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/ADT/MapVector.h"
25 #include "llvm/IR/BasicBlock.h"
26 #include "llvm/IR/CallSite.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/InstIterator.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Statepoint.h"
37 #include "llvm/IR/Value.h"
38 #include "llvm/IR/Verifier.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Transforms/Scalar.h"
42 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 #include "llvm/Transforms/Utils/Cloning.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
47 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
51 // Print the liveset found at the insert location
52 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
54 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden,
56 // Print out the base pointers for debugging
57 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden,
60 // Cost threshold measuring when it is profitable to rematerialize value instead
62 static cl::opt<unsigned>
63 RematerializationThreshold("spp-rematerialization-threshold", cl::Hidden,
67 static bool ClobberNonLive = true;
69 static bool ClobberNonLive = false;
71 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
72 cl::location(ClobberNonLive),
75 static cl::opt<bool> UseDeoptBundles("rs4gc-use-deopt-bundles", cl::Hidden,
78 AllowStatepointWithNoDeoptInfo("rs4gc-allow-statepoint-with-no-deopt-info",
79 cl::Hidden, cl::init(true));
82 struct RewriteStatepointsForGC : public ModulePass {
83 static char ID; // Pass identification, replacement for typeid
85 RewriteStatepointsForGC() : ModulePass(ID) {
86 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
88 bool runOnFunction(Function &F);
89 bool runOnModule(Module &M) override {
92 Changed |= runOnFunction(F);
95 // stripNonValidAttributes asserts that shouldRewriteStatepointsIn
96 // returns true for at least one function in the module. Since at least
97 // one function changed, we know that the precondition is satisfied.
98 stripNonValidAttributes(M);
104 void getAnalysisUsage(AnalysisUsage &AU) const override {
105 // We add and rewrite a bunch of instructions, but don't really do much
106 // else. We could in theory preserve a lot more analyses here.
107 AU.addRequired<DominatorTreeWrapperPass>();
108 AU.addRequired<TargetTransformInfoWrapperPass>();
111 /// The IR fed into RewriteStatepointsForGC may have had attributes implying
112 /// dereferenceability that are no longer valid/correct after
113 /// RewriteStatepointsForGC has run. This is because semantically, after
114 /// RewriteStatepointsForGC runs, all calls to gc.statepoint "free" the entire
115 /// heap. stripNonValidAttributes (conservatively) restores correctness
116 /// by erasing all attributes in the module that externally imply
117 /// dereferenceability.
118 /// Similar reasoning also applies to the noalias attributes. gc.statepoint
119 /// can touch the entire heap including noalias objects.
120 void stripNonValidAttributes(Module &M);
122 // Helpers for stripNonValidAttributes
123 void stripNonValidAttributesFromBody(Function &F);
124 void stripNonValidAttributesFromPrototype(Function &F);
128 char RewriteStatepointsForGC::ID = 0;
130 ModulePass *llvm::createRewriteStatepointsForGCPass() {
131 return new RewriteStatepointsForGC();
134 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
135 "Make relocations explicit at statepoints", false, false)
136 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
137 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
138 "Make relocations explicit at statepoints", false, false)
141 struct GCPtrLivenessData {
142 /// Values defined in this block.
143 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet;
144 /// Values used in this block (and thus live); does not included values
145 /// killed within this block.
146 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet;
148 /// Values live into this basic block (i.e. used by any
149 /// instruction in this basic block or ones reachable from here)
150 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn;
152 /// Values live out of this basic block (i.e. live into
153 /// any successor block)
154 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut;
157 // The type of the internal cache used inside the findBasePointers family
158 // of functions. From the callers perspective, this is an opaque type and
159 // should not be inspected.
161 // In the actual implementation this caches two relations:
162 // - The base relation itself (i.e. this pointer is based on that one)
163 // - The base defining value relation (i.e. before base_phi insertion)
164 // Generally, after the execution of a full findBasePointer call, only the
165 // base relation will remain. Internally, we add a mixture of the two
166 // types, then update all the second type to the first type
167 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
168 typedef DenseSet<Value *> StatepointLiveSetTy;
169 typedef DenseMap<AssertingVH<Instruction>, AssertingVH<Value>>
170 RematerializedValueMapTy;
172 struct PartiallyConstructedSafepointRecord {
173 /// The set of values known to be live across this safepoint
174 StatepointLiveSetTy LiveSet;
176 /// Mapping from live pointers to a base-defining-value
177 DenseMap<Value *, Value *> PointerToBase;
179 /// The *new* gc.statepoint instruction itself. This produces the token
180 /// that normal path gc.relocates and the gc.result are tied to.
181 Instruction *StatepointToken;
183 /// Instruction to which exceptional gc relocates are attached
184 /// Makes it easier to iterate through them during relocationViaAlloca.
185 Instruction *UnwindToken;
187 /// Record live values we are rematerialized instead of relocating.
188 /// They are not included into 'LiveSet' field.
189 /// Maps rematerialized copy to it's original value.
190 RematerializedValueMapTy RematerializedValues;
194 static ArrayRef<Use> GetDeoptBundleOperands(ImmutableCallSite CS) {
195 assert(UseDeoptBundles && "Should not be called otherwise!");
197 Optional<OperandBundleUse> DeoptBundle = CS.getOperandBundle("deopt");
199 if (!DeoptBundle.hasValue()) {
200 assert(AllowStatepointWithNoDeoptInfo &&
201 "Found non-leaf call without deopt info!");
205 return DeoptBundle.getValue().Inputs;
208 /// Compute the live-in set for every basic block in the function
209 static void computeLiveInValues(DominatorTree &DT, Function &F,
210 GCPtrLivenessData &Data);
212 /// Given results from the dataflow liveness computation, find the set of live
213 /// Values at a particular instruction.
214 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
215 StatepointLiveSetTy &out);
217 // TODO: Once we can get to the GCStrategy, this becomes
218 // Optional<bool> isGCManagedPointer(const Value *V) const override {
220 static bool isGCPointerType(Type *T) {
221 if (auto *PT = dyn_cast<PointerType>(T))
222 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
223 // GC managed heap. We know that a pointer into this heap needs to be
224 // updated and that no other pointer does.
225 return (1 == PT->getAddressSpace());
229 // Return true if this type is one which a) is a gc pointer or contains a GC
230 // pointer and b) is of a type this code expects to encounter as a live value.
231 // (The insertion code will assert that a type which matches (a) and not (b)
232 // is not encountered.)
233 static bool isHandledGCPointerType(Type *T) {
234 // We fully support gc pointers
235 if (isGCPointerType(T))
237 // We partially support vectors of gc pointers. The code will assert if it
238 // can't handle something.
239 if (auto VT = dyn_cast<VectorType>(T))
240 if (isGCPointerType(VT->getElementType()))
246 /// Returns true if this type contains a gc pointer whether we know how to
247 /// handle that type or not.
248 static bool containsGCPtrType(Type *Ty) {
249 if (isGCPointerType(Ty))
251 if (VectorType *VT = dyn_cast<VectorType>(Ty))
252 return isGCPointerType(VT->getScalarType());
253 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
254 return containsGCPtrType(AT->getElementType());
255 if (StructType *ST = dyn_cast<StructType>(Ty))
256 return std::any_of(ST->subtypes().begin(), ST->subtypes().end(),
261 // Returns true if this is a type which a) is a gc pointer or contains a GC
262 // pointer and b) is of a type which the code doesn't expect (i.e. first class
263 // aggregates). Used to trip assertions.
264 static bool isUnhandledGCPointerType(Type *Ty) {
265 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
269 static bool order_by_name(Value *a, Value *b) {
270 if (a->hasName() && b->hasName()) {
271 return -1 == a->getName().compare(b->getName());
272 } else if (a->hasName() && !b->hasName()) {
274 } else if (!a->hasName() && b->hasName()) {
277 // Better than nothing, but not stable
282 // Return the name of the value suffixed with the provided value, or if the
283 // value didn't have a name, the default value specified.
284 static std::string suffixed_name_or(Value *V, StringRef Suffix,
285 StringRef DefaultName) {
286 return V->hasName() ? (V->getName() + Suffix).str() : DefaultName.str();
289 // Conservatively identifies any definitions which might be live at the
290 // given instruction. The analysis is performed immediately before the
291 // given instruction. Values defined by that instruction are not considered
292 // live. Values used by that instruction are considered live.
293 static void analyzeParsePointLiveness(
294 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
295 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
296 Instruction *inst = CS.getInstruction();
298 StatepointLiveSetTy LiveSet;
299 findLiveSetAtInst(inst, OriginalLivenessData, LiveSet);
302 // Note: This output is used by several of the test cases
303 // The order of elements in a set is not stable, put them in a vec and sort
305 SmallVector<Value *, 64> Temp;
306 Temp.insert(Temp.end(), LiveSet.begin(), LiveSet.end());
307 std::sort(Temp.begin(), Temp.end(), order_by_name);
308 errs() << "Live Variables:\n";
309 for (Value *V : Temp)
310 dbgs() << " " << V->getName() << " " << *V << "\n";
312 if (PrintLiveSetSize) {
313 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
314 errs() << "Number live values: " << LiveSet.size() << "\n";
316 result.LiveSet = LiveSet;
319 static bool isKnownBaseResult(Value *V);
321 /// A single base defining value - An immediate base defining value for an
322 /// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'.
323 /// For instructions which have multiple pointer [vector] inputs or that
324 /// transition between vector and scalar types, there is no immediate base
325 /// defining value. The 'base defining value' for 'Def' is the transitive
326 /// closure of this relation stopping at the first instruction which has no
327 /// immediate base defining value. The b.d.v. might itself be a base pointer,
328 /// but it can also be an arbitrary derived pointer.
329 struct BaseDefiningValueResult {
330 /// Contains the value which is the base defining value.
332 /// True if the base defining value is also known to be an actual base
334 const bool IsKnownBase;
335 BaseDefiningValueResult(Value *BDV, bool IsKnownBase)
336 : BDV(BDV), IsKnownBase(IsKnownBase) {
338 // Check consistency between new and old means of checking whether a BDV is
340 bool MustBeBase = isKnownBaseResult(BDV);
341 assert(!MustBeBase || MustBeBase == IsKnownBase);
347 static BaseDefiningValueResult findBaseDefiningValue(Value *I);
349 /// Return a base defining value for the 'Index' element of the given vector
350 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
351 /// 'I'. As an optimization, this method will try to determine when the
352 /// element is known to already be a base pointer. If this can be established,
353 /// the second value in the returned pair will be true. Note that either a
354 /// vector or a pointer typed value can be returned. For the former, the
355 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
356 /// If the later, the return pointer is a BDV (or possibly a base) for the
357 /// particular element in 'I'.
358 static BaseDefiningValueResult
359 findBaseDefiningValueOfVector(Value *I) {
360 assert(I->getType()->isVectorTy() &&
361 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
362 "Illegal to ask for the base pointer of a non-pointer type");
364 // Each case parallels findBaseDefiningValue below, see that code for
365 // detailed motivation.
367 if (isa<Argument>(I))
368 // An incoming argument to the function is a base pointer
369 return BaseDefiningValueResult(I, true);
371 // We shouldn't see the address of a global as a vector value?
372 assert(!isa<GlobalVariable>(I) &&
373 "unexpected global variable found in base of vector");
375 // inlining could possibly introduce phi node that contains
376 // undef if callee has multiple returns
377 if (isa<UndefValue>(I))
378 // utterly meaningless, but useful for dealing with partially optimized
380 return BaseDefiningValueResult(I, true);
382 // Due to inheritance, this must be _after_ the global variable and undef
384 if (Constant *Con = dyn_cast<Constant>(I)) {
385 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
386 "order of checks wrong!");
387 assert(Con->isNullValue() && "null is the only case which makes sense");
388 return BaseDefiningValueResult(Con, true);
391 if (isa<LoadInst>(I))
392 return BaseDefiningValueResult(I, true);
394 if (isa<InsertElementInst>(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 return BaseDefiningValueResult(I, false);
400 if (isa<ShuffleVectorInst>(I))
401 // We don't know whether this vector contains entirely base pointers or
402 // not. To be conservatively correct, we treat it as a BDV and will
403 // duplicate code as needed to construct a parallel vector of bases.
404 // TODO: There a number of local optimizations which could be applied here
405 // for particular sufflevector patterns.
406 return BaseDefiningValueResult(I, false);
408 // A PHI or Select is a base defining value. The outer findBasePointer
409 // algorithm is responsible for constructing a base value for this BDV.
410 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
411 "unknown vector instruction - no base found for vector element");
412 return BaseDefiningValueResult(I, false);
415 /// Helper function for findBasePointer - Will return a value which either a)
416 /// defines the base pointer for the input, b) blocks the simple search
417 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
418 /// from pointer to vector type or back.
419 static BaseDefiningValueResult findBaseDefiningValue(Value *I) {
420 if (I->getType()->isVectorTy())
421 return findBaseDefiningValueOfVector(I);
423 assert(I->getType()->isPointerTy() &&
424 "Illegal to ask for the base pointer of a non-pointer type");
426 if (isa<Argument>(I))
427 // An incoming argument to the function is a base pointer
428 // We should have never reached here if this argument isn't an gc value
429 return BaseDefiningValueResult(I, true);
431 if (isa<GlobalVariable>(I))
433 return BaseDefiningValueResult(I, true);
435 // inlining could possibly introduce phi node that contains
436 // undef if callee has multiple returns
437 if (isa<UndefValue>(I))
438 // utterly meaningless, but useful for dealing with
439 // partially optimized code.
440 return BaseDefiningValueResult(I, true);
442 // Due to inheritance, this must be _after_ the global variable and undef
444 if (isa<Constant>(I)) {
445 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
446 "order of checks wrong!");
447 // Note: Even for frontends which don't have constant references, we can
448 // see constants appearing after optimizations. A simple example is
449 // specialization of an address computation on null feeding into a merge
450 // point where the actual use of the now-constant input is protected by
451 // another null check. (e.g. test4 in constants.ll)
452 return BaseDefiningValueResult(I, true);
455 if (CastInst *CI = dyn_cast<CastInst>(I)) {
456 Value *Def = CI->stripPointerCasts();
457 // If we find a cast instruction here, it means we've found a cast which is
458 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
459 // handle int->ptr conversion.
460 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
461 return findBaseDefiningValue(Def);
464 if (isa<LoadInst>(I))
465 // The value loaded is an gc base itself
466 return BaseDefiningValueResult(I, true);
469 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
470 // The base of this GEP is the base
471 return findBaseDefiningValue(GEP->getPointerOperand());
473 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
474 switch (II->getIntrinsicID()) {
475 case Intrinsic::experimental_gc_result_ptr:
477 // fall through to general call handling
479 case Intrinsic::experimental_gc_statepoint:
480 case Intrinsic::experimental_gc_result_float:
481 case Intrinsic::experimental_gc_result_int:
482 llvm_unreachable("these don't produce pointers");
483 case Intrinsic::experimental_gc_relocate: {
484 // Rerunning safepoint insertion after safepoints are already
485 // inserted is not supported. It could probably be made to work,
486 // but why are you doing this? There's no good reason.
487 llvm_unreachable("repeat safepoint insertion is not supported");
489 case Intrinsic::gcroot:
490 // Currently, this mechanism hasn't been extended to work with gcroot.
491 // There's no reason it couldn't be, but I haven't thought about the
492 // implications much.
494 "interaction with the gcroot mechanism is not supported");
497 // We assume that functions in the source language only return base
498 // pointers. This should probably be generalized via attributes to support
499 // both source language and internal functions.
500 if (isa<CallInst>(I) || isa<InvokeInst>(I))
501 return BaseDefiningValueResult(I, true);
503 // I have absolutely no idea how to implement this part yet. It's not
504 // necessarily hard, I just haven't really looked at it yet.
505 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
507 if (isa<AtomicCmpXchgInst>(I))
508 // A CAS is effectively a atomic store and load combined under a
509 // predicate. From the perspective of base pointers, we just treat it
511 return BaseDefiningValueResult(I, true);
513 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
514 "binary ops which don't apply to pointers");
516 // The aggregate ops. Aggregates can either be in the heap or on the
517 // stack, but in either case, this is simply a field load. As a result,
518 // this is a defining definition of the base just like a load is.
519 if (isa<ExtractValueInst>(I))
520 return BaseDefiningValueResult(I, true);
522 // We should never see an insert vector since that would require we be
523 // tracing back a struct value not a pointer value.
524 assert(!isa<InsertValueInst>(I) &&
525 "Base pointer for a struct is meaningless");
527 // An extractelement produces a base result exactly when it's input does.
528 // We may need to insert a parallel instruction to extract the appropriate
529 // element out of the base vector corresponding to the input. Given this,
530 // it's analogous to the phi and select case even though it's not a merge.
531 if (isa<ExtractElementInst>(I))
532 // Note: There a lot of obvious peephole cases here. This are deliberately
533 // handled after the main base pointer inference algorithm to make writing
534 // test cases to exercise that code easier.
535 return BaseDefiningValueResult(I, false);
537 // The last two cases here don't return a base pointer. Instead, they
538 // return a value which dynamically selects from among several base
539 // derived pointers (each with it's own base potentially). It's the job of
540 // the caller to resolve these.
541 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
542 "missing instruction case in findBaseDefiningValing");
543 return BaseDefiningValueResult(I, false);
546 /// Returns the base defining value for this value.
547 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
548 Value *&Cached = Cache[I];
550 Cached = findBaseDefiningValue(I).BDV;
551 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
552 << Cached->getName() << "\n");
554 assert(Cache[I] != nullptr);
558 /// Return a base pointer for this value if known. Otherwise, return it's
559 /// base defining value.
560 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
561 Value *Def = findBaseDefiningValueCached(I, Cache);
562 auto Found = Cache.find(Def);
563 if (Found != Cache.end()) {
564 // Either a base-of relation, or a self reference. Caller must check.
565 return Found->second;
567 // Only a BDV available
571 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
572 /// is it known to be a base pointer? Or do we need to continue searching.
573 static bool isKnownBaseResult(Value *V) {
574 if (!isa<PHINode>(V) && !isa<SelectInst>(V) &&
575 !isa<ExtractElementInst>(V) && !isa<InsertElementInst>(V) &&
576 !isa<ShuffleVectorInst>(V)) {
577 // no recursion possible
580 if (isa<Instruction>(V) &&
581 cast<Instruction>(V)->getMetadata("is_base_value")) {
582 // This is a previously inserted base phi or select. We know
583 // that this is a base value.
587 // We need to keep searching
592 /// Models the state of a single base defining value in the findBasePointer
593 /// algorithm for determining where a new instruction is needed to propagate
594 /// the base of this BDV.
597 enum Status { Unknown, Base, Conflict };
599 BDVState(Status s, Value *b = nullptr) : status(s), base(b) {
600 assert(status != Base || b);
602 explicit BDVState(Value *b) : status(Base), base(b) {}
603 BDVState() : status(Unknown), base(nullptr) {}
605 Status getStatus() const { return status; }
606 Value *getBase() const { return base; }
608 bool isBase() const { return getStatus() == Base; }
609 bool isUnknown() const { return getStatus() == Unknown; }
610 bool isConflict() const { return getStatus() == Conflict; }
612 bool operator==(const BDVState &other) const {
613 return base == other.base && status == other.status;
616 bool operator!=(const BDVState &other) const { return !(*this == other); }
619 void dump() const { print(dbgs()); dbgs() << '\n'; }
621 void print(raw_ostream &OS) const {
633 OS << " (" << base << " - "
634 << (base ? base->getName() : "nullptr") << "): ";
639 AssertingVH<Value> base; // non null only if status == base
644 static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
651 // Values of type BDVState form a lattice, and this is a helper
652 // class that implementes the meet operation. The meat of the meet
653 // operation is implemented in MeetBDVStates::pureMeet
654 class MeetBDVStates {
656 /// Initializes the currentResult to the TOP state so that if can be met with
657 /// any other state to produce that state.
660 // Destructively meet the current result with the given BDVState
661 void meetWith(BDVState otherState) {
662 currentResult = meet(otherState, currentResult);
665 BDVState getResult() const { return currentResult; }
668 BDVState currentResult;
670 /// Perform a meet operation on two elements of the BDVState lattice.
671 static BDVState meet(BDVState LHS, BDVState RHS) {
672 assert((pureMeet(LHS, RHS) == pureMeet(RHS, LHS)) &&
673 "math is wrong: meet does not commute!");
674 BDVState Result = pureMeet(LHS, RHS);
675 DEBUG(dbgs() << "meet of " << LHS << " with " << RHS
676 << " produced " << Result << "\n");
680 static BDVState pureMeet(const BDVState &stateA, const BDVState &stateB) {
681 switch (stateA.getStatus()) {
682 case BDVState::Unknown:
686 assert(stateA.getBase() && "can't be null");
687 if (stateB.isUnknown())
690 if (stateB.isBase()) {
691 if (stateA.getBase() == stateB.getBase()) {
692 assert(stateA == stateB && "equality broken!");
695 return BDVState(BDVState::Conflict);
697 assert(stateB.isConflict() && "only three states!");
698 return BDVState(BDVState::Conflict);
700 case BDVState::Conflict:
703 llvm_unreachable("only three states!");
709 /// For a given value or instruction, figure out what base ptr it's derived
710 /// from. For gc objects, this is simply itself. On success, returns a value
711 /// which is the base pointer. (This is reliable and can be used for
712 /// relocation.) On failure, returns nullptr.
713 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
714 Value *def = findBaseOrBDV(I, cache);
716 if (isKnownBaseResult(def)) {
720 // Here's the rough algorithm:
721 // - For every SSA value, construct a mapping to either an actual base
722 // pointer or a PHI which obscures the base pointer.
723 // - Construct a mapping from PHI to unknown TOP state. Use an
724 // optimistic algorithm to propagate base pointer information. Lattice
729 // When algorithm terminates, all PHIs will either have a single concrete
730 // base or be in a conflict state.
731 // - For every conflict, insert a dummy PHI node without arguments. Add
732 // these to the base[Instruction] = BasePtr mapping. For every
733 // non-conflict, add the actual base.
734 // - For every conflict, add arguments for the base[a] of each input
737 // Note: A simpler form of this would be to add the conflict form of all
738 // PHIs without running the optimistic algorithm. This would be
739 // analogous to pessimistic data flow and would likely lead to an
740 // overall worse solution.
743 auto isExpectedBDVType = [](Value *BDV) {
744 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) ||
745 isa<ExtractElementInst>(BDV) || isa<InsertElementInst>(BDV);
749 // Once populated, will contain a mapping from each potentially non-base BDV
750 // to a lattice value (described above) which corresponds to that BDV.
751 // We use the order of insertion (DFS over the def/use graph) to provide a
752 // stable deterministic ordering for visiting DenseMaps (which are unordered)
753 // below. This is important for deterministic compilation.
754 MapVector<Value *, BDVState> States;
756 // Recursively fill in all base defining values reachable from the initial
757 // one for which we don't already know a definite base value for
759 SmallVector<Value*, 16> Worklist;
760 Worklist.push_back(def);
761 States.insert(std::make_pair(def, BDVState()));
762 while (!Worklist.empty()) {
763 Value *Current = Worklist.pop_back_val();
764 assert(!isKnownBaseResult(Current) && "why did it get added?");
766 auto visitIncomingValue = [&](Value *InVal) {
767 Value *Base = findBaseOrBDV(InVal, cache);
768 if (isKnownBaseResult(Base))
769 // Known bases won't need new instructions introduced and can be
772 assert(isExpectedBDVType(Base) && "the only non-base values "
773 "we see should be base defining values");
774 if (States.insert(std::make_pair(Base, BDVState())).second)
775 Worklist.push_back(Base);
777 if (PHINode *Phi = dyn_cast<PHINode>(Current)) {
778 for (Value *InVal : Phi->incoming_values())
779 visitIncomingValue(InVal);
780 } else if (SelectInst *Sel = dyn_cast<SelectInst>(Current)) {
781 visitIncomingValue(Sel->getTrueValue());
782 visitIncomingValue(Sel->getFalseValue());
783 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
784 visitIncomingValue(EE->getVectorOperand());
785 } else if (auto *IE = dyn_cast<InsertElementInst>(Current)) {
786 visitIncomingValue(IE->getOperand(0)); // vector operand
787 visitIncomingValue(IE->getOperand(1)); // scalar operand
789 // There is one known class of instructions we know we don't handle.
790 assert(isa<ShuffleVectorInst>(Current));
791 llvm_unreachable("unimplemented instruction case");
797 DEBUG(dbgs() << "States after initialization:\n");
798 for (auto Pair : States) {
799 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
803 // Return a phi state for a base defining value. We'll generate a new
804 // base state for known bases and expect to find a cached state otherwise.
805 auto getStateForBDV = [&](Value *baseValue) {
806 if (isKnownBaseResult(baseValue))
807 return BDVState(baseValue);
808 auto I = States.find(baseValue);
809 assert(I != States.end() && "lookup failed!");
813 bool progress = true;
816 const size_t oldSize = States.size();
819 // We're only changing values in this loop, thus safe to keep iterators.
820 // Since this is computing a fixed point, the order of visit does not
821 // effect the result. TODO: We could use a worklist here and make this run
823 for (auto Pair : States) {
824 Value *BDV = Pair.first;
825 assert(!isKnownBaseResult(BDV) && "why did it get added?");
827 // Given an input value for the current instruction, return a BDVState
828 // instance which represents the BDV of that value.
829 auto getStateForInput = [&](Value *V) mutable {
830 Value *BDV = findBaseOrBDV(V, cache);
831 return getStateForBDV(BDV);
834 MeetBDVStates calculateMeet;
835 if (SelectInst *select = dyn_cast<SelectInst>(BDV)) {
836 calculateMeet.meetWith(getStateForInput(select->getTrueValue()));
837 calculateMeet.meetWith(getStateForInput(select->getFalseValue()));
838 } else if (PHINode *Phi = dyn_cast<PHINode>(BDV)) {
839 for (Value *Val : Phi->incoming_values())
840 calculateMeet.meetWith(getStateForInput(Val));
841 } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) {
842 // The 'meet' for an extractelement is slightly trivial, but it's still
843 // useful in that it drives us to conflict if our input is.
844 calculateMeet.meetWith(getStateForInput(EE->getVectorOperand()));
846 // Given there's a inherent type mismatch between the operands, will
847 // *always* produce Conflict.
848 auto *IE = cast<InsertElementInst>(BDV);
849 calculateMeet.meetWith(getStateForInput(IE->getOperand(0)));
850 calculateMeet.meetWith(getStateForInput(IE->getOperand(1)));
853 BDVState oldState = States[BDV];
854 BDVState newState = calculateMeet.getResult();
855 if (oldState != newState) {
857 States[BDV] = newState;
861 assert(oldSize == States.size() &&
862 "fixed point shouldn't be adding any new nodes to state");
866 DEBUG(dbgs() << "States after meet iteration:\n");
867 for (auto Pair : States) {
868 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
872 // Insert Phis for all conflicts
873 // TODO: adjust naming patterns to avoid this order of iteration dependency
874 for (auto Pair : States) {
875 Instruction *I = cast<Instruction>(Pair.first);
876 BDVState State = Pair.second;
877 assert(!isKnownBaseResult(I) && "why did it get added?");
878 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
880 // extractelement instructions are a bit special in that we may need to
881 // insert an extract even when we know an exact base for the instruction.
882 // The problem is that we need to convert from a vector base to a scalar
883 // base for the particular indice we're interested in.
884 if (State.isBase() && isa<ExtractElementInst>(I) &&
885 isa<VectorType>(State.getBase()->getType())) {
886 auto *EE = cast<ExtractElementInst>(I);
887 // TODO: In many cases, the new instruction is just EE itself. We should
888 // exploit this, but can't do it here since it would break the invariant
889 // about the BDV not being known to be a base.
890 auto *BaseInst = ExtractElementInst::Create(State.getBase(),
891 EE->getIndexOperand(),
893 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
894 States[I] = BDVState(BDVState::Base, BaseInst);
897 // Since we're joining a vector and scalar base, they can never be the
898 // same. As a result, we should always see insert element having reached
899 // the conflict state.
900 if (isa<InsertElementInst>(I)) {
901 assert(State.isConflict());
904 if (!State.isConflict())
907 /// Create and insert a new instruction which will represent the base of
908 /// the given instruction 'I'.
909 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
910 if (isa<PHINode>(I)) {
911 BasicBlock *BB = I->getParent();
912 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
913 assert(NumPreds > 0 && "how did we reach here");
914 std::string Name = suffixed_name_or(I, ".base", "base_phi");
915 return PHINode::Create(I->getType(), NumPreds, Name, I);
916 } else if (SelectInst *Sel = dyn_cast<SelectInst>(I)) {
917 // The undef will be replaced later
918 UndefValue *Undef = UndefValue::get(Sel->getType());
919 std::string Name = suffixed_name_or(I, ".base", "base_select");
920 return SelectInst::Create(Sel->getCondition(), Undef,
922 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
923 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
924 std::string Name = suffixed_name_or(I, ".base", "base_ee");
925 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
928 auto *IE = cast<InsertElementInst>(I);
929 UndefValue *VecUndef = UndefValue::get(IE->getOperand(0)->getType());
930 UndefValue *ScalarUndef = UndefValue::get(IE->getOperand(1)->getType());
931 std::string Name = suffixed_name_or(I, ".base", "base_ie");
932 return InsertElementInst::Create(VecUndef, ScalarUndef,
933 IE->getOperand(2), Name, IE);
937 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
938 // Add metadata marking this as a base value
939 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
940 States[I] = BDVState(BDVState::Conflict, BaseInst);
943 // Returns a instruction which produces the base pointer for a given
944 // instruction. The instruction is assumed to be an input to one of the BDVs
945 // seen in the inference algorithm above. As such, we must either already
946 // know it's base defining value is a base, or have inserted a new
947 // instruction to propagate the base of it's BDV and have entered that newly
948 // introduced instruction into the state table. In either case, we are
949 // assured to be able to determine an instruction which produces it's base
951 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
952 Value *BDV = findBaseOrBDV(Input, cache);
953 Value *Base = nullptr;
954 if (isKnownBaseResult(BDV)) {
957 // Either conflict or base.
958 assert(States.count(BDV));
959 Base = States[BDV].getBase();
961 assert(Base && "can't be null");
962 // The cast is needed since base traversal may strip away bitcasts
963 if (Base->getType() != Input->getType() &&
965 Base = new BitCastInst(Base, Input->getType(), "cast",
971 // Fixup all the inputs of the new PHIs. Visit order needs to be
972 // deterministic and predictable because we're naming newly created
974 for (auto Pair : States) {
975 Instruction *BDV = cast<Instruction>(Pair.first);
976 BDVState State = Pair.second;
978 assert(!isKnownBaseResult(BDV) && "why did it get added?");
979 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
980 if (!State.isConflict())
983 if (PHINode *basephi = dyn_cast<PHINode>(State.getBase())) {
984 PHINode *phi = cast<PHINode>(BDV);
985 unsigned NumPHIValues = phi->getNumIncomingValues();
986 for (unsigned i = 0; i < NumPHIValues; i++) {
987 Value *InVal = phi->getIncomingValue(i);
988 BasicBlock *InBB = phi->getIncomingBlock(i);
990 // If we've already seen InBB, add the same incoming value
991 // we added for it earlier. The IR verifier requires phi
992 // nodes with multiple entries from the same basic block
993 // to have the same incoming value for each of those
994 // entries. If we don't do this check here and basephi
995 // has a different type than base, we'll end up adding two
996 // bitcasts (and hence two distinct values) as incoming
997 // values for the same basic block.
999 int blockIndex = basephi->getBasicBlockIndex(InBB);
1000 if (blockIndex != -1) {
1001 Value *oldBase = basephi->getIncomingValue(blockIndex);
1002 basephi->addIncoming(oldBase, InBB);
1005 Value *Base = getBaseForInput(InVal, nullptr);
1006 // In essence this assert states: the only way two
1007 // values incoming from the same basic block may be
1008 // different is by being different bitcasts of the same
1009 // value. A cleanup that remains TODO is changing
1010 // findBaseOrBDV to return an llvm::Value of the correct
1011 // type (and still remain pure). This will remove the
1012 // need to add bitcasts.
1013 assert(Base->stripPointerCasts() == oldBase->stripPointerCasts() &&
1014 "sanity -- findBaseOrBDV should be pure!");
1019 // Find the instruction which produces the base for each input. We may
1020 // need to insert a bitcast in the incoming block.
1021 // TODO: Need to split critical edges if insertion is needed
1022 Value *Base = getBaseForInput(InVal, InBB->getTerminator());
1023 basephi->addIncoming(Base, InBB);
1025 assert(basephi->getNumIncomingValues() == NumPHIValues);
1026 } else if (SelectInst *BaseSel = dyn_cast<SelectInst>(State.getBase())) {
1027 SelectInst *Sel = cast<SelectInst>(BDV);
1028 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
1029 // something more safe and less hacky.
1030 for (int i = 1; i <= 2; i++) {
1031 Value *InVal = Sel->getOperand(i);
1032 // Find the instruction which produces the base for each input. We may
1033 // need to insert a bitcast.
1034 Value *Base = getBaseForInput(InVal, BaseSel);
1035 BaseSel->setOperand(i, Base);
1037 } else if (auto *BaseEE = dyn_cast<ExtractElementInst>(State.getBase())) {
1038 Value *InVal = cast<ExtractElementInst>(BDV)->getVectorOperand();
1039 // Find the instruction which produces the base for each input. We may
1040 // need to insert a bitcast.
1041 Value *Base = getBaseForInput(InVal, BaseEE);
1042 BaseEE->setOperand(0, Base);
1044 auto *BaseIE = cast<InsertElementInst>(State.getBase());
1045 auto *BdvIE = cast<InsertElementInst>(BDV);
1046 auto UpdateOperand = [&](int OperandIdx) {
1047 Value *InVal = BdvIE->getOperand(OperandIdx);
1048 Value *Base = getBaseForInput(InVal, BaseIE);
1049 BaseIE->setOperand(OperandIdx, Base);
1051 UpdateOperand(0); // vector operand
1052 UpdateOperand(1); // scalar operand
1057 // Now that we're done with the algorithm, see if we can optimize the
1058 // results slightly by reducing the number of new instructions needed.
1059 // Arguably, this should be integrated into the algorithm above, but
1060 // doing as a post process step is easier to reason about for the moment.
1061 DenseMap<Value *, Value *> ReverseMap;
1062 SmallPtrSet<Instruction *, 16> NewInsts;
1063 SmallSetVector<AssertingVH<Instruction>, 16> Worklist;
1064 // Note: We need to visit the states in a deterministic order. We uses the
1065 // Keys we sorted above for this purpose. Note that we are papering over a
1066 // bigger problem with the algorithm above - it's visit order is not
1067 // deterministic. A larger change is needed to fix this.
1068 for (auto Pair : States) {
1069 auto *BDV = Pair.first;
1070 auto State = Pair.second;
1071 Value *Base = State.getBase();
1072 assert(BDV && Base);
1073 assert(!isKnownBaseResult(BDV) && "why did it get added?");
1074 assert(isKnownBaseResult(Base) &&
1075 "must be something we 'know' is a base pointer");
1076 if (!State.isConflict())
1079 ReverseMap[Base] = BDV;
1080 if (auto *BaseI = dyn_cast<Instruction>(Base)) {
1081 NewInsts.insert(BaseI);
1082 Worklist.insert(BaseI);
1085 auto ReplaceBaseInstWith = [&](Value *BDV, Instruction *BaseI,
1086 Value *Replacement) {
1087 // Add users which are new instructions (excluding self references)
1088 for (User *U : BaseI->users())
1089 if (auto *UI = dyn_cast<Instruction>(U))
1090 if (NewInsts.count(UI) && UI != BaseI)
1091 Worklist.insert(UI);
1092 // Then do the actual replacement
1093 NewInsts.erase(BaseI);
1094 ReverseMap.erase(BaseI);
1095 BaseI->replaceAllUsesWith(Replacement);
1096 assert(States.count(BDV));
1097 assert(States[BDV].isConflict() && States[BDV].getBase() == BaseI);
1098 States[BDV] = BDVState(BDVState::Conflict, Replacement);
1099 BaseI->eraseFromParent();
1101 const DataLayout &DL = cast<Instruction>(def)->getModule()->getDataLayout();
1102 while (!Worklist.empty()) {
1103 Instruction *BaseI = Worklist.pop_back_val();
1104 assert(NewInsts.count(BaseI));
1105 Value *Bdv = ReverseMap[BaseI];
1106 if (auto *BdvI = dyn_cast<Instruction>(Bdv))
1107 if (BaseI->isIdenticalTo(BdvI)) {
1108 DEBUG(dbgs() << "Identical Base: " << *BaseI << "\n");
1109 ReplaceBaseInstWith(Bdv, BaseI, Bdv);
1112 if (Value *V = SimplifyInstruction(BaseI, DL)) {
1113 DEBUG(dbgs() << "Base " << *BaseI << " simplified to " << *V << "\n");
1114 ReplaceBaseInstWith(Bdv, BaseI, V);
1119 // Cache all of our results so we can cheaply reuse them
1120 // NOTE: This is actually two caches: one of the base defining value
1121 // relation and one of the base pointer relation! FIXME
1122 for (auto Pair : States) {
1123 auto *BDV = Pair.first;
1124 Value *base = Pair.second.getBase();
1125 assert(BDV && base);
1127 std::string fromstr = cache.count(BDV) ? cache[BDV]->getName() : "none";
1128 DEBUG(dbgs() << "Updating base value cache"
1129 << " for: " << BDV->getName()
1130 << " from: " << fromstr
1131 << " to: " << base->getName() << "\n");
1133 if (cache.count(BDV)) {
1134 // Once we transition from the BDV relation being store in the cache to
1135 // the base relation being stored, it must be stable
1136 assert((!isKnownBaseResult(cache[BDV]) || cache[BDV] == base) &&
1137 "base relation should be stable");
1141 assert(cache.find(def) != cache.end());
1145 // For a set of live pointers (base and/or derived), identify the base
1146 // pointer of the object which they are derived from. This routine will
1147 // mutate the IR graph as needed to make the 'base' pointer live at the
1148 // definition site of 'derived'. This ensures that any use of 'derived' can
1149 // also use 'base'. This may involve the insertion of a number of
1150 // additional PHI nodes.
1152 // preconditions: live is a set of pointer type Values
1154 // side effects: may insert PHI nodes into the existing CFG, will preserve
1155 // CFG, will not remove or mutate any existing nodes
1157 // post condition: PointerToBase contains one (derived, base) pair for every
1158 // pointer in live. Note that derived can be equal to base if the original
1159 // pointer was a base pointer.
1161 findBasePointers(const StatepointLiveSetTy &live,
1162 DenseMap<Value *, Value *> &PointerToBase,
1163 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1164 // For the naming of values inserted to be deterministic - which makes for
1165 // much cleaner and more stable tests - we need to assign an order to the
1166 // live values. DenseSets do not provide a deterministic order across runs.
1167 SmallVector<Value *, 64> Temp;
1168 Temp.insert(Temp.end(), live.begin(), live.end());
1169 std::sort(Temp.begin(), Temp.end(), order_by_name);
1170 for (Value *ptr : Temp) {
1171 Value *base = findBasePointer(ptr, DVCache);
1172 assert(base && "failed to find base pointer");
1173 PointerToBase[ptr] = base;
1174 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1175 DT->dominates(cast<Instruction>(base)->getParent(),
1176 cast<Instruction>(ptr)->getParent())) &&
1177 "The base we found better dominate the derived pointer");
1179 // If you see this trip and like to live really dangerously, the code should
1180 // be correct, just with idioms the verifier can't handle. You can try
1181 // disabling the verifier at your own substantial risk.
1182 assert(!isa<ConstantPointerNull>(base) &&
1183 "the relocation code needs adjustment to handle the relocation of "
1184 "a null pointer constant without causing false positives in the "
1185 "safepoint ir verifier.");
1189 /// Find the required based pointers (and adjust the live set) for the given
1191 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1193 PartiallyConstructedSafepointRecord &result) {
1194 DenseMap<Value *, Value *> PointerToBase;
1195 findBasePointers(result.LiveSet, PointerToBase, &DT, DVCache);
1197 if (PrintBasePointers) {
1198 // Note: Need to print these in a stable order since this is checked in
1200 errs() << "Base Pairs (w/o Relocation):\n";
1201 SmallVector<Value *, 64> Temp;
1202 Temp.reserve(PointerToBase.size());
1203 for (auto Pair : PointerToBase) {
1204 Temp.push_back(Pair.first);
1206 std::sort(Temp.begin(), Temp.end(), order_by_name);
1207 for (Value *Ptr : Temp) {
1208 Value *Base = PointerToBase[Ptr];
1209 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
1214 result.PointerToBase = PointerToBase;
1217 /// Given an updated version of the dataflow liveness results, update the
1218 /// liveset and base pointer maps for the call site CS.
1219 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1221 PartiallyConstructedSafepointRecord &result);
1223 static void recomputeLiveInValues(
1224 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1225 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1226 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1227 // again. The old values are still live and will help it stabilize quickly.
1228 GCPtrLivenessData RevisedLivenessData;
1229 computeLiveInValues(DT, F, RevisedLivenessData);
1230 for (size_t i = 0; i < records.size(); i++) {
1231 struct PartiallyConstructedSafepointRecord &info = records[i];
1232 const CallSite &CS = toUpdate[i];
1233 recomputeLiveInValues(RevisedLivenessData, CS, info);
1237 // When inserting gc.relocate and gc.result calls, we need to ensure there are
1238 // no uses of the original value / return value between the gc.statepoint and
1239 // the gc.relocate / gc.result call. One case which can arise is a phi node
1240 // starting one of the successor blocks. We also need to be able to insert the
1241 // gc.relocates only on the path which goes through the statepoint. We might
1242 // need to split an edge to make this possible.
1244 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1245 DominatorTree &DT) {
1246 BasicBlock *Ret = BB;
1247 if (!BB->getUniquePredecessor())
1248 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1250 // Now that 'Ret' has unique predecessor we can safely remove all phi nodes
1252 FoldSingleEntryPHINodes(Ret);
1253 assert(!isa<PHINode>(Ret->begin()) &&
1254 "All PHI nodes should have been removed!");
1256 // At this point, we can safely insert a gc.relocate or gc.result as the first
1257 // instruction in Ret if needed.
1261 // Create new attribute set containing only attributes which can be transferred
1262 // from original call to the safepoint.
1263 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1266 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1267 unsigned Index = AS.getSlotIndex(Slot);
1269 if (Index == AttributeSet::ReturnIndex ||
1270 Index == AttributeSet::FunctionIndex) {
1272 for (Attribute Attr : make_range(AS.begin(Slot), AS.end(Slot))) {
1274 // Do not allow certain attributes - just skip them
1275 // Safepoint can not be read only or read none.
1276 if (Attr.hasAttribute(Attribute::ReadNone) ||
1277 Attr.hasAttribute(Attribute::ReadOnly))
1280 // These attributes control the generation of the gc.statepoint call /
1281 // invoke itself; and once the gc.statepoint is in place, they're of no
1283 if (Attr.hasAttribute("statepoint-num-patch-bytes") ||
1284 Attr.hasAttribute("statepoint-id"))
1287 Ret = Ret.addAttributes(
1288 AS.getContext(), Index,
1289 AttributeSet::get(AS.getContext(), Index, AttrBuilder(Attr)));
1293 // Just skip parameter attributes for now
1299 /// Helper function to place all gc relocates necessary for the given
1302 /// liveVariables - list of variables to be relocated.
1303 /// liveStart - index of the first live variable.
1304 /// basePtrs - base pointers.
1305 /// statepointToken - statepoint instruction to which relocates should be
1307 /// Builder - Llvm IR builder to be used to construct new calls.
1308 static void CreateGCRelocates(ArrayRef<Value *> LiveVariables,
1309 const int LiveStart,
1310 ArrayRef<Value *> BasePtrs,
1311 Instruction *StatepointToken,
1312 IRBuilder<> Builder) {
1313 if (LiveVariables.empty())
1316 auto FindIndex = [](ArrayRef<Value *> LiveVec, Value *Val) {
1317 auto ValIt = std::find(LiveVec.begin(), LiveVec.end(), Val);
1318 assert(ValIt != LiveVec.end() && "Val not found in LiveVec!");
1319 size_t Index = std::distance(LiveVec.begin(), ValIt);
1320 assert(Index < LiveVec.size() && "Bug in std::find?");
1324 // All gc_relocate are set to i8 addrspace(1)* type. We originally generated
1325 // unique declarations for each pointer type, but this proved problematic
1326 // because the intrinsic mangling code is incomplete and fragile. Since
1327 // we're moving towards a single unified pointer type anyways, we can just
1328 // cast everything to an i8* of the right address space. A bitcast is added
1329 // later to convert gc_relocate to the actual value's type.
1330 Module *M = StatepointToken->getModule();
1331 auto AS = cast<PointerType>(LiveVariables[0]->getType())->getAddressSpace();
1332 Type *Types[] = {Type::getInt8PtrTy(M->getContext(), AS)};
1333 Value *GCRelocateDecl =
1334 Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, Types);
1336 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1337 // Generate the gc.relocate call and save the result
1339 Builder.getInt32(LiveStart + FindIndex(LiveVariables, BasePtrs[i]));
1340 Value *LiveIdx = Builder.getInt32(LiveStart + i);
1342 // only specify a debug name if we can give a useful one
1343 CallInst *Reloc = Builder.CreateCall(
1344 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1345 suffixed_name_or(LiveVariables[i], ".relocated", ""));
1346 // Trick CodeGen into thinking there are lots of free registers at this
1348 Reloc->setCallingConv(CallingConv::Cold);
1354 /// This struct is used to defer RAUWs and `eraseFromParent` s. Using this
1355 /// avoids having to worry about keeping around dangling pointers to Values.
1356 class DeferredReplacement {
1357 AssertingVH<Instruction> Old;
1358 AssertingVH<Instruction> New;
1361 explicit DeferredReplacement(Instruction *Old, Instruction *New) :
1362 Old(Old), New(New) {
1363 assert(Old != New && "Not allowed!");
1366 /// Does the task represented by this instance.
1367 void doReplacement() {
1368 Instruction *OldI = Old;
1369 Instruction *NewI = New;
1371 assert(OldI != NewI && "Disallowed at construction?!");
1377 OldI->replaceAllUsesWith(NewI);
1378 OldI->eraseFromParent();
1384 makeStatepointExplicitImpl(const CallSite CS, /* to replace */
1385 const SmallVectorImpl<Value *> &BasePtrs,
1386 const SmallVectorImpl<Value *> &LiveVariables,
1387 PartiallyConstructedSafepointRecord &Result,
1388 std::vector<DeferredReplacement> &Replacements) {
1389 assert(BasePtrs.size() == LiveVariables.size());
1390 assert((UseDeoptBundles || isStatepoint(CS)) &&
1391 "This method expects to be rewriting a statepoint");
1393 // Then go ahead and use the builder do actually do the inserts. We insert
1394 // immediately before the previous instruction under the assumption that all
1395 // arguments will be available here. We can't insert afterwards since we may
1396 // be replacing a terminator.
1397 Instruction *InsertBefore = CS.getInstruction();
1398 IRBuilder<> Builder(InsertBefore);
1400 ArrayRef<Value *> GCArgs(LiveVariables);
1401 uint64_t StatepointID = 0xABCDEF00;
1402 uint32_t NumPatchBytes = 0;
1403 uint32_t Flags = uint32_t(StatepointFlags::None);
1405 ArrayRef<Use> CallArgs;
1406 ArrayRef<Use> DeoptArgs;
1407 ArrayRef<Use> TransitionArgs;
1409 Value *CallTarget = nullptr;
1411 if (UseDeoptBundles) {
1412 CallArgs = {CS.arg_begin(), CS.arg_end()};
1413 DeoptArgs = GetDeoptBundleOperands(CS);
1414 // TODO: we don't fill in TransitionArgs or Flags in this branch, but we
1415 // could have an operand bundle for that too.
1416 AttributeSet OriginalAttrs = CS.getAttributes();
1418 Attribute AttrID = OriginalAttrs.getAttribute(AttributeSet::FunctionIndex,
1420 if (AttrID.isStringAttribute())
1421 AttrID.getValueAsString().getAsInteger(10, StatepointID);
1423 Attribute AttrNumPatchBytes = OriginalAttrs.getAttribute(
1424 AttributeSet::FunctionIndex, "statepoint-num-patch-bytes");
1425 if (AttrNumPatchBytes.isStringAttribute())
1426 AttrNumPatchBytes.getValueAsString().getAsInteger(10, NumPatchBytes);
1428 CallTarget = CS.getCalledValue();
1430 // This branch will be gone soon, and we will soon only support the
1431 // UseDeoptBundles == true configuration.
1432 Statepoint OldSP(CS);
1433 StatepointID = OldSP.getID();
1434 NumPatchBytes = OldSP.getNumPatchBytes();
1435 Flags = OldSP.getFlags();
1437 CallArgs = {OldSP.arg_begin(), OldSP.arg_end()};
1438 DeoptArgs = {OldSP.vm_state_begin(), OldSP.vm_state_end()};
1439 TransitionArgs = {OldSP.gc_transition_args_begin(),
1440 OldSP.gc_transition_args_end()};
1441 CallTarget = OldSP.getCalledValue();
1444 // Create the statepoint given all the arguments
1445 Instruction *Token = nullptr;
1446 AttributeSet ReturnAttrs;
1448 CallInst *ToReplace = cast<CallInst>(CS.getInstruction());
1449 CallInst *Call = Builder.CreateGCStatepointCall(
1450 StatepointID, NumPatchBytes, CallTarget, Flags, CallArgs,
1451 TransitionArgs, DeoptArgs, GCArgs, "safepoint_token");
1453 Call->setTailCall(ToReplace->isTailCall());
1454 Call->setCallingConv(ToReplace->getCallingConv());
1456 // Currently we will fail on parameter attributes and on certain
1457 // function attributes.
1458 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes());
1459 // In case if we can handle this set of attributes - set up function attrs
1460 // directly on statepoint and return attrs later for gc_result intrinsic.
1461 Call->setAttributes(NewAttrs.getFnAttributes());
1462 ReturnAttrs = NewAttrs.getRetAttributes();
1466 // Put the following gc_result and gc_relocate calls immediately after the
1467 // the old call (which we're about to delete)
1468 assert(ToReplace->getNextNode() && "Not a terminator, must have next!");
1469 Builder.SetInsertPoint(ToReplace->getNextNode());
1470 Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc());
1472 InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction());
1474 // Insert the new invoke into the old block. We'll remove the old one in a
1475 // moment at which point this will become the new terminator for the
1477 InvokeInst *Invoke = Builder.CreateGCStatepointInvoke(
1478 StatepointID, NumPatchBytes, CallTarget, ToReplace->getNormalDest(),
1479 ToReplace->getUnwindDest(), Flags, CallArgs, TransitionArgs, DeoptArgs,
1480 GCArgs, "statepoint_token");
1482 Invoke->setCallingConv(ToReplace->getCallingConv());
1484 // Currently we will fail on parameter attributes and on certain
1485 // function attributes.
1486 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes());
1487 // In case if we can handle this set of attributes - set up function attrs
1488 // directly on statepoint and return attrs later for gc_result intrinsic.
1489 Invoke->setAttributes(NewAttrs.getFnAttributes());
1490 ReturnAttrs = NewAttrs.getRetAttributes();
1494 // Generate gc relocates in exceptional path
1495 BasicBlock *UnwindBlock = ToReplace->getUnwindDest();
1496 assert(!isa<PHINode>(UnwindBlock->begin()) &&
1497 UnwindBlock->getUniquePredecessor() &&
1498 "can't safely insert in this block!");
1500 Builder.SetInsertPoint(&*UnwindBlock->getFirstInsertionPt());
1501 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1503 // Extract second element from landingpad return value. We will attach
1504 // exceptional gc relocates to it.
1505 Instruction *ExceptionalToken =
1506 cast<Instruction>(Builder.CreateExtractValue(
1507 UnwindBlock->getLandingPadInst(), 1, "relocate_token"));
1508 Result.UnwindToken = ExceptionalToken;
1510 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1511 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, ExceptionalToken,
1514 // Generate gc relocates and returns for normal block
1515 BasicBlock *NormalDest = ToReplace->getNormalDest();
1516 assert(!isa<PHINode>(NormalDest->begin()) &&
1517 NormalDest->getUniquePredecessor() &&
1518 "can't safely insert in this block!");
1520 Builder.SetInsertPoint(&*NormalDest->getFirstInsertionPt());
1522 // gc relocates will be generated later as if it were regular call
1525 assert(Token && "Should be set in one of the above branches!");
1527 if (UseDeoptBundles) {
1528 Token->setName("statepoint_token");
1529 if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
1531 CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "";
1532 CallInst *GCResult = Builder.CreateGCResult(Token, CS.getType(), Name);
1533 GCResult->setAttributes(CS.getAttributes().getRetAttributes());
1535 // We cannot RAUW or delete CS.getInstruction() because it could be in the
1536 // live set of some other safepoint, in which case that safepoint's
1537 // PartiallyConstructedSafepointRecord will hold a raw pointer to this
1538 // llvm::Instruction. Instead, we defer the replacement and deletion to
1539 // after the live sets have been made explicit in the IR, and we no longer
1540 // have raw pointers to worry about.
1541 Replacements.emplace_back(CS.getInstruction(), GCResult);
1543 Replacements.emplace_back(CS.getInstruction(), nullptr);
1546 assert(!CS.getInstruction()->hasNUsesOrMore(2) &&
1547 "only valid use before rewrite is gc.result");
1548 assert(!CS.getInstruction()->hasOneUse() ||
1549 isGCResult(cast<Instruction>(*CS.getInstruction()->user_begin())));
1551 // Take the name of the original statepoint token if there was one.
1552 Token->takeName(CS.getInstruction());
1554 // Update the gc.result of the original statepoint (if any) to use the newly
1555 // inserted statepoint. This is safe to do here since the token can't be
1556 // considered a live reference.
1557 CS.getInstruction()->replaceAllUsesWith(Token);
1558 CS.getInstruction()->eraseFromParent();
1561 Result.StatepointToken = Token;
1563 // Second, create a gc.relocate for every live variable
1564 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1565 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, Token, Builder);
1569 struct NameOrdering {
1573 bool operator()(NameOrdering const &a, NameOrdering const &b) {
1574 return -1 == a.Derived->getName().compare(b.Derived->getName());
1579 static void StabilizeOrder(SmallVectorImpl<Value *> &BaseVec,
1580 SmallVectorImpl<Value *> &LiveVec) {
1581 assert(BaseVec.size() == LiveVec.size());
1583 SmallVector<NameOrdering, 64> Temp;
1584 for (size_t i = 0; i < BaseVec.size(); i++) {
1586 v.Base = BaseVec[i];
1587 v.Derived = LiveVec[i];
1591 std::sort(Temp.begin(), Temp.end(), NameOrdering());
1592 for (size_t i = 0; i < BaseVec.size(); i++) {
1593 BaseVec[i] = Temp[i].Base;
1594 LiveVec[i] = Temp[i].Derived;
1598 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1599 // which make the relocations happening at this safepoint explicit.
1601 // WARNING: Does not do any fixup to adjust users of the original live
1602 // values. That's the callers responsibility.
1604 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS,
1605 PartiallyConstructedSafepointRecord &Result,
1606 std::vector<DeferredReplacement> &Replacements) {
1607 const auto &LiveSet = Result.LiveSet;
1608 const auto &PointerToBase = Result.PointerToBase;
1610 // Convert to vector for efficient cross referencing.
1611 SmallVector<Value *, 64> BaseVec, LiveVec;
1612 LiveVec.reserve(LiveSet.size());
1613 BaseVec.reserve(LiveSet.size());
1614 for (Value *L : LiveSet) {
1615 LiveVec.push_back(L);
1616 assert(PointerToBase.count(L));
1617 Value *Base = PointerToBase.find(L)->second;
1618 BaseVec.push_back(Base);
1620 assert(LiveVec.size() == BaseVec.size());
1622 // To make the output IR slightly more stable (for use in diffs), ensure a
1623 // fixed order of the values in the safepoint (by sorting the value name).
1624 // The order is otherwise meaningless.
1625 StabilizeOrder(BaseVec, LiveVec);
1627 // Do the actual rewriting and delete the old statepoint
1628 makeStatepointExplicitImpl(CS, BaseVec, LiveVec, Result, Replacements);
1631 // Helper function for the relocationViaAlloca.
1633 // It receives iterator to the statepoint gc relocates and emits a store to the
1634 // assigned location (via allocaMap) for the each one of them. It adds the
1635 // visited values into the visitedLiveValues set, which we will later use them
1636 // for sanity checking.
1638 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1639 DenseMap<Value *, Value *> &AllocaMap,
1640 DenseSet<Value *> &VisitedLiveValues) {
1642 for (User *U : GCRelocs) {
1643 if (!isa<IntrinsicInst>(U))
1646 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1648 // We only care about relocates
1649 if (RelocatedValue->getIntrinsicID() !=
1650 Intrinsic::experimental_gc_relocate) {
1654 GCRelocateOperands RelocateOperands(RelocatedValue);
1655 Value *OriginalValue =
1656 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1657 assert(AllocaMap.count(OriginalValue));
1658 Value *Alloca = AllocaMap[OriginalValue];
1660 // Emit store into the related alloca
1661 // All gc_relocates are i8 addrspace(1)* typed, and it must be bitcasted to
1662 // the correct type according to alloca.
1663 assert(RelocatedValue->getNextNode() &&
1664 "Should always have one since it's not a terminator");
1665 IRBuilder<> Builder(RelocatedValue->getNextNode());
1666 Value *CastedRelocatedValue =
1667 Builder.CreateBitCast(RelocatedValue,
1668 cast<AllocaInst>(Alloca)->getAllocatedType(),
1669 suffixed_name_or(RelocatedValue, ".casted", ""));
1671 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1672 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1675 VisitedLiveValues.insert(OriginalValue);
1680 // Helper function for the "relocationViaAlloca". Similar to the
1681 // "insertRelocationStores" but works for rematerialized values.
1683 insertRematerializationStores(
1684 RematerializedValueMapTy RematerializedValues,
1685 DenseMap<Value *, Value *> &AllocaMap,
1686 DenseSet<Value *> &VisitedLiveValues) {
1688 for (auto RematerializedValuePair: RematerializedValues) {
1689 Instruction *RematerializedValue = RematerializedValuePair.first;
1690 Value *OriginalValue = RematerializedValuePair.second;
1692 assert(AllocaMap.count(OriginalValue) &&
1693 "Can not find alloca for rematerialized value");
1694 Value *Alloca = AllocaMap[OriginalValue];
1696 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1697 Store->insertAfter(RematerializedValue);
1700 VisitedLiveValues.insert(OriginalValue);
1705 /// Do all the relocation update via allocas and mem2reg
1706 static void relocationViaAlloca(
1707 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1708 ArrayRef<PartiallyConstructedSafepointRecord> Records) {
1710 // record initial number of (static) allocas; we'll check we have the same
1711 // number when we get done.
1712 int InitialAllocaNum = 0;
1713 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1715 if (isa<AllocaInst>(*I))
1719 // TODO-PERF: change data structures, reserve
1720 DenseMap<Value *, Value *> AllocaMap;
1721 SmallVector<AllocaInst *, 200> PromotableAllocas;
1722 // Used later to chack that we have enough allocas to store all values
1723 std::size_t NumRematerializedValues = 0;
1724 PromotableAllocas.reserve(Live.size());
1726 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1727 // "PromotableAllocas"
1728 auto emitAllocaFor = [&](Value *LiveValue) {
1729 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1730 F.getEntryBlock().getFirstNonPHI());
1731 AllocaMap[LiveValue] = Alloca;
1732 PromotableAllocas.push_back(Alloca);
1735 // Emit alloca for each live gc pointer
1736 for (Value *V : Live)
1739 // Emit allocas for rematerialized values
1740 for (const auto &Info : Records)
1741 for (auto RematerializedValuePair : Info.RematerializedValues) {
1742 Value *OriginalValue = RematerializedValuePair.second;
1743 if (AllocaMap.count(OriginalValue) != 0)
1746 emitAllocaFor(OriginalValue);
1747 ++NumRematerializedValues;
1750 // The next two loops are part of the same conceptual operation. We need to
1751 // insert a store to the alloca after the original def and at each
1752 // redefinition. We need to insert a load before each use. These are split
1753 // into distinct loops for performance reasons.
1755 // Update gc pointer after each statepoint: either store a relocated value or
1756 // null (if no relocated value was found for this gc pointer and it is not a
1757 // gc_result). This must happen before we update the statepoint with load of
1758 // alloca otherwise we lose the link between statepoint and old def.
1759 for (const auto &Info : Records) {
1760 Value *Statepoint = Info.StatepointToken;
1762 // This will be used for consistency check
1763 DenseSet<Value *> VisitedLiveValues;
1765 // Insert stores for normal statepoint gc relocates
1766 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1768 // In case if it was invoke statepoint
1769 // we will insert stores for exceptional path gc relocates.
1770 if (isa<InvokeInst>(Statepoint)) {
1771 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1775 // Do similar thing with rematerialized values
1776 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1779 if (ClobberNonLive) {
1780 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1781 // the gc.statepoint. This will turn some subtle GC problems into
1782 // slightly easier to debug SEGVs. Note that on large IR files with
1783 // lots of gc.statepoints this is extremely costly both memory and time
1785 SmallVector<AllocaInst *, 64> ToClobber;
1786 for (auto Pair : AllocaMap) {
1787 Value *Def = Pair.first;
1788 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1790 // This value was relocated
1791 if (VisitedLiveValues.count(Def)) {
1794 ToClobber.push_back(Alloca);
1797 auto InsertClobbersAt = [&](Instruction *IP) {
1798 for (auto *AI : ToClobber) {
1799 auto AIType = cast<PointerType>(AI->getType());
1800 auto PT = cast<PointerType>(AIType->getElementType());
1801 Constant *CPN = ConstantPointerNull::get(PT);
1802 StoreInst *Store = new StoreInst(CPN, AI);
1803 Store->insertBefore(IP);
1807 // Insert the clobbering stores. These may get intermixed with the
1808 // gc.results and gc.relocates, but that's fine.
1809 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1810 InsertClobbersAt(&*II->getNormalDest()->getFirstInsertionPt());
1811 InsertClobbersAt(&*II->getUnwindDest()->getFirstInsertionPt());
1813 InsertClobbersAt(cast<Instruction>(Statepoint)->getNextNode());
1818 // Update use with load allocas and add store for gc_relocated.
1819 for (auto Pair : AllocaMap) {
1820 Value *Def = Pair.first;
1821 Value *Alloca = Pair.second;
1823 // We pre-record the uses of allocas so that we dont have to worry about
1824 // later update that changes the user information..
1826 SmallVector<Instruction *, 20> Uses;
1827 // PERF: trade a linear scan for repeated reallocation
1828 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1829 for (User *U : Def->users()) {
1830 if (!isa<ConstantExpr>(U)) {
1831 // If the def has a ConstantExpr use, then the def is either a
1832 // ConstantExpr use itself or null. In either case
1833 // (recursively in the first, directly in the second), the oop
1834 // it is ultimately dependent on is null and this particular
1835 // use does not need to be fixed up.
1836 Uses.push_back(cast<Instruction>(U));
1840 std::sort(Uses.begin(), Uses.end());
1841 auto Last = std::unique(Uses.begin(), Uses.end());
1842 Uses.erase(Last, Uses.end());
1844 for (Instruction *Use : Uses) {
1845 if (isa<PHINode>(Use)) {
1846 PHINode *Phi = cast<PHINode>(Use);
1847 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1848 if (Def == Phi->getIncomingValue(i)) {
1849 LoadInst *Load = new LoadInst(
1850 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1851 Phi->setIncomingValue(i, Load);
1855 LoadInst *Load = new LoadInst(Alloca, "", Use);
1856 Use->replaceUsesOfWith(Def, Load);
1860 // Emit store for the initial gc value. Store must be inserted after load,
1861 // otherwise store will be in alloca's use list and an extra load will be
1862 // inserted before it.
1863 StoreInst *Store = new StoreInst(Def, Alloca);
1864 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1865 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1866 // InvokeInst is a TerminatorInst so the store need to be inserted
1867 // into its normal destination block.
1868 BasicBlock *NormalDest = Invoke->getNormalDest();
1869 Store->insertBefore(NormalDest->getFirstNonPHI());
1871 assert(!Inst->isTerminator() &&
1872 "The only TerminatorInst that can produce a value is "
1873 "InvokeInst which is handled above.");
1874 Store->insertAfter(Inst);
1877 assert(isa<Argument>(Def));
1878 Store->insertAfter(cast<Instruction>(Alloca));
1882 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1883 "we must have the same allocas with lives");
1884 if (!PromotableAllocas.empty()) {
1885 // Apply mem2reg to promote alloca to SSA
1886 PromoteMemToReg(PromotableAllocas, DT);
1890 for (auto &I : F.getEntryBlock())
1891 if (isa<AllocaInst>(I))
1893 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1897 /// Implement a unique function which doesn't require we sort the input
1898 /// vector. Doing so has the effect of changing the output of a couple of
1899 /// tests in ways which make them less useful in testing fused safepoints.
1900 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1901 SmallSet<T, 8> Seen;
1902 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1903 return !Seen.insert(V).second;
1907 /// Insert holders so that each Value is obviously live through the entire
1908 /// lifetime of the call.
1909 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1910 SmallVectorImpl<CallInst *> &Holders) {
1912 // No values to hold live, might as well not insert the empty holder
1915 Module *M = CS.getInstruction()->getModule();
1916 // Use a dummy vararg function to actually hold the values live
1917 Function *Func = cast<Function>(M->getOrInsertFunction(
1918 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1920 // For call safepoints insert dummy calls right after safepoint
1921 Holders.push_back(CallInst::Create(Func, Values, "",
1922 &*++CS.getInstruction()->getIterator()));
1925 // For invoke safepooints insert dummy calls both in normal and
1926 // exceptional destination blocks
1927 auto *II = cast<InvokeInst>(CS.getInstruction());
1928 Holders.push_back(CallInst::Create(
1929 Func, Values, "", &*II->getNormalDest()->getFirstInsertionPt()));
1930 Holders.push_back(CallInst::Create(
1931 Func, Values, "", &*II->getUnwindDest()->getFirstInsertionPt()));
1934 static void findLiveReferences(
1935 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1936 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1937 GCPtrLivenessData OriginalLivenessData;
1938 computeLiveInValues(DT, F, OriginalLivenessData);
1939 for (size_t i = 0; i < records.size(); i++) {
1940 struct PartiallyConstructedSafepointRecord &info = records[i];
1941 const CallSite &CS = toUpdate[i];
1942 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1946 /// Remove any vector of pointers from the live set by scalarizing them over the
1947 /// statepoint instruction. Adds the scalarized pieces to the live set. It
1948 /// would be preferable to include the vector in the statepoint itself, but
1949 /// the lowering code currently does not handle that. Extending it would be
1950 /// slightly non-trivial since it requires a format change. Given how rare
1951 /// such cases are (for the moment?) scalarizing is an acceptable compromise.
1952 static void splitVectorValues(Instruction *StatepointInst,
1953 StatepointLiveSetTy &LiveSet,
1954 DenseMap<Value *, Value *>& PointerToBase,
1955 DominatorTree &DT) {
1956 SmallVector<Value *, 16> ToSplit;
1957 for (Value *V : LiveSet)
1958 if (isa<VectorType>(V->getType()))
1959 ToSplit.push_back(V);
1961 if (ToSplit.empty())
1964 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1966 Function &F = *(StatepointInst->getParent()->getParent());
1968 DenseMap<Value *, AllocaInst *> AllocaMap;
1969 // First is normal return, second is exceptional return (invoke only)
1970 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1971 for (Value *V : ToSplit) {
1972 AllocaInst *Alloca =
1973 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1974 AllocaMap[V] = Alloca;
1976 VectorType *VT = cast<VectorType>(V->getType());
1977 IRBuilder<> Builder(StatepointInst);
1978 SmallVector<Value *, 16> Elements;
1979 for (unsigned i = 0; i < VT->getNumElements(); i++)
1980 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1981 ElementMapping[V] = Elements;
1983 auto InsertVectorReform = [&](Instruction *IP) {
1984 Builder.SetInsertPoint(IP);
1985 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1986 Value *ResultVec = UndefValue::get(VT);
1987 for (unsigned i = 0; i < VT->getNumElements(); i++)
1988 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1989 Builder.getInt32(i));
1993 if (isa<CallInst>(StatepointInst)) {
1994 BasicBlock::iterator Next(StatepointInst);
1996 Instruction *IP = &*(Next);
1997 Replacements[V].first = InsertVectorReform(IP);
1998 Replacements[V].second = nullptr;
2000 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
2001 // We've already normalized - check that we don't have shared destination
2003 BasicBlock *NormalDest = Invoke->getNormalDest();
2004 assert(!isa<PHINode>(NormalDest->begin()));
2005 BasicBlock *UnwindDest = Invoke->getUnwindDest();
2006 assert(!isa<PHINode>(UnwindDest->begin()));
2007 // Insert insert element sequences in both successors
2008 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
2009 Replacements[V].first = InsertVectorReform(IP);
2010 IP = &*(UnwindDest->getFirstInsertionPt());
2011 Replacements[V].second = InsertVectorReform(IP);
2015 for (Value *V : ToSplit) {
2016 AllocaInst *Alloca = AllocaMap[V];
2018 // Capture all users before we start mutating use lists
2019 SmallVector<Instruction *, 16> Users;
2020 for (User *U : V->users())
2021 Users.push_back(cast<Instruction>(U));
2023 for (Instruction *I : Users) {
2024 if (auto Phi = dyn_cast<PHINode>(I)) {
2025 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
2026 if (V == Phi->getIncomingValue(i)) {
2027 LoadInst *Load = new LoadInst(
2028 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
2029 Phi->setIncomingValue(i, Load);
2032 LoadInst *Load = new LoadInst(Alloca, "", I);
2033 I->replaceUsesOfWith(V, Load);
2037 // Store the original value and the replacement value into the alloca
2038 StoreInst *Store = new StoreInst(V, Alloca);
2039 if (auto I = dyn_cast<Instruction>(V))
2040 Store->insertAfter(I);
2042 Store->insertAfter(Alloca);
2044 // Normal return for invoke, or call return
2045 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
2046 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
2047 // Unwind return for invoke only
2048 Replacement = cast_or_null<Instruction>(Replacements[V].second);
2050 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
2053 // apply mem2reg to promote alloca to SSA
2054 SmallVector<AllocaInst *, 16> Allocas;
2055 for (Value *V : ToSplit)
2056 Allocas.push_back(AllocaMap[V]);
2057 PromoteMemToReg(Allocas, DT);
2059 // Update our tracking of live pointers and base mappings to account for the
2060 // changes we just made.
2061 for (Value *V : ToSplit) {
2062 auto &Elements = ElementMapping[V];
2065 LiveSet.insert(Elements.begin(), Elements.end());
2066 // We need to update the base mapping as well.
2067 assert(PointerToBase.count(V));
2068 Value *OldBase = PointerToBase[V];
2069 auto &BaseElements = ElementMapping[OldBase];
2070 PointerToBase.erase(V);
2071 assert(Elements.size() == BaseElements.size());
2072 for (unsigned i = 0; i < Elements.size(); i++) {
2073 Value *Elem = Elements[i];
2074 PointerToBase[Elem] = BaseElements[i];
2079 // Helper function for the "rematerializeLiveValues". It walks use chain
2080 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
2081 // values are visited (currently it is GEP's and casts). Returns true if it
2082 // successfully reached "BaseValue" and false otherwise.
2083 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
2085 static bool findRematerializableChainToBasePointer(
2086 SmallVectorImpl<Instruction*> &ChainToBase,
2087 Value *CurrentValue, Value *BaseValue) {
2089 // We have found a base value
2090 if (CurrentValue == BaseValue) {
2094 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
2095 ChainToBase.push_back(GEP);
2096 return findRematerializableChainToBasePointer(ChainToBase,
2097 GEP->getPointerOperand(),
2101 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
2102 Value *Def = CI->stripPointerCasts();
2104 // This two checks are basically similar. First one is here for the
2105 // consistency with findBasePointers logic.
2106 assert(!isa<CastInst>(Def) && "not a pointer cast found");
2107 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
2110 ChainToBase.push_back(CI);
2111 return findRematerializableChainToBasePointer(ChainToBase, Def, BaseValue);
2114 // Not supported instruction in the chain
2118 // Helper function for the "rematerializeLiveValues". Compute cost of the use
2119 // chain we are going to rematerialize.
2121 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
2122 TargetTransformInfo &TTI) {
2125 for (Instruction *Instr : Chain) {
2126 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
2127 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
2128 "non noop cast is found during rematerialization");
2130 Type *SrcTy = CI->getOperand(0)->getType();
2131 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
2133 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
2134 // Cost of the address calculation
2135 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
2136 Cost += TTI.getAddressComputationCost(ValTy);
2138 // And cost of the GEP itself
2139 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
2140 // allowed for the external usage)
2141 if (!GEP->hasAllConstantIndices())
2145 llvm_unreachable("unsupported instruciton type during rematerialization");
2152 // From the statepoint live set pick values that are cheaper to recompute then
2153 // to relocate. Remove this values from the live set, rematerialize them after
2154 // statepoint and record them in "Info" structure. Note that similar to
2155 // relocated values we don't do any user adjustments here.
2156 static void rematerializeLiveValues(CallSite CS,
2157 PartiallyConstructedSafepointRecord &Info,
2158 TargetTransformInfo &TTI) {
2159 const unsigned int ChainLengthThreshold = 10;
2161 // Record values we are going to delete from this statepoint live set.
2162 // We can not di this in following loop due to iterator invalidation.
2163 SmallVector<Value *, 32> LiveValuesToBeDeleted;
2165 for (Value *LiveValue: Info.LiveSet) {
2166 // For each live pointer find it's defining chain
2167 SmallVector<Instruction *, 3> ChainToBase;
2168 assert(Info.PointerToBase.count(LiveValue));
2170 findRematerializableChainToBasePointer(ChainToBase,
2172 Info.PointerToBase[LiveValue]);
2173 // Nothing to do, or chain is too long
2175 ChainToBase.size() == 0 ||
2176 ChainToBase.size() > ChainLengthThreshold)
2179 // Compute cost of this chain
2180 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
2181 // TODO: We can also account for cases when we will be able to remove some
2182 // of the rematerialized values by later optimization passes. I.e if
2183 // we rematerialized several intersecting chains. Or if original values
2184 // don't have any uses besides this statepoint.
2186 // For invokes we need to rematerialize each chain twice - for normal and
2187 // for unwind basic blocks. Model this by multiplying cost by two.
2188 if (CS.isInvoke()) {
2191 // If it's too expensive - skip it
2192 if (Cost >= RematerializationThreshold)
2195 // Remove value from the live set
2196 LiveValuesToBeDeleted.push_back(LiveValue);
2198 // Clone instructions and record them inside "Info" structure
2200 // Walk backwards to visit top-most instructions first
2201 std::reverse(ChainToBase.begin(), ChainToBase.end());
2203 // Utility function which clones all instructions from "ChainToBase"
2204 // and inserts them before "InsertBefore". Returns rematerialized value
2205 // which should be used after statepoint.
2206 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
2207 Instruction *LastClonedValue = nullptr;
2208 Instruction *LastValue = nullptr;
2209 for (Instruction *Instr: ChainToBase) {
2210 // Only GEP's and casts are suported as we need to be careful to not
2211 // introduce any new uses of pointers not in the liveset.
2212 // Note that it's fine to introduce new uses of pointers which were
2213 // otherwise not used after this statepoint.
2214 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2216 Instruction *ClonedValue = Instr->clone();
2217 ClonedValue->insertBefore(InsertBefore);
2218 ClonedValue->setName(Instr->getName() + ".remat");
2220 // If it is not first instruction in the chain then it uses previously
2221 // cloned value. We should update it to use cloned value.
2222 if (LastClonedValue) {
2224 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2226 // Assert that cloned instruction does not use any instructions from
2227 // this chain other than LastClonedValue
2228 for (auto OpValue : ClonedValue->operand_values()) {
2229 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2230 ChainToBase.end() &&
2231 "incorrect use in rematerialization chain");
2236 LastClonedValue = ClonedValue;
2239 assert(LastClonedValue);
2240 return LastClonedValue;
2243 // Different cases for calls and invokes. For invokes we need to clone
2244 // instructions both on normal and unwind path.
2246 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2247 assert(InsertBefore);
2248 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2249 Info.RematerializedValues[RematerializedValue] = LiveValue;
2251 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2253 Instruction *NormalInsertBefore =
2254 &*Invoke->getNormalDest()->getFirstInsertionPt();
2255 Instruction *UnwindInsertBefore =
2256 &*Invoke->getUnwindDest()->getFirstInsertionPt();
2258 Instruction *NormalRematerializedValue =
2259 rematerializeChain(NormalInsertBefore);
2260 Instruction *UnwindRematerializedValue =
2261 rematerializeChain(UnwindInsertBefore);
2263 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2264 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2268 // Remove rematerializaed values from the live set
2269 for (auto LiveValue: LiveValuesToBeDeleted) {
2270 Info.LiveSet.erase(LiveValue);
2274 static bool insertParsePoints(Function &F, DominatorTree &DT,
2275 TargetTransformInfo &TTI,
2276 SmallVectorImpl<CallSite> &ToUpdate) {
2278 // sanity check the input
2279 std::set<CallSite> Uniqued;
2280 Uniqued.insert(ToUpdate.begin(), ToUpdate.end());
2281 assert(Uniqued.size() == ToUpdate.size() && "no duplicates please!");
2283 for (CallSite CS : ToUpdate) {
2284 assert(CS.getInstruction()->getParent()->getParent() == &F);
2285 assert((UseDeoptBundles || isStatepoint(CS)) &&
2286 "expected to already be a deopt statepoint");
2290 // When inserting gc.relocates for invokes, we need to be able to insert at
2291 // the top of the successor blocks. See the comment on
2292 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2293 // may restructure the CFG.
2294 for (CallSite CS : ToUpdate) {
2297 auto *II = cast<InvokeInst>(CS.getInstruction());
2298 normalizeForInvokeSafepoint(II->getNormalDest(), II->getParent(), DT);
2299 normalizeForInvokeSafepoint(II->getUnwindDest(), II->getParent(), DT);
2302 // A list of dummy calls added to the IR to keep various values obviously
2303 // live in the IR. We'll remove all of these when done.
2304 SmallVector<CallInst *, 64> Holders;
2306 // Insert a dummy call with all of the arguments to the vm_state we'll need
2307 // for the actual safepoint insertion. This ensures reference arguments in
2308 // the deopt argument list are considered live through the safepoint (and
2309 // thus makes sure they get relocated.)
2310 for (CallSite CS : ToUpdate) {
2311 SmallVector<Value *, 64> DeoptValues;
2313 iterator_range<const Use *> DeoptStateRange =
2315 ? iterator_range<const Use *>(GetDeoptBundleOperands(CS))
2316 : iterator_range<const Use *>(Statepoint(CS).vm_state_args());
2318 for (Value *Arg : DeoptStateRange) {
2319 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2320 "support for FCA unimplemented");
2321 if (isHandledGCPointerType(Arg->getType()))
2322 DeoptValues.push_back(Arg);
2325 insertUseHolderAfter(CS, DeoptValues, Holders);
2328 SmallVector<PartiallyConstructedSafepointRecord, 64> Records(ToUpdate.size());
2330 // A) Identify all gc pointers which are statically live at the given call
2332 findLiveReferences(F, DT, ToUpdate, Records);
2334 // B) Find the base pointers for each live pointer
2335 /* scope for caching */ {
2336 // Cache the 'defining value' relation used in the computation and
2337 // insertion of base phis and selects. This ensures that we don't insert
2338 // large numbers of duplicate base_phis.
2339 DefiningValueMapTy DVCache;
2341 for (size_t i = 0; i < Records.size(); i++) {
2342 PartiallyConstructedSafepointRecord &info = Records[i];
2343 findBasePointers(DT, DVCache, ToUpdate[i], info);
2345 } // end of cache scope
2347 // The base phi insertion logic (for any safepoint) may have inserted new
2348 // instructions which are now live at some safepoint. The simplest such
2351 // phi a <-- will be a new base_phi here
2352 // safepoint 1 <-- that needs to be live here
2356 // We insert some dummy calls after each safepoint to definitely hold live
2357 // the base pointers which were identified for that safepoint. We'll then
2358 // ask liveness for _every_ base inserted to see what is now live. Then we
2359 // remove the dummy calls.
2360 Holders.reserve(Holders.size() + Records.size());
2361 for (size_t i = 0; i < Records.size(); i++) {
2362 PartiallyConstructedSafepointRecord &Info = Records[i];
2364 SmallVector<Value *, 128> Bases;
2365 for (auto Pair : Info.PointerToBase)
2366 Bases.push_back(Pair.second);
2368 insertUseHolderAfter(ToUpdate[i], Bases, Holders);
2371 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2372 // need to rerun liveness. We may *also* have inserted new defs, but that's
2373 // not the key issue.
2374 recomputeLiveInValues(F, DT, ToUpdate, Records);
2376 if (PrintBasePointers) {
2377 for (auto &Info : Records) {
2378 errs() << "Base Pairs: (w/Relocation)\n";
2379 for (auto Pair : Info.PointerToBase)
2380 errs() << " derived %" << Pair.first->getName() << " base %"
2381 << Pair.second->getName() << "\n";
2385 for (CallInst *CI : Holders)
2386 CI->eraseFromParent();
2390 // Do a limited scalarization of any live at safepoint vector values which
2391 // contain pointers. This enables this pass to run after vectorization at
2392 // the cost of some possible performance loss. TODO: it would be nice to
2393 // natively support vectors all the way through the backend so we don't need
2394 // to scalarize here.
2395 for (size_t i = 0; i < Records.size(); i++) {
2396 PartiallyConstructedSafepointRecord &Info = Records[i];
2397 Instruction *Statepoint = ToUpdate[i].getInstruction();
2398 splitVectorValues(cast<Instruction>(Statepoint), Info.LiveSet,
2399 Info.PointerToBase, DT);
2402 // In order to reduce live set of statepoint we might choose to rematerialize
2403 // some values instead of relocating them. This is purely an optimization and
2404 // does not influence correctness.
2405 for (size_t i = 0; i < Records.size(); i++)
2406 rematerializeLiveValues(ToUpdate[i], Records[i], TTI);
2408 // We need this to safely RAUW and delete call or invoke return values that
2409 // may themselves be live over a statepoint. For details, please see usage in
2410 // makeStatepointExplicitImpl.
2411 std::vector<DeferredReplacement> Replacements;
2413 // Now run through and replace the existing statepoints with new ones with
2414 // the live variables listed. We do not yet update uses of the values being
2415 // relocated. We have references to live variables that need to
2416 // survive to the last iteration of this loop. (By construction, the
2417 // previous statepoint can not be a live variable, thus we can and remove
2418 // the old statepoint calls as we go.)
2419 for (size_t i = 0; i < Records.size(); i++)
2420 makeStatepointExplicit(DT, ToUpdate[i], Records[i], Replacements);
2422 ToUpdate.clear(); // prevent accident use of invalid CallSites
2424 for (auto &PR : Replacements)
2427 Replacements.clear();
2429 for (auto &Info : Records) {
2430 // These live sets may contain state Value pointers, since we replaced calls
2431 // with operand bundles with calls wrapped in gc.statepoint, and some of
2432 // those calls may have been def'ing live gc pointers. Clear these out to
2433 // avoid accidentally using them.
2435 // TODO: We should create a separate data structure that does not contain
2436 // these live sets, and migrate to using that data structure from this point
2438 Info.LiveSet.clear();
2439 Info.PointerToBase.clear();
2442 // Do all the fixups of the original live variables to their relocated selves
2443 SmallVector<Value *, 128> Live;
2444 for (size_t i = 0; i < Records.size(); i++) {
2445 PartiallyConstructedSafepointRecord &Info = Records[i];
2447 // We can't simply save the live set from the original insertion. One of
2448 // the live values might be the result of a call which needs a safepoint.
2449 // That Value* no longer exists and we need to use the new gc_result.
2450 // Thankfully, the live set is embedded in the statepoint (and updated), so
2451 // we just grab that.
2452 Statepoint Statepoint(Info.StatepointToken);
2453 Live.insert(Live.end(), Statepoint.gc_args_begin(),
2454 Statepoint.gc_args_end());
2456 // Do some basic sanity checks on our liveness results before performing
2457 // relocation. Relocation can and will turn mistakes in liveness results
2458 // into non-sensical code which is must harder to debug.
2459 // TODO: It would be nice to test consistency as well
2460 assert(DT.isReachableFromEntry(Info.StatepointToken->getParent()) &&
2461 "statepoint must be reachable or liveness is meaningless");
2462 for (Value *V : Statepoint.gc_args()) {
2463 if (!isa<Instruction>(V))
2464 // Non-instruction values trivial dominate all possible uses
2466 auto *LiveInst = cast<Instruction>(V);
2467 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2468 "unreachable values should never be live");
2469 assert(DT.dominates(LiveInst, Info.StatepointToken) &&
2470 "basic SSA liveness expectation violated by liveness analysis");
2474 unique_unsorted(Live);
2478 for (auto *Ptr : Live)
2479 assert(isGCPointerType(Ptr->getType()) && "must be a gc pointer type");
2482 relocationViaAlloca(F, DT, Live, Records);
2483 return !Records.empty();
2486 // Handles both return values and arguments for Functions and CallSites.
2487 template <typename AttrHolder>
2488 static void RemoveNonValidAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2491 if (AH.getDereferenceableBytes(Index))
2492 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2493 AH.getDereferenceableBytes(Index)));
2494 if (AH.getDereferenceableOrNullBytes(Index))
2495 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2496 AH.getDereferenceableOrNullBytes(Index)));
2497 if (AH.doesNotAlias(Index))
2498 R.addAttribute(Attribute::NoAlias);
2501 AH.setAttributes(AH.getAttributes().removeAttributes(
2502 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2506 RewriteStatepointsForGC::stripNonValidAttributesFromPrototype(Function &F) {
2507 LLVMContext &Ctx = F.getContext();
2509 for (Argument &A : F.args())
2510 if (isa<PointerType>(A.getType()))
2511 RemoveNonValidAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2513 if (isa<PointerType>(F.getReturnType()))
2514 RemoveNonValidAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2517 void RewriteStatepointsForGC::stripNonValidAttributesFromBody(Function &F) {
2521 LLVMContext &Ctx = F.getContext();
2522 MDBuilder Builder(Ctx);
2524 for (Instruction &I : instructions(F)) {
2525 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2526 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2527 bool IsImmutableTBAA =
2528 MD->getNumOperands() == 4 &&
2529 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2531 if (!IsImmutableTBAA)
2532 continue; // no work to do, MD_tbaa is already marked mutable
2534 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2535 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2537 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2539 MDNode *MutableTBAA =
2540 Builder.createTBAAStructTagNode(Base, Access, Offset);
2541 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2544 if (CallSite CS = CallSite(&I)) {
2545 for (int i = 0, e = CS.arg_size(); i != e; i++)
2546 if (isa<PointerType>(CS.getArgument(i)->getType()))
2547 RemoveNonValidAttrAtIndex(Ctx, CS, i + 1);
2548 if (isa<PointerType>(CS.getType()))
2549 RemoveNonValidAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2554 /// Returns true if this function should be rewritten by this pass. The main
2555 /// point of this function is as an extension point for custom logic.
2556 static bool shouldRewriteStatepointsIn(Function &F) {
2557 // TODO: This should check the GCStrategy
2559 const char *FunctionGCName = F.getGC();
2560 const StringRef StatepointExampleName("statepoint-example");
2561 const StringRef CoreCLRName("coreclr");
2562 return (StatepointExampleName == FunctionGCName) ||
2563 (CoreCLRName == FunctionGCName);
2568 void RewriteStatepointsForGC::stripNonValidAttributes(Module &M) {
2570 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2574 for (Function &F : M)
2575 stripNonValidAttributesFromPrototype(F);
2577 for (Function &F : M)
2578 stripNonValidAttributesFromBody(F);
2581 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2582 // Nothing to do for declarations.
2583 if (F.isDeclaration() || F.empty())
2586 // Policy choice says not to rewrite - the most common reason is that we're
2587 // compiling code without a GCStrategy.
2588 if (!shouldRewriteStatepointsIn(F))
2591 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2592 TargetTransformInfo &TTI =
2593 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2595 auto NeedsRewrite = [](Instruction &I) {
2596 if (UseDeoptBundles) {
2597 if (ImmutableCallSite CS = ImmutableCallSite(&I))
2598 return !callsGCLeafFunction(CS);
2602 return isStatepoint(I);
2605 // Gather all the statepoints which need rewritten. Be careful to only
2606 // consider those in reachable code since we need to ask dominance queries
2607 // when rewriting. We'll delete the unreachable ones in a moment.
2608 SmallVector<CallSite, 64> ParsePointNeeded;
2609 bool HasUnreachableStatepoint = false;
2610 for (Instruction &I : instructions(F)) {
2611 // TODO: only the ones with the flag set!
2612 if (NeedsRewrite(I)) {
2613 if (DT.isReachableFromEntry(I.getParent()))
2614 ParsePointNeeded.push_back(CallSite(&I));
2616 HasUnreachableStatepoint = true;
2620 bool MadeChange = false;
2622 // Delete any unreachable statepoints so that we don't have unrewritten
2623 // statepoints surviving this pass. This makes testing easier and the
2624 // resulting IR less confusing to human readers. Rather than be fancy, we
2625 // just reuse a utility function which removes the unreachable blocks.
2626 if (HasUnreachableStatepoint)
2627 MadeChange |= removeUnreachableBlocks(F);
2629 // Return early if no work to do.
2630 if (ParsePointNeeded.empty())
2633 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2634 // These are created by LCSSA. They have the effect of increasing the size
2635 // of liveness sets for no good reason. It may be harder to do this post
2636 // insertion since relocations and base phis can confuse things.
2637 for (BasicBlock &BB : F)
2638 if (BB.getUniquePredecessor()) {
2640 FoldSingleEntryPHINodes(&BB);
2643 // Before we start introducing relocations, we want to tweak the IR a bit to
2644 // avoid unfortunate code generation effects. The main example is that we
2645 // want to try to make sure the comparison feeding a branch is after any
2646 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2647 // values feeding a branch after relocation. This is semantically correct,
2648 // but results in extra register pressure since both the pre-relocation and
2649 // post-relocation copies must be available in registers. For code without
2650 // relocations this is handled elsewhere, but teaching the scheduler to
2651 // reverse the transform we're about to do would be slightly complex.
2652 // Note: This may extend the live range of the inputs to the icmp and thus
2653 // increase the liveset of any statepoint we move over. This is profitable
2654 // as long as all statepoints are in rare blocks. If we had in-register
2655 // lowering for live values this would be a much safer transform.
2656 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2657 if (auto *BI = dyn_cast<BranchInst>(TI))
2658 if (BI->isConditional())
2659 return dyn_cast<Instruction>(BI->getCondition());
2660 // TODO: Extend this to handle switches
2663 for (BasicBlock &BB : F) {
2664 TerminatorInst *TI = BB.getTerminator();
2665 if (auto *Cond = getConditionInst(TI))
2666 // TODO: Handle more than just ICmps here. We should be able to move
2667 // most instructions without side effects or memory access.
2668 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2670 Cond->moveBefore(TI);
2674 MadeChange |= insertParsePoints(F, DT, TTI, ParsePointNeeded);
2678 // liveness computation via standard dataflow
2679 // -------------------------------------------------------------------
2681 // TODO: Consider using bitvectors for liveness, the set of potentially
2682 // interesting values should be small and easy to pre-compute.
2684 /// Compute the live-in set for the location rbegin starting from
2685 /// the live-out set of the basic block
2686 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2687 BasicBlock::reverse_iterator rend,
2688 DenseSet<Value *> &LiveTmp) {
2690 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2691 Instruction *I = &*ritr;
2693 // KILL/Def - Remove this definition from LiveIn
2696 // Don't consider *uses* in PHI nodes, we handle their contribution to
2697 // predecessor blocks when we seed the LiveOut sets
2698 if (isa<PHINode>(I))
2701 // USE - Add to the LiveIn set for this instruction
2702 for (Value *V : I->operands()) {
2703 assert(!isUnhandledGCPointerType(V->getType()) &&
2704 "support for FCA unimplemented");
2705 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2706 // The choice to exclude all things constant here is slightly subtle.
2707 // There are two independent reasons:
2708 // - We assume that things which are constant (from LLVM's definition)
2709 // do not move at runtime. For example, the address of a global
2710 // variable is fixed, even though it's contents may not be.
2711 // - Second, we can't disallow arbitrary inttoptr constants even
2712 // if the language frontend does. Optimization passes are free to
2713 // locally exploit facts without respect to global reachability. This
2714 // can create sections of code which are dynamically unreachable and
2715 // contain just about anything. (see constants.ll in tests)
2722 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2724 for (BasicBlock *Succ : successors(BB)) {
2725 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2726 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2727 PHINode *Phi = cast<PHINode>(&*I);
2728 Value *V = Phi->getIncomingValueForBlock(BB);
2729 assert(!isUnhandledGCPointerType(V->getType()) &&
2730 "support for FCA unimplemented");
2731 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2738 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2739 DenseSet<Value *> KillSet;
2740 for (Instruction &I : *BB)
2741 if (isHandledGCPointerType(I.getType()))
2747 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2748 /// sanity check for the liveness computation.
2749 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2750 TerminatorInst *TI, bool TermOkay = false) {
2751 for (Value *V : Live) {
2752 if (auto *I = dyn_cast<Instruction>(V)) {
2753 // The terminator can be a member of the LiveOut set. LLVM's definition
2754 // of instruction dominance states that V does not dominate itself. As
2755 // such, we need to special case this to allow it.
2756 if (TermOkay && TI == I)
2758 assert(DT.dominates(I, TI) &&
2759 "basic SSA liveness expectation violated by liveness analysis");
2764 /// Check that all the liveness sets used during the computation of liveness
2765 /// obey basic SSA properties. This is useful for finding cases where we miss
2767 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2769 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2770 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2771 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2775 static void computeLiveInValues(DominatorTree &DT, Function &F,
2776 GCPtrLivenessData &Data) {
2778 SmallSetVector<BasicBlock *, 200> Worklist;
2779 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2780 // We use a SetVector so that we don't have duplicates in the worklist.
2781 Worklist.insert(pred_begin(BB), pred_end(BB));
2783 auto NextItem = [&]() {
2784 BasicBlock *BB = Worklist.back();
2785 Worklist.pop_back();
2789 // Seed the liveness for each individual block
2790 for (BasicBlock &BB : F) {
2791 Data.KillSet[&BB] = computeKillSet(&BB);
2792 Data.LiveSet[&BB].clear();
2793 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2796 for (Value *Kill : Data.KillSet[&BB])
2797 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2800 Data.LiveOut[&BB] = DenseSet<Value *>();
2801 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2802 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2803 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2804 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2805 if (!Data.LiveIn[&BB].empty())
2806 AddPredsToWorklist(&BB);
2809 // Propagate that liveness until stable
2810 while (!Worklist.empty()) {
2811 BasicBlock *BB = NextItem();
2813 // Compute our new liveout set, then exit early if it hasn't changed
2814 // despite the contribution of our successor.
2815 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2816 const auto OldLiveOutSize = LiveOut.size();
2817 for (BasicBlock *Succ : successors(BB)) {
2818 assert(Data.LiveIn.count(Succ));
2819 set_union(LiveOut, Data.LiveIn[Succ]);
2821 // assert OutLiveOut is a subset of LiveOut
2822 if (OldLiveOutSize == LiveOut.size()) {
2823 // If the sets are the same size, then we didn't actually add anything
2824 // when unioning our successors LiveIn Thus, the LiveIn of this block
2828 Data.LiveOut[BB] = LiveOut;
2830 // Apply the effects of this basic block
2831 DenseSet<Value *> LiveTmp = LiveOut;
2832 set_union(LiveTmp, Data.LiveSet[BB]);
2833 set_subtract(LiveTmp, Data.KillSet[BB]);
2835 assert(Data.LiveIn.count(BB));
2836 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2837 // assert: OldLiveIn is a subset of LiveTmp
2838 if (OldLiveIn.size() != LiveTmp.size()) {
2839 Data.LiveIn[BB] = LiveTmp;
2840 AddPredsToWorklist(BB);
2842 } // while( !worklist.empty() )
2845 // Sanity check our output against SSA properties. This helps catch any
2846 // missing kills during the above iteration.
2847 for (BasicBlock &BB : F) {
2848 checkBasicSSA(DT, Data, BB);
2853 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2854 StatepointLiveSetTy &Out) {
2856 BasicBlock *BB = Inst->getParent();
2858 // Note: The copy is intentional and required
2859 assert(Data.LiveOut.count(BB));
2860 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2862 // We want to handle the statepoint itself oddly. It's
2863 // call result is not live (normal), nor are it's arguments
2864 // (unless they're used again later). This adjustment is
2865 // specifically what we need to relocate
2866 BasicBlock::reverse_iterator rend(Inst->getIterator());
2867 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2868 LiveOut.erase(Inst);
2869 Out.insert(LiveOut.begin(), LiveOut.end());
2872 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2874 PartiallyConstructedSafepointRecord &Info) {
2875 Instruction *Inst = CS.getInstruction();
2876 StatepointLiveSetTy Updated;
2877 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2880 DenseSet<Value *> Bases;
2881 for (auto KVPair : Info.PointerToBase) {
2882 Bases.insert(KVPair.second);
2885 // We may have base pointers which are now live that weren't before. We need
2886 // to update the PointerToBase structure to reflect this.
2887 for (auto V : Updated)
2888 if (!Info.PointerToBase.count(V)) {
2889 assert(Bases.count(V) && "can't find base for unexpected live value");
2890 Info.PointerToBase[V] = V;
2895 for (auto V : Updated) {
2896 assert(Info.PointerToBase.count(V) &&
2897 "must be able to find base for live value");
2901 // Remove any stale base mappings - this can happen since our liveness is
2902 // more precise then the one inherent in the base pointer analysis
2903 DenseSet<Value *> ToErase;
2904 for (auto KVPair : Info.PointerToBase)
2905 if (!Updated.count(KVPair.first))
2906 ToErase.insert(KVPair.first);
2907 for (auto V : ToErase)
2908 Info.PointerToBase.erase(V);
2911 for (auto KVPair : Info.PointerToBase)
2912 assert(Updated.count(KVPair.first) && "record for non-live value");
2915 Info.LiveSet = Updated;