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 (isa<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>(I) &&
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 // Returns a instruction which produces the base pointer for a given
950 // instruction. The instruction is assumed to be an input to one of the BDVs
951 // seen in the inference algorithm above. As such, we must either already
952 // know it's base defining value is a base, or have inserted a new
953 // instruction to propagate the base of it's BDV and have entered that newly
954 // introduced instruction into the state table. In either case, we are
955 // assured to be able to determine an instruction which produces it's base
957 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
958 Value *BDV = findBaseOrBDV(Input, cache);
959 Value *Base = nullptr;
960 if (isKnownBaseResult(BDV)) {
963 // Either conflict or base.
964 assert(states.count(BDV));
965 Base = states[BDV].getBase();
967 assert(Base && "can't be null");
968 // The cast is needed since base traversal may strip away bitcasts
969 if (Base->getType() != Input->getType() &&
971 Base = new BitCastInst(Base, Input->getType(), "cast",
977 // Fixup all the inputs of the new PHIs
978 for (auto Pair : states) {
979 Instruction *v = cast<Instruction>(Pair.first);
980 BDVState state = Pair.second;
982 assert(!isKnownBaseResult(v) && "why did it get added?");
983 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
984 if (!state.isConflict())
987 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
988 PHINode *phi = cast<PHINode>(v);
989 unsigned NumPHIValues = phi->getNumIncomingValues();
990 for (unsigned i = 0; i < NumPHIValues; i++) {
991 Value *InVal = phi->getIncomingValue(i);
992 BasicBlock *InBB = phi->getIncomingBlock(i);
994 // If we've already seen InBB, add the same incoming value
995 // we added for it earlier. The IR verifier requires phi
996 // nodes with multiple entries from the same basic block
997 // to have the same incoming value for each of those
998 // entries. If we don't do this check here and basephi
999 // has a different type than base, we'll end up adding two
1000 // bitcasts (and hence two distinct values) as incoming
1001 // values for the same basic block.
1003 int blockIndex = basephi->getBasicBlockIndex(InBB);
1004 if (blockIndex != -1) {
1005 Value *oldBase = basephi->getIncomingValue(blockIndex);
1006 basephi->addIncoming(oldBase, InBB);
1009 Value *Base = getBaseForInput(InVal, nullptr);
1010 // In essence this assert states: the only way two
1011 // values incoming from the same basic block may be
1012 // different is by being different bitcasts of the same
1013 // value. A cleanup that remains TODO is changing
1014 // findBaseOrBDV to return an llvm::Value of the correct
1015 // type (and still remain pure). This will remove the
1016 // need to add bitcasts.
1017 assert(Base->stripPointerCasts() == oldBase->stripPointerCasts() &&
1018 "sanity -- findBaseOrBDV should be pure!");
1023 // Find the instruction which produces the base for each input. We may
1024 // need to insert a bitcast in the incoming block.
1025 // TODO: Need to split critical edges if insertion is needed
1026 Value *Base = getBaseForInput(InVal, InBB->getTerminator());
1027 basephi->addIncoming(Base, InBB);
1029 assert(basephi->getNumIncomingValues() == NumPHIValues);
1030 } else if (SelectInst *BaseSel = dyn_cast<SelectInst>(state.getBase())) {
1031 SelectInst *Sel = cast<SelectInst>(v);
1032 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
1033 // something more safe and less hacky.
1034 for (int i = 1; i <= 2; i++) {
1035 Value *InVal = Sel->getOperand(i);
1036 // Find the instruction which produces the base for each input. We may
1037 // need to insert a bitcast.
1038 Value *Base = getBaseForInput(InVal, BaseSel);
1039 BaseSel->setOperand(i, Base);
1042 auto *BaseEE = cast<ExtractElementInst>(state.getBase());
1043 Value *InVal = cast<ExtractElementInst>(v)->getVectorOperand();
1044 // Find the instruction which produces the base for each input. We may
1045 // need to insert a bitcast.
1046 Value *Base = getBaseForInput(InVal, BaseEE);
1047 BaseEE->setOperand(0, Base);
1051 // Now that we're done with the algorithm, see if we can optimize the
1052 // results slightly by reducing the number of new instructions needed.
1053 // Arguably, this should be integrated into the algorithm above, but
1054 // doing as a post process step is easier to reason about for the moment.
1055 DenseMap<Value *, Value *> ReverseMap;
1056 SmallPtrSet<Instruction *, 16> NewInsts;
1057 SmallSetVector<AssertingVH<Instruction>, 16> Worklist;
1058 // Note: We need to visit the states in a deterministic order. We uses the
1059 // Keys we sorted above for this purpose. Note that we are papering over a
1060 // bigger problem with the algorithm above - it's visit order is not
1061 // deterministic. A larger change is needed to fix this.
1062 for (auto Key : Keys) {
1064 auto State = states[Key];
1065 Value *Base = State.getBase();
1067 assert(!isKnownBaseResult(V) && "why did it get added?");
1068 assert(isKnownBaseResult(Base) &&
1069 "must be something we 'know' is a base pointer");
1070 if (!State.isConflict())
1073 ReverseMap[Base] = V;
1074 if (auto *BaseI = dyn_cast<Instruction>(Base)) {
1075 NewInsts.insert(BaseI);
1076 Worklist.insert(BaseI);
1079 auto ReplaceBaseInstWith = [&](Value *BDV, Instruction *BaseI,
1080 Value *Replacement) {
1081 // Add users which are new instructions (excluding self references)
1082 for (User *U : BaseI->users())
1083 if (auto *UI = dyn_cast<Instruction>(U))
1084 if (NewInsts.count(UI) && UI != BaseI)
1085 Worklist.insert(UI);
1086 // Then do the actual replacement
1087 NewInsts.erase(BaseI);
1088 ReverseMap.erase(BaseI);
1089 BaseI->replaceAllUsesWith(Replacement);
1090 BaseI->eraseFromParent();
1091 assert(states.count(BDV));
1092 assert(states[BDV].isConflict() && states[BDV].getBase() == BaseI);
1093 states[BDV] = BDVState(BDVState::Conflict, Replacement);
1095 const DataLayout &DL = cast<Instruction>(def)->getModule()->getDataLayout();
1096 while (!Worklist.empty()) {
1097 Instruction *BaseI = Worklist.pop_back_val();
1098 assert(NewInsts.count(BaseI));
1099 Value *Bdv = ReverseMap[BaseI];
1100 if (auto *BdvI = dyn_cast<Instruction>(Bdv))
1101 if (BaseI->isIdenticalTo(BdvI)) {
1102 DEBUG(dbgs() << "Identical Base: " << *BaseI << "\n");
1103 ReplaceBaseInstWith(Bdv, BaseI, Bdv);
1106 if (Value *V = SimplifyInstruction(BaseI, DL)) {
1107 DEBUG(dbgs() << "Base " << *BaseI << " simplified to " << *V << "\n");
1108 ReplaceBaseInstWith(Bdv, BaseI, V);
1113 // Cache all of our results so we can cheaply reuse them
1114 // NOTE: This is actually two caches: one of the base defining value
1115 // relation and one of the base pointer relation! FIXME
1116 for (auto item : states) {
1117 Value *v = item.first;
1118 Value *base = item.second.getBase();
1121 std::string fromstr =
1122 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
1124 DEBUG(dbgs() << "Updating base value cache"
1125 << " for: " << (v->hasName() ? v->getName() : "")
1126 << " from: " << fromstr
1127 << " to: " << (base->hasName() ? base->getName() : "") << "\n");
1129 if (cache.count(v)) {
1130 // Once we transition from the BDV relation being store in the cache to
1131 // the base relation being stored, it must be stable
1132 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
1133 "base relation should be stable");
1137 assert(cache.find(def) != cache.end());
1141 // For a set of live pointers (base and/or derived), identify the base
1142 // pointer of the object which they are derived from. This routine will
1143 // mutate the IR graph as needed to make the 'base' pointer live at the
1144 // definition site of 'derived'. This ensures that any use of 'derived' can
1145 // also use 'base'. This may involve the insertion of a number of
1146 // additional PHI nodes.
1148 // preconditions: live is a set of pointer type Values
1150 // side effects: may insert PHI nodes into the existing CFG, will preserve
1151 // CFG, will not remove or mutate any existing nodes
1153 // post condition: PointerToBase contains one (derived, base) pair for every
1154 // pointer in live. Note that derived can be equal to base if the original
1155 // pointer was a base pointer.
1157 findBasePointers(const StatepointLiveSetTy &live,
1158 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
1159 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1160 // For the naming of values inserted to be deterministic - which makes for
1161 // much cleaner and more stable tests - we need to assign an order to the
1162 // live values. DenseSets do not provide a deterministic order across runs.
1163 SmallVector<Value *, 64> Temp;
1164 Temp.insert(Temp.end(), live.begin(), live.end());
1165 std::sort(Temp.begin(), Temp.end(), order_by_name);
1166 for (Value *ptr : Temp) {
1167 Value *base = findBasePointer(ptr, DVCache);
1168 assert(base && "failed to find base pointer");
1169 PointerToBase[ptr] = base;
1170 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1171 DT->dominates(cast<Instruction>(base)->getParent(),
1172 cast<Instruction>(ptr)->getParent())) &&
1173 "The base we found better dominate the derived pointer");
1175 // If you see this trip and like to live really dangerously, the code should
1176 // be correct, just with idioms the verifier can't handle. You can try
1177 // disabling the verifier at your own substantial risk.
1178 assert(!isa<ConstantPointerNull>(base) &&
1179 "the relocation code needs adjustment to handle the relocation of "
1180 "a null pointer constant without causing false positives in the "
1181 "safepoint ir verifier.");
1185 /// Find the required based pointers (and adjust the live set) for the given
1187 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1189 PartiallyConstructedSafepointRecord &result) {
1190 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
1191 findBasePointers(result.liveset, PointerToBase, &DT, DVCache);
1193 if (PrintBasePointers) {
1194 // Note: Need to print these in a stable order since this is checked in
1196 errs() << "Base Pairs (w/o Relocation):\n";
1197 SmallVector<Value *, 64> Temp;
1198 Temp.reserve(PointerToBase.size());
1199 for (auto Pair : PointerToBase) {
1200 Temp.push_back(Pair.first);
1202 std::sort(Temp.begin(), Temp.end(), order_by_name);
1203 for (Value *Ptr : Temp) {
1204 Value *Base = PointerToBase[Ptr];
1205 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
1210 result.PointerToBase = PointerToBase;
1213 /// Given an updated version of the dataflow liveness results, update the
1214 /// liveset and base pointer maps for the call site CS.
1215 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1217 PartiallyConstructedSafepointRecord &result);
1219 static void recomputeLiveInValues(
1220 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1221 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1222 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1223 // again. The old values are still live and will help it stabilize quickly.
1224 GCPtrLivenessData RevisedLivenessData;
1225 computeLiveInValues(DT, F, RevisedLivenessData);
1226 for (size_t i = 0; i < records.size(); i++) {
1227 struct PartiallyConstructedSafepointRecord &info = records[i];
1228 const CallSite &CS = toUpdate[i];
1229 recomputeLiveInValues(RevisedLivenessData, CS, info);
1233 // When inserting gc.relocate calls, we need to ensure there are no uses
1234 // of the original value between the gc.statepoint and the gc.relocate call.
1235 // One case which can arise is a phi node starting one of the successor blocks.
1236 // We also need to be able to insert the gc.relocates only on the path which
1237 // goes through the statepoint. We might need to split an edge to make this
1240 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1241 DominatorTree &DT) {
1242 BasicBlock *Ret = BB;
1243 if (!BB->getUniquePredecessor()) {
1244 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1247 // Now that 'ret' has unique predecessor we can safely remove all phi nodes
1249 FoldSingleEntryPHINodes(Ret);
1250 assert(!isa<PHINode>(Ret->begin()));
1252 // At this point, we can safely insert a gc.relocate as the first instruction
1253 // in Ret if needed.
1257 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1258 auto itr = std::find(livevec.begin(), livevec.end(), val);
1259 assert(livevec.end() != itr);
1260 size_t index = std::distance(livevec.begin(), itr);
1261 assert(index < livevec.size());
1265 // Create new attribute set containing only attributes which can be transferred
1266 // from original call to the safepoint.
1267 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1270 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1271 unsigned index = AS.getSlotIndex(Slot);
1273 if (index == AttributeSet::ReturnIndex ||
1274 index == AttributeSet::FunctionIndex) {
1276 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1278 Attribute attr = *it;
1280 // Do not allow certain attributes - just skip them
1281 // Safepoint can not be read only or read none.
1282 if (attr.hasAttribute(Attribute::ReadNone) ||
1283 attr.hasAttribute(Attribute::ReadOnly))
1286 ret = ret.addAttributes(
1287 AS.getContext(), index,
1288 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1292 // Just skip parameter attributes for now
1298 /// Helper function to place all gc relocates necessary for the given
1301 /// liveVariables - list of variables to be relocated.
1302 /// liveStart - index of the first live variable.
1303 /// basePtrs - base pointers.
1304 /// statepointToken - statepoint instruction to which relocates should be
1306 /// Builder - Llvm IR builder to be used to construct new calls.
1307 static void CreateGCRelocates(ArrayRef<llvm::Value *> LiveVariables,
1308 const int LiveStart,
1309 ArrayRef<llvm::Value *> BasePtrs,
1310 Instruction *StatepointToken,
1311 IRBuilder<> Builder) {
1312 if (LiveVariables.empty())
1315 // All gc_relocate are set to i8 addrspace(1)* type. We originally generated
1316 // unique declarations for each pointer type, but this proved problematic
1317 // because the intrinsic mangling code is incomplete and fragile. Since
1318 // we're moving towards a single unified pointer type anyways, we can just
1319 // cast everything to an i8* of the right address space. A bitcast is added
1320 // later to convert gc_relocate to the actual value's type.
1321 Module *M = StatepointToken->getModule();
1322 auto AS = cast<PointerType>(LiveVariables[0]->getType())->getAddressSpace();
1323 Type *Types[] = {Type::getInt8PtrTy(M->getContext(), AS)};
1324 Value *GCRelocateDecl =
1325 Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, Types);
1327 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1328 // Generate the gc.relocate call and save the result
1330 Builder.getInt32(LiveStart + find_index(LiveVariables, BasePtrs[i]));
1332 Builder.getInt32(LiveStart + find_index(LiveVariables, LiveVariables[i]));
1334 // only specify a debug name if we can give a useful one
1335 CallInst *Reloc = Builder.CreateCall(
1336 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1337 LiveVariables[i]->hasName() ? LiveVariables[i]->getName() + ".relocated"
1339 // Trick CodeGen into thinking there are lots of free registers at this
1341 Reloc->setCallingConv(CallingConv::Cold);
1346 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1347 const SmallVectorImpl<llvm::Value *> &basePtrs,
1348 const SmallVectorImpl<llvm::Value *> &liveVariables,
1350 PartiallyConstructedSafepointRecord &result) {
1351 assert(basePtrs.size() == liveVariables.size());
1352 assert(isStatepoint(CS) &&
1353 "This method expects to be rewriting a statepoint");
1355 BasicBlock *BB = CS.getInstruction()->getParent();
1357 Function *F = BB->getParent();
1358 assert(F && "must be set");
1359 Module *M = F->getParent();
1361 assert(M && "must be set");
1363 // We're not changing the function signature of the statepoint since the gc
1364 // arguments go into the var args section.
1365 Function *gc_statepoint_decl = CS.getCalledFunction();
1367 // Then go ahead and use the builder do actually do the inserts. We insert
1368 // immediately before the previous instruction under the assumption that all
1369 // arguments will be available here. We can't insert afterwards since we may
1370 // be replacing a terminator.
1371 Instruction *insertBefore = CS.getInstruction();
1372 IRBuilder<> Builder(insertBefore);
1373 // Copy all of the arguments from the original statepoint - this includes the
1374 // target, call args, and deopt args
1375 SmallVector<llvm::Value *, 64> args;
1376 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1377 // TODO: Clear the 'needs rewrite' flag
1379 // add all the pointers to be relocated (gc arguments)
1380 // Capture the start of the live variable list for use in the gc_relocates
1381 const int live_start = args.size();
1382 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1384 // Create the statepoint given all the arguments
1385 Instruction *token = nullptr;
1386 AttributeSet return_attributes;
1388 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1390 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1391 call->setTailCall(toReplace->isTailCall());
1392 call->setCallingConv(toReplace->getCallingConv());
1394 // Currently we will fail on parameter attributes and on certain
1395 // function attributes.
1396 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1397 // In case if we can handle this set of attributes - set up function attrs
1398 // directly on statepoint and return attrs later for gc_result intrinsic.
1399 call->setAttributes(new_attrs.getFnAttributes());
1400 return_attributes = new_attrs.getRetAttributes();
1404 // Put the following gc_result and gc_relocate calls immediately after the
1405 // the old call (which we're about to delete)
1406 BasicBlock::iterator next(toReplace);
1407 assert(BB->end() != next && "not a terminator, must have next");
1409 Instruction *IP = &*(next);
1410 Builder.SetInsertPoint(IP);
1411 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1414 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1416 // Insert the new invoke into the old block. We'll remove the old one in a
1417 // moment at which point this will become the new terminator for the
1419 InvokeInst *invoke = InvokeInst::Create(
1420 gc_statepoint_decl, toReplace->getNormalDest(),
1421 toReplace->getUnwindDest(), args, "statepoint_token", toReplace->getParent());
1422 invoke->setCallingConv(toReplace->getCallingConv());
1424 // Currently we will fail on parameter attributes and on certain
1425 // function attributes.
1426 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1427 // In case if we can handle this set of attributes - set up function attrs
1428 // directly on statepoint and return attrs later for gc_result intrinsic.
1429 invoke->setAttributes(new_attrs.getFnAttributes());
1430 return_attributes = new_attrs.getRetAttributes();
1434 // Generate gc relocates in exceptional path
1435 BasicBlock *unwindBlock = toReplace->getUnwindDest();
1436 assert(!isa<PHINode>(unwindBlock->begin()) &&
1437 unwindBlock->getUniquePredecessor() &&
1438 "can't safely insert in this block!");
1440 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1441 Builder.SetInsertPoint(IP);
1442 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1444 // Extract second element from landingpad return value. We will attach
1445 // exceptional gc relocates to it.
1446 const unsigned idx = 1;
1447 Instruction *exceptional_token =
1448 cast<Instruction>(Builder.CreateExtractValue(
1449 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1450 result.UnwindToken = exceptional_token;
1452 CreateGCRelocates(liveVariables, live_start, basePtrs,
1453 exceptional_token, Builder);
1455 // Generate gc relocates and returns for normal block
1456 BasicBlock *normalDest = toReplace->getNormalDest();
1457 assert(!isa<PHINode>(normalDest->begin()) &&
1458 normalDest->getUniquePredecessor() &&
1459 "can't safely insert in this block!");
1461 IP = &*(normalDest->getFirstInsertionPt());
1462 Builder.SetInsertPoint(IP);
1464 // gc relocates will be generated later as if it were regular call
1469 // Take the name of the original value call if it had one.
1470 token->takeName(CS.getInstruction());
1472 // The GCResult is already inserted, we just need to find it
1474 Instruction *toReplace = CS.getInstruction();
1475 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1476 "only valid use before rewrite is gc.result");
1477 assert(!toReplace->hasOneUse() ||
1478 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1481 // Update the gc.result of the original statepoint (if any) to use the newly
1482 // inserted statepoint. This is safe to do here since the token can't be
1483 // considered a live reference.
1484 CS.getInstruction()->replaceAllUsesWith(token);
1486 result.StatepointToken = token;
1488 // Second, create a gc.relocate for every live variable
1489 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1493 struct name_ordering {
1496 bool operator()(name_ordering const &a, name_ordering const &b) {
1497 return -1 == a.derived->getName().compare(b.derived->getName());
1501 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1502 SmallVectorImpl<Value *> &livevec) {
1503 assert(basevec.size() == livevec.size());
1505 SmallVector<name_ordering, 64> temp;
1506 for (size_t i = 0; i < basevec.size(); i++) {
1508 v.base = basevec[i];
1509 v.derived = livevec[i];
1512 std::sort(temp.begin(), temp.end(), name_ordering());
1513 for (size_t i = 0; i < basevec.size(); i++) {
1514 basevec[i] = temp[i].base;
1515 livevec[i] = temp[i].derived;
1519 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1520 // which make the relocations happening at this safepoint explicit.
1522 // WARNING: Does not do any fixup to adjust users of the original live
1523 // values. That's the callers responsibility.
1525 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1526 PartiallyConstructedSafepointRecord &result) {
1527 auto liveset = result.liveset;
1528 auto PointerToBase = result.PointerToBase;
1530 // Convert to vector for efficient cross referencing.
1531 SmallVector<Value *, 64> basevec, livevec;
1532 livevec.reserve(liveset.size());
1533 basevec.reserve(liveset.size());
1534 for (Value *L : liveset) {
1535 livevec.push_back(L);
1536 assert(PointerToBase.count(L));
1537 Value *base = PointerToBase[L];
1538 basevec.push_back(base);
1540 assert(livevec.size() == basevec.size());
1542 // To make the output IR slightly more stable (for use in diffs), ensure a
1543 // fixed order of the values in the safepoint (by sorting the value name).
1544 // The order is otherwise meaningless.
1545 stablize_order(basevec, livevec);
1547 // Do the actual rewriting and delete the old statepoint
1548 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1549 CS.getInstruction()->eraseFromParent();
1552 // Helper function for the relocationViaAlloca.
1553 // It receives iterator to the statepoint gc relocates and emits store to the
1555 // location (via allocaMap) for the each one of them.
1556 // Add visited values into the visitedLiveValues set we will later use them
1557 // for sanity check.
1559 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1560 DenseMap<Value *, Value *> &AllocaMap,
1561 DenseSet<Value *> &VisitedLiveValues) {
1563 for (User *U : GCRelocs) {
1564 if (!isa<IntrinsicInst>(U))
1567 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1569 // We only care about relocates
1570 if (RelocatedValue->getIntrinsicID() !=
1571 Intrinsic::experimental_gc_relocate) {
1575 GCRelocateOperands RelocateOperands(RelocatedValue);
1576 Value *OriginalValue =
1577 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1578 assert(AllocaMap.count(OriginalValue));
1579 Value *Alloca = AllocaMap[OriginalValue];
1581 // Emit store into the related alloca
1582 // All gc_relocate are i8 addrspace(1)* typed, and it must be bitcasted to
1583 // the correct type according to alloca.
1584 assert(RelocatedValue->getNextNode() && "Should always have one since it's not a terminator");
1585 IRBuilder<> Builder(RelocatedValue->getNextNode());
1586 Value *CastedRelocatedValue =
1587 Builder.CreateBitCast(RelocatedValue, cast<AllocaInst>(Alloca)->getAllocatedType(),
1588 RelocatedValue->hasName() ? RelocatedValue->getName() + ".casted" : "");
1590 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1591 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1594 VisitedLiveValues.insert(OriginalValue);
1599 // Helper function for the "relocationViaAlloca". Similar to the
1600 // "insertRelocationStores" but works for rematerialized values.
1602 insertRematerializationStores(
1603 RematerializedValueMapTy RematerializedValues,
1604 DenseMap<Value *, Value *> &AllocaMap,
1605 DenseSet<Value *> &VisitedLiveValues) {
1607 for (auto RematerializedValuePair: RematerializedValues) {
1608 Instruction *RematerializedValue = RematerializedValuePair.first;
1609 Value *OriginalValue = RematerializedValuePair.second;
1611 assert(AllocaMap.count(OriginalValue) &&
1612 "Can not find alloca for rematerialized value");
1613 Value *Alloca = AllocaMap[OriginalValue];
1615 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1616 Store->insertAfter(RematerializedValue);
1619 VisitedLiveValues.insert(OriginalValue);
1624 /// do all the relocation update via allocas and mem2reg
1625 static void relocationViaAlloca(
1626 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1627 ArrayRef<struct PartiallyConstructedSafepointRecord> Records) {
1629 // record initial number of (static) allocas; we'll check we have the same
1630 // number when we get done.
1631 int InitialAllocaNum = 0;
1632 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1634 if (isa<AllocaInst>(*I))
1638 // TODO-PERF: change data structures, reserve
1639 DenseMap<Value *, Value *> AllocaMap;
1640 SmallVector<AllocaInst *, 200> PromotableAllocas;
1641 // Used later to chack that we have enough allocas to store all values
1642 std::size_t NumRematerializedValues = 0;
1643 PromotableAllocas.reserve(Live.size());
1645 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1646 // "PromotableAllocas"
1647 auto emitAllocaFor = [&](Value *LiveValue) {
1648 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1649 F.getEntryBlock().getFirstNonPHI());
1650 AllocaMap[LiveValue] = Alloca;
1651 PromotableAllocas.push_back(Alloca);
1654 // emit alloca for each live gc pointer
1655 for (unsigned i = 0; i < Live.size(); i++) {
1656 emitAllocaFor(Live[i]);
1659 // emit allocas for rematerialized values
1660 for (size_t i = 0; i < Records.size(); i++) {
1661 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1663 for (auto RematerializedValuePair : Info.RematerializedValues) {
1664 Value *OriginalValue = RematerializedValuePair.second;
1665 if (AllocaMap.count(OriginalValue) != 0)
1668 emitAllocaFor(OriginalValue);
1669 ++NumRematerializedValues;
1673 // The next two loops are part of the same conceptual operation. We need to
1674 // insert a store to the alloca after the original def and at each
1675 // redefinition. We need to insert a load before each use. These are split
1676 // into distinct loops for performance reasons.
1678 // update gc pointer after each statepoint
1679 // either store a relocated value or null (if no relocated value found for
1680 // this gc pointer and it is not a gc_result)
1681 // this must happen before we update the statepoint with load of alloca
1682 // otherwise we lose the link between statepoint and old def
1683 for (size_t i = 0; i < Records.size(); i++) {
1684 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1685 Value *Statepoint = Info.StatepointToken;
1687 // This will be used for consistency check
1688 DenseSet<Value *> VisitedLiveValues;
1690 // Insert stores for normal statepoint gc relocates
1691 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1693 // In case if it was invoke statepoint
1694 // we will insert stores for exceptional path gc relocates.
1695 if (isa<InvokeInst>(Statepoint)) {
1696 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1700 // Do similar thing with rematerialized values
1701 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1704 if (ClobberNonLive) {
1705 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1706 // the gc.statepoint. This will turn some subtle GC problems into
1707 // slightly easier to debug SEGVs. Note that on large IR files with
1708 // lots of gc.statepoints this is extremely costly both memory and time
1710 SmallVector<AllocaInst *, 64> ToClobber;
1711 for (auto Pair : AllocaMap) {
1712 Value *Def = Pair.first;
1713 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1715 // This value was relocated
1716 if (VisitedLiveValues.count(Def)) {
1719 ToClobber.push_back(Alloca);
1722 auto InsertClobbersAt = [&](Instruction *IP) {
1723 for (auto *AI : ToClobber) {
1724 auto AIType = cast<PointerType>(AI->getType());
1725 auto PT = cast<PointerType>(AIType->getElementType());
1726 Constant *CPN = ConstantPointerNull::get(PT);
1727 StoreInst *Store = new StoreInst(CPN, AI);
1728 Store->insertBefore(IP);
1732 // Insert the clobbering stores. These may get intermixed with the
1733 // gc.results and gc.relocates, but that's fine.
1734 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1735 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1736 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1738 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1740 InsertClobbersAt(Next);
1744 // update use with load allocas and add store for gc_relocated
1745 for (auto Pair : AllocaMap) {
1746 Value *Def = Pair.first;
1747 Value *Alloca = Pair.second;
1749 // we pre-record the uses of allocas so that we dont have to worry about
1751 // that change the user information.
1752 SmallVector<Instruction *, 20> Uses;
1753 // PERF: trade a linear scan for repeated reallocation
1754 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1755 for (User *U : Def->users()) {
1756 if (!isa<ConstantExpr>(U)) {
1757 // If the def has a ConstantExpr use, then the def is either a
1758 // ConstantExpr use itself or null. In either case
1759 // (recursively in the first, directly in the second), the oop
1760 // it is ultimately dependent on is null and this particular
1761 // use does not need to be fixed up.
1762 Uses.push_back(cast<Instruction>(U));
1766 std::sort(Uses.begin(), Uses.end());
1767 auto Last = std::unique(Uses.begin(), Uses.end());
1768 Uses.erase(Last, Uses.end());
1770 for (Instruction *Use : Uses) {
1771 if (isa<PHINode>(Use)) {
1772 PHINode *Phi = cast<PHINode>(Use);
1773 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1774 if (Def == Phi->getIncomingValue(i)) {
1775 LoadInst *Load = new LoadInst(
1776 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1777 Phi->setIncomingValue(i, Load);
1781 LoadInst *Load = new LoadInst(Alloca, "", Use);
1782 Use->replaceUsesOfWith(Def, Load);
1786 // emit store for the initial gc value
1787 // store must be inserted after load, otherwise store will be in alloca's
1788 // use list and an extra load will be inserted before it
1789 StoreInst *Store = new StoreInst(Def, Alloca);
1790 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1791 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1792 // InvokeInst is a TerminatorInst so the store need to be inserted
1793 // into its normal destination block.
1794 BasicBlock *NormalDest = Invoke->getNormalDest();
1795 Store->insertBefore(NormalDest->getFirstNonPHI());
1797 assert(!Inst->isTerminator() &&
1798 "The only TerminatorInst that can produce a value is "
1799 "InvokeInst which is handled above.");
1800 Store->insertAfter(Inst);
1803 assert(isa<Argument>(Def));
1804 Store->insertAfter(cast<Instruction>(Alloca));
1808 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1809 "we must have the same allocas with lives");
1810 if (!PromotableAllocas.empty()) {
1811 // apply mem2reg to promote alloca to SSA
1812 PromoteMemToReg(PromotableAllocas, DT);
1816 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1818 if (isa<AllocaInst>(*I))
1820 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1824 /// Implement a unique function which doesn't require we sort the input
1825 /// vector. Doing so has the effect of changing the output of a couple of
1826 /// tests in ways which make them less useful in testing fused safepoints.
1827 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1828 SmallSet<T, 8> Seen;
1829 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1830 return !Seen.insert(V).second;
1834 /// Insert holders so that each Value is obviously live through the entire
1835 /// lifetime of the call.
1836 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1837 SmallVectorImpl<CallInst *> &Holders) {
1839 // No values to hold live, might as well not insert the empty holder
1842 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1843 // Use a dummy vararg function to actually hold the values live
1844 Function *Func = cast<Function>(M->getOrInsertFunction(
1845 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1847 // For call safepoints insert dummy calls right after safepoint
1848 BasicBlock::iterator Next(CS.getInstruction());
1850 Holders.push_back(CallInst::Create(Func, Values, "", Next));
1853 // For invoke safepooints insert dummy calls both in normal and
1854 // exceptional destination blocks
1855 auto *II = cast<InvokeInst>(CS.getInstruction());
1856 Holders.push_back(CallInst::Create(
1857 Func, Values, "", II->getNormalDest()->getFirstInsertionPt()));
1858 Holders.push_back(CallInst::Create(
1859 Func, Values, "", II->getUnwindDest()->getFirstInsertionPt()));
1862 static void findLiveReferences(
1863 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1864 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1865 GCPtrLivenessData OriginalLivenessData;
1866 computeLiveInValues(DT, F, OriginalLivenessData);
1867 for (size_t i = 0; i < records.size(); i++) {
1868 struct PartiallyConstructedSafepointRecord &info = records[i];
1869 const CallSite &CS = toUpdate[i];
1870 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1874 /// Remove any vector of pointers from the liveset by scalarizing them over the
1875 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1876 /// would be preferable to include the vector in the statepoint itself, but
1877 /// the lowering code currently does not handle that. Extending it would be
1878 /// slightly non-trivial since it requires a format change. Given how rare
1879 /// such cases are (for the moment?) scalarizing is an acceptable compromise.
1880 static void splitVectorValues(Instruction *StatepointInst,
1881 StatepointLiveSetTy &LiveSet,
1882 DenseMap<Value *, Value *>& PointerToBase,
1883 DominatorTree &DT) {
1884 SmallVector<Value *, 16> ToSplit;
1885 for (Value *V : LiveSet)
1886 if (isa<VectorType>(V->getType()))
1887 ToSplit.push_back(V);
1889 if (ToSplit.empty())
1892 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1894 Function &F = *(StatepointInst->getParent()->getParent());
1896 DenseMap<Value *, AllocaInst *> AllocaMap;
1897 // First is normal return, second is exceptional return (invoke only)
1898 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1899 for (Value *V : ToSplit) {
1900 AllocaInst *Alloca =
1901 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1902 AllocaMap[V] = Alloca;
1904 VectorType *VT = cast<VectorType>(V->getType());
1905 IRBuilder<> Builder(StatepointInst);
1906 SmallVector<Value *, 16> Elements;
1907 for (unsigned i = 0; i < VT->getNumElements(); i++)
1908 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1909 ElementMapping[V] = Elements;
1911 auto InsertVectorReform = [&](Instruction *IP) {
1912 Builder.SetInsertPoint(IP);
1913 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1914 Value *ResultVec = UndefValue::get(VT);
1915 for (unsigned i = 0; i < VT->getNumElements(); i++)
1916 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1917 Builder.getInt32(i));
1921 if (isa<CallInst>(StatepointInst)) {
1922 BasicBlock::iterator Next(StatepointInst);
1924 Instruction *IP = &*(Next);
1925 Replacements[V].first = InsertVectorReform(IP);
1926 Replacements[V].second = nullptr;
1928 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1929 // We've already normalized - check that we don't have shared destination
1931 BasicBlock *NormalDest = Invoke->getNormalDest();
1932 assert(!isa<PHINode>(NormalDest->begin()));
1933 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1934 assert(!isa<PHINode>(UnwindDest->begin()));
1935 // Insert insert element sequences in both successors
1936 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1937 Replacements[V].first = InsertVectorReform(IP);
1938 IP = &*(UnwindDest->getFirstInsertionPt());
1939 Replacements[V].second = InsertVectorReform(IP);
1943 for (Value *V : ToSplit) {
1944 AllocaInst *Alloca = AllocaMap[V];
1946 // Capture all users before we start mutating use lists
1947 SmallVector<Instruction *, 16> Users;
1948 for (User *U : V->users())
1949 Users.push_back(cast<Instruction>(U));
1951 for (Instruction *I : Users) {
1952 if (auto Phi = dyn_cast<PHINode>(I)) {
1953 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1954 if (V == Phi->getIncomingValue(i)) {
1955 LoadInst *Load = new LoadInst(
1956 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1957 Phi->setIncomingValue(i, Load);
1960 LoadInst *Load = new LoadInst(Alloca, "", I);
1961 I->replaceUsesOfWith(V, Load);
1965 // Store the original value and the replacement value into the alloca
1966 StoreInst *Store = new StoreInst(V, Alloca);
1967 if (auto I = dyn_cast<Instruction>(V))
1968 Store->insertAfter(I);
1970 Store->insertAfter(Alloca);
1972 // Normal return for invoke, or call return
1973 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1974 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1975 // Unwind return for invoke only
1976 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1978 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1981 // apply mem2reg to promote alloca to SSA
1982 SmallVector<AllocaInst *, 16> Allocas;
1983 for (Value *V : ToSplit)
1984 Allocas.push_back(AllocaMap[V]);
1985 PromoteMemToReg(Allocas, DT);
1987 // Update our tracking of live pointers and base mappings to account for the
1988 // changes we just made.
1989 for (Value *V : ToSplit) {
1990 auto &Elements = ElementMapping[V];
1993 LiveSet.insert(Elements.begin(), Elements.end());
1994 // We need to update the base mapping as well.
1995 assert(PointerToBase.count(V));
1996 Value *OldBase = PointerToBase[V];
1997 auto &BaseElements = ElementMapping[OldBase];
1998 PointerToBase.erase(V);
1999 assert(Elements.size() == BaseElements.size());
2000 for (unsigned i = 0; i < Elements.size(); i++) {
2001 Value *Elem = Elements[i];
2002 PointerToBase[Elem] = BaseElements[i];
2007 // Helper function for the "rematerializeLiveValues". It walks use chain
2008 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
2009 // values are visited (currently it is GEP's and casts). Returns true if it
2010 // successfully reached "BaseValue" and false otherwise.
2011 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
2013 static bool findRematerializableChainToBasePointer(
2014 SmallVectorImpl<Instruction*> &ChainToBase,
2015 Value *CurrentValue, Value *BaseValue) {
2017 // We have found a base value
2018 if (CurrentValue == BaseValue) {
2022 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
2023 ChainToBase.push_back(GEP);
2024 return findRematerializableChainToBasePointer(ChainToBase,
2025 GEP->getPointerOperand(),
2029 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
2030 Value *Def = CI->stripPointerCasts();
2032 // This two checks are basically similar. First one is here for the
2033 // consistency with findBasePointers logic.
2034 assert(!isa<CastInst>(Def) && "not a pointer cast found");
2035 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
2038 ChainToBase.push_back(CI);
2039 return findRematerializableChainToBasePointer(ChainToBase, Def, BaseValue);
2042 // Not supported instruction in the chain
2046 // Helper function for the "rematerializeLiveValues". Compute cost of the use
2047 // chain we are going to rematerialize.
2049 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
2050 TargetTransformInfo &TTI) {
2053 for (Instruction *Instr : Chain) {
2054 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
2055 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
2056 "non noop cast is found during rematerialization");
2058 Type *SrcTy = CI->getOperand(0)->getType();
2059 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
2061 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
2062 // Cost of the address calculation
2063 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
2064 Cost += TTI.getAddressComputationCost(ValTy);
2066 // And cost of the GEP itself
2067 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
2068 // allowed for the external usage)
2069 if (!GEP->hasAllConstantIndices())
2073 llvm_unreachable("unsupported instruciton type during rematerialization");
2080 // From the statepoint liveset pick values that are cheaper to recompute then to
2081 // relocate. Remove this values from the liveset, rematerialize them after
2082 // statepoint and record them in "Info" structure. Note that similar to
2083 // relocated values we don't do any user adjustments here.
2084 static void rematerializeLiveValues(CallSite CS,
2085 PartiallyConstructedSafepointRecord &Info,
2086 TargetTransformInfo &TTI) {
2087 const unsigned int ChainLengthThreshold = 10;
2089 // Record values we are going to delete from this statepoint live set.
2090 // We can not di this in following loop due to iterator invalidation.
2091 SmallVector<Value *, 32> LiveValuesToBeDeleted;
2093 for (Value *LiveValue: Info.liveset) {
2094 // For each live pointer find it's defining chain
2095 SmallVector<Instruction *, 3> ChainToBase;
2096 assert(Info.PointerToBase.count(LiveValue));
2098 findRematerializableChainToBasePointer(ChainToBase,
2100 Info.PointerToBase[LiveValue]);
2101 // Nothing to do, or chain is too long
2103 ChainToBase.size() == 0 ||
2104 ChainToBase.size() > ChainLengthThreshold)
2107 // Compute cost of this chain
2108 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
2109 // TODO: We can also account for cases when we will be able to remove some
2110 // of the rematerialized values by later optimization passes. I.e if
2111 // we rematerialized several intersecting chains. Or if original values
2112 // don't have any uses besides this statepoint.
2114 // For invokes we need to rematerialize each chain twice - for normal and
2115 // for unwind basic blocks. Model this by multiplying cost by two.
2116 if (CS.isInvoke()) {
2119 // If it's too expensive - skip it
2120 if (Cost >= RematerializationThreshold)
2123 // Remove value from the live set
2124 LiveValuesToBeDeleted.push_back(LiveValue);
2126 // Clone instructions and record them inside "Info" structure
2128 // Walk backwards to visit top-most instructions first
2129 std::reverse(ChainToBase.begin(), ChainToBase.end());
2131 // Utility function which clones all instructions from "ChainToBase"
2132 // and inserts them before "InsertBefore". Returns rematerialized value
2133 // which should be used after statepoint.
2134 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
2135 Instruction *LastClonedValue = nullptr;
2136 Instruction *LastValue = nullptr;
2137 for (Instruction *Instr: ChainToBase) {
2138 // Only GEP's and casts are suported as we need to be careful to not
2139 // introduce any new uses of pointers not in the liveset.
2140 // Note that it's fine to introduce new uses of pointers which were
2141 // otherwise not used after this statepoint.
2142 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2144 Instruction *ClonedValue = Instr->clone();
2145 ClonedValue->insertBefore(InsertBefore);
2146 ClonedValue->setName(Instr->getName() + ".remat");
2148 // If it is not first instruction in the chain then it uses previously
2149 // cloned value. We should update it to use cloned value.
2150 if (LastClonedValue) {
2152 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2154 // Assert that cloned instruction does not use any instructions from
2155 // this chain other than LastClonedValue
2156 for (auto OpValue : ClonedValue->operand_values()) {
2157 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2158 ChainToBase.end() &&
2159 "incorrect use in rematerialization chain");
2164 LastClonedValue = ClonedValue;
2167 assert(LastClonedValue);
2168 return LastClonedValue;
2171 // Different cases for calls and invokes. For invokes we need to clone
2172 // instructions both on normal and unwind path.
2174 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2175 assert(InsertBefore);
2176 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2177 Info.RematerializedValues[RematerializedValue] = LiveValue;
2179 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2181 Instruction *NormalInsertBefore =
2182 Invoke->getNormalDest()->getFirstInsertionPt();
2183 Instruction *UnwindInsertBefore =
2184 Invoke->getUnwindDest()->getFirstInsertionPt();
2186 Instruction *NormalRematerializedValue =
2187 rematerializeChain(NormalInsertBefore);
2188 Instruction *UnwindRematerializedValue =
2189 rematerializeChain(UnwindInsertBefore);
2191 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2192 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2196 // Remove rematerializaed values from the live set
2197 for (auto LiveValue: LiveValuesToBeDeleted) {
2198 Info.liveset.erase(LiveValue);
2202 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
2203 SmallVectorImpl<CallSite> &toUpdate) {
2205 // sanity check the input
2206 std::set<CallSite> uniqued;
2207 uniqued.insert(toUpdate.begin(), toUpdate.end());
2208 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
2210 for (size_t i = 0; i < toUpdate.size(); i++) {
2211 CallSite &CS = toUpdate[i];
2212 assert(CS.getInstruction()->getParent()->getParent() == &F);
2213 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
2217 // When inserting gc.relocates for invokes, we need to be able to insert at
2218 // the top of the successor blocks. See the comment on
2219 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2220 // may restructure the CFG.
2221 for (CallSite CS : toUpdate) {
2224 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
2225 normalizeForInvokeSafepoint(invoke->getNormalDest(), invoke->getParent(),
2227 normalizeForInvokeSafepoint(invoke->getUnwindDest(), invoke->getParent(),
2231 // A list of dummy calls added to the IR to keep various values obviously
2232 // live in the IR. We'll remove all of these when done.
2233 SmallVector<CallInst *, 64> holders;
2235 // Insert a dummy call with all of the arguments to the vm_state we'll need
2236 // for the actual safepoint insertion. This ensures reference arguments in
2237 // the deopt argument list are considered live through the safepoint (and
2238 // thus makes sure they get relocated.)
2239 for (size_t i = 0; i < toUpdate.size(); i++) {
2240 CallSite &CS = toUpdate[i];
2241 Statepoint StatepointCS(CS);
2243 SmallVector<Value *, 64> DeoptValues;
2244 for (Use &U : StatepointCS.vm_state_args()) {
2245 Value *Arg = cast<Value>(&U);
2246 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2247 "support for FCA unimplemented");
2248 if (isHandledGCPointerType(Arg->getType()))
2249 DeoptValues.push_back(Arg);
2251 insertUseHolderAfter(CS, DeoptValues, holders);
2254 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
2255 records.reserve(toUpdate.size());
2256 for (size_t i = 0; i < toUpdate.size(); i++) {
2257 struct PartiallyConstructedSafepointRecord info;
2258 records.push_back(info);
2260 assert(records.size() == toUpdate.size());
2262 // A) Identify all gc pointers which are statically live at the given call
2264 findLiveReferences(F, DT, P, toUpdate, records);
2266 // B) Find the base pointers for each live pointer
2267 /* scope for caching */ {
2268 // Cache the 'defining value' relation used in the computation and
2269 // insertion of base phis and selects. This ensures that we don't insert
2270 // large numbers of duplicate base_phis.
2271 DefiningValueMapTy DVCache;
2273 for (size_t i = 0; i < records.size(); i++) {
2274 struct PartiallyConstructedSafepointRecord &info = records[i];
2275 CallSite &CS = toUpdate[i];
2276 findBasePointers(DT, DVCache, CS, info);
2278 } // end of cache scope
2280 // The base phi insertion logic (for any safepoint) may have inserted new
2281 // instructions which are now live at some safepoint. The simplest such
2284 // phi a <-- will be a new base_phi here
2285 // safepoint 1 <-- that needs to be live here
2289 // We insert some dummy calls after each safepoint to definitely hold live
2290 // the base pointers which were identified for that safepoint. We'll then
2291 // ask liveness for _every_ base inserted to see what is now live. Then we
2292 // remove the dummy calls.
2293 holders.reserve(holders.size() + records.size());
2294 for (size_t i = 0; i < records.size(); i++) {
2295 struct PartiallyConstructedSafepointRecord &info = records[i];
2296 CallSite &CS = toUpdate[i];
2298 SmallVector<Value *, 128> Bases;
2299 for (auto Pair : info.PointerToBase) {
2300 Bases.push_back(Pair.second);
2302 insertUseHolderAfter(CS, Bases, holders);
2305 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2306 // need to rerun liveness. We may *also* have inserted new defs, but that's
2307 // not the key issue.
2308 recomputeLiveInValues(F, DT, P, toUpdate, records);
2310 if (PrintBasePointers) {
2311 for (size_t i = 0; i < records.size(); i++) {
2312 struct PartiallyConstructedSafepointRecord &info = records[i];
2313 errs() << "Base Pairs: (w/Relocation)\n";
2314 for (auto Pair : info.PointerToBase) {
2315 errs() << " derived %" << Pair.first->getName() << " base %"
2316 << Pair.second->getName() << "\n";
2320 for (size_t i = 0; i < holders.size(); i++) {
2321 holders[i]->eraseFromParent();
2322 holders[i] = nullptr;
2326 // Do a limited scalarization of any live at safepoint vector values which
2327 // contain pointers. This enables this pass to run after vectorization at
2328 // the cost of some possible performance loss. TODO: it would be nice to
2329 // natively support vectors all the way through the backend so we don't need
2330 // to scalarize here.
2331 for (size_t i = 0; i < records.size(); i++) {
2332 struct PartiallyConstructedSafepointRecord &info = records[i];
2333 Instruction *statepoint = toUpdate[i].getInstruction();
2334 splitVectorValues(cast<Instruction>(statepoint), info.liveset,
2335 info.PointerToBase, DT);
2338 // In order to reduce live set of statepoint we might choose to rematerialize
2339 // some values instead of relocating them. This is purely an optimization and
2340 // does not influence correctness.
2341 TargetTransformInfo &TTI =
2342 P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2344 for (size_t i = 0; i < records.size(); i++) {
2345 struct PartiallyConstructedSafepointRecord &info = records[i];
2346 CallSite &CS = toUpdate[i];
2348 rematerializeLiveValues(CS, info, TTI);
2351 // Now run through and replace the existing statepoints with new ones with
2352 // the live variables listed. We do not yet update uses of the values being
2353 // relocated. We have references to live variables that need to
2354 // survive to the last iteration of this loop. (By construction, the
2355 // previous statepoint can not be a live variable, thus we can and remove
2356 // the old statepoint calls as we go.)
2357 for (size_t i = 0; i < records.size(); i++) {
2358 struct PartiallyConstructedSafepointRecord &info = records[i];
2359 CallSite &CS = toUpdate[i];
2360 makeStatepointExplicit(DT, CS, P, info);
2362 toUpdate.clear(); // prevent accident use of invalid CallSites
2364 // Do all the fixups of the original live variables to their relocated selves
2365 SmallVector<Value *, 128> live;
2366 for (size_t i = 0; i < records.size(); i++) {
2367 struct PartiallyConstructedSafepointRecord &info = records[i];
2368 // We can't simply save the live set from the original insertion. One of
2369 // the live values might be the result of a call which needs a safepoint.
2370 // That Value* no longer exists and we need to use the new gc_result.
2371 // Thankfully, the liveset is embedded in the statepoint (and updated), so
2372 // we just grab that.
2373 Statepoint statepoint(info.StatepointToken);
2374 live.insert(live.end(), statepoint.gc_args_begin(),
2375 statepoint.gc_args_end());
2377 // Do some basic sanity checks on our liveness results before performing
2378 // relocation. Relocation can and will turn mistakes in liveness results
2379 // into non-sensical code which is must harder to debug.
2380 // TODO: It would be nice to test consistency as well
2381 assert(DT.isReachableFromEntry(info.StatepointToken->getParent()) &&
2382 "statepoint must be reachable or liveness is meaningless");
2383 for (Value *V : statepoint.gc_args()) {
2384 if (!isa<Instruction>(V))
2385 // Non-instruction values trivial dominate all possible uses
2387 auto LiveInst = cast<Instruction>(V);
2388 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2389 "unreachable values should never be live");
2390 assert(DT.dominates(LiveInst, info.StatepointToken) &&
2391 "basic SSA liveness expectation violated by liveness analysis");
2395 unique_unsorted(live);
2399 for (auto ptr : live) {
2400 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
2404 relocationViaAlloca(F, DT, live, records);
2405 return !records.empty();
2408 // Handles both return values and arguments for Functions and CallSites.
2409 template <typename AttrHolder>
2410 static void RemoveDerefAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2413 if (AH.getDereferenceableBytes(Index))
2414 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2415 AH.getDereferenceableBytes(Index)));
2416 if (AH.getDereferenceableOrNullBytes(Index))
2417 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2418 AH.getDereferenceableOrNullBytes(Index)));
2421 AH.setAttributes(AH.getAttributes().removeAttributes(
2422 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2426 RewriteStatepointsForGC::stripDereferenceabilityInfoFromPrototype(Function &F) {
2427 LLVMContext &Ctx = F.getContext();
2429 for (Argument &A : F.args())
2430 if (isa<PointerType>(A.getType()))
2431 RemoveDerefAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2433 if (isa<PointerType>(F.getReturnType()))
2434 RemoveDerefAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2437 void RewriteStatepointsForGC::stripDereferenceabilityInfoFromBody(Function &F) {
2441 LLVMContext &Ctx = F.getContext();
2442 MDBuilder Builder(Ctx);
2444 for (Instruction &I : instructions(F)) {
2445 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2446 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2447 bool IsImmutableTBAA =
2448 MD->getNumOperands() == 4 &&
2449 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2451 if (!IsImmutableTBAA)
2452 continue; // no work to do, MD_tbaa is already marked mutable
2454 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2455 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2457 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2459 MDNode *MutableTBAA =
2460 Builder.createTBAAStructTagNode(Base, Access, Offset);
2461 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2464 if (CallSite CS = CallSite(&I)) {
2465 for (int i = 0, e = CS.arg_size(); i != e; i++)
2466 if (isa<PointerType>(CS.getArgument(i)->getType()))
2467 RemoveDerefAttrAtIndex(Ctx, CS, i + 1);
2468 if (isa<PointerType>(CS.getType()))
2469 RemoveDerefAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2474 /// Returns true if this function should be rewritten by this pass. The main
2475 /// point of this function is as an extension point for custom logic.
2476 static bool shouldRewriteStatepointsIn(Function &F) {
2477 // TODO: This should check the GCStrategy
2479 const char *FunctionGCName = F.getGC();
2480 const StringRef StatepointExampleName("statepoint-example");
2481 const StringRef CoreCLRName("coreclr");
2482 return (StatepointExampleName == FunctionGCName) ||
2483 (CoreCLRName == FunctionGCName);
2488 void RewriteStatepointsForGC::stripDereferenceabilityInfo(Module &M) {
2490 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2494 for (Function &F : M)
2495 stripDereferenceabilityInfoFromPrototype(F);
2497 for (Function &F : M)
2498 stripDereferenceabilityInfoFromBody(F);
2501 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2502 // Nothing to do for declarations.
2503 if (F.isDeclaration() || F.empty())
2506 // Policy choice says not to rewrite - the most common reason is that we're
2507 // compiling code without a GCStrategy.
2508 if (!shouldRewriteStatepointsIn(F))
2511 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2513 // Gather all the statepoints which need rewritten. Be careful to only
2514 // consider those in reachable code since we need to ask dominance queries
2515 // when rewriting. We'll delete the unreachable ones in a moment.
2516 SmallVector<CallSite, 64> ParsePointNeeded;
2517 bool HasUnreachableStatepoint = false;
2518 for (Instruction &I : instructions(F)) {
2519 // TODO: only the ones with the flag set!
2520 if (isStatepoint(I)) {
2521 if (DT.isReachableFromEntry(I.getParent()))
2522 ParsePointNeeded.push_back(CallSite(&I));
2524 HasUnreachableStatepoint = true;
2528 bool MadeChange = false;
2530 // Delete any unreachable statepoints so that we don't have unrewritten
2531 // statepoints surviving this pass. This makes testing easier and the
2532 // resulting IR less confusing to human readers. Rather than be fancy, we
2533 // just reuse a utility function which removes the unreachable blocks.
2534 if (HasUnreachableStatepoint)
2535 MadeChange |= removeUnreachableBlocks(F);
2537 // Return early if no work to do.
2538 if (ParsePointNeeded.empty())
2541 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2542 // These are created by LCSSA. They have the effect of increasing the size
2543 // of liveness sets for no good reason. It may be harder to do this post
2544 // insertion since relocations and base phis can confuse things.
2545 for (BasicBlock &BB : F)
2546 if (BB.getUniquePredecessor()) {
2548 FoldSingleEntryPHINodes(&BB);
2551 // Before we start introducing relocations, we want to tweak the IR a bit to
2552 // avoid unfortunate code generation effects. The main example is that we
2553 // want to try to make sure the comparison feeding a branch is after any
2554 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2555 // values feeding a branch after relocation. This is semantically correct,
2556 // but results in extra register pressure since both the pre-relocation and
2557 // post-relocation copies must be available in registers. For code without
2558 // relocations this is handled elsewhere, but teaching the scheduler to
2559 // reverse the transform we're about to do would be slightly complex.
2560 // Note: This may extend the live range of the inputs to the icmp and thus
2561 // increase the liveset of any statepoint we move over. This is profitable
2562 // as long as all statepoints are in rare blocks. If we had in-register
2563 // lowering for live values this would be a much safer transform.
2564 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2565 if (auto *BI = dyn_cast<BranchInst>(TI))
2566 if (BI->isConditional())
2567 return dyn_cast<Instruction>(BI->getCondition());
2568 // TODO: Extend this to handle switches
2571 for (BasicBlock &BB : F) {
2572 TerminatorInst *TI = BB.getTerminator();
2573 if (auto *Cond = getConditionInst(TI))
2574 // TODO: Handle more than just ICmps here. We should be able to move
2575 // most instructions without side effects or memory access.
2576 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2578 Cond->moveBefore(TI);
2582 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
2586 // liveness computation via standard dataflow
2587 // -------------------------------------------------------------------
2589 // TODO: Consider using bitvectors for liveness, the set of potentially
2590 // interesting values should be small and easy to pre-compute.
2592 /// Compute the live-in set for the location rbegin starting from
2593 /// the live-out set of the basic block
2594 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2595 BasicBlock::reverse_iterator rend,
2596 DenseSet<Value *> &LiveTmp) {
2598 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2599 Instruction *I = &*ritr;
2601 // KILL/Def - Remove this definition from LiveIn
2604 // Don't consider *uses* in PHI nodes, we handle their contribution to
2605 // predecessor blocks when we seed the LiveOut sets
2606 if (isa<PHINode>(I))
2609 // USE - Add to the LiveIn set for this instruction
2610 for (Value *V : I->operands()) {
2611 assert(!isUnhandledGCPointerType(V->getType()) &&
2612 "support for FCA unimplemented");
2613 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2614 // The choice to exclude all things constant here is slightly subtle.
2615 // There are two independent reasons:
2616 // - We assume that things which are constant (from LLVM's definition)
2617 // do not move at runtime. For example, the address of a global
2618 // variable is fixed, even though it's contents may not be.
2619 // - Second, we can't disallow arbitrary inttoptr constants even
2620 // if the language frontend does. Optimization passes are free to
2621 // locally exploit facts without respect to global reachability. This
2622 // can create sections of code which are dynamically unreachable and
2623 // contain just about anything. (see constants.ll in tests)
2630 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2632 for (BasicBlock *Succ : successors(BB)) {
2633 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2634 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2635 PHINode *Phi = cast<PHINode>(&*I);
2636 Value *V = Phi->getIncomingValueForBlock(BB);
2637 assert(!isUnhandledGCPointerType(V->getType()) &&
2638 "support for FCA unimplemented");
2639 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2646 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2647 DenseSet<Value *> KillSet;
2648 for (Instruction &I : *BB)
2649 if (isHandledGCPointerType(I.getType()))
2655 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2656 /// sanity check for the liveness computation.
2657 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2658 TerminatorInst *TI, bool TermOkay = false) {
2659 for (Value *V : Live) {
2660 if (auto *I = dyn_cast<Instruction>(V)) {
2661 // The terminator can be a member of the LiveOut set. LLVM's definition
2662 // of instruction dominance states that V does not dominate itself. As
2663 // such, we need to special case this to allow it.
2664 if (TermOkay && TI == I)
2666 assert(DT.dominates(I, TI) &&
2667 "basic SSA liveness expectation violated by liveness analysis");
2672 /// Check that all the liveness sets used during the computation of liveness
2673 /// obey basic SSA properties. This is useful for finding cases where we miss
2675 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2677 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2678 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2679 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2683 static void computeLiveInValues(DominatorTree &DT, Function &F,
2684 GCPtrLivenessData &Data) {
2686 SmallSetVector<BasicBlock *, 200> Worklist;
2687 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2688 // We use a SetVector so that we don't have duplicates in the worklist.
2689 Worklist.insert(pred_begin(BB), pred_end(BB));
2691 auto NextItem = [&]() {
2692 BasicBlock *BB = Worklist.back();
2693 Worklist.pop_back();
2697 // Seed the liveness for each individual block
2698 for (BasicBlock &BB : F) {
2699 Data.KillSet[&BB] = computeKillSet(&BB);
2700 Data.LiveSet[&BB].clear();
2701 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2704 for (Value *Kill : Data.KillSet[&BB])
2705 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2708 Data.LiveOut[&BB] = DenseSet<Value *>();
2709 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2710 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2711 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2712 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2713 if (!Data.LiveIn[&BB].empty())
2714 AddPredsToWorklist(&BB);
2717 // Propagate that liveness until stable
2718 while (!Worklist.empty()) {
2719 BasicBlock *BB = NextItem();
2721 // Compute our new liveout set, then exit early if it hasn't changed
2722 // despite the contribution of our successor.
2723 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2724 const auto OldLiveOutSize = LiveOut.size();
2725 for (BasicBlock *Succ : successors(BB)) {
2726 assert(Data.LiveIn.count(Succ));
2727 set_union(LiveOut, Data.LiveIn[Succ]);
2729 // assert OutLiveOut is a subset of LiveOut
2730 if (OldLiveOutSize == LiveOut.size()) {
2731 // If the sets are the same size, then we didn't actually add anything
2732 // when unioning our successors LiveIn Thus, the LiveIn of this block
2736 Data.LiveOut[BB] = LiveOut;
2738 // Apply the effects of this basic block
2739 DenseSet<Value *> LiveTmp = LiveOut;
2740 set_union(LiveTmp, Data.LiveSet[BB]);
2741 set_subtract(LiveTmp, Data.KillSet[BB]);
2743 assert(Data.LiveIn.count(BB));
2744 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2745 // assert: OldLiveIn is a subset of LiveTmp
2746 if (OldLiveIn.size() != LiveTmp.size()) {
2747 Data.LiveIn[BB] = LiveTmp;
2748 AddPredsToWorklist(BB);
2750 } // while( !worklist.empty() )
2753 // Sanity check our output against SSA properties. This helps catch any
2754 // missing kills during the above iteration.
2755 for (BasicBlock &BB : F) {
2756 checkBasicSSA(DT, Data, BB);
2761 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2762 StatepointLiveSetTy &Out) {
2764 BasicBlock *BB = Inst->getParent();
2766 // Note: The copy is intentional and required
2767 assert(Data.LiveOut.count(BB));
2768 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2770 // We want to handle the statepoint itself oddly. It's
2771 // call result is not live (normal), nor are it's arguments
2772 // (unless they're used again later). This adjustment is
2773 // specifically what we need to relocate
2774 BasicBlock::reverse_iterator rend(Inst);
2775 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2776 LiveOut.erase(Inst);
2777 Out.insert(LiveOut.begin(), LiveOut.end());
2780 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2782 PartiallyConstructedSafepointRecord &Info) {
2783 Instruction *Inst = CS.getInstruction();
2784 StatepointLiveSetTy Updated;
2785 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2788 DenseSet<Value *> Bases;
2789 for (auto KVPair : Info.PointerToBase) {
2790 Bases.insert(KVPair.second);
2793 // We may have base pointers which are now live that weren't before. We need
2794 // to update the PointerToBase structure to reflect this.
2795 for (auto V : Updated)
2796 if (!Info.PointerToBase.count(V)) {
2797 assert(Bases.count(V) && "can't find base for unexpected live value");
2798 Info.PointerToBase[V] = V;
2803 for (auto V : Updated) {
2804 assert(Info.PointerToBase.count(V) &&
2805 "must be able to find base for live value");
2809 // Remove any stale base mappings - this can happen since our liveness is
2810 // more precise then the one inherent in the base pointer analysis
2811 DenseSet<Value *> ToErase;
2812 for (auto KVPair : Info.PointerToBase)
2813 if (!Updated.count(KVPair.first))
2814 ToErase.insert(KVPair.first);
2815 for (auto V : ToErase)
2816 Info.PointerToBase.erase(V);
2819 for (auto KVPair : Info.PointerToBase)
2820 assert(Updated.count(KVPair.first) && "record for non-live value");
2823 Info.liveset = Updated;