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 bool isKnownBaseResult(Value *V);
292 /// A single base defining value - An immediate base defining value for an
293 /// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'.
294 /// For instructions which have multiple pointer [vector] inputs or that
295 /// transition between vector and scalar types, there is no immediate base
296 /// defining value. The 'base defining value' for 'Def' is the transitive
297 /// closure of this relation stopping at the first instruction which has no
298 /// immediate base defining value. The b.d.v. might itself be a base pointer,
299 /// but it can also be an arbitrary derived pointer.
300 struct BaseDefiningValueResult {
301 /// Contains the value which is the base defining value.
303 /// True if the base defining value is also known to be an actual base
305 const bool IsKnownBase;
306 BaseDefiningValueResult(Value *BDV, bool IsKnownBase)
307 : BDV(BDV), IsKnownBase(IsKnownBase) {
309 // Check consistency between new and old means of checking whether a BDV is
311 bool MustBeBase = isKnownBaseResult(BDV);
312 assert(!MustBeBase || MustBeBase == IsKnownBase);
318 static BaseDefiningValueResult findBaseDefiningValue(Value *I);
320 /// Return a base defining value for the 'Index' element of the given vector
321 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
322 /// 'I'. As an optimization, this method will try to determine when the
323 /// element is known to already be a base pointer. If this can be established,
324 /// the second value in the returned pair will be true. Note that either a
325 /// vector or a pointer typed value can be returned. For the former, the
326 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
327 /// If the later, the return pointer is a BDV (or possibly a base) for the
328 /// particular element in 'I'.
329 static BaseDefiningValueResult
330 findBaseDefiningValueOfVector(Value *I, Value *Index = nullptr) {
331 assert(I->getType()->isVectorTy() &&
332 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
333 "Illegal to ask for the base pointer of a non-pointer type");
335 // Each case parallels findBaseDefiningValue below, see that code for
336 // detailed motivation.
338 if (isa<Argument>(I))
339 // An incoming argument to the function is a base pointer
340 return BaseDefiningValueResult(I, true);
342 // We shouldn't see the address of a global as a vector value?
343 assert(!isa<GlobalVariable>(I) &&
344 "unexpected global variable found in base of vector");
346 // inlining could possibly introduce phi node that contains
347 // undef if callee has multiple returns
348 if (isa<UndefValue>(I))
349 // utterly meaningless, but useful for dealing with partially optimized
351 return BaseDefiningValueResult(I, true);
353 // Due to inheritance, this must be _after_ the global variable and undef
355 if (Constant *Con = dyn_cast<Constant>(I)) {
356 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
357 "order of checks wrong!");
358 assert(Con->isNullValue() && "null is the only case which makes sense");
359 return BaseDefiningValueResult(Con, true);
362 if (isa<LoadInst>(I))
363 return BaseDefiningValueResult(I, true);
365 // For an insert element, we might be able to look through it if we know
366 // something about the indexes.
367 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(I)) {
369 Value *InsertIndex = IEI->getOperand(2);
370 // This index is inserting the value, look for its BDV
371 if (InsertIndex == Index)
372 return findBaseDefiningValue(IEI->getOperand(1));
373 // Both constant, and can't be equal per above. This insert is definitely
374 // not relevant, look back at the rest of the vector and keep trying.
375 if (isa<ConstantInt>(Index) && isa<ConstantInt>(InsertIndex))
376 return findBaseDefiningValueOfVector(IEI->getOperand(0), Index);
379 // If both inputs to the insertelement are known bases, then so is the
380 // insertelement itself. NOTE: This should be handled within the generic
381 // base pointer inference code and after http://reviews.llvm.org/D12583,
382 // will be. However, when strengthening asserts I needed to add this to
383 // keep an existing test passing which was 'working'. FIXME
384 if (findBaseDefiningValue(IEI->getOperand(0)).IsKnownBase &&
385 findBaseDefiningValue(IEI->getOperand(1)).IsKnownBase)
386 return BaseDefiningValueResult(IEI, true);
388 // We don't know whether this vector contains entirely base pointers or
389 // not. To be conservatively correct, we treat it as a BDV and will
390 // duplicate code as needed to construct a parallel vector of bases.
391 return BaseDefiningValueResult(IEI, false);
394 if (isa<ShuffleVectorInst>(I))
395 // We don't know whether this vector contains entirely base pointers or
396 // not. To be conservatively correct, we treat it as a BDV and will
397 // duplicate code as needed to construct a parallel vector of bases.
398 // TODO: There a number of local optimizations which could be applied here
399 // for particular sufflevector patterns.
400 return BaseDefiningValueResult(I, false);
402 // A PHI or Select is a base defining value. The outer findBasePointer
403 // algorithm is responsible for constructing a base value for this BDV.
404 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
405 "unknown vector instruction - no base found for vector element");
406 return BaseDefiningValueResult(I, false);
409 /// Helper function for findBasePointer - Will return a value which either a)
410 /// defines the base pointer for the input, b) blocks the simple search
411 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
412 /// from pointer to vector type or back.
413 static BaseDefiningValueResult findBaseDefiningValue(Value *I) {
414 if (I->getType()->isVectorTy())
415 return findBaseDefiningValueOfVector(I);
417 assert(I->getType()->isPointerTy() &&
418 "Illegal to ask for the base pointer of a non-pointer type");
420 if (isa<Argument>(I))
421 // An incoming argument to the function is a base pointer
422 // We should have never reached here if this argument isn't an gc value
423 return BaseDefiningValueResult(I, true);
425 if (isa<GlobalVariable>(I))
427 return BaseDefiningValueResult(I, true);
429 // inlining could possibly introduce phi node that contains
430 // undef if callee has multiple returns
431 if (isa<UndefValue>(I))
432 // utterly meaningless, but useful for dealing with
433 // partially optimized code.
434 return BaseDefiningValueResult(I, true);
436 // Due to inheritance, this must be _after_ the global variable and undef
438 if (Constant *Con = dyn_cast<Constant>(I)) {
439 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
440 "order of checks wrong!");
441 // Note: Finding a constant base for something marked for relocation
442 // doesn't really make sense. The most likely case is either a) some
443 // screwed up the address space usage or b) your validating against
444 // compiled C++ code w/o the proper separation. The only real exception
445 // is a null pointer. You could have generic code written to index of
446 // off a potentially null value and have proven it null. We also use
447 // null pointers in dead paths of relocation phis (which we might later
448 // want to find a base pointer for).
449 assert(isa<ConstantPointerNull>(Con) &&
450 "null is the only case which makes sense");
451 return BaseDefiningValueResult(I, true);
454 if (CastInst *CI = dyn_cast<CastInst>(I)) {
455 Value *Def = CI->stripPointerCasts();
456 // If we find a cast instruction here, it means we've found a cast which is
457 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
458 // handle int->ptr conversion.
459 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
460 return findBaseDefiningValue(Def);
463 if (isa<LoadInst>(I))
464 // The value loaded is an gc base itself
465 return BaseDefiningValueResult(I, true);
468 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
469 // The base of this GEP is the base
470 return findBaseDefiningValue(GEP->getPointerOperand());
472 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
473 switch (II->getIntrinsicID()) {
474 case Intrinsic::experimental_gc_result_ptr:
476 // fall through to general call handling
478 case Intrinsic::experimental_gc_statepoint:
479 case Intrinsic::experimental_gc_result_float:
480 case Intrinsic::experimental_gc_result_int:
481 llvm_unreachable("these don't produce pointers");
482 case Intrinsic::experimental_gc_relocate: {
483 // Rerunning safepoint insertion after safepoints are already
484 // inserted is not supported. It could probably be made to work,
485 // but why are you doing this? There's no good reason.
486 llvm_unreachable("repeat safepoint insertion is not supported");
488 case Intrinsic::gcroot:
489 // Currently, this mechanism hasn't been extended to work with gcroot.
490 // There's no reason it couldn't be, but I haven't thought about the
491 // implications much.
493 "interaction with the gcroot mechanism is not supported");
496 // We assume that functions in the source language only return base
497 // pointers. This should probably be generalized via attributes to support
498 // both source language and internal functions.
499 if (isa<CallInst>(I) || isa<InvokeInst>(I))
500 return BaseDefiningValueResult(I, true);
502 // I have absolutely no idea how to implement this part yet. It's not
503 // necessarily hard, I just haven't really looked at it yet.
504 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
506 if (isa<AtomicCmpXchgInst>(I))
507 // A CAS is effectively a atomic store and load combined under a
508 // predicate. From the perspective of base pointers, we just treat it
510 return BaseDefiningValueResult(I, true);
512 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
513 "binary ops which don't apply to pointers");
515 // The aggregate ops. Aggregates can either be in the heap or on the
516 // stack, but in either case, this is simply a field load. As a result,
517 // this is a defining definition of the base just like a load is.
518 if (isa<ExtractValueInst>(I))
519 return BaseDefiningValueResult(I, true);
521 // We should never see an insert vector since that would require we be
522 // tracing back a struct value not a pointer value.
523 assert(!isa<InsertValueInst>(I) &&
524 "Base pointer for a struct is meaningless");
526 // An extractelement produces a base result exactly when it's input does.
527 // We may need to insert a parallel instruction to extract the appropriate
528 // element out of the base vector corresponding to the input. Given this,
529 // it's analogous to the phi and select case even though it's not a merge.
530 if (auto *EEI = dyn_cast<ExtractElementInst>(I)) {
531 Value *VectorOperand = EEI->getVectorOperand();
532 Value *Index = EEI->getIndexOperand();
533 auto VecResult = findBaseDefiningValueOfVector(VectorOperand, Index);
534 Value *VectorBase = VecResult.BDV;
535 if (VectorBase->getType()->isPointerTy())
536 // We found a BDV for this specific element with the vector. This is an
537 // optimization, but in practice it covers most of the useful cases
538 // created via scalarization. Note: The peephole optimization here is
539 // currently needed for correctness since the general algorithm doesn't
540 // yet handle insertelements. That will change shortly.
541 return BaseDefiningValueResult(VectorBase, VecResult.IsKnownBase);
543 assert(VectorBase->getType()->isVectorTy());
544 // Otherwise, we have an instruction which potentially produces a
545 // derived pointer and we need findBasePointers to clone code for us
546 // such that we can create an instruction which produces the
547 // accompanying base pointer.
548 return BaseDefiningValueResult(I, VecResult.IsKnownBase);
552 // The last two cases here don't return a base pointer. Instead, they
553 // return a value which dynamically selects from among several base
554 // derived pointers (each with it's own base potentially). It's the job of
555 // the caller to resolve these.
556 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
557 "missing instruction case in findBaseDefiningValing");
558 return BaseDefiningValueResult(I, false);
561 /// Returns the base defining value for this value.
562 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
563 Value *&Cached = Cache[I];
565 Cached = findBaseDefiningValue(I).BDV;
566 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
567 << Cached->getName() << "\n");
569 assert(Cache[I] != nullptr);
573 /// Return a base pointer for this value if known. Otherwise, return it's
574 /// base defining value.
575 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
576 Value *Def = findBaseDefiningValueCached(I, Cache);
577 auto Found = Cache.find(Def);
578 if (Found != Cache.end()) {
579 // Either a base-of relation, or a self reference. Caller must check.
580 return Found->second;
582 // Only a BDV available
586 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
587 /// is it known to be a base pointer? Or do we need to continue searching.
588 static bool isKnownBaseResult(Value *V) {
589 if (!isa<PHINode>(V) && !isa<SelectInst>(V) && !isa<ExtractElementInst>(V)) {
590 // no recursion possible
593 if (isa<Instruction>(V) &&
594 cast<Instruction>(V)->getMetadata("is_base_value")) {
595 // This is a previously inserted base phi or select. We know
596 // that this is a base value.
600 // We need to keep searching
605 /// Models the state of a single base defining value in the findBasePointer
606 /// algorithm for determining where a new instruction is needed to propagate
607 /// the base of this BDV.
610 enum Status { Unknown, Base, Conflict };
612 BDVState(Status s, Value *b = nullptr) : status(s), base(b) {
613 assert(status != Base || b);
615 explicit BDVState(Value *b) : status(Base), base(b) {}
616 BDVState() : status(Unknown), base(nullptr) {}
618 Status getStatus() const { return status; }
619 Value *getBase() const { return base; }
621 bool isBase() const { return getStatus() == Base; }
622 bool isUnknown() const { return getStatus() == Unknown; }
623 bool isConflict() const { return getStatus() == Conflict; }
625 bool operator==(const BDVState &other) const {
626 return base == other.base && status == other.status;
629 bool operator!=(const BDVState &other) const { return !(*this == other); }
632 void dump() const { print(dbgs()); dbgs() << '\n'; }
634 void print(raw_ostream &OS) const {
646 OS << " (" << base << " - "
647 << (base ? base->getName() : "nullptr") << "): ";
652 Value *base; // non null only if status == base
657 static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
664 typedef DenseMap<Value *, BDVState> ConflictStateMapTy;
665 // Values of type BDVState form a lattice, and this is a helper
666 // class that implementes the meet operation. The meat of the meet
667 // operation is implemented in MeetBDVStates::pureMeet
668 class MeetBDVStates {
670 /// Initializes the currentResult to the TOP state so that if can be met with
671 /// any other state to produce that state.
674 // Destructively meet the current result with the given BDVState
675 void meetWith(BDVState otherState) {
676 currentResult = meet(otherState, currentResult);
679 BDVState getResult() const { return currentResult; }
682 BDVState currentResult;
684 /// Perform a meet operation on two elements of the BDVState lattice.
685 static BDVState meet(BDVState LHS, BDVState RHS) {
686 assert((pureMeet(LHS, RHS) == pureMeet(RHS, LHS)) &&
687 "math is wrong: meet does not commute!");
688 BDVState Result = pureMeet(LHS, RHS);
689 DEBUG(dbgs() << "meet of " << LHS << " with " << RHS
690 << " produced " << Result << "\n");
694 static BDVState pureMeet(const BDVState &stateA, const BDVState &stateB) {
695 switch (stateA.getStatus()) {
696 case BDVState::Unknown:
700 assert(stateA.getBase() && "can't be null");
701 if (stateB.isUnknown())
704 if (stateB.isBase()) {
705 if (stateA.getBase() == stateB.getBase()) {
706 assert(stateA == stateB && "equality broken!");
709 return BDVState(BDVState::Conflict);
711 assert(stateB.isConflict() && "only three states!");
712 return BDVState(BDVState::Conflict);
714 case BDVState::Conflict:
717 llvm_unreachable("only three states!");
723 /// For a given value or instruction, figure out what base ptr it's derived
724 /// from. For gc objects, this is simply itself. On success, returns a value
725 /// which is the base pointer. (This is reliable and can be used for
726 /// relocation.) On failure, returns nullptr.
727 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
728 Value *def = findBaseOrBDV(I, cache);
730 if (isKnownBaseResult(def)) {
734 // Here's the rough algorithm:
735 // - For every SSA value, construct a mapping to either an actual base
736 // pointer or a PHI which obscures the base pointer.
737 // - Construct a mapping from PHI to unknown TOP state. Use an
738 // optimistic algorithm to propagate base pointer information. Lattice
743 // When algorithm terminates, all PHIs will either have a single concrete
744 // base or be in a conflict state.
745 // - For every conflict, insert a dummy PHI node without arguments. Add
746 // these to the base[Instruction] = BasePtr mapping. For every
747 // non-conflict, add the actual base.
748 // - For every conflict, add arguments for the base[a] of each input
751 // Note: A simpler form of this would be to add the conflict form of all
752 // PHIs without running the optimistic algorithm. This would be
753 // analogous to pessimistic data flow and would likely lead to an
754 // overall worse solution.
757 auto isExpectedBDVType = [](Value *BDV) {
758 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) || isa<ExtractElementInst>(BDV);
762 // Once populated, will contain a mapping from each potentially non-base BDV
763 // to a lattice value (described above) which corresponds to that BDV.
764 ConflictStateMapTy states;
765 // Recursively fill in all phis & selects reachable from the initial one
766 // for which we don't already know a definite base value for
768 DenseSet<Value *> Visited;
769 SmallVector<Value*, 16> Worklist;
770 Worklist.push_back(def);
772 while (!Worklist.empty()) {
773 Value *Current = Worklist.pop_back_val();
774 assert(!isKnownBaseResult(Current) && "why did it get added?");
776 auto visitIncomingValue = [&](Value *InVal) {
777 Value *Base = findBaseOrBDV(InVal, cache);
778 if (isKnownBaseResult(Base))
779 // Known bases won't need new instructions introduced and can be
782 assert(isExpectedBDVType(Base) && "the only non-base values "
783 "we see should be base defining values");
784 if (Visited.insert(Base).second)
785 Worklist.push_back(Base);
787 if (PHINode *Phi = dyn_cast<PHINode>(Current)) {
788 for (Value *InVal : Phi->incoming_values())
789 visitIncomingValue(InVal);
790 } else if (SelectInst *Sel = dyn_cast<SelectInst>(Current)) {
791 visitIncomingValue(Sel->getTrueValue());
792 visitIncomingValue(Sel->getFalseValue());
793 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
794 visitIncomingValue(EE->getVectorOperand());
796 // There are two classes of instructions we know we don't handle.
797 assert(isa<ShuffleVectorInst>(Current) ||
798 isa<InsertElementInst>(Current));
799 llvm_unreachable("unimplemented instruction case");
802 // The frontier of visited instructions are the ones we might need to
803 // duplicate, so fill in the starting state for the optimistic algorithm
805 for (Value *BDV : Visited) {
806 states[BDV] = BDVState();
811 DEBUG(dbgs() << "States after initialization:\n");
812 for (auto Pair : states) {
813 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
817 // Return a phi state for a base defining value. We'll generate a new
818 // base state for known bases and expect to find a cached state otherwise.
819 auto getStateForBDV = [&](Value *baseValue) {
820 if (isKnownBaseResult(baseValue))
821 return BDVState(baseValue);
822 auto I = states.find(baseValue);
823 assert(I != states.end() && "lookup failed!");
827 bool progress = true;
830 size_t oldSize = states.size();
833 // We're only changing keys in this loop, thus safe to keep iterators
834 for (auto Pair : states) {
835 Value *v = Pair.first;
836 assert(!isKnownBaseResult(v) && "why did it get added?");
838 // Given an input value for the current instruction, return a BDVState
839 // instance which represents the BDV of that value.
840 auto getStateForInput = [&](Value *V) mutable {
841 Value *BDV = findBaseOrBDV(V, cache);
842 return getStateForBDV(BDV);
845 MeetBDVStates calculateMeet;
846 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
847 calculateMeet.meetWith(getStateForInput(select->getTrueValue()));
848 calculateMeet.meetWith(getStateForInput(select->getFalseValue()));
849 } else if (PHINode *Phi = dyn_cast<PHINode>(v)) {
850 for (Value *Val : Phi->incoming_values())
851 calculateMeet.meetWith(getStateForInput(Val));
853 // The 'meet' for an extractelement is slightly trivial, but it's still
854 // useful in that it drives us to conflict if our input is.
855 auto *EE = cast<ExtractElementInst>(v);
856 calculateMeet.meetWith(getStateForInput(EE->getVectorOperand()));
860 BDVState oldState = states[v];
861 BDVState newState = calculateMeet.getResult();
862 if (oldState != newState) {
864 states[v] = newState;
868 assert(oldSize <= states.size());
869 assert(oldSize == states.size() || progress);
873 DEBUG(dbgs() << "States after meet iteration:\n");
874 for (auto Pair : states) {
875 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
879 // Insert Phis for all conflicts
880 // We want to keep naming deterministic in the loop that follows, so
881 // sort the keys before iteration. This is useful in allowing us to
882 // write stable tests. Note that there is no invalidation issue here.
883 SmallVector<Value *, 16> Keys;
884 Keys.reserve(states.size());
885 for (auto Pair : states) {
886 Value *V = Pair.first;
889 std::sort(Keys.begin(), Keys.end(), order_by_name);
890 // TODO: adjust naming patterns to avoid this order of iteration dependency
891 for (Value *V : Keys) {
892 Instruction *I = cast<Instruction>(V);
893 BDVState State = states[I];
894 assert(!isKnownBaseResult(I) && "why did it get added?");
895 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
897 // extractelement instructions are a bit special in that we may need to
898 // insert an extract even when we know an exact base for the instruction.
899 // The problem is that we need to convert from a vector base to a scalar
900 // base for the particular indice we're interested in.
901 if (State.isBase() && isa<ExtractElementInst>(I) &&
902 isa<VectorType>(State.getBase()->getType())) {
903 auto *EE = cast<ExtractElementInst>(I);
904 // TODO: In many cases, the new instruction is just EE itself. We should
905 // exploit this, but can't do it here since it would break the invariant
906 // about the BDV not being known to be a base.
907 auto *BaseInst = ExtractElementInst::Create(State.getBase(),
908 EE->getIndexOperand(),
910 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
911 states[I] = BDVState(BDVState::Base, BaseInst);
914 if (!State.isConflict())
917 /// Create and insert a new instruction which will represent the base of
918 /// the given instruction 'I'.
919 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
920 if (isa<PHINode>(I)) {
921 BasicBlock *BB = I->getParent();
922 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
923 assert(NumPreds > 0 && "how did we reach here");
924 std::string Name = I->hasName() ?
925 (I->getName() + ".base").str() : "base_phi";
926 return PHINode::Create(I->getType(), NumPreds, Name, I);
927 } else if (SelectInst *Sel = dyn_cast<SelectInst>(I)) {
928 // The undef will be replaced later
929 UndefValue *Undef = UndefValue::get(Sel->getType());
930 std::string Name = I->hasName() ?
931 (I->getName() + ".base").str() : "base_select";
932 return SelectInst::Create(Sel->getCondition(), Undef,
935 auto *EE = cast<ExtractElementInst>(I);
936 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
937 std::string Name = I->hasName() ?
938 (I->getName() + ".base").str() : "base_ee";
939 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
943 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
944 // Add metadata marking this as a base value
945 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
946 states[I] = BDVState(BDVState::Conflict, BaseInst);
949 // Fixup all the inputs of the new PHIs
950 for (auto Pair : states) {
951 Instruction *v = cast<Instruction>(Pair.first);
952 BDVState state = Pair.second;
954 assert(!isKnownBaseResult(v) && "why did it get added?");
955 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
956 if (!state.isConflict())
959 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
960 PHINode *phi = cast<PHINode>(v);
961 unsigned NumPHIValues = phi->getNumIncomingValues();
962 for (unsigned i = 0; i < NumPHIValues; i++) {
963 Value *InVal = phi->getIncomingValue(i);
964 BasicBlock *InBB = phi->getIncomingBlock(i);
966 // If we've already seen InBB, add the same incoming value
967 // we added for it earlier. The IR verifier requires phi
968 // nodes with multiple entries from the same basic block
969 // to have the same incoming value for each of those
970 // entries. If we don't do this check here and basephi
971 // has a different type than base, we'll end up adding two
972 // bitcasts (and hence two distinct values) as incoming
973 // values for the same basic block.
975 int blockIndex = basephi->getBasicBlockIndex(InBB);
976 if (blockIndex != -1) {
977 Value *oldBase = basephi->getIncomingValue(blockIndex);
978 basephi->addIncoming(oldBase, InBB);
980 Value *base = findBaseOrBDV(InVal, cache);
981 if (!isKnownBaseResult(base)) {
982 // Either conflict or base.
983 assert(states.count(base));
984 base = states[base].getBase();
985 assert(base != nullptr && "unknown BDVState!");
988 // In essence this assert states: the only way two
989 // values incoming from the same basic block may be
990 // different is by being different bitcasts of the same
991 // value. A cleanup that remains TODO is changing
992 // findBaseOrBDV to return an llvm::Value of the correct
993 // type (and still remain pure). This will remove the
994 // need to add bitcasts.
995 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
996 "sanity -- findBaseOrBDV should be pure!");
1001 // Find either the defining value for the PHI or the normal base for
1003 Value *base = findBaseOrBDV(InVal, cache);
1004 if (!isKnownBaseResult(base)) {
1005 // Either conflict or base.
1006 assert(states.count(base));
1007 base = states[base].getBase();
1008 assert(base != nullptr && "unknown BDVState!");
1010 assert(base && "can't be null");
1011 // Must use original input BB since base may not be Instruction
1012 // The cast is needed since base traversal may strip away bitcasts
1013 if (base->getType() != basephi->getType()) {
1014 base = new BitCastInst(base, basephi->getType(), "cast",
1015 InBB->getTerminator());
1017 basephi->addIncoming(base, InBB);
1019 assert(basephi->getNumIncomingValues() == NumPHIValues);
1020 } else if (SelectInst *basesel = dyn_cast<SelectInst>(state.getBase())) {
1021 SelectInst *sel = cast<SelectInst>(v);
1022 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
1023 // something more safe and less hacky.
1024 for (int i = 1; i <= 2; i++) {
1025 Value *InVal = sel->getOperand(i);
1026 // Find either the defining value for the PHI or the normal base for
1028 Value *base = findBaseOrBDV(InVal, cache);
1029 if (!isKnownBaseResult(base)) {
1030 // Either conflict or base.
1031 assert(states.count(base));
1032 base = states[base].getBase();
1033 assert(base != nullptr && "unknown BDVState!");
1035 assert(base && "can't be null");
1036 // Must use original input BB since base may not be Instruction
1037 // The cast is needed since base traversal may strip away bitcasts
1038 if (base->getType() != basesel->getType()) {
1039 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
1041 basesel->setOperand(i, base);
1044 auto *BaseEE = cast<ExtractElementInst>(state.getBase());
1045 Value *InVal = cast<ExtractElementInst>(v)->getVectorOperand();
1046 Value *Base = findBaseOrBDV(InVal, cache);
1047 if (!isKnownBaseResult(Base)) {
1048 // Either conflict or base.
1049 assert(states.count(Base));
1050 Base = states[Base].getBase();
1051 assert(Base != nullptr && "unknown BDVState!");
1053 assert(Base && "can't be null");
1054 BaseEE->setOperand(0, Base);
1058 // Now that we're done with the algorithm, see if we can optimize the
1059 // results slightly by reducing the number of new instructions needed.
1060 // Arguably, this should be integrated into the algorithm above, but
1061 // doing as a post process step is easier to reason about for the moment.
1062 DenseMap<Value *, Value *> ReverseMap;
1063 SmallPtrSet<Instruction *, 16> NewInsts;
1064 SmallSetVector<AssertingVH<Instruction>, 16> Worklist;
1065 // Note: We need to visit the states in a deterministic order. We uses the
1066 // Keys we sorted above for this purpose. Note that we are papering over a
1067 // bigger problem with the algorithm above - it's visit order is not
1068 // deterministic. A larger change is needed to fix this.
1069 for (auto Key : Keys) {
1071 auto State = states[Key];
1072 Value *Base = State.getBase();
1074 assert(!isKnownBaseResult(V) && "why did it get added?");
1075 assert(isKnownBaseResult(Base) &&
1076 "must be something we 'know' is a base pointer");
1077 if (!State.isConflict())
1080 ReverseMap[Base] = V;
1081 if (auto *BaseI = dyn_cast<Instruction>(Base)) {
1082 NewInsts.insert(BaseI);
1083 Worklist.insert(BaseI);
1086 auto ReplaceBaseInstWith = [&](Value *BDV, Instruction *BaseI,
1087 Value *Replacement) {
1088 // Add users which are new instructions (excluding self references)
1089 for (User *U : BaseI->users())
1090 if (auto *UI = dyn_cast<Instruction>(U))
1091 if (NewInsts.count(UI) && UI != BaseI)
1092 Worklist.insert(UI);
1093 // Then do the actual replacement
1094 NewInsts.erase(BaseI);
1095 ReverseMap.erase(BaseI);
1096 BaseI->replaceAllUsesWith(Replacement);
1097 BaseI->eraseFromParent();
1098 assert(states.count(BDV));
1099 assert(states[BDV].isConflict() && states[BDV].getBase() == BaseI);
1100 states[BDV] = BDVState(BDVState::Conflict, Replacement);
1102 const DataLayout &DL = cast<Instruction>(def)->getModule()->getDataLayout();
1103 while (!Worklist.empty()) {
1104 Instruction *BaseI = Worklist.pop_back_val();
1105 assert(NewInsts.count(BaseI));
1106 Value *Bdv = ReverseMap[BaseI];
1107 if (auto *BdvI = dyn_cast<Instruction>(Bdv))
1108 if (BaseI->isIdenticalTo(BdvI)) {
1109 DEBUG(dbgs() << "Identical Base: " << *BaseI << "\n");
1110 ReplaceBaseInstWith(Bdv, BaseI, Bdv);
1113 if (Value *V = SimplifyInstruction(BaseI, DL)) {
1114 DEBUG(dbgs() << "Base " << *BaseI << " simplified to " << *V << "\n");
1115 ReplaceBaseInstWith(Bdv, BaseI, V);
1120 // Cache all of our results so we can cheaply reuse them
1121 // NOTE: This is actually two caches: one of the base defining value
1122 // relation and one of the base pointer relation! FIXME
1123 for (auto item : states) {
1124 Value *v = item.first;
1125 Value *base = item.second.getBase();
1128 std::string fromstr =
1129 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
1131 DEBUG(dbgs() << "Updating base value cache"
1132 << " for: " << (v->hasName() ? v->getName() : "")
1133 << " from: " << fromstr
1134 << " to: " << (base->hasName() ? base->getName() : "") << "\n");
1136 if (cache.count(v)) {
1137 // Once we transition from the BDV relation being store in the cache to
1138 // the base relation being stored, it must be stable
1139 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
1140 "base relation should be stable");
1144 assert(cache.find(def) != cache.end());
1148 // For a set of live pointers (base and/or derived), identify the base
1149 // pointer of the object which they are derived from. This routine will
1150 // mutate the IR graph as needed to make the 'base' pointer live at the
1151 // definition site of 'derived'. This ensures that any use of 'derived' can
1152 // also use 'base'. This may involve the insertion of a number of
1153 // additional PHI nodes.
1155 // preconditions: live is a set of pointer type Values
1157 // side effects: may insert PHI nodes into the existing CFG, will preserve
1158 // CFG, will not remove or mutate any existing nodes
1160 // post condition: PointerToBase contains one (derived, base) pair for every
1161 // pointer in live. Note that derived can be equal to base if the original
1162 // pointer was a base pointer.
1164 findBasePointers(const StatepointLiveSetTy &live,
1165 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
1166 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1167 // For the naming of values inserted to be deterministic - which makes for
1168 // much cleaner and more stable tests - we need to assign an order to the
1169 // live values. DenseSets do not provide a deterministic order across runs.
1170 SmallVector<Value *, 64> Temp;
1171 Temp.insert(Temp.end(), live.begin(), live.end());
1172 std::sort(Temp.begin(), Temp.end(), order_by_name);
1173 for (Value *ptr : Temp) {
1174 Value *base = findBasePointer(ptr, DVCache);
1175 assert(base && "failed to find base pointer");
1176 PointerToBase[ptr] = base;
1177 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1178 DT->dominates(cast<Instruction>(base)->getParent(),
1179 cast<Instruction>(ptr)->getParent())) &&
1180 "The base we found better dominate the derived pointer");
1182 // If you see this trip and like to live really dangerously, the code should
1183 // be correct, just with idioms the verifier can't handle. You can try
1184 // disabling the verifier at your own substantial risk.
1185 assert(!isa<ConstantPointerNull>(base) &&
1186 "the relocation code needs adjustment to handle the relocation of "
1187 "a null pointer constant without causing false positives in the "
1188 "safepoint ir verifier.");
1192 /// Find the required based pointers (and adjust the live set) for the given
1194 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1196 PartiallyConstructedSafepointRecord &result) {
1197 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
1198 findBasePointers(result.liveset, PointerToBase, &DT, DVCache);
1200 if (PrintBasePointers) {
1201 // Note: Need to print these in a stable order since this is checked in
1203 errs() << "Base Pairs (w/o Relocation):\n";
1204 SmallVector<Value *, 64> Temp;
1205 Temp.reserve(PointerToBase.size());
1206 for (auto Pair : PointerToBase) {
1207 Temp.push_back(Pair.first);
1209 std::sort(Temp.begin(), Temp.end(), order_by_name);
1210 for (Value *Ptr : Temp) {
1211 Value *Base = PointerToBase[Ptr];
1212 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
1217 result.PointerToBase = PointerToBase;
1220 /// Given an updated version of the dataflow liveness results, update the
1221 /// liveset and base pointer maps for the call site CS.
1222 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1224 PartiallyConstructedSafepointRecord &result);
1226 static void recomputeLiveInValues(
1227 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1228 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1229 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1230 // again. The old values are still live and will help it stabilize quickly.
1231 GCPtrLivenessData RevisedLivenessData;
1232 computeLiveInValues(DT, F, RevisedLivenessData);
1233 for (size_t i = 0; i < records.size(); i++) {
1234 struct PartiallyConstructedSafepointRecord &info = records[i];
1235 const CallSite &CS = toUpdate[i];
1236 recomputeLiveInValues(RevisedLivenessData, CS, info);
1240 // When inserting gc.relocate calls, we need to ensure there are no uses
1241 // of the original value between the gc.statepoint and the gc.relocate call.
1242 // One case which can arise is a phi node starting one of the successor blocks.
1243 // We also need to be able to insert the gc.relocates only on the path which
1244 // goes through the statepoint. We might need to split an edge to make this
1247 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1248 DominatorTree &DT) {
1249 BasicBlock *Ret = BB;
1250 if (!BB->getUniquePredecessor()) {
1251 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1254 // Now that 'ret' has unique predecessor we can safely remove all phi nodes
1256 FoldSingleEntryPHINodes(Ret);
1257 assert(!isa<PHINode>(Ret->begin()));
1259 // At this point, we can safely insert a gc.relocate as the first instruction
1260 // in Ret if needed.
1264 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1265 auto itr = std::find(livevec.begin(), livevec.end(), val);
1266 assert(livevec.end() != itr);
1267 size_t index = std::distance(livevec.begin(), itr);
1268 assert(index < livevec.size());
1272 // Create new attribute set containing only attributes which can be transferred
1273 // from original call to the safepoint.
1274 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1277 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1278 unsigned index = AS.getSlotIndex(Slot);
1280 if (index == AttributeSet::ReturnIndex ||
1281 index == AttributeSet::FunctionIndex) {
1283 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1285 Attribute attr = *it;
1287 // Do not allow certain attributes - just skip them
1288 // Safepoint can not be read only or read none.
1289 if (attr.hasAttribute(Attribute::ReadNone) ||
1290 attr.hasAttribute(Attribute::ReadOnly))
1293 ret = ret.addAttributes(
1294 AS.getContext(), index,
1295 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1299 // Just skip parameter attributes for now
1305 /// Helper function to place all gc relocates necessary for the given
1308 /// liveVariables - list of variables to be relocated.
1309 /// liveStart - index of the first live variable.
1310 /// basePtrs - base pointers.
1311 /// statepointToken - statepoint instruction to which relocates should be
1313 /// Builder - Llvm IR builder to be used to construct new calls.
1314 static void CreateGCRelocates(ArrayRef<llvm::Value *> LiveVariables,
1315 const int LiveStart,
1316 ArrayRef<llvm::Value *> BasePtrs,
1317 Instruction *StatepointToken,
1318 IRBuilder<> Builder) {
1319 if (LiveVariables.empty())
1322 // All gc_relocate are set to i8 addrspace(1)* type. We originally generated
1323 // unique declarations for each pointer type, but this proved problematic
1324 // because the intrinsic mangling code is incomplete and fragile. Since
1325 // we're moving towards a single unified pointer type anyways, we can just
1326 // cast everything to an i8* of the right address space. A bitcast is added
1327 // later to convert gc_relocate to the actual value's type.
1328 Module *M = StatepointToken->getModule();
1329 auto AS = cast<PointerType>(LiveVariables[0]->getType())->getAddressSpace();
1330 Type *Types[] = {Type::getInt8PtrTy(M->getContext(), AS)};
1331 Value *GCRelocateDecl =
1332 Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, Types);
1334 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1335 // Generate the gc.relocate call and save the result
1337 Builder.getInt32(LiveStart + find_index(LiveVariables, BasePtrs[i]));
1339 Builder.getInt32(LiveStart + find_index(LiveVariables, LiveVariables[i]));
1341 // only specify a debug name if we can give a useful one
1342 CallInst *Reloc = Builder.CreateCall(
1343 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1344 LiveVariables[i]->hasName() ? LiveVariables[i]->getName() + ".relocated"
1346 // Trick CodeGen into thinking there are lots of free registers at this
1348 Reloc->setCallingConv(CallingConv::Cold);
1353 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1354 const SmallVectorImpl<llvm::Value *> &basePtrs,
1355 const SmallVectorImpl<llvm::Value *> &liveVariables,
1357 PartiallyConstructedSafepointRecord &result) {
1358 assert(basePtrs.size() == liveVariables.size());
1359 assert(isStatepoint(CS) &&
1360 "This method expects to be rewriting a statepoint");
1362 BasicBlock *BB = CS.getInstruction()->getParent();
1364 Function *F = BB->getParent();
1365 assert(F && "must be set");
1366 Module *M = F->getParent();
1368 assert(M && "must be set");
1370 // We're not changing the function signature of the statepoint since the gc
1371 // arguments go into the var args section.
1372 Function *gc_statepoint_decl = CS.getCalledFunction();
1374 // Then go ahead and use the builder do actually do the inserts. We insert
1375 // immediately before the previous instruction under the assumption that all
1376 // arguments will be available here. We can't insert afterwards since we may
1377 // be replacing a terminator.
1378 Instruction *insertBefore = CS.getInstruction();
1379 IRBuilder<> Builder(insertBefore);
1380 // Copy all of the arguments from the original statepoint - this includes the
1381 // target, call args, and deopt args
1382 SmallVector<llvm::Value *, 64> args;
1383 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1384 // TODO: Clear the 'needs rewrite' flag
1386 // add all the pointers to be relocated (gc arguments)
1387 // Capture the start of the live variable list for use in the gc_relocates
1388 const int live_start = args.size();
1389 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1391 // Create the statepoint given all the arguments
1392 Instruction *token = nullptr;
1393 AttributeSet return_attributes;
1395 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1397 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1398 call->setTailCall(toReplace->isTailCall());
1399 call->setCallingConv(toReplace->getCallingConv());
1401 // Currently we will fail on parameter attributes and on certain
1402 // function attributes.
1403 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1404 // In case if we can handle this set of attributes - set up function attrs
1405 // directly on statepoint and return attrs later for gc_result intrinsic.
1406 call->setAttributes(new_attrs.getFnAttributes());
1407 return_attributes = new_attrs.getRetAttributes();
1411 // Put the following gc_result and gc_relocate calls immediately after the
1412 // the old call (which we're about to delete)
1413 BasicBlock::iterator next(toReplace);
1414 assert(BB->end() != next && "not a terminator, must have next");
1416 Instruction *IP = &*(next);
1417 Builder.SetInsertPoint(IP);
1418 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1421 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1423 // Insert the new invoke into the old block. We'll remove the old one in a
1424 // moment at which point this will become the new terminator for the
1426 InvokeInst *invoke = InvokeInst::Create(
1427 gc_statepoint_decl, toReplace->getNormalDest(),
1428 toReplace->getUnwindDest(), args, "statepoint_token", toReplace->getParent());
1429 invoke->setCallingConv(toReplace->getCallingConv());
1431 // Currently we will fail on parameter attributes and on certain
1432 // function attributes.
1433 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1434 // In case if we can handle this set of attributes - set up function attrs
1435 // directly on statepoint and return attrs later for gc_result intrinsic.
1436 invoke->setAttributes(new_attrs.getFnAttributes());
1437 return_attributes = new_attrs.getRetAttributes();
1441 // Generate gc relocates in exceptional path
1442 BasicBlock *unwindBlock = toReplace->getUnwindDest();
1443 assert(!isa<PHINode>(unwindBlock->begin()) &&
1444 unwindBlock->getUniquePredecessor() &&
1445 "can't safely insert in this block!");
1447 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1448 Builder.SetInsertPoint(IP);
1449 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1451 // Extract second element from landingpad return value. We will attach
1452 // exceptional gc relocates to it.
1453 const unsigned idx = 1;
1454 Instruction *exceptional_token =
1455 cast<Instruction>(Builder.CreateExtractValue(
1456 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1457 result.UnwindToken = exceptional_token;
1459 CreateGCRelocates(liveVariables, live_start, basePtrs,
1460 exceptional_token, Builder);
1462 // Generate gc relocates and returns for normal block
1463 BasicBlock *normalDest = toReplace->getNormalDest();
1464 assert(!isa<PHINode>(normalDest->begin()) &&
1465 normalDest->getUniquePredecessor() &&
1466 "can't safely insert in this block!");
1468 IP = &*(normalDest->getFirstInsertionPt());
1469 Builder.SetInsertPoint(IP);
1471 // gc relocates will be generated later as if it were regular call
1476 // Take the name of the original value call if it had one.
1477 token->takeName(CS.getInstruction());
1479 // The GCResult is already inserted, we just need to find it
1481 Instruction *toReplace = CS.getInstruction();
1482 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1483 "only valid use before rewrite is gc.result");
1484 assert(!toReplace->hasOneUse() ||
1485 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1488 // Update the gc.result of the original statepoint (if any) to use the newly
1489 // inserted statepoint. This is safe to do here since the token can't be
1490 // considered a live reference.
1491 CS.getInstruction()->replaceAllUsesWith(token);
1493 result.StatepointToken = token;
1495 // Second, create a gc.relocate for every live variable
1496 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1500 struct name_ordering {
1503 bool operator()(name_ordering const &a, name_ordering const &b) {
1504 return -1 == a.derived->getName().compare(b.derived->getName());
1508 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1509 SmallVectorImpl<Value *> &livevec) {
1510 assert(basevec.size() == livevec.size());
1512 SmallVector<name_ordering, 64> temp;
1513 for (size_t i = 0; i < basevec.size(); i++) {
1515 v.base = basevec[i];
1516 v.derived = livevec[i];
1519 std::sort(temp.begin(), temp.end(), name_ordering());
1520 for (size_t i = 0; i < basevec.size(); i++) {
1521 basevec[i] = temp[i].base;
1522 livevec[i] = temp[i].derived;
1526 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1527 // which make the relocations happening at this safepoint explicit.
1529 // WARNING: Does not do any fixup to adjust users of the original live
1530 // values. That's the callers responsibility.
1532 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1533 PartiallyConstructedSafepointRecord &result) {
1534 auto liveset = result.liveset;
1535 auto PointerToBase = result.PointerToBase;
1537 // Convert to vector for efficient cross referencing.
1538 SmallVector<Value *, 64> basevec, livevec;
1539 livevec.reserve(liveset.size());
1540 basevec.reserve(liveset.size());
1541 for (Value *L : liveset) {
1542 livevec.push_back(L);
1543 assert(PointerToBase.count(L));
1544 Value *base = PointerToBase[L];
1545 basevec.push_back(base);
1547 assert(livevec.size() == basevec.size());
1549 // To make the output IR slightly more stable (for use in diffs), ensure a
1550 // fixed order of the values in the safepoint (by sorting the value name).
1551 // The order is otherwise meaningless.
1552 stablize_order(basevec, livevec);
1554 // Do the actual rewriting and delete the old statepoint
1555 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1556 CS.getInstruction()->eraseFromParent();
1559 // Helper function for the relocationViaAlloca.
1560 // It receives iterator to the statepoint gc relocates and emits store to the
1562 // location (via allocaMap) for the each one of them.
1563 // Add visited values into the visitedLiveValues set we will later use them
1564 // for sanity check.
1566 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1567 DenseMap<Value *, Value *> &AllocaMap,
1568 DenseSet<Value *> &VisitedLiveValues) {
1570 for (User *U : GCRelocs) {
1571 if (!isa<IntrinsicInst>(U))
1574 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1576 // We only care about relocates
1577 if (RelocatedValue->getIntrinsicID() !=
1578 Intrinsic::experimental_gc_relocate) {
1582 GCRelocateOperands RelocateOperands(RelocatedValue);
1583 Value *OriginalValue =
1584 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1585 assert(AllocaMap.count(OriginalValue));
1586 Value *Alloca = AllocaMap[OriginalValue];
1588 // Emit store into the related alloca
1589 // All gc_relocate are i8 addrspace(1)* typed, and it must be bitcasted to
1590 // the correct type according to alloca.
1591 assert(RelocatedValue->getNextNode() && "Should always have one since it's not a terminator");
1592 IRBuilder<> Builder(RelocatedValue->getNextNode());
1593 Value *CastedRelocatedValue =
1594 Builder.CreateBitCast(RelocatedValue, cast<AllocaInst>(Alloca)->getAllocatedType(),
1595 RelocatedValue->hasName() ? RelocatedValue->getName() + ".casted" : "");
1597 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1598 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1601 VisitedLiveValues.insert(OriginalValue);
1606 // Helper function for the "relocationViaAlloca". Similar to the
1607 // "insertRelocationStores" but works for rematerialized values.
1609 insertRematerializationStores(
1610 RematerializedValueMapTy RematerializedValues,
1611 DenseMap<Value *, Value *> &AllocaMap,
1612 DenseSet<Value *> &VisitedLiveValues) {
1614 for (auto RematerializedValuePair: RematerializedValues) {
1615 Instruction *RematerializedValue = RematerializedValuePair.first;
1616 Value *OriginalValue = RematerializedValuePair.second;
1618 assert(AllocaMap.count(OriginalValue) &&
1619 "Can not find alloca for rematerialized value");
1620 Value *Alloca = AllocaMap[OriginalValue];
1622 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1623 Store->insertAfter(RematerializedValue);
1626 VisitedLiveValues.insert(OriginalValue);
1631 /// do all the relocation update via allocas and mem2reg
1632 static void relocationViaAlloca(
1633 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1634 ArrayRef<struct PartiallyConstructedSafepointRecord> Records) {
1636 // record initial number of (static) allocas; we'll check we have the same
1637 // number when we get done.
1638 int InitialAllocaNum = 0;
1639 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1641 if (isa<AllocaInst>(*I))
1645 // TODO-PERF: change data structures, reserve
1646 DenseMap<Value *, Value *> AllocaMap;
1647 SmallVector<AllocaInst *, 200> PromotableAllocas;
1648 // Used later to chack that we have enough allocas to store all values
1649 std::size_t NumRematerializedValues = 0;
1650 PromotableAllocas.reserve(Live.size());
1652 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1653 // "PromotableAllocas"
1654 auto emitAllocaFor = [&](Value *LiveValue) {
1655 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1656 F.getEntryBlock().getFirstNonPHI());
1657 AllocaMap[LiveValue] = Alloca;
1658 PromotableAllocas.push_back(Alloca);
1661 // emit alloca for each live gc pointer
1662 for (unsigned i = 0; i < Live.size(); i++) {
1663 emitAllocaFor(Live[i]);
1666 // emit allocas for rematerialized values
1667 for (size_t i = 0; i < Records.size(); i++) {
1668 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1670 for (auto RematerializedValuePair : Info.RematerializedValues) {
1671 Value *OriginalValue = RematerializedValuePair.second;
1672 if (AllocaMap.count(OriginalValue) != 0)
1675 emitAllocaFor(OriginalValue);
1676 ++NumRematerializedValues;
1680 // The next two loops are part of the same conceptual operation. We need to
1681 // insert a store to the alloca after the original def and at each
1682 // redefinition. We need to insert a load before each use. These are split
1683 // into distinct loops for performance reasons.
1685 // update gc pointer after each statepoint
1686 // either store a relocated value or null (if no relocated value found for
1687 // this gc pointer and it is not a gc_result)
1688 // this must happen before we update the statepoint with load of alloca
1689 // otherwise we lose the link between statepoint and old def
1690 for (size_t i = 0; i < Records.size(); i++) {
1691 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1692 Value *Statepoint = Info.StatepointToken;
1694 // This will be used for consistency check
1695 DenseSet<Value *> VisitedLiveValues;
1697 // Insert stores for normal statepoint gc relocates
1698 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1700 // In case if it was invoke statepoint
1701 // we will insert stores for exceptional path gc relocates.
1702 if (isa<InvokeInst>(Statepoint)) {
1703 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1707 // Do similar thing with rematerialized values
1708 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1711 if (ClobberNonLive) {
1712 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1713 // the gc.statepoint. This will turn some subtle GC problems into
1714 // slightly easier to debug SEGVs. Note that on large IR files with
1715 // lots of gc.statepoints this is extremely costly both memory and time
1717 SmallVector<AllocaInst *, 64> ToClobber;
1718 for (auto Pair : AllocaMap) {
1719 Value *Def = Pair.first;
1720 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1722 // This value was relocated
1723 if (VisitedLiveValues.count(Def)) {
1726 ToClobber.push_back(Alloca);
1729 auto InsertClobbersAt = [&](Instruction *IP) {
1730 for (auto *AI : ToClobber) {
1731 auto AIType = cast<PointerType>(AI->getType());
1732 auto PT = cast<PointerType>(AIType->getElementType());
1733 Constant *CPN = ConstantPointerNull::get(PT);
1734 StoreInst *Store = new StoreInst(CPN, AI);
1735 Store->insertBefore(IP);
1739 // Insert the clobbering stores. These may get intermixed with the
1740 // gc.results and gc.relocates, but that's fine.
1741 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1742 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1743 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1745 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1747 InsertClobbersAt(Next);
1751 // update use with load allocas and add store for gc_relocated
1752 for (auto Pair : AllocaMap) {
1753 Value *Def = Pair.first;
1754 Value *Alloca = Pair.second;
1756 // we pre-record the uses of allocas so that we dont have to worry about
1758 // that change the user information.
1759 SmallVector<Instruction *, 20> Uses;
1760 // PERF: trade a linear scan for repeated reallocation
1761 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1762 for (User *U : Def->users()) {
1763 if (!isa<ConstantExpr>(U)) {
1764 // If the def has a ConstantExpr use, then the def is either a
1765 // ConstantExpr use itself or null. In either case
1766 // (recursively in the first, directly in the second), the oop
1767 // it is ultimately dependent on is null and this particular
1768 // use does not need to be fixed up.
1769 Uses.push_back(cast<Instruction>(U));
1773 std::sort(Uses.begin(), Uses.end());
1774 auto Last = std::unique(Uses.begin(), Uses.end());
1775 Uses.erase(Last, Uses.end());
1777 for (Instruction *Use : Uses) {
1778 if (isa<PHINode>(Use)) {
1779 PHINode *Phi = cast<PHINode>(Use);
1780 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1781 if (Def == Phi->getIncomingValue(i)) {
1782 LoadInst *Load = new LoadInst(
1783 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1784 Phi->setIncomingValue(i, Load);
1788 LoadInst *Load = new LoadInst(Alloca, "", Use);
1789 Use->replaceUsesOfWith(Def, Load);
1793 // emit store for the initial gc value
1794 // store must be inserted after load, otherwise store will be in alloca's
1795 // use list and an extra load will be inserted before it
1796 StoreInst *Store = new StoreInst(Def, Alloca);
1797 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1798 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1799 // InvokeInst is a TerminatorInst so the store need to be inserted
1800 // into its normal destination block.
1801 BasicBlock *NormalDest = Invoke->getNormalDest();
1802 Store->insertBefore(NormalDest->getFirstNonPHI());
1804 assert(!Inst->isTerminator() &&
1805 "The only TerminatorInst that can produce a value is "
1806 "InvokeInst which is handled above.");
1807 Store->insertAfter(Inst);
1810 assert(isa<Argument>(Def));
1811 Store->insertAfter(cast<Instruction>(Alloca));
1815 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1816 "we must have the same allocas with lives");
1817 if (!PromotableAllocas.empty()) {
1818 // apply mem2reg to promote alloca to SSA
1819 PromoteMemToReg(PromotableAllocas, DT);
1823 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1825 if (isa<AllocaInst>(*I))
1827 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1831 /// Implement a unique function which doesn't require we sort the input
1832 /// vector. Doing so has the effect of changing the output of a couple of
1833 /// tests in ways which make them less useful in testing fused safepoints.
1834 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1835 SmallSet<T, 8> Seen;
1836 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1837 return !Seen.insert(V).second;
1841 /// Insert holders so that each Value is obviously live through the entire
1842 /// lifetime of the call.
1843 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1844 SmallVectorImpl<CallInst *> &Holders) {
1846 // No values to hold live, might as well not insert the empty holder
1849 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1850 // Use a dummy vararg function to actually hold the values live
1851 Function *Func = cast<Function>(M->getOrInsertFunction(
1852 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1854 // For call safepoints insert dummy calls right after safepoint
1855 BasicBlock::iterator Next(CS.getInstruction());
1857 Holders.push_back(CallInst::Create(Func, Values, "", Next));
1860 // For invoke safepooints insert dummy calls both in normal and
1861 // exceptional destination blocks
1862 auto *II = cast<InvokeInst>(CS.getInstruction());
1863 Holders.push_back(CallInst::Create(
1864 Func, Values, "", II->getNormalDest()->getFirstInsertionPt()));
1865 Holders.push_back(CallInst::Create(
1866 Func, Values, "", II->getUnwindDest()->getFirstInsertionPt()));
1869 static void findLiveReferences(
1870 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1871 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1872 GCPtrLivenessData OriginalLivenessData;
1873 computeLiveInValues(DT, F, OriginalLivenessData);
1874 for (size_t i = 0; i < records.size(); i++) {
1875 struct PartiallyConstructedSafepointRecord &info = records[i];
1876 const CallSite &CS = toUpdate[i];
1877 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1881 /// Remove any vector of pointers from the liveset by scalarizing them over the
1882 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1883 /// would be preferable to include the vector in the statepoint itself, but
1884 /// the lowering code currently does not handle that. Extending it would be
1885 /// slightly non-trivial since it requires a format change. Given how rare
1886 /// such cases are (for the moment?) scalarizing is an acceptable compromise.
1887 static void splitVectorValues(Instruction *StatepointInst,
1888 StatepointLiveSetTy &LiveSet,
1889 DenseMap<Value *, Value *>& PointerToBase,
1890 DominatorTree &DT) {
1891 SmallVector<Value *, 16> ToSplit;
1892 for (Value *V : LiveSet)
1893 if (isa<VectorType>(V->getType()))
1894 ToSplit.push_back(V);
1896 if (ToSplit.empty())
1899 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1901 Function &F = *(StatepointInst->getParent()->getParent());
1903 DenseMap<Value *, AllocaInst *> AllocaMap;
1904 // First is normal return, second is exceptional return (invoke only)
1905 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1906 for (Value *V : ToSplit) {
1907 AllocaInst *Alloca =
1908 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1909 AllocaMap[V] = Alloca;
1911 VectorType *VT = cast<VectorType>(V->getType());
1912 IRBuilder<> Builder(StatepointInst);
1913 SmallVector<Value *, 16> Elements;
1914 for (unsigned i = 0; i < VT->getNumElements(); i++)
1915 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1916 ElementMapping[V] = Elements;
1918 auto InsertVectorReform = [&](Instruction *IP) {
1919 Builder.SetInsertPoint(IP);
1920 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1921 Value *ResultVec = UndefValue::get(VT);
1922 for (unsigned i = 0; i < VT->getNumElements(); i++)
1923 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1924 Builder.getInt32(i));
1928 if (isa<CallInst>(StatepointInst)) {
1929 BasicBlock::iterator Next(StatepointInst);
1931 Instruction *IP = &*(Next);
1932 Replacements[V].first = InsertVectorReform(IP);
1933 Replacements[V].second = nullptr;
1935 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1936 // We've already normalized - check that we don't have shared destination
1938 BasicBlock *NormalDest = Invoke->getNormalDest();
1939 assert(!isa<PHINode>(NormalDest->begin()));
1940 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1941 assert(!isa<PHINode>(UnwindDest->begin()));
1942 // Insert insert element sequences in both successors
1943 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1944 Replacements[V].first = InsertVectorReform(IP);
1945 IP = &*(UnwindDest->getFirstInsertionPt());
1946 Replacements[V].second = InsertVectorReform(IP);
1950 for (Value *V : ToSplit) {
1951 AllocaInst *Alloca = AllocaMap[V];
1953 // Capture all users before we start mutating use lists
1954 SmallVector<Instruction *, 16> Users;
1955 for (User *U : V->users())
1956 Users.push_back(cast<Instruction>(U));
1958 for (Instruction *I : Users) {
1959 if (auto Phi = dyn_cast<PHINode>(I)) {
1960 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1961 if (V == Phi->getIncomingValue(i)) {
1962 LoadInst *Load = new LoadInst(
1963 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1964 Phi->setIncomingValue(i, Load);
1967 LoadInst *Load = new LoadInst(Alloca, "", I);
1968 I->replaceUsesOfWith(V, Load);
1972 // Store the original value and the replacement value into the alloca
1973 StoreInst *Store = new StoreInst(V, Alloca);
1974 if (auto I = dyn_cast<Instruction>(V))
1975 Store->insertAfter(I);
1977 Store->insertAfter(Alloca);
1979 // Normal return for invoke, or call return
1980 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1981 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1982 // Unwind return for invoke only
1983 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1985 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1988 // apply mem2reg to promote alloca to SSA
1989 SmallVector<AllocaInst *, 16> Allocas;
1990 for (Value *V : ToSplit)
1991 Allocas.push_back(AllocaMap[V]);
1992 PromoteMemToReg(Allocas, DT);
1994 // Update our tracking of live pointers and base mappings to account for the
1995 // changes we just made.
1996 for (Value *V : ToSplit) {
1997 auto &Elements = ElementMapping[V];
2000 LiveSet.insert(Elements.begin(), Elements.end());
2001 // We need to update the base mapping as well.
2002 assert(PointerToBase.count(V));
2003 Value *OldBase = PointerToBase[V];
2004 auto &BaseElements = ElementMapping[OldBase];
2005 PointerToBase.erase(V);
2006 assert(Elements.size() == BaseElements.size());
2007 for (unsigned i = 0; i < Elements.size(); i++) {
2008 Value *Elem = Elements[i];
2009 PointerToBase[Elem] = BaseElements[i];
2014 // Helper function for the "rematerializeLiveValues". It walks use chain
2015 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
2016 // values are visited (currently it is GEP's and casts). Returns true if it
2017 // successfully reached "BaseValue" and false otherwise.
2018 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
2020 static bool findRematerializableChainToBasePointer(
2021 SmallVectorImpl<Instruction*> &ChainToBase,
2022 Value *CurrentValue, Value *BaseValue) {
2024 // We have found a base value
2025 if (CurrentValue == BaseValue) {
2029 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
2030 ChainToBase.push_back(GEP);
2031 return findRematerializableChainToBasePointer(ChainToBase,
2032 GEP->getPointerOperand(),
2036 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
2037 Value *Def = CI->stripPointerCasts();
2039 // This two checks are basically similar. First one is here for the
2040 // consistency with findBasePointers logic.
2041 assert(!isa<CastInst>(Def) && "not a pointer cast found");
2042 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
2045 ChainToBase.push_back(CI);
2046 return findRematerializableChainToBasePointer(ChainToBase, Def, BaseValue);
2049 // Not supported instruction in the chain
2053 // Helper function for the "rematerializeLiveValues". Compute cost of the use
2054 // chain we are going to rematerialize.
2056 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
2057 TargetTransformInfo &TTI) {
2060 for (Instruction *Instr : Chain) {
2061 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
2062 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
2063 "non noop cast is found during rematerialization");
2065 Type *SrcTy = CI->getOperand(0)->getType();
2066 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
2068 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
2069 // Cost of the address calculation
2070 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
2071 Cost += TTI.getAddressComputationCost(ValTy);
2073 // And cost of the GEP itself
2074 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
2075 // allowed for the external usage)
2076 if (!GEP->hasAllConstantIndices())
2080 llvm_unreachable("unsupported instruciton type during rematerialization");
2087 // From the statepoint liveset pick values that are cheaper to recompute then to
2088 // relocate. Remove this values from the liveset, rematerialize them after
2089 // statepoint and record them in "Info" structure. Note that similar to
2090 // relocated values we don't do any user adjustments here.
2091 static void rematerializeLiveValues(CallSite CS,
2092 PartiallyConstructedSafepointRecord &Info,
2093 TargetTransformInfo &TTI) {
2094 const unsigned int ChainLengthThreshold = 10;
2096 // Record values we are going to delete from this statepoint live set.
2097 // We can not di this in following loop due to iterator invalidation.
2098 SmallVector<Value *, 32> LiveValuesToBeDeleted;
2100 for (Value *LiveValue: Info.liveset) {
2101 // For each live pointer find it's defining chain
2102 SmallVector<Instruction *, 3> ChainToBase;
2103 assert(Info.PointerToBase.count(LiveValue));
2105 findRematerializableChainToBasePointer(ChainToBase,
2107 Info.PointerToBase[LiveValue]);
2108 // Nothing to do, or chain is too long
2110 ChainToBase.size() == 0 ||
2111 ChainToBase.size() > ChainLengthThreshold)
2114 // Compute cost of this chain
2115 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
2116 // TODO: We can also account for cases when we will be able to remove some
2117 // of the rematerialized values by later optimization passes. I.e if
2118 // we rematerialized several intersecting chains. Or if original values
2119 // don't have any uses besides this statepoint.
2121 // For invokes we need to rematerialize each chain twice - for normal and
2122 // for unwind basic blocks. Model this by multiplying cost by two.
2123 if (CS.isInvoke()) {
2126 // If it's too expensive - skip it
2127 if (Cost >= RematerializationThreshold)
2130 // Remove value from the live set
2131 LiveValuesToBeDeleted.push_back(LiveValue);
2133 // Clone instructions and record them inside "Info" structure
2135 // Walk backwards to visit top-most instructions first
2136 std::reverse(ChainToBase.begin(), ChainToBase.end());
2138 // Utility function which clones all instructions from "ChainToBase"
2139 // and inserts them before "InsertBefore". Returns rematerialized value
2140 // which should be used after statepoint.
2141 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
2142 Instruction *LastClonedValue = nullptr;
2143 Instruction *LastValue = nullptr;
2144 for (Instruction *Instr: ChainToBase) {
2145 // Only GEP's and casts are suported as we need to be careful to not
2146 // introduce any new uses of pointers not in the liveset.
2147 // Note that it's fine to introduce new uses of pointers which were
2148 // otherwise not used after this statepoint.
2149 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2151 Instruction *ClonedValue = Instr->clone();
2152 ClonedValue->insertBefore(InsertBefore);
2153 ClonedValue->setName(Instr->getName() + ".remat");
2155 // If it is not first instruction in the chain then it uses previously
2156 // cloned value. We should update it to use cloned value.
2157 if (LastClonedValue) {
2159 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2161 // Assert that cloned instruction does not use any instructions from
2162 // this chain other than LastClonedValue
2163 for (auto OpValue : ClonedValue->operand_values()) {
2164 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2165 ChainToBase.end() &&
2166 "incorrect use in rematerialization chain");
2171 LastClonedValue = ClonedValue;
2174 assert(LastClonedValue);
2175 return LastClonedValue;
2178 // Different cases for calls and invokes. For invokes we need to clone
2179 // instructions both on normal and unwind path.
2181 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2182 assert(InsertBefore);
2183 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2184 Info.RematerializedValues[RematerializedValue] = LiveValue;
2186 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2188 Instruction *NormalInsertBefore =
2189 Invoke->getNormalDest()->getFirstInsertionPt();
2190 Instruction *UnwindInsertBefore =
2191 Invoke->getUnwindDest()->getFirstInsertionPt();
2193 Instruction *NormalRematerializedValue =
2194 rematerializeChain(NormalInsertBefore);
2195 Instruction *UnwindRematerializedValue =
2196 rematerializeChain(UnwindInsertBefore);
2198 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2199 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2203 // Remove rematerializaed values from the live set
2204 for (auto LiveValue: LiveValuesToBeDeleted) {
2205 Info.liveset.erase(LiveValue);
2209 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
2210 SmallVectorImpl<CallSite> &toUpdate) {
2212 // sanity check the input
2213 std::set<CallSite> uniqued;
2214 uniqued.insert(toUpdate.begin(), toUpdate.end());
2215 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
2217 for (size_t i = 0; i < toUpdate.size(); i++) {
2218 CallSite &CS = toUpdate[i];
2219 assert(CS.getInstruction()->getParent()->getParent() == &F);
2220 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
2224 // When inserting gc.relocates for invokes, we need to be able to insert at
2225 // the top of the successor blocks. See the comment on
2226 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2227 // may restructure the CFG.
2228 for (CallSite CS : toUpdate) {
2231 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
2232 normalizeForInvokeSafepoint(invoke->getNormalDest(), invoke->getParent(),
2234 normalizeForInvokeSafepoint(invoke->getUnwindDest(), invoke->getParent(),
2238 // A list of dummy calls added to the IR to keep various values obviously
2239 // live in the IR. We'll remove all of these when done.
2240 SmallVector<CallInst *, 64> holders;
2242 // Insert a dummy call with all of the arguments to the vm_state we'll need
2243 // for the actual safepoint insertion. This ensures reference arguments in
2244 // the deopt argument list are considered live through the safepoint (and
2245 // thus makes sure they get relocated.)
2246 for (size_t i = 0; i < toUpdate.size(); i++) {
2247 CallSite &CS = toUpdate[i];
2248 Statepoint StatepointCS(CS);
2250 SmallVector<Value *, 64> DeoptValues;
2251 for (Use &U : StatepointCS.vm_state_args()) {
2252 Value *Arg = cast<Value>(&U);
2253 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2254 "support for FCA unimplemented");
2255 if (isHandledGCPointerType(Arg->getType()))
2256 DeoptValues.push_back(Arg);
2258 insertUseHolderAfter(CS, DeoptValues, holders);
2261 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
2262 records.reserve(toUpdate.size());
2263 for (size_t i = 0; i < toUpdate.size(); i++) {
2264 struct PartiallyConstructedSafepointRecord info;
2265 records.push_back(info);
2267 assert(records.size() == toUpdate.size());
2269 // A) Identify all gc pointers which are statically live at the given call
2271 findLiveReferences(F, DT, P, toUpdate, records);
2273 // B) Find the base pointers for each live pointer
2274 /* scope for caching */ {
2275 // Cache the 'defining value' relation used in the computation and
2276 // insertion of base phis and selects. This ensures that we don't insert
2277 // large numbers of duplicate base_phis.
2278 DefiningValueMapTy DVCache;
2280 for (size_t i = 0; i < records.size(); i++) {
2281 struct PartiallyConstructedSafepointRecord &info = records[i];
2282 CallSite &CS = toUpdate[i];
2283 findBasePointers(DT, DVCache, CS, info);
2285 } // end of cache scope
2287 // The base phi insertion logic (for any safepoint) may have inserted new
2288 // instructions which are now live at some safepoint. The simplest such
2291 // phi a <-- will be a new base_phi here
2292 // safepoint 1 <-- that needs to be live here
2296 // We insert some dummy calls after each safepoint to definitely hold live
2297 // the base pointers which were identified for that safepoint. We'll then
2298 // ask liveness for _every_ base inserted to see what is now live. Then we
2299 // remove the dummy calls.
2300 holders.reserve(holders.size() + records.size());
2301 for (size_t i = 0; i < records.size(); i++) {
2302 struct PartiallyConstructedSafepointRecord &info = records[i];
2303 CallSite &CS = toUpdate[i];
2305 SmallVector<Value *, 128> Bases;
2306 for (auto Pair : info.PointerToBase) {
2307 Bases.push_back(Pair.second);
2309 insertUseHolderAfter(CS, Bases, holders);
2312 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2313 // need to rerun liveness. We may *also* have inserted new defs, but that's
2314 // not the key issue.
2315 recomputeLiveInValues(F, DT, P, toUpdate, records);
2317 if (PrintBasePointers) {
2318 for (size_t i = 0; i < records.size(); i++) {
2319 struct PartiallyConstructedSafepointRecord &info = records[i];
2320 errs() << "Base Pairs: (w/Relocation)\n";
2321 for (auto Pair : info.PointerToBase) {
2322 errs() << " derived %" << Pair.first->getName() << " base %"
2323 << Pair.second->getName() << "\n";
2327 for (size_t i = 0; i < holders.size(); i++) {
2328 holders[i]->eraseFromParent();
2329 holders[i] = nullptr;
2333 // Do a limited scalarization of any live at safepoint vector values which
2334 // contain pointers. This enables this pass to run after vectorization at
2335 // the cost of some possible performance loss. TODO: it would be nice to
2336 // natively support vectors all the way through the backend so we don't need
2337 // to scalarize here.
2338 for (size_t i = 0; i < records.size(); i++) {
2339 struct PartiallyConstructedSafepointRecord &info = records[i];
2340 Instruction *statepoint = toUpdate[i].getInstruction();
2341 splitVectorValues(cast<Instruction>(statepoint), info.liveset,
2342 info.PointerToBase, DT);
2345 // In order to reduce live set of statepoint we might choose to rematerialize
2346 // some values instead of relocating them. This is purely an optimization and
2347 // does not influence correctness.
2348 TargetTransformInfo &TTI =
2349 P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2351 for (size_t i = 0; i < records.size(); i++) {
2352 struct PartiallyConstructedSafepointRecord &info = records[i];
2353 CallSite &CS = toUpdate[i];
2355 rematerializeLiveValues(CS, info, TTI);
2358 // Now run through and replace the existing statepoints with new ones with
2359 // the live variables listed. We do not yet update uses of the values being
2360 // relocated. We have references to live variables that need to
2361 // survive to the last iteration of this loop. (By construction, the
2362 // previous statepoint can not be a live variable, thus we can and remove
2363 // the old statepoint calls as we go.)
2364 for (size_t i = 0; i < records.size(); i++) {
2365 struct PartiallyConstructedSafepointRecord &info = records[i];
2366 CallSite &CS = toUpdate[i];
2367 makeStatepointExplicit(DT, CS, P, info);
2369 toUpdate.clear(); // prevent accident use of invalid CallSites
2371 // Do all the fixups of the original live variables to their relocated selves
2372 SmallVector<Value *, 128> live;
2373 for (size_t i = 0; i < records.size(); i++) {
2374 struct PartiallyConstructedSafepointRecord &info = records[i];
2375 // We can't simply save the live set from the original insertion. One of
2376 // the live values might be the result of a call which needs a safepoint.
2377 // That Value* no longer exists and we need to use the new gc_result.
2378 // Thankfully, the liveset is embedded in the statepoint (and updated), so
2379 // we just grab that.
2380 Statepoint statepoint(info.StatepointToken);
2381 live.insert(live.end(), statepoint.gc_args_begin(),
2382 statepoint.gc_args_end());
2384 // Do some basic sanity checks on our liveness results before performing
2385 // relocation. Relocation can and will turn mistakes in liveness results
2386 // into non-sensical code which is must harder to debug.
2387 // TODO: It would be nice to test consistency as well
2388 assert(DT.isReachableFromEntry(info.StatepointToken->getParent()) &&
2389 "statepoint must be reachable or liveness is meaningless");
2390 for (Value *V : statepoint.gc_args()) {
2391 if (!isa<Instruction>(V))
2392 // Non-instruction values trivial dominate all possible uses
2394 auto LiveInst = cast<Instruction>(V);
2395 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2396 "unreachable values should never be live");
2397 assert(DT.dominates(LiveInst, info.StatepointToken) &&
2398 "basic SSA liveness expectation violated by liveness analysis");
2402 unique_unsorted(live);
2406 for (auto ptr : live) {
2407 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
2411 relocationViaAlloca(F, DT, live, records);
2412 return !records.empty();
2415 // Handles both return values and arguments for Functions and CallSites.
2416 template <typename AttrHolder>
2417 static void RemoveDerefAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2420 if (AH.getDereferenceableBytes(Index))
2421 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2422 AH.getDereferenceableBytes(Index)));
2423 if (AH.getDereferenceableOrNullBytes(Index))
2424 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2425 AH.getDereferenceableOrNullBytes(Index)));
2428 AH.setAttributes(AH.getAttributes().removeAttributes(
2429 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2433 RewriteStatepointsForGC::stripDereferenceabilityInfoFromPrototype(Function &F) {
2434 LLVMContext &Ctx = F.getContext();
2436 for (Argument &A : F.args())
2437 if (isa<PointerType>(A.getType()))
2438 RemoveDerefAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2440 if (isa<PointerType>(F.getReturnType()))
2441 RemoveDerefAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2444 void RewriteStatepointsForGC::stripDereferenceabilityInfoFromBody(Function &F) {
2448 LLVMContext &Ctx = F.getContext();
2449 MDBuilder Builder(Ctx);
2451 for (Instruction &I : instructions(F)) {
2452 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2453 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2454 bool IsImmutableTBAA =
2455 MD->getNumOperands() == 4 &&
2456 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2458 if (!IsImmutableTBAA)
2459 continue; // no work to do, MD_tbaa is already marked mutable
2461 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2462 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2464 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2466 MDNode *MutableTBAA =
2467 Builder.createTBAAStructTagNode(Base, Access, Offset);
2468 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2471 if (CallSite CS = CallSite(&I)) {
2472 for (int i = 0, e = CS.arg_size(); i != e; i++)
2473 if (isa<PointerType>(CS.getArgument(i)->getType()))
2474 RemoveDerefAttrAtIndex(Ctx, CS, i + 1);
2475 if (isa<PointerType>(CS.getType()))
2476 RemoveDerefAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2481 /// Returns true if this function should be rewritten by this pass. The main
2482 /// point of this function is as an extension point for custom logic.
2483 static bool shouldRewriteStatepointsIn(Function &F) {
2484 // TODO: This should check the GCStrategy
2486 const char *FunctionGCName = F.getGC();
2487 const StringRef StatepointExampleName("statepoint-example");
2488 const StringRef CoreCLRName("coreclr");
2489 return (StatepointExampleName == FunctionGCName) ||
2490 (CoreCLRName == FunctionGCName);
2495 void RewriteStatepointsForGC::stripDereferenceabilityInfo(Module &M) {
2497 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2501 for (Function &F : M)
2502 stripDereferenceabilityInfoFromPrototype(F);
2504 for (Function &F : M)
2505 stripDereferenceabilityInfoFromBody(F);
2508 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2509 // Nothing to do for declarations.
2510 if (F.isDeclaration() || F.empty())
2513 // Policy choice says not to rewrite - the most common reason is that we're
2514 // compiling code without a GCStrategy.
2515 if (!shouldRewriteStatepointsIn(F))
2518 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2520 // Gather all the statepoints which need rewritten. Be careful to only
2521 // consider those in reachable code since we need to ask dominance queries
2522 // when rewriting. We'll delete the unreachable ones in a moment.
2523 SmallVector<CallSite, 64> ParsePointNeeded;
2524 bool HasUnreachableStatepoint = false;
2525 for (Instruction &I : instructions(F)) {
2526 // TODO: only the ones with the flag set!
2527 if (isStatepoint(I)) {
2528 if (DT.isReachableFromEntry(I.getParent()))
2529 ParsePointNeeded.push_back(CallSite(&I));
2531 HasUnreachableStatepoint = true;
2535 bool MadeChange = false;
2537 // Delete any unreachable statepoints so that we don't have unrewritten
2538 // statepoints surviving this pass. This makes testing easier and the
2539 // resulting IR less confusing to human readers. Rather than be fancy, we
2540 // just reuse a utility function which removes the unreachable blocks.
2541 if (HasUnreachableStatepoint)
2542 MadeChange |= removeUnreachableBlocks(F);
2544 // Return early if no work to do.
2545 if (ParsePointNeeded.empty())
2548 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2549 // These are created by LCSSA. They have the effect of increasing the size
2550 // of liveness sets for no good reason. It may be harder to do this post
2551 // insertion since relocations and base phis can confuse things.
2552 for (BasicBlock &BB : F)
2553 if (BB.getUniquePredecessor()) {
2555 FoldSingleEntryPHINodes(&BB);
2558 // Before we start introducing relocations, we want to tweak the IR a bit to
2559 // avoid unfortunate code generation effects. The main example is that we
2560 // want to try to make sure the comparison feeding a branch is after any
2561 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2562 // values feeding a branch after relocation. This is semantically correct,
2563 // but results in extra register pressure since both the pre-relocation and
2564 // post-relocation copies must be available in registers. For code without
2565 // relocations this is handled elsewhere, but teaching the scheduler to
2566 // reverse the transform we're about to do would be slightly complex.
2567 // Note: This may extend the live range of the inputs to the icmp and thus
2568 // increase the liveset of any statepoint we move over. This is profitable
2569 // as long as all statepoints are in rare blocks. If we had in-register
2570 // lowering for live values this would be a much safer transform.
2571 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2572 if (auto *BI = dyn_cast<BranchInst>(TI))
2573 if (BI->isConditional())
2574 return dyn_cast<Instruction>(BI->getCondition());
2575 // TODO: Extend this to handle switches
2578 for (BasicBlock &BB : F) {
2579 TerminatorInst *TI = BB.getTerminator();
2580 if (auto *Cond = getConditionInst(TI))
2581 // TODO: Handle more than just ICmps here. We should be able to move
2582 // most instructions without side effects or memory access.
2583 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2585 Cond->moveBefore(TI);
2589 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
2593 // liveness computation via standard dataflow
2594 // -------------------------------------------------------------------
2596 // TODO: Consider using bitvectors for liveness, the set of potentially
2597 // interesting values should be small and easy to pre-compute.
2599 /// Compute the live-in set for the location rbegin starting from
2600 /// the live-out set of the basic block
2601 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2602 BasicBlock::reverse_iterator rend,
2603 DenseSet<Value *> &LiveTmp) {
2605 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2606 Instruction *I = &*ritr;
2608 // KILL/Def - Remove this definition from LiveIn
2611 // Don't consider *uses* in PHI nodes, we handle their contribution to
2612 // predecessor blocks when we seed the LiveOut sets
2613 if (isa<PHINode>(I))
2616 // USE - Add to the LiveIn set for this instruction
2617 for (Value *V : I->operands()) {
2618 assert(!isUnhandledGCPointerType(V->getType()) &&
2619 "support for FCA unimplemented");
2620 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2621 // The choice to exclude all things constant here is slightly subtle.
2622 // There are two independent reasons:
2623 // - We assume that things which are constant (from LLVM's definition)
2624 // do not move at runtime. For example, the address of a global
2625 // variable is fixed, even though it's contents may not be.
2626 // - Second, we can't disallow arbitrary inttoptr constants even
2627 // if the language frontend does. Optimization passes are free to
2628 // locally exploit facts without respect to global reachability. This
2629 // can create sections of code which are dynamically unreachable and
2630 // contain just about anything. (see constants.ll in tests)
2637 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2639 for (BasicBlock *Succ : successors(BB)) {
2640 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2641 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2642 PHINode *Phi = cast<PHINode>(&*I);
2643 Value *V = Phi->getIncomingValueForBlock(BB);
2644 assert(!isUnhandledGCPointerType(V->getType()) &&
2645 "support for FCA unimplemented");
2646 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2653 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2654 DenseSet<Value *> KillSet;
2655 for (Instruction &I : *BB)
2656 if (isHandledGCPointerType(I.getType()))
2662 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2663 /// sanity check for the liveness computation.
2664 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2665 TerminatorInst *TI, bool TermOkay = false) {
2666 for (Value *V : Live) {
2667 if (auto *I = dyn_cast<Instruction>(V)) {
2668 // The terminator can be a member of the LiveOut set. LLVM's definition
2669 // of instruction dominance states that V does not dominate itself. As
2670 // such, we need to special case this to allow it.
2671 if (TermOkay && TI == I)
2673 assert(DT.dominates(I, TI) &&
2674 "basic SSA liveness expectation violated by liveness analysis");
2679 /// Check that all the liveness sets used during the computation of liveness
2680 /// obey basic SSA properties. This is useful for finding cases where we miss
2682 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2684 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2685 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2686 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2690 static void computeLiveInValues(DominatorTree &DT, Function &F,
2691 GCPtrLivenessData &Data) {
2693 SmallSetVector<BasicBlock *, 200> Worklist;
2694 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2695 // We use a SetVector so that we don't have duplicates in the worklist.
2696 Worklist.insert(pred_begin(BB), pred_end(BB));
2698 auto NextItem = [&]() {
2699 BasicBlock *BB = Worklist.back();
2700 Worklist.pop_back();
2704 // Seed the liveness for each individual block
2705 for (BasicBlock &BB : F) {
2706 Data.KillSet[&BB] = computeKillSet(&BB);
2707 Data.LiveSet[&BB].clear();
2708 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2711 for (Value *Kill : Data.KillSet[&BB])
2712 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2715 Data.LiveOut[&BB] = DenseSet<Value *>();
2716 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2717 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2718 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2719 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2720 if (!Data.LiveIn[&BB].empty())
2721 AddPredsToWorklist(&BB);
2724 // Propagate that liveness until stable
2725 while (!Worklist.empty()) {
2726 BasicBlock *BB = NextItem();
2728 // Compute our new liveout set, then exit early if it hasn't changed
2729 // despite the contribution of our successor.
2730 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2731 const auto OldLiveOutSize = LiveOut.size();
2732 for (BasicBlock *Succ : successors(BB)) {
2733 assert(Data.LiveIn.count(Succ));
2734 set_union(LiveOut, Data.LiveIn[Succ]);
2736 // assert OutLiveOut is a subset of LiveOut
2737 if (OldLiveOutSize == LiveOut.size()) {
2738 // If the sets are the same size, then we didn't actually add anything
2739 // when unioning our successors LiveIn Thus, the LiveIn of this block
2743 Data.LiveOut[BB] = LiveOut;
2745 // Apply the effects of this basic block
2746 DenseSet<Value *> LiveTmp = LiveOut;
2747 set_union(LiveTmp, Data.LiveSet[BB]);
2748 set_subtract(LiveTmp, Data.KillSet[BB]);
2750 assert(Data.LiveIn.count(BB));
2751 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2752 // assert: OldLiveIn is a subset of LiveTmp
2753 if (OldLiveIn.size() != LiveTmp.size()) {
2754 Data.LiveIn[BB] = LiveTmp;
2755 AddPredsToWorklist(BB);
2757 } // while( !worklist.empty() )
2760 // Sanity check our output against SSA properties. This helps catch any
2761 // missing kills during the above iteration.
2762 for (BasicBlock &BB : F) {
2763 checkBasicSSA(DT, Data, BB);
2768 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2769 StatepointLiveSetTy &Out) {
2771 BasicBlock *BB = Inst->getParent();
2773 // Note: The copy is intentional and required
2774 assert(Data.LiveOut.count(BB));
2775 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2777 // We want to handle the statepoint itself oddly. It's
2778 // call result is not live (normal), nor are it's arguments
2779 // (unless they're used again later). This adjustment is
2780 // specifically what we need to relocate
2781 BasicBlock::reverse_iterator rend(Inst);
2782 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2783 LiveOut.erase(Inst);
2784 Out.insert(LiveOut.begin(), LiveOut.end());
2787 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2789 PartiallyConstructedSafepointRecord &Info) {
2790 Instruction *Inst = CS.getInstruction();
2791 StatepointLiveSetTy Updated;
2792 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2795 DenseSet<Value *> Bases;
2796 for (auto KVPair : Info.PointerToBase) {
2797 Bases.insert(KVPair.second);
2800 // We may have base pointers which are now live that weren't before. We need
2801 // to update the PointerToBase structure to reflect this.
2802 for (auto V : Updated)
2803 if (!Info.PointerToBase.count(V)) {
2804 assert(Bases.count(V) && "can't find base for unexpected live value");
2805 Info.PointerToBase[V] = V;
2810 for (auto V : Updated) {
2811 assert(Info.PointerToBase.count(V) &&
2812 "must be able to find base for live value");
2816 // Remove any stale base mappings - this can happen since our liveness is
2817 // more precise then the one inherent in the base pointer analysis
2818 DenseSet<Value *> ToErase;
2819 for (auto KVPair : Info.PointerToBase)
2820 if (!Updated.count(KVPair.first))
2821 ToErase.insert(KVPair.first);
2822 for (auto V : ToErase)
2823 Info.PointerToBase.erase(V);
2826 for (auto KVPair : Info.PointerToBase)
2827 assert(Updated.count(KVPair.first) && "record for non-live value");
2830 Info.liveset = Updated;