1 //===- RewriteStatepointsForGC.cpp - Make GC relocations explicit ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Rewrite an existing set of gc.statepoints such that they make potential
11 // relocations performed by the garbage collector explicit in the IR.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Pass.h"
16 #include "llvm/Analysis/CFG.h"
17 #include "llvm/ADT/SetOperations.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/DenseSet.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/IR/BasicBlock.h"
22 #include "llvm/IR/CallSite.h"
23 #include "llvm/IR/Dominators.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/InstIterator.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/Module.h"
31 #include "llvm/IR/Statepoint.h"
32 #include "llvm/IR/Value.h"
33 #include "llvm/IR/Verifier.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Transforms/Scalar.h"
37 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
38 #include "llvm/Transforms/Utils/Cloning.h"
39 #include "llvm/Transforms/Utils/Local.h"
40 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
42 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
46 // Print tracing output
47 static cl::opt<bool> TraceLSP("trace-rewrite-statepoints", cl::Hidden,
50 // Print the liveset found at the insert location
51 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
53 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden,
55 // Print out the base pointers for debugging
56 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden,
60 static bool ClobberNonLive = true;
62 static bool ClobberNonLive = false;
64 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
65 cl::location(ClobberNonLive),
69 struct RewriteStatepointsForGC : public FunctionPass {
70 static char ID; // Pass identification, replacement for typeid
72 RewriteStatepointsForGC() : FunctionPass(ID) {
73 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
75 bool runOnFunction(Function &F) override;
77 void getAnalysisUsage(AnalysisUsage &AU) const override {
78 // We add and rewrite a bunch of instructions, but don't really do much
79 // else. We could in theory preserve a lot more analyses here.
80 AU.addRequired<DominatorTreeWrapperPass>();
85 char RewriteStatepointsForGC::ID = 0;
87 FunctionPass *llvm::createRewriteStatepointsForGCPass() {
88 return new RewriteStatepointsForGC();
91 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
92 "Make relocations explicit at statepoints", false, false)
93 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
94 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
95 "Make relocations explicit at statepoints", false, false)
98 struct GCPtrLivenessData {
99 /// Values defined in this block.
100 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet;
101 /// Values used in this block (and thus live); does not included values
102 /// killed within this block.
103 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet;
105 /// Values live into this basic block (i.e. used by any
106 /// instruction in this basic block or ones reachable from here)
107 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn;
109 /// Values live out of this basic block (i.e. live into
110 /// any successor block)
111 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut;
114 // The type of the internal cache used inside the findBasePointers family
115 // of functions. From the callers perspective, this is an opaque type and
116 // should not be inspected.
118 // In the actual implementation this caches two relations:
119 // - The base relation itself (i.e. this pointer is based on that one)
120 // - The base defining value relation (i.e. before base_phi insertion)
121 // Generally, after the execution of a full findBasePointer call, only the
122 // base relation will remain. Internally, we add a mixture of the two
123 // types, then update all the second type to the first type
124 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
125 typedef DenseSet<llvm::Value *> StatepointLiveSetTy;
127 struct PartiallyConstructedSafepointRecord {
128 /// The set of values known to be live accross this safepoint
129 StatepointLiveSetTy liveset;
131 /// Mapping from live pointers to a base-defining-value
132 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
134 /// The *new* gc.statepoint instruction itself. This produces the token
135 /// that normal path gc.relocates and the gc.result are tied to.
136 Instruction *StatepointToken;
138 /// Instruction to which exceptional gc relocates are attached
139 /// Makes it easier to iterate through them during relocationViaAlloca.
140 Instruction *UnwindToken;
144 /// Compute the live-in set for every basic block in the function
145 static void computeLiveInValues(DominatorTree &DT, Function &F,
146 GCPtrLivenessData &Data);
148 /// Given results from the dataflow liveness computation, find the set of live
149 /// Values at a particular instruction.
150 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
151 StatepointLiveSetTy &out);
153 // TODO: Once we can get to the GCStrategy, this becomes
154 // Optional<bool> isGCManagedPointer(const Value *V) const override {
156 static bool isGCPointerType(const Type *T) {
157 if (const PointerType *PT = dyn_cast<PointerType>(T))
158 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
159 // GC managed heap. We know that a pointer into this heap needs to be
160 // updated and that no other pointer does.
161 return (1 == PT->getAddressSpace());
165 // Return true if this type is one which a) is a gc pointer or contains a GC
166 // pointer and b) is of a type this code expects to encounter as a live value.
167 // (The insertion code will assert that a type which matches (a) and not (b)
168 // is not encountered.)
169 static bool isHandledGCPointerType(Type *T) {
170 // We fully support gc pointers
171 if (isGCPointerType(T))
173 // We partially support vectors of gc pointers. The code will assert if it
174 // can't handle something.
175 if (auto VT = dyn_cast<VectorType>(T))
176 if (isGCPointerType(VT->getElementType()))
182 /// Returns true if this type contains a gc pointer whether we know how to
183 /// handle that type or not.
184 static bool containsGCPtrType(Type *Ty) {
185 if (isGCPointerType(Ty))
187 if (VectorType *VT = dyn_cast<VectorType>(Ty))
188 return isGCPointerType(VT->getScalarType());
189 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
190 return containsGCPtrType(AT->getElementType());
191 if (StructType *ST = dyn_cast<StructType>(Ty))
193 ST->subtypes().begin(), ST->subtypes().end(),
194 [](Type *SubType) { return containsGCPtrType(SubType); });
198 // Returns true if this is a type which a) is a gc pointer or contains a GC
199 // pointer and b) is of a type which the code doesn't expect (i.e. first class
200 // aggregates). Used to trip assertions.
201 static bool isUnhandledGCPointerType(Type *Ty) {
202 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
206 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
207 if (a->hasName() && b->hasName()) {
208 return -1 == a->getName().compare(b->getName());
209 } else if (a->hasName() && !b->hasName()) {
211 } else if (!a->hasName() && b->hasName()) {
214 // Better than nothing, but not stable
219 // Conservatively identifies any definitions which might be live at the
220 // given instruction. The analysis is performed immediately before the
221 // given instruction. Values defined by that instruction are not considered
222 // live. Values used by that instruction are considered live.
223 static void analyzeParsePointLiveness(
224 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
225 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
226 Instruction *inst = CS.getInstruction();
228 StatepointLiveSetTy liveset;
229 findLiveSetAtInst(inst, OriginalLivenessData, liveset);
232 // Note: This output is used by several of the test cases
233 // The order of elemtns in a set is not stable, put them in a vec and sort
235 SmallVector<Value *, 64> temp;
236 temp.insert(temp.end(), liveset.begin(), liveset.end());
237 std::sort(temp.begin(), temp.end(), order_by_name);
238 errs() << "Live Variables:\n";
239 for (Value *V : temp) {
240 errs() << " " << V->getName(); // no newline
244 if (PrintLiveSetSize) {
245 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
246 errs() << "Number live values: " << liveset.size() << "\n";
248 result.liveset = liveset;
251 /// If we can trivially determine that this vector contains only base pointers,
252 /// return the base instruction.
253 static Value *findBaseOfVector(Value *I) {
254 assert(I->getType()->isVectorTy() &&
255 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
256 "Illegal to ask for the base pointer of a non-pointer type");
258 // Each case parallels findBaseDefiningValue below, see that code for
259 // detailed motivation.
261 if (isa<Argument>(I))
262 // An incoming argument to the function is a base pointer
265 // We shouldn't see the address of a global as a vector value?
266 assert(!isa<GlobalVariable>(I) &&
267 "unexpected global variable found in base of vector");
269 // inlining could possibly introduce phi node that contains
270 // undef if callee has multiple returns
271 if (isa<UndefValue>(I))
272 // utterly meaningless, but useful for dealing with partially optimized
276 // Due to inheritance, this must be _after_ the global variable and undef
278 if (Constant *Con = dyn_cast<Constant>(I)) {
279 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
280 "order of checks wrong!");
281 assert(Con->isNullValue() && "null is the only case which makes sense");
285 if (isa<LoadInst>(I))
288 // Note: This code is currently rather incomplete. We are essentially only
289 // handling cases where the vector element is trivially a base pointer. We
290 // need to update the entire base pointer construction algorithm to know how
291 // to track vector elements and potentially scalarize, but the case which
292 // would motivate the work hasn't shown up in real workloads yet.
293 llvm_unreachable("no base found for vector element");
296 /// Helper function for findBasePointer - Will return a value which either a)
297 /// defines the base pointer for the input or b) blocks the simple search
298 /// (i.e. a PHI or Select of two derived pointers)
299 static Value *findBaseDefiningValue(Value *I) {
300 assert(I->getType()->isPointerTy() &&
301 "Illegal to ask for the base pointer of a non-pointer type");
303 // This case is a bit of a hack - it only handles extracts from vectors which
304 // trivially contain only base pointers. See note inside the function for
305 // how to improve this.
306 if (auto *EEI = dyn_cast<ExtractElementInst>(I)) {
307 Value *VectorOperand = EEI->getVectorOperand();
308 Value *VectorBase = findBaseOfVector(VectorOperand);
310 assert(VectorBase && "extract element not known to be a trivial base");
314 if (isa<Argument>(I))
315 // An incoming argument to the function is a base pointer
316 // We should have never reached here if this argument isn't an gc value
319 if (isa<GlobalVariable>(I))
323 // inlining could possibly introduce phi node that contains
324 // undef if callee has multiple returns
325 if (isa<UndefValue>(I))
326 // utterly meaningless, but useful for dealing with
327 // partially optimized code.
330 // Due to inheritance, this must be _after_ the global variable and undef
332 if (Constant *Con = dyn_cast<Constant>(I)) {
333 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
334 "order of checks wrong!");
335 // Note: Finding a constant base for something marked for relocation
336 // doesn't really make sense. The most likely case is either a) some
337 // screwed up the address space usage or b) your validating against
338 // compiled C++ code w/o the proper separation. The only real exception
339 // is a null pointer. You could have generic code written to index of
340 // off a potentially null value and have proven it null. We also use
341 // null pointers in dead paths of relocation phis (which we might later
342 // want to find a base pointer for).
343 assert(isa<ConstantPointerNull>(Con) &&
344 "null is the only case which makes sense");
348 if (CastInst *CI = dyn_cast<CastInst>(I)) {
349 Value *Def = CI->stripPointerCasts();
350 // If we find a cast instruction here, it means we've found a cast which is
351 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
352 // handle int->ptr conversion.
353 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
354 return findBaseDefiningValue(Def);
357 if (isa<LoadInst>(I))
358 return I; // The value loaded is an gc base itself
360 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
361 // The base of this GEP is the base
362 return findBaseDefiningValue(GEP->getPointerOperand());
364 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
365 switch (II->getIntrinsicID()) {
366 case Intrinsic::experimental_gc_result_ptr:
368 // fall through to general call handling
370 case Intrinsic::experimental_gc_statepoint:
371 case Intrinsic::experimental_gc_result_float:
372 case Intrinsic::experimental_gc_result_int:
373 llvm_unreachable("these don't produce pointers");
374 case Intrinsic::experimental_gc_relocate: {
375 // Rerunning safepoint insertion after safepoints are already
376 // inserted is not supported. It could probably be made to work,
377 // but why are you doing this? There's no good reason.
378 llvm_unreachable("repeat safepoint insertion is not supported");
380 case Intrinsic::gcroot:
381 // Currently, this mechanism hasn't been extended to work with gcroot.
382 // There's no reason it couldn't be, but I haven't thought about the
383 // implications much.
385 "interaction with the gcroot mechanism is not supported");
388 // We assume that functions in the source language only return base
389 // pointers. This should probably be generalized via attributes to support
390 // both source language and internal functions.
391 if (isa<CallInst>(I) || isa<InvokeInst>(I))
394 // I have absolutely no idea how to implement this part yet. It's not
395 // neccessarily hard, I just haven't really looked at it yet.
396 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
398 if (isa<AtomicCmpXchgInst>(I))
399 // A CAS is effectively a atomic store and load combined under a
400 // predicate. From the perspective of base pointers, we just treat it
404 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
405 "binary ops which don't apply to pointers");
407 // The aggregate ops. Aggregates can either be in the heap or on the
408 // stack, but in either case, this is simply a field load. As a result,
409 // this is a defining definition of the base just like a load is.
410 if (isa<ExtractValueInst>(I))
413 // We should never see an insert vector since that would require we be
414 // tracing back a struct value not a pointer value.
415 assert(!isa<InsertValueInst>(I) &&
416 "Base pointer for a struct is meaningless");
418 // The last two cases here don't return a base pointer. Instead, they
419 // return a value which dynamically selects from amoung several base
420 // derived pointers (each with it's own base potentially). It's the job of
421 // the caller to resolve these.
422 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
423 "missing instruction case in findBaseDefiningValing");
427 /// Returns the base defining value for this value.
428 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
429 Value *&Cached = Cache[I];
431 Cached = findBaseDefiningValue(I);
433 assert(Cache[I] != nullptr);
436 dbgs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
442 /// Return a base pointer for this value if known. Otherwise, return it's
443 /// base defining value.
444 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
445 Value *Def = findBaseDefiningValueCached(I, Cache);
446 auto Found = Cache.find(Def);
447 if (Found != Cache.end()) {
448 // Either a base-of relation, or a self reference. Caller must check.
449 return Found->second;
451 // Only a BDV available
455 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
456 /// is it known to be a base pointer? Or do we need to continue searching.
457 static bool isKnownBaseResult(Value *V) {
458 if (!isa<PHINode>(V) && !isa<SelectInst>(V)) {
459 // no recursion possible
462 if (isa<Instruction>(V) &&
463 cast<Instruction>(V)->getMetadata("is_base_value")) {
464 // This is a previously inserted base phi or select. We know
465 // that this is a base value.
469 // We need to keep searching
473 // TODO: find a better name for this
477 enum Status { Unknown, Base, Conflict };
479 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
480 assert(status != Base || b);
482 PhiState(Value *b) : status(Base), base(b) {}
483 PhiState() : status(Unknown), base(nullptr) {}
485 Status getStatus() const { return status; }
486 Value *getBase() const { return base; }
488 bool isBase() const { return getStatus() == Base; }
489 bool isUnknown() const { return getStatus() == Unknown; }
490 bool isConflict() const { return getStatus() == Conflict; }
492 bool operator==(const PhiState &other) const {
493 return base == other.base && status == other.status;
496 bool operator!=(const PhiState &other) const { return !(*this == other); }
499 errs() << status << " (" << base << " - "
500 << (base ? base->getName() : "nullptr") << "): ";
505 Value *base; // non null only if status == base
508 typedef DenseMap<Value *, PhiState> ConflictStateMapTy;
509 // Values of type PhiState form a lattice, and this is a helper
510 // class that implementes the meet operation. The meat of the meet
511 // operation is implemented in MeetPhiStates::pureMeet
512 class MeetPhiStates {
514 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
515 explicit MeetPhiStates(const ConflictStateMapTy &phiStates)
516 : phiStates(phiStates) {}
518 // Destructively meet the current result with the base V. V can
519 // either be a merge instruction (SelectInst / PHINode), in which
520 // case its status is looked up in the phiStates map; or a regular
521 // SSA value, in which case it is assumed to be a base.
522 void meetWith(Value *V) {
523 PhiState otherState = getStateForBDV(V);
524 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
525 MeetPhiStates::pureMeet(currentResult, otherState)) &&
526 "math is wrong: meet does not commute!");
527 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
530 PhiState getResult() const { return currentResult; }
533 const ConflictStateMapTy &phiStates;
534 PhiState currentResult;
536 /// Return a phi state for a base defining value. We'll generate a new
537 /// base state for known bases and expect to find a cached state otherwise
538 PhiState getStateForBDV(Value *baseValue) {
539 if (isKnownBaseResult(baseValue)) {
540 return PhiState(baseValue);
542 return lookupFromMap(baseValue);
546 PhiState lookupFromMap(Value *V) {
547 auto I = phiStates.find(V);
548 assert(I != phiStates.end() && "lookup failed!");
552 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
553 switch (stateA.getStatus()) {
554 case PhiState::Unknown:
558 assert(stateA.getBase() && "can't be null");
559 if (stateB.isUnknown())
562 if (stateB.isBase()) {
563 if (stateA.getBase() == stateB.getBase()) {
564 assert(stateA == stateB && "equality broken!");
567 return PhiState(PhiState::Conflict);
569 assert(stateB.isConflict() && "only three states!");
570 return PhiState(PhiState::Conflict);
572 case PhiState::Conflict:
575 llvm_unreachable("only three states!");
579 /// For a given value or instruction, figure out what base ptr it's derived
580 /// from. For gc objects, this is simply itself. On success, returns a value
581 /// which is the base pointer. (This is reliable and can be used for
582 /// relocation.) On failure, returns nullptr.
583 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
584 Value *def = findBaseOrBDV(I, cache);
586 if (isKnownBaseResult(def)) {
590 // Here's the rough algorithm:
591 // - For every SSA value, construct a mapping to either an actual base
592 // pointer or a PHI which obscures the base pointer.
593 // - Construct a mapping from PHI to unknown TOP state. Use an
594 // optimistic algorithm to propagate base pointer information. Lattice
599 // When algorithm terminates, all PHIs will either have a single concrete
600 // base or be in a conflict state.
601 // - For every conflict, insert a dummy PHI node without arguments. Add
602 // these to the base[Instruction] = BasePtr mapping. For every
603 // non-conflict, add the actual base.
604 // - For every conflict, add arguments for the base[a] of each input
607 // Note: A simpler form of this would be to add the conflict form of all
608 // PHIs without running the optimistic algorithm. This would be
609 // analougous to pessimistic data flow and would likely lead to an
610 // overall worse solution.
612 ConflictStateMapTy states;
613 states[def] = PhiState();
614 // Recursively fill in all phis & selects reachable from the initial one
615 // for which we don't already know a definite base value for
616 // TODO: This should be rewritten with a worklist
620 // Since we're adding elements to 'states' as we run, we can't keep
621 // iterators into the set.
622 SmallVector<Value *, 16> Keys;
623 Keys.reserve(states.size());
624 for (auto Pair : states) {
625 Value *V = Pair.first;
628 for (Value *v : Keys) {
629 assert(!isKnownBaseResult(v) && "why did it get added?");
630 if (PHINode *phi = dyn_cast<PHINode>(v)) {
631 assert(phi->getNumIncomingValues() > 0 &&
632 "zero input phis are illegal");
633 for (Value *InVal : phi->incoming_values()) {
634 Value *local = findBaseOrBDV(InVal, cache);
635 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
636 states[local] = PhiState();
640 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
641 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
642 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
643 states[local] = PhiState();
646 local = findBaseOrBDV(sel->getFalseValue(), cache);
647 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
648 states[local] = PhiState();
656 errs() << "States after initialization:\n";
657 for (auto Pair : states) {
658 Instruction *v = cast<Instruction>(Pair.first);
659 PhiState state = Pair.second;
665 // TODO: come back and revisit the state transitions around inputs which
666 // have reached conflict state. The current version seems too conservative.
668 bool progress = true;
671 size_t oldSize = states.size();
674 // We're only changing keys in this loop, thus safe to keep iterators
675 for (auto Pair : states) {
676 MeetPhiStates calculateMeet(states);
677 Value *v = Pair.first;
678 assert(!isKnownBaseResult(v) && "why did it get added?");
679 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
680 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
681 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
683 for (Value *Val : cast<PHINode>(v)->incoming_values())
684 calculateMeet.meetWith(findBaseOrBDV(Val, cache));
686 PhiState oldState = states[v];
687 PhiState newState = calculateMeet.getResult();
688 if (oldState != newState) {
690 states[v] = newState;
694 assert(oldSize <= states.size());
695 assert(oldSize == states.size() || progress);
699 errs() << "States after meet iteration:\n";
700 for (auto Pair : states) {
701 Instruction *v = cast<Instruction>(Pair.first);
702 PhiState state = Pair.second;
708 // Insert Phis for all conflicts
709 // We want to keep naming deterministic in the loop that follows, so
710 // sort the keys before iteration. This is useful in allowing us to
711 // write stable tests. Note that there is no invalidation issue here.
712 SmallVector<Value *, 16> Keys;
713 Keys.reserve(states.size());
714 for (auto Pair : states) {
715 Value *V = Pair.first;
718 std::sort(Keys.begin(), Keys.end(), order_by_name);
719 // TODO: adjust naming patterns to avoid this order of iteration dependency
720 for (Value *V : Keys) {
721 Instruction *v = cast<Instruction>(V);
722 PhiState state = states[V];
723 assert(!isKnownBaseResult(v) && "why did it get added?");
724 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
725 if (!state.isConflict())
728 if (isa<PHINode>(v)) {
730 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
731 assert(num_preds > 0 && "how did we reach here");
732 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
733 // Add metadata marking this as a base value
734 auto *const_1 = ConstantInt::get(
736 v->getParent()->getParent()->getParent()->getContext()),
738 auto MDConst = ConstantAsMetadata::get(const_1);
739 MDNode *md = MDNode::get(
740 v->getParent()->getParent()->getParent()->getContext(), MDConst);
741 phi->setMetadata("is_base_value", md);
742 states[v] = PhiState(PhiState::Conflict, phi);
744 SelectInst *sel = cast<SelectInst>(v);
745 // The undef will be replaced later
746 UndefValue *undef = UndefValue::get(sel->getType());
747 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
748 undef, "base_select", sel);
749 // Add metadata marking this as a base value
750 auto *const_1 = ConstantInt::get(
752 v->getParent()->getParent()->getParent()->getContext()),
754 auto MDConst = ConstantAsMetadata::get(const_1);
755 MDNode *md = MDNode::get(
756 v->getParent()->getParent()->getParent()->getContext(), MDConst);
757 basesel->setMetadata("is_base_value", md);
758 states[v] = PhiState(PhiState::Conflict, basesel);
762 // Fixup all the inputs of the new PHIs
763 for (auto Pair : states) {
764 Instruction *v = cast<Instruction>(Pair.first);
765 PhiState state = Pair.second;
767 assert(!isKnownBaseResult(v) && "why did it get added?");
768 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
769 if (!state.isConflict())
772 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
773 PHINode *phi = cast<PHINode>(v);
774 unsigned NumPHIValues = phi->getNumIncomingValues();
775 for (unsigned i = 0; i < NumPHIValues; i++) {
776 Value *InVal = phi->getIncomingValue(i);
777 BasicBlock *InBB = phi->getIncomingBlock(i);
779 // If we've already seen InBB, add the same incoming value
780 // we added for it earlier. The IR verifier requires phi
781 // nodes with multiple entries from the same basic block
782 // to have the same incoming value for each of those
783 // entries. If we don't do this check here and basephi
784 // has a different type than base, we'll end up adding two
785 // bitcasts (and hence two distinct values) as incoming
786 // values for the same basic block.
788 int blockIndex = basephi->getBasicBlockIndex(InBB);
789 if (blockIndex != -1) {
790 Value *oldBase = basephi->getIncomingValue(blockIndex);
791 basephi->addIncoming(oldBase, InBB);
793 Value *base = findBaseOrBDV(InVal, cache);
794 if (!isKnownBaseResult(base)) {
795 // Either conflict or base.
796 assert(states.count(base));
797 base = states[base].getBase();
798 assert(base != nullptr && "unknown PhiState!");
801 // In essense this assert states: the only way two
802 // values incoming from the same basic block may be
803 // different is by being different bitcasts of the same
804 // value. A cleanup that remains TODO is changing
805 // findBaseOrBDV to return an llvm::Value of the correct
806 // type (and still remain pure). This will remove the
807 // need to add bitcasts.
808 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
809 "sanity -- findBaseOrBDV should be pure!");
814 // Find either the defining value for the PHI or the normal base for
816 Value *base = findBaseOrBDV(InVal, cache);
817 if (!isKnownBaseResult(base)) {
818 // Either conflict or base.
819 assert(states.count(base));
820 base = states[base].getBase();
821 assert(base != nullptr && "unknown PhiState!");
823 assert(base && "can't be null");
824 // Must use original input BB since base may not be Instruction
825 // The cast is needed since base traversal may strip away bitcasts
826 if (base->getType() != basephi->getType()) {
827 base = new BitCastInst(base, basephi->getType(), "cast",
828 InBB->getTerminator());
830 basephi->addIncoming(base, InBB);
832 assert(basephi->getNumIncomingValues() == NumPHIValues);
834 SelectInst *basesel = cast<SelectInst>(state.getBase());
835 SelectInst *sel = cast<SelectInst>(v);
836 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
837 // something more safe and less hacky.
838 for (int i = 1; i <= 2; i++) {
839 Value *InVal = sel->getOperand(i);
840 // Find either the defining value for the PHI or the normal base for
842 Value *base = findBaseOrBDV(InVal, cache);
843 if (!isKnownBaseResult(base)) {
844 // Either conflict or base.
845 assert(states.count(base));
846 base = states[base].getBase();
847 assert(base != nullptr && "unknown PhiState!");
849 assert(base && "can't be null");
850 // Must use original input BB since base may not be Instruction
851 // The cast is needed since base traversal may strip away bitcasts
852 if (base->getType() != basesel->getType()) {
853 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
855 basesel->setOperand(i, base);
860 // Cache all of our results so we can cheaply reuse them
861 // NOTE: This is actually two caches: one of the base defining value
862 // relation and one of the base pointer relation! FIXME
863 for (auto item : states) {
864 Value *v = item.first;
865 Value *base = item.second.getBase();
867 assert(!isKnownBaseResult(v) && "why did it get added?");
870 std::string fromstr =
871 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
873 errs() << "Updating base value cache"
874 << " for: " << (v->hasName() ? v->getName() : "")
875 << " from: " << fromstr
876 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
879 assert(isKnownBaseResult(base) &&
880 "must be something we 'know' is a base pointer");
881 if (cache.count(v)) {
882 // Once we transition from the BDV relation being store in the cache to
883 // the base relation being stored, it must be stable
884 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
885 "base relation should be stable");
889 assert(cache.find(def) != cache.end());
893 // For a set of live pointers (base and/or derived), identify the base
894 // pointer of the object which they are derived from. This routine will
895 // mutate the IR graph as needed to make the 'base' pointer live at the
896 // definition site of 'derived'. This ensures that any use of 'derived' can
897 // also use 'base'. This may involve the insertion of a number of
898 // additional PHI nodes.
900 // preconditions: live is a set of pointer type Values
902 // side effects: may insert PHI nodes into the existing CFG, will preserve
903 // CFG, will not remove or mutate any existing nodes
905 // post condition: PointerToBase contains one (derived, base) pair for every
906 // pointer in live. Note that derived can be equal to base if the original
907 // pointer was a base pointer.
909 findBasePointers(const StatepointLiveSetTy &live,
910 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
911 DominatorTree *DT, DefiningValueMapTy &DVCache) {
912 // For the naming of values inserted to be deterministic - which makes for
913 // much cleaner and more stable tests - we need to assign an order to the
914 // live values. DenseSets do not provide a deterministic order across runs.
915 SmallVector<Value *, 64> Temp;
916 Temp.insert(Temp.end(), live.begin(), live.end());
917 std::sort(Temp.begin(), Temp.end(), order_by_name);
918 for (Value *ptr : Temp) {
919 Value *base = findBasePointer(ptr, DVCache);
920 assert(base && "failed to find base pointer");
921 PointerToBase[ptr] = base;
922 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
923 DT->dominates(cast<Instruction>(base)->getParent(),
924 cast<Instruction>(ptr)->getParent())) &&
925 "The base we found better dominate the derived pointer");
927 // If you see this trip and like to live really dangerously, the code should
928 // be correct, just with idioms the verifier can't handle. You can try
929 // disabling the verifier at your own substaintial risk.
930 assert(!isa<ConstantPointerNull>(base) &&
931 "the relocation code needs adjustment to handle the relocation of "
932 "a null pointer constant without causing false positives in the "
933 "safepoint ir verifier.");
937 /// Find the required based pointers (and adjust the live set) for the given
939 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
941 PartiallyConstructedSafepointRecord &result) {
942 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
943 findBasePointers(result.liveset, PointerToBase, &DT, DVCache);
945 if (PrintBasePointers) {
946 // Note: Need to print these in a stable order since this is checked in
948 errs() << "Base Pairs (w/o Relocation):\n";
949 SmallVector<Value *, 64> Temp;
950 Temp.reserve(PointerToBase.size());
951 for (auto Pair : PointerToBase) {
952 Temp.push_back(Pair.first);
954 std::sort(Temp.begin(), Temp.end(), order_by_name);
955 for (Value *Ptr : Temp) {
956 Value *Base = PointerToBase[Ptr];
957 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
962 result.PointerToBase = PointerToBase;
965 /// Given an updated version of the dataflow liveness results, update the
966 /// liveset and base pointer maps for the call site CS.
967 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
969 PartiallyConstructedSafepointRecord &result);
971 static void recomputeLiveInValues(
972 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
973 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
974 // TODO-PERF: reuse the original liveness, then simply run the dataflow
975 // again. The old values are still live and will help it stablize quickly.
976 GCPtrLivenessData RevisedLivenessData;
977 computeLiveInValues(DT, F, RevisedLivenessData);
978 for (size_t i = 0; i < records.size(); i++) {
979 struct PartiallyConstructedSafepointRecord &info = records[i];
980 const CallSite &CS = toUpdate[i];
981 recomputeLiveInValues(RevisedLivenessData, CS, info);
985 // When inserting gc.relocate calls, we need to ensure there are no uses
986 // of the original value between the gc.statepoint and the gc.relocate call.
987 // One case which can arise is a phi node starting one of the successor blocks.
988 // We also need to be able to insert the gc.relocates only on the path which
989 // goes through the statepoint. We might need to split an edge to make this
992 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent, Pass *P) {
993 DominatorTree *DT = nullptr;
994 if (auto *DTP = P->getAnalysisIfAvailable<DominatorTreeWrapperPass>())
995 DT = &DTP->getDomTree();
997 BasicBlock *Ret = BB;
998 if (!BB->getUniquePredecessor()) {
999 Ret = SplitBlockPredecessors(BB, InvokeParent, "", nullptr, DT);
1002 // Now that 'ret' has unique predecessor we can safely remove all phi nodes
1004 FoldSingleEntryPHINodes(Ret);
1005 assert(!isa<PHINode>(Ret->begin()));
1007 // At this point, we can safely insert a gc.relocate as the first instruction
1008 // in Ret if needed.
1012 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1013 auto itr = std::find(livevec.begin(), livevec.end(), val);
1014 assert(livevec.end() != itr);
1015 size_t index = std::distance(livevec.begin(), itr);
1016 assert(index < livevec.size());
1020 // Create new attribute set containing only attributes which can be transfered
1021 // from original call to the safepoint.
1022 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1025 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1026 unsigned index = AS.getSlotIndex(Slot);
1028 if (index == AttributeSet::ReturnIndex ||
1029 index == AttributeSet::FunctionIndex) {
1031 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1033 Attribute attr = *it;
1035 // Do not allow certain attributes - just skip them
1036 // Safepoint can not be read only or read none.
1037 if (attr.hasAttribute(Attribute::ReadNone) ||
1038 attr.hasAttribute(Attribute::ReadOnly))
1041 ret = ret.addAttributes(
1042 AS.getContext(), index,
1043 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1047 // Just skip parameter attributes for now
1053 /// Helper function to place all gc relocates necessary for the given
1056 /// liveVariables - list of variables to be relocated.
1057 /// liveStart - index of the first live variable.
1058 /// basePtrs - base pointers.
1059 /// statepointToken - statepoint instruction to which relocates should be
1061 /// Builder - Llvm IR builder to be used to construct new calls.
1062 static void CreateGCRelocates(ArrayRef<llvm::Value *> liveVariables,
1063 const int liveStart,
1064 ArrayRef<llvm::Value *> basePtrs,
1065 Instruction *statepointToken,
1066 IRBuilder<> Builder) {
1067 SmallVector<Instruction *, 64> NewDefs;
1068 NewDefs.reserve(liveVariables.size());
1070 Module *M = statepointToken->getParent()->getParent()->getParent();
1072 for (unsigned i = 0; i < liveVariables.size(); i++) {
1073 // We generate a (potentially) unique declaration for every pointer type
1074 // combination. This results is some blow up the function declarations in
1075 // the IR, but removes the need for argument bitcasts which shrinks the IR
1076 // greatly and makes it much more readable.
1077 SmallVector<Type *, 1> types; // one per 'any' type
1078 types.push_back(liveVariables[i]->getType()); // result type
1079 Value *gc_relocate_decl = Intrinsic::getDeclaration(
1080 M, Intrinsic::experimental_gc_relocate, types);
1082 // Generate the gc.relocate call and save the result
1084 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1085 liveStart + find_index(liveVariables, basePtrs[i]));
1086 Value *liveIdx = ConstantInt::get(
1087 Type::getInt32Ty(M->getContext()),
1088 liveStart + find_index(liveVariables, liveVariables[i]));
1090 // only specify a debug name if we can give a useful one
1091 Value *reloc = Builder.CreateCall3(
1092 gc_relocate_decl, statepointToken, baseIdx, liveIdx,
1093 liveVariables[i]->hasName() ? liveVariables[i]->getName() + ".relocated"
1095 // Trick CodeGen into thinking there are lots of free registers at this
1097 cast<CallInst>(reloc)->setCallingConv(CallingConv::Cold);
1099 NewDefs.push_back(cast<Instruction>(reloc));
1101 assert(NewDefs.size() == liveVariables.size() &&
1102 "missing or extra redefinition at safepoint");
1106 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1107 const SmallVectorImpl<llvm::Value *> &basePtrs,
1108 const SmallVectorImpl<llvm::Value *> &liveVariables,
1110 PartiallyConstructedSafepointRecord &result) {
1111 assert(basePtrs.size() == liveVariables.size());
1112 assert(isStatepoint(CS) &&
1113 "This method expects to be rewriting a statepoint");
1115 BasicBlock *BB = CS.getInstruction()->getParent();
1117 Function *F = BB->getParent();
1118 assert(F && "must be set");
1119 Module *M = F->getParent();
1121 assert(M && "must be set");
1123 // We're not changing the function signature of the statepoint since the gc
1124 // arguments go into the var args section.
1125 Function *gc_statepoint_decl = CS.getCalledFunction();
1127 // Then go ahead and use the builder do actually do the inserts. We insert
1128 // immediately before the previous instruction under the assumption that all
1129 // arguments will be available here. We can't insert afterwards since we may
1130 // be replacing a terminator.
1131 Instruction *insertBefore = CS.getInstruction();
1132 IRBuilder<> Builder(insertBefore);
1133 // Copy all of the arguments from the original statepoint - this includes the
1134 // target, call args, and deopt args
1135 SmallVector<llvm::Value *, 64> args;
1136 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1137 // TODO: Clear the 'needs rewrite' flag
1139 // add all the pointers to be relocated (gc arguments)
1140 // Capture the start of the live variable list for use in the gc_relocates
1141 const int live_start = args.size();
1142 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1144 // Create the statepoint given all the arguments
1145 Instruction *token = nullptr;
1146 AttributeSet return_attributes;
1148 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1150 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1151 call->setTailCall(toReplace->isTailCall());
1152 call->setCallingConv(toReplace->getCallingConv());
1154 // Currently we will fail on parameter attributes and on certain
1155 // function attributes.
1156 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1157 // In case if we can handle this set of sttributes - set up function attrs
1158 // directly on statepoint and return attrs later for gc_result intrinsic.
1159 call->setAttributes(new_attrs.getFnAttributes());
1160 return_attributes = new_attrs.getRetAttributes();
1164 // Put the following gc_result and gc_relocate calls immediately after the
1165 // the old call (which we're about to delete)
1166 BasicBlock::iterator next(toReplace);
1167 assert(BB->end() != next && "not a terminator, must have next");
1169 Instruction *IP = &*(next);
1170 Builder.SetInsertPoint(IP);
1171 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1174 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1176 // Insert the new invoke into the old block. We'll remove the old one in a
1177 // moment at which point this will become the new terminator for the
1179 InvokeInst *invoke = InvokeInst::Create(
1180 gc_statepoint_decl, toReplace->getNormalDest(),
1181 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1182 invoke->setCallingConv(toReplace->getCallingConv());
1184 // Currently we will fail on parameter attributes and on certain
1185 // function attributes.
1186 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1187 // In case if we can handle this set of sttributes - set up function attrs
1188 // directly on statepoint and return attrs later for gc_result intrinsic.
1189 invoke->setAttributes(new_attrs.getFnAttributes());
1190 return_attributes = new_attrs.getRetAttributes();
1194 // Generate gc relocates in exceptional path
1195 BasicBlock *unwindBlock = toReplace->getUnwindDest();
1196 assert(!isa<PHINode>(unwindBlock->begin()) &&
1197 unwindBlock->getUniquePredecessor() &&
1198 "can't safely insert in this block!");
1200 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1201 Builder.SetInsertPoint(IP);
1202 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1204 // Extract second element from landingpad return value. We will attach
1205 // exceptional gc relocates to it.
1206 const unsigned idx = 1;
1207 Instruction *exceptional_token =
1208 cast<Instruction>(Builder.CreateExtractValue(
1209 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1210 result.UnwindToken = exceptional_token;
1212 // Just throw away return value. We will use the one we got for normal
1214 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1215 exceptional_token, Builder);
1217 // Generate gc relocates and returns for normal block
1218 BasicBlock *normalDest = toReplace->getNormalDest();
1219 assert(!isa<PHINode>(normalDest->begin()) &&
1220 normalDest->getUniquePredecessor() &&
1221 "can't safely insert in this block!");
1223 IP = &*(normalDest->getFirstInsertionPt());
1224 Builder.SetInsertPoint(IP);
1226 // gc relocates will be generated later as if it were regular call
1231 // Take the name of the original value call if it had one.
1232 token->takeName(CS.getInstruction());
1234 // The GCResult is already inserted, we just need to find it
1236 Instruction *toReplace = CS.getInstruction();
1237 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1238 "only valid use before rewrite is gc.result");
1239 assert(!toReplace->hasOneUse() ||
1240 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1243 // Update the gc.result of the original statepoint (if any) to use the newly
1244 // inserted statepoint. This is safe to do here since the token can't be
1245 // considered a live reference.
1246 CS.getInstruction()->replaceAllUsesWith(token);
1248 result.StatepointToken = token;
1250 // Second, create a gc.relocate for every live variable
1251 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1255 struct name_ordering {
1258 bool operator()(name_ordering const &a, name_ordering const &b) {
1259 return -1 == a.derived->getName().compare(b.derived->getName());
1263 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1264 SmallVectorImpl<Value *> &livevec) {
1265 assert(basevec.size() == livevec.size());
1267 SmallVector<name_ordering, 64> temp;
1268 for (size_t i = 0; i < basevec.size(); i++) {
1270 v.base = basevec[i];
1271 v.derived = livevec[i];
1274 std::sort(temp.begin(), temp.end(), name_ordering());
1275 for (size_t i = 0; i < basevec.size(); i++) {
1276 basevec[i] = temp[i].base;
1277 livevec[i] = temp[i].derived;
1281 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1282 // which make the relocations happening at this safepoint explicit.
1284 // WARNING: Does not do any fixup to adjust users of the original live
1285 // values. That's the callers responsibility.
1287 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1288 PartiallyConstructedSafepointRecord &result) {
1289 auto liveset = result.liveset;
1290 auto PointerToBase = result.PointerToBase;
1292 // Convert to vector for efficient cross referencing.
1293 SmallVector<Value *, 64> basevec, livevec;
1294 livevec.reserve(liveset.size());
1295 basevec.reserve(liveset.size());
1296 for (Value *L : liveset) {
1297 livevec.push_back(L);
1299 assert(PointerToBase.find(L) != PointerToBase.end());
1300 Value *base = PointerToBase[L];
1301 basevec.push_back(base);
1303 assert(livevec.size() == basevec.size());
1305 // To make the output IR slightly more stable (for use in diffs), ensure a
1306 // fixed order of the values in the safepoint (by sorting the value name).
1307 // The order is otherwise meaningless.
1308 stablize_order(basevec, livevec);
1310 // Do the actual rewriting and delete the old statepoint
1311 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1312 CS.getInstruction()->eraseFromParent();
1315 // Helper function for the relocationViaAlloca.
1316 // It receives iterator to the statepoint gc relocates and emits store to the
1318 // location (via allocaMap) for the each one of them.
1319 // Add visited values into the visitedLiveValues set we will later use them
1320 // for sanity check.
1322 insertRelocationStores(iterator_range<Value::user_iterator> gcRelocs,
1323 DenseMap<Value *, Value *> &allocaMap,
1324 DenseSet<Value *> &visitedLiveValues) {
1326 for (User *U : gcRelocs) {
1327 if (!isa<IntrinsicInst>(U))
1330 IntrinsicInst *relocatedValue = cast<IntrinsicInst>(U);
1332 // We only care about relocates
1333 if (relocatedValue->getIntrinsicID() !=
1334 Intrinsic::experimental_gc_relocate) {
1338 GCRelocateOperands relocateOperands(relocatedValue);
1339 Value *originalValue =
1340 const_cast<Value *>(relocateOperands.getDerivedPtr());
1341 assert(allocaMap.count(originalValue));
1342 Value *alloca = allocaMap[originalValue];
1344 // Emit store into the related alloca
1345 StoreInst *store = new StoreInst(relocatedValue, alloca);
1346 store->insertAfter(relocatedValue);
1349 visitedLiveValues.insert(originalValue);
1354 /// do all the relocation update via allocas and mem2reg
1355 static void relocationViaAlloca(
1356 Function &F, DominatorTree &DT, ArrayRef<Value *> live,
1357 ArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1359 // record initial number of (static) allocas; we'll check we have the same
1360 // number when we get done.
1361 int InitialAllocaNum = 0;
1362 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1364 if (isa<AllocaInst>(*I))
1368 // TODO-PERF: change data structures, reserve
1369 DenseMap<Value *, Value *> allocaMap;
1370 SmallVector<AllocaInst *, 200> PromotableAllocas;
1371 PromotableAllocas.reserve(live.size());
1373 // emit alloca for each live gc pointer
1374 for (unsigned i = 0; i < live.size(); i++) {
1375 Value *liveValue = live[i];
1376 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1377 F.getEntryBlock().getFirstNonPHI());
1378 allocaMap[liveValue] = alloca;
1379 PromotableAllocas.push_back(alloca);
1382 // The next two loops are part of the same conceptual operation. We need to
1383 // insert a store to the alloca after the original def and at each
1384 // redefinition. We need to insert a load before each use. These are split
1385 // into distinct loops for performance reasons.
1387 // update gc pointer after each statepoint
1388 // either store a relocated value or null (if no relocated value found for
1389 // this gc pointer and it is not a gc_result)
1390 // this must happen before we update the statepoint with load of alloca
1391 // otherwise we lose the link between statepoint and old def
1392 for (size_t i = 0; i < records.size(); i++) {
1393 const struct PartiallyConstructedSafepointRecord &info = records[i];
1394 Value *Statepoint = info.StatepointToken;
1396 // This will be used for consistency check
1397 DenseSet<Value *> visitedLiveValues;
1399 // Insert stores for normal statepoint gc relocates
1400 insertRelocationStores(Statepoint->users(), allocaMap, visitedLiveValues);
1402 // In case if it was invoke statepoint
1403 // we will insert stores for exceptional path gc relocates.
1404 if (isa<InvokeInst>(Statepoint)) {
1405 insertRelocationStores(info.UnwindToken->users(), allocaMap,
1409 if (ClobberNonLive) {
1410 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1411 // the gc.statepoint. This will turn some subtle GC problems into
1412 // slightly easier to debug SEGVs. Note that on large IR files with
1413 // lots of gc.statepoints this is extremely costly both memory and time
1415 SmallVector<AllocaInst *, 64> ToClobber;
1416 for (auto Pair : allocaMap) {
1417 Value *Def = Pair.first;
1418 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1420 // This value was relocated
1421 if (visitedLiveValues.count(Def)) {
1424 ToClobber.push_back(Alloca);
1427 auto InsertClobbersAt = [&](Instruction *IP) {
1428 for (auto *AI : ToClobber) {
1429 auto AIType = cast<PointerType>(AI->getType());
1430 auto PT = cast<PointerType>(AIType->getElementType());
1431 Constant *CPN = ConstantPointerNull::get(PT);
1432 StoreInst *store = new StoreInst(CPN, AI);
1433 store->insertBefore(IP);
1437 // Insert the clobbering stores. These may get intermixed with the
1438 // gc.results and gc.relocates, but that's fine.
1439 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1440 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1441 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1443 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1445 InsertClobbersAt(Next);
1449 // update use with load allocas and add store for gc_relocated
1450 for (auto Pair : allocaMap) {
1451 Value *def = Pair.first;
1452 Value *alloca = Pair.second;
1454 // we pre-record the uses of allocas so that we dont have to worry about
1456 // that change the user information.
1457 SmallVector<Instruction *, 20> uses;
1458 // PERF: trade a linear scan for repeated reallocation
1459 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1460 for (User *U : def->users()) {
1461 if (!isa<ConstantExpr>(U)) {
1462 // If the def has a ConstantExpr use, then the def is either a
1463 // ConstantExpr use itself or null. In either case
1464 // (recursively in the first, directly in the second), the oop
1465 // it is ultimately dependent on is null and this particular
1466 // use does not need to be fixed up.
1467 uses.push_back(cast<Instruction>(U));
1471 std::sort(uses.begin(), uses.end());
1472 auto last = std::unique(uses.begin(), uses.end());
1473 uses.erase(last, uses.end());
1475 for (Instruction *use : uses) {
1476 if (isa<PHINode>(use)) {
1477 PHINode *phi = cast<PHINode>(use);
1478 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1479 if (def == phi->getIncomingValue(i)) {
1480 LoadInst *load = new LoadInst(
1481 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1482 phi->setIncomingValue(i, load);
1486 LoadInst *load = new LoadInst(alloca, "", use);
1487 use->replaceUsesOfWith(def, load);
1491 // emit store for the initial gc value
1492 // store must be inserted after load, otherwise store will be in alloca's
1493 // use list and an extra load will be inserted before it
1494 StoreInst *store = new StoreInst(def, alloca);
1495 if (Instruction *inst = dyn_cast<Instruction>(def)) {
1496 if (InvokeInst *invoke = dyn_cast<InvokeInst>(inst)) {
1497 // InvokeInst is a TerminatorInst so the store need to be inserted
1498 // into its normal destination block.
1499 BasicBlock *normalDest = invoke->getNormalDest();
1500 store->insertBefore(normalDest->getFirstNonPHI());
1502 assert(!inst->isTerminator() &&
1503 "The only TerminatorInst that can produce a value is "
1504 "InvokeInst which is handled above.");
1505 store->insertAfter(inst);
1508 assert(isa<Argument>(def));
1509 store->insertAfter(cast<Instruction>(alloca));
1513 assert(PromotableAllocas.size() == live.size() &&
1514 "we must have the same allocas with lives");
1515 if (!PromotableAllocas.empty()) {
1516 // apply mem2reg to promote alloca to SSA
1517 PromoteMemToReg(PromotableAllocas, DT);
1521 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1523 if (isa<AllocaInst>(*I))
1525 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1529 /// Implement a unique function which doesn't require we sort the input
1530 /// vector. Doing so has the effect of changing the output of a couple of
1531 /// tests in ways which make them less useful in testing fused safepoints.
1532 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1534 SmallVector<T, 128> TempVec;
1535 TempVec.reserve(Vec.size());
1536 for (auto Element : Vec)
1537 TempVec.push_back(Element);
1539 for (auto V : TempVec) {
1540 if (Seen.insert(V).second) {
1546 /// Insert holders so that each Value is obviously live through the entire
1547 /// lifetime of the call.
1548 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1549 SmallVectorImpl<CallInst *> &Holders) {
1551 // No values to hold live, might as well not insert the empty holder
1554 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1555 // Use a dummy vararg function to actually hold the values live
1556 Function *Func = cast<Function>(M->getOrInsertFunction(
1557 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1559 // For call safepoints insert dummy calls right after safepoint
1560 BasicBlock::iterator Next(CS.getInstruction());
1562 Holders.push_back(CallInst::Create(Func, Values, "", Next));
1565 // For invoke safepooints insert dummy calls both in normal and
1566 // exceptional destination blocks
1567 auto *II = cast<InvokeInst>(CS.getInstruction());
1568 Holders.push_back(CallInst::Create(
1569 Func, Values, "", II->getNormalDest()->getFirstInsertionPt()));
1570 Holders.push_back(CallInst::Create(
1571 Func, Values, "", II->getUnwindDest()->getFirstInsertionPt()));
1574 static void findLiveReferences(
1575 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1576 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1577 GCPtrLivenessData OriginalLivenessData;
1578 computeLiveInValues(DT, F, OriginalLivenessData);
1579 for (size_t i = 0; i < records.size(); i++) {
1580 struct PartiallyConstructedSafepointRecord &info = records[i];
1581 const CallSite &CS = toUpdate[i];
1582 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1586 /// Remove any vector of pointers from the liveset by scalarizing them over the
1587 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1588 /// would be preferrable to include the vector in the statepoint itself, but
1589 /// the lowering code currently does not handle that. Extending it would be
1590 /// slightly non-trivial since it requires a format change. Given how rare
1591 /// such cases are (for the moment?) scalarizing is an acceptable comprimise.
1592 static void splitVectorValues(Instruction *StatepointInst,
1593 StatepointLiveSetTy &LiveSet, DominatorTree &DT) {
1594 SmallVector<Value *, 16> ToSplit;
1595 for (Value *V : LiveSet)
1596 if (isa<VectorType>(V->getType()))
1597 ToSplit.push_back(V);
1599 if (ToSplit.empty())
1602 Function &F = *(StatepointInst->getParent()->getParent());
1604 DenseMap<Value *, AllocaInst *> AllocaMap;
1605 // First is normal return, second is exceptional return (invoke only)
1606 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1607 for (Value *V : ToSplit) {
1610 AllocaInst *Alloca =
1611 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1612 AllocaMap[V] = Alloca;
1614 VectorType *VT = cast<VectorType>(V->getType());
1615 IRBuilder<> Builder(StatepointInst);
1616 SmallVector<Value *, 16> Elements;
1617 for (unsigned i = 0; i < VT->getNumElements(); i++)
1618 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1619 LiveSet.insert(Elements.begin(), Elements.end());
1621 auto InsertVectorReform = [&](Instruction *IP) {
1622 Builder.SetInsertPoint(IP);
1623 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1624 Value *ResultVec = UndefValue::get(VT);
1625 for (unsigned i = 0; i < VT->getNumElements(); i++)
1626 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1627 Builder.getInt32(i));
1631 if (isa<CallInst>(StatepointInst)) {
1632 BasicBlock::iterator Next(StatepointInst);
1634 Instruction *IP = &*(Next);
1635 Replacements[V].first = InsertVectorReform(IP);
1636 Replacements[V].second = nullptr;
1638 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1639 // We've already normalized - check that we don't have shared destination
1641 BasicBlock *NormalDest = Invoke->getNormalDest();
1642 assert(!isa<PHINode>(NormalDest->begin()));
1643 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1644 assert(!isa<PHINode>(UnwindDest->begin()));
1645 // Insert insert element sequences in both successors
1646 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1647 Replacements[V].first = InsertVectorReform(IP);
1648 IP = &*(UnwindDest->getFirstInsertionPt());
1649 Replacements[V].second = InsertVectorReform(IP);
1652 for (Value *V : ToSplit) {
1653 AllocaInst *Alloca = AllocaMap[V];
1655 // Capture all users before we start mutating use lists
1656 SmallVector<Instruction *, 16> Users;
1657 for (User *U : V->users())
1658 Users.push_back(cast<Instruction>(U));
1660 for (Instruction *I : Users) {
1661 if (auto Phi = dyn_cast<PHINode>(I)) {
1662 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1663 if (V == Phi->getIncomingValue(i)) {
1664 LoadInst *Load = new LoadInst(
1665 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1666 Phi->setIncomingValue(i, Load);
1669 LoadInst *Load = new LoadInst(Alloca, "", I);
1670 I->replaceUsesOfWith(V, Load);
1674 // Store the original value and the replacement value into the alloca
1675 StoreInst *Store = new StoreInst(V, Alloca);
1676 if (auto I = dyn_cast<Instruction>(V))
1677 Store->insertAfter(I);
1679 Store->insertAfter(Alloca);
1681 // Normal return for invoke, or call return
1682 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1683 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1684 // Unwind return for invoke only
1685 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1687 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1690 // apply mem2reg to promote alloca to SSA
1691 SmallVector<AllocaInst *, 16> Allocas;
1692 for (Value *V : ToSplit)
1693 Allocas.push_back(AllocaMap[V]);
1694 PromoteMemToReg(Allocas, DT);
1697 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1698 SmallVectorImpl<CallSite> &toUpdate) {
1700 // sanity check the input
1701 std::set<CallSite> uniqued;
1702 uniqued.insert(toUpdate.begin(), toUpdate.end());
1703 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1705 for (size_t i = 0; i < toUpdate.size(); i++) {
1706 CallSite &CS = toUpdate[i];
1707 assert(CS.getInstruction()->getParent()->getParent() == &F);
1708 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1712 // When inserting gc.relocates for invokes, we need to be able to insert at
1713 // the top of the successor blocks. See the comment on
1714 // normalForInvokeSafepoint on exactly what is needed. Note that this step
1715 // may restructure the CFG.
1716 for (CallSite CS : toUpdate) {
1719 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1720 normalizeForInvokeSafepoint(invoke->getNormalDest(), invoke->getParent(),
1722 normalizeForInvokeSafepoint(invoke->getUnwindDest(), invoke->getParent(),
1726 // A list of dummy calls added to the IR to keep various values obviously
1727 // live in the IR. We'll remove all of these when done.
1728 SmallVector<CallInst *, 64> holders;
1730 // Insert a dummy call with all of the arguments to the vm_state we'll need
1731 // for the actual safepoint insertion. This ensures reference arguments in
1732 // the deopt argument list are considered live through the safepoint (and
1733 // thus makes sure they get relocated.)
1734 for (size_t i = 0; i < toUpdate.size(); i++) {
1735 CallSite &CS = toUpdate[i];
1736 Statepoint StatepointCS(CS);
1738 SmallVector<Value *, 64> DeoptValues;
1739 for (Use &U : StatepointCS.vm_state_args()) {
1740 Value *Arg = cast<Value>(&U);
1741 assert(!isUnhandledGCPointerType(Arg->getType()) &&
1742 "support for FCA unimplemented");
1743 if (isHandledGCPointerType(Arg->getType()))
1744 DeoptValues.push_back(Arg);
1746 insertUseHolderAfter(CS, DeoptValues, holders);
1749 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
1750 records.reserve(toUpdate.size());
1751 for (size_t i = 0; i < toUpdate.size(); i++) {
1752 struct PartiallyConstructedSafepointRecord info;
1753 records.push_back(info);
1755 assert(records.size() == toUpdate.size());
1757 // A) Identify all gc pointers which are staticly live at the given call
1759 findLiveReferences(F, DT, P, toUpdate, records);
1761 // Do a limited scalarization of any live at safepoint vector values which
1762 // contain pointers. This enables this pass to run after vectorization at
1763 // the cost of some possible performance loss. TODO: it would be nice to
1764 // natively support vectors all the way through the backend so we don't need
1765 // to scalarize here.
1766 for (size_t i = 0; i < records.size(); i++) {
1767 struct PartiallyConstructedSafepointRecord &info = records[i];
1768 Instruction *statepoint = toUpdate[i].getInstruction();
1769 splitVectorValues(cast<Instruction>(statepoint), info.liveset, DT);
1772 // B) Find the base pointers for each live pointer
1773 /* scope for caching */ {
1774 // Cache the 'defining value' relation used in the computation and
1775 // insertion of base phis and selects. This ensures that we don't insert
1776 // large numbers of duplicate base_phis.
1777 DefiningValueMapTy DVCache;
1779 for (size_t i = 0; i < records.size(); i++) {
1780 struct PartiallyConstructedSafepointRecord &info = records[i];
1781 CallSite &CS = toUpdate[i];
1782 findBasePointers(DT, DVCache, CS, info);
1784 } // end of cache scope
1786 // The base phi insertion logic (for any safepoint) may have inserted new
1787 // instructions which are now live at some safepoint. The simplest such
1790 // phi a <-- will be a new base_phi here
1791 // safepoint 1 <-- that needs to be live here
1795 // We insert some dummy calls after each safepoint to definitely hold live
1796 // the base pointers which were identified for that safepoint. We'll then
1797 // ask liveness for _every_ base inserted to see what is now live. Then we
1798 // remove the dummy calls.
1799 holders.reserve(holders.size() + records.size());
1800 for (size_t i = 0; i < records.size(); i++) {
1801 struct PartiallyConstructedSafepointRecord &info = records[i];
1802 CallSite &CS = toUpdate[i];
1804 SmallVector<Value *, 128> Bases;
1805 for (auto Pair : info.PointerToBase) {
1806 Bases.push_back(Pair.second);
1808 insertUseHolderAfter(CS, Bases, holders);
1811 // By selecting base pointers, we've effectively inserted new uses. Thus, we
1812 // need to rerun liveness. We may *also* have inserted new defs, but that's
1813 // not the key issue.
1814 recomputeLiveInValues(F, DT, P, toUpdate, records);
1816 if (PrintBasePointers) {
1817 for (size_t i = 0; i < records.size(); i++) {
1818 struct PartiallyConstructedSafepointRecord &info = records[i];
1819 errs() << "Base Pairs: (w/Relocation)\n";
1820 for (auto Pair : info.PointerToBase) {
1821 errs() << " derived %" << Pair.first->getName() << " base %"
1822 << Pair.second->getName() << "\n";
1826 for (size_t i = 0; i < holders.size(); i++) {
1827 holders[i]->eraseFromParent();
1828 holders[i] = nullptr;
1832 // Now run through and replace the existing statepoints with new ones with
1833 // the live variables listed. We do not yet update uses of the values being
1834 // relocated. We have references to live variables that need to
1835 // survive to the last iteration of this loop. (By construction, the
1836 // previous statepoint can not be a live variable, thus we can and remove
1837 // the old statepoint calls as we go.)
1838 for (size_t i = 0; i < records.size(); i++) {
1839 struct PartiallyConstructedSafepointRecord &info = records[i];
1840 CallSite &CS = toUpdate[i];
1841 makeStatepointExplicit(DT, CS, P, info);
1843 toUpdate.clear(); // prevent accident use of invalid CallSites
1845 // Do all the fixups of the original live variables to their relocated selves
1846 SmallVector<Value *, 128> live;
1847 for (size_t i = 0; i < records.size(); i++) {
1848 struct PartiallyConstructedSafepointRecord &info = records[i];
1849 // We can't simply save the live set from the original insertion. One of
1850 // the live values might be the result of a call which needs a safepoint.
1851 // That Value* no longer exists and we need to use the new gc_result.
1852 // Thankfully, the liveset is embedded in the statepoint (and updated), so
1853 // we just grab that.
1854 Statepoint statepoint(info.StatepointToken);
1855 live.insert(live.end(), statepoint.gc_args_begin(),
1856 statepoint.gc_args_end());
1858 // Do some basic sanity checks on our liveness results before performing
1859 // relocation. Relocation can and will turn mistakes in liveness results
1860 // into non-sensical code which is must harder to debug.
1861 // TODO: It would be nice to test consistency as well
1862 assert(DT.isReachableFromEntry(info.StatepointToken->getParent()) &&
1863 "statepoint must be reachable or liveness is meaningless");
1864 for (Value *V : statepoint.gc_args()) {
1865 if (!isa<Instruction>(V))
1866 // Non-instruction values trivial dominate all possible uses
1868 auto LiveInst = cast<Instruction>(V);
1869 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
1870 "unreachable values should never be live");
1871 assert(DT.dominates(LiveInst, info.StatepointToken) &&
1872 "basic SSA liveness expectation violated by liveness analysis");
1876 unique_unsorted(live);
1880 for (auto ptr : live) {
1881 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
1885 relocationViaAlloca(F, DT, live, records);
1886 return !records.empty();
1889 /// Returns true if this function should be rewritten by this pass. The main
1890 /// point of this function is as an extension point for custom logic.
1891 static bool shouldRewriteStatepointsIn(Function &F) {
1892 // TODO: This should check the GCStrategy
1894 const std::string StatepointExampleName("statepoint-example");
1895 return StatepointExampleName == F.getGC();
1900 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
1901 // Nothing to do for declarations.
1902 if (F.isDeclaration() || F.empty())
1905 // Policy choice says not to rewrite - the most common reason is that we're
1906 // compiling code without a GCStrategy.
1907 if (!shouldRewriteStatepointsIn(F))
1910 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1912 // Gather all the statepoints which need rewritten. Be careful to only
1913 // consider those in reachable code since we need to ask dominance queries
1914 // when rewriting. We'll delete the unreachable ones in a moment.
1915 SmallVector<CallSite, 64> ParsePointNeeded;
1916 bool HasUnreachableStatepoint = false;
1917 for (Instruction &I : inst_range(F)) {
1918 // TODO: only the ones with the flag set!
1919 if (isStatepoint(I)) {
1920 if (DT.isReachableFromEntry(I.getParent()))
1921 ParsePointNeeded.push_back(CallSite(&I));
1923 HasUnreachableStatepoint = true;
1927 bool MadeChange = false;
1929 // Delete any unreachable statepoints so that we don't have unrewritten
1930 // statepoints surviving this pass. This makes testing easier and the
1931 // resulting IR less confusing to human readers. Rather than be fancy, we
1932 // just reuse a utility function which removes the unreachable blocks.
1933 if (HasUnreachableStatepoint)
1934 MadeChange |= removeUnreachableBlocks(F);
1936 // Return early if no work to do.
1937 if (ParsePointNeeded.empty())
1940 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
1941 // These are created by LCSSA. They have the effect of increasing the size
1942 // of liveness sets for no good reason. It may be harder to do this post
1943 // insertion since relocations and base phis can confuse things.
1944 for (BasicBlock &BB : F)
1945 if (BB.getUniquePredecessor()) {
1947 FoldSingleEntryPHINodes(&BB);
1950 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
1954 // liveness computation via standard dataflow
1955 // -------------------------------------------------------------------
1957 // TODO: Consider using bitvectors for liveness, the set of potentially
1958 // interesting values should be small and easy to pre-compute.
1960 /// Compute the live-in set for the location rbegin starting from
1961 /// the live-out set of the basic block
1962 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
1963 BasicBlock::reverse_iterator rend,
1964 DenseSet<Value *> &LiveTmp) {
1966 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
1967 Instruction *I = &*ritr;
1969 // KILL/Def - Remove this definition from LiveIn
1972 // Don't consider *uses* in PHI nodes, we handle their contribution to
1973 // predecessor blocks when we seed the LiveOut sets
1974 if (isa<PHINode>(I))
1977 // USE - Add to the LiveIn set for this instruction
1978 for (Value *V : I->operands()) {
1979 assert(!isUnhandledGCPointerType(V->getType()) &&
1980 "support for FCA unimplemented");
1981 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
1982 // The choice to exclude all things constant here is slightly subtle.
1983 // There are two idependent reasons:
1984 // - We assume that things which are constant (from LLVM's definition)
1985 // do not move at runtime. For example, the address of a global
1986 // variable is fixed, even though it's contents may not be.
1987 // - Second, we can't disallow arbitrary inttoptr constants even
1988 // if the language frontend does. Optimization passes are free to
1989 // locally exploit facts without respect to global reachability. This
1990 // can create sections of code which are dynamically unreachable and
1991 // contain just about anything. (see constants.ll in tests)
1998 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2000 for (BasicBlock *Succ : successors(BB)) {
2001 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2002 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2003 PHINode *Phi = cast<PHINode>(&*I);
2004 Value *V = Phi->getIncomingValueForBlock(BB);
2005 assert(!isUnhandledGCPointerType(V->getType()) &&
2006 "support for FCA unimplemented");
2007 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2014 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2015 DenseSet<Value *> KillSet;
2016 for (Instruction &I : *BB)
2017 if (isHandledGCPointerType(I.getType()))
2023 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2024 /// sanity check for the liveness computation.
2025 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2026 TerminatorInst *TI, bool TermOkay = false) {
2027 for (Value *V : Live) {
2028 if (auto *I = dyn_cast<Instruction>(V)) {
2029 // The terminator can be a member of the LiveOut set. LLVM's definition
2030 // of instruction dominance states that V does not dominate itself. As
2031 // such, we need to special case this to allow it.
2032 if (TermOkay && TI == I)
2034 assert(DT.dominates(I, TI) &&
2035 "basic SSA liveness expectation violated by liveness analysis");
2040 /// Check that all the liveness sets used during the computation of liveness
2041 /// obey basic SSA properties. This is useful for finding cases where we miss
2043 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2045 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2046 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2047 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2051 static void computeLiveInValues(DominatorTree &DT, Function &F,
2052 GCPtrLivenessData &Data) {
2054 SmallSetVector<BasicBlock *, 200> Worklist;
2055 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2056 // We use a SetVector so that we don't have duplicates in the worklist.
2057 Worklist.insert(pred_begin(BB), pred_end(BB));
2059 auto NextItem = [&]() {
2060 BasicBlock *BB = Worklist.back();
2061 Worklist.pop_back();
2065 // Seed the liveness for each individual block
2066 for (BasicBlock &BB : F) {
2067 Data.KillSet[&BB] = computeKillSet(&BB);
2068 Data.LiveSet[&BB].clear();
2069 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2072 for (Value *Kill : Data.KillSet[&BB])
2073 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2076 Data.LiveOut[&BB] = DenseSet<Value *>();
2077 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2078 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2079 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2080 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2081 if (!Data.LiveIn[&BB].empty())
2082 AddPredsToWorklist(&BB);
2085 // Propagate that liveness until stable
2086 while (!Worklist.empty()) {
2087 BasicBlock *BB = NextItem();
2089 // Compute our new liveout set, then exit early if it hasn't changed
2090 // despite the contribution of our successor.
2091 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2092 const auto OldLiveOutSize = LiveOut.size();
2093 for (BasicBlock *Succ : successors(BB)) {
2094 assert(Data.LiveIn.count(Succ));
2095 set_union(LiveOut, Data.LiveIn[Succ]);
2097 // assert OutLiveOut is a subset of LiveOut
2098 if (OldLiveOutSize == LiveOut.size()) {
2099 // If the sets are the same size, then we didn't actually add anything
2100 // when unioning our successors LiveIn Thus, the LiveIn of this block
2104 Data.LiveOut[BB] = LiveOut;
2106 // Apply the effects of this basic block
2107 DenseSet<Value *> LiveTmp = LiveOut;
2108 set_union(LiveTmp, Data.LiveSet[BB]);
2109 set_subtract(LiveTmp, Data.KillSet[BB]);
2111 assert(Data.LiveIn.count(BB));
2112 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2113 // assert: OldLiveIn is a subset of LiveTmp
2114 if (OldLiveIn.size() != LiveTmp.size()) {
2115 Data.LiveIn[BB] = LiveTmp;
2116 AddPredsToWorklist(BB);
2118 } // while( !worklist.empty() )
2121 // Sanity check our ouput against SSA properties. This helps catch any
2122 // missing kills during the above iteration.
2123 for (BasicBlock &BB : F) {
2124 checkBasicSSA(DT, Data, BB);
2129 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2130 StatepointLiveSetTy &Out) {
2132 BasicBlock *BB = Inst->getParent();
2134 // Note: The copy is intentional and required
2135 assert(Data.LiveOut.count(BB));
2136 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2138 // We want to handle the statepoint itself oddly. It's
2139 // call result is not live (normal), nor are it's arguments
2140 // (unless they're used again later). This adjustment is
2141 // specifically what we need to relocate
2142 BasicBlock::reverse_iterator rend(Inst);
2143 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2144 LiveOut.erase(Inst);
2145 Out.insert(LiveOut.begin(), LiveOut.end());
2148 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2150 PartiallyConstructedSafepointRecord &Info) {
2151 Instruction *Inst = CS.getInstruction();
2152 StatepointLiveSetTy Updated;
2153 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2156 DenseSet<Value *> Bases;
2157 for (auto KVPair : Info.PointerToBase) {
2158 Bases.insert(KVPair.second);
2161 // We may have base pointers which are now live that weren't before. We need
2162 // to update the PointerToBase structure to reflect this.
2163 for (auto V : Updated)
2164 if (!Info.PointerToBase.count(V)) {
2165 assert(Bases.count(V) && "can't find base for unexpected live value");
2166 Info.PointerToBase[V] = V;
2171 for (auto V : Updated) {
2172 assert(Info.PointerToBase.count(V) &&
2173 "must be able to find base for live value");
2177 // Remove any stale base mappings - this can happen since our liveness is
2178 // more precise then the one inherent in the base pointer analysis
2179 DenseSet<Value *> ToErase;
2180 for (auto KVPair : Info.PointerToBase)
2181 if (!Updated.count(KVPair.first))
2182 ToErase.insert(KVPair.first);
2183 for (auto V : ToErase)
2184 Info.PointerToBase.erase(V);
2187 for (auto KVPair : Info.PointerToBase)
2188 assert(Updated.count(KVPair.first) && "record for non-live value");
2191 Info.liveset = Updated;