043 assert(index < livevec.size());
1047 // Create new attribute set containing only attributes which can be transfered
1048 // from original call to the safepoint.
1049 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1052 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1053 unsigned index = AS.getSlotIndex(Slot);
1055 if (index == AttributeSet::ReturnIndex ||
1056 index == AttributeSet::FunctionIndex) {
1058 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1060 Attribute attr = *it;
1062 // Do not allow certain attributes - just skip them
1063 // Safepoint can not be read only or read none.
1064 if (attr.hasAttribute(Attribute::ReadNone) ||
1065 attr.hasAttribute(Attribute::ReadOnly))
1068 ret = ret.addAttributes(
1069 AS.getContext(), index,
1070 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1074 // Just skip parameter attributes for now
1080 /// Helper function to place all gc relocates necessary for the given
1083 /// liveVariables - list of variables to be relocated.
1084 /// liveStart - index of the first live variable.
1085 /// basePtrs - base pointers.
1086 /// statepointToken - statepoint instruction to which relocates should be
1088 /// Builder - Llvm IR builder to be used to construct new calls.
1089 static void CreateGCRelocates(ArrayRef<llvm::Value *> LiveVariables,
1090 const int LiveStart,
1091 ArrayRef<llvm::Value *> BasePtrs,
1092 Instruction *StatepointToken,
1093 IRBuilder<> Builder) {
1094 SmallVector<Instruction *, 64> NewDefs;
1095 NewDefs.reserve(LiveVariables.size());
1097 Module *M = StatepointToken->getParent()->getParent()->getParent();
1099 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1100 // We generate a (potentially) unique declaration for every pointer type
1101 // combination. This results is some blow up the function declarations in
1102 // the IR, but removes the need for argument bitcasts which shrinks the IR
1103 // greatly and makes it much more readable.
1104 SmallVector<Type *, 1> Types; // one per 'any' type
1105 // All gc_relocate are set to i8 addrspace(1)* type. This could help avoid
1106 // cases where the actual value's type mangling is not supported by llvm. A
1107 // bitcast is added later to convert gc_relocate to the actual value's type.
1108 Types.push_back(Type::getInt8PtrTy(M->getContext(), 1));
1109 Value *GCRelocateDecl = Intrinsic::getDeclaration(
1110 M, Intrinsic::experimental_gc_relocate, Types);
1112 // Generate the gc.relocate call and save the result
1114 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1115 LiveStart + find_index(LiveVariables, BasePtrs[i]));
1116 Value *LiveIdx = ConstantInt::get(
1117 Type::getInt32Ty(M->getContext()),
1118 LiveStart + find_index(LiveVariables, LiveVariables[i]));
1120 // only specify a debug name if we can give a useful one
1121 Value *Reloc = Builder.CreateCall3(
1122 GCRelocateDecl, StatepointToken, BaseIdx, LiveIdx,
1123 LiveVariables[i]->hasName() ? LiveVariables[i]->getName() + ".relocated"
1125 // Trick CodeGen into thinking there are lots of free registers at this
1127 cast<CallInst>(Reloc)->setCallingConv(CallingConv::Cold);
1129 NewDefs.push_back(cast<Instruction>(Reloc));
1131 assert(NewDefs.size() == LiveVariables.size() &&
1132 "missing or extra redefinition at safepoint");
1136 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1137 const SmallVectorImpl<llvm::Value *> &basePtrs,
1138 const SmallVectorImpl<llvm::Value *> &liveVariables,
1140 PartiallyConstructedSafepointRecord &result) {
1141 assert(basePtrs.size() == liveVariables.size());
1142 assert(isStatepoint(CS) &&
1143 "This method expects to be rewriting a statepoint");
1145 BasicBlock *BB = CS.getInstruction()->getParent();
1147 Function *F = BB->getParent();
1148 assert(F && "must be set");
1149 Module *M = F->getParent();
1151 assert(M && "must be set");
1153 // We're not changing the function signature of the statepoint since the gc
1154 // arguments go into the var args section.
1155 Function *gc_statepoint_decl = CS.getCalledFunction();
1157 // Then go ahead and use the builder do actually do the inserts. We insert
1158 // immediately before the previous instruction under the assumption that all
1159 // arguments will be available here. We can't insert afterwards since we may
1160 // be replacing a terminator.
1161 Instruction *insertBefore = CS.getInstruction();
1162 IRBuilder<> Builder(insertBefore);
1163 // Copy all of the arguments from the original statepoint - this includes the
1164 // target, call args, and deopt args
1165 SmallVector<llvm::Value *, 64> args;
1166 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1167 // TODO: Clear the 'needs rewrite' flag
1169 // add all the pointers to be relocated (gc arguments)
1170 // Capture the start of the live variable list for use in the gc_relocates
1171 const int live_start = args.size();
1172 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1174 // Create the statepoint given all the arguments
1175 Instruction *token = nullptr;
1176 AttributeSet return_attributes;
1178 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1180 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1181 call->setTailCall(toReplace->isTailCall());
1182 call->setCallingConv(toReplace->getCallingConv());
1184 // Currently we will fail on parameter attributes and on certain
1185 // function attributes.
1186 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1187 // In case if we can handle this set of sttributes - set up function attrs
1188 // directly on statepoint and return attrs later for gc_result intrinsic.
1189 call->setAttributes(new_attrs.getFnAttributes());
1190 return_attributes = new_attrs.getRetAttributes();
1194 // Put the following gc_result and gc_relocate calls immediately after the
1195 // the old call (which we're about to delete)
1196 BasicBlock::iterator next(toReplace);
1197 assert(BB->end() != next && "not a terminator, must have next");
1199 Instruction *IP = &*(next);
1200 Builder.SetInsertPoint(IP);
1201 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1204 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1206 // Insert the new invoke into the old block. We'll remove the old one in a
1207 // moment at which point this will become the new terminator for the
1209 InvokeInst *invoke = InvokeInst::Create(
1210 gc_statepoint_decl, toReplace->getNormalDest(),
1211 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1212 invoke->setCallingConv(toReplace->getCallingConv());
1214 // Currently we will fail on parameter attributes and on certain
1215 // function attributes.
1216 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1217 // In case if we can handle this set of sttributes - set up function attrs
1218 // directly on statepoint and return attrs later for gc_result intrinsic.
1219 invoke->setAttributes(new_attrs.getFnAttributes());
1220 return_attributes = new_attrs.getRetAttributes();
1224 // Generate gc relocates in exceptional path
1225 BasicBlock *unwindBlock = toReplace->getUnwindDest();
1226 assert(!isa<PHINode>(unwindBlock->begin()) &&
1227 unwindBlock->getUniquePredecessor() &&
1228 "can't safely insert in this block!");
1230 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1231 Builder.SetInsertPoint(IP);
1232 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1234 // Extract second element from landingpad return value. We will attach
1235 // exceptional gc relocates to it.
1236 const unsigned idx = 1;
1237 Instruction *exceptional_token =
1238 cast<Instruction>(Builder.CreateExtractValue(
1239 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1240 result.UnwindToken = exceptional_token;
1242 // Just throw away return value. We will use the one we got for normal
1244 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1245 exceptional_token, Builder);
1247 // Generate gc relocates and returns for normal block
1248 BasicBlock *normalDest = toReplace->getNormalDest();
1249 assert(!isa<PHINode>(normalDest->begin()) &&
1250 normalDest->getUniquePredecessor() &&
1251 "can't safely insert in this block!");
1253 IP = &*(normalDest->getFirstInsertionPt());
1254 Builder.SetInsertPoint(IP);
1256 // gc relocates will be generated later as if it were regular call
1261 // Take the name of the original value call if it had one.
1262 token->takeName(CS.getInstruction());
1264 // The GCResult is already inserted, we just need to find it
1266 Instruction *toReplace = CS.getInstruction();
1267 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1268 "only valid use before rewrite is gc.result");
1269 assert(!toReplace->hasOneUse() ||
1270 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1273 // Update the gc.result of the original statepoint (if any) to use the newly
1274 // inserted statepoint. This is safe to do here since the token can't be
1275 // considered a live reference.
1276 CS.getInstruction()->replaceAllUsesWith(token);
1278 result.StatepointToken = token;
1280 // Second, create a gc.relocate for every live variable
1281 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1285 struct name_ordering {
1288 bool operator()(name_ordering const &a, name_ordering const &b) {
1289 return -1 == a.derived->getName().compare(b.derived->getName());
1293 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1294 SmallVectorImpl<Value *> &livevec) {
1295 assert(basevec.size() == livevec.size());
1297 SmallVector<name_ordering, 64> temp;
1298 for (size_t i = 0; i < basevec.size(); i++) {
1300 v.base = basevec[i];
1301 v.derived = livevec[i];
1304 std::sort(temp.begin(), temp.end(), name_ordering());
1305 for (size_t i = 0; i < basevec.size(); i++) {
1306 basevec[i] = temp[i].base;
1307 livevec[i] = temp[i].derived;
1311 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1312 // which make the relocations happening at this safepoint explicit.
1314 // WARNING: Does not do any fixup to adjust users of the original live
1315 // values. That's the callers responsibility.
1317 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1318 PartiallyConstructedSafepointRecord &result) {
1319 auto liveset = result.liveset;
1320 auto PointerToBase = result.PointerToBase;
1322 // Convert to vector for efficient cross referencing.
1323 SmallVector<Value *, 64> basevec, livevec;
1324 livevec.reserve(liveset.size());
1325 basevec.reserve(liveset.size());
1326 for (Value *L : liveset) {
1327 livevec.push_back(L);
1329 assert(PointerToBase.find(L) != PointerToBase.end());
1330 Value *base = PointerToBase[L];
1331 basevec.push_back(base);
1333 assert(livevec.size() == basevec.size());
1335 // To make the output IR slightly more stable (for use in diffs), ensure a
1336 // fixed order of the values in the safepoint (by sorting the value name).
1337 // The order is otherwise meaningless.
1338 stablize_order(basevec, livevec);
1340 // Do the actual rewriting and delete the old statepoint
1341 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1342 CS.getInstruction()->eraseFromParent();
1345 // Helper function for the relocationViaAlloca.
1346 // It receives iterator to the statepoint gc relocates and emits store to the
1348 // location (via allocaMap) for the each one of them.
1349 // Add visited values into the visitedLiveValues set we will later use them
1350 // for sanity check.
1352 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1353 DenseMap<Value *, Value *> &AllocaMap,
1354 DenseSet<Value *> &VisitedLiveValues) {
1356 for (User *U : GCRelocs) {
1357 if (!isa<IntrinsicInst>(U))
1360 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1362 // We only care about relocates
1363 if (RelocatedValue->getIntrinsicID() !=
1364 Intrinsic::experimental_gc_relocate) {
1368 GCRelocateOperands RelocateOperands(RelocatedValue);
1369 Value *OriginalValue =
1370 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1371 assert(AllocaMap.count(OriginalValue));
1372 Value *Alloca = AllocaMap[OriginalValue];
1374 // Emit store into the related alloca
1375 // All gc_relocate are i8 addrspace(1)* typed, and it must be bitcasted to
1376 // the correct type according to alloca.
1377 assert(RelocatedValue->getNextNode() && "Should always have one since it's not a terminator");
1378 IRBuilder<> Builder(RelocatedValue->getNextNode());
1379 Value *CastedRelocatedValue =
1380 Builder.CreateBitCast(RelocatedValue, cast<AllocaInst>(Alloca)->getAllocatedType(),
1381 RelocatedValue->hasName() ? RelocatedValue->getName() + ".casted" : "");
1383 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1384 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1387 VisitedLiveValues.insert(OriginalValue);
1392 /// do all the relocation update via allocas and mem2reg
1393 static void relocationViaAlloca(
1394 Function &F, DominatorTree &DT, ArrayRef<Value *> live,
1395 ArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1397 // record initial number of (static) allocas; we'll check we have the same
1398 // number when we get done.
1399 int InitialAllocaNum = 0;
1400 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1402 if (isa<AllocaInst>(*I))
1406 // TODO-PERF: change data structures, reserve
1407 DenseMap<Value *, Value *> allocaMap;
1408 SmallVector<AllocaInst *, 200> PromotableAllocas;
1409 PromotableAllocas.reserve(live.size());
1411 // emit alloca for each live gc pointer
1412 for (unsigned i = 0; i < live.size(); i++) {
1413 Value *liveValue = live[i];
1414 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1415 F.getEntryBlock().getFirstNonPHI());
1416 allocaMap[liveValue] = alloca;
1417 PromotableAllocas.push_back(alloca);
1420 // The next two loops are part of the same conceptual operation. We need to
1421 // insert a store to the alloca after the original def and at each
1422 // redefinition. We need to insert a load before each use. These are split
1423 // into distinct loops for performance reasons.
1425 // update gc pointer after each statepoint
1426 // either store a relocated value or null (if no relocated value found for
1427 // this gc pointer and it is not a gc_result)
1428 // this must happen before we update the statepoint with load of alloca
1429 // otherwise we lose the link between statepoint and old def
1430 for (size_t i = 0; i < records.size(); i++) {
1431 const struct PartiallyConstructedSafepointRecord &info = records[i];
1432 Value *Statepoint = info.StatepointToken;
1434 // This will be used for consistency check
1435 DenseSet<Value *> visitedLiveValues;
1437 // Insert stores for normal statepoint gc relocates
1438 insertRelocationStores(Statepoint->users(), allocaMap, visitedLiveValues);
1440 // In case if it was invoke statepoint
1441 // we will insert stores for exceptional path gc relocates.
1442 if (isa<InvokeInst>(Statepoint)) {
1443 insertRelocationStores(info.UnwindToken->users(), allocaMap,
1447 if (ClobberNonLive) {
1448 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1449 // the gc.statepoint. This will turn some subtle GC problems into
1450 // slightly easier to debug SEGVs. Note that on large IR files with
1451 // lots of gc.statepoints this is extremely costly both memory and time
1453 SmallVector<AllocaInst *, 64> ToClobber;
1454 for (auto Pair : allocaMap) {
1455 Value *Def = Pair.first;
1456 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1458 // This value was relocated
1459 if (visitedLiveValues.count(Def)) {
1462 ToClobber.push_back(Alloca);
1465 auto InsertClobbersAt = [&](Instruction *IP) {
1466 for (auto *AI : ToClobber) {
1467 auto AIType = cast<PointerType>(AI->getType());
1468 auto PT = cast<PointerType>(AIType->getElementType());
1469 Constant *CPN = ConstantPointerNull::get(PT);
1470 StoreInst *store = new StoreInst(CPN, AI);
1471 store->insertBefore(IP);
1475 // Insert the clobbering stores. These may get intermixed with the
1476 // gc.results and gc.relocates, but that's fine.
1477 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1478 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1479 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1481 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1483 InsertClobbersAt(Next);
1487 // update use with load allocas and add store for gc_relocated
1488 for (auto Pair : allocaMap) {
1489 Value *def = Pair.first;
1490 Value *alloca = Pair.second;
1492 // we pre-record the uses of allocas so that we dont have to worry about
1494 // that change the user information.
1495 SmallVector<Instruction *, 20> uses;
1496 // PERF: trade a linear scan for repeated reallocation
1497 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1498 for (User *U : def->users()) {
1499 if (!isa<ConstantExpr>(U)) {
1500 // If the def has a ConstantExpr use, then the def is either a
1501 // ConstantExpr use itself or null. In either case
1502 // (recursively in the first, directly in the second), the oop
1503 // it is ultimately dependent on is null and this particular
1504 // use does not need to be fixed up.
1505 uses.push_back(cast<Instruction>(U));
1509 std::sort(uses.begin(), uses.end());
1510 auto last = std::unique(uses.begin(), uses.end());
1511 uses.erase(last, uses.end());
1513 for (Instruction *use : uses) {
1514 if (isa<PHINode>(use)) {
1515 PHINode *phi = cast<PHINode>(use);
1516 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1517 if (def == phi->getIncomingValue(i)) {
1518 LoadInst *load = new LoadInst(
1519 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1520 phi->setIncomingValue(i, load);
1524 LoadInst *load = new LoadInst(alloca, "", use);
1525 use->replaceUsesOfWith(def, load);
1529 // emit store for the initial gc value
1530 // store must be inserted after load, otherwise store will be in alloca's
1531 // use list and an extra load will be inserted before it
1532 StoreInst *store = new StoreInst(def, alloca);
1533 if (Instruction *inst = dyn_cast<Instruction>(def)) {
1534 if (InvokeInst *invoke = dyn_cast<InvokeInst>(inst)) {
1535 // InvokeInst is a TerminatorInst so the store need to be inserted
1536 // into its normal destination block.
1537 BasicBlock *normalDest = invoke->getNormalDest();
1538 store->insertBefore(normalDest->getFirstNonPHI());
1540 assert(!inst->isTerminator() &&
1541 "The only TerminatorInst that can produce a value is "
1542 "InvokeInst which is handled above.");
1543 store->insertAfter(inst);
1546 assert(isa<Argument>(def));
1547 store->insertAfter(cast<Instruction>(alloca));
1551 assert(PromotableAllocas.size() == live.size() &&
1552 "we must have the same allocas with lives");
1553 if (!PromotableAllocas.empty()) {
1554 // apply mem2reg to promote alloca to SSA
1555 PromoteMemToReg(PromotableAllocas, DT);
1559 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1561 if (isa<AllocaInst>(*I))
1563 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1567 /// Implement a unique function which doesn't require we sort the input
1568 /// vector. Doing so has the effect of changing the output of a couple of
1569 /// tests in ways which make them less useful in testing fused safepoints.
1570 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1572 SmallVector<T, 128> TempVec;
1573 TempVec.reserve(Vec.size());
1574 for (auto Element : Vec)
1575 TempVec.push_back(Element);
1577 for (auto V : TempVec) {
1578 if (Seen.insert(V).second) {
1584 /// Insert holders so that each Value is obviously live through the entire
1585 /// lifetime of the call.
1586 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1587 SmallVectorImpl<CallInst *> &Holders) {
1589 // No values to hold live, might as well not insert the empty holder
1592 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1593 // Use a dummy vararg function to actually hold the values live
1594 Function *Func = cast<Function>(M->getOrInsertFunction(
1595 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1597 // For call safepoints insert dummy calls right after safepoint
1598 BasicBlock::iterator Next(CS.getInstruction());
1600 Holders.push_back(CallInst::Create(Func, Values, "", Next));
1603 // For invoke safepooints insert dummy calls both in normal and
1604 // exceptional destination blocks
1605 auto *II = cast<InvokeInst>(CS.getInstruction());
1606 Holders.push_back(CallInst::Create(
1607 Func, Values, "", II->getNormalDest()->getFirstInsertionPt()));
1608 Holders.push_back(CallInst::Create(
1609 Func, Values, "", II->getUnwindDest()->getFirstInsertionPt()));
1612 static void findLiveReferences(
1613 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1614 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1615 GCPtrLivenessData OriginalLivenessData;
1616 computeLiveInValues(DT, F, OriginalLivenessData);
1617 for (size_t i = 0; i < records.size(); i++) {
1618 struct PartiallyConstructedSafepointRecord &info = records[i];
1619 const CallSite &CS = toUpdate[i];
1620 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1624 /// Remove any vector of pointers from the liveset by scalarizing them over the
1625 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1626 /// would be preferrable to include the vector in the statepoint itself, but
1627 /// the lowering code currently does not handle that. Extending it would be
1628 /// slightly non-trivial since it requires a format change. Given how rare
1629 /// such cases are (for the moment?) scalarizing is an acceptable comprimise.
1630 static void splitVectorValues(Instruction *StatepointInst,
1631 StatepointLiveSetTy &LiveSet, DominatorTree &DT) {
1632 SmallVector<Value *, 16> ToSplit;
1633 for (Value *V : LiveSet)
1634 if (isa<VectorType>(V->getType()))
1635 ToSplit.push_back(V);
1637 if (ToSplit.empty())
1640 Function &F = *(StatepointInst->getParent()->getParent());
1642 DenseMap<Value *, AllocaInst *> AllocaMap;
1643 // First is normal return, second is exceptional return (invoke only)
1644 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1645 for (Value *V : ToSplit) {
1648 AllocaInst *Alloca =
1649 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1650 AllocaMap[V] = Alloca;
1652 VectorType *VT = cast<VectorType>(V->getType());
1653 IRBuilder<> Builder(StatepointInst);
1654 SmallVector<Value *, 16> Elements;
1655 for (unsigned i = 0; i < VT->getNumElements(); i++)
1656 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1657 LiveSet.insert(Elements.begin(), Elements.end());
1659 auto InsertVectorReform = [&](Instruction *IP) {
1660 Builder.SetInsertPoint(IP);
1661 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1662 Value *ResultVec = UndefValue::get(VT);
1663 for (unsigned i = 0; i < VT->getNumElements(); i++)
1664 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1665 Builder.getInt32(i));
1669 if (isa<CallInst>(StatepointInst)) {
1670 BasicBlock::iterator Next(StatepointInst);
1672 Instruction *IP = &*(Next);
1673 Replacements[V].first = InsertVectorReform(IP);
1674 Replacements[V].second = nullptr;
1676 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1677 // We've already normalized - check that we don't have shared destination
1679 BasicBlock *NormalDest = Invoke->getNormalDest();
1680 assert(!isa<PHINode>(NormalDest->begin()));
1681 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1682 assert(!isa<PHINode>(UnwindDest->begin()));
1683 // Insert insert element sequences in both successors
1684 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1685 Replacements[V].first = InsertVectorReform(IP);
1686 IP = &*(UnwindDest->getFirstInsertionPt());
1687 Replacements[V].second = InsertVectorReform(IP);
1690 for (Value *V : ToSplit) {
1691 AllocaInst *Alloca = AllocaMap[V];
1693 // Capture all users before we start mutating use lists
1694 SmallVector<Instruction *, 16> Users;
1695 for (User *U : V->users())
1696 Users.push_back(cast<Instruction>(U));
1698 for (Instruction *I : Users) {
1699 if (auto Phi = dyn_cast<PHINode>(I)) {
1700 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1701 if (V == Phi->getIncomingValue(i)) {
1702 LoadInst *Load = new LoadInst(
1703 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1704 Phi->setIncomingValue(i, Load);
1707 LoadInst *Load = new LoadInst(Alloca, "", I);
1708 I->replaceUsesOfWith(V, Load);
1712 // Store the original value and the replacement value into the alloca
1713 StoreInst *Store = new StoreInst(V, Alloca);
1714 if (auto I = dyn_cast<Instruction>(V))
1715 Store->insertAfter(I);
1717 Store->insertAfter(Alloca);
1719 // Normal return for invoke, or call return
1720 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1721 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1722 // Unwind return for invoke only
1723 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1725 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1728 // apply mem2reg to promote alloca to SSA
1729 SmallVector<AllocaInst *, 16> Allocas;
1730 for (Value *V : ToSplit)
1731 Allocas.push_back(AllocaMap[V]);
1732 PromoteMemToReg(Allocas, DT);
1735 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1736 SmallVectorImpl<CallSite> &toUpdate) {
1738 // sanity check the input
1739 std::set<CallSite> uniqued;
1740 uniqued.insert(toUpdate.begin(), toUpdate.end());
1741 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1743 for (size_t i = 0; i < toUpdate.size(); i++) {
1744 CallSite &CS = toUpdate[i];
1745 assert(CS.getInstruction()->getParent()->getParent() == &F);
1746 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1750 // When inserting gc.relocates for invokes, we need to be able to insert at
1751 // the top of the successor blocks. See the comment on
1752 // normalForInvokeSafepoint on exactly what is needed. Note that this step
1753 // may restructure the CFG.
1754 for (CallSite CS : toUpdate) {
1757 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1758 normalizeForInvokeSafepoint(invoke->getNormalDest(), invoke->getParent(),
1760 normalizeForInvokeSafepoint(invoke->getUnwindDest(), invoke->getParent(),
1764 // A list of dummy calls added to the IR to keep various values obviously
1765 // live in the IR. We'll remove all of these when done.
1766 SmallVector<CallInst *, 64> holders;
1768 // Insert a dummy call with all of the arguments to the vm_state we'll need
1769 // for the actual safepoint insertion. This ensures reference arguments in
1770 // the deopt argument list are considered live through the safepoint (and
1771 // thus makes sure they get relocated.)
1772 for (size_t i = 0; i < toUpdate.size(); i++) {
1773 CallSite &CS = toUpdate[i];
1774 Statepoint StatepointCS(CS);
1776 SmallVector<Value *, 64> DeoptValues;
1777 for (Use &U : StatepointCS.vm_state_args()) {
1778 Value *Arg = cast<Value>(&U);
1779 assert(!isUnhandledGCPointerType(Arg->getType()) &&
1780 "support for FCA unimplemented");
1781 if (isHandledGCPointerType(Arg->getType()))
1782 DeoptValues.push_back(Arg);
1784 insertUseHolderAfter(CS, DeoptValues, holders);
1787 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
1788 records.reserve(toUpdate.size());
1789 for (size_t i = 0; i < toUpdate.size(); i++) {
1790 struct PartiallyConstructedSafepointRecord info;
1791 records.push_back(info);
1793 assert(records.size() == toUpdate.size());
1795 // A) Identify all gc pointers which are staticly live at the given call
1797 findLiveReferences(F, DT, P, toUpdate, records);
1799 // Do a limited scalarization of any live at safepoint vector values which
1800 // contain pointers. This enables this pass to run after vectorization at
1801 // the cost of some possible performance loss. TODO: it would be nice to
1802 // natively support vectors all the way through the backend so we don't need
1803 // to scalarize here.
1804 for (size_t i = 0; i < records.size(); i++) {
1805 struct PartiallyConstructedSafepointRecord &info = records[i];
1806 Instruction *statepoint = toUpdate[i].getInstruction();
1807 splitVectorValues(cast<Instruction>(statepoint), info.liveset, DT);
1810 // B) Find the base pointers for each live pointer
1811 /* scope for caching */ {
1812 // Cache the 'defining value' relation used in the computation and
1813 // insertion of base phis and selects. This ensures that we don't insert
1814 // large numbers of duplicate base_phis.
1815 DefiningValueMapTy DVCache;
1817 for (size_t i = 0; i < records.size(); i++) {
1818 struct PartiallyConstructedSafepointRecord &info = records[i];
1819 CallSite &CS = toUpdate[i];
1820 findBasePointers(DT, DVCache, CS, info);
1822 } // end of cache scope
1824 // The base phi insertion logic (for any safepoint) may have inserted new
1825 // instructions which are now live at some safepoint. The simplest such
1828 // phi a <-- will be a new base_phi here
1829 // safepoint 1 <-- that needs to be live here
1833 // We insert some dummy calls after each safepoint to definitely hold live
1834 // the base pointers which were identified for that safepoint. We'll then
1835 // ask liveness for _every_ base inserted to see what is now live. Then we
1836 // remove the dummy calls.
1837 holders.reserve(holders.size() + records.size());
1838 for (size_t i = 0; i < records.size(); i++) {
1839 struct PartiallyConstructedSafepointRecord &info = records[i];
1840 CallSite &CS = toUpdate[i];
1842 SmallVector<Value *, 128> Bases;
1843 for (auto Pair : info.PointerToBase) {
1844 Bases.push_back(Pair.second);
1846 insertUseHolderAfter(CS, Bases, holders);
1849 // By selecting base pointers, we've effectively inserted new uses. Thus, we
1850 // need to rerun liveness. We may *also* have inserted new defs, but that's
1851 // not the key issue.
1852 recomputeLiveInValues(F, DT, P, toUpdate, records);
1854 if (PrintBasePointers) {
1855 for (size_t i = 0; i < records.size(); i++) {
1856 struct PartiallyConstructedSafepointRecord &info = records[i];
1857 errs() << "Base Pairs: (w/Relocation)\n";
1858 for (auto Pair : info.PointerToBase) {
1859 errs() << " derived %" << Pair.first->getName() << " base %"
1860 << Pair.second->getName() << "\n";
1864 for (size_t i = 0; i < holders.size(); i++) {
1865 holders[i]->eraseFromParent();
1866 holders[i] = nullptr;
1870 // Now run through and replace the existing statepoints with new ones with
1871 // the live variables listed. We do not yet update uses of the values being
1872 // relocated. We have references to live variables that need to
1873 // survive to the last iteration of this loop. (By construction, the
1874 // previous statepoint can not be a live variable, thus we can and remove
1875 // the old statepoint calls as we go.)
1876 for (size_t i = 0; i < records.size(); i++) {
1877 struct PartiallyConstructedSafepointRecord &info = records[i];
1878 CallSite &CS = toUpdate[i];
1879 makeStatepointExplicit(DT, CS, P, info);
1881 toUpdate.clear(); // prevent accident use of invalid CallSites
1883 // Do all the fixups of the original live variables to their relocated selves
1884 SmallVector<Value *, 128> live;
1885 for (size_t i = 0; i < records.size(); i++) {
1886 struct PartiallyConstructedSafepointRecord &info = records[i];
1887 // We can't simply save the live set from the original insertion. One of
1888 // the live values might be the result of a call which needs a safepoint.
1889 // That Value* no longer exists and we need to use the new gc_result.
1890 // Thankfully, the liveset is embedded in the statepoint (and updated), so
1891 // we just grab that.
1892 Statepoint statepoint(info.StatepointToken);
1893 live.insert(live.end(), statepoint.gc_args_begin(),
1894 statepoint.gc_args_end());
1896 // Do some basic sanity checks on our liveness results before performing
1897 // relocation. Relocation can and will turn mistakes in liveness results
1898 // into non-sensical code which is must harder to debug.
1899 // TODO: It would be nice to test consistency as well
1900 assert(DT.isReachableFromEntry(info.StatepointToken->getParent()) &&
1901 "statepoint must be reachable or liveness is meaningless");
1902 for (Value *V : statepoint.gc_args()) {
1903 if (!isa<Instruction>(V))
1904 // Non-instruction values trivial dominate all possible uses
1906 auto LiveInst = cast<Instruction>(V);
1907 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
1908 "unreachable values should never be live");
1909 assert(DT.dominates(LiveInst, info.StatepointToken) &&
1910 "basic SSA liveness expectation violated by liveness analysis");
1914 unique_unsorted(live);
1918 for (auto ptr : live) {
1919 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
1923 relocationViaAlloca(F, DT, live, records);
1924 return !records.empty();
1927 /// Returns true if this function should be rewritten by this pass. The main
1928 /// point of this function is as an extension point for custom logic.
1929 static bool shouldRewriteStatepointsIn(Function &F) {
1930 // TODO: This should check the GCStrategy
1932 const std::string StatepointExampleName("statepoint-example");
1933 return StatepointExampleName == F.getGC();
1938 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
1939 // Nothing to do for declarations.
1940 if (F.isDeclaration() || F.empty())
1943 // Policy choice says not to rewrite - the most common reason is that we're
1944 // compiling code without a GCStrategy.
1945 if (!shouldRewriteStatepointsIn(F))
1948 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1950 // Gather all the statepoints which need rewritten. Be careful to only
1951 // consider those in reachable code since we need to ask dominance queries
1952 // when rewriting. We'll delete the unreachable ones in a moment.
1953 SmallVector<CallSite, 64> ParsePointNeeded;
1954 bool HasUnreachableStatepoint = false;
1955 for (Instruction &I : inst_range(F)) {
1956 // TODO: only the ones with the flag set!
1957 if (isStatepoint(I)) {
1958 if (DT.isReachableFromEntry(I.getParent()))
1959 ParsePointNeeded.push_back(CallSite(&I));
1961 HasUnreachableStatepoint = true;
1965 bool MadeChange = false;
1967 // Delete any unreachable statepoints so that we don't have unrewritten
1968 // statepoints surviving this pass. This makes testing easier and the
1969 // resulting IR less confusing to human readers. Rather than be fancy, we
1970 // just reuse a utility function which removes the unreachable blocks.
1971 if (HasUnreachableStatepoint)
1972 MadeChange |= removeUnreachableBlocks(F);
1974 // Return early if no work to do.
1975 if (ParsePointNeeded.empty())
1978 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
1979 // These are created by LCSSA. They have the effect of increasing the size
1980 // of liveness sets for no good reason. It may be harder to do this post
1981 // insertion since relocations and base phis can confuse things.
1982 for (BasicBlock &BB : F)
1983 if (BB.getUniquePredecessor()) {
1985 FoldSingleEntryPHINodes(&BB);
1988 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
1992 // liveness computation via standard dataflow
1993 // -------------------------------------------------------------------
1995 // TODO: Consider using bitvectors for liveness, the set of potentially
1996 // interesting values should be small and easy to pre-compute.
1998 /// Compute the live-in set for the location rbegin starting from
1999 /// the live-out set of the basic block
2000 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2001 BasicBlock::reverse_iterator rend,
2002 DenseSet<Value *> &LiveTmp) {
2004 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2005 Instruction *I = &*ritr;
2007 // KILL/Def - Remove this definition from LiveIn
2010 // Don't consider *uses* in PHI nodes, we handle their contribution to
2011 // predecessor blocks when we seed the LiveOut sets
2012 if (isa<PHINode>(I))
2015 // USE - Add to the LiveIn set for this instruction
2016 for (Value *V : I->operands()) {
2017 assert(!isUnhandledGCPointerType(V->getType()) &&
2018 "support for FCA unimplemented");
2019 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2020 // The choice to exclude all things constant here is slightly subtle.
2021 // There are two idependent reasons:
2022 // - We assume that things which are constant (from LLVM's definition)
2023 // do not move at runtime. For example, the address of a global
2024 // variable is fixed, even though it's contents may not be.
2025 // - Second, we can't disallow arbitrary inttoptr constants even
2026 // if the language frontend does. Optimization passes are free to
2027 // locally exploit facts without respect to global reachability. This
2028 // can create sections of code which are dynamically unreachable and
2029 // contain just about anything. (see constants.ll in tests)
2036 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2038 for (BasicBlock *Succ : successors(BB)) {
2039 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2040 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2041 PHINode *Phi = cast<PHINode>(&*I);
2042 Value *V = Phi->getIncomingValueForBlock(BB);
2043 assert(!isUnhandledGCPointerType(V->getType()) &&
2044 "support for FCA unimplemented");
2045 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2052 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2053 DenseSet<Value *> KillSet;
2054 for (Instruction &I : *BB)
2055 if (isHandledGCPointerType(I.getType()))
2061 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2062 /// sanity check for the liveness computation.
2063 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2064 TerminatorInst *TI, bool TermOkay = false) {
2065 for (Value *V : Live) {
2066 if (auto *I = dyn_cast<Instruction>(V)) {
2067 // The terminator can be a member of the LiveOut set. LLVM's definition
2068 // of instruction dominance states that V does not dominate itself. As
2069 // such, we need to special case this to allow it.
2070 if (TermOkay && TI == I)
2072 assert(DT.dominates(I, TI) &&
2073 "basic SSA liveness expectation violated by liveness analysis");
2078 /// Check that all the liveness sets used during the computation of liveness
2079 /// obey basic SSA properties. This is useful for finding cases where we miss
2081 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2083 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2084 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2085 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2089 static void computeLiveInValues(DominatorTree &DT, Function &F,
2090 GCPtrLivenessData &Data) {
2092 SmallSetVector<BasicBlock *, 200> Worklist;
2093 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2094 // We use a SetVector so that we don't have duplicates in the worklist.
2095 Worklist.insert(pred_begin(BB), pred_end(BB));
2097 auto NextItem = [&]() {
2098 BasicBlock *BB = Worklist.back();
2099 Worklist.pop_back();
2103 // Seed the liveness for each individual block
2104 for (BasicBlock &BB : F) {
2105 Data.KillSet[&BB] = computeKillSet(&BB);
2106 Data.LiveSet[&BB].clear();
2107 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2110 for (Value *Kill : Data.KillSet[&BB])
2111 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2114 Data.LiveOut[&BB] = DenseSet<Value *>();
2115 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2116 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2117 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2118 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2119 if (!Data.LiveIn[&BB].empty())
2120 AddPredsToWorklist(&BB);
2123 // Propagate that liveness until stable
2124 while (!Worklist.empty()) {
2125 BasicBlock *BB = NextItem();
2127 // Compute our new liveout set, then exit early if it hasn't changed
2128 // despite the contribution of our successor.
2129 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2130 const auto OldLiveOutSize = LiveOut.size();
2131 for (BasicBlock *Succ : successors(BB)) {
2132 assert(Data.LiveIn.count(Succ));
2133 set_union(LiveOut, Data.LiveIn[Succ]);
2135 // assert OutLiveOut is a subset of LiveOut
2136 if (OldLiveOutSize == LiveOut.size()) {
2137 // If the sets are the same size, then we didn't actually add anything
2138 // when unioning our successors LiveIn Thus, the LiveIn of this block
2142 Data.LiveOut[BB] = LiveOut;
2144 // Apply the effects of this basic block
2145 DenseSet<Value *> LiveTmp = LiveOut;
2146 set_union(LiveTmp, Data.LiveSet[BB]);
2147 set_subtract(LiveTmp, Data.KillSet[BB]);
2149 assert(Data.LiveIn.count(BB));
2150 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2151 // assert: OldLiveIn is a subset of LiveTmp
2152 if (OldLiveIn.size() != LiveTmp.size()) {
2153 Data.LiveIn[BB] = LiveTmp;
2154 AddPredsToWorklist(BB);
2156 } // while( !worklist.empty() )
2159 // Sanity check our ouput against SSA properties. This helps catch any
2160 // missing kills during the above iteration.
2161 for (BasicBlock &BB : F) {
2162 checkBasicSSA(DT, Data, BB);
2167 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2168 StatepointLiveSetTy &Out) {
2170 BasicBlock *BB = Inst->getParent();
2172 // Note: The copy is intentional and required
2173 assert(Data.LiveOut.count(BB));
2174 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2176 // We want to handle the statepoint itself oddly. It's
2177 // call result is not live (normal), nor are it's arguments
2178 // (unless they're used again later). This adjustment is
2179 // specifically what we need to relocate
2180 BasicBlock::reverse_iterator rend(Inst);
2181 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2182 LiveOut.erase(Inst);
2183 Out.insert(LiveOut.begin(), LiveOut.end());
2186 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2188 PartiallyConstructedSafepointRecord &Info) {
2189 Instruction *Inst = CS.getInstruction();
2190 StatepointLiveSetTy Updated;
2191 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2194 DenseSet<Value *> Bases;
2195 for (auto KVPair : Info.PointerToBase) {
2196 Bases.insert(KVPair.second);
2199 // We may have base pointers which are now live that weren't before. We need
2200 // to update the PointerToBase structure to reflect this.
2201 for (auto V : Updated)
2202 if (!Info.PointerToBase.count(V)) {
2203 assert(Bases.count(V) && "can't find base for unexpected live value");
2204 Info.PointerToBase[V] = V;
2209 for (auto V : Updated) {
2210 assert(Info.PointerToBase.count(V) &&
2211 "must be able to find base for live value");
2215 // Remove any stale base mappings - this can happen since our liveness is
2216 // more precise then the one inherent in the base pointer analysis
2217 DenseSet<Value *> ToErase;
2218 for (auto KVPair : Info.PointerToBase)
2219 if (!Updated.count(KVPair.first))
2220 ToErase.insert(KVPair.first);
2221 for (auto V : ToErase)
2222 Info.PointerToBase.erase(V);
2225 for (auto KVPair : Info.PointerToBase)
2226 assert(Updated.count(KVPair.first) && "record for non-live value");
2229 Info.liveset = Updated;