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/IR/BasicBlock.h"
21 #include "llvm/IR/CallSite.h"
22 #include "llvm/IR/Dominators.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/IRBuilder.h"
25 #include "llvm/IR/InstIterator.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/Intrinsics.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/IR/Statepoint.h"
31 #include "llvm/IR/Value.h"
32 #include "llvm/IR/Verifier.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Transforms/Scalar.h"
36 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
37 #include "llvm/Transforms/Utils/Cloning.h"
38 #include "llvm/Transforms/Utils/Local.h"
39 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
41 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
45 // Print tracing output
46 static cl::opt<bool> TraceLSP("trace-rewrite-statepoints", cl::Hidden,
49 // Print the liveset found at the insert location
50 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
52 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size",
53 cl::Hidden, cl::init(false));
54 // Print out the base pointers for debugging
55 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers",
56 cl::Hidden, cl::init(false));
59 struct RewriteStatepointsForGC : public FunctionPass {
60 static char ID; // Pass identification, replacement for typeid
62 RewriteStatepointsForGC() : FunctionPass(ID) {
63 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
65 bool runOnFunction(Function &F) override;
67 void getAnalysisUsage(AnalysisUsage &AU) const override {
68 // We add and rewrite a bunch of instructions, but don't really do much
69 // else. We could in theory preserve a lot more analyses here.
70 AU.addRequired<DominatorTreeWrapperPass>();
75 char RewriteStatepointsForGC::ID = 0;
77 FunctionPass *llvm::createRewriteStatepointsForGCPass() {
78 return new RewriteStatepointsForGC();
81 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
82 "Make relocations explicit at statepoints", false, false)
83 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
84 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
85 "Make relocations explicit at statepoints", false, false)
88 // The type of the internal cache used inside the findBasePointers family
89 // of functions. From the callers perspective, this is an opaque type and
90 // should not be inspected.
92 // In the actual implementation this caches two relations:
93 // - The base relation itself (i.e. this pointer is based on that one)
94 // - The base defining value relation (i.e. before base_phi insertion)
95 // Generally, after the execution of a full findBasePointer call, only the
96 // base relation will remain. Internally, we add a mixture of the two
97 // types, then update all the second type to the first type
98 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
99 typedef DenseSet<llvm::Value *> StatepointLiveSetTy;
101 struct PartiallyConstructedSafepointRecord {
102 /// The set of values known to be live accross this safepoint
103 StatepointLiveSetTy liveset;
105 /// Mapping from live pointers to a base-defining-value
106 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
108 /// Any new values which were added to the IR during base pointer analysis
109 /// for this safepoint
110 DenseSet<llvm::Value *> NewInsertedDefs;
112 /// The *new* gc.statepoint instruction itself. This produces the token
113 /// that normal path gc.relocates and the gc.result are tied to.
114 Instruction *StatepointToken;
116 /// Instruction to which exceptional gc relocates are attached
117 /// Makes it easier to iterate through them during relocationViaAlloca.
118 Instruction *UnwindToken;
122 // TODO: Once we can get to the GCStrategy, this becomes
123 // Optional<bool> isGCManagedPointer(const Value *V) const override {
125 static bool isGCPointerType(const Type *T) {
126 if (const PointerType *PT = dyn_cast<PointerType>(T))
127 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
128 // GC managed heap. We know that a pointer into this heap needs to be
129 // updated and that no other pointer does.
130 return (1 == PT->getAddressSpace());
134 // Return true if this type is one which a) is a gc pointer or contains a GC
135 // pointer and b) is of a type this code expects to encounter as a live value.
136 // (The insertion code will assert that a type which matches (a) and not (b)
137 // is not encountered.)
138 static bool isHandledGCPointerType(Type *T) {
139 // We fully support gc pointers
140 if (isGCPointerType(T))
142 // We partially support vectors of gc pointers. The code will assert if it
143 // can't handle something.
144 if (auto VT = dyn_cast<VectorType>(T))
145 if (isGCPointerType(VT->getElementType()))
151 /// Returns true if this type contains a gc pointer whether we know how to
152 /// handle that type or not.
153 static bool containsGCPtrType(Type *Ty) {
154 if(isGCPointerType(Ty))
156 if (VectorType *VT = dyn_cast<VectorType>(Ty))
157 return isGCPointerType(VT->getScalarType());
158 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
159 return containsGCPtrType(AT->getElementType());
160 if (StructType *ST = dyn_cast<StructType>(Ty))
161 return std::any_of(ST->subtypes().begin(), ST->subtypes().end(),
163 return containsGCPtrType(SubType);
168 // Returns true if this is a type which a) is a gc pointer or contains a GC
169 // pointer and b) is of a type which the code doesn't expect (i.e. first class
170 // aggregates). Used to trip assertions.
171 static bool isUnhandledGCPointerType(Type *Ty) {
172 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
176 /// Return true if the Value is a gc reference type which is potentially used
177 /// after the instruction 'loc'. This is only used with the edge reachability
178 /// liveness code. Note: It is assumed the V dominates loc.
179 static bool isLiveGCReferenceAt(Value &V, Instruction *Loc, DominatorTree &DT,
181 if (!isHandledGCPointerType(V.getType()))
187 // Given assumption that V dominates loc, this may be live
191 // Conservatively identifies any definitions which might be live at the
192 // given instruction. The analysis is performed immediately before the
193 // given instruction. Values defined by that instruction are not considered
194 // live. Values used by that instruction are considered live.
196 // preconditions: valid IR graph, term is either a terminator instruction or
197 // a call instruction, pred is the basic block of term, DT, LI are valid
199 // side effects: none, does not mutate IR
201 // postconditions: populates liveValues as discussed above
202 static void findLiveGCValuesAtInst(Instruction *term, BasicBlock *pred,
203 DominatorTree &DT, LoopInfo *LI,
204 StatepointLiveSetTy &liveValues) {
207 assert(isa<CallInst>(term) || isa<InvokeInst>(term) || term->isTerminator());
209 Function *F = pred->getParent();
211 auto is_live_gc_reference =
212 [&](Value &V) { return isLiveGCReferenceAt(V, term, DT, LI); };
214 // Are there any gc pointer arguments live over this point? This needs to be
215 // special cased since arguments aren't defined in basic blocks.
216 for (Argument &arg : F->args()) {
217 assert(!isUnhandledGCPointerType(arg.getType()) &&
218 "support for FCA unimplemented");
220 if (is_live_gc_reference(arg)) {
221 liveValues.insert(&arg);
225 // Walk through all dominating blocks - the ones which can contain
226 // definitions used in this block - and check to see if any of the values
227 // they define are used in locations potentially reachable from the
228 // interesting instruction.
229 BasicBlock *BBI = pred;
232 errs() << "[LSP] Looking at dominating block " << pred->getName() << "\n";
234 assert(DT.dominates(BBI, pred));
235 assert(isPotentiallyReachable(BBI, pred, &DT) &&
236 "dominated block must be reachable");
238 // Walk through the instructions in dominating blocks and keep any
239 // that have a use potentially reachable from the block we're
240 // considering putting the safepoint in
241 for (Instruction &inst : *BBI) {
243 errs() << "[LSP] Looking at instruction ";
247 if (pred == BBI && (&inst) == term) {
249 errs() << "[LSP] stopped because we encountered the safepoint "
253 // If we're in the block which defines the interesting instruction,
254 // we don't want to include any values as live which are defined
255 // _after_ the interesting line or as part of the line itself
256 // i.e. "term" is the call instruction for a call safepoint, the
257 // results of the call should not be considered live in that stackmap
261 assert(!isUnhandledGCPointerType(inst.getType()) &&
262 "support for FCA unimplemented");
264 if (is_live_gc_reference(inst)) {
266 errs() << "[LSP] found live value for this safepoint ";
270 liveValues.insert(&inst);
273 if (!DT.getNode(BBI)->getIDom()) {
274 assert(BBI == &F->getEntryBlock() &&
275 "failed to find a dominator for something other than "
279 BBI = DT.getNode(BBI)->getIDom()->getBlock();
283 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
284 if (a->hasName() && b->hasName()) {
285 return -1 == a->getName().compare(b->getName());
286 } else if (a->hasName() && !b->hasName()) {
288 } else if (!a->hasName() && b->hasName()) {
291 // Better than nothing, but not stable
296 /// Find the initial live set. Note that due to base pointer
297 /// insertion, the live set may be incomplete.
299 analyzeParsePointLiveness(DominatorTree &DT, const CallSite &CS,
300 PartiallyConstructedSafepointRecord &result) {
301 Instruction *inst = CS.getInstruction();
303 BasicBlock *BB = inst->getParent();
304 StatepointLiveSetTy liveset;
305 findLiveGCValuesAtInst(inst, BB, DT, nullptr, liveset);
308 // Note: This output is used by several of the test cases
309 // The order of elemtns in a set is not stable, put them in a vec and sort
311 SmallVector<Value *, 64> temp;
312 temp.insert(temp.end(), liveset.begin(), liveset.end());
313 std::sort(temp.begin(), temp.end(), order_by_name);
314 errs() << "Live Variables:\n";
315 for (Value *V : temp) {
316 errs() << " " << V->getName(); // no newline
320 if (PrintLiveSetSize) {
321 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
322 errs() << "Number live values: " << liveset.size() << "\n";
324 result.liveset = liveset;
327 /// If we can trivially determine that this vector contains only base pointers,
328 /// return the base instruction.
329 static Value *findBaseOfVector(Value *I) {
330 assert(I->getType()->isVectorTy() &&
331 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
332 "Illegal to ask for the base pointer of a non-pointer type");
334 // Each case parallels findBaseDefiningValue below, see that code for
335 // detailed motivation.
337 if (isa<Argument>(I))
338 // An incoming argument to the function is a base pointer
341 // We shouldn't see the address of a global as a vector value?
342 assert(!isa<GlobalVariable>(I) &&
343 "unexpected global variable found in base of vector");
345 // inlining could possibly introduce phi node that contains
346 // undef if callee has multiple returns
347 if (isa<UndefValue>(I))
348 // utterly meaningless, but useful for dealing with partially optimized
352 // Due to inheritance, this must be _after_ the global variable and undef
354 if (Constant *Con = dyn_cast<Constant>(I)) {
355 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
356 "order of checks wrong!");
357 assert(Con->isNullValue() && "null is the only case which makes sense");
361 if (isa<LoadInst>(I))
364 // Note: This code is currently rather incomplete. We are essentially only
365 // handling cases where the vector element is trivially a base pointer. We
366 // need to update the entire base pointer construction algorithm to know how
367 // to track vector elements and potentially scalarize, but the case which
368 // would motivate the work hasn't shown up in real workloads yet.
369 llvm_unreachable("no base found for vector element");
372 /// Helper function for findBasePointer - Will return a value which either a)
373 /// defines the base pointer for the input or b) blocks the simple search
374 /// (i.e. a PHI or Select of two derived pointers)
375 static Value *findBaseDefiningValue(Value *I) {
376 assert(I->getType()->isPointerTy() &&
377 "Illegal to ask for the base pointer of a non-pointer type");
379 // This case is a bit of a hack - it only handles extracts from vectors which
380 // trivially contain only base pointers. See note inside the function for
381 // how to improve this.
382 if (auto *EEI = dyn_cast<ExtractElementInst>(I)) {
383 Value *VectorOperand = EEI->getVectorOperand();
384 Value *VectorBase = findBaseOfVector(VectorOperand);
386 assert(VectorBase && "extract element not known to be a trivial base");
390 if (isa<Argument>(I))
391 // An incoming argument to the function is a base pointer
392 // We should have never reached here if this argument isn't an gc value
395 if (isa<GlobalVariable>(I))
399 // inlining could possibly introduce phi node that contains
400 // undef if callee has multiple returns
401 if (isa<UndefValue>(I))
402 // utterly meaningless, but useful for dealing with
403 // partially optimized code.
406 // Due to inheritance, this must be _after_ the global variable and undef
408 if (Constant *Con = dyn_cast<Constant>(I)) {
409 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
410 "order of checks wrong!");
411 // Note: Finding a constant base for something marked for relocation
412 // doesn't really make sense. The most likely case is either a) some
413 // screwed up the address space usage or b) your validating against
414 // compiled C++ code w/o the proper separation. The only real exception
415 // is a null pointer. You could have generic code written to index of
416 // off a potentially null value and have proven it null. We also use
417 // null pointers in dead paths of relocation phis (which we might later
418 // want to find a base pointer for).
419 assert(isa<ConstantPointerNull>(Con) &&
420 "null is the only case which makes sense");
424 if (CastInst *CI = dyn_cast<CastInst>(I)) {
425 Value *Def = CI->stripPointerCasts();
426 // If we find a cast instruction here, it means we've found a cast which is
427 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
428 // handle int->ptr conversion.
429 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
430 return findBaseDefiningValue(Def);
433 if (isa<LoadInst>(I))
434 return I; // The value loaded is an gc base itself
436 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
437 // The base of this GEP is the base
438 return findBaseDefiningValue(GEP->getPointerOperand());
440 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
441 switch (II->getIntrinsicID()) {
442 case Intrinsic::experimental_gc_result_ptr:
444 // fall through to general call handling
446 case Intrinsic::experimental_gc_statepoint:
447 case Intrinsic::experimental_gc_result_float:
448 case Intrinsic::experimental_gc_result_int:
449 llvm_unreachable("these don't produce pointers");
450 case Intrinsic::experimental_gc_relocate: {
451 // Rerunning safepoint insertion after safepoints are already
452 // inserted is not supported. It could probably be made to work,
453 // but why are you doing this? There's no good reason.
454 llvm_unreachable("repeat safepoint insertion is not supported");
456 case Intrinsic::gcroot:
457 // Currently, this mechanism hasn't been extended to work with gcroot.
458 // There's no reason it couldn't be, but I haven't thought about the
459 // implications much.
461 "interaction with the gcroot mechanism is not supported");
464 // We assume that functions in the source language only return base
465 // pointers. This should probably be generalized via attributes to support
466 // both source language and internal functions.
467 if (isa<CallInst>(I) || isa<InvokeInst>(I))
470 // I have absolutely no idea how to implement this part yet. It's not
471 // neccessarily hard, I just haven't really looked at it yet.
472 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
474 if (isa<AtomicCmpXchgInst>(I))
475 // A CAS is effectively a atomic store and load combined under a
476 // predicate. From the perspective of base pointers, we just treat it
480 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
481 "binary ops which don't apply to pointers");
483 // The aggregate ops. Aggregates can either be in the heap or on the
484 // stack, but in either case, this is simply a field load. As a result,
485 // this is a defining definition of the base just like a load is.
486 if (isa<ExtractValueInst>(I))
489 // We should never see an insert vector since that would require we be
490 // tracing back a struct value not a pointer value.
491 assert(!isa<InsertValueInst>(I) &&
492 "Base pointer for a struct is meaningless");
494 // The last two cases here don't return a base pointer. Instead, they
495 // return a value which dynamically selects from amoung several base
496 // derived pointers (each with it's own base potentially). It's the job of
497 // the caller to resolve these.
498 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
499 "missing instruction case in findBaseDefiningValing");
503 /// Returns the base defining value for this value.
504 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
505 Value *&Cached = Cache[I];
507 Cached = findBaseDefiningValue(I);
509 assert(Cache[I] != nullptr);
512 dbgs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
518 /// Return a base pointer for this value if known. Otherwise, return it's
519 /// base defining value.
520 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
521 Value *Def = findBaseDefiningValueCached(I, Cache);
522 auto Found = Cache.find(Def);
523 if (Found != Cache.end()) {
524 // Either a base-of relation, or a self reference. Caller must check.
525 return Found->second;
527 // Only a BDV available
531 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
532 /// is it known to be a base pointer? Or do we need to continue searching.
533 static bool isKnownBaseResult(Value *V) {
534 if (!isa<PHINode>(V) && !isa<SelectInst>(V)) {
535 // no recursion possible
538 if (isa<Instruction>(V) &&
539 cast<Instruction>(V)->getMetadata("is_base_value")) {
540 // This is a previously inserted base phi or select. We know
541 // that this is a base value.
545 // We need to keep searching
549 // TODO: find a better name for this
553 enum Status { Unknown, Base, Conflict };
555 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
556 assert(status != Base || b);
558 PhiState(Value *b) : status(Base), base(b) {}
559 PhiState() : status(Unknown), base(nullptr) {}
561 Status getStatus() const { return status; }
562 Value *getBase() const { return base; }
564 bool isBase() const { return getStatus() == Base; }
565 bool isUnknown() const { return getStatus() == Unknown; }
566 bool isConflict() const { return getStatus() == Conflict; }
568 bool operator==(const PhiState &other) const {
569 return base == other.base && status == other.status;
572 bool operator!=(const PhiState &other) const { return !(*this == other); }
575 errs() << status << " (" << base << " - "
576 << (base ? base->getName() : "nullptr") << "): ";
581 Value *base; // non null only if status == base
584 typedef DenseMap<Value *, PhiState> ConflictStateMapTy;
585 // Values of type PhiState form a lattice, and this is a helper
586 // class that implementes the meet operation. The meat of the meet
587 // operation is implemented in MeetPhiStates::pureMeet
588 class MeetPhiStates {
590 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
591 explicit MeetPhiStates(const ConflictStateMapTy &phiStates)
592 : phiStates(phiStates) {}
594 // Destructively meet the current result with the base V. V can
595 // either be a merge instruction (SelectInst / PHINode), in which
596 // case its status is looked up in the phiStates map; or a regular
597 // SSA value, in which case it is assumed to be a base.
598 void meetWith(Value *V) {
599 PhiState otherState = getStateForBDV(V);
600 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
601 MeetPhiStates::pureMeet(currentResult, otherState)) &&
602 "math is wrong: meet does not commute!");
603 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
606 PhiState getResult() const { return currentResult; }
609 const ConflictStateMapTy &phiStates;
610 PhiState currentResult;
612 /// Return a phi state for a base defining value. We'll generate a new
613 /// base state for known bases and expect to find a cached state otherwise
614 PhiState getStateForBDV(Value *baseValue) {
615 if (isKnownBaseResult(baseValue)) {
616 return PhiState(baseValue);
618 return lookupFromMap(baseValue);
622 PhiState lookupFromMap(Value *V) {
623 auto I = phiStates.find(V);
624 assert(I != phiStates.end() && "lookup failed!");
628 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
629 switch (stateA.getStatus()) {
630 case PhiState::Unknown:
634 assert(stateA.getBase() && "can't be null");
635 if (stateB.isUnknown())
638 if (stateB.isBase()) {
639 if (stateA.getBase() == stateB.getBase()) {
640 assert(stateA == stateB && "equality broken!");
643 return PhiState(PhiState::Conflict);
645 assert(stateB.isConflict() && "only three states!");
646 return PhiState(PhiState::Conflict);
648 case PhiState::Conflict:
651 llvm_unreachable("only three states!");
655 /// For a given value or instruction, figure out what base ptr it's derived
656 /// from. For gc objects, this is simply itself. On success, returns a value
657 /// which is the base pointer. (This is reliable and can be used for
658 /// relocation.) On failure, returns nullptr.
659 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache,
660 DenseSet<llvm::Value *> &NewInsertedDefs) {
661 Value *def = findBaseOrBDV(I, cache);
663 if (isKnownBaseResult(def)) {
667 // Here's the rough algorithm:
668 // - For every SSA value, construct a mapping to either an actual base
669 // pointer or a PHI which obscures the base pointer.
670 // - Construct a mapping from PHI to unknown TOP state. Use an
671 // optimistic algorithm to propagate base pointer information. Lattice
676 // When algorithm terminates, all PHIs will either have a single concrete
677 // base or be in a conflict state.
678 // - For every conflict, insert a dummy PHI node without arguments. Add
679 // these to the base[Instruction] = BasePtr mapping. For every
680 // non-conflict, add the actual base.
681 // - For every conflict, add arguments for the base[a] of each input
684 // Note: A simpler form of this would be to add the conflict form of all
685 // PHIs without running the optimistic algorithm. This would be
686 // analougous to pessimistic data flow and would likely lead to an
687 // overall worse solution.
689 ConflictStateMapTy states;
690 states[def] = PhiState();
691 // Recursively fill in all phis & selects reachable from the initial one
692 // for which we don't already know a definite base value for
693 // TODO: This should be rewritten with a worklist
697 // Since we're adding elements to 'states' as we run, we can't keep
698 // iterators into the set.
699 SmallVector<Value*, 16> Keys;
700 Keys.reserve(states.size());
701 for (auto Pair : states) {
702 Value *V = Pair.first;
705 for (Value *v : Keys) {
706 assert(!isKnownBaseResult(v) && "why did it get added?");
707 if (PHINode *phi = dyn_cast<PHINode>(v)) {
708 assert(phi->getNumIncomingValues() > 0 &&
709 "zero input phis are illegal");
710 for (Value *InVal : phi->incoming_values()) {
711 Value *local = findBaseOrBDV(InVal, cache);
712 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
713 states[local] = PhiState();
717 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
718 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
719 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
720 states[local] = PhiState();
723 local = findBaseOrBDV(sel->getFalseValue(), cache);
724 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
725 states[local] = PhiState();
733 errs() << "States after initialization:\n";
734 for (auto Pair : states) {
735 Instruction *v = cast<Instruction>(Pair.first);
736 PhiState state = Pair.second;
742 // TODO: come back and revisit the state transitions around inputs which
743 // have reached conflict state. The current version seems too conservative.
745 bool progress = true;
748 size_t oldSize = states.size();
751 // We're only changing keys in this loop, thus safe to keep iterators
752 for (auto Pair : states) {
753 MeetPhiStates calculateMeet(states);
754 Value *v = Pair.first;
755 assert(!isKnownBaseResult(v) && "why did it get added?");
756 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
757 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
758 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
760 for (Value *Val : cast<PHINode>(v)->incoming_values())
761 calculateMeet.meetWith(findBaseOrBDV(Val, cache));
763 PhiState oldState = states[v];
764 PhiState newState = calculateMeet.getResult();
765 if (oldState != newState) {
767 states[v] = newState;
771 assert(oldSize <= states.size());
772 assert(oldSize == states.size() || progress);
776 errs() << "States after meet iteration:\n";
777 for (auto Pair : states) {
778 Instruction *v = cast<Instruction>(Pair.first);
779 PhiState state = Pair.second;
785 // Insert Phis for all conflicts
786 // We want to keep naming deterministic in the loop that follows, so
787 // sort the keys before iteration. This is useful in allowing us to
788 // write stable tests. Note that there is no invalidation issue here.
789 SmallVector<Value*, 16> Keys;
790 Keys.reserve(states.size());
791 for (auto Pair : states) {
792 Value *V = Pair.first;
795 std::sort(Keys.begin(), Keys.end(), order_by_name);
796 // TODO: adjust naming patterns to avoid this order of iteration dependency
797 for (Value *V : Keys) {
798 Instruction *v = cast<Instruction>(V);
799 PhiState state = states[V];
800 assert(!isKnownBaseResult(v) && "why did it get added?");
801 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
802 if (!state.isConflict())
805 if (isa<PHINode>(v)) {
807 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
808 assert(num_preds > 0 && "how did we reach here");
809 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
810 NewInsertedDefs.insert(phi);
811 // Add metadata marking this as a base value
812 auto *const_1 = ConstantInt::get(
814 v->getParent()->getParent()->getParent()->getContext()),
816 auto MDConst = ConstantAsMetadata::get(const_1);
817 MDNode *md = MDNode::get(
818 v->getParent()->getParent()->getParent()->getContext(), MDConst);
819 phi->setMetadata("is_base_value", md);
820 states[v] = PhiState(PhiState::Conflict, phi);
822 SelectInst *sel = cast<SelectInst>(v);
823 // The undef will be replaced later
824 UndefValue *undef = UndefValue::get(sel->getType());
825 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
826 undef, "base_select", sel);
827 NewInsertedDefs.insert(basesel);
828 // Add metadata marking this as a base value
829 auto *const_1 = ConstantInt::get(
831 v->getParent()->getParent()->getParent()->getContext()),
833 auto MDConst = ConstantAsMetadata::get(const_1);
834 MDNode *md = MDNode::get(
835 v->getParent()->getParent()->getParent()->getContext(), MDConst);
836 basesel->setMetadata("is_base_value", md);
837 states[v] = PhiState(PhiState::Conflict, basesel);
841 // Fixup all the inputs of the new PHIs
842 for (auto Pair : states) {
843 Instruction *v = cast<Instruction>(Pair.first);
844 PhiState state = Pair.second;
846 assert(!isKnownBaseResult(v) && "why did it get added?");
847 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
848 if (!state.isConflict())
851 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
852 PHINode *phi = cast<PHINode>(v);
853 unsigned NumPHIValues = phi->getNumIncomingValues();
854 for (unsigned i = 0; i < NumPHIValues; i++) {
855 Value *InVal = phi->getIncomingValue(i);
856 BasicBlock *InBB = phi->getIncomingBlock(i);
858 // If we've already seen InBB, add the same incoming value
859 // we added for it earlier. The IR verifier requires phi
860 // nodes with multiple entries from the same basic block
861 // to have the same incoming value for each of those
862 // entries. If we don't do this check here and basephi
863 // has a different type than base, we'll end up adding two
864 // bitcasts (and hence two distinct values) as incoming
865 // values for the same basic block.
867 int blockIndex = basephi->getBasicBlockIndex(InBB);
868 if (blockIndex != -1) {
869 Value *oldBase = basephi->getIncomingValue(blockIndex);
870 basephi->addIncoming(oldBase, InBB);
872 Value *base = findBaseOrBDV(InVal, cache);
873 if (!isKnownBaseResult(base)) {
874 // Either conflict or base.
875 assert(states.count(base));
876 base = states[base].getBase();
877 assert(base != nullptr && "unknown PhiState!");
878 assert(NewInsertedDefs.count(base) &&
879 "should have already added this in a prev. iteration!");
882 // In essense this assert states: the only way two
883 // values incoming from the same basic block may be
884 // different is by being different bitcasts of the same
885 // value. A cleanup that remains TODO is changing
886 // findBaseOrBDV to return an llvm::Value of the correct
887 // type (and still remain pure). This will remove the
888 // need to add bitcasts.
889 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
890 "sanity -- findBaseOrBDV should be pure!");
895 // Find either the defining value for the PHI or the normal base for
897 Value *base = findBaseOrBDV(InVal, cache);
898 if (!isKnownBaseResult(base)) {
899 // Either conflict or base.
900 assert(states.count(base));
901 base = states[base].getBase();
902 assert(base != nullptr && "unknown PhiState!");
904 assert(base && "can't be null");
905 // Must use original input BB since base may not be Instruction
906 // The cast is needed since base traversal may strip away bitcasts
907 if (base->getType() != basephi->getType()) {
908 base = new BitCastInst(base, basephi->getType(), "cast",
909 InBB->getTerminator());
910 NewInsertedDefs.insert(base);
912 basephi->addIncoming(base, InBB);
914 assert(basephi->getNumIncomingValues() == NumPHIValues);
916 SelectInst *basesel = cast<SelectInst>(state.getBase());
917 SelectInst *sel = cast<SelectInst>(v);
918 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
919 // something more safe and less hacky.
920 for (int i = 1; i <= 2; i++) {
921 Value *InVal = sel->getOperand(i);
922 // Find either the defining value for the PHI or the normal base for
924 Value *base = findBaseOrBDV(InVal, cache);
925 if (!isKnownBaseResult(base)) {
926 // Either conflict or base.
927 assert(states.count(base));
928 base = states[base].getBase();
929 assert(base != nullptr && "unknown PhiState!");
931 assert(base && "can't be null");
932 // Must use original input BB since base may not be Instruction
933 // The cast is needed since base traversal may strip away bitcasts
934 if (base->getType() != basesel->getType()) {
935 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
936 NewInsertedDefs.insert(base);
938 basesel->setOperand(i, base);
943 // Cache all of our results so we can cheaply reuse them
944 // NOTE: This is actually two caches: one of the base defining value
945 // relation and one of the base pointer relation! FIXME
946 for (auto item : states) {
947 Value *v = item.first;
948 Value *base = item.second.getBase();
950 assert(!isKnownBaseResult(v) && "why did it get added?");
953 std::string fromstr =
954 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
956 errs() << "Updating base value cache"
957 << " for: " << (v->hasName() ? v->getName() : "")
958 << " from: " << fromstr
959 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
962 assert(isKnownBaseResult(base) &&
963 "must be something we 'know' is a base pointer");
964 if (cache.count(v)) {
965 // Once we transition from the BDV relation being store in the cache to
966 // the base relation being stored, it must be stable
967 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
968 "base relation should be stable");
972 assert(cache.find(def) != cache.end());
976 // For a set of live pointers (base and/or derived), identify the base
977 // pointer of the object which they are derived from. This routine will
978 // mutate the IR graph as needed to make the 'base' pointer live at the
979 // definition site of 'derived'. This ensures that any use of 'derived' can
980 // also use 'base'. This may involve the insertion of a number of
981 // additional PHI nodes.
983 // preconditions: live is a set of pointer type Values
985 // side effects: may insert PHI nodes into the existing CFG, will preserve
986 // CFG, will not remove or mutate any existing nodes
988 // post condition: PointerToBase contains one (derived, base) pair for every
989 // pointer in live. Note that derived can be equal to base if the original
990 // pointer was a base pointer.
991 static void findBasePointers(const StatepointLiveSetTy &live,
992 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
993 DominatorTree *DT, DefiningValueMapTy &DVCache,
994 DenseSet<llvm::Value *> &NewInsertedDefs) {
995 // For the naming of values inserted to be deterministic - which makes for
996 // much cleaner and more stable tests - we need to assign an order to the
997 // live values. DenseSets do not provide a deterministic order across runs.
998 SmallVector<Value*, 64> Temp;
999 Temp.insert(Temp.end(), live.begin(), live.end());
1000 std::sort(Temp.begin(), Temp.end(), order_by_name);
1001 for (Value *ptr : Temp) {
1002 Value *base = findBasePointer(ptr, DVCache, NewInsertedDefs);
1003 assert(base && "failed to find base pointer");
1004 PointerToBase[ptr] = base;
1005 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1006 DT->dominates(cast<Instruction>(base)->getParent(),
1007 cast<Instruction>(ptr)->getParent())) &&
1008 "The base we found better dominate the derived pointer");
1010 // If you see this trip and like to live really dangerously, the code should
1011 // be correct, just with idioms the verifier can't handle. You can try
1012 // disabling the verifier at your own substaintial risk.
1013 assert(!isa<ConstantPointerNull>(base) &&
1014 "the relocation code needs adjustment to handle the relocation of "
1015 "a null pointer constant without causing false positives in the "
1016 "safepoint ir verifier.");
1020 /// Find the required based pointers (and adjust the live set) for the given
1022 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1024 PartiallyConstructedSafepointRecord &result) {
1025 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
1026 DenseSet<llvm::Value *> NewInsertedDefs;
1027 findBasePointers(result.liveset, PointerToBase, &DT, DVCache, NewInsertedDefs);
1029 if (PrintBasePointers) {
1030 // Note: Need to print these in a stable order since this is checked in
1032 errs() << "Base Pairs (w/o Relocation):\n";
1033 SmallVector<Value*, 64> Temp;
1034 Temp.reserve(PointerToBase.size());
1035 for (auto Pair : PointerToBase) {
1036 Temp.push_back(Pair.first);
1038 std::sort(Temp.begin(), Temp.end(), order_by_name);
1039 for (Value *Ptr : Temp) {
1040 Value *Base = PointerToBase[Ptr];
1041 errs() << " derived %" << Ptr->getName() << " base %"
1042 << Base->getName() << "\n";
1046 result.PointerToBase = PointerToBase;
1047 result.NewInsertedDefs = NewInsertedDefs;
1050 /// Check for liveness of items in the insert defs and add them to the live
1051 /// and base pointer sets
1052 static void fixupLiveness(DominatorTree &DT, const CallSite &CS,
1053 const DenseSet<Value *> &allInsertedDefs,
1054 PartiallyConstructedSafepointRecord &result) {
1055 Instruction *inst = CS.getInstruction();
1057 auto liveset = result.liveset;
1058 auto PointerToBase = result.PointerToBase;
1060 auto is_live_gc_reference =
1061 [&](Value &V) { return isLiveGCReferenceAt(V, inst, DT, nullptr); };
1063 // For each new definition, check to see if a) the definition dominates the
1064 // instruction we're interested in, and b) one of the uses of that definition
1065 // is edge-reachable from the instruction we're interested in. This is the
1066 // same definition of liveness we used in the intial liveness analysis
1067 for (Value *newDef : allInsertedDefs) {
1068 if (liveset.count(newDef)) {
1069 // already live, no action needed
1073 // PERF: Use DT to check instruction domination might not be good for
1074 // compilation time, and we could change to optimal solution if this
1075 // turn to be a issue
1076 if (!DT.dominates(cast<Instruction>(newDef), inst)) {
1077 // can't possibly be live at inst
1081 if (is_live_gc_reference(*newDef)) {
1082 // Add the live new defs into liveset and PointerToBase
1083 liveset.insert(newDef);
1084 PointerToBase[newDef] = newDef;
1088 result.liveset = liveset;
1089 result.PointerToBase = PointerToBase;
1092 static void fixupLiveReferences(
1093 Function &F, DominatorTree &DT, Pass *P,
1094 const DenseSet<llvm::Value *> &allInsertedDefs,
1095 ArrayRef<CallSite> toUpdate,
1096 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1097 for (size_t i = 0; i < records.size(); i++) {
1098 struct PartiallyConstructedSafepointRecord &info = records[i];
1099 const CallSite &CS = toUpdate[i];
1100 fixupLiveness(DT, CS, allInsertedDefs, info);
1104 // Normalize basic block to make it ready to be target of invoke statepoint.
1105 // It means spliting it to have single predecessor. Return newly created BB
1106 // ready to be successor of invoke statepoint.
1107 static BasicBlock *normalizeBBForInvokeSafepoint(BasicBlock *BB,
1108 BasicBlock *InvokeParent,
1110 BasicBlock *ret = BB;
1112 if (!BB->getUniquePredecessor()) {
1113 ret = SplitBlockPredecessors(BB, InvokeParent, "");
1116 // Another requirement for such basic blocks is to not have any phi nodes.
1117 // Since we just ensured that new BB will have single predecessor,
1118 // all phi nodes in it will have one value. Here it would be naturall place
1120 // remove them all. But we can not do this because we are risking to remove
1121 // one of the values stored in liveset of another statepoint. We will do it
1122 // later after placing all safepoints.
1127 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1128 auto itr = std::find(livevec.begin(), livevec.end(), val);
1129 assert(livevec.end() != itr);
1130 size_t index = std::distance(livevec.begin(), itr);
1131 assert(index < livevec.size());
1135 // Create new attribute set containing only attributes which can be transfered
1136 // from original call to the safepoint.
1137 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1140 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1141 unsigned index = AS.getSlotIndex(Slot);
1143 if (index == AttributeSet::ReturnIndex ||
1144 index == AttributeSet::FunctionIndex) {
1146 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1148 Attribute attr = *it;
1150 // Do not allow certain attributes - just skip them
1151 // Safepoint can not be read only or read none.
1152 if (attr.hasAttribute(Attribute::ReadNone) ||
1153 attr.hasAttribute(Attribute::ReadOnly))
1156 ret = ret.addAttributes(
1157 AS.getContext(), index,
1158 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1162 // Just skip parameter attributes for now
1168 /// Helper function to place all gc relocates necessary for the given
1171 /// liveVariables - list of variables to be relocated.
1172 /// liveStart - index of the first live variable.
1173 /// basePtrs - base pointers.
1174 /// statepointToken - statepoint instruction to which relocates should be
1176 /// Builder - Llvm IR builder to be used to construct new calls.
1177 static void CreateGCRelocates(ArrayRef<llvm::Value *> liveVariables,
1178 const int liveStart,
1179 ArrayRef<llvm::Value *> basePtrs,
1180 Instruction *statepointToken,
1181 IRBuilder<> Builder) {
1182 SmallVector<Instruction *, 64> NewDefs;
1183 NewDefs.reserve(liveVariables.size());
1185 Module *M = statepointToken->getParent()->getParent()->getParent();
1187 for (unsigned i = 0; i < liveVariables.size(); i++) {
1188 // We generate a (potentially) unique declaration for every pointer type
1189 // combination. This results is some blow up the function declarations in
1190 // the IR, but removes the need for argument bitcasts which shrinks the IR
1191 // greatly and makes it much more readable.
1192 SmallVector<Type *, 1> types; // one per 'any' type
1193 types.push_back(liveVariables[i]->getType()); // result type
1194 Value *gc_relocate_decl = Intrinsic::getDeclaration(
1195 M, Intrinsic::experimental_gc_relocate, types);
1197 // Generate the gc.relocate call and save the result
1199 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1200 liveStart + find_index(liveVariables, basePtrs[i]));
1201 Value *liveIdx = ConstantInt::get(
1202 Type::getInt32Ty(M->getContext()),
1203 liveStart + find_index(liveVariables, liveVariables[i]));
1205 // only specify a debug name if we can give a useful one
1206 Value *reloc = Builder.CreateCall3(
1207 gc_relocate_decl, statepointToken, baseIdx, liveIdx,
1208 liveVariables[i]->hasName() ? liveVariables[i]->getName() + ".relocated"
1210 // Trick CodeGen into thinking there are lots of free registers at this
1212 cast<CallInst>(reloc)->setCallingConv(CallingConv::Cold);
1214 NewDefs.push_back(cast<Instruction>(reloc));
1216 assert(NewDefs.size() == liveVariables.size() &&
1217 "missing or extra redefinition at safepoint");
1221 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1222 const SmallVectorImpl<llvm::Value *> &basePtrs,
1223 const SmallVectorImpl<llvm::Value *> &liveVariables,
1225 PartiallyConstructedSafepointRecord &result) {
1226 assert(basePtrs.size() == liveVariables.size());
1227 assert(isStatepoint(CS) &&
1228 "This method expects to be rewriting a statepoint");
1230 BasicBlock *BB = CS.getInstruction()->getParent();
1232 Function *F = BB->getParent();
1233 assert(F && "must be set");
1234 Module *M = F->getParent();
1236 assert(M && "must be set");
1238 // We're not changing the function signature of the statepoint since the gc
1239 // arguments go into the var args section.
1240 Function *gc_statepoint_decl = CS.getCalledFunction();
1242 // Then go ahead and use the builder do actually do the inserts. We insert
1243 // immediately before the previous instruction under the assumption that all
1244 // arguments will be available here. We can't insert afterwards since we may
1245 // be replacing a terminator.
1246 Instruction *insertBefore = CS.getInstruction();
1247 IRBuilder<> Builder(insertBefore);
1248 // Copy all of the arguments from the original statepoint - this includes the
1249 // target, call args, and deopt args
1250 SmallVector<llvm::Value *, 64> args;
1251 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1252 // TODO: Clear the 'needs rewrite' flag
1254 // add all the pointers to be relocated (gc arguments)
1255 // Capture the start of the live variable list for use in the gc_relocates
1256 const int live_start = args.size();
1257 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1259 // Create the statepoint given all the arguments
1260 Instruction *token = nullptr;
1261 AttributeSet return_attributes;
1263 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1265 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1266 call->setTailCall(toReplace->isTailCall());
1267 call->setCallingConv(toReplace->getCallingConv());
1269 // Currently we will fail on parameter attributes and on certain
1270 // function attributes.
1271 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1272 // In case if we can handle this set of sttributes - set up function attrs
1273 // directly on statepoint and return attrs later for gc_result intrinsic.
1274 call->setAttributes(new_attrs.getFnAttributes());
1275 return_attributes = new_attrs.getRetAttributes();
1279 // Put the following gc_result and gc_relocate calls immediately after the
1280 // the old call (which we're about to delete)
1281 BasicBlock::iterator next(toReplace);
1282 assert(BB->end() != next && "not a terminator, must have next");
1284 Instruction *IP = &*(next);
1285 Builder.SetInsertPoint(IP);
1286 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1289 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1291 // Insert the new invoke into the old block. We'll remove the old one in a
1292 // moment at which point this will become the new terminator for the
1294 InvokeInst *invoke = InvokeInst::Create(
1295 gc_statepoint_decl, toReplace->getNormalDest(),
1296 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1297 invoke->setCallingConv(toReplace->getCallingConv());
1299 // Currently we will fail on parameter attributes and on certain
1300 // function attributes.
1301 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1302 // In case if we can handle this set of sttributes - set up function attrs
1303 // directly on statepoint and return attrs later for gc_result intrinsic.
1304 invoke->setAttributes(new_attrs.getFnAttributes());
1305 return_attributes = new_attrs.getRetAttributes();
1309 // Generate gc relocates in exceptional path
1310 BasicBlock *unwindBlock = normalizeBBForInvokeSafepoint(
1311 toReplace->getUnwindDest(), invoke->getParent(), P);
1313 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1314 Builder.SetInsertPoint(IP);
1315 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1317 // Extract second element from landingpad return value. We will attach
1318 // exceptional gc relocates to it.
1319 const unsigned idx = 1;
1320 Instruction *exceptional_token =
1321 cast<Instruction>(Builder.CreateExtractValue(
1322 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1323 result.UnwindToken = exceptional_token;
1325 // Just throw away return value. We will use the one we got for normal
1327 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1328 exceptional_token, Builder);
1330 // Generate gc relocates and returns for normal block
1331 BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
1332 toReplace->getNormalDest(), invoke->getParent(), P);
1334 IP = &*(normalDest->getFirstInsertionPt());
1335 Builder.SetInsertPoint(IP);
1337 // gc relocates will be generated later as if it were regular call
1342 // Take the name of the original value call if it had one.
1343 token->takeName(CS.getInstruction());
1345 // The GCResult is already inserted, we just need to find it
1347 Instruction *toReplace = CS.getInstruction();
1348 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1349 "only valid use before rewrite is gc.result");
1350 assert(!toReplace->hasOneUse() ||
1351 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1354 // Update the gc.result of the original statepoint (if any) to use the newly
1355 // inserted statepoint. This is safe to do here since the token can't be
1356 // considered a live reference.
1357 CS.getInstruction()->replaceAllUsesWith(token);
1359 result.StatepointToken = token;
1361 // Second, create a gc.relocate for every live variable
1362 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1367 struct name_ordering {
1370 bool operator()(name_ordering const &a, name_ordering const &b) {
1371 return -1 == a.derived->getName().compare(b.derived->getName());
1375 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1376 SmallVectorImpl<Value *> &livevec) {
1377 assert(basevec.size() == livevec.size());
1379 SmallVector<name_ordering, 64> temp;
1380 for (size_t i = 0; i < basevec.size(); i++) {
1382 v.base = basevec[i];
1383 v.derived = livevec[i];
1386 std::sort(temp.begin(), temp.end(), name_ordering());
1387 for (size_t i = 0; i < basevec.size(); i++) {
1388 basevec[i] = temp[i].base;
1389 livevec[i] = temp[i].derived;
1393 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1394 // which make the relocations happening at this safepoint explicit.
1396 // WARNING: Does not do any fixup to adjust users of the original live
1397 // values. That's the callers responsibility.
1399 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1400 PartiallyConstructedSafepointRecord &result) {
1401 auto liveset = result.liveset;
1402 auto PointerToBase = result.PointerToBase;
1404 // Convert to vector for efficient cross referencing.
1405 SmallVector<Value *, 64> basevec, livevec;
1406 livevec.reserve(liveset.size());
1407 basevec.reserve(liveset.size());
1408 for (Value *L : liveset) {
1409 livevec.push_back(L);
1411 assert(PointerToBase.find(L) != PointerToBase.end());
1412 Value *base = PointerToBase[L];
1413 basevec.push_back(base);
1415 assert(livevec.size() == basevec.size());
1417 // To make the output IR slightly more stable (for use in diffs), ensure a
1418 // fixed order of the values in the safepoint (by sorting the value name).
1419 // The order is otherwise meaningless.
1420 stablize_order(basevec, livevec);
1422 // Do the actual rewriting and delete the old statepoint
1423 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1424 CS.getInstruction()->eraseFromParent();
1427 // Helper function for the relocationViaAlloca.
1428 // It receives iterator to the statepoint gc relocates and emits store to the
1430 // location (via allocaMap) for the each one of them.
1431 // Add visited values into the visitedLiveValues set we will later use them
1432 // for sanity check.
1434 insertRelocationStores(iterator_range<Value::user_iterator> gcRelocs,
1435 DenseMap<Value *, Value *> &allocaMap,
1436 DenseSet<Value *> &visitedLiveValues) {
1438 for (User *U : gcRelocs) {
1439 if (!isa<IntrinsicInst>(U))
1442 IntrinsicInst *relocatedValue = cast<IntrinsicInst>(U);
1444 // We only care about relocates
1445 if (relocatedValue->getIntrinsicID() !=
1446 Intrinsic::experimental_gc_relocate) {
1450 GCRelocateOperands relocateOperands(relocatedValue);
1451 Value *originalValue = const_cast<Value *>(relocateOperands.derivedPtr());
1452 assert(allocaMap.count(originalValue));
1453 Value *alloca = allocaMap[originalValue];
1455 // Emit store into the related alloca
1456 StoreInst *store = new StoreInst(relocatedValue, alloca);
1457 store->insertAfter(relocatedValue);
1460 visitedLiveValues.insert(originalValue);
1465 /// do all the relocation update via allocas and mem2reg
1466 static void relocationViaAlloca(
1467 Function &F, DominatorTree &DT, ArrayRef<Value *> live,
1468 ArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1470 // record initial number of (static) allocas; we'll check we have the same
1471 // number when we get done.
1472 int InitialAllocaNum = 0;
1473 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end();
1475 if (isa<AllocaInst>(*I))
1479 // TODO-PERF: change data structures, reserve
1480 DenseMap<Value *, Value *> allocaMap;
1481 SmallVector<AllocaInst *, 200> PromotableAllocas;
1482 PromotableAllocas.reserve(live.size());
1484 // emit alloca for each live gc pointer
1485 for (unsigned i = 0; i < live.size(); i++) {
1486 Value *liveValue = live[i];
1487 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1488 F.getEntryBlock().getFirstNonPHI());
1489 allocaMap[liveValue] = alloca;
1490 PromotableAllocas.push_back(alloca);
1493 // The next two loops are part of the same conceptual operation. We need to
1494 // insert a store to the alloca after the original def and at each
1495 // redefinition. We need to insert a load before each use. These are split
1496 // into distinct loops for performance reasons.
1498 // update gc pointer after each statepoint
1499 // either store a relocated value or null (if no relocated value found for
1500 // this gc pointer and it is not a gc_result)
1501 // this must happen before we update the statepoint with load of alloca
1502 // otherwise we lose the link between statepoint and old def
1503 for (size_t i = 0; i < records.size(); i++) {
1504 const struct PartiallyConstructedSafepointRecord &info = records[i];
1505 Value *Statepoint = info.StatepointToken;
1507 // This will be used for consistency check
1508 DenseSet<Value *> visitedLiveValues;
1510 // Insert stores for normal statepoint gc relocates
1511 insertRelocationStores(Statepoint->users(), allocaMap, visitedLiveValues);
1513 // In case if it was invoke statepoint
1514 // we will insert stores for exceptional path gc relocates.
1515 if (isa<InvokeInst>(Statepoint)) {
1516 insertRelocationStores(info.UnwindToken->users(),
1517 allocaMap, visitedLiveValues);
1521 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1522 // the gc.statepoint. This will turn some subtle GC problems into slightly
1523 // easier to debug SEGVs
1524 SmallVector<AllocaInst *, 64> ToClobber;
1525 for (auto Pair : allocaMap) {
1526 Value *Def = Pair.first;
1527 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1529 // This value was relocated
1530 if (visitedLiveValues.count(Def)) {
1533 ToClobber.push_back(Alloca);
1536 auto InsertClobbersAt = [&](Instruction *IP) {
1537 for (auto *AI : ToClobber) {
1538 auto AIType = cast<PointerType>(AI->getType());
1539 auto PT = cast<PointerType>(AIType->getElementType());
1540 Constant *CPN = ConstantPointerNull::get(PT);
1541 StoreInst *store = new StoreInst(CPN, AI);
1542 store->insertBefore(IP);
1546 // Insert the clobbering stores. These may get intermixed with the
1547 // gc.results and gc.relocates, but that's fine.
1548 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1549 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1550 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1552 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1554 InsertClobbersAt(Next);
1558 // update use with load allocas and add store for gc_relocated
1559 for (auto Pair : allocaMap) {
1560 Value *def = Pair.first;
1561 Value *alloca = Pair.second;
1563 // we pre-record the uses of allocas so that we dont have to worry about
1565 // that change the user information.
1566 SmallVector<Instruction *, 20> uses;
1567 // PERF: trade a linear scan for repeated reallocation
1568 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1569 for (User *U : def->users()) {
1570 if (!isa<ConstantExpr>(U)) {
1571 // If the def has a ConstantExpr use, then the def is either a
1572 // ConstantExpr use itself or null. In either case
1573 // (recursively in the first, directly in the second), the oop
1574 // it is ultimately dependent on is null and this particular
1575 // use does not need to be fixed up.
1576 uses.push_back(cast<Instruction>(U));
1580 std::sort(uses.begin(), uses.end());
1581 auto last = std::unique(uses.begin(), uses.end());
1582 uses.erase(last, uses.end());
1584 for (Instruction *use : uses) {
1585 if (isa<PHINode>(use)) {
1586 PHINode *phi = cast<PHINode>(use);
1587 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1588 if (def == phi->getIncomingValue(i)) {
1589 LoadInst *load = new LoadInst(
1590 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1591 phi->setIncomingValue(i, load);
1595 LoadInst *load = new LoadInst(alloca, "", use);
1596 use->replaceUsesOfWith(def, load);
1600 // emit store for the initial gc value
1601 // store must be inserted after load, otherwise store will be in alloca's
1602 // use list and an extra load will be inserted before it
1603 StoreInst *store = new StoreInst(def, alloca);
1604 if (Instruction *inst = dyn_cast<Instruction>(def)) {
1605 if (InvokeInst *invoke = dyn_cast<InvokeInst>(inst)) {
1606 // InvokeInst is a TerminatorInst so the store need to be inserted
1607 // into its normal destination block.
1608 BasicBlock *normalDest = invoke->getNormalDest();
1609 store->insertBefore(normalDest->getFirstNonPHI());
1611 assert(!inst->isTerminator() &&
1612 "The only TerminatorInst that can produce a value is "
1613 "InvokeInst which is handled above.");
1614 store->insertAfter(inst);
1617 assert((isa<Argument>(def) || isa<GlobalVariable>(def) ||
1618 isa<ConstantPointerNull>(def)) &&
1619 "Must be argument or global");
1620 store->insertAfter(cast<Instruction>(alloca));
1624 assert(PromotableAllocas.size() == live.size() &&
1625 "we must have the same allocas with lives");
1626 if (!PromotableAllocas.empty()) {
1627 // apply mem2reg to promote alloca to SSA
1628 PromoteMemToReg(PromotableAllocas, DT);
1632 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end();
1634 if (isa<AllocaInst>(*I))
1636 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1640 /// Implement a unique function which doesn't require we sort the input
1641 /// vector. Doing so has the effect of changing the output of a couple of
1642 /// tests in ways which make them less useful in testing fused safepoints.
1643 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1645 SmallVector<T, 128> TempVec;
1646 TempVec.reserve(Vec.size());
1647 for (auto Element : Vec)
1648 TempVec.push_back(Element);
1650 for (auto V : TempVec) {
1651 if (Seen.insert(V).second) {
1657 static Function *getUseHolder(Module &M) {
1658 FunctionType *ftype =
1659 FunctionType::get(Type::getVoidTy(M.getContext()), true);
1660 Function *Func = cast<Function>(M.getOrInsertFunction("__tmp_use", ftype));
1664 /// Insert holders so that each Value is obviously live through the entire
1665 /// liftetime of the call.
1666 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1667 SmallVectorImpl<CallInst *> &holders) {
1668 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1669 Function *Func = getUseHolder(*M);
1671 // For call safepoints insert dummy calls right after safepoint
1672 BasicBlock::iterator next(CS.getInstruction());
1674 CallInst *base_holder = CallInst::Create(Func, Values, "", next);
1675 holders.push_back(base_holder);
1676 } else if (CS.isInvoke()) {
1677 // For invoke safepooints insert dummy calls both in normal and
1678 // exceptional destination blocks
1679 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1680 CallInst *normal_holder = CallInst::Create(
1681 Func, Values, "", invoke->getNormalDest()->getFirstInsertionPt());
1682 CallInst *unwind_holder = CallInst::Create(
1683 Func, Values, "", invoke->getUnwindDest()->getFirstInsertionPt());
1684 holders.push_back(normal_holder);
1685 holders.push_back(unwind_holder);
1687 llvm_unreachable("unsupported call type");
1690 static void findLiveReferences(
1691 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1692 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1693 for (size_t i = 0; i < records.size(); i++) {
1694 struct PartiallyConstructedSafepointRecord &info = records[i];
1695 const CallSite &CS = toUpdate[i];
1696 analyzeParsePointLiveness(DT, CS, info);
1700 static void addBasesAsLiveValues(StatepointLiveSetTy &liveset,
1701 DenseMap<Value *, Value *> &PointerToBase) {
1702 // Identify any base pointers which are used in this safepoint, but not
1703 // themselves relocated. We need to relocate them so that later inserted
1704 // safepoints can get the properly relocated base register.
1705 DenseSet<Value *> missing;
1706 for (Value *L : liveset) {
1707 assert(PointerToBase.find(L) != PointerToBase.end());
1708 Value *base = PointerToBase[L];
1710 if (liveset.find(base) == liveset.end()) {
1711 assert(PointerToBase.find(base) == PointerToBase.end());
1712 // uniqued by set insert
1713 missing.insert(base);
1717 // Note that we want these at the end of the list, otherwise
1718 // register placement gets screwed up once we lower to STATEPOINT
1719 // instructions. This is an utter hack, but there doesn't seem to be a
1721 for (Value *base : missing) {
1723 liveset.insert(base);
1724 PointerToBase[base] = base;
1726 assert(liveset.size() == PointerToBase.size());
1729 /// Remove any vector of pointers from the liveset by scalarizing them over the
1730 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1731 /// would be preferrable to include the vector in the statepoint itself, but
1732 /// the lowering code currently does not handle that. Extending it would be
1733 /// slightly non-trivial since it requires a format change. Given how rare
1734 /// such cases are (for the moment?) scalarizing is an acceptable comprimise.
1735 static void splitVectorValues(Instruction *StatepointInst,
1736 StatepointLiveSetTy& LiveSet, DominatorTree &DT) {
1737 SmallVector<Value *, 16> ToSplit;
1738 for (Value *V : LiveSet)
1739 if (isa<VectorType>(V->getType()))
1740 ToSplit.push_back(V);
1742 if (ToSplit.empty())
1745 Function &F = *(StatepointInst->getParent()->getParent());
1747 DenseMap<Value*, AllocaInst*> AllocaMap;
1748 // First is normal return, second is exceptional return (invoke only)
1749 DenseMap<Value*, std::pair<Value*,Value*>> Replacements;
1750 for (Value *V : ToSplit) {
1753 AllocaInst *Alloca = new AllocaInst(V->getType(), "",
1754 F.getEntryBlock().getFirstNonPHI());
1755 AllocaMap[V] = Alloca;
1757 VectorType *VT = cast<VectorType>(V->getType());
1758 IRBuilder<> Builder(StatepointInst);
1759 SmallVector<Value*, 16> Elements;
1760 for (unsigned i = 0; i < VT->getNumElements(); i++)
1761 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1762 LiveSet.insert(Elements.begin(), Elements.end());
1764 auto InsertVectorReform = [&](Instruction *IP) {
1765 Builder.SetInsertPoint(IP);
1766 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1767 Value *ResultVec = UndefValue::get(VT);
1768 for (unsigned i = 0; i < VT->getNumElements(); i++)
1769 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1770 Builder.getInt32(i));
1774 if (isa<CallInst>(StatepointInst)) {
1775 BasicBlock::iterator Next(StatepointInst);
1777 Instruction *IP = &*(Next);
1778 Replacements[V].first = InsertVectorReform(IP);
1779 Replacements[V].second = nullptr;
1781 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1782 // We've already normalized - check that we don't have shared destination
1784 BasicBlock *NormalDest = Invoke->getNormalDest();
1785 assert(!isa<PHINode>(NormalDest->begin()));
1786 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1787 assert(!isa<PHINode>(UnwindDest->begin()));
1788 // Insert insert element sequences in both successors
1789 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1790 Replacements[V].first = InsertVectorReform(IP);
1791 IP = &*(UnwindDest->getFirstInsertionPt());
1792 Replacements[V].second = InsertVectorReform(IP);
1795 for (Value *V : ToSplit) {
1796 AllocaInst *Alloca = AllocaMap[V];
1798 // Capture all users before we start mutating use lists
1799 SmallVector<Instruction*, 16> Users;
1800 for (User *U : V->users())
1801 Users.push_back(cast<Instruction>(U));
1803 for (Instruction *I : Users) {
1804 if (auto Phi = dyn_cast<PHINode>(I)) {
1805 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1806 if (V == Phi->getIncomingValue(i)) {
1807 LoadInst *Load = new LoadInst(Alloca, "",
1808 Phi->getIncomingBlock(i)->getTerminator());
1809 Phi->setIncomingValue(i, Load);
1812 LoadInst *Load = new LoadInst(Alloca, "", I);
1813 I->replaceUsesOfWith(V, Load);
1817 // Store the original value and the replacement value into the alloca
1818 StoreInst *Store = new StoreInst(V, Alloca);
1819 if (auto I = dyn_cast<Instruction>(V))
1820 Store->insertAfter(I);
1822 Store->insertAfter(Alloca);
1824 // Normal return for invoke, or call return
1825 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1826 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1827 // Unwind return for invoke only
1828 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1830 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1833 // apply mem2reg to promote alloca to SSA
1834 SmallVector<AllocaInst*, 16> Allocas;
1835 for (Value *V : ToSplit)
1836 Allocas.push_back(AllocaMap[V]);
1837 PromoteMemToReg(Allocas, DT);
1840 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1841 SmallVectorImpl<CallSite> &toUpdate) {
1843 // sanity check the input
1844 std::set<CallSite> uniqued;
1845 uniqued.insert(toUpdate.begin(), toUpdate.end());
1846 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1848 for (size_t i = 0; i < toUpdate.size(); i++) {
1849 CallSite &CS = toUpdate[i];
1850 assert(CS.getInstruction()->getParent()->getParent() == &F);
1851 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1855 // A list of dummy calls added to the IR to keep various values obviously
1856 // live in the IR. We'll remove all of these when done.
1857 SmallVector<CallInst *, 64> holders;
1859 // Insert a dummy call with all of the arguments to the vm_state we'll need
1860 // for the actual safepoint insertion. This ensures reference arguments in
1861 // the deopt argument list are considered live through the safepoint (and
1862 // thus makes sure they get relocated.)
1863 for (size_t i = 0; i < toUpdate.size(); i++) {
1864 CallSite &CS = toUpdate[i];
1865 Statepoint StatepointCS(CS);
1867 SmallVector<Value *, 64> DeoptValues;
1868 for (Use &U : StatepointCS.vm_state_args()) {
1869 Value *Arg = cast<Value>(&U);
1870 assert(!isUnhandledGCPointerType(Arg->getType()) &&
1871 "support for FCA unimplemented");
1872 if (isHandledGCPointerType(Arg->getType()))
1873 DeoptValues.push_back(Arg);
1875 insertUseHolderAfter(CS, DeoptValues, holders);
1878 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
1879 records.reserve(toUpdate.size());
1880 for (size_t i = 0; i < toUpdate.size(); i++) {
1881 struct PartiallyConstructedSafepointRecord info;
1882 records.push_back(info);
1884 assert(records.size() == toUpdate.size());
1886 // A) Identify all gc pointers which are staticly live at the given call
1888 findLiveReferences(F, DT, P, toUpdate, records);
1890 // Do a limited scalarization of any live at safepoint vector values which
1891 // contain pointers. This enables this pass to run after vectorization at
1892 // the cost of some possible performance loss. TODO: it would be nice to
1893 // natively support vectors all the way through the backend so we don't need
1894 // to scalarize here.
1895 for (size_t i = 0; i < records.size(); i++) {
1896 struct PartiallyConstructedSafepointRecord &info = records[i];
1897 Instruction *statepoint = toUpdate[i].getInstruction();
1898 splitVectorValues(cast<Instruction>(statepoint), info.liveset, DT);
1901 // B) Find the base pointers for each live pointer
1902 /* scope for caching */ {
1903 // Cache the 'defining value' relation used in the computation and
1904 // insertion of base phis and selects. This ensures that we don't insert
1905 // large numbers of duplicate base_phis.
1906 DefiningValueMapTy DVCache;
1908 for (size_t i = 0; i < records.size(); i++) {
1909 struct PartiallyConstructedSafepointRecord &info = records[i];
1910 CallSite &CS = toUpdate[i];
1911 findBasePointers(DT, DVCache, CS, info);
1913 } // end of cache scope
1915 // The base phi insertion logic (for any safepoint) may have inserted new
1916 // instructions which are now live at some safepoint. The simplest such
1919 // phi a <-- will be a new base_phi here
1920 // safepoint 1 <-- that needs to be live here
1924 DenseSet<llvm::Value *> allInsertedDefs;
1925 for (size_t i = 0; i < records.size(); i++) {
1926 struct PartiallyConstructedSafepointRecord &info = records[i];
1927 allInsertedDefs.insert(info.NewInsertedDefs.begin(),
1928 info.NewInsertedDefs.end());
1931 // We insert some dummy calls after each safepoint to definitely hold live
1932 // the base pointers which were identified for that safepoint. We'll then
1933 // ask liveness for _every_ base inserted to see what is now live. Then we
1934 // remove the dummy calls.
1935 holders.reserve(holders.size() + records.size());
1936 for (size_t i = 0; i < records.size(); i++) {
1937 struct PartiallyConstructedSafepointRecord &info = records[i];
1938 CallSite &CS = toUpdate[i];
1940 SmallVector<Value *, 128> Bases;
1941 for (auto Pair : info.PointerToBase) {
1942 Bases.push_back(Pair.second);
1944 insertUseHolderAfter(CS, Bases, holders);
1947 // Add the bases explicitly to the live vector set. This may result in a few
1948 // extra relocations, but the base has to be available whenever a pointer
1949 // derived from it is used. Thus, we need it to be part of the statepoint's
1950 // gc arguments list. TODO: Introduce an explicit notion (in the following
1951 // code) of the GC argument list as seperate from the live Values at a
1952 // given statepoint.
1953 for (size_t i = 0; i < records.size(); i++) {
1954 struct PartiallyConstructedSafepointRecord &info = records[i];
1955 addBasesAsLiveValues(info.liveset, info.PointerToBase);
1958 // If we inserted any new values, we need to adjust our notion of what is
1959 // live at a particular safepoint.
1960 if (!allInsertedDefs.empty()) {
1961 fixupLiveReferences(F, DT, P, allInsertedDefs, toUpdate, records);
1963 if (PrintBasePointers) {
1964 for (size_t i = 0; i < records.size(); i++) {
1965 struct PartiallyConstructedSafepointRecord &info = records[i];
1966 errs() << "Base Pairs: (w/Relocation)\n";
1967 for (auto Pair : info.PointerToBase) {
1968 errs() << " derived %" << Pair.first->getName() << " base %"
1969 << Pair.second->getName() << "\n";
1973 for (size_t i = 0; i < holders.size(); i++) {
1974 holders[i]->eraseFromParent();
1975 holders[i] = nullptr;
1979 // Now run through and replace the existing statepoints with new ones with
1980 // the live variables listed. We do not yet update uses of the values being
1981 // relocated. We have references to live variables that need to
1982 // survive to the last iteration of this loop. (By construction, the
1983 // previous statepoint can not be a live variable, thus we can and remove
1984 // the old statepoint calls as we go.)
1985 for (size_t i = 0; i < records.size(); i++) {
1986 struct PartiallyConstructedSafepointRecord &info = records[i];
1987 CallSite &CS = toUpdate[i];
1988 makeStatepointExplicit(DT, CS, P, info);
1990 toUpdate.clear(); // prevent accident use of invalid CallSites
1992 // In case if we inserted relocates in a different basic block than the
1993 // original safepoint (this can happen for invokes). We need to be sure that
1994 // original values were not used in any of the phi nodes at the
1995 // beginning of basic block containing them. Because we know that all such
1996 // blocks will have single predecessor we can safely assume that all phi
1997 // nodes have single entry (because of normalizeBBForInvokeSafepoint).
1998 // Just remove them all here.
1999 for (size_t i = 0; i < records.size(); i++) {
2000 Instruction *I = records[i].StatepointToken;
2002 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
2003 FoldSingleEntryPHINodes(invoke->getNormalDest());
2004 assert(!isa<PHINode>(invoke->getNormalDest()->begin()));
2006 FoldSingleEntryPHINodes(invoke->getUnwindDest());
2007 assert(!isa<PHINode>(invoke->getUnwindDest()->begin()));
2011 // Do all the fixups of the original live variables to their relocated selves
2012 SmallVector<Value *, 128> live;
2013 for (size_t i = 0; i < records.size(); i++) {
2014 struct PartiallyConstructedSafepointRecord &info = records[i];
2015 // We can't simply save the live set from the original insertion. One of
2016 // the live values might be the result of a call which needs a safepoint.
2017 // That Value* no longer exists and we need to use the new gc_result.
2018 // Thankfully, the liveset is embedded in the statepoint (and updated), so
2019 // we just grab that.
2020 Statepoint statepoint(info.StatepointToken);
2021 live.insert(live.end(), statepoint.gc_args_begin(),
2022 statepoint.gc_args_end());
2024 unique_unsorted(live);
2028 for (auto ptr : live) {
2029 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
2033 relocationViaAlloca(F, DT, live, records);
2034 return !records.empty();
2037 /// Returns true if this function should be rewritten by this pass. The main
2038 /// point of this function is as an extension point for custom logic.
2039 static bool shouldRewriteStatepointsIn(Function &F) {
2040 // TODO: This should check the GCStrategy
2042 const std::string StatepointExampleName("statepoint-example");
2043 return StatepointExampleName == F.getGC();
2048 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2049 // Nothing to do for declarations.
2050 if (F.isDeclaration() || F.empty())
2053 // Policy choice says not to rewrite - the most common reason is that we're
2054 // compiling code without a GCStrategy.
2055 if (!shouldRewriteStatepointsIn(F))
2058 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2060 // Gather all the statepoints which need rewritten. Be careful to only
2061 // consider those in reachable code since we need to ask dominance queries
2062 // when rewriting. We'll delete the unreachable ones in a moment.
2063 SmallVector<CallSite, 64> ParsePointNeeded;
2064 bool HasUnreachableStatepoint = false;
2065 for (Instruction &I : inst_range(F)) {
2066 // TODO: only the ones with the flag set!
2067 if (isStatepoint(I)) {
2068 if (DT.isReachableFromEntry(I.getParent()))
2069 ParsePointNeeded.push_back(CallSite(&I));
2071 HasUnreachableStatepoint = true;
2075 bool MadeChange = false;
2077 // Delete any unreachable statepoints so that we don't have unrewritten
2078 // statepoints surviving this pass. This makes testing easier and the
2079 // resulting IR less confusing to human readers. Rather than be fancy, we
2080 // just reuse a utility function which removes the unreachable blocks.
2081 if (HasUnreachableStatepoint)
2082 MadeChange |= removeUnreachableBlocks(F);
2084 // Return early if no work to do.
2085 if (ParsePointNeeded.empty())
2088 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2089 // These are created by LCSSA. They have the effect of increasing the size
2090 // of liveness sets for no good reason. It may be harder to do this post
2091 // insertion since relocations and base phis can confuse things.
2092 for (BasicBlock &BB : F)
2093 if (BB.getUniquePredecessor()) {
2095 FoldSingleEntryPHINodes(&BB);
2098 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);