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/TargetTransformInfo.h"
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/ADT/DenseSet.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/StringRef.h"
23 #include "llvm/ADT/MapVector.h"
24 #include "llvm/IR/BasicBlock.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/InstIterator.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/Intrinsics.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/Statepoint.h"
36 #include "llvm/IR/Value.h"
37 #include "llvm/IR/Verifier.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Transforms/Scalar.h"
41 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
42 #include "llvm/Transforms/Utils/Cloning.h"
43 #include "llvm/Transforms/Utils/Local.h"
44 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
46 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
50 // Print the liveset found at the insert location
51 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
53 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden,
55 // Print out the base pointers for debugging
56 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden,
59 // Cost threshold measuring when it is profitable to rematerialize value instead
61 static cl::opt<unsigned>
62 RematerializationThreshold("spp-rematerialization-threshold", cl::Hidden,
65 #ifdef EXPENSIVE_CHECKS
66 static bool ClobberNonLive = true;
68 static bool ClobberNonLive = false;
70 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
71 cl::location(ClobberNonLive),
75 AllowStatepointWithNoDeoptInfo("rs4gc-allow-statepoint-with-no-deopt-info",
76 cl::Hidden, cl::init(true));
79 struct RewriteStatepointsForGC : public ModulePass {
80 static char ID; // Pass identification, replacement for typeid
82 RewriteStatepointsForGC() : ModulePass(ID) {
83 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
85 bool runOnFunction(Function &F);
86 bool runOnModule(Module &M) override {
89 Changed |= runOnFunction(F);
92 // stripNonValidAttributes asserts that shouldRewriteStatepointsIn
93 // returns true for at least one function in the module. Since at least
94 // one function changed, we know that the precondition is satisfied.
95 stripNonValidAttributes(M);
101 void getAnalysisUsage(AnalysisUsage &AU) const override {
102 // We add and rewrite a bunch of instructions, but don't really do much
103 // else. We could in theory preserve a lot more analyses here.
104 AU.addRequired<DominatorTreeWrapperPass>();
105 AU.addRequired<TargetTransformInfoWrapperPass>();
108 /// The IR fed into RewriteStatepointsForGC may have had attributes implying
109 /// dereferenceability that are no longer valid/correct after
110 /// RewriteStatepointsForGC has run. This is because semantically, after
111 /// RewriteStatepointsForGC runs, all calls to gc.statepoint "free" the entire
112 /// heap. stripNonValidAttributes (conservatively) restores correctness
113 /// by erasing all attributes in the module that externally imply
114 /// dereferenceability.
115 /// Similar reasoning also applies to the noalias attributes. gc.statepoint
116 /// can touch the entire heap including noalias objects.
117 void stripNonValidAttributes(Module &M);
119 // Helpers for stripNonValidAttributes
120 void stripNonValidAttributesFromBody(Function &F);
121 void stripNonValidAttributesFromPrototype(Function &F);
125 char RewriteStatepointsForGC::ID = 0;
127 ModulePass *llvm::createRewriteStatepointsForGCPass() {
128 return new RewriteStatepointsForGC();
131 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
132 "Make relocations explicit at statepoints", false, false)
133 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
134 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
135 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
136 "Make relocations explicit at statepoints", false, false)
139 struct GCPtrLivenessData {
140 /// Values defined in this block.
141 MapVector<BasicBlock *, SetVector<Value *>> KillSet;
142 /// Values used in this block (and thus live); does not included values
143 /// killed within this block.
144 MapVector<BasicBlock *, SetVector<Value *>> LiveSet;
146 /// Values live into this basic block (i.e. used by any
147 /// instruction in this basic block or ones reachable from here)
148 MapVector<BasicBlock *, SetVector<Value *>> LiveIn;
150 /// Values live out of this basic block (i.e. live into
151 /// any successor block)
152 MapVector<BasicBlock *, SetVector<Value *>> LiveOut;
155 // The type of the internal cache used inside the findBasePointers family
156 // of functions. From the callers perspective, this is an opaque type and
157 // should not be inspected.
159 // In the actual implementation this caches two relations:
160 // - The base relation itself (i.e. this pointer is based on that one)
161 // - The base defining value relation (i.e. before base_phi insertion)
162 // Generally, after the execution of a full findBasePointer call, only the
163 // base relation will remain. Internally, we add a mixture of the two
164 // types, then update all the second type to the first type
165 typedef MapVector<Value *, Value *> DefiningValueMapTy;
166 typedef SetVector<Value *> StatepointLiveSetTy;
167 typedef MapVector<AssertingVH<Instruction>, AssertingVH<Value>>
168 RematerializedValueMapTy;
170 struct PartiallyConstructedSafepointRecord {
171 /// The set of values known to be live across this safepoint
172 StatepointLiveSetTy LiveSet;
174 /// Mapping from live pointers to a base-defining-value
175 MapVector<Value *, Value *> PointerToBase;
177 /// The *new* gc.statepoint instruction itself. This produces the token
178 /// that normal path gc.relocates and the gc.result are tied to.
179 Instruction *StatepointToken;
181 /// Instruction to which exceptional gc relocates are attached
182 /// Makes it easier to iterate through them during relocationViaAlloca.
183 Instruction *UnwindToken;
185 /// Record live values we are rematerialized instead of relocating.
186 /// They are not included into 'LiveSet' field.
187 /// Maps rematerialized copy to it's original value.
188 RematerializedValueMapTy RematerializedValues;
192 static ArrayRef<Use> GetDeoptBundleOperands(ImmutableCallSite CS) {
193 Optional<OperandBundleUse> DeoptBundle =
194 CS.getOperandBundle(LLVMContext::OB_deopt);
196 if (!DeoptBundle.hasValue()) {
197 assert(AllowStatepointWithNoDeoptInfo &&
198 "Found non-leaf call without deopt info!");
202 return DeoptBundle.getValue().Inputs;
205 /// Compute the live-in set for every basic block in the function
206 static void computeLiveInValues(DominatorTree &DT, Function &F,
207 GCPtrLivenessData &Data);
209 /// Given results from the dataflow liveness computation, find the set of live
210 /// Values at a particular instruction.
211 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
212 StatepointLiveSetTy &out);
214 // TODO: Once we can get to the GCStrategy, this becomes
215 // Optional<bool> isGCManagedPointer(const Type *Ty) const override {
217 static bool isGCPointerType(Type *T) {
218 if (auto *PT = dyn_cast<PointerType>(T))
219 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
220 // GC managed heap. We know that a pointer into this heap needs to be
221 // updated and that no other pointer does.
222 return PT->getAddressSpace() == 1;
226 // Return true if this type is one which a) is a gc pointer or contains a GC
227 // pointer and b) is of a type this code expects to encounter as a live value.
228 // (The insertion code will assert that a type which matches (a) and not (b)
229 // is not encountered.)
230 static bool isHandledGCPointerType(Type *T) {
231 // We fully support gc pointers
232 if (isGCPointerType(T))
234 // We partially support vectors of gc pointers. The code will assert if it
235 // can't handle something.
236 if (auto VT = dyn_cast<VectorType>(T))
237 if (isGCPointerType(VT->getElementType()))
243 /// Returns true if this type contains a gc pointer whether we know how to
244 /// handle that type or not.
245 static bool containsGCPtrType(Type *Ty) {
246 if (isGCPointerType(Ty))
248 if (VectorType *VT = dyn_cast<VectorType>(Ty))
249 return isGCPointerType(VT->getScalarType());
250 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
251 return containsGCPtrType(AT->getElementType());
252 if (StructType *ST = dyn_cast<StructType>(Ty))
253 return any_of(ST->subtypes(), containsGCPtrType);
257 // Returns true if this is a type which a) is a gc pointer or contains a GC
258 // pointer and b) is of a type which the code doesn't expect (i.e. first class
259 // aggregates). Used to trip assertions.
260 static bool isUnhandledGCPointerType(Type *Ty) {
261 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
265 // Return the name of the value suffixed with the provided value, or if the
266 // value didn't have a name, the default value specified.
267 static std::string suffixed_name_or(Value *V, StringRef Suffix,
268 StringRef DefaultName) {
269 return V->hasName() ? (V->getName() + Suffix).str() : DefaultName.str();
272 // Conservatively identifies any definitions which might be live at the
273 // given instruction. The analysis is performed immediately before the
274 // given instruction. Values defined by that instruction are not considered
275 // live. Values used by that instruction are considered live.
277 analyzeParsePointLiveness(DominatorTree &DT,
278 GCPtrLivenessData &OriginalLivenessData, CallSite CS,
279 PartiallyConstructedSafepointRecord &Result) {
280 Instruction *Inst = CS.getInstruction();
282 StatepointLiveSetTy LiveSet;
283 findLiveSetAtInst(Inst, OriginalLivenessData, LiveSet);
286 dbgs() << "Live Variables:\n";
287 for (Value *V : LiveSet)
288 dbgs() << " " << V->getName() << " " << *V << "\n";
290 if (PrintLiveSetSize) {
291 dbgs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
292 dbgs() << "Number live values: " << LiveSet.size() << "\n";
294 Result.LiveSet = LiveSet;
297 static bool isKnownBaseResult(Value *V);
299 /// A single base defining value - An immediate base defining value for an
300 /// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'.
301 /// For instructions which have multiple pointer [vector] inputs or that
302 /// transition between vector and scalar types, there is no immediate base
303 /// defining value. The 'base defining value' for 'Def' is the transitive
304 /// closure of this relation stopping at the first instruction which has no
305 /// immediate base defining value. The b.d.v. might itself be a base pointer,
306 /// but it can also be an arbitrary derived pointer.
307 struct BaseDefiningValueResult {
308 /// Contains the value which is the base defining value.
310 /// True if the base defining value is also known to be an actual base
312 const bool IsKnownBase;
313 BaseDefiningValueResult(Value *BDV, bool IsKnownBase)
314 : BDV(BDV), IsKnownBase(IsKnownBase) {
316 // Check consistency between new and old means of checking whether a BDV is
318 bool MustBeBase = isKnownBaseResult(BDV);
319 assert(!MustBeBase || MustBeBase == IsKnownBase);
325 static BaseDefiningValueResult findBaseDefiningValue(Value *I);
327 /// Return a base defining value for the 'Index' element of the given vector
328 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
329 /// 'I'. As an optimization, this method will try to determine when the
330 /// element is known to already be a base pointer. If this can be established,
331 /// the second value in the returned pair will be true. Note that either a
332 /// vector or a pointer typed value can be returned. For the former, the
333 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
334 /// If the later, the return pointer is a BDV (or possibly a base) for the
335 /// particular element in 'I'.
336 static BaseDefiningValueResult
337 findBaseDefiningValueOfVector(Value *I) {
338 // Each case parallels findBaseDefiningValue below, see that code for
339 // detailed motivation.
341 if (isa<Argument>(I))
342 // An incoming argument to the function is a base pointer
343 return BaseDefiningValueResult(I, true);
345 if (isa<Constant>(I))
346 // Base of constant vector consists only of constant null pointers.
347 // For reasoning see similar case inside 'findBaseDefiningValue' function.
348 return BaseDefiningValueResult(ConstantAggregateZero::get(I->getType()),
351 if (isa<LoadInst>(I))
352 return BaseDefiningValueResult(I, true);
354 if (isa<InsertElementInst>(I))
355 // We don't know whether this vector contains entirely base pointers or
356 // not. To be conservatively correct, we treat it as a BDV and will
357 // duplicate code as needed to construct a parallel vector of bases.
358 return BaseDefiningValueResult(I, false);
360 if (isa<ShuffleVectorInst>(I))
361 // We don't know whether this vector contains entirely base pointers or
362 // not. To be conservatively correct, we treat it as a BDV and will
363 // duplicate code as needed to construct a parallel vector of bases.
364 // TODO: There a number of local optimizations which could be applied here
365 // for particular sufflevector patterns.
366 return BaseDefiningValueResult(I, false);
368 // The behavior of getelementptr instructions is the same for vector and
369 // non-vector data types.
370 if (auto *GEP = dyn_cast<GetElementPtrInst>(I))
371 return findBaseDefiningValue(GEP->getPointerOperand());
373 // A PHI or Select is a base defining value. The outer findBasePointer
374 // algorithm is responsible for constructing a base value for this BDV.
375 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
376 "unknown vector instruction - no base found for vector element");
377 return BaseDefiningValueResult(I, false);
380 /// Helper function for findBasePointer - Will return a value which either a)
381 /// defines the base pointer for the input, b) blocks the simple search
382 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
383 /// from pointer to vector type or back.
384 static BaseDefiningValueResult findBaseDefiningValue(Value *I) {
385 assert(I->getType()->isPtrOrPtrVectorTy() &&
386 "Illegal to ask for the base pointer of a non-pointer type");
388 if (I->getType()->isVectorTy())
389 return findBaseDefiningValueOfVector(I);
391 if (isa<Argument>(I))
392 // An incoming argument to the function is a base pointer
393 // We should have never reached here if this argument isn't an gc value
394 return BaseDefiningValueResult(I, true);
396 if (isa<Constant>(I)) {
397 // We assume that objects with a constant base (e.g. a global) can't move
398 // and don't need to be reported to the collector because they are always
399 // live. Besides global references, all kinds of constants (e.g. undef,
400 // constant expressions, null pointers) can be introduced by the inliner or
401 // the optimizer, especially on dynamically dead paths.
402 // Here we treat all of them as having single null base. By doing this we
403 // trying to avoid problems reporting various conflicts in a form of
404 // "phi (const1, const2)" or "phi (const, regular gc ptr)".
405 // See constant.ll file for relevant test cases.
407 return BaseDefiningValueResult(
408 ConstantPointerNull::get(cast<PointerType>(I->getType())), true);
411 if (CastInst *CI = dyn_cast<CastInst>(I)) {
412 Value *Def = CI->stripPointerCasts();
413 // If stripping pointer casts changes the address space there is an
414 // addrspacecast in between.
415 assert(cast<PointerType>(Def->getType())->getAddressSpace() ==
416 cast<PointerType>(CI->getType())->getAddressSpace() &&
417 "unsupported addrspacecast");
418 // If we find a cast instruction here, it means we've found a cast which is
419 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
420 // handle int->ptr conversion.
421 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
422 return findBaseDefiningValue(Def);
425 if (isa<LoadInst>(I))
426 // The value loaded is an gc base itself
427 return BaseDefiningValueResult(I, true);
430 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
431 // The base of this GEP is the base
432 return findBaseDefiningValue(GEP->getPointerOperand());
434 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
435 switch (II->getIntrinsicID()) {
437 // fall through to general call handling
439 case Intrinsic::experimental_gc_statepoint:
440 llvm_unreachable("statepoints don't produce pointers");
441 case Intrinsic::experimental_gc_relocate: {
442 // Rerunning safepoint insertion after safepoints are already
443 // inserted is not supported. It could probably be made to work,
444 // but why are you doing this? There's no good reason.
445 llvm_unreachable("repeat safepoint insertion is not supported");
447 case Intrinsic::gcroot:
448 // Currently, this mechanism hasn't been extended to work with gcroot.
449 // There's no reason it couldn't be, but I haven't thought about the
450 // implications much.
452 "interaction with the gcroot mechanism is not supported");
455 // We assume that functions in the source language only return base
456 // pointers. This should probably be generalized via attributes to support
457 // both source language and internal functions.
458 if (isa<CallInst>(I) || isa<InvokeInst>(I))
459 return BaseDefiningValueResult(I, true);
461 // TODO: I have absolutely no idea how to implement this part yet. It's not
462 // necessarily hard, I just haven't really looked at it yet.
463 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
465 if (isa<AtomicCmpXchgInst>(I))
466 // A CAS is effectively a atomic store and load combined under a
467 // predicate. From the perspective of base pointers, we just treat it
469 return BaseDefiningValueResult(I, true);
471 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
472 "binary ops which don't apply to pointers");
474 // The aggregate ops. Aggregates can either be in the heap or on the
475 // stack, but in either case, this is simply a field load. As a result,
476 // this is a defining definition of the base just like a load is.
477 if (isa<ExtractValueInst>(I))
478 return BaseDefiningValueResult(I, true);
480 // We should never see an insert vector since that would require we be
481 // tracing back a struct value not a pointer value.
482 assert(!isa<InsertValueInst>(I) &&
483 "Base pointer for a struct is meaningless");
485 // An extractelement produces a base result exactly when it's input does.
486 // We may need to insert a parallel instruction to extract the appropriate
487 // element out of the base vector corresponding to the input. Given this,
488 // it's analogous to the phi and select case even though it's not a merge.
489 if (isa<ExtractElementInst>(I))
490 // Note: There a lot of obvious peephole cases here. This are deliberately
491 // handled after the main base pointer inference algorithm to make writing
492 // test cases to exercise that code easier.
493 return BaseDefiningValueResult(I, false);
495 // The last two cases here don't return a base pointer. Instead, they
496 // return a value which dynamically selects from among several base
497 // derived pointers (each with it's own base potentially). It's the job of
498 // the caller to resolve these.
499 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
500 "missing instruction case in findBaseDefiningValing");
501 return BaseDefiningValueResult(I, false);
504 /// Returns the base defining value for this value.
505 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
506 Value *&Cached = Cache[I];
508 Cached = findBaseDefiningValue(I).BDV;
509 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
510 << Cached->getName() << "\n");
512 assert(Cache[I] != nullptr);
516 /// Return a base pointer for this value if known. Otherwise, return it's
517 /// base defining value.
518 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
519 Value *Def = findBaseDefiningValueCached(I, Cache);
520 auto Found = Cache.find(Def);
521 if (Found != Cache.end()) {
522 // Either a base-of relation, or a self reference. Caller must check.
523 return Found->second;
525 // Only a BDV available
529 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
530 /// is it known to be a base pointer? Or do we need to continue searching.
531 static bool isKnownBaseResult(Value *V) {
532 if (!isa<PHINode>(V) && !isa<SelectInst>(V) &&
533 !isa<ExtractElementInst>(V) && !isa<InsertElementInst>(V) &&
534 !isa<ShuffleVectorInst>(V)) {
535 // no recursion possible
538 if (isa<Instruction>(V) &&
539 cast<Instruction>(V)->getMetadata("is_base_value")) {
540 // This is a previously inserted base phi or select. We know
541 // that this is a base value.
545 // We need to keep searching
550 /// Models the state of a single base defining value in the findBasePointer
551 /// algorithm for determining where a new instruction is needed to propagate
552 /// the base of this BDV.
555 enum Status { Unknown, Base, Conflict };
557 BDVState() : Status(Unknown), BaseValue(nullptr) {}
559 explicit BDVState(Status Status, Value *BaseValue = nullptr)
560 : Status(Status), BaseValue(BaseValue) {
561 assert(Status != Base || BaseValue);
564 explicit BDVState(Value *BaseValue) : Status(Base), BaseValue(BaseValue) {}
566 Status getStatus() const { return Status; }
567 Value *getBaseValue() const { return BaseValue; }
569 bool isBase() const { return getStatus() == Base; }
570 bool isUnknown() const { return getStatus() == Unknown; }
571 bool isConflict() const { return getStatus() == Conflict; }
573 bool operator==(const BDVState &Other) const {
574 return BaseValue == Other.BaseValue && Status == Other.Status;
577 bool operator!=(const BDVState &other) const { return !(*this == other); }
585 void print(raw_ostream &OS) const {
586 switch (getStatus()) {
597 OS << " (" << getBaseValue() << " - "
598 << (getBaseValue() ? getBaseValue()->getName() : "nullptr") << "): ";
603 AssertingVH<Value> BaseValue; // Non-null only if Status == Base.
608 static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
614 static BDVState meetBDVStateImpl(const BDVState &LHS, const BDVState &RHS) {
615 switch (LHS.getStatus()) {
616 case BDVState::Unknown:
620 assert(LHS.getBaseValue() && "can't be null");
625 if (LHS.getBaseValue() == RHS.getBaseValue()) {
626 assert(LHS == RHS && "equality broken!");
629 return BDVState(BDVState::Conflict);
631 assert(RHS.isConflict() && "only three states!");
632 return BDVState(BDVState::Conflict);
634 case BDVState::Conflict:
637 llvm_unreachable("only three states!");
640 // Values of type BDVState form a lattice, and this function implements the meet
642 static BDVState meetBDVState(const BDVState &LHS, const BDVState &RHS) {
643 BDVState Result = meetBDVStateImpl(LHS, RHS);
644 assert(Result == meetBDVStateImpl(RHS, LHS) &&
645 "Math is wrong: meet does not commute!");
649 /// For a given value or instruction, figure out what base ptr its derived from.
650 /// For gc objects, this is simply itself. On success, returns a value which is
651 /// the base pointer. (This is reliable and can be used for relocation.) On
652 /// failure, returns nullptr.
653 static Value *findBasePointer(Value *I, DefiningValueMapTy &Cache) {
654 Value *Def = findBaseOrBDV(I, Cache);
656 if (isKnownBaseResult(Def))
659 // Here's the rough algorithm:
660 // - For every SSA value, construct a mapping to either an actual base
661 // pointer or a PHI which obscures the base pointer.
662 // - Construct a mapping from PHI to unknown TOP state. Use an
663 // optimistic algorithm to propagate base pointer information. Lattice
668 // When algorithm terminates, all PHIs will either have a single concrete
669 // base or be in a conflict state.
670 // - For every conflict, insert a dummy PHI node without arguments. Add
671 // these to the base[Instruction] = BasePtr mapping. For every
672 // non-conflict, add the actual base.
673 // - For every conflict, add arguments for the base[a] of each input
676 // Note: A simpler form of this would be to add the conflict form of all
677 // PHIs without running the optimistic algorithm. This would be
678 // analogous to pessimistic data flow and would likely lead to an
679 // overall worse solution.
682 auto isExpectedBDVType = [](Value *BDV) {
683 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) ||
684 isa<ExtractElementInst>(BDV) || isa<InsertElementInst>(BDV) ||
685 isa<ShuffleVectorInst>(BDV);
689 // Once populated, will contain a mapping from each potentially non-base BDV
690 // to a lattice value (described above) which corresponds to that BDV.
691 // We use the order of insertion (DFS over the def/use graph) to provide a
692 // stable deterministic ordering for visiting DenseMaps (which are unordered)
693 // below. This is important for deterministic compilation.
694 MapVector<Value *, BDVState> States;
696 // Recursively fill in all base defining values reachable from the initial
697 // one for which we don't already know a definite base value for
699 SmallVector<Value*, 16> Worklist;
700 Worklist.push_back(Def);
701 States.insert({Def, BDVState()});
702 while (!Worklist.empty()) {
703 Value *Current = Worklist.pop_back_val();
704 assert(!isKnownBaseResult(Current) && "why did it get added?");
706 auto visitIncomingValue = [&](Value *InVal) {
707 Value *Base = findBaseOrBDV(InVal, Cache);
708 if (isKnownBaseResult(Base))
709 // Known bases won't need new instructions introduced and can be
712 assert(isExpectedBDVType(Base) && "the only non-base values "
713 "we see should be base defining values");
714 if (States.insert(std::make_pair(Base, BDVState())).second)
715 Worklist.push_back(Base);
717 if (PHINode *PN = dyn_cast<PHINode>(Current)) {
718 for (Value *InVal : PN->incoming_values())
719 visitIncomingValue(InVal);
720 } else if (SelectInst *SI = dyn_cast<SelectInst>(Current)) {
721 visitIncomingValue(SI->getTrueValue());
722 visitIncomingValue(SI->getFalseValue());
723 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
724 visitIncomingValue(EE->getVectorOperand());
725 } else if (auto *IE = dyn_cast<InsertElementInst>(Current)) {
726 visitIncomingValue(IE->getOperand(0)); // vector operand
727 visitIncomingValue(IE->getOperand(1)); // scalar operand
728 } else if (auto *SV = dyn_cast<ShuffleVectorInst>(Current)) {
729 visitIncomingValue(SV->getOperand(0));
730 visitIncomingValue(SV->getOperand(1));
733 llvm_unreachable("Unimplemented instruction case");
739 DEBUG(dbgs() << "States after initialization:\n");
740 for (auto Pair : States) {
741 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
745 // Return a phi state for a base defining value. We'll generate a new
746 // base state for known bases and expect to find a cached state otherwise.
747 auto getStateForBDV = [&](Value *baseValue) {
748 if (isKnownBaseResult(baseValue))
749 return BDVState(baseValue);
750 auto I = States.find(baseValue);
751 assert(I != States.end() && "lookup failed!");
755 bool Progress = true;
758 const size_t OldSize = States.size();
761 // We're only changing values in this loop, thus safe to keep iterators.
762 // Since this is computing a fixed point, the order of visit does not
763 // effect the result. TODO: We could use a worklist here and make this run
765 for (auto Pair : States) {
766 Value *BDV = Pair.first;
767 assert(!isKnownBaseResult(BDV) && "why did it get added?");
769 // Given an input value for the current instruction, return a BDVState
770 // instance which represents the BDV of that value.
771 auto getStateForInput = [&](Value *V) mutable {
772 Value *BDV = findBaseOrBDV(V, Cache);
773 return getStateForBDV(BDV);
777 if (SelectInst *SI = dyn_cast<SelectInst>(BDV)) {
778 NewState = meetBDVState(NewState, getStateForInput(SI->getTrueValue()));
780 meetBDVState(NewState, getStateForInput(SI->getFalseValue()));
781 } else if (PHINode *PN = dyn_cast<PHINode>(BDV)) {
782 for (Value *Val : PN->incoming_values())
783 NewState = meetBDVState(NewState, getStateForInput(Val));
784 } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) {
785 // The 'meet' for an extractelement is slightly trivial, but it's still
786 // useful in that it drives us to conflict if our input is.
788 meetBDVState(NewState, getStateForInput(EE->getVectorOperand()));
789 } else if (auto *IE = dyn_cast<InsertElementInst>(BDV)){
790 // Given there's a inherent type mismatch between the operands, will
791 // *always* produce Conflict.
792 NewState = meetBDVState(NewState, getStateForInput(IE->getOperand(0)));
793 NewState = meetBDVState(NewState, getStateForInput(IE->getOperand(1)));
795 // The only instance this does not return a Conflict is when both the
796 // vector operands are the same vector.
797 auto *SV = cast<ShuffleVectorInst>(BDV);
798 NewState = meetBDVState(NewState, getStateForInput(SV->getOperand(0)));
799 NewState = meetBDVState(NewState, getStateForInput(SV->getOperand(1)));
802 BDVState OldState = States[BDV];
803 if (OldState != NewState) {
805 States[BDV] = NewState;
809 assert(OldSize == States.size() &&
810 "fixed point shouldn't be adding any new nodes to state");
814 DEBUG(dbgs() << "States after meet iteration:\n");
815 for (auto Pair : States) {
816 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
820 // Insert Phis for all conflicts
821 // TODO: adjust naming patterns to avoid this order of iteration dependency
822 for (auto Pair : States) {
823 Instruction *I = cast<Instruction>(Pair.first);
824 BDVState State = Pair.second;
825 assert(!isKnownBaseResult(I) && "why did it get added?");
826 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
828 // extractelement instructions are a bit special in that we may need to
829 // insert an extract even when we know an exact base for the instruction.
830 // The problem is that we need to convert from a vector base to a scalar
831 // base for the particular indice we're interested in.
832 if (State.isBase() && isa<ExtractElementInst>(I) &&
833 isa<VectorType>(State.getBaseValue()->getType())) {
834 auto *EE = cast<ExtractElementInst>(I);
835 // TODO: In many cases, the new instruction is just EE itself. We should
836 // exploit this, but can't do it here since it would break the invariant
837 // about the BDV not being known to be a base.
838 auto *BaseInst = ExtractElementInst::Create(
839 State.getBaseValue(), EE->getIndexOperand(), "base_ee", EE);
840 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
841 States[I] = BDVState(BDVState::Base, BaseInst);
844 // Since we're joining a vector and scalar base, they can never be the
845 // same. As a result, we should always see insert element having reached
846 // the conflict state.
847 assert(!isa<InsertElementInst>(I) || State.isConflict());
849 if (!State.isConflict())
852 /// Create and insert a new instruction which will represent the base of
853 /// the given instruction 'I'.
854 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
855 if (isa<PHINode>(I)) {
856 BasicBlock *BB = I->getParent();
857 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
858 assert(NumPreds > 0 && "how did we reach here");
859 std::string Name = suffixed_name_or(I, ".base", "base_phi");
860 return PHINode::Create(I->getType(), NumPreds, Name, I);
861 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
862 // The undef will be replaced later
863 UndefValue *Undef = UndefValue::get(SI->getType());
864 std::string Name = suffixed_name_or(I, ".base", "base_select");
865 return SelectInst::Create(SI->getCondition(), Undef, Undef, Name, SI);
866 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
867 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
868 std::string Name = suffixed_name_or(I, ".base", "base_ee");
869 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
871 } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
872 UndefValue *VecUndef = UndefValue::get(IE->getOperand(0)->getType());
873 UndefValue *ScalarUndef = UndefValue::get(IE->getOperand(1)->getType());
874 std::string Name = suffixed_name_or(I, ".base", "base_ie");
875 return InsertElementInst::Create(VecUndef, ScalarUndef,
876 IE->getOperand(2), Name, IE);
878 auto *SV = cast<ShuffleVectorInst>(I);
879 UndefValue *VecUndef = UndefValue::get(SV->getOperand(0)->getType());
880 std::string Name = suffixed_name_or(I, ".base", "base_sv");
881 return new ShuffleVectorInst(VecUndef, VecUndef, SV->getOperand(2),
885 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
886 // Add metadata marking this as a base value
887 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
888 States[I] = BDVState(BDVState::Conflict, BaseInst);
891 // Returns a instruction which produces the base pointer for a given
892 // instruction. The instruction is assumed to be an input to one of the BDVs
893 // seen in the inference algorithm above. As such, we must either already
894 // know it's base defining value is a base, or have inserted a new
895 // instruction to propagate the base of it's BDV and have entered that newly
896 // introduced instruction into the state table. In either case, we are
897 // assured to be able to determine an instruction which produces it's base
899 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
900 Value *BDV = findBaseOrBDV(Input, Cache);
901 Value *Base = nullptr;
902 if (isKnownBaseResult(BDV)) {
905 // Either conflict or base.
906 assert(States.count(BDV));
907 Base = States[BDV].getBaseValue();
909 assert(Base && "Can't be null");
910 // The cast is needed since base traversal may strip away bitcasts
911 if (Base->getType() != Input->getType() && InsertPt)
912 Base = new BitCastInst(Base, Input->getType(), "cast", InsertPt);
916 // Fixup all the inputs of the new PHIs. Visit order needs to be
917 // deterministic and predictable because we're naming newly created
919 for (auto Pair : States) {
920 Instruction *BDV = cast<Instruction>(Pair.first);
921 BDVState State = Pair.second;
923 assert(!isKnownBaseResult(BDV) && "why did it get added?");
924 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
925 if (!State.isConflict())
928 if (PHINode *BasePHI = dyn_cast<PHINode>(State.getBaseValue())) {
929 PHINode *PN = cast<PHINode>(BDV);
930 unsigned NumPHIValues = PN->getNumIncomingValues();
931 for (unsigned i = 0; i < NumPHIValues; i++) {
932 Value *InVal = PN->getIncomingValue(i);
933 BasicBlock *InBB = PN->getIncomingBlock(i);
935 // If we've already seen InBB, add the same incoming value
936 // we added for it earlier. The IR verifier requires phi
937 // nodes with multiple entries from the same basic block
938 // to have the same incoming value for each of those
939 // entries. If we don't do this check here and basephi
940 // has a different type than base, we'll end up adding two
941 // bitcasts (and hence two distinct values) as incoming
942 // values for the same basic block.
944 int BlockIndex = BasePHI->getBasicBlockIndex(InBB);
945 if (BlockIndex != -1) {
946 Value *OldBase = BasePHI->getIncomingValue(BlockIndex);
947 BasePHI->addIncoming(OldBase, InBB);
950 Value *Base = getBaseForInput(InVal, nullptr);
951 // In essence this assert states: the only way two values
952 // incoming from the same basic block may be different is by
953 // being different bitcasts of the same value. A cleanup
954 // that remains TODO is changing findBaseOrBDV to return an
955 // llvm::Value of the correct type (and still remain pure).
956 // This will remove the need to add bitcasts.
957 assert(Base->stripPointerCasts() == OldBase->stripPointerCasts() &&
958 "Sanity -- findBaseOrBDV should be pure!");
963 // Find the instruction which produces the base for each input. We may
964 // need to insert a bitcast in the incoming block.
965 // TODO: Need to split critical edges if insertion is needed
966 Value *Base = getBaseForInput(InVal, InBB->getTerminator());
967 BasePHI->addIncoming(Base, InBB);
969 assert(BasePHI->getNumIncomingValues() == NumPHIValues);
970 } else if (SelectInst *BaseSI =
971 dyn_cast<SelectInst>(State.getBaseValue())) {
972 SelectInst *SI = cast<SelectInst>(BDV);
974 // Find the instruction which produces the base for each input.
975 // We may need to insert a bitcast.
976 BaseSI->setTrueValue(getBaseForInput(SI->getTrueValue(), BaseSI));
977 BaseSI->setFalseValue(getBaseForInput(SI->getFalseValue(), BaseSI));
978 } else if (auto *BaseEE =
979 dyn_cast<ExtractElementInst>(State.getBaseValue())) {
980 Value *InVal = cast<ExtractElementInst>(BDV)->getVectorOperand();
981 // Find the instruction which produces the base for each input. We may
982 // need to insert a bitcast.
983 BaseEE->setOperand(0, getBaseForInput(InVal, BaseEE));
984 } else if (auto *BaseIE = dyn_cast<InsertElementInst>(State.getBaseValue())){
985 auto *BdvIE = cast<InsertElementInst>(BDV);
986 auto UpdateOperand = [&](int OperandIdx) {
987 Value *InVal = BdvIE->getOperand(OperandIdx);
988 Value *Base = getBaseForInput(InVal, BaseIE);
989 BaseIE->setOperand(OperandIdx, Base);
991 UpdateOperand(0); // vector operand
992 UpdateOperand(1); // scalar operand
994 auto *BaseSV = cast<ShuffleVectorInst>(State.getBaseValue());
995 auto *BdvSV = cast<ShuffleVectorInst>(BDV);
996 auto UpdateOperand = [&](int OperandIdx) {
997 Value *InVal = BdvSV->getOperand(OperandIdx);
998 Value *Base = getBaseForInput(InVal, BaseSV);
999 BaseSV->setOperand(OperandIdx, Base);
1001 UpdateOperand(0); // vector operand
1002 UpdateOperand(1); // vector operand
1006 // Cache all of our results so we can cheaply reuse them
1007 // NOTE: This is actually two caches: one of the base defining value
1008 // relation and one of the base pointer relation! FIXME
1009 for (auto Pair : States) {
1010 auto *BDV = Pair.first;
1011 Value *Base = Pair.second.getBaseValue();
1012 assert(BDV && Base);
1013 assert(!isKnownBaseResult(BDV) && "why did it get added?");
1015 DEBUG(dbgs() << "Updating base value cache"
1016 << " for: " << BDV->getName() << " from: "
1017 << (Cache.count(BDV) ? Cache[BDV]->getName().str() : "none")
1018 << " to: " << Base->getName() << "\n");
1020 if (Cache.count(BDV)) {
1021 assert(isKnownBaseResult(Base) &&
1022 "must be something we 'know' is a base pointer");
1023 // Once we transition from the BDV relation being store in the Cache to
1024 // the base relation being stored, it must be stable
1025 assert((!isKnownBaseResult(Cache[BDV]) || Cache[BDV] == Base) &&
1026 "base relation should be stable");
1030 assert(Cache.count(Def));
1034 // For a set of live pointers (base and/or derived), identify the base
1035 // pointer of the object which they are derived from. This routine will
1036 // mutate the IR graph as needed to make the 'base' pointer live at the
1037 // definition site of 'derived'. This ensures that any use of 'derived' can
1038 // also use 'base'. This may involve the insertion of a number of
1039 // additional PHI nodes.
1041 // preconditions: live is a set of pointer type Values
1043 // side effects: may insert PHI nodes into the existing CFG, will preserve
1044 // CFG, will not remove or mutate any existing nodes
1046 // post condition: PointerToBase contains one (derived, base) pair for every
1047 // pointer in live. Note that derived can be equal to base if the original
1048 // pointer was a base pointer.
1050 findBasePointers(const StatepointLiveSetTy &live,
1051 MapVector<Value *, Value *> &PointerToBase,
1052 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1053 for (Value *ptr : live) {
1054 Value *base = findBasePointer(ptr, DVCache);
1055 assert(base && "failed to find base pointer");
1056 PointerToBase[ptr] = base;
1057 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1058 DT->dominates(cast<Instruction>(base)->getParent(),
1059 cast<Instruction>(ptr)->getParent())) &&
1060 "The base we found better dominate the derived pointer");
1064 /// Find the required based pointers (and adjust the live set) for the given
1066 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1068 PartiallyConstructedSafepointRecord &result) {
1069 MapVector<Value *, Value *> PointerToBase;
1070 findBasePointers(result.LiveSet, PointerToBase, &DT, DVCache);
1072 if (PrintBasePointers) {
1073 errs() << "Base Pairs (w/o Relocation):\n";
1074 for (auto &Pair : PointerToBase) {
1075 errs() << " derived ";
1076 Pair.first->printAsOperand(errs(), false);
1078 Pair.second->printAsOperand(errs(), false);
1083 result.PointerToBase = PointerToBase;
1086 /// Given an updated version of the dataflow liveness results, update the
1087 /// liveset and base pointer maps for the call site CS.
1088 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1090 PartiallyConstructedSafepointRecord &result);
1092 static void recomputeLiveInValues(
1093 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1094 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1095 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1096 // again. The old values are still live and will help it stabilize quickly.
1097 GCPtrLivenessData RevisedLivenessData;
1098 computeLiveInValues(DT, F, RevisedLivenessData);
1099 for (size_t i = 0; i < records.size(); i++) {
1100 struct PartiallyConstructedSafepointRecord &info = records[i];
1101 recomputeLiveInValues(RevisedLivenessData, toUpdate[i], info);
1105 // When inserting gc.relocate and gc.result calls, we need to ensure there are
1106 // no uses of the original value / return value between the gc.statepoint and
1107 // the gc.relocate / gc.result call. One case which can arise is a phi node
1108 // starting one of the successor blocks. We also need to be able to insert the
1109 // gc.relocates only on the path which goes through the statepoint. We might
1110 // need to split an edge to make this possible.
1112 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1113 DominatorTree &DT) {
1114 BasicBlock *Ret = BB;
1115 if (!BB->getUniquePredecessor())
1116 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1118 // Now that 'Ret' has unique predecessor we can safely remove all phi nodes
1120 FoldSingleEntryPHINodes(Ret);
1121 assert(!isa<PHINode>(Ret->begin()) &&
1122 "All PHI nodes should have been removed!");
1124 // At this point, we can safely insert a gc.relocate or gc.result as the first
1125 // instruction in Ret if needed.
1129 // Create new attribute set containing only attributes which can be transferred
1130 // from original call to the safepoint.
1131 static AttributeList legalizeCallAttributes(AttributeList AL) {
1135 // Remove the readonly, readnone, and statepoint function attributes.
1136 AttrBuilder FnAttrs = AL.getFnAttributes();
1137 FnAttrs.removeAttribute(Attribute::ReadNone);
1138 FnAttrs.removeAttribute(Attribute::ReadOnly);
1139 for (Attribute A : AL.getFnAttributes()) {
1140 if (isStatepointDirectiveAttr(A))
1144 // Just skip parameter and return attributes for now
1145 LLVMContext &Ctx = AL.getContext();
1146 return AttributeList::get(Ctx, AttributeList::FunctionIndex,
1147 AttributeSet::get(Ctx, FnAttrs));
1150 /// Helper function to place all gc relocates necessary for the given
1153 /// liveVariables - list of variables to be relocated.
1154 /// liveStart - index of the first live variable.
1155 /// basePtrs - base pointers.
1156 /// statepointToken - statepoint instruction to which relocates should be
1158 /// Builder - Llvm IR builder to be used to construct new calls.
1159 static void CreateGCRelocates(ArrayRef<Value *> LiveVariables,
1160 const int LiveStart,
1161 ArrayRef<Value *> BasePtrs,
1162 Instruction *StatepointToken,
1163 IRBuilder<> Builder) {
1164 if (LiveVariables.empty())
1167 auto FindIndex = [](ArrayRef<Value *> LiveVec, Value *Val) {
1168 auto ValIt = find(LiveVec, Val);
1169 assert(ValIt != LiveVec.end() && "Val not found in LiveVec!");
1170 size_t Index = std::distance(LiveVec.begin(), ValIt);
1171 assert(Index < LiveVec.size() && "Bug in std::find?");
1174 Module *M = StatepointToken->getModule();
1176 // All gc_relocate are generated as i8 addrspace(1)* (or a vector type whose
1177 // element type is i8 addrspace(1)*). We originally generated unique
1178 // declarations for each pointer type, but this proved problematic because
1179 // the intrinsic mangling code is incomplete and fragile. Since we're moving
1180 // towards a single unified pointer type anyways, we can just cast everything
1181 // to an i8* of the right address space. A bitcast is added later to convert
1182 // gc_relocate to the actual value's type.
1183 auto getGCRelocateDecl = [&] (Type *Ty) {
1184 assert(isHandledGCPointerType(Ty));
1185 auto AS = Ty->getScalarType()->getPointerAddressSpace();
1186 Type *NewTy = Type::getInt8PtrTy(M->getContext(), AS);
1187 if (auto *VT = dyn_cast<VectorType>(Ty))
1188 NewTy = VectorType::get(NewTy, VT->getNumElements());
1189 return Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate,
1193 // Lazily populated map from input types to the canonicalized form mentioned
1194 // in the comment above. This should probably be cached somewhere more
1196 DenseMap<Type*, Value*> TypeToDeclMap;
1198 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1199 // Generate the gc.relocate call and save the result
1201 Builder.getInt32(LiveStart + FindIndex(LiveVariables, BasePtrs[i]));
1202 Value *LiveIdx = Builder.getInt32(LiveStart + i);
1204 Type *Ty = LiveVariables[i]->getType();
1205 if (!TypeToDeclMap.count(Ty))
1206 TypeToDeclMap[Ty] = getGCRelocateDecl(Ty);
1207 Value *GCRelocateDecl = TypeToDeclMap[Ty];
1209 // only specify a debug name if we can give a useful one
1210 CallInst *Reloc = Builder.CreateCall(
1211 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1212 suffixed_name_or(LiveVariables[i], ".relocated", ""));
1213 // Trick CodeGen into thinking there are lots of free registers at this
1215 Reloc->setCallingConv(CallingConv::Cold);
1221 /// This struct is used to defer RAUWs and `eraseFromParent` s. Using this
1222 /// avoids having to worry about keeping around dangling pointers to Values.
1223 class DeferredReplacement {
1224 AssertingVH<Instruction> Old;
1225 AssertingVH<Instruction> New;
1226 bool IsDeoptimize = false;
1228 DeferredReplacement() {}
1231 static DeferredReplacement createRAUW(Instruction *Old, Instruction *New) {
1232 assert(Old != New && Old && New &&
1233 "Cannot RAUW equal values or to / from null!");
1235 DeferredReplacement D;
1241 static DeferredReplacement createDelete(Instruction *ToErase) {
1242 DeferredReplacement D;
1247 static DeferredReplacement createDeoptimizeReplacement(Instruction *Old) {
1249 auto *F = cast<CallInst>(Old)->getCalledFunction();
1250 assert(F && F->getIntrinsicID() == Intrinsic::experimental_deoptimize &&
1251 "Only way to construct a deoptimize deferred replacement");
1253 DeferredReplacement D;
1255 D.IsDeoptimize = true;
1259 /// Does the task represented by this instance.
1260 void doReplacement() {
1261 Instruction *OldI = Old;
1262 Instruction *NewI = New;
1264 assert(OldI != NewI && "Disallowed at construction?!");
1265 assert((!IsDeoptimize || !New) &&
1266 "Deoptimize instrinsics are not replaced!");
1272 OldI->replaceAllUsesWith(NewI);
1275 // Note: we've inserted instructions, so the call to llvm.deoptimize may
1276 // not necessarilly be followed by the matching return.
1277 auto *RI = cast<ReturnInst>(OldI->getParent()->getTerminator());
1278 new UnreachableInst(RI->getContext(), RI);
1279 RI->eraseFromParent();
1282 OldI->eraseFromParent();
1287 static StringRef getDeoptLowering(CallSite CS) {
1288 const char *DeoptLowering = "deopt-lowering";
1289 if (CS.hasFnAttr(DeoptLowering)) {
1290 // FIXME: CallSite has a *really* confusing interface around attributes
1292 const AttributeList &CSAS = CS.getAttributes();
1293 if (CSAS.hasAttribute(AttributeList::FunctionIndex, DeoptLowering))
1294 return CSAS.getAttribute(AttributeList::FunctionIndex, DeoptLowering)
1295 .getValueAsString();
1296 Function *F = CS.getCalledFunction();
1297 assert(F && F->hasFnAttribute(DeoptLowering));
1298 return F->getFnAttribute(DeoptLowering).getValueAsString();
1300 return "live-through";
1305 makeStatepointExplicitImpl(const CallSite CS, /* to replace */
1306 const SmallVectorImpl<Value *> &BasePtrs,
1307 const SmallVectorImpl<Value *> &LiveVariables,
1308 PartiallyConstructedSafepointRecord &Result,
1309 std::vector<DeferredReplacement> &Replacements) {
1310 assert(BasePtrs.size() == LiveVariables.size());
1312 // Then go ahead and use the builder do actually do the inserts. We insert
1313 // immediately before the previous instruction under the assumption that all
1314 // arguments will be available here. We can't insert afterwards since we may
1315 // be replacing a terminator.
1316 Instruction *InsertBefore = CS.getInstruction();
1317 IRBuilder<> Builder(InsertBefore);
1319 ArrayRef<Value *> GCArgs(LiveVariables);
1320 uint64_t StatepointID = StatepointDirectives::DefaultStatepointID;
1321 uint32_t NumPatchBytes = 0;
1322 uint32_t Flags = uint32_t(StatepointFlags::None);
1324 ArrayRef<Use> CallArgs(CS.arg_begin(), CS.arg_end());
1325 ArrayRef<Use> DeoptArgs = GetDeoptBundleOperands(CS);
1326 ArrayRef<Use> TransitionArgs;
1327 if (auto TransitionBundle =
1328 CS.getOperandBundle(LLVMContext::OB_gc_transition)) {
1329 Flags |= uint32_t(StatepointFlags::GCTransition);
1330 TransitionArgs = TransitionBundle->Inputs;
1333 // Instead of lowering calls to @llvm.experimental.deoptimize as normal calls
1334 // with a return value, we lower then as never returning calls to
1335 // __llvm_deoptimize that are followed by unreachable to get better codegen.
1336 bool IsDeoptimize = false;
1338 StatepointDirectives SD =
1339 parseStatepointDirectivesFromAttrs(CS.getAttributes());
1340 if (SD.NumPatchBytes)
1341 NumPatchBytes = *SD.NumPatchBytes;
1342 if (SD.StatepointID)
1343 StatepointID = *SD.StatepointID;
1345 // Pass through the requested lowering if any. The default is live-through.
1346 StringRef DeoptLowering = getDeoptLowering(CS);
1347 if (DeoptLowering.equals("live-in"))
1348 Flags |= uint32_t(StatepointFlags::DeoptLiveIn);
1350 assert(DeoptLowering.equals("live-through") && "Unsupported value!");
1353 Value *CallTarget = CS.getCalledValue();
1354 if (Function *F = dyn_cast<Function>(CallTarget)) {
1355 if (F->getIntrinsicID() == Intrinsic::experimental_deoptimize) {
1356 // Calls to llvm.experimental.deoptimize are lowered to calls to the
1357 // __llvm_deoptimize symbol. We want to resolve this now, since the
1358 // verifier does not allow taking the address of an intrinsic function.
1360 SmallVector<Type *, 8> DomainTy;
1361 for (Value *Arg : CallArgs)
1362 DomainTy.push_back(Arg->getType());
1363 auto *FTy = FunctionType::get(Type::getVoidTy(F->getContext()), DomainTy,
1364 /* isVarArg = */ false);
1366 // Note: CallTarget can be a bitcast instruction of a symbol if there are
1367 // calls to @llvm.experimental.deoptimize with different argument types in
1368 // the same module. This is fine -- we assume the frontend knew what it
1369 // was doing when generating this kind of IR.
1371 F->getParent()->getOrInsertFunction("__llvm_deoptimize", FTy);
1373 IsDeoptimize = true;
1377 // Create the statepoint given all the arguments
1378 Instruction *Token = nullptr;
1380 CallInst *ToReplace = cast<CallInst>(CS.getInstruction());
1381 CallInst *Call = Builder.CreateGCStatepointCall(
1382 StatepointID, NumPatchBytes, CallTarget, Flags, CallArgs,
1383 TransitionArgs, DeoptArgs, GCArgs, "safepoint_token");
1385 Call->setTailCallKind(ToReplace->getTailCallKind());
1386 Call->setCallingConv(ToReplace->getCallingConv());
1388 // Currently we will fail on parameter attributes and on certain
1389 // function attributes. In case if we can handle this set of attributes -
1390 // set up function attrs directly on statepoint and return attrs later for
1391 // gc_result intrinsic.
1392 Call->setAttributes(legalizeCallAttributes(ToReplace->getAttributes()));
1396 // Put the following gc_result and gc_relocate calls immediately after the
1397 // the old call (which we're about to delete)
1398 assert(ToReplace->getNextNode() && "Not a terminator, must have next!");
1399 Builder.SetInsertPoint(ToReplace->getNextNode());
1400 Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc());
1402 InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction());
1404 // Insert the new invoke into the old block. We'll remove the old one in a
1405 // moment at which point this will become the new terminator for the
1407 InvokeInst *Invoke = Builder.CreateGCStatepointInvoke(
1408 StatepointID, NumPatchBytes, CallTarget, ToReplace->getNormalDest(),
1409 ToReplace->getUnwindDest(), Flags, CallArgs, TransitionArgs, DeoptArgs,
1410 GCArgs, "statepoint_token");
1412 Invoke->setCallingConv(ToReplace->getCallingConv());
1414 // Currently we will fail on parameter attributes and on certain
1415 // function attributes. In case if we can handle this set of attributes -
1416 // set up function attrs directly on statepoint and return attrs later for
1417 // gc_result intrinsic.
1418 Invoke->setAttributes(legalizeCallAttributes(ToReplace->getAttributes()));
1422 // Generate gc relocates in exceptional path
1423 BasicBlock *UnwindBlock = ToReplace->getUnwindDest();
1424 assert(!isa<PHINode>(UnwindBlock->begin()) &&
1425 UnwindBlock->getUniquePredecessor() &&
1426 "can't safely insert in this block!");
1428 Builder.SetInsertPoint(&*UnwindBlock->getFirstInsertionPt());
1429 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1431 // Attach exceptional gc relocates to the landingpad.
1432 Instruction *ExceptionalToken = UnwindBlock->getLandingPadInst();
1433 Result.UnwindToken = ExceptionalToken;
1435 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1436 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, ExceptionalToken,
1439 // Generate gc relocates and returns for normal block
1440 BasicBlock *NormalDest = ToReplace->getNormalDest();
1441 assert(!isa<PHINode>(NormalDest->begin()) &&
1442 NormalDest->getUniquePredecessor() &&
1443 "can't safely insert in this block!");
1445 Builder.SetInsertPoint(&*NormalDest->getFirstInsertionPt());
1447 // gc relocates will be generated later as if it were regular call
1450 assert(Token && "Should be set in one of the above branches!");
1453 // If we're wrapping an @llvm.experimental.deoptimize in a statepoint, we
1454 // transform the tail-call like structure to a call to a void function
1455 // followed by unreachable to get better codegen.
1456 Replacements.push_back(
1457 DeferredReplacement::createDeoptimizeReplacement(CS.getInstruction()));
1459 Token->setName("statepoint_token");
1460 if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
1462 CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "";
1463 CallInst *GCResult = Builder.CreateGCResult(Token, CS.getType(), Name);
1464 GCResult->setAttributes(
1465 AttributeList::get(GCResult->getContext(), AttributeList::ReturnIndex,
1466 CS.getAttributes().getRetAttributes()));
1468 // We cannot RAUW or delete CS.getInstruction() because it could be in the
1469 // live set of some other safepoint, in which case that safepoint's
1470 // PartiallyConstructedSafepointRecord will hold a raw pointer to this
1471 // llvm::Instruction. Instead, we defer the replacement and deletion to
1472 // after the live sets have been made explicit in the IR, and we no longer
1473 // have raw pointers to worry about.
1474 Replacements.emplace_back(
1475 DeferredReplacement::createRAUW(CS.getInstruction(), GCResult));
1477 Replacements.emplace_back(
1478 DeferredReplacement::createDelete(CS.getInstruction()));
1482 Result.StatepointToken = Token;
1484 // Second, create a gc.relocate for every live variable
1485 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1486 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, Token, Builder);
1489 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1490 // which make the relocations happening at this safepoint explicit.
1492 // WARNING: Does not do any fixup to adjust users of the original live
1493 // values. That's the callers responsibility.
1495 makeStatepointExplicit(DominatorTree &DT, CallSite CS,
1496 PartiallyConstructedSafepointRecord &Result,
1497 std::vector<DeferredReplacement> &Replacements) {
1498 const auto &LiveSet = Result.LiveSet;
1499 const auto &PointerToBase = Result.PointerToBase;
1501 // Convert to vector for efficient cross referencing.
1502 SmallVector<Value *, 64> BaseVec, LiveVec;
1503 LiveVec.reserve(LiveSet.size());
1504 BaseVec.reserve(LiveSet.size());
1505 for (Value *L : LiveSet) {
1506 LiveVec.push_back(L);
1507 assert(PointerToBase.count(L));
1508 Value *Base = PointerToBase.find(L)->second;
1509 BaseVec.push_back(Base);
1511 assert(LiveVec.size() == BaseVec.size());
1513 // Do the actual rewriting and delete the old statepoint
1514 makeStatepointExplicitImpl(CS, BaseVec, LiveVec, Result, Replacements);
1517 // Helper function for the relocationViaAlloca.
1519 // It receives iterator to the statepoint gc relocates and emits a store to the
1520 // assigned location (via allocaMap) for the each one of them. It adds the
1521 // visited values into the visitedLiveValues set, which we will later use them
1522 // for sanity checking.
1524 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1525 DenseMap<Value *, Value *> &AllocaMap,
1526 DenseSet<Value *> &VisitedLiveValues) {
1528 for (User *U : GCRelocs) {
1529 GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U);
1533 Value *OriginalValue = Relocate->getDerivedPtr();
1534 assert(AllocaMap.count(OriginalValue));
1535 Value *Alloca = AllocaMap[OriginalValue];
1537 // Emit store into the related alloca
1538 // All gc_relocates are i8 addrspace(1)* typed, and it must be bitcasted to
1539 // the correct type according to alloca.
1540 assert(Relocate->getNextNode() &&
1541 "Should always have one since it's not a terminator");
1542 IRBuilder<> Builder(Relocate->getNextNode());
1543 Value *CastedRelocatedValue =
1544 Builder.CreateBitCast(Relocate,
1545 cast<AllocaInst>(Alloca)->getAllocatedType(),
1546 suffixed_name_or(Relocate, ".casted", ""));
1548 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1549 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1552 VisitedLiveValues.insert(OriginalValue);
1557 // Helper function for the "relocationViaAlloca". Similar to the
1558 // "insertRelocationStores" but works for rematerialized values.
1559 static void insertRematerializationStores(
1560 const RematerializedValueMapTy &RematerializedValues,
1561 DenseMap<Value *, Value *> &AllocaMap,
1562 DenseSet<Value *> &VisitedLiveValues) {
1564 for (auto RematerializedValuePair: RematerializedValues) {
1565 Instruction *RematerializedValue = RematerializedValuePair.first;
1566 Value *OriginalValue = RematerializedValuePair.second;
1568 assert(AllocaMap.count(OriginalValue) &&
1569 "Can not find alloca for rematerialized value");
1570 Value *Alloca = AllocaMap[OriginalValue];
1572 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1573 Store->insertAfter(RematerializedValue);
1576 VisitedLiveValues.insert(OriginalValue);
1581 /// Do all the relocation update via allocas and mem2reg
1582 static void relocationViaAlloca(
1583 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1584 ArrayRef<PartiallyConstructedSafepointRecord> Records) {
1586 // record initial number of (static) allocas; we'll check we have the same
1587 // number when we get done.
1588 int InitialAllocaNum = 0;
1589 for (Instruction &I : F.getEntryBlock())
1590 if (isa<AllocaInst>(I))
1594 // TODO-PERF: change data structures, reserve
1595 DenseMap<Value *, Value *> AllocaMap;
1596 SmallVector<AllocaInst *, 200> PromotableAllocas;
1597 // Used later to chack that we have enough allocas to store all values
1598 std::size_t NumRematerializedValues = 0;
1599 PromotableAllocas.reserve(Live.size());
1601 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1602 // "PromotableAllocas"
1603 const DataLayout &DL = F.getParent()->getDataLayout();
1604 auto emitAllocaFor = [&](Value *LiveValue) {
1605 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(),
1606 DL.getAllocaAddrSpace(), "",
1607 F.getEntryBlock().getFirstNonPHI());
1608 AllocaMap[LiveValue] = Alloca;
1609 PromotableAllocas.push_back(Alloca);
1612 // Emit alloca for each live gc pointer
1613 for (Value *V : Live)
1616 // Emit allocas for rematerialized values
1617 for (const auto &Info : Records)
1618 for (auto RematerializedValuePair : Info.RematerializedValues) {
1619 Value *OriginalValue = RematerializedValuePair.second;
1620 if (AllocaMap.count(OriginalValue) != 0)
1623 emitAllocaFor(OriginalValue);
1624 ++NumRematerializedValues;
1627 // The next two loops are part of the same conceptual operation. We need to
1628 // insert a store to the alloca after the original def and at each
1629 // redefinition. We need to insert a load before each use. These are split
1630 // into distinct loops for performance reasons.
1632 // Update gc pointer after each statepoint: either store a relocated value or
1633 // null (if no relocated value was found for this gc pointer and it is not a
1634 // gc_result). This must happen before we update the statepoint with load of
1635 // alloca otherwise we lose the link between statepoint and old def.
1636 for (const auto &Info : Records) {
1637 Value *Statepoint = Info.StatepointToken;
1639 // This will be used for consistency check
1640 DenseSet<Value *> VisitedLiveValues;
1642 // Insert stores for normal statepoint gc relocates
1643 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1645 // In case if it was invoke statepoint
1646 // we will insert stores for exceptional path gc relocates.
1647 if (isa<InvokeInst>(Statepoint)) {
1648 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1652 // Do similar thing with rematerialized values
1653 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1656 if (ClobberNonLive) {
1657 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1658 // the gc.statepoint. This will turn some subtle GC problems into
1659 // slightly easier to debug SEGVs. Note that on large IR files with
1660 // lots of gc.statepoints this is extremely costly both memory and time
1662 SmallVector<AllocaInst *, 64> ToClobber;
1663 for (auto Pair : AllocaMap) {
1664 Value *Def = Pair.first;
1665 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1667 // This value was relocated
1668 if (VisitedLiveValues.count(Def)) {
1671 ToClobber.push_back(Alloca);
1674 auto InsertClobbersAt = [&](Instruction *IP) {
1675 for (auto *AI : ToClobber) {
1676 auto PT = cast<PointerType>(AI->getAllocatedType());
1677 Constant *CPN = ConstantPointerNull::get(PT);
1678 StoreInst *Store = new StoreInst(CPN, AI);
1679 Store->insertBefore(IP);
1683 // Insert the clobbering stores. These may get intermixed with the
1684 // gc.results and gc.relocates, but that's fine.
1685 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1686 InsertClobbersAt(&*II->getNormalDest()->getFirstInsertionPt());
1687 InsertClobbersAt(&*II->getUnwindDest()->getFirstInsertionPt());
1689 InsertClobbersAt(cast<Instruction>(Statepoint)->getNextNode());
1694 // Update use with load allocas and add store for gc_relocated.
1695 for (auto Pair : AllocaMap) {
1696 Value *Def = Pair.first;
1697 Value *Alloca = Pair.second;
1699 // We pre-record the uses of allocas so that we dont have to worry about
1700 // later update that changes the user information..
1702 SmallVector<Instruction *, 20> Uses;
1703 // PERF: trade a linear scan for repeated reallocation
1704 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1705 for (User *U : Def->users()) {
1706 if (!isa<ConstantExpr>(U)) {
1707 // If the def has a ConstantExpr use, then the def is either a
1708 // ConstantExpr use itself or null. In either case
1709 // (recursively in the first, directly in the second), the oop
1710 // it is ultimately dependent on is null and this particular
1711 // use does not need to be fixed up.
1712 Uses.push_back(cast<Instruction>(U));
1716 std::sort(Uses.begin(), Uses.end());
1717 auto Last = std::unique(Uses.begin(), Uses.end());
1718 Uses.erase(Last, Uses.end());
1720 for (Instruction *Use : Uses) {
1721 if (isa<PHINode>(Use)) {
1722 PHINode *Phi = cast<PHINode>(Use);
1723 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1724 if (Def == Phi->getIncomingValue(i)) {
1725 LoadInst *Load = new LoadInst(
1726 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1727 Phi->setIncomingValue(i, Load);
1731 LoadInst *Load = new LoadInst(Alloca, "", Use);
1732 Use->replaceUsesOfWith(Def, Load);
1736 // Emit store for the initial gc value. Store must be inserted after load,
1737 // otherwise store will be in alloca's use list and an extra load will be
1738 // inserted before it.
1739 StoreInst *Store = new StoreInst(Def, Alloca);
1740 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1741 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1742 // InvokeInst is a TerminatorInst so the store need to be inserted
1743 // into its normal destination block.
1744 BasicBlock *NormalDest = Invoke->getNormalDest();
1745 Store->insertBefore(NormalDest->getFirstNonPHI());
1747 assert(!Inst->isTerminator() &&
1748 "The only TerminatorInst that can produce a value is "
1749 "InvokeInst which is handled above.");
1750 Store->insertAfter(Inst);
1753 assert(isa<Argument>(Def));
1754 Store->insertAfter(cast<Instruction>(Alloca));
1758 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1759 "we must have the same allocas with lives");
1760 if (!PromotableAllocas.empty()) {
1761 // Apply mem2reg to promote alloca to SSA
1762 PromoteMemToReg(PromotableAllocas, DT);
1766 for (auto &I : F.getEntryBlock())
1767 if (isa<AllocaInst>(I))
1769 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1773 /// Implement a unique function which doesn't require we sort the input
1774 /// vector. Doing so has the effect of changing the output of a couple of
1775 /// tests in ways which make them less useful in testing fused safepoints.
1776 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1777 SmallSet<T, 8> Seen;
1778 Vec.erase(remove_if(Vec, [&](const T &V) { return !Seen.insert(V).second; }),
1782 /// Insert holders so that each Value is obviously live through the entire
1783 /// lifetime of the call.
1784 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1785 SmallVectorImpl<CallInst *> &Holders) {
1787 // No values to hold live, might as well not insert the empty holder
1790 Module *M = CS.getInstruction()->getModule();
1791 // Use a dummy vararg function to actually hold the values live
1792 Function *Func = cast<Function>(M->getOrInsertFunction(
1793 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1795 // For call safepoints insert dummy calls right after safepoint
1796 Holders.push_back(CallInst::Create(Func, Values, "",
1797 &*++CS.getInstruction()->getIterator()));
1800 // For invoke safepooints insert dummy calls both in normal and
1801 // exceptional destination blocks
1802 auto *II = cast<InvokeInst>(CS.getInstruction());
1803 Holders.push_back(CallInst::Create(
1804 Func, Values, "", &*II->getNormalDest()->getFirstInsertionPt()));
1805 Holders.push_back(CallInst::Create(
1806 Func, Values, "", &*II->getUnwindDest()->getFirstInsertionPt()));
1809 static void findLiveReferences(
1810 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1811 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1812 GCPtrLivenessData OriginalLivenessData;
1813 computeLiveInValues(DT, F, OriginalLivenessData);
1814 for (size_t i = 0; i < records.size(); i++) {
1815 struct PartiallyConstructedSafepointRecord &info = records[i];
1816 analyzeParsePointLiveness(DT, OriginalLivenessData, toUpdate[i], info);
1820 // Helper function for the "rematerializeLiveValues". It walks use chain
1821 // starting from the "CurrentValue" until it reaches the root of the chain, i.e.
1822 // the base or a value it cannot process. Only "simple" values are processed
1823 // (currently it is GEP's and casts). The returned root is examined by the
1824 // callers of findRematerializableChainToBasePointer. Fills "ChainToBase" array
1825 // with all visited values.
1826 static Value* findRematerializableChainToBasePointer(
1827 SmallVectorImpl<Instruction*> &ChainToBase,
1828 Value *CurrentValue) {
1830 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
1831 ChainToBase.push_back(GEP);
1832 return findRematerializableChainToBasePointer(ChainToBase,
1833 GEP->getPointerOperand());
1836 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
1837 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
1840 ChainToBase.push_back(CI);
1841 return findRematerializableChainToBasePointer(ChainToBase,
1845 // We have reached the root of the chain, which is either equal to the base or
1846 // is the first unsupported value along the use chain.
1847 return CurrentValue;
1850 // Helper function for the "rematerializeLiveValues". Compute cost of the use
1851 // chain we are going to rematerialize.
1853 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
1854 TargetTransformInfo &TTI) {
1857 for (Instruction *Instr : Chain) {
1858 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
1859 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
1860 "non noop cast is found during rematerialization");
1862 Type *SrcTy = CI->getOperand(0)->getType();
1863 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy, CI);
1865 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
1866 // Cost of the address calculation
1867 Type *ValTy = GEP->getSourceElementType();
1868 Cost += TTI.getAddressComputationCost(ValTy);
1870 // And cost of the GEP itself
1871 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
1872 // allowed for the external usage)
1873 if (!GEP->hasAllConstantIndices())
1877 llvm_unreachable("unsupported instruciton type during rematerialization");
1884 static bool AreEquivalentPhiNodes(PHINode &OrigRootPhi, PHINode &AlternateRootPhi) {
1886 unsigned PhiNum = OrigRootPhi.getNumIncomingValues();
1887 if (PhiNum != AlternateRootPhi.getNumIncomingValues() ||
1888 OrigRootPhi.getParent() != AlternateRootPhi.getParent())
1890 // Map of incoming values and their corresponding basic blocks of
1892 SmallDenseMap<Value *, BasicBlock *, 8> CurrentIncomingValues;
1893 for (unsigned i = 0; i < PhiNum; i++)
1894 CurrentIncomingValues[OrigRootPhi.getIncomingValue(i)] =
1895 OrigRootPhi.getIncomingBlock(i);
1897 // Both current and base PHIs should have same incoming values and
1898 // the same basic blocks corresponding to the incoming values.
1899 for (unsigned i = 0; i < PhiNum; i++) {
1901 CurrentIncomingValues.find(AlternateRootPhi.getIncomingValue(i));
1902 if (CIVI == CurrentIncomingValues.end())
1904 BasicBlock *CurrentIncomingBB = CIVI->second;
1905 if (CurrentIncomingBB != AlternateRootPhi.getIncomingBlock(i))
1912 // From the statepoint live set pick values that are cheaper to recompute then
1913 // to relocate. Remove this values from the live set, rematerialize them after
1914 // statepoint and record them in "Info" structure. Note that similar to
1915 // relocated values we don't do any user adjustments here.
1916 static void rematerializeLiveValues(CallSite CS,
1917 PartiallyConstructedSafepointRecord &Info,
1918 TargetTransformInfo &TTI) {
1919 const unsigned int ChainLengthThreshold = 10;
1921 // Record values we are going to delete from this statepoint live set.
1922 // We can not di this in following loop due to iterator invalidation.
1923 SmallVector<Value *, 32> LiveValuesToBeDeleted;
1925 for (Value *LiveValue: Info.LiveSet) {
1926 // For each live pointer find it's defining chain
1927 SmallVector<Instruction *, 3> ChainToBase;
1928 assert(Info.PointerToBase.count(LiveValue));
1929 Value *RootOfChain =
1930 findRematerializableChainToBasePointer(ChainToBase,
1933 // Nothing to do, or chain is too long
1934 if ( ChainToBase.size() == 0 ||
1935 ChainToBase.size() > ChainLengthThreshold)
1938 // Handle the scenario where the RootOfChain is not equal to the
1939 // Base Value, but they are essentially the same phi values.
1940 if (RootOfChain != Info.PointerToBase[LiveValue]) {
1941 PHINode *OrigRootPhi = dyn_cast<PHINode>(RootOfChain);
1942 PHINode *AlternateRootPhi = dyn_cast<PHINode>(Info.PointerToBase[LiveValue]);
1943 if (!OrigRootPhi || !AlternateRootPhi)
1945 // PHI nodes that have the same incoming values, and belonging to the same
1946 // basic blocks are essentially the same SSA value. When the original phi
1947 // has incoming values with different base pointers, the original phi is
1948 // marked as conflict, and an additional `AlternateRootPhi` with the same
1949 // incoming values get generated by the findBasePointer function. We need
1950 // to identify the newly generated AlternateRootPhi (.base version of phi)
1951 // and RootOfChain (the original phi node itself) are the same, so that we
1952 // can rematerialize the gep and casts. This is a workaround for the
1953 // deficieny in the findBasePointer algorithm.
1954 if (!AreEquivalentPhiNodes(*OrigRootPhi, *AlternateRootPhi))
1956 // Now that the phi nodes are proved to be the same, assert that
1957 // findBasePointer's newly generated AlternateRootPhi is present in the
1958 // liveset of the call.
1959 assert(Info.LiveSet.count(AlternateRootPhi));
1961 // Compute cost of this chain
1962 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
1963 // TODO: We can also account for cases when we will be able to remove some
1964 // of the rematerialized values by later optimization passes. I.e if
1965 // we rematerialized several intersecting chains. Or if original values
1966 // don't have any uses besides this statepoint.
1968 // For invokes we need to rematerialize each chain twice - for normal and
1969 // for unwind basic blocks. Model this by multiplying cost by two.
1970 if (CS.isInvoke()) {
1973 // If it's too expensive - skip it
1974 if (Cost >= RematerializationThreshold)
1977 // Remove value from the live set
1978 LiveValuesToBeDeleted.push_back(LiveValue);
1980 // Clone instructions and record them inside "Info" structure
1982 // Walk backwards to visit top-most instructions first
1983 std::reverse(ChainToBase.begin(), ChainToBase.end());
1985 // Utility function which clones all instructions from "ChainToBase"
1986 // and inserts them before "InsertBefore". Returns rematerialized value
1987 // which should be used after statepoint.
1988 auto rematerializeChain = [&ChainToBase](
1989 Instruction *InsertBefore, Value *RootOfChain, Value *AlternateLiveBase) {
1990 Instruction *LastClonedValue = nullptr;
1991 Instruction *LastValue = nullptr;
1992 for (Instruction *Instr: ChainToBase) {
1993 // Only GEP's and casts are suported as we need to be careful to not
1994 // introduce any new uses of pointers not in the liveset.
1995 // Note that it's fine to introduce new uses of pointers which were
1996 // otherwise not used after this statepoint.
1997 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
1999 Instruction *ClonedValue = Instr->clone();
2000 ClonedValue->insertBefore(InsertBefore);
2001 ClonedValue->setName(Instr->getName() + ".remat");
2003 // If it is not first instruction in the chain then it uses previously
2004 // cloned value. We should update it to use cloned value.
2005 if (LastClonedValue) {
2007 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2009 for (auto OpValue : ClonedValue->operand_values()) {
2010 // Assert that cloned instruction does not use any instructions from
2011 // this chain other than LastClonedValue
2012 assert(!is_contained(ChainToBase, OpValue) &&
2013 "incorrect use in rematerialization chain");
2014 // Assert that the cloned instruction does not use the RootOfChain
2015 // or the AlternateLiveBase.
2016 assert(OpValue != RootOfChain && OpValue != AlternateLiveBase);
2020 // For the first instruction, replace the use of unrelocated base i.e.
2021 // RootOfChain/OrigRootPhi, with the corresponding PHI present in the
2022 // live set. They have been proved to be the same PHI nodes. Note
2023 // that the *only* use of the RootOfChain in the ChainToBase list is
2024 // the first Value in the list.
2025 if (RootOfChain != AlternateLiveBase)
2026 ClonedValue->replaceUsesOfWith(RootOfChain, AlternateLiveBase);
2029 LastClonedValue = ClonedValue;
2032 assert(LastClonedValue);
2033 return LastClonedValue;
2036 // Different cases for calls and invokes. For invokes we need to clone
2037 // instructions both on normal and unwind path.
2039 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2040 assert(InsertBefore);
2041 Instruction *RematerializedValue = rematerializeChain(
2042 InsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2043 Info.RematerializedValues[RematerializedValue] = LiveValue;
2045 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2047 Instruction *NormalInsertBefore =
2048 &*Invoke->getNormalDest()->getFirstInsertionPt();
2049 Instruction *UnwindInsertBefore =
2050 &*Invoke->getUnwindDest()->getFirstInsertionPt();
2052 Instruction *NormalRematerializedValue = rematerializeChain(
2053 NormalInsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2054 Instruction *UnwindRematerializedValue = rematerializeChain(
2055 UnwindInsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2057 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2058 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2062 // Remove rematerializaed values from the live set
2063 for (auto LiveValue: LiveValuesToBeDeleted) {
2064 Info.LiveSet.remove(LiveValue);
2068 static bool insertParsePoints(Function &F, DominatorTree &DT,
2069 TargetTransformInfo &TTI,
2070 SmallVectorImpl<CallSite> &ToUpdate) {
2072 // sanity check the input
2073 std::set<CallSite> Uniqued;
2074 Uniqued.insert(ToUpdate.begin(), ToUpdate.end());
2075 assert(Uniqued.size() == ToUpdate.size() && "no duplicates please!");
2077 for (CallSite CS : ToUpdate)
2078 assert(CS.getInstruction()->getFunction() == &F);
2081 // When inserting gc.relocates for invokes, we need to be able to insert at
2082 // the top of the successor blocks. See the comment on
2083 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2084 // may restructure the CFG.
2085 for (CallSite CS : ToUpdate) {
2088 auto *II = cast<InvokeInst>(CS.getInstruction());
2089 normalizeForInvokeSafepoint(II->getNormalDest(), II->getParent(), DT);
2090 normalizeForInvokeSafepoint(II->getUnwindDest(), II->getParent(), DT);
2093 // A list of dummy calls added to the IR to keep various values obviously
2094 // live in the IR. We'll remove all of these when done.
2095 SmallVector<CallInst *, 64> Holders;
2097 // Insert a dummy call with all of the arguments to the vm_state we'll need
2098 // for the actual safepoint insertion. This ensures reference arguments in
2099 // the deopt argument list are considered live through the safepoint (and
2100 // thus makes sure they get relocated.)
2101 for (CallSite CS : ToUpdate) {
2102 SmallVector<Value *, 64> DeoptValues;
2104 for (Value *Arg : GetDeoptBundleOperands(CS)) {
2105 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2106 "support for FCA unimplemented");
2107 if (isHandledGCPointerType(Arg->getType()))
2108 DeoptValues.push_back(Arg);
2111 insertUseHolderAfter(CS, DeoptValues, Holders);
2114 SmallVector<PartiallyConstructedSafepointRecord, 64> Records(ToUpdate.size());
2116 // A) Identify all gc pointers which are statically live at the given call
2118 findLiveReferences(F, DT, ToUpdate, Records);
2120 // B) Find the base pointers for each live pointer
2121 /* scope for caching */ {
2122 // Cache the 'defining value' relation used in the computation and
2123 // insertion of base phis and selects. This ensures that we don't insert
2124 // large numbers of duplicate base_phis.
2125 DefiningValueMapTy DVCache;
2127 for (size_t i = 0; i < Records.size(); i++) {
2128 PartiallyConstructedSafepointRecord &info = Records[i];
2129 findBasePointers(DT, DVCache, ToUpdate[i], info);
2131 } // end of cache scope
2133 // The base phi insertion logic (for any safepoint) may have inserted new
2134 // instructions which are now live at some safepoint. The simplest such
2137 // phi a <-- will be a new base_phi here
2138 // safepoint 1 <-- that needs to be live here
2142 // We insert some dummy calls after each safepoint to definitely hold live
2143 // the base pointers which were identified for that safepoint. We'll then
2144 // ask liveness for _every_ base inserted to see what is now live. Then we
2145 // remove the dummy calls.
2146 Holders.reserve(Holders.size() + Records.size());
2147 for (size_t i = 0; i < Records.size(); i++) {
2148 PartiallyConstructedSafepointRecord &Info = Records[i];
2150 SmallVector<Value *, 128> Bases;
2151 for (auto Pair : Info.PointerToBase)
2152 Bases.push_back(Pair.second);
2154 insertUseHolderAfter(ToUpdate[i], Bases, Holders);
2157 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2158 // need to rerun liveness. We may *also* have inserted new defs, but that's
2159 // not the key issue.
2160 recomputeLiveInValues(F, DT, ToUpdate, Records);
2162 if (PrintBasePointers) {
2163 for (auto &Info : Records) {
2164 errs() << "Base Pairs: (w/Relocation)\n";
2165 for (auto Pair : Info.PointerToBase) {
2166 errs() << " derived ";
2167 Pair.first->printAsOperand(errs(), false);
2169 Pair.second->printAsOperand(errs(), false);
2175 // It is possible that non-constant live variables have a constant base. For
2176 // example, a GEP with a variable offset from a global. In this case we can
2177 // remove it from the liveset. We already don't add constants to the liveset
2178 // because we assume they won't move at runtime and the GC doesn't need to be
2179 // informed about them. The same reasoning applies if the base is constant.
2180 // Note that the relocation placement code relies on this filtering for
2181 // correctness as it expects the base to be in the liveset, which isn't true
2182 // if the base is constant.
2183 for (auto &Info : Records)
2184 for (auto &BasePair : Info.PointerToBase)
2185 if (isa<Constant>(BasePair.second))
2186 Info.LiveSet.remove(BasePair.first);
2188 for (CallInst *CI : Holders)
2189 CI->eraseFromParent();
2193 // In order to reduce live set of statepoint we might choose to rematerialize
2194 // some values instead of relocating them. This is purely an optimization and
2195 // does not influence correctness.
2196 for (size_t i = 0; i < Records.size(); i++)
2197 rematerializeLiveValues(ToUpdate[i], Records[i], TTI);
2199 // We need this to safely RAUW and delete call or invoke return values that
2200 // may themselves be live over a statepoint. For details, please see usage in
2201 // makeStatepointExplicitImpl.
2202 std::vector<DeferredReplacement> Replacements;
2204 // Now run through and replace the existing statepoints with new ones with
2205 // the live variables listed. We do not yet update uses of the values being
2206 // relocated. We have references to live variables that need to
2207 // survive to the last iteration of this loop. (By construction, the
2208 // previous statepoint can not be a live variable, thus we can and remove
2209 // the old statepoint calls as we go.)
2210 for (size_t i = 0; i < Records.size(); i++)
2211 makeStatepointExplicit(DT, ToUpdate[i], Records[i], Replacements);
2213 ToUpdate.clear(); // prevent accident use of invalid CallSites
2215 for (auto &PR : Replacements)
2218 Replacements.clear();
2220 for (auto &Info : Records) {
2221 // These live sets may contain state Value pointers, since we replaced calls
2222 // with operand bundles with calls wrapped in gc.statepoint, and some of
2223 // those calls may have been def'ing live gc pointers. Clear these out to
2224 // avoid accidentally using them.
2226 // TODO: We should create a separate data structure that does not contain
2227 // these live sets, and migrate to using that data structure from this point
2229 Info.LiveSet.clear();
2230 Info.PointerToBase.clear();
2233 // Do all the fixups of the original live variables to their relocated selves
2234 SmallVector<Value *, 128> Live;
2235 for (size_t i = 0; i < Records.size(); i++) {
2236 PartiallyConstructedSafepointRecord &Info = Records[i];
2238 // We can't simply save the live set from the original insertion. One of
2239 // the live values might be the result of a call which needs a safepoint.
2240 // That Value* no longer exists and we need to use the new gc_result.
2241 // Thankfully, the live set is embedded in the statepoint (and updated), so
2242 // we just grab that.
2243 Statepoint Statepoint(Info.StatepointToken);
2244 Live.insert(Live.end(), Statepoint.gc_args_begin(),
2245 Statepoint.gc_args_end());
2247 // Do some basic sanity checks on our liveness results before performing
2248 // relocation. Relocation can and will turn mistakes in liveness results
2249 // into non-sensical code which is must harder to debug.
2250 // TODO: It would be nice to test consistency as well
2251 assert(DT.isReachableFromEntry(Info.StatepointToken->getParent()) &&
2252 "statepoint must be reachable or liveness is meaningless");
2253 for (Value *V : Statepoint.gc_args()) {
2254 if (!isa<Instruction>(V))
2255 // Non-instruction values trivial dominate all possible uses
2257 auto *LiveInst = cast<Instruction>(V);
2258 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2259 "unreachable values should never be live");
2260 assert(DT.dominates(LiveInst, Info.StatepointToken) &&
2261 "basic SSA liveness expectation violated by liveness analysis");
2265 unique_unsorted(Live);
2269 for (auto *Ptr : Live)
2270 assert(isHandledGCPointerType(Ptr->getType()) &&
2271 "must be a gc pointer type");
2274 relocationViaAlloca(F, DT, Live, Records);
2275 return !Records.empty();
2278 // Handles both return values and arguments for Functions and CallSites.
2279 template <typename AttrHolder>
2280 static void RemoveNonValidAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2283 if (AH.getDereferenceableBytes(Index))
2284 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2285 AH.getDereferenceableBytes(Index)));
2286 if (AH.getDereferenceableOrNullBytes(Index))
2287 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2288 AH.getDereferenceableOrNullBytes(Index)));
2289 if (AH.getAttributes().hasAttribute(Index, Attribute::NoAlias))
2290 R.addAttribute(Attribute::NoAlias);
2293 AH.setAttributes(AH.getAttributes().removeAttributes(Ctx, Index, R));
2297 RewriteStatepointsForGC::stripNonValidAttributesFromPrototype(Function &F) {
2298 LLVMContext &Ctx = F.getContext();
2300 for (Argument &A : F.args())
2301 if (isa<PointerType>(A.getType()))
2302 RemoveNonValidAttrAtIndex(Ctx, F,
2303 A.getArgNo() + AttributeList::FirstArgIndex);
2305 if (isa<PointerType>(F.getReturnType()))
2306 RemoveNonValidAttrAtIndex(Ctx, F, AttributeList::ReturnIndex);
2309 void RewriteStatepointsForGC::stripNonValidAttributesFromBody(Function &F) {
2313 LLVMContext &Ctx = F.getContext();
2314 MDBuilder Builder(Ctx);
2316 for (Instruction &I : instructions(F)) {
2317 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2318 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2319 bool IsImmutableTBAA =
2320 MD->getNumOperands() == 4 &&
2321 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2323 if (!IsImmutableTBAA)
2324 continue; // no work to do, MD_tbaa is already marked mutable
2326 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2327 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2329 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2331 MDNode *MutableTBAA =
2332 Builder.createTBAAStructTagNode(Base, Access, Offset);
2333 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2336 if (CallSite CS = CallSite(&I)) {
2337 for (int i = 0, e = CS.arg_size(); i != e; i++)
2338 if (isa<PointerType>(CS.getArgument(i)->getType()))
2339 RemoveNonValidAttrAtIndex(Ctx, CS, i + AttributeList::FirstArgIndex);
2340 if (isa<PointerType>(CS.getType()))
2341 RemoveNonValidAttrAtIndex(Ctx, CS, AttributeList::ReturnIndex);
2346 /// Returns true if this function should be rewritten by this pass. The main
2347 /// point of this function is as an extension point for custom logic.
2348 static bool shouldRewriteStatepointsIn(Function &F) {
2349 // TODO: This should check the GCStrategy
2351 const auto &FunctionGCName = F.getGC();
2352 const StringRef StatepointExampleName("statepoint-example");
2353 const StringRef CoreCLRName("coreclr");
2354 return (StatepointExampleName == FunctionGCName) ||
2355 (CoreCLRName == FunctionGCName);
2360 void RewriteStatepointsForGC::stripNonValidAttributes(Module &M) {
2362 assert(any_of(M, shouldRewriteStatepointsIn) && "precondition!");
2365 for (Function &F : M)
2366 stripNonValidAttributesFromPrototype(F);
2368 for (Function &F : M)
2369 stripNonValidAttributesFromBody(F);
2372 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2373 // Nothing to do for declarations.
2374 if (F.isDeclaration() || F.empty())
2377 // Policy choice says not to rewrite - the most common reason is that we're
2378 // compiling code without a GCStrategy.
2379 if (!shouldRewriteStatepointsIn(F))
2382 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2383 TargetTransformInfo &TTI =
2384 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2386 auto NeedsRewrite = [](Instruction &I) {
2387 if (ImmutableCallSite CS = ImmutableCallSite(&I))
2388 return !callsGCLeafFunction(CS) && !isStatepoint(CS);
2392 // Gather all the statepoints which need rewritten. Be careful to only
2393 // consider those in reachable code since we need to ask dominance queries
2394 // when rewriting. We'll delete the unreachable ones in a moment.
2395 SmallVector<CallSite, 64> ParsePointNeeded;
2396 bool HasUnreachableStatepoint = false;
2397 for (Instruction &I : instructions(F)) {
2398 // TODO: only the ones with the flag set!
2399 if (NeedsRewrite(I)) {
2400 if (DT.isReachableFromEntry(I.getParent()))
2401 ParsePointNeeded.push_back(CallSite(&I));
2403 HasUnreachableStatepoint = true;
2407 bool MadeChange = false;
2409 // Delete any unreachable statepoints so that we don't have unrewritten
2410 // statepoints surviving this pass. This makes testing easier and the
2411 // resulting IR less confusing to human readers. Rather than be fancy, we
2412 // just reuse a utility function which removes the unreachable blocks.
2413 if (HasUnreachableStatepoint)
2414 MadeChange |= removeUnreachableBlocks(F);
2416 // Return early if no work to do.
2417 if (ParsePointNeeded.empty())
2420 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2421 // These are created by LCSSA. They have the effect of increasing the size
2422 // of liveness sets for no good reason. It may be harder to do this post
2423 // insertion since relocations and base phis can confuse things.
2424 for (BasicBlock &BB : F)
2425 if (BB.getUniquePredecessor()) {
2427 FoldSingleEntryPHINodes(&BB);
2430 // Before we start introducing relocations, we want to tweak the IR a bit to
2431 // avoid unfortunate code generation effects. The main example is that we
2432 // want to try to make sure the comparison feeding a branch is after any
2433 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2434 // values feeding a branch after relocation. This is semantically correct,
2435 // but results in extra register pressure since both the pre-relocation and
2436 // post-relocation copies must be available in registers. For code without
2437 // relocations this is handled elsewhere, but teaching the scheduler to
2438 // reverse the transform we're about to do would be slightly complex.
2439 // Note: This may extend the live range of the inputs to the icmp and thus
2440 // increase the liveset of any statepoint we move over. This is profitable
2441 // as long as all statepoints are in rare blocks. If we had in-register
2442 // lowering for live values this would be a much safer transform.
2443 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2444 if (auto *BI = dyn_cast<BranchInst>(TI))
2445 if (BI->isConditional())
2446 return dyn_cast<Instruction>(BI->getCondition());
2447 // TODO: Extend this to handle switches
2450 for (BasicBlock &BB : F) {
2451 TerminatorInst *TI = BB.getTerminator();
2452 if (auto *Cond = getConditionInst(TI))
2453 // TODO: Handle more than just ICmps here. We should be able to move
2454 // most instructions without side effects or memory access.
2455 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2457 Cond->moveBefore(TI);
2461 MadeChange |= insertParsePoints(F, DT, TTI, ParsePointNeeded);
2465 // liveness computation via standard dataflow
2466 // -------------------------------------------------------------------
2468 // TODO: Consider using bitvectors for liveness, the set of potentially
2469 // interesting values should be small and easy to pre-compute.
2471 /// Compute the live-in set for the location rbegin starting from
2472 /// the live-out set of the basic block
2473 static void computeLiveInValues(BasicBlock::reverse_iterator Begin,
2474 BasicBlock::reverse_iterator End,
2475 SetVector<Value *> &LiveTmp) {
2476 for (auto &I : make_range(Begin, End)) {
2477 // KILL/Def - Remove this definition from LiveIn
2480 // Don't consider *uses* in PHI nodes, we handle their contribution to
2481 // predecessor blocks when we seed the LiveOut sets
2482 if (isa<PHINode>(I))
2485 // USE - Add to the LiveIn set for this instruction
2486 for (Value *V : I.operands()) {
2487 assert(!isUnhandledGCPointerType(V->getType()) &&
2488 "support for FCA unimplemented");
2489 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2490 // The choice to exclude all things constant here is slightly subtle.
2491 // There are two independent reasons:
2492 // - We assume that things which are constant (from LLVM's definition)
2493 // do not move at runtime. For example, the address of a global
2494 // variable is fixed, even though it's contents may not be.
2495 // - Second, we can't disallow arbitrary inttoptr constants even
2496 // if the language frontend does. Optimization passes are free to
2497 // locally exploit facts without respect to global reachability. This
2498 // can create sections of code which are dynamically unreachable and
2499 // contain just about anything. (see constants.ll in tests)
2506 static void computeLiveOutSeed(BasicBlock *BB, SetVector<Value *> &LiveTmp) {
2507 for (BasicBlock *Succ : successors(BB)) {
2508 for (auto &I : *Succ) {
2509 PHINode *PN = dyn_cast<PHINode>(&I);
2513 Value *V = PN->getIncomingValueForBlock(BB);
2514 assert(!isUnhandledGCPointerType(V->getType()) &&
2515 "support for FCA unimplemented");
2516 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V))
2522 static SetVector<Value *> computeKillSet(BasicBlock *BB) {
2523 SetVector<Value *> KillSet;
2524 for (Instruction &I : *BB)
2525 if (isHandledGCPointerType(I.getType()))
2531 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2532 /// sanity check for the liveness computation.
2533 static void checkBasicSSA(DominatorTree &DT, SetVector<Value *> &Live,
2534 TerminatorInst *TI, bool TermOkay = false) {
2535 for (Value *V : Live) {
2536 if (auto *I = dyn_cast<Instruction>(V)) {
2537 // The terminator can be a member of the LiveOut set. LLVM's definition
2538 // of instruction dominance states that V does not dominate itself. As
2539 // such, we need to special case this to allow it.
2540 if (TermOkay && TI == I)
2542 assert(DT.dominates(I, TI) &&
2543 "basic SSA liveness expectation violated by liveness analysis");
2548 /// Check that all the liveness sets used during the computation of liveness
2549 /// obey basic SSA properties. This is useful for finding cases where we miss
2551 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2553 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2554 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2555 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2559 static void computeLiveInValues(DominatorTree &DT, Function &F,
2560 GCPtrLivenessData &Data) {
2561 SmallSetVector<BasicBlock *, 32> Worklist;
2563 // Seed the liveness for each individual block
2564 for (BasicBlock &BB : F) {
2565 Data.KillSet[&BB] = computeKillSet(&BB);
2566 Data.LiveSet[&BB].clear();
2567 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2570 for (Value *Kill : Data.KillSet[&BB])
2571 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2574 Data.LiveOut[&BB] = SetVector<Value *>();
2575 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2576 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2577 Data.LiveIn[&BB].set_union(Data.LiveOut[&BB]);
2578 Data.LiveIn[&BB].set_subtract(Data.KillSet[&BB]);
2579 if (!Data.LiveIn[&BB].empty())
2580 Worklist.insert(pred_begin(&BB), pred_end(&BB));
2583 // Propagate that liveness until stable
2584 while (!Worklist.empty()) {
2585 BasicBlock *BB = Worklist.pop_back_val();
2587 // Compute our new liveout set, then exit early if it hasn't changed despite
2588 // the contribution of our successor.
2589 SetVector<Value *> LiveOut = Data.LiveOut[BB];
2590 const auto OldLiveOutSize = LiveOut.size();
2591 for (BasicBlock *Succ : successors(BB)) {
2592 assert(Data.LiveIn.count(Succ));
2593 LiveOut.set_union(Data.LiveIn[Succ]);
2595 // assert OutLiveOut is a subset of LiveOut
2596 if (OldLiveOutSize == LiveOut.size()) {
2597 // If the sets are the same size, then we didn't actually add anything
2598 // when unioning our successors LiveIn. Thus, the LiveIn of this block
2602 Data.LiveOut[BB] = LiveOut;
2604 // Apply the effects of this basic block
2605 SetVector<Value *> LiveTmp = LiveOut;
2606 LiveTmp.set_union(Data.LiveSet[BB]);
2607 LiveTmp.set_subtract(Data.KillSet[BB]);
2609 assert(Data.LiveIn.count(BB));
2610 const SetVector<Value *> &OldLiveIn = Data.LiveIn[BB];
2611 // assert: OldLiveIn is a subset of LiveTmp
2612 if (OldLiveIn.size() != LiveTmp.size()) {
2613 Data.LiveIn[BB] = LiveTmp;
2614 Worklist.insert(pred_begin(BB), pred_end(BB));
2616 } // while (!Worklist.empty())
2619 // Sanity check our output against SSA properties. This helps catch any
2620 // missing kills during the above iteration.
2621 for (BasicBlock &BB : F)
2622 checkBasicSSA(DT, Data, BB);
2626 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2627 StatepointLiveSetTy &Out) {
2629 BasicBlock *BB = Inst->getParent();
2631 // Note: The copy is intentional and required
2632 assert(Data.LiveOut.count(BB));
2633 SetVector<Value *> LiveOut = Data.LiveOut[BB];
2635 // We want to handle the statepoint itself oddly. It's
2636 // call result is not live (normal), nor are it's arguments
2637 // (unless they're used again later). This adjustment is
2638 // specifically what we need to relocate
2639 computeLiveInValues(BB->rbegin(), ++Inst->getIterator().getReverse(),
2641 LiveOut.remove(Inst);
2642 Out.insert(LiveOut.begin(), LiveOut.end());
2645 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2647 PartiallyConstructedSafepointRecord &Info) {
2648 Instruction *Inst = CS.getInstruction();
2649 StatepointLiveSetTy Updated;
2650 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2653 DenseSet<Value *> Bases;
2654 for (auto KVPair : Info.PointerToBase)
2655 Bases.insert(KVPair.second);
2658 // We may have base pointers which are now live that weren't before. We need
2659 // to update the PointerToBase structure to reflect this.
2660 for (auto V : Updated)
2661 if (Info.PointerToBase.insert({V, V}).second) {
2662 assert(Bases.count(V) && "Can't find base for unexpected live value!");
2667 for (auto V : Updated)
2668 assert(Info.PointerToBase.count(V) &&
2669 "Must be able to find base for live value!");
2672 // Remove any stale base mappings - this can happen since our liveness is
2673 // more precise then the one inherent in the base pointer analysis.
2674 DenseSet<Value *> ToErase;
2675 for (auto KVPair : Info.PointerToBase)
2676 if (!Updated.count(KVPair.first))
2677 ToErase.insert(KVPair.first);
2679 for (auto *V : ToErase)
2680 Info.PointerToBase.erase(V);
2683 for (auto KVPair : Info.PointerToBase)
2684 assert(Updated.count(KVPair.first) && "record for non-live value");
2687 Info.LiveSet = Updated;