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 /// Any new values which were added to the IR during base pointer analysis
135 /// for this safepoint
136 DenseSet<llvm::Value *> NewInsertedDefs;
138 /// The *new* gc.statepoint instruction itself. This produces the token
139 /// that normal path gc.relocates and the gc.result are tied to.
140 Instruction *StatepointToken;
142 /// Instruction to which exceptional gc relocates are attached
143 /// Makes it easier to iterate through them during relocationViaAlloca.
144 Instruction *UnwindToken;
148 /// Compute the live-in set for every basic block in the function
149 static void computeLiveInValues(DominatorTree &DT, Function &F,
150 GCPtrLivenessData &Data);
152 /// Given results from the dataflow liveness computation, find the set of live
153 /// Values at a particular instruction.
154 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
155 StatepointLiveSetTy &out);
157 // TODO: Once we can get to the GCStrategy, this becomes
158 // Optional<bool> isGCManagedPointer(const Value *V) const override {
160 static bool isGCPointerType(const Type *T) {
161 if (const PointerType *PT = dyn_cast<PointerType>(T))
162 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
163 // GC managed heap. We know that a pointer into this heap needs to be
164 // updated and that no other pointer does.
165 return (1 == PT->getAddressSpace());
169 // Return true if this type is one which a) is a gc pointer or contains a GC
170 // pointer and b) is of a type this code expects to encounter as a live value.
171 // (The insertion code will assert that a type which matches (a) and not (b)
172 // is not encountered.)
173 static bool isHandledGCPointerType(Type *T) {
174 // We fully support gc pointers
175 if (isGCPointerType(T))
177 // We partially support vectors of gc pointers. The code will assert if it
178 // can't handle something.
179 if (auto VT = dyn_cast<VectorType>(T))
180 if (isGCPointerType(VT->getElementType()))
186 /// Returns true if this type contains a gc pointer whether we know how to
187 /// handle that type or not.
188 static bool containsGCPtrType(Type *Ty) {
189 if (isGCPointerType(Ty))
191 if (VectorType *VT = dyn_cast<VectorType>(Ty))
192 return isGCPointerType(VT->getScalarType());
193 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
194 return containsGCPtrType(AT->getElementType());
195 if (StructType *ST = dyn_cast<StructType>(Ty))
197 ST->subtypes().begin(), ST->subtypes().end(),
198 [](Type *SubType) { return containsGCPtrType(SubType); });
202 // Returns true if this is a type which a) is a gc pointer or contains a GC
203 // pointer and b) is of a type which the code doesn't expect (i.e. first class
204 // aggregates). Used to trip assertions.
205 static bool isUnhandledGCPointerType(Type *Ty) {
206 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
210 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
211 if (a->hasName() && b->hasName()) {
212 return -1 == a->getName().compare(b->getName());
213 } else if (a->hasName() && !b->hasName()) {
215 } else if (!a->hasName() && b->hasName()) {
218 // Better than nothing, but not stable
223 // Conservatively identifies any definitions which might be live at the
224 // given instruction. The analysis is performed immediately before the
225 // given instruction. Values defined by that instruction are not considered
226 // live. Values used by that instruction are considered live.
227 static void analyzeParsePointLiveness(
228 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
229 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
230 Instruction *inst = CS.getInstruction();
232 StatepointLiveSetTy liveset;
233 findLiveSetAtInst(inst, OriginalLivenessData, liveset);
236 // Note: This output is used by several of the test cases
237 // The order of elemtns in a set is not stable, put them in a vec and sort
239 SmallVector<Value *, 64> temp;
240 temp.insert(temp.end(), liveset.begin(), liveset.end());
241 std::sort(temp.begin(), temp.end(), order_by_name);
242 errs() << "Live Variables:\n";
243 for (Value *V : temp) {
244 errs() << " " << V->getName(); // no newline
248 if (PrintLiveSetSize) {
249 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
250 errs() << "Number live values: " << liveset.size() << "\n";
252 result.liveset = liveset;
255 /// If we can trivially determine that this vector contains only base pointers,
256 /// return the base instruction.
257 static Value *findBaseOfVector(Value *I) {
258 assert(I->getType()->isVectorTy() &&
259 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
260 "Illegal to ask for the base pointer of a non-pointer type");
262 // Each case parallels findBaseDefiningValue below, see that code for
263 // detailed motivation.
265 if (isa<Argument>(I))
266 // An incoming argument to the function is a base pointer
269 // We shouldn't see the address of a global as a vector value?
270 assert(!isa<GlobalVariable>(I) &&
271 "unexpected global variable found in base of vector");
273 // inlining could possibly introduce phi node that contains
274 // undef if callee has multiple returns
275 if (isa<UndefValue>(I))
276 // utterly meaningless, but useful for dealing with partially optimized
280 // Due to inheritance, this must be _after_ the global variable and undef
282 if (Constant *Con = dyn_cast<Constant>(I)) {
283 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
284 "order of checks wrong!");
285 assert(Con->isNullValue() && "null is the only case which makes sense");
289 if (isa<LoadInst>(I))
292 // Note: This code is currently rather incomplete. We are essentially only
293 // handling cases where the vector element is trivially a base pointer. We
294 // need to update the entire base pointer construction algorithm to know how
295 // to track vector elements and potentially scalarize, but the case which
296 // would motivate the work hasn't shown up in real workloads yet.
297 llvm_unreachable("no base found for vector element");
300 /// Helper function for findBasePointer - Will return a value which either a)
301 /// defines the base pointer for the input or b) blocks the simple search
302 /// (i.e. a PHI or Select of two derived pointers)
303 static Value *findBaseDefiningValue(Value *I) {
304 assert(I->getType()->isPointerTy() &&
305 "Illegal to ask for the base pointer of a non-pointer type");
307 // This case is a bit of a hack - it only handles extracts from vectors which
308 // trivially contain only base pointers. See note inside the function for
309 // how to improve this.
310 if (auto *EEI = dyn_cast<ExtractElementInst>(I)) {
311 Value *VectorOperand = EEI->getVectorOperand();
312 Value *VectorBase = findBaseOfVector(VectorOperand);
314 assert(VectorBase && "extract element not known to be a trivial base");
318 if (isa<Argument>(I))
319 // An incoming argument to the function is a base pointer
320 // We should have never reached here if this argument isn't an gc value
323 if (isa<GlobalVariable>(I))
327 // inlining could possibly introduce phi node that contains
328 // undef if callee has multiple returns
329 if (isa<UndefValue>(I))
330 // utterly meaningless, but useful for dealing with
331 // partially optimized code.
334 // Due to inheritance, this must be _after_ the global variable and undef
336 if (Constant *Con = dyn_cast<Constant>(I)) {
337 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
338 "order of checks wrong!");
339 // Note: Finding a constant base for something marked for relocation
340 // doesn't really make sense. The most likely case is either a) some
341 // screwed up the address space usage or b) your validating against
342 // compiled C++ code w/o the proper separation. The only real exception
343 // is a null pointer. You could have generic code written to index of
344 // off a potentially null value and have proven it null. We also use
345 // null pointers in dead paths of relocation phis (which we might later
346 // want to find a base pointer for).
347 assert(isa<ConstantPointerNull>(Con) &&
348 "null is the only case which makes sense");
352 if (CastInst *CI = dyn_cast<CastInst>(I)) {
353 Value *Def = CI->stripPointerCasts();
354 // If we find a cast instruction here, it means we've found a cast which is
355 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
356 // handle int->ptr conversion.
357 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
358 return findBaseDefiningValue(Def);
361 if (isa<LoadInst>(I))
362 return I; // The value loaded is an gc base itself
364 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
365 // The base of this GEP is the base
366 return findBaseDefiningValue(GEP->getPointerOperand());
368 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
369 switch (II->getIntrinsicID()) {
370 case Intrinsic::experimental_gc_result_ptr:
372 // fall through to general call handling
374 case Intrinsic::experimental_gc_statepoint:
375 case Intrinsic::experimental_gc_result_float:
376 case Intrinsic::experimental_gc_result_int:
377 llvm_unreachable("these don't produce pointers");
378 case Intrinsic::experimental_gc_relocate: {
379 // Rerunning safepoint insertion after safepoints are already
380 // inserted is not supported. It could probably be made to work,
381 // but why are you doing this? There's no good reason.
382 llvm_unreachable("repeat safepoint insertion is not supported");
384 case Intrinsic::gcroot:
385 // Currently, this mechanism hasn't been extended to work with gcroot.
386 // There's no reason it couldn't be, but I haven't thought about the
387 // implications much.
389 "interaction with the gcroot mechanism is not supported");
392 // We assume that functions in the source language only return base
393 // pointers. This should probably be generalized via attributes to support
394 // both source language and internal functions.
395 if (isa<CallInst>(I) || isa<InvokeInst>(I))
398 // I have absolutely no idea how to implement this part yet. It's not
399 // neccessarily hard, I just haven't really looked at it yet.
400 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
402 if (isa<AtomicCmpXchgInst>(I))
403 // A CAS is effectively a atomic store and load combined under a
404 // predicate. From the perspective of base pointers, we just treat it
408 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
409 "binary ops which don't apply to pointers");
411 // The aggregate ops. Aggregates can either be in the heap or on the
412 // stack, but in either case, this is simply a field load. As a result,
413 // this is a defining definition of the base just like a load is.
414 if (isa<ExtractValueInst>(I))
417 // We should never see an insert vector since that would require we be
418 // tracing back a struct value not a pointer value.
419 assert(!isa<InsertValueInst>(I) &&
420 "Base pointer for a struct is meaningless");
422 // The last two cases here don't return a base pointer. Instead, they
423 // return a value which dynamically selects from amoung several base
424 // derived pointers (each with it's own base potentially). It's the job of
425 // the caller to resolve these.
426 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
427 "missing instruction case in findBaseDefiningValing");
431 /// Returns the base defining value for this value.
432 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
433 Value *&Cached = Cache[I];
435 Cached = findBaseDefiningValue(I);
437 assert(Cache[I] != nullptr);
440 dbgs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
446 /// Return a base pointer for this value if known. Otherwise, return it's
447 /// base defining value.
448 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
449 Value *Def = findBaseDefiningValueCached(I, Cache);
450 auto Found = Cache.find(Def);
451 if (Found != Cache.end()) {
452 // Either a base-of relation, or a self reference. Caller must check.
453 return Found->second;
455 // Only a BDV available
459 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
460 /// is it known to be a base pointer? Or do we need to continue searching.
461 static bool isKnownBaseResult(Value *V) {
462 if (!isa<PHINode>(V) && !isa<SelectInst>(V)) {
463 // no recursion possible
466 if (isa<Instruction>(V) &&
467 cast<Instruction>(V)->getMetadata("is_base_value")) {
468 // This is a previously inserted base phi or select. We know
469 // that this is a base value.
473 // We need to keep searching
477 // TODO: find a better name for this
481 enum Status { Unknown, Base, Conflict };
483 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
484 assert(status != Base || b);
486 PhiState(Value *b) : status(Base), base(b) {}
487 PhiState() : status(Unknown), base(nullptr) {}
489 Status getStatus() const { return status; }
490 Value *getBase() const { return base; }
492 bool isBase() const { return getStatus() == Base; }
493 bool isUnknown() const { return getStatus() == Unknown; }
494 bool isConflict() const { return getStatus() == Conflict; }
496 bool operator==(const PhiState &other) const {
497 return base == other.base && status == other.status;
500 bool operator!=(const PhiState &other) const { return !(*this == other); }
503 errs() << status << " (" << base << " - "
504 << (base ? base->getName() : "nullptr") << "): ";
509 Value *base; // non null only if status == base
512 typedef DenseMap<Value *, PhiState> ConflictStateMapTy;
513 // Values of type PhiState form a lattice, and this is a helper
514 // class that implementes the meet operation. The meat of the meet
515 // operation is implemented in MeetPhiStates::pureMeet
516 class MeetPhiStates {
518 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
519 explicit MeetPhiStates(const ConflictStateMapTy &phiStates)
520 : phiStates(phiStates) {}
522 // Destructively meet the current result with the base V. V can
523 // either be a merge instruction (SelectInst / PHINode), in which
524 // case its status is looked up in the phiStates map; or a regular
525 // SSA value, in which case it is assumed to be a base.
526 void meetWith(Value *V) {
527 PhiState otherState = getStateForBDV(V);
528 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
529 MeetPhiStates::pureMeet(currentResult, otherState)) &&
530 "math is wrong: meet does not commute!");
531 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
534 PhiState getResult() const { return currentResult; }
537 const ConflictStateMapTy &phiStates;
538 PhiState currentResult;
540 /// Return a phi state for a base defining value. We'll generate a new
541 /// base state for known bases and expect to find a cached state otherwise
542 PhiState getStateForBDV(Value *baseValue) {
543 if (isKnownBaseResult(baseValue)) {
544 return PhiState(baseValue);
546 return lookupFromMap(baseValue);
550 PhiState lookupFromMap(Value *V) {
551 auto I = phiStates.find(V);
552 assert(I != phiStates.end() && "lookup failed!");
556 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
557 switch (stateA.getStatus()) {
558 case PhiState::Unknown:
562 assert(stateA.getBase() && "can't be null");
563 if (stateB.isUnknown())
566 if (stateB.isBase()) {
567 if (stateA.getBase() == stateB.getBase()) {
568 assert(stateA == stateB && "equality broken!");
571 return PhiState(PhiState::Conflict);
573 assert(stateB.isConflict() && "only three states!");
574 return PhiState(PhiState::Conflict);
576 case PhiState::Conflict:
579 llvm_unreachable("only three states!");
583 /// For a given value or instruction, figure out what base ptr it's derived
584 /// from. For gc objects, this is simply itself. On success, returns a value
585 /// which is the base pointer. (This is reliable and can be used for
586 /// relocation.) On failure, returns nullptr.
587 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache,
588 DenseSet<llvm::Value *> &NewInsertedDefs) {
589 Value *def = findBaseOrBDV(I, cache);
591 if (isKnownBaseResult(def)) {
595 // Here's the rough algorithm:
596 // - For every SSA value, construct a mapping to either an actual base
597 // pointer or a PHI which obscures the base pointer.
598 // - Construct a mapping from PHI to unknown TOP state. Use an
599 // optimistic algorithm to propagate base pointer information. Lattice
604 // When algorithm terminates, all PHIs will either have a single concrete
605 // base or be in a conflict state.
606 // - For every conflict, insert a dummy PHI node without arguments. Add
607 // these to the base[Instruction] = BasePtr mapping. For every
608 // non-conflict, add the actual base.
609 // - For every conflict, add arguments for the base[a] of each input
612 // Note: A simpler form of this would be to add the conflict form of all
613 // PHIs without running the optimistic algorithm. This would be
614 // analougous to pessimistic data flow and would likely lead to an
615 // overall worse solution.
617 ConflictStateMapTy states;
618 states[def] = PhiState();
619 // Recursively fill in all phis & selects reachable from the initial one
620 // for which we don't already know a definite base value for
621 // TODO: This should be rewritten with a worklist
625 // Since we're adding elements to 'states' as we run, we can't keep
626 // iterators into the set.
627 SmallVector<Value *, 16> Keys;
628 Keys.reserve(states.size());
629 for (auto Pair : states) {
630 Value *V = Pair.first;
633 for (Value *v : Keys) {
634 assert(!isKnownBaseResult(v) && "why did it get added?");
635 if (PHINode *phi = dyn_cast<PHINode>(v)) {
636 assert(phi->getNumIncomingValues() > 0 &&
637 "zero input phis are illegal");
638 for (Value *InVal : phi->incoming_values()) {
639 Value *local = findBaseOrBDV(InVal, cache);
640 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
641 states[local] = PhiState();
645 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
646 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
647 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
648 states[local] = PhiState();
651 local = findBaseOrBDV(sel->getFalseValue(), cache);
652 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
653 states[local] = PhiState();
661 errs() << "States after initialization:\n";
662 for (auto Pair : states) {
663 Instruction *v = cast<Instruction>(Pair.first);
664 PhiState state = Pair.second;
670 // TODO: come back and revisit the state transitions around inputs which
671 // have reached conflict state. The current version seems too conservative.
673 bool progress = true;
676 size_t oldSize = states.size();
679 // We're only changing keys in this loop, thus safe to keep iterators
680 for (auto Pair : states) {
681 MeetPhiStates calculateMeet(states);
682 Value *v = Pair.first;
683 assert(!isKnownBaseResult(v) && "why did it get added?");
684 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
685 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
686 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
688 for (Value *Val : cast<PHINode>(v)->incoming_values())
689 calculateMeet.meetWith(findBaseOrBDV(Val, cache));
691 PhiState oldState = states[v];
692 PhiState newState = calculateMeet.getResult();
693 if (oldState != newState) {
695 states[v] = newState;
699 assert(oldSize <= states.size());
700 assert(oldSize == states.size() || progress);
704 errs() << "States after meet iteration:\n";
705 for (auto Pair : states) {
706 Instruction *v = cast<Instruction>(Pair.first);
707 PhiState state = Pair.second;
713 // Insert Phis for all conflicts
714 // We want to keep naming deterministic in the loop that follows, so
715 // sort the keys before iteration. This is useful in allowing us to
716 // write stable tests. Note that there is no invalidation issue here.
717 SmallVector<Value *, 16> Keys;
718 Keys.reserve(states.size());
719 for (auto Pair : states) {
720 Value *V = Pair.first;
723 std::sort(Keys.begin(), Keys.end(), order_by_name);
724 // TODO: adjust naming patterns to avoid this order of iteration dependency
725 for (Value *V : Keys) {
726 Instruction *v = cast<Instruction>(V);
727 PhiState state = states[V];
728 assert(!isKnownBaseResult(v) && "why did it get added?");
729 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
730 if (!state.isConflict())
733 if (isa<PHINode>(v)) {
735 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
736 assert(num_preds > 0 && "how did we reach here");
737 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
738 NewInsertedDefs.insert(phi);
739 // Add metadata marking this as a base value
740 auto *const_1 = ConstantInt::get(
742 v->getParent()->getParent()->getParent()->getContext()),
744 auto MDConst = ConstantAsMetadata::get(const_1);
745 MDNode *md = MDNode::get(
746 v->getParent()->getParent()->getParent()->getContext(), MDConst);
747 phi->setMetadata("is_base_value", md);
748 states[v] = PhiState(PhiState::Conflict, phi);
750 SelectInst *sel = cast<SelectInst>(v);
751 // The undef will be replaced later
752 UndefValue *undef = UndefValue::get(sel->getType());
753 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
754 undef, "base_select", sel);
755 NewInsertedDefs.insert(basesel);
756 // Add metadata marking this as a base value
757 auto *const_1 = ConstantInt::get(
759 v->getParent()->getParent()->getParent()->getContext()),
761 auto MDConst = ConstantAsMetadata::get(const_1);
762 MDNode *md = MDNode::get(
763 v->getParent()->getParent()->getParent()->getContext(), MDConst);
764 basesel->setMetadata("is_base_value", md);
765 states[v] = PhiState(PhiState::Conflict, basesel);
769 // Fixup all the inputs of the new PHIs
770 for (auto Pair : states) {
771 Instruction *v = cast<Instruction>(Pair.first);
772 PhiState state = Pair.second;
774 assert(!isKnownBaseResult(v) && "why did it get added?");
775 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
776 if (!state.isConflict())
779 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
780 PHINode *phi = cast<PHINode>(v);
781 unsigned NumPHIValues = phi->getNumIncomingValues();
782 for (unsigned i = 0; i < NumPHIValues; i++) {
783 Value *InVal = phi->getIncomingValue(i);
784 BasicBlock *InBB = phi->getIncomingBlock(i);
786 // If we've already seen InBB, add the same incoming value
787 // we added for it earlier. The IR verifier requires phi
788 // nodes with multiple entries from the same basic block
789 // to have the same incoming value for each of those
790 // entries. If we don't do this check here and basephi
791 // has a different type than base, we'll end up adding two
792 // bitcasts (and hence two distinct values) as incoming
793 // values for the same basic block.
795 int blockIndex = basephi->getBasicBlockIndex(InBB);
796 if (blockIndex != -1) {
797 Value *oldBase = basephi->getIncomingValue(blockIndex);
798 basephi->addIncoming(oldBase, InBB);
800 Value *base = findBaseOrBDV(InVal, cache);
801 if (!isKnownBaseResult(base)) {
802 // Either conflict or base.
803 assert(states.count(base));
804 base = states[base].getBase();
805 assert(base != nullptr && "unknown PhiState!");
806 assert(NewInsertedDefs.count(base) &&
807 "should have already added this in a prev. iteration!");
810 // In essense this assert states: the only way two
811 // values incoming from the same basic block may be
812 // different is by being different bitcasts of the same
813 // value. A cleanup that remains TODO is changing
814 // findBaseOrBDV to return an llvm::Value of the correct
815 // type (and still remain pure). This will remove the
816 // need to add bitcasts.
817 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
818 "sanity -- findBaseOrBDV should be pure!");
823 // Find either the defining value for the PHI or the normal base for
825 Value *base = findBaseOrBDV(InVal, cache);
826 if (!isKnownBaseResult(base)) {
827 // Either conflict or base.
828 assert(states.count(base));
829 base = states[base].getBase();
830 assert(base != nullptr && "unknown PhiState!");
832 assert(base && "can't be null");
833 // Must use original input BB since base may not be Instruction
834 // The cast is needed since base traversal may strip away bitcasts
835 if (base->getType() != basephi->getType()) {
836 base = new BitCastInst(base, basephi->getType(), "cast",
837 InBB->getTerminator());
838 NewInsertedDefs.insert(base);
840 basephi->addIncoming(base, InBB);
842 assert(basephi->getNumIncomingValues() == NumPHIValues);
844 SelectInst *basesel = cast<SelectInst>(state.getBase());
845 SelectInst *sel = cast<SelectInst>(v);
846 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
847 // something more safe and less hacky.
848 for (int i = 1; i <= 2; i++) {
849 Value *InVal = sel->getOperand(i);
850 // Find either the defining value for the PHI or the normal base for
852 Value *base = findBaseOrBDV(InVal, cache);
853 if (!isKnownBaseResult(base)) {
854 // Either conflict or base.
855 assert(states.count(base));
856 base = states[base].getBase();
857 assert(base != nullptr && "unknown PhiState!");
859 assert(base && "can't be null");
860 // Must use original input BB since base may not be Instruction
861 // The cast is needed since base traversal may strip away bitcasts
862 if (base->getType() != basesel->getType()) {
863 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
864 NewInsertedDefs.insert(base);
866 basesel->setOperand(i, base);
871 // Cache all of our results so we can cheaply reuse them
872 // NOTE: This is actually two caches: one of the base defining value
873 // relation and one of the base pointer relation! FIXME
874 for (auto item : states) {
875 Value *v = item.first;
876 Value *base = item.second.getBase();
878 assert(!isKnownBaseResult(v) && "why did it get added?");
881 std::string fromstr =
882 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
884 errs() << "Updating base value cache"
885 << " for: " << (v->hasName() ? v->getName() : "")
886 << " from: " << fromstr
887 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
890 assert(isKnownBaseResult(base) &&
891 "must be something we 'know' is a base pointer");
892 if (cache.count(v)) {
893 // Once we transition from the BDV relation being store in the cache to
894 // the base relation being stored, it must be stable
895 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
896 "base relation should be stable");
900 assert(cache.find(def) != cache.end());
904 // For a set of live pointers (base and/or derived), identify the base
905 // pointer of the object which they are derived from. This routine will
906 // mutate the IR graph as needed to make the 'base' pointer live at the
907 // definition site of 'derived'. This ensures that any use of 'derived' can
908 // also use 'base'. This may involve the insertion of a number of
909 // additional PHI nodes.
911 // preconditions: live is a set of pointer type Values
913 // side effects: may insert PHI nodes into the existing CFG, will preserve
914 // CFG, will not remove or mutate any existing nodes
916 // post condition: PointerToBase contains one (derived, base) pair for every
917 // pointer in live. Note that derived can be equal to base if the original
918 // pointer was a base pointer.
920 findBasePointers(const StatepointLiveSetTy &live,
921 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
922 DominatorTree *DT, DefiningValueMapTy &DVCache,
923 DenseSet<llvm::Value *> &NewInsertedDefs) {
924 // For the naming of values inserted to be deterministic - which makes for
925 // much cleaner and more stable tests - we need to assign an order to the
926 // live values. DenseSets do not provide a deterministic order across runs.
927 SmallVector<Value *, 64> Temp;
928 Temp.insert(Temp.end(), live.begin(), live.end());
929 std::sort(Temp.begin(), Temp.end(), order_by_name);
930 for (Value *ptr : Temp) {
931 Value *base = findBasePointer(ptr, DVCache, NewInsertedDefs);
932 assert(base && "failed to find base pointer");
933 PointerToBase[ptr] = base;
934 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
935 DT->dominates(cast<Instruction>(base)->getParent(),
936 cast<Instruction>(ptr)->getParent())) &&
937 "The base we found better dominate the derived pointer");
939 // If you see this trip and like to live really dangerously, the code should
940 // be correct, just with idioms the verifier can't handle. You can try
941 // disabling the verifier at your own substaintial risk.
942 assert(!isa<ConstantPointerNull>(base) &&
943 "the relocation code needs adjustment to handle the relocation of "
944 "a null pointer constant without causing false positives in the "
945 "safepoint ir verifier.");
949 /// Find the required based pointers (and adjust the live set) for the given
951 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
953 PartiallyConstructedSafepointRecord &result) {
954 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
955 DenseSet<llvm::Value *> NewInsertedDefs;
956 findBasePointers(result.liveset, PointerToBase, &DT, DVCache,
959 if (PrintBasePointers) {
960 // Note: Need to print these in a stable order since this is checked in
962 errs() << "Base Pairs (w/o Relocation):\n";
963 SmallVector<Value *, 64> Temp;
964 Temp.reserve(PointerToBase.size());
965 for (auto Pair : PointerToBase) {
966 Temp.push_back(Pair.first);
968 std::sort(Temp.begin(), Temp.end(), order_by_name);
969 for (Value *Ptr : Temp) {
970 Value *Base = PointerToBase[Ptr];
971 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
976 result.PointerToBase = PointerToBase;
977 result.NewInsertedDefs = NewInsertedDefs;
980 /// Given an updated version of the dataflow liveness results, update the
981 /// liveset and base pointer maps for the call site CS.
982 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
984 PartiallyConstructedSafepointRecord &result);
986 static void recomputeLiveInValues(
987 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
988 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
989 // TODO-PERF: reuse the original liveness, then simply run the dataflow
990 // again. The old values are still live and will help it stablize quickly.
991 GCPtrLivenessData RevisedLivenessData;
992 computeLiveInValues(DT, F, RevisedLivenessData);
993 for (size_t i = 0; i < records.size(); i++) {
994 struct PartiallyConstructedSafepointRecord &info = records[i];
995 const CallSite &CS = toUpdate[i];
996 recomputeLiveInValues(RevisedLivenessData, CS, info);
1000 // Normalize basic block to make it ready to be target of invoke statepoint.
1001 // It means spliting it to have single predecessor. Return newly created BB
1002 // ready to be successor of invoke statepoint.
1003 static BasicBlock *normalizeBBForInvokeSafepoint(BasicBlock *BB,
1004 BasicBlock *InvokeParent,
1006 BasicBlock *ret = BB;
1008 if (!BB->getUniquePredecessor()) {
1009 ret = SplitBlockPredecessors(BB, InvokeParent, "");
1012 // Another requirement for such basic blocks is to not have any phi nodes.
1013 // Since we just ensured that new BB will have single predecessor,
1014 // all phi nodes in it will have one value. Here it would be naturall place
1016 // remove them all. But we can not do this because we are risking to remove
1017 // one of the values stored in liveset of another statepoint. We will do it
1018 // later after placing all safepoints.
1023 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1024 auto itr = std::find(livevec.begin(), livevec.end(), val);
1025 assert(livevec.end() != itr);
1026 size_t index = std::distance(livevec.begin(), itr);
1027 assert(index < livevec.size());
1031 // Create new attribute set containing only attributes which can be transfered
1032 // from original call to the safepoint.
1033 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1036 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1037 unsigned index = AS.getSlotIndex(Slot);
1039 if (index == AttributeSet::ReturnIndex ||
1040 index == AttributeSet::FunctionIndex) {
1042 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1044 Attribute attr = *it;
1046 // Do not allow certain attributes - just skip them
1047 // Safepoint can not be read only or read none.
1048 if (attr.hasAttribute(Attribute::ReadNone) ||
1049 attr.hasAttribute(Attribute::ReadOnly))
1052 ret = ret.addAttributes(
1053 AS.getContext(), index,
1054 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1058 // Just skip parameter attributes for now
1064 /// Helper function to place all gc relocates necessary for the given
1067 /// liveVariables - list of variables to be relocated.
1068 /// liveStart - index of the first live variable.
1069 /// basePtrs - base pointers.
1070 /// statepointToken - statepoint instruction to which relocates should be
1072 /// Builder - Llvm IR builder to be used to construct new calls.
1073 static void CreateGCRelocates(ArrayRef<llvm::Value *> liveVariables,
1074 const int liveStart,
1075 ArrayRef<llvm::Value *> basePtrs,
1076 Instruction *statepointToken,
1077 IRBuilder<> Builder) {
1078 SmallVector<Instruction *, 64> NewDefs;
1079 NewDefs.reserve(liveVariables.size());
1081 Module *M = statepointToken->getParent()->getParent()->getParent();
1083 for (unsigned i = 0; i < liveVariables.size(); i++) {
1084 // We generate a (potentially) unique declaration for every pointer type
1085 // combination. This results is some blow up the function declarations in
1086 // the IR, but removes the need for argument bitcasts which shrinks the IR
1087 // greatly and makes it much more readable.
1088 SmallVector<Type *, 1> types; // one per 'any' type
1089 types.push_back(liveVariables[i]->getType()); // result type
1090 Value *gc_relocate_decl = Intrinsic::getDeclaration(
1091 M, Intrinsic::experimental_gc_relocate, types);
1093 // Generate the gc.relocate call and save the result
1095 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1096 liveStart + find_index(liveVariables, basePtrs[i]));
1097 Value *liveIdx = ConstantInt::get(
1098 Type::getInt32Ty(M->getContext()),
1099 liveStart + find_index(liveVariables, liveVariables[i]));
1101 // only specify a debug name if we can give a useful one
1102 Value *reloc = Builder.CreateCall3(
1103 gc_relocate_decl, statepointToken, baseIdx, liveIdx,
1104 liveVariables[i]->hasName() ? liveVariables[i]->getName() + ".relocated"
1106 // Trick CodeGen into thinking there are lots of free registers at this
1108 cast<CallInst>(reloc)->setCallingConv(CallingConv::Cold);
1110 NewDefs.push_back(cast<Instruction>(reloc));
1112 assert(NewDefs.size() == liveVariables.size() &&
1113 "missing or extra redefinition at safepoint");
1117 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1118 const SmallVectorImpl<llvm::Value *> &basePtrs,
1119 const SmallVectorImpl<llvm::Value *> &liveVariables,
1121 PartiallyConstructedSafepointRecord &result) {
1122 assert(basePtrs.size() == liveVariables.size());
1123 assert(isStatepoint(CS) &&
1124 "This method expects to be rewriting a statepoint");
1126 BasicBlock *BB = CS.getInstruction()->getParent();
1128 Function *F = BB->getParent();
1129 assert(F && "must be set");
1130 Module *M = F->getParent();
1132 assert(M && "must be set");
1134 // We're not changing the function signature of the statepoint since the gc
1135 // arguments go into the var args section.
1136 Function *gc_statepoint_decl = CS.getCalledFunction();
1138 // Then go ahead and use the builder do actually do the inserts. We insert
1139 // immediately before the previous instruction under the assumption that all
1140 // arguments will be available here. We can't insert afterwards since we may
1141 // be replacing a terminator.
1142 Instruction *insertBefore = CS.getInstruction();
1143 IRBuilder<> Builder(insertBefore);
1144 // Copy all of the arguments from the original statepoint - this includes the
1145 // target, call args, and deopt args
1146 SmallVector<llvm::Value *, 64> args;
1147 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1148 // TODO: Clear the 'needs rewrite' flag
1150 // add all the pointers to be relocated (gc arguments)
1151 // Capture the start of the live variable list for use in the gc_relocates
1152 const int live_start = args.size();
1153 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1155 // Create the statepoint given all the arguments
1156 Instruction *token = nullptr;
1157 AttributeSet return_attributes;
1159 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1161 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1162 call->setTailCall(toReplace->isTailCall());
1163 call->setCallingConv(toReplace->getCallingConv());
1165 // Currently we will fail on parameter attributes and on certain
1166 // function attributes.
1167 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1168 // In case if we can handle this set of sttributes - set up function attrs
1169 // directly on statepoint and return attrs later for gc_result intrinsic.
1170 call->setAttributes(new_attrs.getFnAttributes());
1171 return_attributes = new_attrs.getRetAttributes();
1175 // Put the following gc_result and gc_relocate calls immediately after the
1176 // the old call (which we're about to delete)
1177 BasicBlock::iterator next(toReplace);
1178 assert(BB->end() != next && "not a terminator, must have next");
1180 Instruction *IP = &*(next);
1181 Builder.SetInsertPoint(IP);
1182 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1185 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1187 // Insert the new invoke into the old block. We'll remove the old one in a
1188 // moment at which point this will become the new terminator for the
1190 InvokeInst *invoke = InvokeInst::Create(
1191 gc_statepoint_decl, toReplace->getNormalDest(),
1192 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1193 invoke->setCallingConv(toReplace->getCallingConv());
1195 // Currently we will fail on parameter attributes and on certain
1196 // function attributes.
1197 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1198 // In case if we can handle this set of sttributes - set up function attrs
1199 // directly on statepoint and return attrs later for gc_result intrinsic.
1200 invoke->setAttributes(new_attrs.getFnAttributes());
1201 return_attributes = new_attrs.getRetAttributes();
1205 // Generate gc relocates in exceptional path
1206 BasicBlock *unwindBlock = normalizeBBForInvokeSafepoint(
1207 toReplace->getUnwindDest(), invoke->getParent(), P);
1209 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1210 Builder.SetInsertPoint(IP);
1211 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1213 // Extract second element from landingpad return value. We will attach
1214 // exceptional gc relocates to it.
1215 const unsigned idx = 1;
1216 Instruction *exceptional_token =
1217 cast<Instruction>(Builder.CreateExtractValue(
1218 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1219 result.UnwindToken = exceptional_token;
1221 // Just throw away return value. We will use the one we got for normal
1223 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1224 exceptional_token, Builder);
1226 // Generate gc relocates and returns for normal block
1227 BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
1228 toReplace->getNormalDest(), invoke->getParent(), P);
1230 IP = &*(normalDest->getFirstInsertionPt());
1231 Builder.SetInsertPoint(IP);
1233 // gc relocates will be generated later as if it were regular call
1238 // Take the name of the original value call if it had one.
1239 token->takeName(CS.getInstruction());
1241 // The GCResult is already inserted, we just need to find it
1243 Instruction *toReplace = CS.getInstruction();
1244 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1245 "only valid use before rewrite is gc.result");
1246 assert(!toReplace->hasOneUse() ||
1247 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1250 // Update the gc.result of the original statepoint (if any) to use the newly
1251 // inserted statepoint. This is safe to do here since the token can't be
1252 // considered a live reference.
1253 CS.getInstruction()->replaceAllUsesWith(token);
1255 result.StatepointToken = token;
1257 // Second, create a gc.relocate for every live variable
1258 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1262 struct name_ordering {
1265 bool operator()(name_ordering const &a, name_ordering const &b) {
1266 return -1 == a.derived->getName().compare(b.derived->getName());
1270 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1271 SmallVectorImpl<Value *> &livevec) {
1272 assert(basevec.size() == livevec.size());
1274 SmallVector<name_ordering, 64> temp;
1275 for (size_t i = 0; i < basevec.size(); i++) {
1277 v.base = basevec[i];
1278 v.derived = livevec[i];
1281 std::sort(temp.begin(), temp.end(), name_ordering());
1282 for (size_t i = 0; i < basevec.size(); i++) {
1283 basevec[i] = temp[i].base;
1284 livevec[i] = temp[i].derived;
1288 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1289 // which make the relocations happening at this safepoint explicit.
1291 // WARNING: Does not do any fixup to adjust users of the original live
1292 // values. That's the callers responsibility.
1294 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1295 PartiallyConstructedSafepointRecord &result) {
1296 auto liveset = result.liveset;
1297 auto PointerToBase = result.PointerToBase;
1299 // Convert to vector for efficient cross referencing.
1300 SmallVector<Value *, 64> basevec, livevec;
1301 livevec.reserve(liveset.size());
1302 basevec.reserve(liveset.size());
1303 for (Value *L : liveset) {
1304 livevec.push_back(L);
1306 assert(PointerToBase.find(L) != PointerToBase.end());
1307 Value *base = PointerToBase[L];
1308 basevec.push_back(base);
1310 assert(livevec.size() == basevec.size());
1312 // To make the output IR slightly more stable (for use in diffs), ensure a
1313 // fixed order of the values in the safepoint (by sorting the value name).
1314 // The order is otherwise meaningless.
1315 stablize_order(basevec, livevec);
1317 // Do the actual rewriting and delete the old statepoint
1318 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1319 CS.getInstruction()->eraseFromParent();
1322 // Helper function for the relocationViaAlloca.
1323 // It receives iterator to the statepoint gc relocates and emits store to the
1325 // location (via allocaMap) for the each one of them.
1326 // Add visited values into the visitedLiveValues set we will later use them
1327 // for sanity check.
1329 insertRelocationStores(iterator_range<Value::user_iterator> gcRelocs,
1330 DenseMap<Value *, Value *> &allocaMap,
1331 DenseSet<Value *> &visitedLiveValues) {
1333 for (User *U : gcRelocs) {
1334 if (!isa<IntrinsicInst>(U))
1337 IntrinsicInst *relocatedValue = cast<IntrinsicInst>(U);
1339 // We only care about relocates
1340 if (relocatedValue->getIntrinsicID() !=
1341 Intrinsic::experimental_gc_relocate) {
1345 GCRelocateOperands relocateOperands(relocatedValue);
1346 Value *originalValue = const_cast<Value *>(relocateOperands.derivedPtr());
1347 assert(allocaMap.count(originalValue));
1348 Value *alloca = allocaMap[originalValue];
1350 // Emit store into the related alloca
1351 StoreInst *store = new StoreInst(relocatedValue, alloca);
1352 store->insertAfter(relocatedValue);
1355 visitedLiveValues.insert(originalValue);
1360 /// do all the relocation update via allocas and mem2reg
1361 static void relocationViaAlloca(
1362 Function &F, DominatorTree &DT, ArrayRef<Value *> live,
1363 ArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1365 // record initial number of (static) allocas; we'll check we have the same
1366 // number when we get done.
1367 int InitialAllocaNum = 0;
1368 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1370 if (isa<AllocaInst>(*I))
1374 // TODO-PERF: change data structures, reserve
1375 DenseMap<Value *, Value *> allocaMap;
1376 SmallVector<AllocaInst *, 200> PromotableAllocas;
1377 PromotableAllocas.reserve(live.size());
1379 // emit alloca for each live gc pointer
1380 for (unsigned i = 0; i < live.size(); i++) {
1381 Value *liveValue = live[i];
1382 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1383 F.getEntryBlock().getFirstNonPHI());
1384 allocaMap[liveValue] = alloca;
1385 PromotableAllocas.push_back(alloca);
1388 // The next two loops are part of the same conceptual operation. We need to
1389 // insert a store to the alloca after the original def and at each
1390 // redefinition. We need to insert a load before each use. These are split
1391 // into distinct loops for performance reasons.
1393 // update gc pointer after each statepoint
1394 // either store a relocated value or null (if no relocated value found for
1395 // this gc pointer and it is not a gc_result)
1396 // this must happen before we update the statepoint with load of alloca
1397 // otherwise we lose the link between statepoint and old def
1398 for (size_t i = 0; i < records.size(); i++) {
1399 const struct PartiallyConstructedSafepointRecord &info = records[i];
1400 Value *Statepoint = info.StatepointToken;
1402 // This will be used for consistency check
1403 DenseSet<Value *> visitedLiveValues;
1405 // Insert stores for normal statepoint gc relocates
1406 insertRelocationStores(Statepoint->users(), allocaMap, visitedLiveValues);
1408 // In case if it was invoke statepoint
1409 // we will insert stores for exceptional path gc relocates.
1410 if (isa<InvokeInst>(Statepoint)) {
1411 insertRelocationStores(info.UnwindToken->users(), allocaMap,
1415 if (ClobberNonLive) {
1416 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1417 // the gc.statepoint. This will turn some subtle GC problems into
1418 // slightly easier to debug SEGVs. Note that on large IR files with
1419 // lots of gc.statepoints this is extremely costly both memory and time
1421 SmallVector<AllocaInst *, 64> ToClobber;
1422 for (auto Pair : allocaMap) {
1423 Value *Def = Pair.first;
1424 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1426 // This value was relocated
1427 if (visitedLiveValues.count(Def)) {
1430 ToClobber.push_back(Alloca);
1433 auto InsertClobbersAt = [&](Instruction *IP) {
1434 for (auto *AI : ToClobber) {
1435 auto AIType = cast<PointerType>(AI->getType());
1436 auto PT = cast<PointerType>(AIType->getElementType());
1437 Constant *CPN = ConstantPointerNull::get(PT);
1438 StoreInst *store = new StoreInst(CPN, AI);
1439 store->insertBefore(IP);
1443 // Insert the clobbering stores. These may get intermixed with the
1444 // gc.results and gc.relocates, but that's fine.
1445 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1446 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1447 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1449 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1451 InsertClobbersAt(Next);
1455 // update use with load allocas and add store for gc_relocated
1456 for (auto Pair : allocaMap) {
1457 Value *def = Pair.first;
1458 Value *alloca = Pair.second;
1460 // we pre-record the uses of allocas so that we dont have to worry about
1462 // that change the user information.
1463 SmallVector<Instruction *, 20> uses;
1464 // PERF: trade a linear scan for repeated reallocation
1465 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1466 for (User *U : def->users()) {
1467 if (!isa<ConstantExpr>(U)) {
1468 // If the def has a ConstantExpr use, then the def is either a
1469 // ConstantExpr use itself or null. In either case
1470 // (recursively in the first, directly in the second), the oop
1471 // it is ultimately dependent on is null and this particular
1472 // use does not need to be fixed up.
1473 uses.push_back(cast<Instruction>(U));
1477 std::sort(uses.begin(), uses.end());
1478 auto last = std::unique(uses.begin(), uses.end());
1479 uses.erase(last, uses.end());
1481 for (Instruction *use : uses) {
1482 if (isa<PHINode>(use)) {
1483 PHINode *phi = cast<PHINode>(use);
1484 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1485 if (def == phi->getIncomingValue(i)) {
1486 LoadInst *load = new LoadInst(
1487 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1488 phi->setIncomingValue(i, load);
1492 LoadInst *load = new LoadInst(alloca, "", use);
1493 use->replaceUsesOfWith(def, load);
1497 // emit store for the initial gc value
1498 // store must be inserted after load, otherwise store will be in alloca's
1499 // use list and an extra load will be inserted before it
1500 StoreInst *store = new StoreInst(def, alloca);
1501 if (Instruction *inst = dyn_cast<Instruction>(def)) {
1502 if (InvokeInst *invoke = dyn_cast<InvokeInst>(inst)) {
1503 // InvokeInst is a TerminatorInst so the store need to be inserted
1504 // into its normal destination block.
1505 BasicBlock *normalDest = invoke->getNormalDest();
1506 store->insertBefore(normalDest->getFirstNonPHI());
1508 assert(!inst->isTerminator() &&
1509 "The only TerminatorInst that can produce a value is "
1510 "InvokeInst which is handled above.");
1511 store->insertAfter(inst);
1514 assert((isa<Argument>(def) || isa<GlobalVariable>(def) ||
1515 isa<ConstantPointerNull>(def)) &&
1516 "Must be argument or global");
1517 store->insertAfter(cast<Instruction>(alloca));
1521 assert(PromotableAllocas.size() == live.size() &&
1522 "we must have the same allocas with lives");
1523 if (!PromotableAllocas.empty()) {
1524 // apply mem2reg to promote alloca to SSA
1525 PromoteMemToReg(PromotableAllocas, DT);
1529 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1531 if (isa<AllocaInst>(*I))
1533 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1537 /// Implement a unique function which doesn't require we sort the input
1538 /// vector. Doing so has the effect of changing the output of a couple of
1539 /// tests in ways which make them less useful in testing fused safepoints.
1540 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1542 SmallVector<T, 128> TempVec;
1543 TempVec.reserve(Vec.size());
1544 for (auto Element : Vec)
1545 TempVec.push_back(Element);
1547 for (auto V : TempVec) {
1548 if (Seen.insert(V).second) {
1554 static Function *getUseHolder(Module &M) {
1555 FunctionType *ftype =
1556 FunctionType::get(Type::getVoidTy(M.getContext()), true);
1557 Function *Func = cast<Function>(M.getOrInsertFunction("__tmp_use", ftype));
1561 /// Insert holders so that each Value is obviously live through the entire
1562 /// liftetime of the call.
1563 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1564 SmallVectorImpl<CallInst *> &holders) {
1565 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1566 Function *Func = getUseHolder(*M);
1568 // For call safepoints insert dummy calls right after safepoint
1569 BasicBlock::iterator next(CS.getInstruction());
1571 CallInst *base_holder = CallInst::Create(Func, Values, "", next);
1572 holders.push_back(base_holder);
1573 } else if (CS.isInvoke()) {
1574 // For invoke safepooints insert dummy calls both in normal and
1575 // exceptional destination blocks
1576 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1577 CallInst *normal_holder = CallInst::Create(
1578 Func, Values, "", invoke->getNormalDest()->getFirstInsertionPt());
1579 CallInst *unwind_holder = CallInst::Create(
1580 Func, Values, "", invoke->getUnwindDest()->getFirstInsertionPt());
1581 holders.push_back(normal_holder);
1582 holders.push_back(unwind_holder);
1584 llvm_unreachable("unsupported call type");
1587 static void findLiveReferences(
1588 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1589 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1590 GCPtrLivenessData OriginalLivenessData;
1591 computeLiveInValues(DT, F, OriginalLivenessData);
1592 for (size_t i = 0; i < records.size(); i++) {
1593 struct PartiallyConstructedSafepointRecord &info = records[i];
1594 const CallSite &CS = toUpdate[i];
1595 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1599 /// Remove any vector of pointers from the liveset by scalarizing them over the
1600 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1601 /// would be preferrable to include the vector in the statepoint itself, but
1602 /// the lowering code currently does not handle that. Extending it would be
1603 /// slightly non-trivial since it requires a format change. Given how rare
1604 /// such cases are (for the moment?) scalarizing is an acceptable comprimise.
1605 static void splitVectorValues(Instruction *StatepointInst,
1606 StatepointLiveSetTy &LiveSet, DominatorTree &DT) {
1607 SmallVector<Value *, 16> ToSplit;
1608 for (Value *V : LiveSet)
1609 if (isa<VectorType>(V->getType()))
1610 ToSplit.push_back(V);
1612 if (ToSplit.empty())
1615 Function &F = *(StatepointInst->getParent()->getParent());
1617 DenseMap<Value *, AllocaInst *> AllocaMap;
1618 // First is normal return, second is exceptional return (invoke only)
1619 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1620 for (Value *V : ToSplit) {
1623 AllocaInst *Alloca =
1624 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1625 AllocaMap[V] = Alloca;
1627 VectorType *VT = cast<VectorType>(V->getType());
1628 IRBuilder<> Builder(StatepointInst);
1629 SmallVector<Value *, 16> Elements;
1630 for (unsigned i = 0; i < VT->getNumElements(); i++)
1631 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1632 LiveSet.insert(Elements.begin(), Elements.end());
1634 auto InsertVectorReform = [&](Instruction *IP) {
1635 Builder.SetInsertPoint(IP);
1636 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1637 Value *ResultVec = UndefValue::get(VT);
1638 for (unsigned i = 0; i < VT->getNumElements(); i++)
1639 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1640 Builder.getInt32(i));
1644 if (isa<CallInst>(StatepointInst)) {
1645 BasicBlock::iterator Next(StatepointInst);
1647 Instruction *IP = &*(Next);
1648 Replacements[V].first = InsertVectorReform(IP);
1649 Replacements[V].second = nullptr;
1651 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1652 // We've already normalized - check that we don't have shared destination
1654 BasicBlock *NormalDest = Invoke->getNormalDest();
1655 assert(!isa<PHINode>(NormalDest->begin()));
1656 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1657 assert(!isa<PHINode>(UnwindDest->begin()));
1658 // Insert insert element sequences in both successors
1659 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1660 Replacements[V].first = InsertVectorReform(IP);
1661 IP = &*(UnwindDest->getFirstInsertionPt());
1662 Replacements[V].second = InsertVectorReform(IP);
1665 for (Value *V : ToSplit) {
1666 AllocaInst *Alloca = AllocaMap[V];
1668 // Capture all users before we start mutating use lists
1669 SmallVector<Instruction *, 16> Users;
1670 for (User *U : V->users())
1671 Users.push_back(cast<Instruction>(U));
1673 for (Instruction *I : Users) {
1674 if (auto Phi = dyn_cast<PHINode>(I)) {
1675 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1676 if (V == Phi->getIncomingValue(i)) {
1677 LoadInst *Load = new LoadInst(
1678 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1679 Phi->setIncomingValue(i, Load);
1682 LoadInst *Load = new LoadInst(Alloca, "", I);
1683 I->replaceUsesOfWith(V, Load);
1687 // Store the original value and the replacement value into the alloca
1688 StoreInst *Store = new StoreInst(V, Alloca);
1689 if (auto I = dyn_cast<Instruction>(V))
1690 Store->insertAfter(I);
1692 Store->insertAfter(Alloca);
1694 // Normal return for invoke, or call return
1695 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1696 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1697 // Unwind return for invoke only
1698 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1700 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1703 // apply mem2reg to promote alloca to SSA
1704 SmallVector<AllocaInst *, 16> Allocas;
1705 for (Value *V : ToSplit)
1706 Allocas.push_back(AllocaMap[V]);
1707 PromoteMemToReg(Allocas, DT);
1710 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1711 SmallVectorImpl<CallSite> &toUpdate) {
1713 // sanity check the input
1714 std::set<CallSite> uniqued;
1715 uniqued.insert(toUpdate.begin(), toUpdate.end());
1716 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1718 for (size_t i = 0; i < toUpdate.size(); i++) {
1719 CallSite &CS = toUpdate[i];
1720 assert(CS.getInstruction()->getParent()->getParent() == &F);
1721 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1725 // A list of dummy calls added to the IR to keep various values obviously
1726 // live in the IR. We'll remove all of these when done.
1727 SmallVector<CallInst *, 64> holders;
1729 // Insert a dummy call with all of the arguments to the vm_state we'll need
1730 // for the actual safepoint insertion. This ensures reference arguments in
1731 // the deopt argument list are considered live through the safepoint (and
1732 // thus makes sure they get relocated.)
1733 for (size_t i = 0; i < toUpdate.size(); i++) {
1734 CallSite &CS = toUpdate[i];
1735 Statepoint StatepointCS(CS);
1737 SmallVector<Value *, 64> DeoptValues;
1738 for (Use &U : StatepointCS.vm_state_args()) {
1739 Value *Arg = cast<Value>(&U);
1740 assert(!isUnhandledGCPointerType(Arg->getType()) &&
1741 "support for FCA unimplemented");
1742 if (isHandledGCPointerType(Arg->getType()))
1743 DeoptValues.push_back(Arg);
1745 insertUseHolderAfter(CS, DeoptValues, holders);
1748 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
1749 records.reserve(toUpdate.size());
1750 for (size_t i = 0; i < toUpdate.size(); i++) {
1751 struct PartiallyConstructedSafepointRecord info;
1752 records.push_back(info);
1754 assert(records.size() == toUpdate.size());
1756 // A) Identify all gc pointers which are staticly live at the given call
1758 findLiveReferences(F, DT, P, toUpdate, records);
1760 // Do a limited scalarization of any live at safepoint vector values which
1761 // contain pointers. This enables this pass to run after vectorization at
1762 // the cost of some possible performance loss. TODO: it would be nice to
1763 // natively support vectors all the way through the backend so we don't need
1764 // to scalarize here.
1765 for (size_t i = 0; i < records.size(); i++) {
1766 struct PartiallyConstructedSafepointRecord &info = records[i];
1767 Instruction *statepoint = toUpdate[i].getInstruction();
1768 splitVectorValues(cast<Instruction>(statepoint), info.liveset, DT);
1771 // B) Find the base pointers for each live pointer
1772 /* scope for caching */ {
1773 // Cache the 'defining value' relation used in the computation and
1774 // insertion of base phis and selects. This ensures that we don't insert
1775 // large numbers of duplicate base_phis.
1776 DefiningValueMapTy DVCache;
1778 for (size_t i = 0; i < records.size(); i++) {
1779 struct PartiallyConstructedSafepointRecord &info = records[i];
1780 CallSite &CS = toUpdate[i];
1781 findBasePointers(DT, DVCache, CS, info);
1783 } // end of cache scope
1785 // The base phi insertion logic (for any safepoint) may have inserted new
1786 // instructions which are now live at some safepoint. The simplest such
1789 // phi a <-- will be a new base_phi here
1790 // safepoint 1 <-- that needs to be live here
1794 DenseSet<llvm::Value *> allInsertedDefs;
1795 for (size_t i = 0; i < records.size(); i++) {
1796 struct PartiallyConstructedSafepointRecord &info = records[i];
1797 allInsertedDefs.insert(info.NewInsertedDefs.begin(),
1798 info.NewInsertedDefs.end());
1801 // We insert some dummy calls after each safepoint to definitely hold live
1802 // the base pointers which were identified for that safepoint. We'll then
1803 // ask liveness for _every_ base inserted to see what is now live. Then we
1804 // remove the dummy calls.
1805 holders.reserve(holders.size() + records.size());
1806 for (size_t i = 0; i < records.size(); i++) {
1807 struct PartiallyConstructedSafepointRecord &info = records[i];
1808 CallSite &CS = toUpdate[i];
1810 SmallVector<Value *, 128> Bases;
1811 for (auto Pair : info.PointerToBase) {
1812 Bases.push_back(Pair.second);
1814 insertUseHolderAfter(CS, Bases, holders);
1817 // By selecting base pointers, we've effectively inserted new uses. Thus, we
1818 // need to rerun liveness. We may *also* have inserted new defs, but that's
1819 // not the key issue.
1820 recomputeLiveInValues(F, DT, P, toUpdate, records);
1822 if (PrintBasePointers) {
1823 for (size_t i = 0; i < records.size(); i++) {
1824 struct PartiallyConstructedSafepointRecord &info = records[i];
1825 errs() << "Base Pairs: (w/Relocation)\n";
1826 for (auto Pair : info.PointerToBase) {
1827 errs() << " derived %" << Pair.first->getName() << " base %"
1828 << Pair.second->getName() << "\n";
1832 for (size_t i = 0; i < holders.size(); i++) {
1833 holders[i]->eraseFromParent();
1834 holders[i] = nullptr;
1838 // Now run through and replace the existing statepoints with new ones with
1839 // the live variables listed. We do not yet update uses of the values being
1840 // relocated. We have references to live variables that need to
1841 // survive to the last iteration of this loop. (By construction, the
1842 // previous statepoint can not be a live variable, thus we can and remove
1843 // the old statepoint calls as we go.)
1844 for (size_t i = 0; i < records.size(); i++) {
1845 struct PartiallyConstructedSafepointRecord &info = records[i];
1846 CallSite &CS = toUpdate[i];
1847 makeStatepointExplicit(DT, CS, P, info);
1849 toUpdate.clear(); // prevent accident use of invalid CallSites
1851 // In case if we inserted relocates in a different basic block than the
1852 // original safepoint (this can happen for invokes). We need to be sure that
1853 // original values were not used in any of the phi nodes at the
1854 // beginning of basic block containing them. Because we know that all such
1855 // blocks will have single predecessor we can safely assume that all phi
1856 // nodes have single entry (because of normalizeBBForInvokeSafepoint).
1857 // Just remove them all here.
1858 for (size_t i = 0; i < records.size(); i++) {
1859 Instruction *I = records[i].StatepointToken;
1861 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
1862 FoldSingleEntryPHINodes(invoke->getNormalDest());
1863 assert(!isa<PHINode>(invoke->getNormalDest()->begin()));
1865 FoldSingleEntryPHINodes(invoke->getUnwindDest());
1866 assert(!isa<PHINode>(invoke->getUnwindDest()->begin()));
1870 // Do all the fixups of the original live variables to their relocated selves
1871 SmallVector<Value *, 128> live;
1872 for (size_t i = 0; i < records.size(); i++) {
1873 struct PartiallyConstructedSafepointRecord &info = records[i];
1874 // We can't simply save the live set from the original insertion. One of
1875 // the live values might be the result of a call which needs a safepoint.
1876 // That Value* no longer exists and we need to use the new gc_result.
1877 // Thankfully, the liveset is embedded in the statepoint (and updated), so
1878 // we just grab that.
1879 Statepoint statepoint(info.StatepointToken);
1880 live.insert(live.end(), statepoint.gc_args_begin(),
1881 statepoint.gc_args_end());
1883 unique_unsorted(live);
1887 for (auto ptr : live) {
1888 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
1892 relocationViaAlloca(F, DT, live, records);
1893 return !records.empty();
1896 /// Returns true if this function should be rewritten by this pass. The main
1897 /// point of this function is as an extension point for custom logic.
1898 static bool shouldRewriteStatepointsIn(Function &F) {
1899 // TODO: This should check the GCStrategy
1901 const std::string StatepointExampleName("statepoint-example");
1902 return StatepointExampleName == F.getGC();
1907 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
1908 // Nothing to do for declarations.
1909 if (F.isDeclaration() || F.empty())
1912 // Policy choice says not to rewrite - the most common reason is that we're
1913 // compiling code without a GCStrategy.
1914 if (!shouldRewriteStatepointsIn(F))
1917 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1919 // Gather all the statepoints which need rewritten. Be careful to only
1920 // consider those in reachable code since we need to ask dominance queries
1921 // when rewriting. We'll delete the unreachable ones in a moment.
1922 SmallVector<CallSite, 64> ParsePointNeeded;
1923 bool HasUnreachableStatepoint = false;
1924 for (Instruction &I : inst_range(F)) {
1925 // TODO: only the ones with the flag set!
1926 if (isStatepoint(I)) {
1927 if (DT.isReachableFromEntry(I.getParent()))
1928 ParsePointNeeded.push_back(CallSite(&I));
1930 HasUnreachableStatepoint = true;
1934 bool MadeChange = false;
1936 // Delete any unreachable statepoints so that we don't have unrewritten
1937 // statepoints surviving this pass. This makes testing easier and the
1938 // resulting IR less confusing to human readers. Rather than be fancy, we
1939 // just reuse a utility function which removes the unreachable blocks.
1940 if (HasUnreachableStatepoint)
1941 MadeChange |= removeUnreachableBlocks(F);
1943 // Return early if no work to do.
1944 if (ParsePointNeeded.empty())
1947 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
1948 // These are created by LCSSA. They have the effect of increasing the size
1949 // of liveness sets for no good reason. It may be harder to do this post
1950 // insertion since relocations and base phis can confuse things.
1951 for (BasicBlock &BB : F)
1952 if (BB.getUniquePredecessor()) {
1954 FoldSingleEntryPHINodes(&BB);
1957 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
1961 // liveness computation via standard dataflow
1962 // -------------------------------------------------------------------
1964 // TODO: Consider using bitvectors for liveness, the set of potentially
1965 // interesting values should be small and easy to pre-compute.
1967 /// Is this value a constant consisting of entirely null values?
1968 static bool isConstantNull(Value *V) {
1969 return isa<Constant>(V) && cast<Constant>(V)->isNullValue();
1972 /// Compute the live-in set for the location rbegin starting from
1973 /// the live-out set of the basic block
1974 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
1975 BasicBlock::reverse_iterator rend,
1976 DenseSet<Value *> &LiveTmp) {
1978 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
1979 Instruction *I = &*ritr;
1981 // KILL/Def - Remove this definition from LiveIn
1984 // Don't consider *uses* in PHI nodes, we handle their contribution to
1985 // predecessor blocks when we seed the LiveOut sets
1986 if (isa<PHINode>(I))
1989 // USE - Add to the LiveIn set for this instruction
1990 for (Value *V : I->operands()) {
1991 assert(!isUnhandledGCPointerType(V->getType()) &&
1992 "support for FCA unimplemented");
1993 if (isHandledGCPointerType(V->getType()) && !isConstantNull(V) &&
1994 !isa<UndefValue>(V)) {
1995 // The choice to exclude null and undef is arbitrary here. Reconsider?
2002 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2004 for (BasicBlock *Succ : successors(BB)) {
2005 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2006 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2007 PHINode *Phi = cast<PHINode>(&*I);
2008 Value *V = Phi->getIncomingValueForBlock(BB);
2009 assert(!isUnhandledGCPointerType(V->getType()) &&
2010 "support for FCA unimplemented");
2011 if (isHandledGCPointerType(V->getType()) && !isConstantNull(V) &&
2012 !isa<UndefValue>(V)) {
2013 // The choice to exclude null and undef is arbitrary here. Reconsider?
2020 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2021 DenseSet<Value *> KillSet;
2022 for (Instruction &I : *BB)
2023 if (isHandledGCPointerType(I.getType()))
2029 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2030 /// sanity check for the liveness computation.
2031 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2032 TerminatorInst *TI, bool TermOkay = false) {
2033 for (Value *V : Live) {
2034 if (auto *I = dyn_cast<Instruction>(V)) {
2035 // The terminator can be a member of the LiveOut set. LLVM's definition
2036 // of instruction dominance states that V does not dominate itself. As
2037 // such, we need to special case this to allow it.
2038 if (TermOkay && TI == I)
2040 assert(DT.dominates(I, TI) &&
2041 "basic SSA liveness expectation violated by liveness analysis");
2046 /// Check that all the liveness sets used during the computation of liveness
2047 /// obey basic SSA properties. This is useful for finding cases where we miss
2049 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2051 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2052 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2053 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2057 static void computeLiveInValues(DominatorTree &DT, Function &F,
2058 GCPtrLivenessData &Data) {
2060 SmallSetVector<BasicBlock *, 200> Worklist;
2061 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2062 // We use a SetVector so that we don't have duplicates in the worklist.
2063 Worklist.insert(pred_begin(BB), pred_end(BB));
2065 auto NextItem = [&]() {
2066 BasicBlock *BB = Worklist.back();
2067 Worklist.pop_back();
2071 // Seed the liveness for each individual block
2072 for (BasicBlock &BB : F) {
2073 Data.KillSet[&BB] = computeKillSet(&BB);
2074 Data.LiveSet[&BB].clear();
2075 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2078 for (Value *Kill : Data.KillSet[&BB])
2079 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2082 Data.LiveOut[&BB] = DenseSet<Value *>();
2083 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2084 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2085 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2086 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2087 if (!Data.LiveIn[&BB].empty())
2088 AddPredsToWorklist(&BB);
2091 // Propagate that liveness until stable
2092 while (!Worklist.empty()) {
2093 BasicBlock *BB = NextItem();
2095 // Compute our new liveout set, then exit early if it hasn't changed
2096 // despite the contribution of our successor.
2097 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2098 const auto OldLiveOutSize = LiveOut.size();
2099 for (BasicBlock *Succ : successors(BB)) {
2100 assert(Data.LiveIn.count(Succ));
2101 set_union(LiveOut, Data.LiveIn[Succ]);
2103 // assert OutLiveOut is a subset of LiveOut
2104 if (OldLiveOutSize == LiveOut.size()) {
2105 // If the sets are the same size, then we didn't actually add anything
2106 // when unioning our successors LiveIn Thus, the LiveIn of this block
2110 Data.LiveOut[BB] = LiveOut;
2112 // Apply the effects of this basic block
2113 DenseSet<Value *> LiveTmp = LiveOut;
2114 set_union(LiveTmp, Data.LiveSet[BB]);
2115 set_subtract(LiveTmp, Data.KillSet[BB]);
2117 assert(Data.LiveIn.count(BB));
2118 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2119 // assert: OldLiveIn is a subset of LiveTmp
2120 if (OldLiveIn.size() != LiveTmp.size()) {
2121 Data.LiveIn[BB] = LiveTmp;
2122 AddPredsToWorklist(BB);
2124 } // while( !worklist.empty() )
2127 // Sanity check our ouput against SSA properties. This helps catch any
2128 // missing kills during the above iteration.
2129 for (BasicBlock &BB : F) {
2130 checkBasicSSA(DT, Data, BB);
2135 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2136 StatepointLiveSetTy &Out) {
2138 BasicBlock *BB = Inst->getParent();
2140 // Note: The copy is intentional and required
2141 assert(Data.LiveOut.count(BB));
2142 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2144 // We want to handle the statepoint itself oddly. It's
2145 // call result is not live (normal), nor are it's arguments
2146 // (unless they're used again later). This adjustment is
2147 // specifically what we need to relocate
2148 BasicBlock::reverse_iterator rend(Inst);
2149 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2150 LiveOut.erase(Inst);
2151 Out.insert(LiveOut.begin(), LiveOut.end());
2154 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2156 PartiallyConstructedSafepointRecord &Info) {
2157 Instruction *Inst = CS.getInstruction();
2158 StatepointLiveSetTy Updated;
2159 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2162 DenseSet<Value *> Bases;
2163 for (auto KVPair : Info.PointerToBase) {
2164 Bases.insert(KVPair.second);
2167 // We may have base pointers which are now live that weren't before. We need
2168 // to update the PointerToBase structure to reflect this.
2169 for (auto V : Updated)
2170 if (!Info.PointerToBase.count(V)) {
2171 assert(Bases.count(V) && "can't find base for unexpected live value");
2172 Info.PointerToBase[V] = V;
2177 for (auto V : Updated) {
2178 assert(Info.PointerToBase.count(V) &&
2179 "must be able to find base for live value");
2183 // Remove any stale base mappings - this can happen since our liveness is
2184 // more precise then the one inherent in the base pointer analysis
2185 DenseSet<Value *> ToErase;
2186 for (auto KVPair : Info.PointerToBase)
2187 if (!Updated.count(KVPair.first))
2188 ToErase.insert(KVPair.first);
2189 for (auto V : ToErase)
2190 Info.PointerToBase.erase(V);
2193 for (auto KVPair : Info.PointerToBase)
2194 assert(Updated.count(KVPair.first) && "record for non-live value");
2197 Info.liveset = Updated;