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);
385 assert(VectorBase && "extract element not known to be a trivial base");
389 if (isa<Argument>(I))
390 // An incoming argument to the function is a base pointer
391 // We should have never reached here if this argument isn't an gc value
394 if (isa<GlobalVariable>(I))
398 // inlining could possibly introduce phi node that contains
399 // undef if callee has multiple returns
400 if (isa<UndefValue>(I))
401 // utterly meaningless, but useful for dealing with
402 // partially optimized code.
405 // Due to inheritance, this must be _after_ the global variable and undef
407 if (Constant *Con = dyn_cast<Constant>(I)) {
408 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
409 "order of checks wrong!");
410 // Note: Finding a constant base for something marked for relocation
411 // doesn't really make sense. The most likely case is either a) some
412 // screwed up the address space usage or b) your validating against
413 // compiled C++ code w/o the proper separation. The only real exception
414 // is a null pointer. You could have generic code written to index of
415 // off a potentially null value and have proven it null. We also use
416 // null pointers in dead paths of relocation phis (which we might later
417 // want to find a base pointer for).
418 assert(isa<ConstantPointerNull>(Con) &&
419 "null is the only case which makes sense");
423 if (CastInst *CI = dyn_cast<CastInst>(I)) {
424 Value *Def = CI->stripPointerCasts();
425 // If we find a cast instruction here, it means we've found a cast which is
426 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
427 // handle int->ptr conversion.
428 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
429 return findBaseDefiningValue(Def);
432 if (isa<LoadInst>(I))
433 return I; // The value loaded is an gc base itself
435 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
436 // The base of this GEP is the base
437 return findBaseDefiningValue(GEP->getPointerOperand());
439 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
440 switch (II->getIntrinsicID()) {
441 case Intrinsic::experimental_gc_result_ptr:
443 // fall through to general call handling
445 case Intrinsic::experimental_gc_statepoint:
446 case Intrinsic::experimental_gc_result_float:
447 case Intrinsic::experimental_gc_result_int:
448 llvm_unreachable("these don't produce pointers");
449 case Intrinsic::experimental_gc_relocate: {
450 // Rerunning safepoint insertion after safepoints are already
451 // inserted is not supported. It could probably be made to work,
452 // but why are you doing this? There's no good reason.
453 llvm_unreachable("repeat safepoint insertion is not supported");
455 case Intrinsic::gcroot:
456 // Currently, this mechanism hasn't been extended to work with gcroot.
457 // There's no reason it couldn't be, but I haven't thought about the
458 // implications much.
460 "interaction with the gcroot mechanism is not supported");
463 // We assume that functions in the source language only return base
464 // pointers. This should probably be generalized via attributes to support
465 // both source language and internal functions.
466 if (isa<CallInst>(I) || isa<InvokeInst>(I))
469 // I have absolutely no idea how to implement this part yet. It's not
470 // neccessarily hard, I just haven't really looked at it yet.
471 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
473 if (isa<AtomicCmpXchgInst>(I))
474 // A CAS is effectively a atomic store and load combined under a
475 // predicate. From the perspective of base pointers, we just treat it
479 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
480 "binary ops which don't apply to pointers");
482 // The aggregate ops. Aggregates can either be in the heap or on the
483 // stack, but in either case, this is simply a field load. As a result,
484 // this is a defining definition of the base just like a load is.
485 if (isa<ExtractValueInst>(I))
488 // We should never see an insert vector since that would require we be
489 // tracing back a struct value not a pointer value.
490 assert(!isa<InsertValueInst>(I) &&
491 "Base pointer for a struct is meaningless");
493 // The last two cases here don't return a base pointer. Instead, they
494 // return a value which dynamically selects from amoung several base
495 // derived pointers (each with it's own base potentially). It's the job of
496 // the caller to resolve these.
497 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
498 "missing instruction case in findBaseDefiningValing");
502 /// Returns the base defining value for this value.
503 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
504 Value *&Cached = Cache[I];
506 Cached = findBaseDefiningValue(I);
508 assert(Cache[I] != nullptr);
511 dbgs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
517 /// Return a base pointer for this value if known. Otherwise, return it's
518 /// base defining value.
519 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
520 Value *Def = findBaseDefiningValueCached(I, Cache);
521 auto Found = Cache.find(Def);
522 if (Found != Cache.end()) {
523 // Either a base-of relation, or a self reference. Caller must check.
524 return Found->second;
526 // Only a BDV available
530 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
531 /// is it known to be a base pointer? Or do we need to continue searching.
532 static bool isKnownBaseResult(Value *V) {
533 if (!isa<PHINode>(V) && !isa<SelectInst>(V)) {
534 // no recursion possible
537 if (isa<Instruction>(V) &&
538 cast<Instruction>(V)->getMetadata("is_base_value")) {
539 // This is a previously inserted base phi or select. We know
540 // that this is a base value.
544 // We need to keep searching
548 // TODO: find a better name for this
552 enum Status { Unknown, Base, Conflict };
554 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
555 assert(status != Base || b);
557 PhiState(Value *b) : status(Base), base(b) {}
558 PhiState() : status(Unknown), base(nullptr) {}
560 Status getStatus() const { return status; }
561 Value *getBase() const { return base; }
563 bool isBase() const { return getStatus() == Base; }
564 bool isUnknown() const { return getStatus() == Unknown; }
565 bool isConflict() const { return getStatus() == Conflict; }
567 bool operator==(const PhiState &other) const {
568 return base == other.base && status == other.status;
571 bool operator!=(const PhiState &other) const { return !(*this == other); }
574 errs() << status << " (" << base << " - "
575 << (base ? base->getName() : "nullptr") << "): ";
580 Value *base; // non null only if status == base
583 typedef DenseMap<Value *, PhiState> ConflictStateMapTy;
584 // Values of type PhiState form a lattice, and this is a helper
585 // class that implementes the meet operation. The meat of the meet
586 // operation is implemented in MeetPhiStates::pureMeet
587 class MeetPhiStates {
589 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
590 explicit MeetPhiStates(const ConflictStateMapTy &phiStates)
591 : phiStates(phiStates) {}
593 // Destructively meet the current result with the base V. V can
594 // either be a merge instruction (SelectInst / PHINode), in which
595 // case its status is looked up in the phiStates map; or a regular
596 // SSA value, in which case it is assumed to be a base.
597 void meetWith(Value *V) {
598 PhiState otherState = getStateForBDV(V);
599 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
600 MeetPhiStates::pureMeet(currentResult, otherState)) &&
601 "math is wrong: meet does not commute!");
602 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
605 PhiState getResult() const { return currentResult; }
608 const ConflictStateMapTy &phiStates;
609 PhiState currentResult;
611 /// Return a phi state for a base defining value. We'll generate a new
612 /// base state for known bases and expect to find a cached state otherwise
613 PhiState getStateForBDV(Value *baseValue) {
614 if (isKnownBaseResult(baseValue)) {
615 return PhiState(baseValue);
617 return lookupFromMap(baseValue);
621 PhiState lookupFromMap(Value *V) {
622 auto I = phiStates.find(V);
623 assert(I != phiStates.end() && "lookup failed!");
627 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
628 switch (stateA.getStatus()) {
629 case PhiState::Unknown:
633 assert(stateA.getBase() && "can't be null");
634 if (stateB.isUnknown())
637 if (stateB.isBase()) {
638 if (stateA.getBase() == stateB.getBase()) {
639 assert(stateA == stateB && "equality broken!");
642 return PhiState(PhiState::Conflict);
644 assert(stateB.isConflict() && "only three states!");
645 return PhiState(PhiState::Conflict);
647 case PhiState::Conflict:
650 llvm_unreachable("only three states!");
654 /// For a given value or instruction, figure out what base ptr it's derived
655 /// from. For gc objects, this is simply itself. On success, returns a value
656 /// which is the base pointer. (This is reliable and can be used for
657 /// relocation.) On failure, returns nullptr.
658 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache,
659 DenseSet<llvm::Value *> &NewInsertedDefs) {
660 Value *def = findBaseOrBDV(I, cache);
662 if (isKnownBaseResult(def)) {
666 // Here's the rough algorithm:
667 // - For every SSA value, construct a mapping to either an actual base
668 // pointer or a PHI which obscures the base pointer.
669 // - Construct a mapping from PHI to unknown TOP state. Use an
670 // optimistic algorithm to propagate base pointer information. Lattice
675 // When algorithm terminates, all PHIs will either have a single concrete
676 // base or be in a conflict state.
677 // - For every conflict, insert a dummy PHI node without arguments. Add
678 // these to the base[Instruction] = BasePtr mapping. For every
679 // non-conflict, add the actual base.
680 // - For every conflict, add arguments for the base[a] of each input
683 // Note: A simpler form of this would be to add the conflict form of all
684 // PHIs without running the optimistic algorithm. This would be
685 // analougous to pessimistic data flow and would likely lead to an
686 // overall worse solution.
688 ConflictStateMapTy states;
689 states[def] = PhiState();
690 // Recursively fill in all phis & selects reachable from the initial one
691 // for which we don't already know a definite base value for
692 // TODO: This should be rewritten with a worklist
696 // Since we're adding elements to 'states' as we run, we can't keep
697 // iterators into the set.
698 SmallVector<Value*, 16> Keys;
699 Keys.reserve(states.size());
700 for (auto Pair : states) {
701 Value *V = Pair.first;
704 for (Value *v : Keys) {
705 assert(!isKnownBaseResult(v) && "why did it get added?");
706 if (PHINode *phi = dyn_cast<PHINode>(v)) {
707 assert(phi->getNumIncomingValues() > 0 &&
708 "zero input phis are illegal");
709 for (Value *InVal : phi->incoming_values()) {
710 Value *local = findBaseOrBDV(InVal, cache);
711 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
712 states[local] = PhiState();
716 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
717 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
718 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
719 states[local] = PhiState();
722 local = findBaseOrBDV(sel->getFalseValue(), cache);
723 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
724 states[local] = PhiState();
732 errs() << "States after initialization:\n";
733 for (auto Pair : states) {
734 Instruction *v = cast<Instruction>(Pair.first);
735 PhiState state = Pair.second;
741 // TODO: come back and revisit the state transitions around inputs which
742 // have reached conflict state. The current version seems too conservative.
744 bool progress = true;
747 size_t oldSize = states.size();
750 // We're only changing keys in this loop, thus safe to keep iterators
751 for (auto Pair : states) {
752 MeetPhiStates calculateMeet(states);
753 Value *v = Pair.first;
754 assert(!isKnownBaseResult(v) && "why did it get added?");
755 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
756 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
757 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
759 for (Value *Val : cast<PHINode>(v)->incoming_values())
760 calculateMeet.meetWith(findBaseOrBDV(Val, cache));
762 PhiState oldState = states[v];
763 PhiState newState = calculateMeet.getResult();
764 if (oldState != newState) {
766 states[v] = newState;
770 assert(oldSize <= states.size());
771 assert(oldSize == states.size() || progress);
775 errs() << "States after meet iteration:\n";
776 for (auto Pair : states) {
777 Instruction *v = cast<Instruction>(Pair.first);
778 PhiState state = Pair.second;
784 // Insert Phis for all conflicts
785 // We want to keep naming deterministic in the loop that follows, so
786 // sort the keys before iteration. This is useful in allowing us to
787 // write stable tests. Note that there is no invalidation issue here.
788 SmallVector<Value*, 16> Keys;
789 Keys.reserve(states.size());
790 for (auto Pair : states) {
791 Value *V = Pair.first;
794 std::sort(Keys.begin(), Keys.end(), order_by_name);
795 // TODO: adjust naming patterns to avoid this order of iteration dependency
796 for (Value *V : Keys) {
797 Instruction *v = cast<Instruction>(V);
798 PhiState state = states[V];
799 assert(!isKnownBaseResult(v) && "why did it get added?");
800 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
801 if (!state.isConflict())
804 if (isa<PHINode>(v)) {
806 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
807 assert(num_preds > 0 && "how did we reach here");
808 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
809 NewInsertedDefs.insert(phi);
810 // Add metadata marking this as a base value
811 auto *const_1 = ConstantInt::get(
813 v->getParent()->getParent()->getParent()->getContext()),
815 auto MDConst = ConstantAsMetadata::get(const_1);
816 MDNode *md = MDNode::get(
817 v->getParent()->getParent()->getParent()->getContext(), MDConst);
818 phi->setMetadata("is_base_value", md);
819 states[v] = PhiState(PhiState::Conflict, phi);
821 SelectInst *sel = cast<SelectInst>(v);
822 // The undef will be replaced later
823 UndefValue *undef = UndefValue::get(sel->getType());
824 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
825 undef, "base_select", sel);
826 NewInsertedDefs.insert(basesel);
827 // Add metadata marking this as a base value
828 auto *const_1 = ConstantInt::get(
830 v->getParent()->getParent()->getParent()->getContext()),
832 auto MDConst = ConstantAsMetadata::get(const_1);
833 MDNode *md = MDNode::get(
834 v->getParent()->getParent()->getParent()->getContext(), MDConst);
835 basesel->setMetadata("is_base_value", md);
836 states[v] = PhiState(PhiState::Conflict, basesel);
840 // Fixup all the inputs of the new PHIs
841 for (auto Pair : states) {
842 Instruction *v = cast<Instruction>(Pair.first);
843 PhiState state = Pair.second;
845 assert(!isKnownBaseResult(v) && "why did it get added?");
846 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
847 if (!state.isConflict())
850 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
851 PHINode *phi = cast<PHINode>(v);
852 unsigned NumPHIValues = phi->getNumIncomingValues();
853 for (unsigned i = 0; i < NumPHIValues; i++) {
854 Value *InVal = phi->getIncomingValue(i);
855 BasicBlock *InBB = phi->getIncomingBlock(i);
857 // If we've already seen InBB, add the same incoming value
858 // we added for it earlier. The IR verifier requires phi
859 // nodes with multiple entries from the same basic block
860 // to have the same incoming value for each of those
861 // entries. If we don't do this check here and basephi
862 // has a different type than base, we'll end up adding two
863 // bitcasts (and hence two distinct values) as incoming
864 // values for the same basic block.
866 int blockIndex = basephi->getBasicBlockIndex(InBB);
867 if (blockIndex != -1) {
868 Value *oldBase = basephi->getIncomingValue(blockIndex);
869 basephi->addIncoming(oldBase, InBB);
871 Value *base = findBaseOrBDV(InVal, cache);
872 if (!isKnownBaseResult(base)) {
873 // Either conflict or base.
874 assert(states.count(base));
875 base = states[base].getBase();
876 assert(base != nullptr && "unknown PhiState!");
877 assert(NewInsertedDefs.count(base) &&
878 "should have already added this in a prev. iteration!");
881 // In essense this assert states: the only way two
882 // values incoming from the same basic block may be
883 // different is by being different bitcasts of the same
884 // value. A cleanup that remains TODO is changing
885 // findBaseOrBDV to return an llvm::Value of the correct
886 // type (and still remain pure). This will remove the
887 // need to add bitcasts.
888 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
889 "sanity -- findBaseOrBDV should be pure!");
894 // Find either the defining value for the PHI or the normal base for
896 Value *base = findBaseOrBDV(InVal, cache);
897 if (!isKnownBaseResult(base)) {
898 // Either conflict or base.
899 assert(states.count(base));
900 base = states[base].getBase();
901 assert(base != nullptr && "unknown PhiState!");
903 assert(base && "can't be null");
904 // Must use original input BB since base may not be Instruction
905 // The cast is needed since base traversal may strip away bitcasts
906 if (base->getType() != basephi->getType()) {
907 base = new BitCastInst(base, basephi->getType(), "cast",
908 InBB->getTerminator());
909 NewInsertedDefs.insert(base);
911 basephi->addIncoming(base, InBB);
913 assert(basephi->getNumIncomingValues() == NumPHIValues);
915 SelectInst *basesel = cast<SelectInst>(state.getBase());
916 SelectInst *sel = cast<SelectInst>(v);
917 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
918 // something more safe and less hacky.
919 for (int i = 1; i <= 2; i++) {
920 Value *InVal = sel->getOperand(i);
921 // Find either the defining value for the PHI or the normal base for
923 Value *base = findBaseOrBDV(InVal, cache);
924 if (!isKnownBaseResult(base)) {
925 // Either conflict or base.
926 assert(states.count(base));
927 base = states[base].getBase();
928 assert(base != nullptr && "unknown PhiState!");
930 assert(base && "can't be null");
931 // Must use original input BB since base may not be Instruction
932 // The cast is needed since base traversal may strip away bitcasts
933 if (base->getType() != basesel->getType()) {
934 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
935 NewInsertedDefs.insert(base);
937 basesel->setOperand(i, base);
942 // Cache all of our results so we can cheaply reuse them
943 // NOTE: This is actually two caches: one of the base defining value
944 // relation and one of the base pointer relation! FIXME
945 for (auto item : states) {
946 Value *v = item.first;
947 Value *base = item.second.getBase();
949 assert(!isKnownBaseResult(v) && "why did it get added?");
952 std::string fromstr =
953 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
955 errs() << "Updating base value cache"
956 << " for: " << (v->hasName() ? v->getName() : "")
957 << " from: " << fromstr
958 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
961 assert(isKnownBaseResult(base) &&
962 "must be something we 'know' is a base pointer");
963 if (cache.count(v)) {
964 // Once we transition from the BDV relation being store in the cache to
965 // the base relation being stored, it must be stable
966 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
967 "base relation should be stable");
971 assert(cache.find(def) != cache.end());
975 // For a set of live pointers (base and/or derived), identify the base
976 // pointer of the object which they are derived from. This routine will
977 // mutate the IR graph as needed to make the 'base' pointer live at the
978 // definition site of 'derived'. This ensures that any use of 'derived' can
979 // also use 'base'. This may involve the insertion of a number of
980 // additional PHI nodes.
982 // preconditions: live is a set of pointer type Values
984 // side effects: may insert PHI nodes into the existing CFG, will preserve
985 // CFG, will not remove or mutate any existing nodes
987 // post condition: PointerToBase contains one (derived, base) pair for every
988 // pointer in live. Note that derived can be equal to base if the original
989 // pointer was a base pointer.
990 static void findBasePointers(const StatepointLiveSetTy &live,
991 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
992 DominatorTree *DT, DefiningValueMapTy &DVCache,
993 DenseSet<llvm::Value *> &NewInsertedDefs) {
994 // For the naming of values inserted to be deterministic - which makes for
995 // much cleaner and more stable tests - we need to assign an order to the
996 // live values. DenseSets do not provide a deterministic order across runs.
997 SmallVector<Value*, 64> Temp;
998 Temp.insert(Temp.end(), live.begin(), live.end());
999 std::sort(Temp.begin(), Temp.end(), order_by_name);
1000 for (Value *ptr : Temp) {
1001 Value *base = findBasePointer(ptr, DVCache, NewInsertedDefs);
1002 assert(base && "failed to find base pointer");
1003 PointerToBase[ptr] = base;
1004 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1005 DT->dominates(cast<Instruction>(base)->getParent(),
1006 cast<Instruction>(ptr)->getParent())) &&
1007 "The base we found better dominate the derived pointer");
1009 // If you see this trip and like to live really dangerously, the code should
1010 // be correct, just with idioms the verifier can't handle. You can try
1011 // disabling the verifier at your own substaintial risk.
1012 assert(!isa<ConstantPointerNull>(base) &&
1013 "the relocation code needs adjustment to handle the relocation of "
1014 "a null pointer constant without causing false positives in the "
1015 "safepoint ir verifier.");
1019 /// Find the required based pointers (and adjust the live set) for the given
1021 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1023 PartiallyConstructedSafepointRecord &result) {
1024 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
1025 DenseSet<llvm::Value *> NewInsertedDefs;
1026 findBasePointers(result.liveset, PointerToBase, &DT, DVCache, NewInsertedDefs);
1028 if (PrintBasePointers) {
1029 // Note: Need to print these in a stable order since this is checked in
1031 errs() << "Base Pairs (w/o Relocation):\n";
1032 SmallVector<Value*, 64> Temp;
1033 Temp.reserve(PointerToBase.size());
1034 for (auto Pair : PointerToBase) {
1035 Temp.push_back(Pair.first);
1037 std::sort(Temp.begin(), Temp.end(), order_by_name);
1038 for (Value *Ptr : Temp) {
1039 Value *Base = PointerToBase[Ptr];
1040 errs() << " derived %" << Ptr->getName() << " base %"
1041 << Base->getName() << "\n";
1045 result.PointerToBase = PointerToBase;
1046 result.NewInsertedDefs = NewInsertedDefs;
1049 /// Check for liveness of items in the insert defs and add them to the live
1050 /// and base pointer sets
1051 static void fixupLiveness(DominatorTree &DT, const CallSite &CS,
1052 const DenseSet<Value *> &allInsertedDefs,
1053 PartiallyConstructedSafepointRecord &result) {
1054 Instruction *inst = CS.getInstruction();
1056 auto liveset = result.liveset;
1057 auto PointerToBase = result.PointerToBase;
1059 auto is_live_gc_reference =
1060 [&](Value &V) { return isLiveGCReferenceAt(V, inst, DT, nullptr); };
1062 // For each new definition, check to see if a) the definition dominates the
1063 // instruction we're interested in, and b) one of the uses of that definition
1064 // is edge-reachable from the instruction we're interested in. This is the
1065 // same definition of liveness we used in the intial liveness analysis
1066 for (Value *newDef : allInsertedDefs) {
1067 if (liveset.count(newDef)) {
1068 // already live, no action needed
1072 // PERF: Use DT to check instruction domination might not be good for
1073 // compilation time, and we could change to optimal solution if this
1074 // turn to be a issue
1075 if (!DT.dominates(cast<Instruction>(newDef), inst)) {
1076 // can't possibly be live at inst
1080 if (is_live_gc_reference(*newDef)) {
1081 // Add the live new defs into liveset and PointerToBase
1082 liveset.insert(newDef);
1083 PointerToBase[newDef] = newDef;
1087 result.liveset = liveset;
1088 result.PointerToBase = PointerToBase;
1091 static void fixupLiveReferences(
1092 Function &F, DominatorTree &DT, Pass *P,
1093 const DenseSet<llvm::Value *> &allInsertedDefs,
1094 ArrayRef<CallSite> toUpdate,
1095 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1096 for (size_t i = 0; i < records.size(); i++) {
1097 struct PartiallyConstructedSafepointRecord &info = records[i];
1098 const CallSite &CS = toUpdate[i];
1099 fixupLiveness(DT, CS, allInsertedDefs, info);
1103 // Normalize basic block to make it ready to be target of invoke statepoint.
1104 // It means spliting it to have single predecessor. Return newly created BB
1105 // ready to be successor of invoke statepoint.
1106 static BasicBlock *normalizeBBForInvokeSafepoint(BasicBlock *BB,
1107 BasicBlock *InvokeParent,
1109 BasicBlock *ret = BB;
1111 if (!BB->getUniquePredecessor()) {
1112 ret = SplitBlockPredecessors(BB, InvokeParent, "");
1115 // Another requirement for such basic blocks is to not have any phi nodes.
1116 // Since we just ensured that new BB will have single predecessor,
1117 // all phi nodes in it will have one value. Here it would be naturall place
1119 // remove them all. But we can not do this because we are risking to remove
1120 // one of the values stored in liveset of another statepoint. We will do it
1121 // later after placing all safepoints.
1126 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1127 auto itr = std::find(livevec.begin(), livevec.end(), val);
1128 assert(livevec.end() != itr);
1129 size_t index = std::distance(livevec.begin(), itr);
1130 assert(index < livevec.size());
1134 // Create new attribute set containing only attributes which can be transfered
1135 // from original call to the safepoint.
1136 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1139 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1140 unsigned index = AS.getSlotIndex(Slot);
1142 if (index == AttributeSet::ReturnIndex ||
1143 index == AttributeSet::FunctionIndex) {
1145 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1147 Attribute attr = *it;
1149 // Do not allow certain attributes - just skip them
1150 // Safepoint can not be read only or read none.
1151 if (attr.hasAttribute(Attribute::ReadNone) ||
1152 attr.hasAttribute(Attribute::ReadOnly))
1155 ret = ret.addAttributes(
1156 AS.getContext(), index,
1157 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1161 // Just skip parameter attributes for now
1167 /// Helper function to place all gc relocates necessary for the given
1170 /// liveVariables - list of variables to be relocated.
1171 /// liveStart - index of the first live variable.
1172 /// basePtrs - base pointers.
1173 /// statepointToken - statepoint instruction to which relocates should be
1175 /// Builder - Llvm IR builder to be used to construct new calls.
1176 static void CreateGCRelocates(ArrayRef<llvm::Value *> liveVariables,
1177 const int liveStart,
1178 ArrayRef<llvm::Value *> basePtrs,
1179 Instruction *statepointToken,
1180 IRBuilder<> Builder) {
1181 SmallVector<Instruction *, 64> NewDefs;
1182 NewDefs.reserve(liveVariables.size());
1184 Module *M = statepointToken->getParent()->getParent()->getParent();
1186 for (unsigned i = 0; i < liveVariables.size(); i++) {
1187 // We generate a (potentially) unique declaration for every pointer type
1188 // combination. This results is some blow up the function declarations in
1189 // the IR, but removes the need for argument bitcasts which shrinks the IR
1190 // greatly and makes it much more readable.
1191 SmallVector<Type *, 1> types; // one per 'any' type
1192 types.push_back(liveVariables[i]->getType()); // result type
1193 Value *gc_relocate_decl = Intrinsic::getDeclaration(
1194 M, Intrinsic::experimental_gc_relocate, types);
1196 // Generate the gc.relocate call and save the result
1198 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1199 liveStart + find_index(liveVariables, basePtrs[i]));
1200 Value *liveIdx = ConstantInt::get(
1201 Type::getInt32Ty(M->getContext()),
1202 liveStart + find_index(liveVariables, liveVariables[i]));
1204 // only specify a debug name if we can give a useful one
1205 Value *reloc = Builder.CreateCall3(
1206 gc_relocate_decl, statepointToken, baseIdx, liveIdx,
1207 liveVariables[i]->hasName() ? liveVariables[i]->getName() + ".relocated"
1209 // Trick CodeGen into thinking there are lots of free registers at this
1211 cast<CallInst>(reloc)->setCallingConv(CallingConv::Cold);
1213 NewDefs.push_back(cast<Instruction>(reloc));
1215 assert(NewDefs.size() == liveVariables.size() &&
1216 "missing or extra redefinition at safepoint");
1220 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1221 const SmallVectorImpl<llvm::Value *> &basePtrs,
1222 const SmallVectorImpl<llvm::Value *> &liveVariables,
1224 PartiallyConstructedSafepointRecord &result) {
1225 assert(basePtrs.size() == liveVariables.size());
1226 assert(isStatepoint(CS) &&
1227 "This method expects to be rewriting a statepoint");
1229 BasicBlock *BB = CS.getInstruction()->getParent();
1231 Function *F = BB->getParent();
1232 assert(F && "must be set");
1233 Module *M = F->getParent();
1235 assert(M && "must be set");
1237 // We're not changing the function signature of the statepoint since the gc
1238 // arguments go into the var args section.
1239 Function *gc_statepoint_decl = CS.getCalledFunction();
1241 // Then go ahead and use the builder do actually do the inserts. We insert
1242 // immediately before the previous instruction under the assumption that all
1243 // arguments will be available here. We can't insert afterwards since we may
1244 // be replacing a terminator.
1245 Instruction *insertBefore = CS.getInstruction();
1246 IRBuilder<> Builder(insertBefore);
1247 // Copy all of the arguments from the original statepoint - this includes the
1248 // target, call args, and deopt args
1249 SmallVector<llvm::Value *, 64> args;
1250 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1251 // TODO: Clear the 'needs rewrite' flag
1253 // add all the pointers to be relocated (gc arguments)
1254 // Capture the start of the live variable list for use in the gc_relocates
1255 const int live_start = args.size();
1256 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1258 // Create the statepoint given all the arguments
1259 Instruction *token = nullptr;
1260 AttributeSet return_attributes;
1262 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1264 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1265 call->setTailCall(toReplace->isTailCall());
1266 call->setCallingConv(toReplace->getCallingConv());
1268 // Currently we will fail on parameter attributes and on certain
1269 // function attributes.
1270 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1271 // In case if we can handle this set of sttributes - set up function attrs
1272 // directly on statepoint and return attrs later for gc_result intrinsic.
1273 call->setAttributes(new_attrs.getFnAttributes());
1274 return_attributes = new_attrs.getRetAttributes();
1278 // Put the following gc_result and gc_relocate calls immediately after the
1279 // the old call (which we're about to delete)
1280 BasicBlock::iterator next(toReplace);
1281 assert(BB->end() != next && "not a terminator, must have next");
1283 Instruction *IP = &*(next);
1284 Builder.SetInsertPoint(IP);
1285 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1288 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1290 // Insert the new invoke into the old block. We'll remove the old one in a
1291 // moment at which point this will become the new terminator for the
1293 InvokeInst *invoke = InvokeInst::Create(
1294 gc_statepoint_decl, toReplace->getNormalDest(),
1295 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1296 invoke->setCallingConv(toReplace->getCallingConv());
1298 // Currently we will fail on parameter attributes and on certain
1299 // function attributes.
1300 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1301 // In case if we can handle this set of sttributes - set up function attrs
1302 // directly on statepoint and return attrs later for gc_result intrinsic.
1303 invoke->setAttributes(new_attrs.getFnAttributes());
1304 return_attributes = new_attrs.getRetAttributes();
1308 // Generate gc relocates in exceptional path
1309 BasicBlock *unwindBlock = normalizeBBForInvokeSafepoint(
1310 toReplace->getUnwindDest(), invoke->getParent(), P);
1312 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1313 Builder.SetInsertPoint(IP);
1314 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1316 // Extract second element from landingpad return value. We will attach
1317 // exceptional gc relocates to it.
1318 const unsigned idx = 1;
1319 Instruction *exceptional_token =
1320 cast<Instruction>(Builder.CreateExtractValue(
1321 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1322 result.UnwindToken = exceptional_token;
1324 // Just throw away return value. We will use the one we got for normal
1326 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1327 exceptional_token, Builder);
1329 // Generate gc relocates and returns for normal block
1330 BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
1331 toReplace->getNormalDest(), invoke->getParent(), P);
1333 IP = &*(normalDest->getFirstInsertionPt());
1334 Builder.SetInsertPoint(IP);
1336 // gc relocates will be generated later as if it were regular call
1341 // Take the name of the original value call if it had one.
1342 token->takeName(CS.getInstruction());
1344 // The GCResult is already inserted, we just need to find it
1346 Instruction *toReplace = CS.getInstruction();
1347 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1348 "only valid use before rewrite is gc.result");
1349 assert(!toReplace->hasOneUse() ||
1350 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1353 // Update the gc.result of the original statepoint (if any) to use the newly
1354 // inserted statepoint. This is safe to do here since the token can't be
1355 // considered a live reference.
1356 CS.getInstruction()->replaceAllUsesWith(token);
1358 result.StatepointToken = token;
1360 // Second, create a gc.relocate for every live variable
1361 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1366 struct name_ordering {
1369 bool operator()(name_ordering const &a, name_ordering const &b) {
1370 return -1 == a.derived->getName().compare(b.derived->getName());
1374 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1375 SmallVectorImpl<Value *> &livevec) {
1376 assert(basevec.size() == livevec.size());
1378 SmallVector<name_ordering, 64> temp;
1379 for (size_t i = 0; i < basevec.size(); i++) {
1381 v.base = basevec[i];
1382 v.derived = livevec[i];
1385 std::sort(temp.begin(), temp.end(), name_ordering());
1386 for (size_t i = 0; i < basevec.size(); i++) {
1387 basevec[i] = temp[i].base;
1388 livevec[i] = temp[i].derived;
1392 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1393 // which make the relocations happening at this safepoint explicit.
1395 // WARNING: Does not do any fixup to adjust users of the original live
1396 // values. That's the callers responsibility.
1398 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1399 PartiallyConstructedSafepointRecord &result) {
1400 auto liveset = result.liveset;
1401 auto PointerToBase = result.PointerToBase;
1403 // Convert to vector for efficient cross referencing.
1404 SmallVector<Value *, 64> basevec, livevec;
1405 livevec.reserve(liveset.size());
1406 basevec.reserve(liveset.size());
1407 for (Value *L : liveset) {
1408 livevec.push_back(L);
1410 assert(PointerToBase.find(L) != PointerToBase.end());
1411 Value *base = PointerToBase[L];
1412 basevec.push_back(base);
1414 assert(livevec.size() == basevec.size());
1416 // To make the output IR slightly more stable (for use in diffs), ensure a
1417 // fixed order of the values in the safepoint (by sorting the value name).
1418 // The order is otherwise meaningless.
1419 stablize_order(basevec, livevec);
1421 // Do the actual rewriting and delete the old statepoint
1422 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1423 CS.getInstruction()->eraseFromParent();
1426 // Helper function for the relocationViaAlloca.
1427 // It receives iterator to the statepoint gc relocates and emits store to the
1429 // location (via allocaMap) for the each one of them.
1430 // Add visited values into the visitedLiveValues set we will later use them
1431 // for sanity check.
1433 insertRelocationStores(iterator_range<Value::user_iterator> gcRelocs,
1434 DenseMap<Value *, Value *> &allocaMap,
1435 DenseSet<Value *> &visitedLiveValues) {
1437 for (User *U : gcRelocs) {
1438 if (!isa<IntrinsicInst>(U))
1441 IntrinsicInst *relocatedValue = cast<IntrinsicInst>(U);
1443 // We only care about relocates
1444 if (relocatedValue->getIntrinsicID() !=
1445 Intrinsic::experimental_gc_relocate) {
1449 GCRelocateOperands relocateOperands(relocatedValue);
1450 Value *originalValue = const_cast<Value *>(relocateOperands.derivedPtr());
1451 assert(allocaMap.count(originalValue));
1452 Value *alloca = allocaMap[originalValue];
1454 // Emit store into the related alloca
1455 StoreInst *store = new StoreInst(relocatedValue, alloca);
1456 store->insertAfter(relocatedValue);
1459 visitedLiveValues.insert(originalValue);
1464 /// do all the relocation update via allocas and mem2reg
1465 static void relocationViaAlloca(
1466 Function &F, DominatorTree &DT, ArrayRef<Value *> live,
1467 ArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1469 // record initial number of (static) allocas; we'll check we have the same
1470 // number when we get done.
1471 int InitialAllocaNum = 0;
1472 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end();
1474 if (isa<AllocaInst>(*I))
1478 // TODO-PERF: change data structures, reserve
1479 DenseMap<Value *, Value *> allocaMap;
1480 SmallVector<AllocaInst *, 200> PromotableAllocas;
1481 PromotableAllocas.reserve(live.size());
1483 // emit alloca for each live gc pointer
1484 for (unsigned i = 0; i < live.size(); i++) {
1485 Value *liveValue = live[i];
1486 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1487 F.getEntryBlock().getFirstNonPHI());
1488 allocaMap[liveValue] = alloca;
1489 PromotableAllocas.push_back(alloca);
1492 // The next two loops are part of the same conceptual operation. We need to
1493 // insert a store to the alloca after the original def and at each
1494 // redefinition. We need to insert a load before each use. These are split
1495 // into distinct loops for performance reasons.
1497 // update gc pointer after each statepoint
1498 // either store a relocated value or null (if no relocated value found for
1499 // this gc pointer and it is not a gc_result)
1500 // this must happen before we update the statepoint with load of alloca
1501 // otherwise we lose the link between statepoint and old def
1502 for (size_t i = 0; i < records.size(); i++) {
1503 const struct PartiallyConstructedSafepointRecord &info = records[i];
1504 Value *Statepoint = info.StatepointToken;
1506 // This will be used for consistency check
1507 DenseSet<Value *> visitedLiveValues;
1509 // Insert stores for normal statepoint gc relocates
1510 insertRelocationStores(Statepoint->users(), allocaMap, visitedLiveValues);
1512 // In case if it was invoke statepoint
1513 // we will insert stores for exceptional path gc relocates.
1514 if (isa<InvokeInst>(Statepoint)) {
1515 insertRelocationStores(info.UnwindToken->users(),
1516 allocaMap, visitedLiveValues);
1520 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1521 // the gc.statepoint. This will turn some subtle GC problems into slightly
1522 // easier to debug SEGVs
1523 SmallVector<AllocaInst *, 64> ToClobber;
1524 for (auto Pair : allocaMap) {
1525 Value *Def = Pair.first;
1526 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1528 // This value was relocated
1529 if (visitedLiveValues.count(Def)) {
1532 ToClobber.push_back(Alloca);
1535 auto InsertClobbersAt = [&](Instruction *IP) {
1536 for (auto *AI : ToClobber) {
1537 auto AIType = cast<PointerType>(AI->getType());
1538 auto PT = cast<PointerType>(AIType->getElementType());
1539 Constant *CPN = ConstantPointerNull::get(PT);
1540 StoreInst *store = new StoreInst(CPN, AI);
1541 store->insertBefore(IP);
1545 // Insert the clobbering stores. These may get intermixed with the
1546 // gc.results and gc.relocates, but that's fine.
1547 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1548 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1549 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1551 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1553 InsertClobbersAt(Next);
1557 // update use with load allocas and add store for gc_relocated
1558 for (auto Pair : allocaMap) {
1559 Value *def = Pair.first;
1560 Value *alloca = Pair.second;
1562 // we pre-record the uses of allocas so that we dont have to worry about
1564 // that change the user information.
1565 SmallVector<Instruction *, 20> uses;
1566 // PERF: trade a linear scan for repeated reallocation
1567 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1568 for (User *U : def->users()) {
1569 if (!isa<ConstantExpr>(U)) {
1570 // If the def has a ConstantExpr use, then the def is either a
1571 // ConstantExpr use itself or null. In either case
1572 // (recursively in the first, directly in the second), the oop
1573 // it is ultimately dependent on is null and this particular
1574 // use does not need to be fixed up.
1575 uses.push_back(cast<Instruction>(U));
1579 std::sort(uses.begin(), uses.end());
1580 auto last = std::unique(uses.begin(), uses.end());
1581 uses.erase(last, uses.end());
1583 for (Instruction *use : uses) {
1584 if (isa<PHINode>(use)) {
1585 PHINode *phi = cast<PHINode>(use);
1586 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1587 if (def == phi->getIncomingValue(i)) {
1588 LoadInst *load = new LoadInst(
1589 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1590 phi->setIncomingValue(i, load);
1594 LoadInst *load = new LoadInst(alloca, "", use);
1595 use->replaceUsesOfWith(def, load);
1599 // emit store for the initial gc value
1600 // store must be inserted after load, otherwise store will be in alloca's
1601 // use list and an extra load will be inserted before it
1602 StoreInst *store = new StoreInst(def, alloca);
1603 if (Instruction *inst = dyn_cast<Instruction>(def)) {
1604 if (InvokeInst *invoke = dyn_cast<InvokeInst>(inst)) {
1605 // InvokeInst is a TerminatorInst so the store need to be inserted
1606 // into its normal destination block.
1607 BasicBlock *normalDest = invoke->getNormalDest();
1608 store->insertBefore(normalDest->getFirstNonPHI());
1610 assert(!inst->isTerminator() &&
1611 "The only TerminatorInst that can produce a value is "
1612 "InvokeInst which is handled above.");
1613 store->insertAfter(inst);
1616 assert((isa<Argument>(def) || isa<GlobalVariable>(def) ||
1617 isa<ConstantPointerNull>(def)) &&
1618 "Must be argument or global");
1619 store->insertAfter(cast<Instruction>(alloca));
1623 assert(PromotableAllocas.size() == live.size() &&
1624 "we must have the same allocas with lives");
1625 if (!PromotableAllocas.empty()) {
1626 // apply mem2reg to promote alloca to SSA
1627 PromoteMemToReg(PromotableAllocas, DT);
1631 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end();
1633 if (isa<AllocaInst>(*I))
1635 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1639 /// Implement a unique function which doesn't require we sort the input
1640 /// vector. Doing so has the effect of changing the output of a couple of
1641 /// tests in ways which make them less useful in testing fused safepoints.
1642 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1644 SmallVector<T, 128> TempVec;
1645 TempVec.reserve(Vec.size());
1646 for (auto Element : Vec)
1647 TempVec.push_back(Element);
1649 for (auto V : TempVec) {
1650 if (Seen.insert(V).second) {
1656 static Function *getUseHolder(Module &M) {
1657 FunctionType *ftype =
1658 FunctionType::get(Type::getVoidTy(M.getContext()), true);
1659 Function *Func = cast<Function>(M.getOrInsertFunction("__tmp_use", ftype));
1663 /// Insert holders so that each Value is obviously live through the entire
1664 /// liftetime of the call.
1665 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1666 SmallVectorImpl<CallInst *> &holders) {
1667 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1668 Function *Func = getUseHolder(*M);
1670 // For call safepoints insert dummy calls right after safepoint
1671 BasicBlock::iterator next(CS.getInstruction());
1673 CallInst *base_holder = CallInst::Create(Func, Values, "", next);
1674 holders.push_back(base_holder);
1675 } else if (CS.isInvoke()) {
1676 // For invoke safepooints insert dummy calls both in normal and
1677 // exceptional destination blocks
1678 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1679 CallInst *normal_holder = CallInst::Create(
1680 Func, Values, "", invoke->getNormalDest()->getFirstInsertionPt());
1681 CallInst *unwind_holder = CallInst::Create(
1682 Func, Values, "", invoke->getUnwindDest()->getFirstInsertionPt());
1683 holders.push_back(normal_holder);
1684 holders.push_back(unwind_holder);
1686 llvm_unreachable("unsupported call type");
1689 static void findLiveReferences(
1690 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1691 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1692 for (size_t i = 0; i < records.size(); i++) {
1693 struct PartiallyConstructedSafepointRecord &info = records[i];
1694 const CallSite &CS = toUpdate[i];
1695 analyzeParsePointLiveness(DT, CS, info);
1699 static void addBasesAsLiveValues(StatepointLiveSetTy &liveset,
1700 DenseMap<Value *, Value *> &PointerToBase) {
1701 // Identify any base pointers which are used in this safepoint, but not
1702 // themselves relocated. We need to relocate them so that later inserted
1703 // safepoints can get the properly relocated base register.
1704 DenseSet<Value *> missing;
1705 for (Value *L : liveset) {
1706 assert(PointerToBase.find(L) != PointerToBase.end());
1707 Value *base = PointerToBase[L];
1709 if (liveset.find(base) == liveset.end()) {
1710 assert(PointerToBase.find(base) == PointerToBase.end());
1711 // uniqued by set insert
1712 missing.insert(base);
1716 // Note that we want these at the end of the list, otherwise
1717 // register placement gets screwed up once we lower to STATEPOINT
1718 // instructions. This is an utter hack, but there doesn't seem to be a
1720 for (Value *base : missing) {
1722 liveset.insert(base);
1723 PointerToBase[base] = base;
1725 assert(liveset.size() == PointerToBase.size());
1728 /// Remove any vector of pointers from the liveset by scalarizing them over the
1729 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1730 /// would be preferrable to include the vector in the statepoint itself, but
1731 /// the lowering code currently does not handle that. Extending it would be
1732 /// slightly non-trivial since it requires a format change. Given how rare
1733 /// such cases are (for the moment?) scalarizing is an acceptable comprimise.
1734 static void splitVectorValues(Instruction *StatepointInst,
1735 StatepointLiveSetTy& LiveSet, DominatorTree &DT) {
1736 SmallVector<Value *, 16> ToSplit;
1737 for (Value *V : LiveSet)
1738 if (isa<VectorType>(V->getType()))
1739 ToSplit.push_back(V);
1741 if (ToSplit.empty())
1744 Function &F = *(StatepointInst->getParent()->getParent());
1746 DenseMap<Value*, AllocaInst*> AllocaMap;
1747 // First is normal return, second is exceptional return (invoke only)
1748 DenseMap<Value*, std::pair<Value*,Value*>> Replacements;
1749 for (Value *V : ToSplit) {
1752 AllocaInst *Alloca = new AllocaInst(V->getType(), "",
1753 F.getEntryBlock().getFirstNonPHI());
1754 AllocaMap[V] = Alloca;
1756 VectorType *VT = cast<VectorType>(V->getType());
1757 IRBuilder<> Builder(StatepointInst);
1758 SmallVector<Value*, 16> Elements;
1759 for (unsigned i = 0; i < VT->getNumElements(); i++)
1760 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1761 LiveSet.insert(Elements.begin(), Elements.end());
1763 auto InsertVectorReform = [&](Instruction *IP) {
1764 Builder.SetInsertPoint(IP);
1765 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1766 Value *ResultVec = UndefValue::get(VT);
1767 for (unsigned i = 0; i < VT->getNumElements(); i++)
1768 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1769 Builder.getInt32(i));
1773 if (isa<CallInst>(StatepointInst)) {
1774 BasicBlock::iterator Next(StatepointInst);
1776 Instruction *IP = &*(Next);
1777 Replacements[V].first = InsertVectorReform(IP);
1778 Replacements[V].second = nullptr;
1780 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1781 // We've already normalized - check that we don't have shared destination
1783 BasicBlock *NormalDest = Invoke->getNormalDest();
1784 assert(!isa<PHINode>(NormalDest->begin()));
1785 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1786 assert(!isa<PHINode>(UnwindDest->begin()));
1787 // Insert insert element sequences in both successors
1788 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1789 Replacements[V].first = InsertVectorReform(IP);
1790 IP = &*(UnwindDest->getFirstInsertionPt());
1791 Replacements[V].second = InsertVectorReform(IP);
1794 for (Value *V : ToSplit) {
1795 AllocaInst *Alloca = AllocaMap[V];
1797 // Capture all users before we start mutating use lists
1798 SmallVector<Instruction*, 16> Users;
1799 for (User *U : V->users())
1800 Users.push_back(cast<Instruction>(U));
1802 for (Instruction *I : Users) {
1803 if (auto Phi = dyn_cast<PHINode>(I)) {
1804 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1805 if (V == Phi->getIncomingValue(i)) {
1806 LoadInst *Load = new LoadInst(Alloca, "",
1807 Phi->getIncomingBlock(i)->getTerminator());
1808 Phi->setIncomingValue(i, Load);
1811 LoadInst *Load = new LoadInst(Alloca, "", I);
1812 I->replaceUsesOfWith(V, Load);
1816 // Store the original value and the replacement value into the alloca
1817 StoreInst *Store = new StoreInst(V, Alloca);
1818 if (auto I = dyn_cast<Instruction>(V))
1819 Store->insertAfter(I);
1821 Store->insertAfter(Alloca);
1823 // Normal return for invoke, or call return
1824 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1825 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1826 // Unwind return for invoke only
1827 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1829 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1832 // apply mem2reg to promote alloca to SSA
1833 SmallVector<AllocaInst*, 16> Allocas;
1834 for (Value *V : ToSplit)
1835 Allocas.push_back(AllocaMap[V]);
1836 PromoteMemToReg(Allocas, DT);
1839 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1840 SmallVectorImpl<CallSite> &toUpdate) {
1842 // sanity check the input
1843 std::set<CallSite> uniqued;
1844 uniqued.insert(toUpdate.begin(), toUpdate.end());
1845 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1847 for (size_t i = 0; i < toUpdate.size(); i++) {
1848 CallSite &CS = toUpdate[i];
1849 assert(CS.getInstruction()->getParent()->getParent() == &F);
1850 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1854 // A list of dummy calls added to the IR to keep various values obviously
1855 // live in the IR. We'll remove all of these when done.
1856 SmallVector<CallInst *, 64> holders;
1858 // Insert a dummy call with all of the arguments to the vm_state we'll need
1859 // for the actual safepoint insertion. This ensures reference arguments in
1860 // the deopt argument list are considered live through the safepoint (and
1861 // thus makes sure they get relocated.)
1862 for (size_t i = 0; i < toUpdate.size(); i++) {
1863 CallSite &CS = toUpdate[i];
1864 Statepoint StatepointCS(CS);
1866 SmallVector<Value *, 64> DeoptValues;
1867 for (Use &U : StatepointCS.vm_state_args()) {
1868 Value *Arg = cast<Value>(&U);
1869 assert(!isUnhandledGCPointerType(Arg->getType()) &&
1870 "support for FCA unimplemented");
1871 if (isHandledGCPointerType(Arg->getType()))
1872 DeoptValues.push_back(Arg);
1874 insertUseHolderAfter(CS, DeoptValues, holders);
1877 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
1878 records.reserve(toUpdate.size());
1879 for (size_t i = 0; i < toUpdate.size(); i++) {
1880 struct PartiallyConstructedSafepointRecord info;
1881 records.push_back(info);
1883 assert(records.size() == toUpdate.size());
1885 // A) Identify all gc pointers which are staticly live at the given call
1887 findLiveReferences(F, DT, P, toUpdate, records);
1889 // Do a limited scalarization of any live at safepoint vector values which
1890 // contain pointers. This enables this pass to run after vectorization at
1891 // the cost of some possible performance loss. TODO: it would be nice to
1892 // natively support vectors all the way through the backend so we don't need
1893 // to scalarize here.
1894 for (size_t i = 0; i < records.size(); i++) {
1895 struct PartiallyConstructedSafepointRecord &info = records[i];
1896 Instruction *statepoint = toUpdate[i].getInstruction();
1897 splitVectorValues(cast<Instruction>(statepoint), info.liveset, DT);
1900 // B) Find the base pointers for each live pointer
1901 /* scope for caching */ {
1902 // Cache the 'defining value' relation used in the computation and
1903 // insertion of base phis and selects. This ensures that we don't insert
1904 // large numbers of duplicate base_phis.
1905 DefiningValueMapTy DVCache;
1907 for (size_t i = 0; i < records.size(); i++) {
1908 struct PartiallyConstructedSafepointRecord &info = records[i];
1909 CallSite &CS = toUpdate[i];
1910 findBasePointers(DT, DVCache, CS, info);
1912 } // end of cache scope
1914 // The base phi insertion logic (for any safepoint) may have inserted new
1915 // instructions which are now live at some safepoint. The simplest such
1918 // phi a <-- will be a new base_phi here
1919 // safepoint 1 <-- that needs to be live here
1923 DenseSet<llvm::Value *> allInsertedDefs;
1924 for (size_t i = 0; i < records.size(); i++) {
1925 struct PartiallyConstructedSafepointRecord &info = records[i];
1926 allInsertedDefs.insert(info.NewInsertedDefs.begin(),
1927 info.NewInsertedDefs.end());
1930 // We insert some dummy calls after each safepoint to definitely hold live
1931 // the base pointers which were identified for that safepoint. We'll then
1932 // ask liveness for _every_ base inserted to see what is now live. Then we
1933 // remove the dummy calls.
1934 holders.reserve(holders.size() + records.size());
1935 for (size_t i = 0; i < records.size(); i++) {
1936 struct PartiallyConstructedSafepointRecord &info = records[i];
1937 CallSite &CS = toUpdate[i];
1939 SmallVector<Value *, 128> Bases;
1940 for (auto Pair : info.PointerToBase) {
1941 Bases.push_back(Pair.second);
1943 insertUseHolderAfter(CS, Bases, holders);
1946 // Add the bases explicitly to the live vector set. This may result in a few
1947 // extra relocations, but the base has to be available whenever a pointer
1948 // derived from it is used. Thus, we need it to be part of the statepoint's
1949 // gc arguments list. TODO: Introduce an explicit notion (in the following
1950 // code) of the GC argument list as seperate from the live Values at a
1951 // given statepoint.
1952 for (size_t i = 0; i < records.size(); i++) {
1953 struct PartiallyConstructedSafepointRecord &info = records[i];
1954 addBasesAsLiveValues(info.liveset, info.PointerToBase);
1957 // If we inserted any new values, we need to adjust our notion of what is
1958 // live at a particular safepoint.
1959 if (!allInsertedDefs.empty()) {
1960 fixupLiveReferences(F, DT, P, allInsertedDefs, toUpdate, records);
1962 if (PrintBasePointers) {
1963 for (size_t i = 0; i < records.size(); i++) {
1964 struct PartiallyConstructedSafepointRecord &info = records[i];
1965 errs() << "Base Pairs: (w/Relocation)\n";
1966 for (auto Pair : info.PointerToBase) {
1967 errs() << " derived %" << Pair.first->getName() << " base %"
1968 << Pair.second->getName() << "\n";
1972 for (size_t i = 0; i < holders.size(); i++) {
1973 holders[i]->eraseFromParent();
1974 holders[i] = nullptr;
1978 // Now run through and replace the existing statepoints with new ones with
1979 // the live variables listed. We do not yet update uses of the values being
1980 // relocated. We have references to live variables that need to
1981 // survive to the last iteration of this loop. (By construction, the
1982 // previous statepoint can not be a live variable, thus we can and remove
1983 // the old statepoint calls as we go.)
1984 for (size_t i = 0; i < records.size(); i++) {
1985 struct PartiallyConstructedSafepointRecord &info = records[i];
1986 CallSite &CS = toUpdate[i];
1987 makeStatepointExplicit(DT, CS, P, info);
1989 toUpdate.clear(); // prevent accident use of invalid CallSites
1991 // In case if we inserted relocates in a different basic block than the
1992 // original safepoint (this can happen for invokes). We need to be sure that
1993 // original values were not used in any of the phi nodes at the
1994 // beginning of basic block containing them. Because we know that all such
1995 // blocks will have single predecessor we can safely assume that all phi
1996 // nodes have single entry (because of normalizeBBForInvokeSafepoint).
1997 // Just remove them all here.
1998 for (size_t i = 0; i < records.size(); i++) {
1999 Instruction *I = records[i].StatepointToken;
2001 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
2002 FoldSingleEntryPHINodes(invoke->getNormalDest());
2003 assert(!isa<PHINode>(invoke->getNormalDest()->begin()));
2005 FoldSingleEntryPHINodes(invoke->getUnwindDest());
2006 assert(!isa<PHINode>(invoke->getUnwindDest()->begin()));
2010 // Do all the fixups of the original live variables to their relocated selves
2011 SmallVector<Value *, 128> live;
2012 for (size_t i = 0; i < records.size(); i++) {
2013 struct PartiallyConstructedSafepointRecord &info = records[i];
2014 // We can't simply save the live set from the original insertion. One of
2015 // the live values might be the result of a call which needs a safepoint.
2016 // That Value* no longer exists and we need to use the new gc_result.
2017 // Thankfully, the liveset is embedded in the statepoint (and updated), so
2018 // we just grab that.
2019 Statepoint statepoint(info.StatepointToken);
2020 live.insert(live.end(), statepoint.gc_args_begin(),
2021 statepoint.gc_args_end());
2023 unique_unsorted(live);
2027 for (auto ptr : live) {
2028 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
2032 relocationViaAlloca(F, DT, live, records);
2033 return !records.empty();
2036 /// Returns true if this function should be rewritten by this pass. The main
2037 /// point of this function is as an extension point for custom logic.
2038 static bool shouldRewriteStatepointsIn(Function &F) {
2039 // TODO: This should check the GCStrategy
2041 const std::string StatepointExampleName("statepoint-example");
2042 return StatepointExampleName == F.getGC();
2047 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2048 // Nothing to do for declarations.
2049 if (F.isDeclaration() || F.empty())
2052 // Policy choice says not to rewrite - the most common reason is that we're
2053 // compiling code without a GCStrategy.
2054 if (!shouldRewriteStatepointsIn(F))
2057 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2059 // Gather all the statepoints which need rewritten. Be careful to only
2060 // consider those in reachable code since we need to ask dominance queries
2061 // when rewriting. We'll delete the unreachable ones in a moment.
2062 SmallVector<CallSite, 64> ParsePointNeeded;
2063 SmallVector<CallSite, 16> UnreachableStatepoints;
2064 for (Instruction &I : inst_range(F)) {
2065 // TODO: only the ones with the flag set!
2066 if (isStatepoint(I)) {
2067 if (DT.isReachableFromEntry(I.getParent()))
2068 ParsePointNeeded.push_back(CallSite(&I));
2070 UnreachableStatepoints.push_back(CallSite(&I));
2074 bool MadeChange = false;
2076 // Delete any unreachable statepoints so that we don't have unrewritten
2077 // statepoints surviving this pass. This makes testing easier and the
2078 // resulting IR less confusing to human readers. Rather than be fancy, we
2079 // just reuse a utility function which removes the unreachable blocks.
2080 if (!UnreachableStatepoints.empty())
2081 MadeChange |= removeUnreachableBlocks(F);
2083 // Return early if no work to do.
2084 if (ParsePointNeeded.empty())
2087 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2088 // These are created by LCSSA. They have the effect of increasing the size
2089 // of liveness sets for no good reason. It may be harder to do this post
2090 // insertion since relocations and base phis can confuse things.
2091 for (BasicBlock &BB : F)
2092 if (BB.getUniquePredecessor()) {
2094 FoldSingleEntryPHINodes(&BB);
2097 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);