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 struct RewriteStatepointsForGC : public FunctionPass {
61 static char ID; // Pass identification, replacement for typeid
63 RewriteStatepointsForGC() : FunctionPass(ID) {
64 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
66 bool runOnFunction(Function &F) override;
68 void getAnalysisUsage(AnalysisUsage &AU) const override {
69 // We add and rewrite a bunch of instructions, but don't really do much
70 // else. We could in theory preserve a lot more analyses here.
71 AU.addRequired<DominatorTreeWrapperPass>();
76 char RewriteStatepointsForGC::ID = 0;
78 FunctionPass *llvm::createRewriteStatepointsForGCPass() {
79 return new RewriteStatepointsForGC();
82 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
83 "Make relocations explicit at statepoints", false, false)
84 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
85 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
86 "Make relocations explicit at statepoints", false, false)
89 struct GCPtrLivenessData {
90 /// Values defined in this block.
91 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet;
92 /// Values used in this block (and thus live); does not included values
93 /// killed within this block.
94 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet;
96 /// Values live into this basic block (i.e. used by any
97 /// instruction in this basic block or ones reachable from here)
98 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn;
100 /// Values live out of this basic block (i.e. live into
101 /// any successor block)
102 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut;
105 // The type of the internal cache used inside the findBasePointers family
106 // of functions. From the callers perspective, this is an opaque type and
107 // should not be inspected.
109 // In the actual implementation this caches two relations:
110 // - The base relation itself (i.e. this pointer is based on that one)
111 // - The base defining value relation (i.e. before base_phi insertion)
112 // Generally, after the execution of a full findBasePointer call, only the
113 // base relation will remain. Internally, we add a mixture of the two
114 // types, then update all the second type to the first type
115 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
116 typedef DenseSet<llvm::Value *> StatepointLiveSetTy;
118 struct PartiallyConstructedSafepointRecord {
119 /// The set of values known to be live accross this safepoint
120 StatepointLiveSetTy liveset;
122 /// Mapping from live pointers to a base-defining-value
123 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
125 /// Any new values which were added to the IR during base pointer analysis
126 /// for this safepoint
127 DenseSet<llvm::Value *> NewInsertedDefs;
129 /// The *new* gc.statepoint instruction itself. This produces the token
130 /// that normal path gc.relocates and the gc.result are tied to.
131 Instruction *StatepointToken;
133 /// Instruction to which exceptional gc relocates are attached
134 /// Makes it easier to iterate through them during relocationViaAlloca.
135 Instruction *UnwindToken;
139 /// Compute the live-in set for every basic block in the function
140 static void computeLiveInValues(DominatorTree &DT, Function &F,
141 GCPtrLivenessData &Data);
143 /// Given results from the dataflow liveness computation, find the set of live
144 /// Values at a particular instruction.
145 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
146 StatepointLiveSetTy &out);
148 // TODO: Once we can get to the GCStrategy, this becomes
149 // Optional<bool> isGCManagedPointer(const Value *V) const override {
151 static bool isGCPointerType(const Type *T) {
152 if (const PointerType *PT = dyn_cast<PointerType>(T))
153 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
154 // GC managed heap. We know that a pointer into this heap needs to be
155 // updated and that no other pointer does.
156 return (1 == PT->getAddressSpace());
160 // Return true if this type is one which a) is a gc pointer or contains a GC
161 // pointer and b) is of a type this code expects to encounter as a live value.
162 // (The insertion code will assert that a type which matches (a) and not (b)
163 // is not encountered.)
164 static bool isHandledGCPointerType(Type *T) {
165 // We fully support gc pointers
166 if (isGCPointerType(T))
168 // We partially support vectors of gc pointers. The code will assert if it
169 // can't handle something.
170 if (auto VT = dyn_cast<VectorType>(T))
171 if (isGCPointerType(VT->getElementType()))
177 /// Returns true if this type contains a gc pointer whether we know how to
178 /// handle that type or not.
179 static bool containsGCPtrType(Type *Ty) {
180 if (isGCPointerType(Ty))
182 if (VectorType *VT = dyn_cast<VectorType>(Ty))
183 return isGCPointerType(VT->getScalarType());
184 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
185 return containsGCPtrType(AT->getElementType());
186 if (StructType *ST = dyn_cast<StructType>(Ty))
188 ST->subtypes().begin(), ST->subtypes().end(),
189 [](Type *SubType) { return containsGCPtrType(SubType); });
193 // Returns true if this is a type which a) is a gc pointer or contains a GC
194 // pointer and b) is of a type which the code doesn't expect (i.e. first class
195 // aggregates). Used to trip assertions.
196 static bool isUnhandledGCPointerType(Type *Ty) {
197 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
201 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
202 if (a->hasName() && b->hasName()) {
203 return -1 == a->getName().compare(b->getName());
204 } else if (a->hasName() && !b->hasName()) {
206 } else if (!a->hasName() && b->hasName()) {
209 // Better than nothing, but not stable
214 // Conservatively identifies any definitions which might be live at the
215 // given instruction. The analysis is performed immediately before the
216 // given instruction. Values defined by that instruction are not considered
217 // live. Values used by that instruction are considered live.
218 static void analyzeParsePointLiveness(
219 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
220 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
221 Instruction *inst = CS.getInstruction();
223 StatepointLiveSetTy liveset;
224 findLiveSetAtInst(inst, OriginalLivenessData, liveset);
227 // Note: This output is used by several of the test cases
228 // The order of elemtns in a set is not stable, put them in a vec and sort
230 SmallVector<Value *, 64> temp;
231 temp.insert(temp.end(), liveset.begin(), liveset.end());
232 std::sort(temp.begin(), temp.end(), order_by_name);
233 errs() << "Live Variables:\n";
234 for (Value *V : temp) {
235 errs() << " " << V->getName(); // no newline
239 if (PrintLiveSetSize) {
240 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
241 errs() << "Number live values: " << liveset.size() << "\n";
243 result.liveset = liveset;
246 /// If we can trivially determine that this vector contains only base pointers,
247 /// return the base instruction.
248 static Value *findBaseOfVector(Value *I) {
249 assert(I->getType()->isVectorTy() &&
250 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
251 "Illegal to ask for the base pointer of a non-pointer type");
253 // Each case parallels findBaseDefiningValue below, see that code for
254 // detailed motivation.
256 if (isa<Argument>(I))
257 // An incoming argument to the function is a base pointer
260 // We shouldn't see the address of a global as a vector value?
261 assert(!isa<GlobalVariable>(I) &&
262 "unexpected global variable found in base of vector");
264 // inlining could possibly introduce phi node that contains
265 // undef if callee has multiple returns
266 if (isa<UndefValue>(I))
267 // utterly meaningless, but useful for dealing with partially optimized
271 // Due to inheritance, this must be _after_ the global variable and undef
273 if (Constant *Con = dyn_cast<Constant>(I)) {
274 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
275 "order of checks wrong!");
276 assert(Con->isNullValue() && "null is the only case which makes sense");
280 if (isa<LoadInst>(I))
283 // Note: This code is currently rather incomplete. We are essentially only
284 // handling cases where the vector element is trivially a base pointer. We
285 // need to update the entire base pointer construction algorithm to know how
286 // to track vector elements and potentially scalarize, but the case which
287 // would motivate the work hasn't shown up in real workloads yet.
288 llvm_unreachable("no base found for vector element");
291 /// Helper function for findBasePointer - Will return a value which either a)
292 /// defines the base pointer for the input or b) blocks the simple search
293 /// (i.e. a PHI or Select of two derived pointers)
294 static Value *findBaseDefiningValue(Value *I) {
295 assert(I->getType()->isPointerTy() &&
296 "Illegal to ask for the base pointer of a non-pointer type");
298 // This case is a bit of a hack - it only handles extracts from vectors which
299 // trivially contain only base pointers. See note inside the function for
300 // how to improve this.
301 if (auto *EEI = dyn_cast<ExtractElementInst>(I)) {
302 Value *VectorOperand = EEI->getVectorOperand();
303 Value *VectorBase = findBaseOfVector(VectorOperand);
305 assert(VectorBase && "extract element not known to be a trivial base");
309 if (isa<Argument>(I))
310 // An incoming argument to the function is a base pointer
311 // We should have never reached here if this argument isn't an gc value
314 if (isa<GlobalVariable>(I))
318 // inlining could possibly introduce phi node that contains
319 // undef if callee has multiple returns
320 if (isa<UndefValue>(I))
321 // utterly meaningless, but useful for dealing with
322 // partially optimized code.
325 // Due to inheritance, this must be _after_ the global variable and undef
327 if (Constant *Con = dyn_cast<Constant>(I)) {
328 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
329 "order of checks wrong!");
330 // Note: Finding a constant base for something marked for relocation
331 // doesn't really make sense. The most likely case is either a) some
332 // screwed up the address space usage or b) your validating against
333 // compiled C++ code w/o the proper separation. The only real exception
334 // is a null pointer. You could have generic code written to index of
335 // off a potentially null value and have proven it null. We also use
336 // null pointers in dead paths of relocation phis (which we might later
337 // want to find a base pointer for).
338 assert(isa<ConstantPointerNull>(Con) &&
339 "null is the only case which makes sense");
343 if (CastInst *CI = dyn_cast<CastInst>(I)) {
344 Value *Def = CI->stripPointerCasts();
345 // If we find a cast instruction here, it means we've found a cast which is
346 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
347 // handle int->ptr conversion.
348 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
349 return findBaseDefiningValue(Def);
352 if (isa<LoadInst>(I))
353 return I; // The value loaded is an gc base itself
355 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
356 // The base of this GEP is the base
357 return findBaseDefiningValue(GEP->getPointerOperand());
359 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
360 switch (II->getIntrinsicID()) {
361 case Intrinsic::experimental_gc_result_ptr:
363 // fall through to general call handling
365 case Intrinsic::experimental_gc_statepoint:
366 case Intrinsic::experimental_gc_result_float:
367 case Intrinsic::experimental_gc_result_int:
368 llvm_unreachable("these don't produce pointers");
369 case Intrinsic::experimental_gc_relocate: {
370 // Rerunning safepoint insertion after safepoints are already
371 // inserted is not supported. It could probably be made to work,
372 // but why are you doing this? There's no good reason.
373 llvm_unreachable("repeat safepoint insertion is not supported");
375 case Intrinsic::gcroot:
376 // Currently, this mechanism hasn't been extended to work with gcroot.
377 // There's no reason it couldn't be, but I haven't thought about the
378 // implications much.
380 "interaction with the gcroot mechanism is not supported");
383 // We assume that functions in the source language only return base
384 // pointers. This should probably be generalized via attributes to support
385 // both source language and internal functions.
386 if (isa<CallInst>(I) || isa<InvokeInst>(I))
389 // I have absolutely no idea how to implement this part yet. It's not
390 // neccessarily hard, I just haven't really looked at it yet.
391 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
393 if (isa<AtomicCmpXchgInst>(I))
394 // A CAS is effectively a atomic store and load combined under a
395 // predicate. From the perspective of base pointers, we just treat it
399 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
400 "binary ops which don't apply to pointers");
402 // The aggregate ops. Aggregates can either be in the heap or on the
403 // stack, but in either case, this is simply a field load. As a result,
404 // this is a defining definition of the base just like a load is.
405 if (isa<ExtractValueInst>(I))
408 // We should never see an insert vector since that would require we be
409 // tracing back a struct value not a pointer value.
410 assert(!isa<InsertValueInst>(I) &&
411 "Base pointer for a struct is meaningless");
413 // The last two cases here don't return a base pointer. Instead, they
414 // return a value which dynamically selects from amoung several base
415 // derived pointers (each with it's own base potentially). It's the job of
416 // the caller to resolve these.
417 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
418 "missing instruction case in findBaseDefiningValing");
422 /// Returns the base defining value for this value.
423 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
424 Value *&Cached = Cache[I];
426 Cached = findBaseDefiningValue(I);
428 assert(Cache[I] != nullptr);
431 dbgs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
437 /// Return a base pointer for this value if known. Otherwise, return it's
438 /// base defining value.
439 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
440 Value *Def = findBaseDefiningValueCached(I, Cache);
441 auto Found = Cache.find(Def);
442 if (Found != Cache.end()) {
443 // Either a base-of relation, or a self reference. Caller must check.
444 return Found->second;
446 // Only a BDV available
450 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
451 /// is it known to be a base pointer? Or do we need to continue searching.
452 static bool isKnownBaseResult(Value *V) {
453 if (!isa<PHINode>(V) && !isa<SelectInst>(V)) {
454 // no recursion possible
457 if (isa<Instruction>(V) &&
458 cast<Instruction>(V)->getMetadata("is_base_value")) {
459 // This is a previously inserted base phi or select. We know
460 // that this is a base value.
464 // We need to keep searching
468 // TODO: find a better name for this
472 enum Status { Unknown, Base, Conflict };
474 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
475 assert(status != Base || b);
477 PhiState(Value *b) : status(Base), base(b) {}
478 PhiState() : status(Unknown), base(nullptr) {}
480 Status getStatus() const { return status; }
481 Value *getBase() const { return base; }
483 bool isBase() const { return getStatus() == Base; }
484 bool isUnknown() const { return getStatus() == Unknown; }
485 bool isConflict() const { return getStatus() == Conflict; }
487 bool operator==(const PhiState &other) const {
488 return base == other.base && status == other.status;
491 bool operator!=(const PhiState &other) const { return !(*this == other); }
494 errs() << status << " (" << base << " - "
495 << (base ? base->getName() : "nullptr") << "): ";
500 Value *base; // non null only if status == base
503 typedef DenseMap<Value *, PhiState> ConflictStateMapTy;
504 // Values of type PhiState form a lattice, and this is a helper
505 // class that implementes the meet operation. The meat of the meet
506 // operation is implemented in MeetPhiStates::pureMeet
507 class MeetPhiStates {
509 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
510 explicit MeetPhiStates(const ConflictStateMapTy &phiStates)
511 : phiStates(phiStates) {}
513 // Destructively meet the current result with the base V. V can
514 // either be a merge instruction (SelectInst / PHINode), in which
515 // case its status is looked up in the phiStates map; or a regular
516 // SSA value, in which case it is assumed to be a base.
517 void meetWith(Value *V) {
518 PhiState otherState = getStateForBDV(V);
519 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
520 MeetPhiStates::pureMeet(currentResult, otherState)) &&
521 "math is wrong: meet does not commute!");
522 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
525 PhiState getResult() const { return currentResult; }
528 const ConflictStateMapTy &phiStates;
529 PhiState currentResult;
531 /// Return a phi state for a base defining value. We'll generate a new
532 /// base state for known bases and expect to find a cached state otherwise
533 PhiState getStateForBDV(Value *baseValue) {
534 if (isKnownBaseResult(baseValue)) {
535 return PhiState(baseValue);
537 return lookupFromMap(baseValue);
541 PhiState lookupFromMap(Value *V) {
542 auto I = phiStates.find(V);
543 assert(I != phiStates.end() && "lookup failed!");
547 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
548 switch (stateA.getStatus()) {
549 case PhiState::Unknown:
553 assert(stateA.getBase() && "can't be null");
554 if (stateB.isUnknown())
557 if (stateB.isBase()) {
558 if (stateA.getBase() == stateB.getBase()) {
559 assert(stateA == stateB && "equality broken!");
562 return PhiState(PhiState::Conflict);
564 assert(stateB.isConflict() && "only three states!");
565 return PhiState(PhiState::Conflict);
567 case PhiState::Conflict:
570 llvm_unreachable("only three states!");
574 /// For a given value or instruction, figure out what base ptr it's derived
575 /// from. For gc objects, this is simply itself. On success, returns a value
576 /// which is the base pointer. (This is reliable and can be used for
577 /// relocation.) On failure, returns nullptr.
578 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache,
579 DenseSet<llvm::Value *> &NewInsertedDefs) {
580 Value *def = findBaseOrBDV(I, cache);
582 if (isKnownBaseResult(def)) {
586 // Here's the rough algorithm:
587 // - For every SSA value, construct a mapping to either an actual base
588 // pointer or a PHI which obscures the base pointer.
589 // - Construct a mapping from PHI to unknown TOP state. Use an
590 // optimistic algorithm to propagate base pointer information. Lattice
595 // When algorithm terminates, all PHIs will either have a single concrete
596 // base or be in a conflict state.
597 // - For every conflict, insert a dummy PHI node without arguments. Add
598 // these to the base[Instruction] = BasePtr mapping. For every
599 // non-conflict, add the actual base.
600 // - For every conflict, add arguments for the base[a] of each input
603 // Note: A simpler form of this would be to add the conflict form of all
604 // PHIs without running the optimistic algorithm. This would be
605 // analougous to pessimistic data flow and would likely lead to an
606 // overall worse solution.
608 ConflictStateMapTy states;
609 states[def] = PhiState();
610 // Recursively fill in all phis & selects reachable from the initial one
611 // for which we don't already know a definite base value for
612 // TODO: This should be rewritten with a worklist
616 // Since we're adding elements to 'states' as we run, we can't keep
617 // iterators into the set.
618 SmallVector<Value *, 16> Keys;
619 Keys.reserve(states.size());
620 for (auto Pair : states) {
621 Value *V = Pair.first;
624 for (Value *v : Keys) {
625 assert(!isKnownBaseResult(v) && "why did it get added?");
626 if (PHINode *phi = dyn_cast<PHINode>(v)) {
627 assert(phi->getNumIncomingValues() > 0 &&
628 "zero input phis are illegal");
629 for (Value *InVal : phi->incoming_values()) {
630 Value *local = findBaseOrBDV(InVal, cache);
631 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
632 states[local] = PhiState();
636 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
637 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
638 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
639 states[local] = PhiState();
642 local = findBaseOrBDV(sel->getFalseValue(), cache);
643 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
644 states[local] = PhiState();
652 errs() << "States after initialization:\n";
653 for (auto Pair : states) {
654 Instruction *v = cast<Instruction>(Pair.first);
655 PhiState state = Pair.second;
661 // TODO: come back and revisit the state transitions around inputs which
662 // have reached conflict state. The current version seems too conservative.
664 bool progress = true;
667 size_t oldSize = states.size();
670 // We're only changing keys in this loop, thus safe to keep iterators
671 for (auto Pair : states) {
672 MeetPhiStates calculateMeet(states);
673 Value *v = Pair.first;
674 assert(!isKnownBaseResult(v) && "why did it get added?");
675 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
676 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
677 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
679 for (Value *Val : cast<PHINode>(v)->incoming_values())
680 calculateMeet.meetWith(findBaseOrBDV(Val, cache));
682 PhiState oldState = states[v];
683 PhiState newState = calculateMeet.getResult();
684 if (oldState != newState) {
686 states[v] = newState;
690 assert(oldSize <= states.size());
691 assert(oldSize == states.size() || progress);
695 errs() << "States after meet iteration:\n";
696 for (auto Pair : states) {
697 Instruction *v = cast<Instruction>(Pair.first);
698 PhiState state = Pair.second;
704 // Insert Phis for all conflicts
705 // We want to keep naming deterministic in the loop that follows, so
706 // sort the keys before iteration. This is useful in allowing us to
707 // write stable tests. Note that there is no invalidation issue here.
708 SmallVector<Value *, 16> Keys;
709 Keys.reserve(states.size());
710 for (auto Pair : states) {
711 Value *V = Pair.first;
714 std::sort(Keys.begin(), Keys.end(), order_by_name);
715 // TODO: adjust naming patterns to avoid this order of iteration dependency
716 for (Value *V : Keys) {
717 Instruction *v = cast<Instruction>(V);
718 PhiState state = states[V];
719 assert(!isKnownBaseResult(v) && "why did it get added?");
720 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
721 if (!state.isConflict())
724 if (isa<PHINode>(v)) {
726 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
727 assert(num_preds > 0 && "how did we reach here");
728 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
729 NewInsertedDefs.insert(phi);
730 // Add metadata marking this as a base value
731 auto *const_1 = ConstantInt::get(
733 v->getParent()->getParent()->getParent()->getContext()),
735 auto MDConst = ConstantAsMetadata::get(const_1);
736 MDNode *md = MDNode::get(
737 v->getParent()->getParent()->getParent()->getContext(), MDConst);
738 phi->setMetadata("is_base_value", md);
739 states[v] = PhiState(PhiState::Conflict, phi);
741 SelectInst *sel = cast<SelectInst>(v);
742 // The undef will be replaced later
743 UndefValue *undef = UndefValue::get(sel->getType());
744 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
745 undef, "base_select", sel);
746 NewInsertedDefs.insert(basesel);
747 // Add metadata marking this as a base value
748 auto *const_1 = ConstantInt::get(
750 v->getParent()->getParent()->getParent()->getContext()),
752 auto MDConst = ConstantAsMetadata::get(const_1);
753 MDNode *md = MDNode::get(
754 v->getParent()->getParent()->getParent()->getContext(), MDConst);
755 basesel->setMetadata("is_base_value", md);
756 states[v] = PhiState(PhiState::Conflict, basesel);
760 // Fixup all the inputs of the new PHIs
761 for (auto Pair : states) {
762 Instruction *v = cast<Instruction>(Pair.first);
763 PhiState state = Pair.second;
765 assert(!isKnownBaseResult(v) && "why did it get added?");
766 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
767 if (!state.isConflict())
770 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
771 PHINode *phi = cast<PHINode>(v);
772 unsigned NumPHIValues = phi->getNumIncomingValues();
773 for (unsigned i = 0; i < NumPHIValues; i++) {
774 Value *InVal = phi->getIncomingValue(i);
775 BasicBlock *InBB = phi->getIncomingBlock(i);
777 // If we've already seen InBB, add the same incoming value
778 // we added for it earlier. The IR verifier requires phi
779 // nodes with multiple entries from the same basic block
780 // to have the same incoming value for each of those
781 // entries. If we don't do this check here and basephi
782 // has a different type than base, we'll end up adding two
783 // bitcasts (and hence two distinct values) as incoming
784 // values for the same basic block.
786 int blockIndex = basephi->getBasicBlockIndex(InBB);
787 if (blockIndex != -1) {
788 Value *oldBase = basephi->getIncomingValue(blockIndex);
789 basephi->addIncoming(oldBase, InBB);
791 Value *base = findBaseOrBDV(InVal, cache);
792 if (!isKnownBaseResult(base)) {
793 // Either conflict or base.
794 assert(states.count(base));
795 base = states[base].getBase();
796 assert(base != nullptr && "unknown PhiState!");
797 assert(NewInsertedDefs.count(base) &&
798 "should have already added this in a prev. iteration!");
801 // In essense this assert states: the only way two
802 // values incoming from the same basic block may be
803 // different is by being different bitcasts of the same
804 // value. A cleanup that remains TODO is changing
805 // findBaseOrBDV to return an llvm::Value of the correct
806 // type (and still remain pure). This will remove the
807 // need to add bitcasts.
808 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
809 "sanity -- findBaseOrBDV should be pure!");
814 // Find either the defining value for the PHI or the normal base for
816 Value *base = findBaseOrBDV(InVal, cache);
817 if (!isKnownBaseResult(base)) {
818 // Either conflict or base.
819 assert(states.count(base));
820 base = states[base].getBase();
821 assert(base != nullptr && "unknown PhiState!");
823 assert(base && "can't be null");
824 // Must use original input BB since base may not be Instruction
825 // The cast is needed since base traversal may strip away bitcasts
826 if (base->getType() != basephi->getType()) {
827 base = new BitCastInst(base, basephi->getType(), "cast",
828 InBB->getTerminator());
829 NewInsertedDefs.insert(base);
831 basephi->addIncoming(base, InBB);
833 assert(basephi->getNumIncomingValues() == NumPHIValues);
835 SelectInst *basesel = cast<SelectInst>(state.getBase());
836 SelectInst *sel = cast<SelectInst>(v);
837 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
838 // something more safe and less hacky.
839 for (int i = 1; i <= 2; i++) {
840 Value *InVal = sel->getOperand(i);
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() != basesel->getType()) {
854 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
855 NewInsertedDefs.insert(base);
857 basesel->setOperand(i, base);
862 // Cache all of our results so we can cheaply reuse them
863 // NOTE: This is actually two caches: one of the base defining value
864 // relation and one of the base pointer relation! FIXME
865 for (auto item : states) {
866 Value *v = item.first;
867 Value *base = item.second.getBase();
869 assert(!isKnownBaseResult(v) && "why did it get added?");
872 std::string fromstr =
873 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
875 errs() << "Updating base value cache"
876 << " for: " << (v->hasName() ? v->getName() : "")
877 << " from: " << fromstr
878 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
881 assert(isKnownBaseResult(base) &&
882 "must be something we 'know' is a base pointer");
883 if (cache.count(v)) {
884 // Once we transition from the BDV relation being store in the cache to
885 // the base relation being stored, it must be stable
886 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
887 "base relation should be stable");
891 assert(cache.find(def) != cache.end());
895 // For a set of live pointers (base and/or derived), identify the base
896 // pointer of the object which they are derived from. This routine will
897 // mutate the IR graph as needed to make the 'base' pointer live at the
898 // definition site of 'derived'. This ensures that any use of 'derived' can
899 // also use 'base'. This may involve the insertion of a number of
900 // additional PHI nodes.
902 // preconditions: live is a set of pointer type Values
904 // side effects: may insert PHI nodes into the existing CFG, will preserve
905 // CFG, will not remove or mutate any existing nodes
907 // post condition: PointerToBase contains one (derived, base) pair for every
908 // pointer in live. Note that derived can be equal to base if the original
909 // pointer was a base pointer.
911 findBasePointers(const StatepointLiveSetTy &live,
912 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
913 DominatorTree *DT, DefiningValueMapTy &DVCache,
914 DenseSet<llvm::Value *> &NewInsertedDefs) {
915 // For the naming of values inserted to be deterministic - which makes for
916 // much cleaner and more stable tests - we need to assign an order to the
917 // live values. DenseSets do not provide a deterministic order across runs.
918 SmallVector<Value *, 64> Temp;
919 Temp.insert(Temp.end(), live.begin(), live.end());
920 std::sort(Temp.begin(), Temp.end(), order_by_name);
921 for (Value *ptr : Temp) {
922 Value *base = findBasePointer(ptr, DVCache, NewInsertedDefs);
923 assert(base && "failed to find base pointer");
924 PointerToBase[ptr] = base;
925 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
926 DT->dominates(cast<Instruction>(base)->getParent(),
927 cast<Instruction>(ptr)->getParent())) &&
928 "The base we found better dominate the derived pointer");
930 // If you see this trip and like to live really dangerously, the code should
931 // be correct, just with idioms the verifier can't handle. You can try
932 // disabling the verifier at your own substaintial risk.
933 assert(!isa<ConstantPointerNull>(base) &&
934 "the relocation code needs adjustment to handle the relocation of "
935 "a null pointer constant without causing false positives in the "
936 "safepoint ir verifier.");
940 /// Find the required based pointers (and adjust the live set) for the given
942 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
944 PartiallyConstructedSafepointRecord &result) {
945 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
946 DenseSet<llvm::Value *> NewInsertedDefs;
947 findBasePointers(result.liveset, PointerToBase, &DT, DVCache,
950 if (PrintBasePointers) {
951 // Note: Need to print these in a stable order since this is checked in
953 errs() << "Base Pairs (w/o Relocation):\n";
954 SmallVector<Value *, 64> Temp;
955 Temp.reserve(PointerToBase.size());
956 for (auto Pair : PointerToBase) {
957 Temp.push_back(Pair.first);
959 std::sort(Temp.begin(), Temp.end(), order_by_name);
960 for (Value *Ptr : Temp) {
961 Value *Base = PointerToBase[Ptr];
962 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
967 result.PointerToBase = PointerToBase;
968 result.NewInsertedDefs = NewInsertedDefs;
971 /// Given an updated version of the dataflow liveness results, update the
972 /// liveset and base pointer maps for the call site CS.
973 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
975 PartiallyConstructedSafepointRecord &result);
977 static void recomputeLiveInValues(
978 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
979 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
980 // TODO-PERF: reuse the original liveness, then simply run the dataflow
981 // again. The old values are still live and will help it stablize quickly.
982 GCPtrLivenessData RevisedLivenessData;
983 computeLiveInValues(DT, F, RevisedLivenessData);
984 for (size_t i = 0; i < records.size(); i++) {
985 struct PartiallyConstructedSafepointRecord &info = records[i];
986 const CallSite &CS = toUpdate[i];
987 recomputeLiveInValues(RevisedLivenessData, CS, info);
991 // Normalize basic block to make it ready to be target of invoke statepoint.
992 // It means spliting it to have single predecessor. Return newly created BB
993 // ready to be successor of invoke statepoint.
994 static BasicBlock *normalizeBBForInvokeSafepoint(BasicBlock *BB,
995 BasicBlock *InvokeParent,
997 BasicBlock *ret = BB;
999 if (!BB->getUniquePredecessor()) {
1000 ret = SplitBlockPredecessors(BB, InvokeParent, "");
1003 // Another requirement for such basic blocks is to not have any phi nodes.
1004 // Since we just ensured that new BB will have single predecessor,
1005 // all phi nodes in it will have one value. Here it would be naturall place
1007 // remove them all. But we can not do this because we are risking to remove
1008 // one of the values stored in liveset of another statepoint. We will do it
1009 // later after placing all safepoints.
1014 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1015 auto itr = std::find(livevec.begin(), livevec.end(), val);
1016 assert(livevec.end() != itr);
1017 size_t index = std::distance(livevec.begin(), itr);
1018 assert(index < livevec.size());
1022 // Create new attribute set containing only attributes which can be transfered
1023 // from original call to the safepoint.
1024 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1027 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1028 unsigned index = AS.getSlotIndex(Slot);
1030 if (index == AttributeSet::ReturnIndex ||
1031 index == AttributeSet::FunctionIndex) {
1033 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1035 Attribute attr = *it;
1037 // Do not allow certain attributes - just skip them
1038 // Safepoint can not be read only or read none.
1039 if (attr.hasAttribute(Attribute::ReadNone) ||
1040 attr.hasAttribute(Attribute::ReadOnly))
1043 ret = ret.addAttributes(
1044 AS.getContext(), index,
1045 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1049 // Just skip parameter attributes for now
1055 /// Helper function to place all gc relocates necessary for the given
1058 /// liveVariables - list of variables to be relocated.
1059 /// liveStart - index of the first live variable.
1060 /// basePtrs - base pointers.
1061 /// statepointToken - statepoint instruction to which relocates should be
1063 /// Builder - Llvm IR builder to be used to construct new calls.
1064 static void CreateGCRelocates(ArrayRef<llvm::Value *> liveVariables,
1065 const int liveStart,
1066 ArrayRef<llvm::Value *> basePtrs,
1067 Instruction *statepointToken,
1068 IRBuilder<> Builder) {
1069 SmallVector<Instruction *, 64> NewDefs;
1070 NewDefs.reserve(liveVariables.size());
1072 Module *M = statepointToken->getParent()->getParent()->getParent();
1074 for (unsigned i = 0; i < liveVariables.size(); i++) {
1075 // We generate a (potentially) unique declaration for every pointer type
1076 // combination. This results is some blow up the function declarations in
1077 // the IR, but removes the need for argument bitcasts which shrinks the IR
1078 // greatly and makes it much more readable.
1079 SmallVector<Type *, 1> types; // one per 'any' type
1080 types.push_back(liveVariables[i]->getType()); // result type
1081 Value *gc_relocate_decl = Intrinsic::getDeclaration(
1082 M, Intrinsic::experimental_gc_relocate, types);
1084 // Generate the gc.relocate call and save the result
1086 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1087 liveStart + find_index(liveVariables, basePtrs[i]));
1088 Value *liveIdx = ConstantInt::get(
1089 Type::getInt32Ty(M->getContext()),
1090 liveStart + find_index(liveVariables, liveVariables[i]));
1092 // only specify a debug name if we can give a useful one
1093 Value *reloc = Builder.CreateCall3(
1094 gc_relocate_decl, statepointToken, baseIdx, liveIdx,
1095 liveVariables[i]->hasName() ? liveVariables[i]->getName() + ".relocated"
1097 // Trick CodeGen into thinking there are lots of free registers at this
1099 cast<CallInst>(reloc)->setCallingConv(CallingConv::Cold);
1101 NewDefs.push_back(cast<Instruction>(reloc));
1103 assert(NewDefs.size() == liveVariables.size() &&
1104 "missing or extra redefinition at safepoint");
1108 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1109 const SmallVectorImpl<llvm::Value *> &basePtrs,
1110 const SmallVectorImpl<llvm::Value *> &liveVariables,
1112 PartiallyConstructedSafepointRecord &result) {
1113 assert(basePtrs.size() == liveVariables.size());
1114 assert(isStatepoint(CS) &&
1115 "This method expects to be rewriting a statepoint");
1117 BasicBlock *BB = CS.getInstruction()->getParent();
1119 Function *F = BB->getParent();
1120 assert(F && "must be set");
1121 Module *M = F->getParent();
1123 assert(M && "must be set");
1125 // We're not changing the function signature of the statepoint since the gc
1126 // arguments go into the var args section.
1127 Function *gc_statepoint_decl = CS.getCalledFunction();
1129 // Then go ahead and use the builder do actually do the inserts. We insert
1130 // immediately before the previous instruction under the assumption that all
1131 // arguments will be available here. We can't insert afterwards since we may
1132 // be replacing a terminator.
1133 Instruction *insertBefore = CS.getInstruction();
1134 IRBuilder<> Builder(insertBefore);
1135 // Copy all of the arguments from the original statepoint - this includes the
1136 // target, call args, and deopt args
1137 SmallVector<llvm::Value *, 64> args;
1138 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1139 // TODO: Clear the 'needs rewrite' flag
1141 // add all the pointers to be relocated (gc arguments)
1142 // Capture the start of the live variable list for use in the gc_relocates
1143 const int live_start = args.size();
1144 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1146 // Create the statepoint given all the arguments
1147 Instruction *token = nullptr;
1148 AttributeSet return_attributes;
1150 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1152 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1153 call->setTailCall(toReplace->isTailCall());
1154 call->setCallingConv(toReplace->getCallingConv());
1156 // Currently we will fail on parameter attributes and on certain
1157 // function attributes.
1158 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1159 // In case if we can handle this set of sttributes - set up function attrs
1160 // directly on statepoint and return attrs later for gc_result intrinsic.
1161 call->setAttributes(new_attrs.getFnAttributes());
1162 return_attributes = new_attrs.getRetAttributes();
1166 // Put the following gc_result and gc_relocate calls immediately after the
1167 // the old call (which we're about to delete)
1168 BasicBlock::iterator next(toReplace);
1169 assert(BB->end() != next && "not a terminator, must have next");
1171 Instruction *IP = &*(next);
1172 Builder.SetInsertPoint(IP);
1173 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1176 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1178 // Insert the new invoke into the old block. We'll remove the old one in a
1179 // moment at which point this will become the new terminator for the
1181 InvokeInst *invoke = InvokeInst::Create(
1182 gc_statepoint_decl, toReplace->getNormalDest(),
1183 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1184 invoke->setCallingConv(toReplace->getCallingConv());
1186 // Currently we will fail on parameter attributes and on certain
1187 // function attributes.
1188 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1189 // In case if we can handle this set of sttributes - set up function attrs
1190 // directly on statepoint and return attrs later for gc_result intrinsic.
1191 invoke->setAttributes(new_attrs.getFnAttributes());
1192 return_attributes = new_attrs.getRetAttributes();
1196 // Generate gc relocates in exceptional path
1197 BasicBlock *unwindBlock = normalizeBBForInvokeSafepoint(
1198 toReplace->getUnwindDest(), invoke->getParent(), P);
1200 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1201 Builder.SetInsertPoint(IP);
1202 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1204 // Extract second element from landingpad return value. We will attach
1205 // exceptional gc relocates to it.
1206 const unsigned idx = 1;
1207 Instruction *exceptional_token =
1208 cast<Instruction>(Builder.CreateExtractValue(
1209 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1210 result.UnwindToken = exceptional_token;
1212 // Just throw away return value. We will use the one we got for normal
1214 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1215 exceptional_token, Builder);
1217 // Generate gc relocates and returns for normal block
1218 BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
1219 toReplace->getNormalDest(), invoke->getParent(), P);
1221 IP = &*(normalDest->getFirstInsertionPt());
1222 Builder.SetInsertPoint(IP);
1224 // gc relocates will be generated later as if it were regular call
1229 // Take the name of the original value call if it had one.
1230 token->takeName(CS.getInstruction());
1232 // The GCResult is already inserted, we just need to find it
1234 Instruction *toReplace = CS.getInstruction();
1235 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1236 "only valid use before rewrite is gc.result");
1237 assert(!toReplace->hasOneUse() ||
1238 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1241 // Update the gc.result of the original statepoint (if any) to use the newly
1242 // inserted statepoint. This is safe to do here since the token can't be
1243 // considered a live reference.
1244 CS.getInstruction()->replaceAllUsesWith(token);
1246 result.StatepointToken = token;
1248 // Second, create a gc.relocate for every live variable
1249 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1253 struct name_ordering {
1256 bool operator()(name_ordering const &a, name_ordering const &b) {
1257 return -1 == a.derived->getName().compare(b.derived->getName());
1261 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1262 SmallVectorImpl<Value *> &livevec) {
1263 assert(basevec.size() == livevec.size());
1265 SmallVector<name_ordering, 64> temp;
1266 for (size_t i = 0; i < basevec.size(); i++) {
1268 v.base = basevec[i];
1269 v.derived = livevec[i];
1272 std::sort(temp.begin(), temp.end(), name_ordering());
1273 for (size_t i = 0; i < basevec.size(); i++) {
1274 basevec[i] = temp[i].base;
1275 livevec[i] = temp[i].derived;
1279 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1280 // which make the relocations happening at this safepoint explicit.
1282 // WARNING: Does not do any fixup to adjust users of the original live
1283 // values. That's the callers responsibility.
1285 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1286 PartiallyConstructedSafepointRecord &result) {
1287 auto liveset = result.liveset;
1288 auto PointerToBase = result.PointerToBase;
1290 // Convert to vector for efficient cross referencing.
1291 SmallVector<Value *, 64> basevec, livevec;
1292 livevec.reserve(liveset.size());
1293 basevec.reserve(liveset.size());
1294 for (Value *L : liveset) {
1295 livevec.push_back(L);
1297 assert(PointerToBase.find(L) != PointerToBase.end());
1298 Value *base = PointerToBase[L];
1299 basevec.push_back(base);
1301 assert(livevec.size() == basevec.size());
1303 // To make the output IR slightly more stable (for use in diffs), ensure a
1304 // fixed order of the values in the safepoint (by sorting the value name).
1305 // The order is otherwise meaningless.
1306 stablize_order(basevec, livevec);
1308 // Do the actual rewriting and delete the old statepoint
1309 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1310 CS.getInstruction()->eraseFromParent();
1313 // Helper function for the relocationViaAlloca.
1314 // It receives iterator to the statepoint gc relocates and emits store to the
1316 // location (via allocaMap) for the each one of them.
1317 // Add visited values into the visitedLiveValues set we will later use them
1318 // for sanity check.
1320 insertRelocationStores(iterator_range<Value::user_iterator> gcRelocs,
1321 DenseMap<Value *, Value *> &allocaMap,
1322 DenseSet<Value *> &visitedLiveValues) {
1324 for (User *U : gcRelocs) {
1325 if (!isa<IntrinsicInst>(U))
1328 IntrinsicInst *relocatedValue = cast<IntrinsicInst>(U);
1330 // We only care about relocates
1331 if (relocatedValue->getIntrinsicID() !=
1332 Intrinsic::experimental_gc_relocate) {
1336 GCRelocateOperands relocateOperands(relocatedValue);
1337 Value *originalValue = const_cast<Value *>(relocateOperands.derivedPtr());
1338 assert(allocaMap.count(originalValue));
1339 Value *alloca = allocaMap[originalValue];
1341 // Emit store into the related alloca
1342 StoreInst *store = new StoreInst(relocatedValue, alloca);
1343 store->insertAfter(relocatedValue);
1346 visitedLiveValues.insert(originalValue);
1351 /// do all the relocation update via allocas and mem2reg
1352 static void relocationViaAlloca(
1353 Function &F, DominatorTree &DT, ArrayRef<Value *> live,
1354 ArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1356 // record initial number of (static) allocas; we'll check we have the same
1357 // number when we get done.
1358 int InitialAllocaNum = 0;
1359 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1361 if (isa<AllocaInst>(*I))
1365 // TODO-PERF: change data structures, reserve
1366 DenseMap<Value *, Value *> allocaMap;
1367 SmallVector<AllocaInst *, 200> PromotableAllocas;
1368 PromotableAllocas.reserve(live.size());
1370 // emit alloca for each live gc pointer
1371 for (unsigned i = 0; i < live.size(); i++) {
1372 Value *liveValue = live[i];
1373 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1374 F.getEntryBlock().getFirstNonPHI());
1375 allocaMap[liveValue] = alloca;
1376 PromotableAllocas.push_back(alloca);
1379 // The next two loops are part of the same conceptual operation. We need to
1380 // insert a store to the alloca after the original def and at each
1381 // redefinition. We need to insert a load before each use. These are split
1382 // into distinct loops for performance reasons.
1384 // update gc pointer after each statepoint
1385 // either store a relocated value or null (if no relocated value found for
1386 // this gc pointer and it is not a gc_result)
1387 // this must happen before we update the statepoint with load of alloca
1388 // otherwise we lose the link between statepoint and old def
1389 for (size_t i = 0; i < records.size(); i++) {
1390 const struct PartiallyConstructedSafepointRecord &info = records[i];
1391 Value *Statepoint = info.StatepointToken;
1393 // This will be used for consistency check
1394 DenseSet<Value *> visitedLiveValues;
1396 // Insert stores for normal statepoint gc relocates
1397 insertRelocationStores(Statepoint->users(), allocaMap, visitedLiveValues);
1399 // In case if it was invoke statepoint
1400 // we will insert stores for exceptional path gc relocates.
1401 if (isa<InvokeInst>(Statepoint)) {
1402 insertRelocationStores(info.UnwindToken->users(), allocaMap,
1407 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1408 // the gc.statepoint. This will turn some subtle GC problems into slightly
1409 // easier to debug SEGVs
1410 SmallVector<AllocaInst *, 64> ToClobber;
1411 for (auto Pair : allocaMap) {
1412 Value *Def = Pair.first;
1413 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1415 // This value was relocated
1416 if (visitedLiveValues.count(Def)) {
1419 ToClobber.push_back(Alloca);
1422 auto InsertClobbersAt = [&](Instruction *IP) {
1423 for (auto *AI : ToClobber) {
1424 auto AIType = cast<PointerType>(AI->getType());
1425 auto PT = cast<PointerType>(AIType->getElementType());
1426 Constant *CPN = ConstantPointerNull::get(PT);
1427 StoreInst *store = new StoreInst(CPN, AI);
1428 store->insertBefore(IP);
1432 // Insert the clobbering stores. These may get intermixed with the
1433 // gc.results and gc.relocates, but that's fine.
1434 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1435 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1436 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1438 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1440 InsertClobbersAt(Next);
1444 // update use with load allocas and add store for gc_relocated
1445 for (auto Pair : allocaMap) {
1446 Value *def = Pair.first;
1447 Value *alloca = Pair.second;
1449 // we pre-record the uses of allocas so that we dont have to worry about
1451 // that change the user information.
1452 SmallVector<Instruction *, 20> uses;
1453 // PERF: trade a linear scan for repeated reallocation
1454 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1455 for (User *U : def->users()) {
1456 if (!isa<ConstantExpr>(U)) {
1457 // If the def has a ConstantExpr use, then the def is either a
1458 // ConstantExpr use itself or null. In either case
1459 // (recursively in the first, directly in the second), the oop
1460 // it is ultimately dependent on is null and this particular
1461 // use does not need to be fixed up.
1462 uses.push_back(cast<Instruction>(U));
1466 std::sort(uses.begin(), uses.end());
1467 auto last = std::unique(uses.begin(), uses.end());
1468 uses.erase(last, uses.end());
1470 for (Instruction *use : uses) {
1471 if (isa<PHINode>(use)) {
1472 PHINode *phi = cast<PHINode>(use);
1473 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1474 if (def == phi->getIncomingValue(i)) {
1475 LoadInst *load = new LoadInst(
1476 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1477 phi->setIncomingValue(i, load);
1481 LoadInst *load = new LoadInst(alloca, "", use);
1482 use->replaceUsesOfWith(def, load);
1486 // emit store for the initial gc value
1487 // store must be inserted after load, otherwise store will be in alloca's
1488 // use list and an extra load will be inserted before it
1489 StoreInst *store = new StoreInst(def, alloca);
1490 if (Instruction *inst = dyn_cast<Instruction>(def)) {
1491 if (InvokeInst *invoke = dyn_cast<InvokeInst>(inst)) {
1492 // InvokeInst is a TerminatorInst so the store need to be inserted
1493 // into its normal destination block.
1494 BasicBlock *normalDest = invoke->getNormalDest();
1495 store->insertBefore(normalDest->getFirstNonPHI());
1497 assert(!inst->isTerminator() &&
1498 "The only TerminatorInst that can produce a value is "
1499 "InvokeInst which is handled above.");
1500 store->insertAfter(inst);
1503 assert((isa<Argument>(def) || isa<GlobalVariable>(def) ||
1504 isa<ConstantPointerNull>(def)) &&
1505 "Must be argument or global");
1506 store->insertAfter(cast<Instruction>(alloca));
1510 assert(PromotableAllocas.size() == live.size() &&
1511 "we must have the same allocas with lives");
1512 if (!PromotableAllocas.empty()) {
1513 // apply mem2reg to promote alloca to SSA
1514 PromoteMemToReg(PromotableAllocas, DT);
1518 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1520 if (isa<AllocaInst>(*I))
1522 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1526 /// Implement a unique function which doesn't require we sort the input
1527 /// vector. Doing so has the effect of changing the output of a couple of
1528 /// tests in ways which make them less useful in testing fused safepoints.
1529 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1531 SmallVector<T, 128> TempVec;
1532 TempVec.reserve(Vec.size());
1533 for (auto Element : Vec)
1534 TempVec.push_back(Element);
1536 for (auto V : TempVec) {
1537 if (Seen.insert(V).second) {
1543 static Function *getUseHolder(Module &M) {
1544 FunctionType *ftype =
1545 FunctionType::get(Type::getVoidTy(M.getContext()), true);
1546 Function *Func = cast<Function>(M.getOrInsertFunction("__tmp_use", ftype));
1550 /// Insert holders so that each Value is obviously live through the entire
1551 /// liftetime of the call.
1552 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1553 SmallVectorImpl<CallInst *> &holders) {
1554 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1555 Function *Func = getUseHolder(*M);
1557 // For call safepoints insert dummy calls right after safepoint
1558 BasicBlock::iterator next(CS.getInstruction());
1560 CallInst *base_holder = CallInst::Create(Func, Values, "", next);
1561 holders.push_back(base_holder);
1562 } else if (CS.isInvoke()) {
1563 // For invoke safepooints insert dummy calls both in normal and
1564 // exceptional destination blocks
1565 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1566 CallInst *normal_holder = CallInst::Create(
1567 Func, Values, "", invoke->getNormalDest()->getFirstInsertionPt());
1568 CallInst *unwind_holder = CallInst::Create(
1569 Func, Values, "", invoke->getUnwindDest()->getFirstInsertionPt());
1570 holders.push_back(normal_holder);
1571 holders.push_back(unwind_holder);
1573 llvm_unreachable("unsupported call type");
1576 static void findLiveReferences(
1577 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1578 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1579 GCPtrLivenessData OriginalLivenessData;
1580 computeLiveInValues(DT, F, OriginalLivenessData);
1581 for (size_t i = 0; i < records.size(); i++) {
1582 struct PartiallyConstructedSafepointRecord &info = records[i];
1583 const CallSite &CS = toUpdate[i];
1584 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1588 /// Remove any vector of pointers from the liveset by scalarizing them over the
1589 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1590 /// would be preferrable to include the vector in the statepoint itself, but
1591 /// the lowering code currently does not handle that. Extending it would be
1592 /// slightly non-trivial since it requires a format change. Given how rare
1593 /// such cases are (for the moment?) scalarizing is an acceptable comprimise.
1594 static void splitVectorValues(Instruction *StatepointInst,
1595 StatepointLiveSetTy &LiveSet, DominatorTree &DT) {
1596 SmallVector<Value *, 16> ToSplit;
1597 for (Value *V : LiveSet)
1598 if (isa<VectorType>(V->getType()))
1599 ToSplit.push_back(V);
1601 if (ToSplit.empty())
1604 Function &F = *(StatepointInst->getParent()->getParent());
1606 DenseMap<Value *, AllocaInst *> AllocaMap;
1607 // First is normal return, second is exceptional return (invoke only)
1608 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1609 for (Value *V : ToSplit) {
1612 AllocaInst *Alloca =
1613 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1614 AllocaMap[V] = Alloca;
1616 VectorType *VT = cast<VectorType>(V->getType());
1617 IRBuilder<> Builder(StatepointInst);
1618 SmallVector<Value *, 16> Elements;
1619 for (unsigned i = 0; i < VT->getNumElements(); i++)
1620 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1621 LiveSet.insert(Elements.begin(), Elements.end());
1623 auto InsertVectorReform = [&](Instruction *IP) {
1624 Builder.SetInsertPoint(IP);
1625 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1626 Value *ResultVec = UndefValue::get(VT);
1627 for (unsigned i = 0; i < VT->getNumElements(); i++)
1628 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1629 Builder.getInt32(i));
1633 if (isa<CallInst>(StatepointInst)) {
1634 BasicBlock::iterator Next(StatepointInst);
1636 Instruction *IP = &*(Next);
1637 Replacements[V].first = InsertVectorReform(IP);
1638 Replacements[V].second = nullptr;
1640 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1641 // We've already normalized - check that we don't have shared destination
1643 BasicBlock *NormalDest = Invoke->getNormalDest();
1644 assert(!isa<PHINode>(NormalDest->begin()));
1645 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1646 assert(!isa<PHINode>(UnwindDest->begin()));
1647 // Insert insert element sequences in both successors
1648 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1649 Replacements[V].first = InsertVectorReform(IP);
1650 IP = &*(UnwindDest->getFirstInsertionPt());
1651 Replacements[V].second = InsertVectorReform(IP);
1654 for (Value *V : ToSplit) {
1655 AllocaInst *Alloca = AllocaMap[V];
1657 // Capture all users before we start mutating use lists
1658 SmallVector<Instruction *, 16> Users;
1659 for (User *U : V->users())
1660 Users.push_back(cast<Instruction>(U));
1662 for (Instruction *I : Users) {
1663 if (auto Phi = dyn_cast<PHINode>(I)) {
1664 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1665 if (V == Phi->getIncomingValue(i)) {
1666 LoadInst *Load = new LoadInst(
1667 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1668 Phi->setIncomingValue(i, Load);
1671 LoadInst *Load = new LoadInst(Alloca, "", I);
1672 I->replaceUsesOfWith(V, Load);
1676 // Store the original value and the replacement value into the alloca
1677 StoreInst *Store = new StoreInst(V, Alloca);
1678 if (auto I = dyn_cast<Instruction>(V))
1679 Store->insertAfter(I);
1681 Store->insertAfter(Alloca);
1683 // Normal return for invoke, or call return
1684 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1685 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1686 // Unwind return for invoke only
1687 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1689 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1692 // apply mem2reg to promote alloca to SSA
1693 SmallVector<AllocaInst *, 16> Allocas;
1694 for (Value *V : ToSplit)
1695 Allocas.push_back(AllocaMap[V]);
1696 PromoteMemToReg(Allocas, DT);
1699 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1700 SmallVectorImpl<CallSite> &toUpdate) {
1702 // sanity check the input
1703 std::set<CallSite> uniqued;
1704 uniqued.insert(toUpdate.begin(), toUpdate.end());
1705 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1707 for (size_t i = 0; i < toUpdate.size(); i++) {
1708 CallSite &CS = toUpdate[i];
1709 assert(CS.getInstruction()->getParent()->getParent() == &F);
1710 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1714 // A list of dummy calls added to the IR to keep various values obviously
1715 // live in the IR. We'll remove all of these when done.
1716 SmallVector<CallInst *, 64> holders;
1718 // Insert a dummy call with all of the arguments to the vm_state we'll need
1719 // for the actual safepoint insertion. This ensures reference arguments in
1720 // the deopt argument list are considered live through the safepoint (and
1721 // thus makes sure they get relocated.)
1722 for (size_t i = 0; i < toUpdate.size(); i++) {
1723 CallSite &CS = toUpdate[i];
1724 Statepoint StatepointCS(CS);
1726 SmallVector<Value *, 64> DeoptValues;
1727 for (Use &U : StatepointCS.vm_state_args()) {
1728 Value *Arg = cast<Value>(&U);
1729 assert(!isUnhandledGCPointerType(Arg->getType()) &&
1730 "support for FCA unimplemented");
1731 if (isHandledGCPointerType(Arg->getType()))
1732 DeoptValues.push_back(Arg);
1734 insertUseHolderAfter(CS, DeoptValues, holders);
1737 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
1738 records.reserve(toUpdate.size());
1739 for (size_t i = 0; i < toUpdate.size(); i++) {
1740 struct PartiallyConstructedSafepointRecord info;
1741 records.push_back(info);
1743 assert(records.size() == toUpdate.size());
1745 // A) Identify all gc pointers which are staticly live at the given call
1747 findLiveReferences(F, DT, P, toUpdate, records);
1749 // Do a limited scalarization of any live at safepoint vector values which
1750 // contain pointers. This enables this pass to run after vectorization at
1751 // the cost of some possible performance loss. TODO: it would be nice to
1752 // natively support vectors all the way through the backend so we don't need
1753 // to scalarize here.
1754 for (size_t i = 0; i < records.size(); i++) {
1755 struct PartiallyConstructedSafepointRecord &info = records[i];
1756 Instruction *statepoint = toUpdate[i].getInstruction();
1757 splitVectorValues(cast<Instruction>(statepoint), info.liveset, DT);
1760 // B) Find the base pointers for each live pointer
1761 /* scope for caching */ {
1762 // Cache the 'defining value' relation used in the computation and
1763 // insertion of base phis and selects. This ensures that we don't insert
1764 // large numbers of duplicate base_phis.
1765 DefiningValueMapTy DVCache;
1767 for (size_t i = 0; i < records.size(); i++) {
1768 struct PartiallyConstructedSafepointRecord &info = records[i];
1769 CallSite &CS = toUpdate[i];
1770 findBasePointers(DT, DVCache, CS, info);
1772 } // end of cache scope
1774 // The base phi insertion logic (for any safepoint) may have inserted new
1775 // instructions which are now live at some safepoint. The simplest such
1778 // phi a <-- will be a new base_phi here
1779 // safepoint 1 <-- that needs to be live here
1783 DenseSet<llvm::Value *> allInsertedDefs;
1784 for (size_t i = 0; i < records.size(); i++) {
1785 struct PartiallyConstructedSafepointRecord &info = records[i];
1786 allInsertedDefs.insert(info.NewInsertedDefs.begin(),
1787 info.NewInsertedDefs.end());
1790 // We insert some dummy calls after each safepoint to definitely hold live
1791 // the base pointers which were identified for that safepoint. We'll then
1792 // ask liveness for _every_ base inserted to see what is now live. Then we
1793 // remove the dummy calls.
1794 holders.reserve(holders.size() + records.size());
1795 for (size_t i = 0; i < records.size(); i++) {
1796 struct PartiallyConstructedSafepointRecord &info = records[i];
1797 CallSite &CS = toUpdate[i];
1799 SmallVector<Value *, 128> Bases;
1800 for (auto Pair : info.PointerToBase) {
1801 Bases.push_back(Pair.second);
1803 insertUseHolderAfter(CS, Bases, holders);
1806 // By selecting base pointers, we've effectively inserted new uses. Thus, we
1807 // need to rerun liveness. We may *also* have inserted new defs, but that's
1808 // not the key issue.
1809 recomputeLiveInValues(F, DT, P, toUpdate, records);
1811 if (PrintBasePointers) {
1812 for (size_t i = 0; i < records.size(); i++) {
1813 struct PartiallyConstructedSafepointRecord &info = records[i];
1814 errs() << "Base Pairs: (w/Relocation)\n";
1815 for (auto Pair : info.PointerToBase) {
1816 errs() << " derived %" << Pair.first->getName() << " base %"
1817 << Pair.second->getName() << "\n";
1821 for (size_t i = 0; i < holders.size(); i++) {
1822 holders[i]->eraseFromParent();
1823 holders[i] = nullptr;
1827 // Now run through and replace the existing statepoints with new ones with
1828 // the live variables listed. We do not yet update uses of the values being
1829 // relocated. We have references to live variables that need to
1830 // survive to the last iteration of this loop. (By construction, the
1831 // previous statepoint can not be a live variable, thus we can and remove
1832 // the old statepoint calls as we go.)
1833 for (size_t i = 0; i < records.size(); i++) {
1834 struct PartiallyConstructedSafepointRecord &info = records[i];
1835 CallSite &CS = toUpdate[i];
1836 makeStatepointExplicit(DT, CS, P, info);
1838 toUpdate.clear(); // prevent accident use of invalid CallSites
1840 // In case if we inserted relocates in a different basic block than the
1841 // original safepoint (this can happen for invokes). We need to be sure that
1842 // original values were not used in any of the phi nodes at the
1843 // beginning of basic block containing them. Because we know that all such
1844 // blocks will have single predecessor we can safely assume that all phi
1845 // nodes have single entry (because of normalizeBBForInvokeSafepoint).
1846 // Just remove them all here.
1847 for (size_t i = 0; i < records.size(); i++) {
1848 Instruction *I = records[i].StatepointToken;
1850 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
1851 FoldSingleEntryPHINodes(invoke->getNormalDest());
1852 assert(!isa<PHINode>(invoke->getNormalDest()->begin()));
1854 FoldSingleEntryPHINodes(invoke->getUnwindDest());
1855 assert(!isa<PHINode>(invoke->getUnwindDest()->begin()));
1859 // Do all the fixups of the original live variables to their relocated selves
1860 SmallVector<Value *, 128> live;
1861 for (size_t i = 0; i < records.size(); i++) {
1862 struct PartiallyConstructedSafepointRecord &info = records[i];
1863 // We can't simply save the live set from the original insertion. One of
1864 // the live values might be the result of a call which needs a safepoint.
1865 // That Value* no longer exists and we need to use the new gc_result.
1866 // Thankfully, the liveset is embedded in the statepoint (and updated), so
1867 // we just grab that.
1868 Statepoint statepoint(info.StatepointToken);
1869 live.insert(live.end(), statepoint.gc_args_begin(),
1870 statepoint.gc_args_end());
1872 unique_unsorted(live);
1876 for (auto ptr : live) {
1877 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
1881 relocationViaAlloca(F, DT, live, records);
1882 return !records.empty();
1885 /// Returns true if this function should be rewritten by this pass. The main
1886 /// point of this function is as an extension point for custom logic.
1887 static bool shouldRewriteStatepointsIn(Function &F) {
1888 // TODO: This should check the GCStrategy
1890 const std::string StatepointExampleName("statepoint-example");
1891 return StatepointExampleName == F.getGC();
1896 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
1897 // Nothing to do for declarations.
1898 if (F.isDeclaration() || F.empty())
1901 // Policy choice says not to rewrite - the most common reason is that we're
1902 // compiling code without a GCStrategy.
1903 if (!shouldRewriteStatepointsIn(F))
1906 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1908 // Gather all the statepoints which need rewritten. Be careful to only
1909 // consider those in reachable code since we need to ask dominance queries
1910 // when rewriting. We'll delete the unreachable ones in a moment.
1911 SmallVector<CallSite, 64> ParsePointNeeded;
1912 bool HasUnreachableStatepoint = false;
1913 for (Instruction &I : inst_range(F)) {
1914 // TODO: only the ones with the flag set!
1915 if (isStatepoint(I)) {
1916 if (DT.isReachableFromEntry(I.getParent()))
1917 ParsePointNeeded.push_back(CallSite(&I));
1919 HasUnreachableStatepoint = true;
1923 bool MadeChange = false;
1925 // Delete any unreachable statepoints so that we don't have unrewritten
1926 // statepoints surviving this pass. This makes testing easier and the
1927 // resulting IR less confusing to human readers. Rather than be fancy, we
1928 // just reuse a utility function which removes the unreachable blocks.
1929 if (HasUnreachableStatepoint)
1930 MadeChange |= removeUnreachableBlocks(F);
1932 // Return early if no work to do.
1933 if (ParsePointNeeded.empty())
1936 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
1937 // These are created by LCSSA. They have the effect of increasing the size
1938 // of liveness sets for no good reason. It may be harder to do this post
1939 // insertion since relocations and base phis can confuse things.
1940 for (BasicBlock &BB : F)
1941 if (BB.getUniquePredecessor()) {
1943 FoldSingleEntryPHINodes(&BB);
1946 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
1950 // liveness computation via standard dataflow
1951 // -------------------------------------------------------------------
1953 // TODO: Consider using bitvectors for liveness, the set of potentially
1954 // interesting values should be small and easy to pre-compute.
1956 /// Is this value a constant consisting of entirely null values?
1957 static bool isConstantNull(Value *V) {
1958 return isa<Constant>(V) && cast<Constant>(V)->isNullValue();
1961 /// Compute the live-in set for the location rbegin starting from
1962 /// the live-out set of the basic block
1963 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
1964 BasicBlock::reverse_iterator rend,
1965 DenseSet<Value *> &LiveTmp) {
1967 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
1968 Instruction *I = &*ritr;
1970 // KILL/Def - Remove this definition from LiveIn
1973 // Don't consider *uses* in PHI nodes, we handle their contribution to
1974 // predecessor blocks when we seed the LiveOut sets
1975 if (isa<PHINode>(I))
1978 // USE - Add to the LiveIn set for this instruction
1979 for (Value *V : I->operands()) {
1980 assert(!isUnhandledGCPointerType(V->getType()) &&
1981 "support for FCA unimplemented");
1982 if (isHandledGCPointerType(V->getType()) && !isConstantNull(V) &&
1983 !isa<UndefValue>(V)) {
1984 // The choice to exclude null and undef is arbitrary here. Reconsider?
1991 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
1993 for (BasicBlock *Succ : successors(BB)) {
1994 const BasicBlock::iterator E(Succ->getFirstNonPHI());
1995 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
1996 PHINode *Phi = cast<PHINode>(&*I);
1997 Value *V = Phi->getIncomingValueForBlock(BB);
1998 assert(!isUnhandledGCPointerType(V->getType()) &&
1999 "support for FCA unimplemented");
2000 if (isHandledGCPointerType(V->getType()) && !isConstantNull(V) &&
2001 !isa<UndefValue>(V)) {
2002 // The choice to exclude null and undef is arbitrary here. Reconsider?
2009 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2010 DenseSet<Value *> KillSet;
2011 for (Instruction &I : *BB)
2012 if (isHandledGCPointerType(I.getType()))
2018 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2019 /// sanity check for the liveness computation.
2020 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2021 TerminatorInst *TI, bool TermOkay = false) {
2022 for (Value *V : Live) {
2023 if (auto *I = dyn_cast<Instruction>(V)) {
2024 // The terminator can be a member of the LiveOut set. LLVM's definition
2025 // of instruction dominance states that V does not dominate itself. As
2026 // such, we need to special case this to allow it.
2027 if (TermOkay && TI == I)
2029 assert(DT.dominates(I, TI) &&
2030 "basic SSA liveness expectation violated by liveness analysis");
2035 /// Check that all the liveness sets used during the computation of liveness
2036 /// obey basic SSA properties. This is useful for finding cases where we miss
2038 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2040 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2041 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2042 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2046 static void computeLiveInValues(DominatorTree &DT, Function &F,
2047 GCPtrLivenessData &Data) {
2049 SmallSetVector<BasicBlock *, 200> Worklist;
2050 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2051 // We use a SetVector so that we don't have duplicates in the worklist.
2052 Worklist.insert(pred_begin(BB), pred_end(BB));
2054 auto NextItem = [&]() {
2055 BasicBlock *BB = Worklist.back();
2056 Worklist.pop_back();
2060 // Seed the liveness for each individual block
2061 for (BasicBlock &BB : F) {
2062 Data.KillSet[&BB] = computeKillSet(&BB);
2063 Data.LiveSet[&BB].clear();
2064 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2067 for (Value *Kill : Data.KillSet[&BB])
2068 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2071 Data.LiveOut[&BB] = DenseSet<Value *>();
2072 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2073 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2074 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2075 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2076 if (!Data.LiveIn[&BB].empty())
2077 AddPredsToWorklist(&BB);
2080 // Propagate that liveness until stable
2081 while (!Worklist.empty()) {
2082 BasicBlock *BB = NextItem();
2084 // Compute our new liveout set, then exit early if it hasn't changed
2085 // despite the contribution of our successor.
2086 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2087 const auto OldLiveOutSize = LiveOut.size();
2088 for (BasicBlock *Succ : successors(BB)) {
2089 assert(Data.LiveIn.count(Succ));
2090 set_union(LiveOut, Data.LiveIn[Succ]);
2092 // assert OutLiveOut is a subset of LiveOut
2093 if (OldLiveOutSize == LiveOut.size()) {
2094 // If the sets are the same size, then we didn't actually add anything
2095 // when unioning our successors LiveIn Thus, the LiveIn of this block
2099 Data.LiveOut[BB] = LiveOut;
2101 // Apply the effects of this basic block
2102 DenseSet<Value *> LiveTmp = LiveOut;
2103 set_union(LiveTmp, Data.LiveSet[BB]);
2104 set_subtract(LiveTmp, Data.KillSet[BB]);
2106 assert(Data.LiveIn.count(BB));
2107 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2108 // assert: OldLiveIn is a subset of LiveTmp
2109 if (OldLiveIn.size() != LiveTmp.size()) {
2110 Data.LiveIn[BB] = LiveTmp;
2111 AddPredsToWorklist(BB);
2113 } // while( !worklist.empty() )
2116 // Sanity check our ouput against SSA properties. This helps catch any
2117 // missing kills during the above iteration.
2118 for (BasicBlock &BB : F) {
2119 checkBasicSSA(DT, Data, BB);
2124 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2125 StatepointLiveSetTy &Out) {
2127 BasicBlock *BB = Inst->getParent();
2129 // Note: The copy is intentional and required
2130 assert(Data.LiveOut.count(BB));
2131 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2133 // We want to handle the statepoint itself oddly. It's
2134 // call result is not live (normal), nor are it's arguments
2135 // (unless they're used again later). This adjustment is
2136 // specifically what we need to relocate
2137 BasicBlock::reverse_iterator rend(Inst);
2138 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2139 LiveOut.erase(Inst);
2140 Out.insert(LiveOut.begin(), LiveOut.end());
2143 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2145 PartiallyConstructedSafepointRecord &Info) {
2146 Instruction *Inst = CS.getInstruction();
2147 StatepointLiveSetTy Updated;
2148 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2151 DenseSet<Value *> Bases;
2152 for (auto KVPair : Info.PointerToBase) {
2153 Bases.insert(KVPair.second);
2156 // We may have base pointers which are now live that weren't before. We need
2157 // to update the PointerToBase structure to reflect this.
2158 for (auto V : Updated)
2159 if (!Info.PointerToBase.count(V)) {
2160 assert(Bases.count(V) && "can't find base for unexpected live value");
2161 Info.PointerToBase[V] = V;
2166 for (auto V : Updated) {
2167 assert(Info.PointerToBase.count(V) &&
2168 "must be able to find base for live value");
2172 // Remove any stale base mappings - this can happen since our liveness is
2173 // more precise then the one inherent in the base pointer analysis
2174 DenseSet<Value *> ToErase;
2175 for (auto KVPair : Info.PointerToBase)
2176 if (!Updated.count(KVPair.first))
2177 ToErase.insert(KVPair.first);
2178 for (auto V : ToErase)
2179 Info.PointerToBase.erase(V);
2182 for (auto KVPair : Info.PointerToBase)
2183 assert(Updated.count(KVPair.first) && "record for non-live value");
2186 Info.liveset = Updated;