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/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/TargetTransformInfo.h"
19 #include "llvm/ADT/SetOperations.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/ADT/DenseSet.h"
22 #include "llvm/ADT/SetVector.h"
23 #include "llvm/ADT/StringRef.h"
24 #include "llvm/IR/BasicBlock.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/InstIterator.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/Intrinsics.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/Statepoint.h"
36 #include "llvm/IR/Value.h"
37 #include "llvm/IR/Verifier.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Transforms/Scalar.h"
41 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
42 #include "llvm/Transforms/Utils/Cloning.h"
43 #include "llvm/Transforms/Utils/Local.h"
44 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
46 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
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,
59 // Cost threshold measuring when it is profitable to rematerialize value instead
61 static cl::opt<unsigned>
62 RematerializationThreshold("spp-rematerialization-threshold", cl::Hidden,
66 static bool ClobberNonLive = true;
68 static bool ClobberNonLive = false;
70 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
71 cl::location(ClobberNonLive),
75 struct RewriteStatepointsForGC : public ModulePass {
76 static char ID; // Pass identification, replacement for typeid
78 RewriteStatepointsForGC() : ModulePass(ID) {
79 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
81 bool runOnFunction(Function &F);
82 bool runOnModule(Module &M) override {
85 Changed |= runOnFunction(F);
88 // stripDereferenceabilityInfo asserts that shouldRewriteStatepointsIn
89 // returns true for at least one function in the module. Since at least
90 // one function changed, we know that the precondition is satisfied.
91 stripDereferenceabilityInfo(M);
97 void getAnalysisUsage(AnalysisUsage &AU) const override {
98 // We add and rewrite a bunch of instructions, but don't really do much
99 // else. We could in theory preserve a lot more analyses here.
100 AU.addRequired<DominatorTreeWrapperPass>();
101 AU.addRequired<TargetTransformInfoWrapperPass>();
104 /// The IR fed into RewriteStatepointsForGC may have had attributes implying
105 /// dereferenceability that are no longer valid/correct after
106 /// RewriteStatepointsForGC has run. This is because semantically, after
107 /// RewriteStatepointsForGC runs, all calls to gc.statepoint "free" the entire
108 /// heap. stripDereferenceabilityInfo (conservatively) restores correctness
109 /// by erasing all attributes in the module that externally imply
110 /// dereferenceability.
112 void stripDereferenceabilityInfo(Module &M);
114 // Helpers for stripDereferenceabilityInfo
115 void stripDereferenceabilityInfoFromBody(Function &F);
116 void stripDereferenceabilityInfoFromPrototype(Function &F);
120 char RewriteStatepointsForGC::ID = 0;
122 ModulePass *llvm::createRewriteStatepointsForGCPass() {
123 return new RewriteStatepointsForGC();
126 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
127 "Make relocations explicit at statepoints", false, false)
128 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
129 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
130 "Make relocations explicit at statepoints", false, false)
133 struct GCPtrLivenessData {
134 /// Values defined in this block.
135 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet;
136 /// Values used in this block (and thus live); does not included values
137 /// killed within this block.
138 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet;
140 /// Values live into this basic block (i.e. used by any
141 /// instruction in this basic block or ones reachable from here)
142 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn;
144 /// Values live out of this basic block (i.e. live into
145 /// any successor block)
146 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut;
149 // The type of the internal cache used inside the findBasePointers family
150 // of functions. From the callers perspective, this is an opaque type and
151 // should not be inspected.
153 // In the actual implementation this caches two relations:
154 // - The base relation itself (i.e. this pointer is based on that one)
155 // - The base defining value relation (i.e. before base_phi insertion)
156 // Generally, after the execution of a full findBasePointer call, only the
157 // base relation will remain. Internally, we add a mixture of the two
158 // types, then update all the second type to the first type
159 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
160 typedef DenseSet<llvm::Value *> StatepointLiveSetTy;
161 typedef DenseMap<Instruction *, Value *> RematerializedValueMapTy;
163 struct PartiallyConstructedSafepointRecord {
164 /// The set of values known to be live across this safepoint
165 StatepointLiveSetTy liveset;
167 /// Mapping from live pointers to a base-defining-value
168 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
170 /// The *new* gc.statepoint instruction itself. This produces the token
171 /// that normal path gc.relocates and the gc.result are tied to.
172 Instruction *StatepointToken;
174 /// Instruction to which exceptional gc relocates are attached
175 /// Makes it easier to iterate through them during relocationViaAlloca.
176 Instruction *UnwindToken;
178 /// Record live values we are rematerialized instead of relocating.
179 /// They are not included into 'liveset' field.
180 /// Maps rematerialized copy to it's original value.
181 RematerializedValueMapTy RematerializedValues;
185 /// Compute the live-in set for every basic block in the function
186 static void computeLiveInValues(DominatorTree &DT, Function &F,
187 GCPtrLivenessData &Data);
189 /// Given results from the dataflow liveness computation, find the set of live
190 /// Values at a particular instruction.
191 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
192 StatepointLiveSetTy &out);
194 // TODO: Once we can get to the GCStrategy, this becomes
195 // Optional<bool> isGCManagedPointer(const Value *V) const override {
197 static bool isGCPointerType(Type *T) {
198 if (auto *PT = dyn_cast<PointerType>(T))
199 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
200 // GC managed heap. We know that a pointer into this heap needs to be
201 // updated and that no other pointer does.
202 return (1 == PT->getAddressSpace());
206 // Return true if this type is one which a) is a gc pointer or contains a GC
207 // pointer and b) is of a type this code expects to encounter as a live value.
208 // (The insertion code will assert that a type which matches (a) and not (b)
209 // is not encountered.)
210 static bool isHandledGCPointerType(Type *T) {
211 // We fully support gc pointers
212 if (isGCPointerType(T))
214 // We partially support vectors of gc pointers. The code will assert if it
215 // can't handle something.
216 if (auto VT = dyn_cast<VectorType>(T))
217 if (isGCPointerType(VT->getElementType()))
223 /// Returns true if this type contains a gc pointer whether we know how to
224 /// handle that type or not.
225 static bool containsGCPtrType(Type *Ty) {
226 if (isGCPointerType(Ty))
228 if (VectorType *VT = dyn_cast<VectorType>(Ty))
229 return isGCPointerType(VT->getScalarType());
230 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
231 return containsGCPtrType(AT->getElementType());
232 if (StructType *ST = dyn_cast<StructType>(Ty))
234 ST->subtypes().begin(), ST->subtypes().end(),
235 [](Type *SubType) { return containsGCPtrType(SubType); });
239 // Returns true if this is a type which a) is a gc pointer or contains a GC
240 // pointer and b) is of a type which the code doesn't expect (i.e. first class
241 // aggregates). Used to trip assertions.
242 static bool isUnhandledGCPointerType(Type *Ty) {
243 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
247 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
248 if (a->hasName() && b->hasName()) {
249 return -1 == a->getName().compare(b->getName());
250 } else if (a->hasName() && !b->hasName()) {
252 } else if (!a->hasName() && b->hasName()) {
255 // Better than nothing, but not stable
260 // Conservatively identifies any definitions which might be live at the
261 // given instruction. The analysis is performed immediately before the
262 // given instruction. Values defined by that instruction are not considered
263 // live. Values used by that instruction are considered live.
264 static void analyzeParsePointLiveness(
265 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
266 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
267 Instruction *inst = CS.getInstruction();
269 StatepointLiveSetTy liveset;
270 findLiveSetAtInst(inst, OriginalLivenessData, liveset);
273 // Note: This output is used by several of the test cases
274 // The order of elements in a set is not stable, put them in a vec and sort
276 SmallVector<Value *, 64> Temp;
277 Temp.insert(Temp.end(), liveset.begin(), liveset.end());
278 std::sort(Temp.begin(), Temp.end(), order_by_name);
279 errs() << "Live Variables:\n";
280 for (Value *V : Temp)
281 dbgs() << " " << V->getName() << " " << *V << "\n";
283 if (PrintLiveSetSize) {
284 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
285 errs() << "Number live values: " << liveset.size() << "\n";
287 result.liveset = liveset;
290 static Value *findBaseDefiningValue(Value *I);
292 /// Return a base defining value for the 'Index' element of the given vector
293 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
294 /// 'I'. As an optimization, this method will try to determine when the
295 /// element is known to already be a base pointer. If this can be established,
296 /// the second value in the returned pair will be true. Note that either a
297 /// vector or a pointer typed value can be returned. For the former, the
298 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
299 /// If the later, the return pointer is a BDV (or possibly a base) for the
300 /// particular element in 'I'.
301 static std::pair<Value *, bool>
302 findBaseDefiningValueOfVector(Value *I, Value *Index = nullptr) {
303 assert(I->getType()->isVectorTy() &&
304 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
305 "Illegal to ask for the base pointer of a non-pointer type");
307 // Each case parallels findBaseDefiningValue below, see that code for
308 // detailed motivation.
310 if (isa<Argument>(I))
311 // An incoming argument to the function is a base pointer
312 return std::make_pair(I, true);
314 // We shouldn't see the address of a global as a vector value?
315 assert(!isa<GlobalVariable>(I) &&
316 "unexpected global variable found in base of vector");
318 // inlining could possibly introduce phi node that contains
319 // undef if callee has multiple returns
320 if (isa<UndefValue>(I))
321 // utterly meaningless, but useful for dealing with partially optimized
323 return std::make_pair(I, true);
325 // Due to inheritance, this must be _after_ the global variable and undef
327 if (Constant *Con = dyn_cast<Constant>(I)) {
328 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
329 "order of checks wrong!");
330 assert(Con->isNullValue() && "null is the only case which makes sense");
331 return std::make_pair(Con, true);
334 if (isa<LoadInst>(I))
335 return std::make_pair(I, true);
337 // For an insert element, we might be able to look through it if we know
338 // something about the indexes.
339 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(I)) {
341 Value *InsertIndex = IEI->getOperand(2);
342 // This index is inserting the value, look for its BDV
343 if (InsertIndex == Index)
344 return std::make_pair(findBaseDefiningValue(IEI->getOperand(1)), false);
345 // Both constant, and can't be equal per above. This insert is definitely
346 // not relevant, look back at the rest of the vector and keep trying.
347 if (isa<ConstantInt>(Index) && isa<ConstantInt>(InsertIndex))
348 return findBaseDefiningValueOfVector(IEI->getOperand(0), Index);
351 // We don't know whether this vector contains entirely base pointers or
352 // not. To be conservatively correct, we treat it as a BDV and will
353 // duplicate code as needed to construct a parallel vector of bases.
354 return std::make_pair(IEI, false);
357 if (isa<ShuffleVectorInst>(I))
358 // We don't know whether this vector contains entirely base pointers or
359 // not. To be conservatively correct, we treat it as a BDV and will
360 // duplicate code as needed to construct a parallel vector of bases.
361 // TODO: There a number of local optimizations which could be applied here
362 // for particular sufflevector patterns.
363 return std::make_pair(I, false);
365 // A PHI or Select is a base defining value. The outer findBasePointer
366 // algorithm is responsible for constructing a base value for this BDV.
367 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
368 "unknown vector instruction - no base found for vector element");
369 return std::make_pair(I, false);
372 static bool isKnownBaseResult(Value *V);
374 /// Helper function for findBasePointer - Will return a value which either a)
375 /// defines the base pointer for the input, b) blocks the simple search
376 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
377 /// from pointer to vector type or back.
378 static Value *findBaseDefiningValue(Value *I) {
379 if (I->getType()->isVectorTy())
380 return findBaseDefiningValueOfVector(I).first;
382 assert(I->getType()->isPointerTy() &&
383 "Illegal to ask for the base pointer of a non-pointer type");
385 if (isa<Argument>(I))
386 // An incoming argument to the function is a base pointer
387 // We should have never reached here if this argument isn't an gc value
390 if (isa<GlobalVariable>(I))
394 // inlining could possibly introduce phi node that contains
395 // undef if callee has multiple returns
396 if (isa<UndefValue>(I))
397 // utterly meaningless, but useful for dealing with
398 // partially optimized code.
401 // Due to inheritance, this must be _after_ the global variable and undef
403 if (Constant *Con = dyn_cast<Constant>(I)) {
404 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
405 "order of checks wrong!");
406 // Note: Finding a constant base for something marked for relocation
407 // doesn't really make sense. The most likely case is either a) some
408 // screwed up the address space usage or b) your validating against
409 // compiled C++ code w/o the proper separation. The only real exception
410 // is a null pointer. You could have generic code written to index of
411 // off a potentially null value and have proven it null. We also use
412 // null pointers in dead paths of relocation phis (which we might later
413 // want to find a base pointer for).
414 assert(isa<ConstantPointerNull>(Con) &&
415 "null is the only case which makes sense");
419 if (CastInst *CI = dyn_cast<CastInst>(I)) {
420 Value *Def = CI->stripPointerCasts();
421 // If we find a cast instruction here, it means we've found a cast which is
422 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
423 // handle int->ptr conversion.
424 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
425 return findBaseDefiningValue(Def);
428 if (isa<LoadInst>(I))
429 return I; // The value loaded is an gc base itself
431 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
432 // The base of this GEP is the base
433 return findBaseDefiningValue(GEP->getPointerOperand());
435 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
436 switch (II->getIntrinsicID()) {
437 case Intrinsic::experimental_gc_result_ptr:
439 // fall through to general call handling
441 case Intrinsic::experimental_gc_statepoint:
442 case Intrinsic::experimental_gc_result_float:
443 case Intrinsic::experimental_gc_result_int:
444 llvm_unreachable("these don't produce pointers");
445 case Intrinsic::experimental_gc_relocate: {
446 // Rerunning safepoint insertion after safepoints are already
447 // inserted is not supported. It could probably be made to work,
448 // but why are you doing this? There's no good reason.
449 llvm_unreachable("repeat safepoint insertion is not supported");
451 case Intrinsic::gcroot:
452 // Currently, this mechanism hasn't been extended to work with gcroot.
453 // There's no reason it couldn't be, but I haven't thought about the
454 // implications much.
456 "interaction with the gcroot mechanism is not supported");
459 // We assume that functions in the source language only return base
460 // pointers. This should probably be generalized via attributes to support
461 // both source language and internal functions.
462 if (isa<CallInst>(I) || isa<InvokeInst>(I))
465 // I have absolutely no idea how to implement this part yet. It's not
466 // necessarily hard, I just haven't really looked at it yet.
467 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
469 if (isa<AtomicCmpXchgInst>(I))
470 // A CAS is effectively a atomic store and load combined under a
471 // predicate. From the perspective of base pointers, we just treat it
475 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
476 "binary ops which don't apply to pointers");
478 // The aggregate ops. Aggregates can either be in the heap or on the
479 // stack, but in either case, this is simply a field load. As a result,
480 // this is a defining definition of the base just like a load is.
481 if (isa<ExtractValueInst>(I))
484 // We should never see an insert vector since that would require we be
485 // tracing back a struct value not a pointer value.
486 assert(!isa<InsertValueInst>(I) &&
487 "Base pointer for a struct is meaningless");
489 // An extractelement produces a base result exactly when it's input does.
490 // We may need to insert a parallel instruction to extract the appropriate
491 // element out of the base vector corresponding to the input. Given this,
492 // it's analogous to the phi and select case even though it's not a merge.
493 if (auto *EEI = dyn_cast<ExtractElementInst>(I)) {
494 Value *VectorOperand = EEI->getVectorOperand();
495 Value *Index = EEI->getIndexOperand();
496 std::pair<Value *, bool> pair =
497 findBaseDefiningValueOfVector(VectorOperand, Index);
498 Value *VectorBase = pair.first;
499 if (VectorBase->getType()->isPointerTy())
500 // We found a BDV for this specific element with the vector. This is an
501 // optimization, but in practice it covers most of the useful cases
502 // created via scalarization. Note: The peephole optimization here is
503 // currently needed for correctness since the general algorithm doesn't
504 // yet handle insertelements. That will change shortly.
507 assert(VectorBase->getType()->isVectorTy());
508 // Otherwise, we have an instruction which potentially produces a
509 // derived pointer and we need findBasePointers to clone code for us
510 // such that we can create an instruction which produces the
511 // accompanying base pointer.
516 // The last two cases here don't return a base pointer. Instead, they
517 // return a value which dynamically selects from among several base
518 // derived pointers (each with it's own base potentially). It's the job of
519 // the caller to resolve these.
520 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
521 "missing instruction case in findBaseDefiningValing");
525 /// Returns the base defining value for this value.
526 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
527 Value *&Cached = Cache[I];
529 Cached = findBaseDefiningValue(I);
530 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
531 << Cached->getName() << "\n");
533 assert(Cache[I] != nullptr);
537 /// Return a base pointer for this value if known. Otherwise, return it's
538 /// base defining value.
539 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
540 Value *Def = findBaseDefiningValueCached(I, Cache);
541 auto Found = Cache.find(Def);
542 if (Found != Cache.end()) {
543 // Either a base-of relation, or a self reference. Caller must check.
544 return Found->second;
546 // Only a BDV available
550 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
551 /// is it known to be a base pointer? Or do we need to continue searching.
552 static bool isKnownBaseResult(Value *V) {
553 if (!isa<PHINode>(V) && !isa<SelectInst>(V) && !isa<ExtractElementInst>(V)) {
554 // no recursion possible
557 if (isa<Instruction>(V) &&
558 cast<Instruction>(V)->getMetadata("is_base_value")) {
559 // This is a previously inserted base phi or select. We know
560 // that this is a base value.
564 // We need to keep searching
569 /// Models the state of a single base defining value in the findBasePointer
570 /// algorithm for determining where a new instruction is needed to propagate
571 /// the base of this BDV.
574 enum Status { Unknown, Base, Conflict };
576 BDVState(Status s, Value *b = nullptr) : status(s), base(b) {
577 assert(status != Base || b);
579 explicit BDVState(Value *b) : status(Base), base(b) {}
580 BDVState() : status(Unknown), base(nullptr) {}
582 Status getStatus() const { return status; }
583 Value *getBase() const { return base; }
585 bool isBase() const { return getStatus() == Base; }
586 bool isUnknown() const { return getStatus() == Unknown; }
587 bool isConflict() const { return getStatus() == Conflict; }
589 bool operator==(const BDVState &other) const {
590 return base == other.base && status == other.status;
593 bool operator!=(const BDVState &other) const { return !(*this == other); }
596 void dump() const { print(dbgs()); dbgs() << '\n'; }
598 void print(raw_ostream &OS) const {
610 OS << " (" << base << " - "
611 << (base ? base->getName() : "nullptr") << "): ";
616 Value *base; // non null only if status == base
619 inline raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
625 typedef DenseMap<Value *, BDVState> ConflictStateMapTy;
626 // Values of type BDVState form a lattice, and this is a helper
627 // class that implementes the meet operation. The meat of the meet
628 // operation is implemented in MeetBDVStates::pureMeet
629 class MeetBDVStates {
631 /// Initializes the currentResult to the TOP state so that if can be met with
632 /// any other state to produce that state.
635 // Destructively meet the current result with the given BDVState
636 void meetWith(BDVState otherState) {
637 currentResult = meet(otherState, currentResult);
640 BDVState getResult() const { return currentResult; }
643 BDVState currentResult;
645 /// Perform a meet operation on two elements of the BDVState lattice.
646 static BDVState meet(BDVState LHS, BDVState RHS) {
647 assert((pureMeet(LHS, RHS) == pureMeet(RHS, LHS)) &&
648 "math is wrong: meet does not commute!");
649 BDVState Result = pureMeet(LHS, RHS);
650 DEBUG(dbgs() << "meet of " << LHS << " with " << RHS
651 << " produced " << Result << "\n");
655 static BDVState pureMeet(const BDVState &stateA, const BDVState &stateB) {
656 switch (stateA.getStatus()) {
657 case BDVState::Unknown:
661 assert(stateA.getBase() && "can't be null");
662 if (stateB.isUnknown())
665 if (stateB.isBase()) {
666 if (stateA.getBase() == stateB.getBase()) {
667 assert(stateA == stateB && "equality broken!");
670 return BDVState(BDVState::Conflict);
672 assert(stateB.isConflict() && "only three states!");
673 return BDVState(BDVState::Conflict);
675 case BDVState::Conflict:
678 llvm_unreachable("only three states!");
682 /// For a given value or instruction, figure out what base ptr it's derived
683 /// from. For gc objects, this is simply itself. On success, returns a value
684 /// which is the base pointer. (This is reliable and can be used for
685 /// relocation.) On failure, returns nullptr.
686 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
687 Value *def = findBaseOrBDV(I, cache);
689 if (isKnownBaseResult(def)) {
693 // Here's the rough algorithm:
694 // - For every SSA value, construct a mapping to either an actual base
695 // pointer or a PHI which obscures the base pointer.
696 // - Construct a mapping from PHI to unknown TOP state. Use an
697 // optimistic algorithm to propagate base pointer information. Lattice
702 // When algorithm terminates, all PHIs will either have a single concrete
703 // base or be in a conflict state.
704 // - For every conflict, insert a dummy PHI node without arguments. Add
705 // these to the base[Instruction] = BasePtr mapping. For every
706 // non-conflict, add the actual base.
707 // - For every conflict, add arguments for the base[a] of each input
710 // Note: A simpler form of this would be to add the conflict form of all
711 // PHIs without running the optimistic algorithm. This would be
712 // analogous to pessimistic data flow and would likely lead to an
713 // overall worse solution.
716 auto isExpectedBDVType = [](Value *BDV) {
717 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) || isa<ExtractElementInst>(BDV);
721 // Once populated, will contain a mapping from each potentially non-base BDV
722 // to a lattice value (described above) which corresponds to that BDV.
723 ConflictStateMapTy states;
724 // Recursively fill in all phis & selects reachable from the initial one
725 // for which we don't already know a definite base value for
727 DenseSet<Value *> Visited;
728 SmallVector<Value*, 16> Worklist;
729 Worklist.push_back(def);
731 while (!Worklist.empty()) {
732 Value *Current = Worklist.pop_back_val();
733 assert(!isKnownBaseResult(Current) && "why did it get added?");
735 auto visitIncomingValue = [&](Value *InVal) {
736 Value *Base = findBaseOrBDV(InVal, cache);
737 if (isKnownBaseResult(Base))
738 // Known bases won't need new instructions introduced and can be
741 assert(isExpectedBDVType(Base) && "the only non-base values "
742 "we see should be base defining values");
743 if (Visited.insert(Base).second)
744 Worklist.push_back(Base);
746 if (PHINode *Phi = dyn_cast<PHINode>(Current)) {
747 for (Value *InVal : Phi->incoming_values())
748 visitIncomingValue(InVal);
749 } else if (SelectInst *Sel = dyn_cast<SelectInst>(Current)) {
750 visitIncomingValue(Sel->getTrueValue());
751 visitIncomingValue(Sel->getFalseValue());
752 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
753 visitIncomingValue(EE->getVectorOperand());
755 // There are two classes of instructions we know we don't handle.
756 assert(isa<ShuffleVectorInst>(Current) ||
757 isa<InsertElementInst>(Current));
758 llvm_unreachable("unimplemented instruction case");
761 // The frontier of visited instructions are the ones we might need to
762 // duplicate, so fill in the starting state for the optimistic algorithm
764 for (Value *BDV : Visited) {
765 states[BDV] = BDVState();
770 DEBUG(dbgs() << "States after initialization:\n");
771 for (auto Pair : states) {
772 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
776 // TODO: come back and revisit the state transitions around inputs which
777 // have reached conflict state. The current version seems too conservative.
779 // Return a phi state for a base defining value. We'll generate a new
780 // base state for known bases and expect to find a cached state otherwise.
781 auto getStateForBDV = [&](Value *baseValue) {
782 if (isKnownBaseResult(baseValue))
783 return BDVState(baseValue);
784 auto I = states.find(baseValue);
785 assert(I != states.end() && "lookup failed!");
789 bool progress = true;
792 size_t oldSize = states.size();
795 // We're only changing keys in this loop, thus safe to keep iterators
796 for (auto Pair : states) {
797 Value *v = Pair.first;
798 assert(!isKnownBaseResult(v) && "why did it get added?");
800 // Given an input value for the current instruction, return a BDVState
801 // instance which represents the BDV of that value.
802 auto getStateForInput = [&](Value *V) mutable {
803 Value *BDV = findBaseOrBDV(V, cache);
804 return getStateForBDV(BDV);
807 MeetBDVStates calculateMeet;
808 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
809 calculateMeet.meetWith(getStateForInput(select->getTrueValue()));
810 calculateMeet.meetWith(getStateForInput(select->getFalseValue()));
811 } else if (PHINode *Phi = dyn_cast<PHINode>(v)) {
812 for (Value *Val : Phi->incoming_values())
813 calculateMeet.meetWith(getStateForInput(Val));
815 // The 'meet' for an extractelement is slightly trivial, but it's still
816 // useful in that it drives us to conflict if our input is.
817 auto *EE = cast<ExtractElementInst>(v);
818 calculateMeet.meetWith(getStateForInput(EE->getVectorOperand()));
822 BDVState oldState = states[v];
823 BDVState newState = calculateMeet.getResult();
824 if (oldState != newState) {
826 states[v] = newState;
830 assert(oldSize <= states.size());
831 assert(oldSize == states.size() || progress);
835 DEBUG(dbgs() << "States after meet iteration:\n");
836 for (auto Pair : states) {
837 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
841 // Insert Phis for all conflicts
842 // We want to keep naming deterministic in the loop that follows, so
843 // sort the keys before iteration. This is useful in allowing us to
844 // write stable tests. Note that there is no invalidation issue here.
845 SmallVector<Value *, 16> Keys;
846 Keys.reserve(states.size());
847 for (auto Pair : states) {
848 Value *V = Pair.first;
851 std::sort(Keys.begin(), Keys.end(), order_by_name);
852 // TODO: adjust naming patterns to avoid this order of iteration dependency
853 for (Value *V : Keys) {
854 Instruction *I = cast<Instruction>(V);
855 BDVState State = states[I];
856 assert(!isKnownBaseResult(I) && "why did it get added?");
857 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
859 // extractelement instructions are a bit special in that we may need to
860 // insert an extract even when we know an exact base for the instruction.
861 // The problem is that we need to convert from a vector base to a scalar
862 // base for the particular indice we're interested in.
863 if (State.isBase() && isa<ExtractElementInst>(I) &&
864 isa<VectorType>(State.getBase()->getType())) {
865 auto *EE = cast<ExtractElementInst>(I);
866 // TODO: In many cases, the new instruction is just EE itself. We should
867 // exploit this, but can't do it here since it would break the invariant
868 // about the BDV not being known to be a base.
869 auto *BaseInst = ExtractElementInst::Create(State.getBase(),
870 EE->getIndexOperand(),
872 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
873 states[I] = BDVState(BDVState::Base, BaseInst);
876 if (!State.isConflict())
879 /// Create and insert a new instruction which will represent the base of
880 /// the given instruction 'I'.
881 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
882 if (isa<PHINode>(I)) {
883 BasicBlock *BB = I->getParent();
884 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
885 assert(NumPreds > 0 && "how did we reach here");
886 std::string Name = I->hasName() ?
887 (I->getName() + ".base").str() : "base_phi";
888 return PHINode::Create(I->getType(), NumPreds, Name, I);
889 } else if (SelectInst *Sel = dyn_cast<SelectInst>(I)) {
890 // The undef will be replaced later
891 UndefValue *Undef = UndefValue::get(Sel->getType());
892 std::string Name = I->hasName() ?
893 (I->getName() + ".base").str() : "base_select";
894 return SelectInst::Create(Sel->getCondition(), Undef,
897 auto *EE = cast<ExtractElementInst>(I);
898 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
899 std::string Name = I->hasName() ?
900 (I->getName() + ".base").str() : "base_ee";
901 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
905 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
906 // Add metadata marking this as a base value
907 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
908 states[I] = BDVState(BDVState::Conflict, BaseInst);
911 // Fixup all the inputs of the new PHIs
912 for (auto Pair : states) {
913 Instruction *v = cast<Instruction>(Pair.first);
914 BDVState state = Pair.second;
916 assert(!isKnownBaseResult(v) && "why did it get added?");
917 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
918 if (!state.isConflict())
921 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
922 PHINode *phi = cast<PHINode>(v);
923 unsigned NumPHIValues = phi->getNumIncomingValues();
924 for (unsigned i = 0; i < NumPHIValues; i++) {
925 Value *InVal = phi->getIncomingValue(i);
926 BasicBlock *InBB = phi->getIncomingBlock(i);
928 // If we've already seen InBB, add the same incoming value
929 // we added for it earlier. The IR verifier requires phi
930 // nodes with multiple entries from the same basic block
931 // to have the same incoming value for each of those
932 // entries. If we don't do this check here and basephi
933 // has a different type than base, we'll end up adding two
934 // bitcasts (and hence two distinct values) as incoming
935 // values for the same basic block.
937 int blockIndex = basephi->getBasicBlockIndex(InBB);
938 if (blockIndex != -1) {
939 Value *oldBase = basephi->getIncomingValue(blockIndex);
940 basephi->addIncoming(oldBase, InBB);
942 Value *base = findBaseOrBDV(InVal, cache);
943 if (!isKnownBaseResult(base)) {
944 // Either conflict or base.
945 assert(states.count(base));
946 base = states[base].getBase();
947 assert(base != nullptr && "unknown BDVState!");
950 // In essence this assert states: the only way two
951 // values incoming from the same basic block may be
952 // different is by being different bitcasts of the same
953 // value. A cleanup that remains TODO is changing
954 // findBaseOrBDV to return an llvm::Value of the correct
955 // type (and still remain pure). This will remove the
956 // need to add bitcasts.
957 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
958 "sanity -- findBaseOrBDV should be pure!");
963 // Find either the defining value for the PHI or the normal base for
965 Value *base = findBaseOrBDV(InVal, cache);
966 if (!isKnownBaseResult(base)) {
967 // Either conflict or base.
968 assert(states.count(base));
969 base = states[base].getBase();
970 assert(base != nullptr && "unknown BDVState!");
972 assert(base && "can't be null");
973 // Must use original input BB since base may not be Instruction
974 // The cast is needed since base traversal may strip away bitcasts
975 if (base->getType() != basephi->getType()) {
976 base = new BitCastInst(base, basephi->getType(), "cast",
977 InBB->getTerminator());
979 basephi->addIncoming(base, InBB);
981 assert(basephi->getNumIncomingValues() == NumPHIValues);
982 } else if (SelectInst *basesel = dyn_cast<SelectInst>(state.getBase())) {
983 SelectInst *sel = cast<SelectInst>(v);
984 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
985 // something more safe and less hacky.
986 for (int i = 1; i <= 2; i++) {
987 Value *InVal = sel->getOperand(i);
988 // Find either the defining value for the PHI or the normal base for
990 Value *base = findBaseOrBDV(InVal, cache);
991 if (!isKnownBaseResult(base)) {
992 // Either conflict or base.
993 assert(states.count(base));
994 base = states[base].getBase();
995 assert(base != nullptr && "unknown BDVState!");
997 assert(base && "can't be null");
998 // Must use original input BB since base may not be Instruction
999 // The cast is needed since base traversal may strip away bitcasts
1000 if (base->getType() != basesel->getType()) {
1001 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
1003 basesel->setOperand(i, base);
1006 auto *BaseEE = cast<ExtractElementInst>(state.getBase());
1007 Value *InVal = cast<ExtractElementInst>(v)->getVectorOperand();
1008 Value *Base = findBaseOrBDV(InVal, cache);
1009 if (!isKnownBaseResult(Base)) {
1010 // Either conflict or base.
1011 assert(states.count(Base));
1012 Base = states[Base].getBase();
1013 assert(Base != nullptr && "unknown BDVState!");
1015 assert(Base && "can't be null");
1016 BaseEE->setOperand(0, Base);
1020 // Now that we're done with the algorithm, see if we can optimize the
1021 // results slightly by reducing the number of new instructions needed.
1022 // Arguably, this should be integrated into the algorithm above, but
1023 // doing as a post process step is easier to reason about for the moment.
1024 DenseMap<Value *, Value *> ReverseMap;
1025 SmallPtrSet<Instruction *, 16> NewInsts;
1026 SmallSetVector<Instruction *, 16> Worklist;
1027 for (auto Item : states) {
1028 Value *V = Item.first;
1029 Value *Base = Item.second.getBase();
1031 assert(!isKnownBaseResult(V) && "why did it get added?");
1032 assert(isKnownBaseResult(Base) &&
1033 "must be something we 'know' is a base pointer");
1034 if (!Item.second.isConflict())
1037 ReverseMap[Base] = V;
1038 if (auto *BaseI = dyn_cast<Instruction>(Base)) {
1039 NewInsts.insert(BaseI);
1040 Worklist.insert(BaseI);
1043 auto PushNewUsers = [&](Instruction *I) {
1044 for (User *U : I->users())
1045 if (auto *UI = dyn_cast<Instruction>(U))
1046 if (NewInsts.count(UI))
1047 Worklist.insert(UI);
1049 const DataLayout &DL = cast<Instruction>(def)->getModule()->getDataLayout();
1050 while (!Worklist.empty()) {
1051 Instruction *BaseI = Worklist.pop_back_val();
1052 assert(NewInsts.count(BaseI));
1053 Value *Bdv = ReverseMap[BaseI];
1054 if (auto *BdvI = dyn_cast<Instruction>(Bdv))
1055 if (BaseI->isIdenticalTo(BdvI)) {
1056 DEBUG(dbgs() << "Identical Base: " << *BaseI << "\n");
1057 PushNewUsers(BaseI);
1058 BaseI->replaceAllUsesWith(Bdv);
1059 BaseI->eraseFromParent();
1060 states[Bdv] = BDVState(BDVState::Conflict, Bdv);
1061 NewInsts.erase(BaseI);
1062 ReverseMap.erase(BaseI);
1065 if (Value *V = SimplifyInstruction(BaseI, DL)) {
1066 DEBUG(dbgs() << "Base " << *BaseI << " simplified to " << *V << "\n");
1067 PushNewUsers(BaseI);
1068 BaseI->replaceAllUsesWith(V);
1069 BaseI->eraseFromParent();
1070 states[Bdv] = BDVState(BDVState::Conflict, V);
1071 NewInsts.erase(BaseI);
1072 ReverseMap.erase(BaseI);
1077 // Cache all of our results so we can cheaply reuse them
1078 // NOTE: This is actually two caches: one of the base defining value
1079 // relation and one of the base pointer relation! FIXME
1080 for (auto item : states) {
1081 Value *v = item.first;
1082 Value *base = item.second.getBase();
1085 std::string fromstr =
1086 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
1088 DEBUG(dbgs() << "Updating base value cache"
1089 << " for: " << (v->hasName() ? v->getName() : "")
1090 << " from: " << fromstr
1091 << " to: " << (base->hasName() ? base->getName() : "") << "\n");
1093 if (cache.count(v)) {
1094 // Once we transition from the BDV relation being store in the cache to
1095 // the base relation being stored, it must be stable
1096 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
1097 "base relation should be stable");
1101 assert(cache.find(def) != cache.end());
1105 // For a set of live pointers (base and/or derived), identify the base
1106 // pointer of the object which they are derived from. This routine will
1107 // mutate the IR graph as needed to make the 'base' pointer live at the
1108 // definition site of 'derived'. This ensures that any use of 'derived' can
1109 // also use 'base'. This may involve the insertion of a number of
1110 // additional PHI nodes.
1112 // preconditions: live is a set of pointer type Values
1114 // side effects: may insert PHI nodes into the existing CFG, will preserve
1115 // CFG, will not remove or mutate any existing nodes
1117 // post condition: PointerToBase contains one (derived, base) pair for every
1118 // pointer in live. Note that derived can be equal to base if the original
1119 // pointer was a base pointer.
1121 findBasePointers(const StatepointLiveSetTy &live,
1122 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
1123 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1124 // For the naming of values inserted to be deterministic - which makes for
1125 // much cleaner and more stable tests - we need to assign an order to the
1126 // live values. DenseSets do not provide a deterministic order across runs.
1127 SmallVector<Value *, 64> Temp;
1128 Temp.insert(Temp.end(), live.begin(), live.end());
1129 std::sort(Temp.begin(), Temp.end(), order_by_name);
1130 for (Value *ptr : Temp) {
1131 Value *base = findBasePointer(ptr, DVCache);
1132 assert(base && "failed to find base pointer");
1133 PointerToBase[ptr] = base;
1134 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1135 DT->dominates(cast<Instruction>(base)->getParent(),
1136 cast<Instruction>(ptr)->getParent())) &&
1137 "The base we found better dominate the derived pointer");
1139 // If you see this trip and like to live really dangerously, the code should
1140 // be correct, just with idioms the verifier can't handle. You can try
1141 // disabling the verifier at your own substantial risk.
1142 assert(!isa<ConstantPointerNull>(base) &&
1143 "the relocation code needs adjustment to handle the relocation of "
1144 "a null pointer constant without causing false positives in the "
1145 "safepoint ir verifier.");
1149 /// Find the required based pointers (and adjust the live set) for the given
1151 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1153 PartiallyConstructedSafepointRecord &result) {
1154 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
1155 findBasePointers(result.liveset, PointerToBase, &DT, DVCache);
1157 if (PrintBasePointers) {
1158 // Note: Need to print these in a stable order since this is checked in
1160 errs() << "Base Pairs (w/o Relocation):\n";
1161 SmallVector<Value *, 64> Temp;
1162 Temp.reserve(PointerToBase.size());
1163 for (auto Pair : PointerToBase) {
1164 Temp.push_back(Pair.first);
1166 std::sort(Temp.begin(), Temp.end(), order_by_name);
1167 for (Value *Ptr : Temp) {
1168 Value *Base = PointerToBase[Ptr];
1169 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
1174 result.PointerToBase = PointerToBase;
1177 /// Given an updated version of the dataflow liveness results, update the
1178 /// liveset and base pointer maps for the call site CS.
1179 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1181 PartiallyConstructedSafepointRecord &result);
1183 static void recomputeLiveInValues(
1184 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1185 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1186 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1187 // again. The old values are still live and will help it stabilize quickly.
1188 GCPtrLivenessData RevisedLivenessData;
1189 computeLiveInValues(DT, F, RevisedLivenessData);
1190 for (size_t i = 0; i < records.size(); i++) {
1191 struct PartiallyConstructedSafepointRecord &info = records[i];
1192 const CallSite &CS = toUpdate[i];
1193 recomputeLiveInValues(RevisedLivenessData, CS, info);
1197 // When inserting gc.relocate calls, we need to ensure there are no uses
1198 // of the original value between the gc.statepoint and the gc.relocate call.
1199 // One case which can arise is a phi node starting one of the successor blocks.
1200 // We also need to be able to insert the gc.relocates only on the path which
1201 // goes through the statepoint. We might need to split an edge to make this
1204 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1205 DominatorTree &DT) {
1206 BasicBlock *Ret = BB;
1207 if (!BB->getUniquePredecessor()) {
1208 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1211 // Now that 'ret' has unique predecessor we can safely remove all phi nodes
1213 FoldSingleEntryPHINodes(Ret);
1214 assert(!isa<PHINode>(Ret->begin()));
1216 // At this point, we can safely insert a gc.relocate as the first instruction
1217 // in Ret if needed.
1221 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1222 auto itr = std::find(livevec.begin(), livevec.end(), val);
1223 assert(livevec.end() != itr);
1224 size_t index = std::distance(livevec.begin(), itr);
1225 assert(index < livevec.size());
1229 // Create new attribute set containing only attributes which can be transferred
1230 // from original call to the safepoint.
1231 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1234 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1235 unsigned index = AS.getSlotIndex(Slot);
1237 if (index == AttributeSet::ReturnIndex ||
1238 index == AttributeSet::FunctionIndex) {
1240 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1242 Attribute attr = *it;
1244 // Do not allow certain attributes - just skip them
1245 // Safepoint can not be read only or read none.
1246 if (attr.hasAttribute(Attribute::ReadNone) ||
1247 attr.hasAttribute(Attribute::ReadOnly))
1250 ret = ret.addAttributes(
1251 AS.getContext(), index,
1252 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1256 // Just skip parameter attributes for now
1262 /// Helper function to place all gc relocates necessary for the given
1265 /// liveVariables - list of variables to be relocated.
1266 /// liveStart - index of the first live variable.
1267 /// basePtrs - base pointers.
1268 /// statepointToken - statepoint instruction to which relocates should be
1270 /// Builder - Llvm IR builder to be used to construct new calls.
1271 static void CreateGCRelocates(ArrayRef<llvm::Value *> LiveVariables,
1272 const int LiveStart,
1273 ArrayRef<llvm::Value *> BasePtrs,
1274 Instruction *StatepointToken,
1275 IRBuilder<> Builder) {
1276 if (LiveVariables.empty())
1279 // All gc_relocate are set to i8 addrspace(1)* type. We originally generated
1280 // unique declarations for each pointer type, but this proved problematic
1281 // because the intrinsic mangling code is incomplete and fragile. Since
1282 // we're moving towards a single unified pointer type anyways, we can just
1283 // cast everything to an i8* of the right address space. A bitcast is added
1284 // later to convert gc_relocate to the actual value's type.
1285 Module *M = StatepointToken->getModule();
1286 auto AS = cast<PointerType>(LiveVariables[0]->getType())->getAddressSpace();
1287 Type *Types[] = {Type::getInt8PtrTy(M->getContext(), AS)};
1288 Value *GCRelocateDecl =
1289 Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, Types);
1291 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1292 // Generate the gc.relocate call and save the result
1294 Builder.getInt32(LiveStart + find_index(LiveVariables, BasePtrs[i]));
1296 Builder.getInt32(LiveStart + find_index(LiveVariables, LiveVariables[i]));
1298 // only specify a debug name if we can give a useful one
1299 CallInst *Reloc = Builder.CreateCall(
1300 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1301 LiveVariables[i]->hasName() ? LiveVariables[i]->getName() + ".relocated"
1303 // Trick CodeGen into thinking there are lots of free registers at this
1305 Reloc->setCallingConv(CallingConv::Cold);
1310 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1311 const SmallVectorImpl<llvm::Value *> &basePtrs,
1312 const SmallVectorImpl<llvm::Value *> &liveVariables,
1314 PartiallyConstructedSafepointRecord &result) {
1315 assert(basePtrs.size() == liveVariables.size());
1316 assert(isStatepoint(CS) &&
1317 "This method expects to be rewriting a statepoint");
1319 BasicBlock *BB = CS.getInstruction()->getParent();
1321 Function *F = BB->getParent();
1322 assert(F && "must be set");
1323 Module *M = F->getParent();
1325 assert(M && "must be set");
1327 // We're not changing the function signature of the statepoint since the gc
1328 // arguments go into the var args section.
1329 Function *gc_statepoint_decl = CS.getCalledFunction();
1331 // Then go ahead and use the builder do actually do the inserts. We insert
1332 // immediately before the previous instruction under the assumption that all
1333 // arguments will be available here. We can't insert afterwards since we may
1334 // be replacing a terminator.
1335 Instruction *insertBefore = CS.getInstruction();
1336 IRBuilder<> Builder(insertBefore);
1337 // Copy all of the arguments from the original statepoint - this includes the
1338 // target, call args, and deopt args
1339 SmallVector<llvm::Value *, 64> args;
1340 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1341 // TODO: Clear the 'needs rewrite' flag
1343 // add all the pointers to be relocated (gc arguments)
1344 // Capture the start of the live variable list for use in the gc_relocates
1345 const int live_start = args.size();
1346 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1348 // Create the statepoint given all the arguments
1349 Instruction *token = nullptr;
1350 AttributeSet return_attributes;
1352 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1354 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1355 call->setTailCall(toReplace->isTailCall());
1356 call->setCallingConv(toReplace->getCallingConv());
1358 // Currently we will fail on parameter attributes and on certain
1359 // function attributes.
1360 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1361 // In case if we can handle this set of attributes - set up function attrs
1362 // directly on statepoint and return attrs later for gc_result intrinsic.
1363 call->setAttributes(new_attrs.getFnAttributes());
1364 return_attributes = new_attrs.getRetAttributes();
1368 // Put the following gc_result and gc_relocate calls immediately after the
1369 // the old call (which we're about to delete)
1370 BasicBlock::iterator next(toReplace);
1371 assert(BB->end() != next && "not a terminator, must have next");
1373 Instruction *IP = &*(next);
1374 Builder.SetInsertPoint(IP);
1375 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1378 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1380 // Insert the new invoke into the old block. We'll remove the old one in a
1381 // moment at which point this will become the new terminator for the
1383 InvokeInst *invoke = InvokeInst::Create(
1384 gc_statepoint_decl, toReplace->getNormalDest(),
1385 toReplace->getUnwindDest(), args, "statepoint_token", toReplace->getParent());
1386 invoke->setCallingConv(toReplace->getCallingConv());
1388 // Currently we will fail on parameter attributes and on certain
1389 // function attributes.
1390 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1391 // In case if we can handle this set of attributes - set up function attrs
1392 // directly on statepoint and return attrs later for gc_result intrinsic.
1393 invoke->setAttributes(new_attrs.getFnAttributes());
1394 return_attributes = new_attrs.getRetAttributes();
1398 // Generate gc relocates in exceptional path
1399 BasicBlock *unwindBlock = toReplace->getUnwindDest();
1400 assert(!isa<PHINode>(unwindBlock->begin()) &&
1401 unwindBlock->getUniquePredecessor() &&
1402 "can't safely insert in this block!");
1404 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1405 Builder.SetInsertPoint(IP);
1406 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1408 // Extract second element from landingpad return value. We will attach
1409 // exceptional gc relocates to it.
1410 const unsigned idx = 1;
1411 Instruction *exceptional_token =
1412 cast<Instruction>(Builder.CreateExtractValue(
1413 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1414 result.UnwindToken = exceptional_token;
1416 CreateGCRelocates(liveVariables, live_start, basePtrs,
1417 exceptional_token, Builder);
1419 // Generate gc relocates and returns for normal block
1420 BasicBlock *normalDest = toReplace->getNormalDest();
1421 assert(!isa<PHINode>(normalDest->begin()) &&
1422 normalDest->getUniquePredecessor() &&
1423 "can't safely insert in this block!");
1425 IP = &*(normalDest->getFirstInsertionPt());
1426 Builder.SetInsertPoint(IP);
1428 // gc relocates will be generated later as if it were regular call
1433 // Take the name of the original value call if it had one.
1434 token->takeName(CS.getInstruction());
1436 // The GCResult is already inserted, we just need to find it
1438 Instruction *toReplace = CS.getInstruction();
1439 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1440 "only valid use before rewrite is gc.result");
1441 assert(!toReplace->hasOneUse() ||
1442 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1445 // Update the gc.result of the original statepoint (if any) to use the newly
1446 // inserted statepoint. This is safe to do here since the token can't be
1447 // considered a live reference.
1448 CS.getInstruction()->replaceAllUsesWith(token);
1450 result.StatepointToken = token;
1452 // Second, create a gc.relocate for every live variable
1453 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1457 struct name_ordering {
1460 bool operator()(name_ordering const &a, name_ordering const &b) {
1461 return -1 == a.derived->getName().compare(b.derived->getName());
1465 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1466 SmallVectorImpl<Value *> &livevec) {
1467 assert(basevec.size() == livevec.size());
1469 SmallVector<name_ordering, 64> temp;
1470 for (size_t i = 0; i < basevec.size(); i++) {
1472 v.base = basevec[i];
1473 v.derived = livevec[i];
1476 std::sort(temp.begin(), temp.end(), name_ordering());
1477 for (size_t i = 0; i < basevec.size(); i++) {
1478 basevec[i] = temp[i].base;
1479 livevec[i] = temp[i].derived;
1483 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1484 // which make the relocations happening at this safepoint explicit.
1486 // WARNING: Does not do any fixup to adjust users of the original live
1487 // values. That's the callers responsibility.
1489 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1490 PartiallyConstructedSafepointRecord &result) {
1491 auto liveset = result.liveset;
1492 auto PointerToBase = result.PointerToBase;
1494 // Convert to vector for efficient cross referencing.
1495 SmallVector<Value *, 64> basevec, livevec;
1496 livevec.reserve(liveset.size());
1497 basevec.reserve(liveset.size());
1498 for (Value *L : liveset) {
1499 livevec.push_back(L);
1500 assert(PointerToBase.count(L));
1501 Value *base = PointerToBase[L];
1502 basevec.push_back(base);
1504 assert(livevec.size() == basevec.size());
1506 // To make the output IR slightly more stable (for use in diffs), ensure a
1507 // fixed order of the values in the safepoint (by sorting the value name).
1508 // The order is otherwise meaningless.
1509 stablize_order(basevec, livevec);
1511 // Do the actual rewriting and delete the old statepoint
1512 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1513 CS.getInstruction()->eraseFromParent();
1516 // Helper function for the relocationViaAlloca.
1517 // It receives iterator to the statepoint gc relocates and emits store to the
1519 // location (via allocaMap) for the each one of them.
1520 // Add visited values into the visitedLiveValues set we will later use them
1521 // for sanity check.
1523 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1524 DenseMap<Value *, Value *> &AllocaMap,
1525 DenseSet<Value *> &VisitedLiveValues) {
1527 for (User *U : GCRelocs) {
1528 if (!isa<IntrinsicInst>(U))
1531 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1533 // We only care about relocates
1534 if (RelocatedValue->getIntrinsicID() !=
1535 Intrinsic::experimental_gc_relocate) {
1539 GCRelocateOperands RelocateOperands(RelocatedValue);
1540 Value *OriginalValue =
1541 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1542 assert(AllocaMap.count(OriginalValue));
1543 Value *Alloca = AllocaMap[OriginalValue];
1545 // Emit store into the related alloca
1546 // All gc_relocate are i8 addrspace(1)* typed, and it must be bitcasted to
1547 // the correct type according to alloca.
1548 assert(RelocatedValue->getNextNode() && "Should always have one since it's not a terminator");
1549 IRBuilder<> Builder(RelocatedValue->getNextNode());
1550 Value *CastedRelocatedValue =
1551 Builder.CreateBitCast(RelocatedValue, cast<AllocaInst>(Alloca)->getAllocatedType(),
1552 RelocatedValue->hasName() ? RelocatedValue->getName() + ".casted" : "");
1554 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1555 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1558 VisitedLiveValues.insert(OriginalValue);
1563 // Helper function for the "relocationViaAlloca". Similar to the
1564 // "insertRelocationStores" but works for rematerialized values.
1566 insertRematerializationStores(
1567 RematerializedValueMapTy RematerializedValues,
1568 DenseMap<Value *, Value *> &AllocaMap,
1569 DenseSet<Value *> &VisitedLiveValues) {
1571 for (auto RematerializedValuePair: RematerializedValues) {
1572 Instruction *RematerializedValue = RematerializedValuePair.first;
1573 Value *OriginalValue = RematerializedValuePair.second;
1575 assert(AllocaMap.count(OriginalValue) &&
1576 "Can not find alloca for rematerialized value");
1577 Value *Alloca = AllocaMap[OriginalValue];
1579 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1580 Store->insertAfter(RematerializedValue);
1583 VisitedLiveValues.insert(OriginalValue);
1588 /// do all the relocation update via allocas and mem2reg
1589 static void relocationViaAlloca(
1590 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1591 ArrayRef<struct PartiallyConstructedSafepointRecord> Records) {
1593 // record initial number of (static) allocas; we'll check we have the same
1594 // number when we get done.
1595 int InitialAllocaNum = 0;
1596 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1598 if (isa<AllocaInst>(*I))
1602 // TODO-PERF: change data structures, reserve
1603 DenseMap<Value *, Value *> AllocaMap;
1604 SmallVector<AllocaInst *, 200> PromotableAllocas;
1605 // Used later to chack that we have enough allocas to store all values
1606 std::size_t NumRematerializedValues = 0;
1607 PromotableAllocas.reserve(Live.size());
1609 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1610 // "PromotableAllocas"
1611 auto emitAllocaFor = [&](Value *LiveValue) {
1612 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1613 F.getEntryBlock().getFirstNonPHI());
1614 AllocaMap[LiveValue] = Alloca;
1615 PromotableAllocas.push_back(Alloca);
1618 // emit alloca for each live gc pointer
1619 for (unsigned i = 0; i < Live.size(); i++) {
1620 emitAllocaFor(Live[i]);
1623 // emit allocas for rematerialized values
1624 for (size_t i = 0; i < Records.size(); i++) {
1625 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1627 for (auto RematerializedValuePair : Info.RematerializedValues) {
1628 Value *OriginalValue = RematerializedValuePair.second;
1629 if (AllocaMap.count(OriginalValue) != 0)
1632 emitAllocaFor(OriginalValue);
1633 ++NumRematerializedValues;
1637 // The next two loops are part of the same conceptual operation. We need to
1638 // insert a store to the alloca after the original def and at each
1639 // redefinition. We need to insert a load before each use. These are split
1640 // into distinct loops for performance reasons.
1642 // update gc pointer after each statepoint
1643 // either store a relocated value or null (if no relocated value found for
1644 // this gc pointer and it is not a gc_result)
1645 // this must happen before we update the statepoint with load of alloca
1646 // otherwise we lose the link between statepoint and old def
1647 for (size_t i = 0; i < Records.size(); i++) {
1648 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1649 Value *Statepoint = Info.StatepointToken;
1651 // This will be used for consistency check
1652 DenseSet<Value *> VisitedLiveValues;
1654 // Insert stores for normal statepoint gc relocates
1655 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1657 // In case if it was invoke statepoint
1658 // we will insert stores for exceptional path gc relocates.
1659 if (isa<InvokeInst>(Statepoint)) {
1660 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1664 // Do similar thing with rematerialized values
1665 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1668 if (ClobberNonLive) {
1669 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1670 // the gc.statepoint. This will turn some subtle GC problems into
1671 // slightly easier to debug SEGVs. Note that on large IR files with
1672 // lots of gc.statepoints this is extremely costly both memory and time
1674 SmallVector<AllocaInst *, 64> ToClobber;
1675 for (auto Pair : AllocaMap) {
1676 Value *Def = Pair.first;
1677 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1679 // This value was relocated
1680 if (VisitedLiveValues.count(Def)) {
1683 ToClobber.push_back(Alloca);
1686 auto InsertClobbersAt = [&](Instruction *IP) {
1687 for (auto *AI : ToClobber) {
1688 auto AIType = cast<PointerType>(AI->getType());
1689 auto PT = cast<PointerType>(AIType->getElementType());
1690 Constant *CPN = ConstantPointerNull::get(PT);
1691 StoreInst *Store = new StoreInst(CPN, AI);
1692 Store->insertBefore(IP);
1696 // Insert the clobbering stores. These may get intermixed with the
1697 // gc.results and gc.relocates, but that's fine.
1698 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1699 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1700 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1702 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1704 InsertClobbersAt(Next);
1708 // update use with load allocas and add store for gc_relocated
1709 for (auto Pair : AllocaMap) {
1710 Value *Def = Pair.first;
1711 Value *Alloca = Pair.second;
1713 // we pre-record the uses of allocas so that we dont have to worry about
1715 // that change the user information.
1716 SmallVector<Instruction *, 20> Uses;
1717 // PERF: trade a linear scan for repeated reallocation
1718 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1719 for (User *U : Def->users()) {
1720 if (!isa<ConstantExpr>(U)) {
1721 // If the def has a ConstantExpr use, then the def is either a
1722 // ConstantExpr use itself or null. In either case
1723 // (recursively in the first, directly in the second), the oop
1724 // it is ultimately dependent on is null and this particular
1725 // use does not need to be fixed up.
1726 Uses.push_back(cast<Instruction>(U));
1730 std::sort(Uses.begin(), Uses.end());
1731 auto Last = std::unique(Uses.begin(), Uses.end());
1732 Uses.erase(Last, Uses.end());
1734 for (Instruction *Use : Uses) {
1735 if (isa<PHINode>(Use)) {
1736 PHINode *Phi = cast<PHINode>(Use);
1737 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1738 if (Def == Phi->getIncomingValue(i)) {
1739 LoadInst *Load = new LoadInst(
1740 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1741 Phi->setIncomingValue(i, Load);
1745 LoadInst *Load = new LoadInst(Alloca, "", Use);
1746 Use->replaceUsesOfWith(Def, Load);
1750 // emit store for the initial gc value
1751 // store must be inserted after load, otherwise store will be in alloca's
1752 // use list and an extra load will be inserted before it
1753 StoreInst *Store = new StoreInst(Def, Alloca);
1754 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1755 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1756 // InvokeInst is a TerminatorInst so the store need to be inserted
1757 // into its normal destination block.
1758 BasicBlock *NormalDest = Invoke->getNormalDest();
1759 Store->insertBefore(NormalDest->getFirstNonPHI());
1761 assert(!Inst->isTerminator() &&
1762 "The only TerminatorInst that can produce a value is "
1763 "InvokeInst which is handled above.");
1764 Store->insertAfter(Inst);
1767 assert(isa<Argument>(Def));
1768 Store->insertAfter(cast<Instruction>(Alloca));
1772 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1773 "we must have the same allocas with lives");
1774 if (!PromotableAllocas.empty()) {
1775 // apply mem2reg to promote alloca to SSA
1776 PromoteMemToReg(PromotableAllocas, DT);
1780 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1782 if (isa<AllocaInst>(*I))
1784 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1788 /// Implement a unique function which doesn't require we sort the input
1789 /// vector. Doing so has the effect of changing the output of a couple of
1790 /// tests in ways which make them less useful in testing fused safepoints.
1791 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1792 SmallSet<T, 8> Seen;
1793 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1794 return !Seen.insert(V).second;
1798 /// Insert holders so that each Value is obviously live through the entire
1799 /// lifetime of the call.
1800 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1801 SmallVectorImpl<CallInst *> &Holders) {
1803 // No values to hold live, might as well not insert the empty holder
1806 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1807 // Use a dummy vararg function to actually hold the values live
1808 Function *Func = cast<Function>(M->getOrInsertFunction(
1809 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1811 // For call safepoints insert dummy calls right after safepoint
1812 BasicBlock::iterator Next(CS.getInstruction());
1814 Holders.push_back(CallInst::Create(Func, Values, "", Next));
1817 // For invoke safepooints insert dummy calls both in normal and
1818 // exceptional destination blocks
1819 auto *II = cast<InvokeInst>(CS.getInstruction());
1820 Holders.push_back(CallInst::Create(
1821 Func, Values, "", II->getNormalDest()->getFirstInsertionPt()));
1822 Holders.push_back(CallInst::Create(
1823 Func, Values, "", II->getUnwindDest()->getFirstInsertionPt()));
1826 static void findLiveReferences(
1827 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1828 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1829 GCPtrLivenessData OriginalLivenessData;
1830 computeLiveInValues(DT, F, OriginalLivenessData);
1831 for (size_t i = 0; i < records.size(); i++) {
1832 struct PartiallyConstructedSafepointRecord &info = records[i];
1833 const CallSite &CS = toUpdate[i];
1834 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1838 /// Remove any vector of pointers from the liveset by scalarizing them over the
1839 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1840 /// would be preferable to include the vector in the statepoint itself, but
1841 /// the lowering code currently does not handle that. Extending it would be
1842 /// slightly non-trivial since it requires a format change. Given how rare
1843 /// such cases are (for the moment?) scalarizing is an acceptable compromise.
1844 static void splitVectorValues(Instruction *StatepointInst,
1845 StatepointLiveSetTy &LiveSet,
1846 DenseMap<Value *, Value *>& PointerToBase,
1847 DominatorTree &DT) {
1848 SmallVector<Value *, 16> ToSplit;
1849 for (Value *V : LiveSet)
1850 if (isa<VectorType>(V->getType()))
1851 ToSplit.push_back(V);
1853 if (ToSplit.empty())
1856 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1858 Function &F = *(StatepointInst->getParent()->getParent());
1860 DenseMap<Value *, AllocaInst *> AllocaMap;
1861 // First is normal return, second is exceptional return (invoke only)
1862 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1863 for (Value *V : ToSplit) {
1864 AllocaInst *Alloca =
1865 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1866 AllocaMap[V] = Alloca;
1868 VectorType *VT = cast<VectorType>(V->getType());
1869 IRBuilder<> Builder(StatepointInst);
1870 SmallVector<Value *, 16> Elements;
1871 for (unsigned i = 0; i < VT->getNumElements(); i++)
1872 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1873 ElementMapping[V] = Elements;
1875 auto InsertVectorReform = [&](Instruction *IP) {
1876 Builder.SetInsertPoint(IP);
1877 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1878 Value *ResultVec = UndefValue::get(VT);
1879 for (unsigned i = 0; i < VT->getNumElements(); i++)
1880 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1881 Builder.getInt32(i));
1885 if (isa<CallInst>(StatepointInst)) {
1886 BasicBlock::iterator Next(StatepointInst);
1888 Instruction *IP = &*(Next);
1889 Replacements[V].first = InsertVectorReform(IP);
1890 Replacements[V].second = nullptr;
1892 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1893 // We've already normalized - check that we don't have shared destination
1895 BasicBlock *NormalDest = Invoke->getNormalDest();
1896 assert(!isa<PHINode>(NormalDest->begin()));
1897 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1898 assert(!isa<PHINode>(UnwindDest->begin()));
1899 // Insert insert element sequences in both successors
1900 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1901 Replacements[V].first = InsertVectorReform(IP);
1902 IP = &*(UnwindDest->getFirstInsertionPt());
1903 Replacements[V].second = InsertVectorReform(IP);
1907 for (Value *V : ToSplit) {
1908 AllocaInst *Alloca = AllocaMap[V];
1910 // Capture all users before we start mutating use lists
1911 SmallVector<Instruction *, 16> Users;
1912 for (User *U : V->users())
1913 Users.push_back(cast<Instruction>(U));
1915 for (Instruction *I : Users) {
1916 if (auto Phi = dyn_cast<PHINode>(I)) {
1917 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1918 if (V == Phi->getIncomingValue(i)) {
1919 LoadInst *Load = new LoadInst(
1920 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1921 Phi->setIncomingValue(i, Load);
1924 LoadInst *Load = new LoadInst(Alloca, "", I);
1925 I->replaceUsesOfWith(V, Load);
1929 // Store the original value and the replacement value into the alloca
1930 StoreInst *Store = new StoreInst(V, Alloca);
1931 if (auto I = dyn_cast<Instruction>(V))
1932 Store->insertAfter(I);
1934 Store->insertAfter(Alloca);
1936 // Normal return for invoke, or call return
1937 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1938 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1939 // Unwind return for invoke only
1940 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1942 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1945 // apply mem2reg to promote alloca to SSA
1946 SmallVector<AllocaInst *, 16> Allocas;
1947 for (Value *V : ToSplit)
1948 Allocas.push_back(AllocaMap[V]);
1949 PromoteMemToReg(Allocas, DT);
1951 // Update our tracking of live pointers and base mappings to account for the
1952 // changes we just made.
1953 for (Value *V : ToSplit) {
1954 auto &Elements = ElementMapping[V];
1957 LiveSet.insert(Elements.begin(), Elements.end());
1958 // We need to update the base mapping as well.
1959 assert(PointerToBase.count(V));
1960 Value *OldBase = PointerToBase[V];
1961 auto &BaseElements = ElementMapping[OldBase];
1962 PointerToBase.erase(V);
1963 assert(Elements.size() == BaseElements.size());
1964 for (unsigned i = 0; i < Elements.size(); i++) {
1965 Value *Elem = Elements[i];
1966 PointerToBase[Elem] = BaseElements[i];
1971 // Helper function for the "rematerializeLiveValues". It walks use chain
1972 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
1973 // values are visited (currently it is GEP's and casts). Returns true if it
1974 // successfully reached "BaseValue" and false otherwise.
1975 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
1977 static bool findRematerializableChainToBasePointer(
1978 SmallVectorImpl<Instruction*> &ChainToBase,
1979 Value *CurrentValue, Value *BaseValue) {
1981 // We have found a base value
1982 if (CurrentValue == BaseValue) {
1986 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
1987 ChainToBase.push_back(GEP);
1988 return findRematerializableChainToBasePointer(ChainToBase,
1989 GEP->getPointerOperand(),
1993 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
1994 Value *Def = CI->stripPointerCasts();
1996 // This two checks are basically similar. First one is here for the
1997 // consistency with findBasePointers logic.
1998 assert(!isa<CastInst>(Def) && "not a pointer cast found");
1999 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
2002 ChainToBase.push_back(CI);
2003 return findRematerializableChainToBasePointer(ChainToBase, Def, BaseValue);
2006 // Not supported instruction in the chain
2010 // Helper function for the "rematerializeLiveValues". Compute cost of the use
2011 // chain we are going to rematerialize.
2013 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
2014 TargetTransformInfo &TTI) {
2017 for (Instruction *Instr : Chain) {
2018 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
2019 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
2020 "non noop cast is found during rematerialization");
2022 Type *SrcTy = CI->getOperand(0)->getType();
2023 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
2025 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
2026 // Cost of the address calculation
2027 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
2028 Cost += TTI.getAddressComputationCost(ValTy);
2030 // And cost of the GEP itself
2031 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
2032 // allowed for the external usage)
2033 if (!GEP->hasAllConstantIndices())
2037 llvm_unreachable("unsupported instruciton type during rematerialization");
2044 // From the statepoint liveset pick values that are cheaper to recompute then to
2045 // relocate. Remove this values from the liveset, rematerialize them after
2046 // statepoint and record them in "Info" structure. Note that similar to
2047 // relocated values we don't do any user adjustments here.
2048 static void rematerializeLiveValues(CallSite CS,
2049 PartiallyConstructedSafepointRecord &Info,
2050 TargetTransformInfo &TTI) {
2051 const unsigned int ChainLengthThreshold = 10;
2053 // Record values we are going to delete from this statepoint live set.
2054 // We can not di this in following loop due to iterator invalidation.
2055 SmallVector<Value *, 32> LiveValuesToBeDeleted;
2057 for (Value *LiveValue: Info.liveset) {
2058 // For each live pointer find it's defining chain
2059 SmallVector<Instruction *, 3> ChainToBase;
2060 assert(Info.PointerToBase.count(LiveValue));
2062 findRematerializableChainToBasePointer(ChainToBase,
2064 Info.PointerToBase[LiveValue]);
2065 // Nothing to do, or chain is too long
2067 ChainToBase.size() == 0 ||
2068 ChainToBase.size() > ChainLengthThreshold)
2071 // Compute cost of this chain
2072 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
2073 // TODO: We can also account for cases when we will be able to remove some
2074 // of the rematerialized values by later optimization passes. I.e if
2075 // we rematerialized several intersecting chains. Or if original values
2076 // don't have any uses besides this statepoint.
2078 // For invokes we need to rematerialize each chain twice - for normal and
2079 // for unwind basic blocks. Model this by multiplying cost by two.
2080 if (CS.isInvoke()) {
2083 // If it's too expensive - skip it
2084 if (Cost >= RematerializationThreshold)
2087 // Remove value from the live set
2088 LiveValuesToBeDeleted.push_back(LiveValue);
2090 // Clone instructions and record them inside "Info" structure
2092 // Walk backwards to visit top-most instructions first
2093 std::reverse(ChainToBase.begin(), ChainToBase.end());
2095 // Utility function which clones all instructions from "ChainToBase"
2096 // and inserts them before "InsertBefore". Returns rematerialized value
2097 // which should be used after statepoint.
2098 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
2099 Instruction *LastClonedValue = nullptr;
2100 Instruction *LastValue = nullptr;
2101 for (Instruction *Instr: ChainToBase) {
2102 // Only GEP's and casts are suported as we need to be careful to not
2103 // introduce any new uses of pointers not in the liveset.
2104 // Note that it's fine to introduce new uses of pointers which were
2105 // otherwise not used after this statepoint.
2106 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2108 Instruction *ClonedValue = Instr->clone();
2109 ClonedValue->insertBefore(InsertBefore);
2110 ClonedValue->setName(Instr->getName() + ".remat");
2112 // If it is not first instruction in the chain then it uses previously
2113 // cloned value. We should update it to use cloned value.
2114 if (LastClonedValue) {
2116 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2118 // Assert that cloned instruction does not use any instructions from
2119 // this chain other than LastClonedValue
2120 for (auto OpValue : ClonedValue->operand_values()) {
2121 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2122 ChainToBase.end() &&
2123 "incorrect use in rematerialization chain");
2128 LastClonedValue = ClonedValue;
2131 assert(LastClonedValue);
2132 return LastClonedValue;
2135 // Different cases for calls and invokes. For invokes we need to clone
2136 // instructions both on normal and unwind path.
2138 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2139 assert(InsertBefore);
2140 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2141 Info.RematerializedValues[RematerializedValue] = LiveValue;
2143 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2145 Instruction *NormalInsertBefore =
2146 Invoke->getNormalDest()->getFirstInsertionPt();
2147 Instruction *UnwindInsertBefore =
2148 Invoke->getUnwindDest()->getFirstInsertionPt();
2150 Instruction *NormalRematerializedValue =
2151 rematerializeChain(NormalInsertBefore);
2152 Instruction *UnwindRematerializedValue =
2153 rematerializeChain(UnwindInsertBefore);
2155 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2156 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2160 // Remove rematerializaed values from the live set
2161 for (auto LiveValue: LiveValuesToBeDeleted) {
2162 Info.liveset.erase(LiveValue);
2166 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
2167 SmallVectorImpl<CallSite> &toUpdate) {
2169 // sanity check the input
2170 std::set<CallSite> uniqued;
2171 uniqued.insert(toUpdate.begin(), toUpdate.end());
2172 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
2174 for (size_t i = 0; i < toUpdate.size(); i++) {
2175 CallSite &CS = toUpdate[i];
2176 assert(CS.getInstruction()->getParent()->getParent() == &F);
2177 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
2181 // When inserting gc.relocates for invokes, we need to be able to insert at
2182 // the top of the successor blocks. See the comment on
2183 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2184 // may restructure the CFG.
2185 for (CallSite CS : toUpdate) {
2188 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
2189 normalizeForInvokeSafepoint(invoke->getNormalDest(), invoke->getParent(),
2191 normalizeForInvokeSafepoint(invoke->getUnwindDest(), invoke->getParent(),
2195 // A list of dummy calls added to the IR to keep various values obviously
2196 // live in the IR. We'll remove all of these when done.
2197 SmallVector<CallInst *, 64> holders;
2199 // Insert a dummy call with all of the arguments to the vm_state we'll need
2200 // for the actual safepoint insertion. This ensures reference arguments in
2201 // the deopt argument list are considered live through the safepoint (and
2202 // thus makes sure they get relocated.)
2203 for (size_t i = 0; i < toUpdate.size(); i++) {
2204 CallSite &CS = toUpdate[i];
2205 Statepoint StatepointCS(CS);
2207 SmallVector<Value *, 64> DeoptValues;
2208 for (Use &U : StatepointCS.vm_state_args()) {
2209 Value *Arg = cast<Value>(&U);
2210 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2211 "support for FCA unimplemented");
2212 if (isHandledGCPointerType(Arg->getType()))
2213 DeoptValues.push_back(Arg);
2215 insertUseHolderAfter(CS, DeoptValues, holders);
2218 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
2219 records.reserve(toUpdate.size());
2220 for (size_t i = 0; i < toUpdate.size(); i++) {
2221 struct PartiallyConstructedSafepointRecord info;
2222 records.push_back(info);
2224 assert(records.size() == toUpdate.size());
2226 // A) Identify all gc pointers which are statically live at the given call
2228 findLiveReferences(F, DT, P, toUpdate, records);
2230 // B) Find the base pointers for each live pointer
2231 /* scope for caching */ {
2232 // Cache the 'defining value' relation used in the computation and
2233 // insertion of base phis and selects. This ensures that we don't insert
2234 // large numbers of duplicate base_phis.
2235 DefiningValueMapTy DVCache;
2237 for (size_t i = 0; i < records.size(); i++) {
2238 struct PartiallyConstructedSafepointRecord &info = records[i];
2239 CallSite &CS = toUpdate[i];
2240 findBasePointers(DT, DVCache, CS, info);
2242 } // end of cache scope
2244 // The base phi insertion logic (for any safepoint) may have inserted new
2245 // instructions which are now live at some safepoint. The simplest such
2248 // phi a <-- will be a new base_phi here
2249 // safepoint 1 <-- that needs to be live here
2253 // We insert some dummy calls after each safepoint to definitely hold live
2254 // the base pointers which were identified for that safepoint. We'll then
2255 // ask liveness for _every_ base inserted to see what is now live. Then we
2256 // remove the dummy calls.
2257 holders.reserve(holders.size() + records.size());
2258 for (size_t i = 0; i < records.size(); i++) {
2259 struct PartiallyConstructedSafepointRecord &info = records[i];
2260 CallSite &CS = toUpdate[i];
2262 SmallVector<Value *, 128> Bases;
2263 for (auto Pair : info.PointerToBase) {
2264 Bases.push_back(Pair.second);
2266 insertUseHolderAfter(CS, Bases, holders);
2269 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2270 // need to rerun liveness. We may *also* have inserted new defs, but that's
2271 // not the key issue.
2272 recomputeLiveInValues(F, DT, P, toUpdate, records);
2274 if (PrintBasePointers) {
2275 for (size_t i = 0; i < records.size(); i++) {
2276 struct PartiallyConstructedSafepointRecord &info = records[i];
2277 errs() << "Base Pairs: (w/Relocation)\n";
2278 for (auto Pair : info.PointerToBase) {
2279 errs() << " derived %" << Pair.first->getName() << " base %"
2280 << Pair.second->getName() << "\n";
2284 for (size_t i = 0; i < holders.size(); i++) {
2285 holders[i]->eraseFromParent();
2286 holders[i] = nullptr;
2290 // Do a limited scalarization of any live at safepoint vector values which
2291 // contain pointers. This enables this pass to run after vectorization at
2292 // the cost of some possible performance loss. TODO: it would be nice to
2293 // natively support vectors all the way through the backend so we don't need
2294 // to scalarize here.
2295 for (size_t i = 0; i < records.size(); i++) {
2296 struct PartiallyConstructedSafepointRecord &info = records[i];
2297 Instruction *statepoint = toUpdate[i].getInstruction();
2298 splitVectorValues(cast<Instruction>(statepoint), info.liveset,
2299 info.PointerToBase, DT);
2302 // In order to reduce live set of statepoint we might choose to rematerialize
2303 // some values instead of relocating them. This is purely an optimization and
2304 // does not influence correctness.
2305 TargetTransformInfo &TTI =
2306 P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2308 for (size_t i = 0; i < records.size(); i++) {
2309 struct PartiallyConstructedSafepointRecord &info = records[i];
2310 CallSite &CS = toUpdate[i];
2312 rematerializeLiveValues(CS, info, TTI);
2315 // Now run through and replace the existing statepoints with new ones with
2316 // the live variables listed. We do not yet update uses of the values being
2317 // relocated. We have references to live variables that need to
2318 // survive to the last iteration of this loop. (By construction, the
2319 // previous statepoint can not be a live variable, thus we can and remove
2320 // the old statepoint calls as we go.)
2321 for (size_t i = 0; i < records.size(); i++) {
2322 struct PartiallyConstructedSafepointRecord &info = records[i];
2323 CallSite &CS = toUpdate[i];
2324 makeStatepointExplicit(DT, CS, P, info);
2326 toUpdate.clear(); // prevent accident use of invalid CallSites
2328 // Do all the fixups of the original live variables to their relocated selves
2329 SmallVector<Value *, 128> live;
2330 for (size_t i = 0; i < records.size(); i++) {
2331 struct PartiallyConstructedSafepointRecord &info = records[i];
2332 // We can't simply save the live set from the original insertion. One of
2333 // the live values might be the result of a call which needs a safepoint.
2334 // That Value* no longer exists and we need to use the new gc_result.
2335 // Thankfully, the liveset is embedded in the statepoint (and updated), so
2336 // we just grab that.
2337 Statepoint statepoint(info.StatepointToken);
2338 live.insert(live.end(), statepoint.gc_args_begin(),
2339 statepoint.gc_args_end());
2341 // Do some basic sanity checks on our liveness results before performing
2342 // relocation. Relocation can and will turn mistakes in liveness results
2343 // into non-sensical code which is must harder to debug.
2344 // TODO: It would be nice to test consistency as well
2345 assert(DT.isReachableFromEntry(info.StatepointToken->getParent()) &&
2346 "statepoint must be reachable or liveness is meaningless");
2347 for (Value *V : statepoint.gc_args()) {
2348 if (!isa<Instruction>(V))
2349 // Non-instruction values trivial dominate all possible uses
2351 auto LiveInst = cast<Instruction>(V);
2352 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2353 "unreachable values should never be live");
2354 assert(DT.dominates(LiveInst, info.StatepointToken) &&
2355 "basic SSA liveness expectation violated by liveness analysis");
2359 unique_unsorted(live);
2363 for (auto ptr : live) {
2364 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
2368 relocationViaAlloca(F, DT, live, records);
2369 return !records.empty();
2372 // Handles both return values and arguments for Functions and CallSites.
2373 template <typename AttrHolder>
2374 static void RemoveDerefAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2377 if (AH.getDereferenceableBytes(Index))
2378 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2379 AH.getDereferenceableBytes(Index)));
2380 if (AH.getDereferenceableOrNullBytes(Index))
2381 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2382 AH.getDereferenceableOrNullBytes(Index)));
2385 AH.setAttributes(AH.getAttributes().removeAttributes(
2386 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2390 RewriteStatepointsForGC::stripDereferenceabilityInfoFromPrototype(Function &F) {
2391 LLVMContext &Ctx = F.getContext();
2393 for (Argument &A : F.args())
2394 if (isa<PointerType>(A.getType()))
2395 RemoveDerefAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2397 if (isa<PointerType>(F.getReturnType()))
2398 RemoveDerefAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2401 void RewriteStatepointsForGC::stripDereferenceabilityInfoFromBody(Function &F) {
2405 LLVMContext &Ctx = F.getContext();
2406 MDBuilder Builder(Ctx);
2408 for (Instruction &I : instructions(F)) {
2409 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2410 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2411 bool IsImmutableTBAA =
2412 MD->getNumOperands() == 4 &&
2413 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2415 if (!IsImmutableTBAA)
2416 continue; // no work to do, MD_tbaa is already marked mutable
2418 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2419 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2421 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2423 MDNode *MutableTBAA =
2424 Builder.createTBAAStructTagNode(Base, Access, Offset);
2425 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2428 if (CallSite CS = CallSite(&I)) {
2429 for (int i = 0, e = CS.arg_size(); i != e; i++)
2430 if (isa<PointerType>(CS.getArgument(i)->getType()))
2431 RemoveDerefAttrAtIndex(Ctx, CS, i + 1);
2432 if (isa<PointerType>(CS.getType()))
2433 RemoveDerefAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2438 /// Returns true if this function should be rewritten by this pass. The main
2439 /// point of this function is as an extension point for custom logic.
2440 static bool shouldRewriteStatepointsIn(Function &F) {
2441 // TODO: This should check the GCStrategy
2443 const char *FunctionGCName = F.getGC();
2444 const StringRef StatepointExampleName("statepoint-example");
2445 const StringRef CoreCLRName("coreclr");
2446 return (StatepointExampleName == FunctionGCName) ||
2447 (CoreCLRName == FunctionGCName);
2452 void RewriteStatepointsForGC::stripDereferenceabilityInfo(Module &M) {
2454 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2458 for (Function &F : M)
2459 stripDereferenceabilityInfoFromPrototype(F);
2461 for (Function &F : M)
2462 stripDereferenceabilityInfoFromBody(F);
2465 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2466 // Nothing to do for declarations.
2467 if (F.isDeclaration() || F.empty())
2470 // Policy choice says not to rewrite - the most common reason is that we're
2471 // compiling code without a GCStrategy.
2472 if (!shouldRewriteStatepointsIn(F))
2475 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2477 // Gather all the statepoints which need rewritten. Be careful to only
2478 // consider those in reachable code since we need to ask dominance queries
2479 // when rewriting. We'll delete the unreachable ones in a moment.
2480 SmallVector<CallSite, 64> ParsePointNeeded;
2481 bool HasUnreachableStatepoint = false;
2482 for (Instruction &I : instructions(F)) {
2483 // TODO: only the ones with the flag set!
2484 if (isStatepoint(I)) {
2485 if (DT.isReachableFromEntry(I.getParent()))
2486 ParsePointNeeded.push_back(CallSite(&I));
2488 HasUnreachableStatepoint = true;
2492 bool MadeChange = false;
2494 // Delete any unreachable statepoints so that we don't have unrewritten
2495 // statepoints surviving this pass. This makes testing easier and the
2496 // resulting IR less confusing to human readers. Rather than be fancy, we
2497 // just reuse a utility function which removes the unreachable blocks.
2498 if (HasUnreachableStatepoint)
2499 MadeChange |= removeUnreachableBlocks(F);
2501 // Return early if no work to do.
2502 if (ParsePointNeeded.empty())
2505 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2506 // These are created by LCSSA. They have the effect of increasing the size
2507 // of liveness sets for no good reason. It may be harder to do this post
2508 // insertion since relocations and base phis can confuse things.
2509 for (BasicBlock &BB : F)
2510 if (BB.getUniquePredecessor()) {
2512 FoldSingleEntryPHINodes(&BB);
2515 // Before we start introducing relocations, we want to tweak the IR a bit to
2516 // avoid unfortunate code generation effects. The main example is that we
2517 // want to try to make sure the comparison feeding a branch is after any
2518 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2519 // values feeding a branch after relocation. This is semantically correct,
2520 // but results in extra register pressure since both the pre-relocation and
2521 // post-relocation copies must be available in registers. For code without
2522 // relocations this is handled elsewhere, but teaching the scheduler to
2523 // reverse the transform we're about to do would be slightly complex.
2524 // Note: This may extend the live range of the inputs to the icmp and thus
2525 // increase the liveset of any statepoint we move over. This is profitable
2526 // as long as all statepoints are in rare blocks. If we had in-register
2527 // lowering for live values this would be a much safer transform.
2528 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2529 if (auto *BI = dyn_cast<BranchInst>(TI))
2530 if (BI->isConditional())
2531 return dyn_cast<Instruction>(BI->getCondition());
2532 // TODO: Extend this to handle switches
2535 for (BasicBlock &BB : F) {
2536 TerminatorInst *TI = BB.getTerminator();
2537 if (auto *Cond = getConditionInst(TI))
2538 // TODO: Handle more than just ICmps here. We should be able to move
2539 // most instructions without side effects or memory access.
2540 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2542 Cond->moveBefore(TI);
2546 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
2550 // liveness computation via standard dataflow
2551 // -------------------------------------------------------------------
2553 // TODO: Consider using bitvectors for liveness, the set of potentially
2554 // interesting values should be small and easy to pre-compute.
2556 /// Compute the live-in set for the location rbegin starting from
2557 /// the live-out set of the basic block
2558 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2559 BasicBlock::reverse_iterator rend,
2560 DenseSet<Value *> &LiveTmp) {
2562 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2563 Instruction *I = &*ritr;
2565 // KILL/Def - Remove this definition from LiveIn
2568 // Don't consider *uses* in PHI nodes, we handle their contribution to
2569 // predecessor blocks when we seed the LiveOut sets
2570 if (isa<PHINode>(I))
2573 // USE - Add to the LiveIn set for this instruction
2574 for (Value *V : I->operands()) {
2575 assert(!isUnhandledGCPointerType(V->getType()) &&
2576 "support for FCA unimplemented");
2577 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2578 // The choice to exclude all things constant here is slightly subtle.
2579 // There are two independent reasons:
2580 // - We assume that things which are constant (from LLVM's definition)
2581 // do not move at runtime. For example, the address of a global
2582 // variable is fixed, even though it's contents may not be.
2583 // - Second, we can't disallow arbitrary inttoptr constants even
2584 // if the language frontend does. Optimization passes are free to
2585 // locally exploit facts without respect to global reachability. This
2586 // can create sections of code which are dynamically unreachable and
2587 // contain just about anything. (see constants.ll in tests)
2594 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2596 for (BasicBlock *Succ : successors(BB)) {
2597 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2598 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2599 PHINode *Phi = cast<PHINode>(&*I);
2600 Value *V = Phi->getIncomingValueForBlock(BB);
2601 assert(!isUnhandledGCPointerType(V->getType()) &&
2602 "support for FCA unimplemented");
2603 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2610 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2611 DenseSet<Value *> KillSet;
2612 for (Instruction &I : *BB)
2613 if (isHandledGCPointerType(I.getType()))
2619 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2620 /// sanity check for the liveness computation.
2621 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2622 TerminatorInst *TI, bool TermOkay = false) {
2623 for (Value *V : Live) {
2624 if (auto *I = dyn_cast<Instruction>(V)) {
2625 // The terminator can be a member of the LiveOut set. LLVM's definition
2626 // of instruction dominance states that V does not dominate itself. As
2627 // such, we need to special case this to allow it.
2628 if (TermOkay && TI == I)
2630 assert(DT.dominates(I, TI) &&
2631 "basic SSA liveness expectation violated by liveness analysis");
2636 /// Check that all the liveness sets used during the computation of liveness
2637 /// obey basic SSA properties. This is useful for finding cases where we miss
2639 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2641 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2642 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2643 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2647 static void computeLiveInValues(DominatorTree &DT, Function &F,
2648 GCPtrLivenessData &Data) {
2650 SmallSetVector<BasicBlock *, 200> Worklist;
2651 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2652 // We use a SetVector so that we don't have duplicates in the worklist.
2653 Worklist.insert(pred_begin(BB), pred_end(BB));
2655 auto NextItem = [&]() {
2656 BasicBlock *BB = Worklist.back();
2657 Worklist.pop_back();
2661 // Seed the liveness for each individual block
2662 for (BasicBlock &BB : F) {
2663 Data.KillSet[&BB] = computeKillSet(&BB);
2664 Data.LiveSet[&BB].clear();
2665 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2668 for (Value *Kill : Data.KillSet[&BB])
2669 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2672 Data.LiveOut[&BB] = DenseSet<Value *>();
2673 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2674 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2675 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2676 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2677 if (!Data.LiveIn[&BB].empty())
2678 AddPredsToWorklist(&BB);
2681 // Propagate that liveness until stable
2682 while (!Worklist.empty()) {
2683 BasicBlock *BB = NextItem();
2685 // Compute our new liveout set, then exit early if it hasn't changed
2686 // despite the contribution of our successor.
2687 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2688 const auto OldLiveOutSize = LiveOut.size();
2689 for (BasicBlock *Succ : successors(BB)) {
2690 assert(Data.LiveIn.count(Succ));
2691 set_union(LiveOut, Data.LiveIn[Succ]);
2693 // assert OutLiveOut is a subset of LiveOut
2694 if (OldLiveOutSize == LiveOut.size()) {
2695 // If the sets are the same size, then we didn't actually add anything
2696 // when unioning our successors LiveIn Thus, the LiveIn of this block
2700 Data.LiveOut[BB] = LiveOut;
2702 // Apply the effects of this basic block
2703 DenseSet<Value *> LiveTmp = LiveOut;
2704 set_union(LiveTmp, Data.LiveSet[BB]);
2705 set_subtract(LiveTmp, Data.KillSet[BB]);
2707 assert(Data.LiveIn.count(BB));
2708 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2709 // assert: OldLiveIn is a subset of LiveTmp
2710 if (OldLiveIn.size() != LiveTmp.size()) {
2711 Data.LiveIn[BB] = LiveTmp;
2712 AddPredsToWorklist(BB);
2714 } // while( !worklist.empty() )
2717 // Sanity check our output against SSA properties. This helps catch any
2718 // missing kills during the above iteration.
2719 for (BasicBlock &BB : F) {
2720 checkBasicSSA(DT, Data, BB);
2725 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2726 StatepointLiveSetTy &Out) {
2728 BasicBlock *BB = Inst->getParent();
2730 // Note: The copy is intentional and required
2731 assert(Data.LiveOut.count(BB));
2732 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2734 // We want to handle the statepoint itself oddly. It's
2735 // call result is not live (normal), nor are it's arguments
2736 // (unless they're used again later). This adjustment is
2737 // specifically what we need to relocate
2738 BasicBlock::reverse_iterator rend(Inst);
2739 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2740 LiveOut.erase(Inst);
2741 Out.insert(LiveOut.begin(), LiveOut.end());
2744 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2746 PartiallyConstructedSafepointRecord &Info) {
2747 Instruction *Inst = CS.getInstruction();
2748 StatepointLiveSetTy Updated;
2749 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2752 DenseSet<Value *> Bases;
2753 for (auto KVPair : Info.PointerToBase) {
2754 Bases.insert(KVPair.second);
2757 // We may have base pointers which are now live that weren't before. We need
2758 // to update the PointerToBase structure to reflect this.
2759 for (auto V : Updated)
2760 if (!Info.PointerToBase.count(V)) {
2761 assert(Bases.count(V) && "can't find base for unexpected live value");
2762 Info.PointerToBase[V] = V;
2767 for (auto V : Updated) {
2768 assert(Info.PointerToBase.count(V) &&
2769 "must be able to find base for live value");
2773 // Remove any stale base mappings - this can happen since our liveness is
2774 // more precise then the one inherent in the base pointer analysis
2775 DenseSet<Value *> ToErase;
2776 for (auto KVPair : Info.PointerToBase)
2777 if (!Updated.count(KVPair.first))
2778 ToErase.insert(KVPair.first);
2779 for (auto V : ToErase)
2780 Info.PointerToBase.erase(V);
2783 for (auto KVPair : Info.PointerToBase)
2784 assert(Updated.count(KVPair.first) && "record for non-live value");
2787 Info.liveset = Updated;