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 call/invoke instructions so as to make potential relocations
11 // performed by the garbage collector explicit in the IR.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/ADT/DenseSet.h"
16 #include "llvm/ADT/MapVector.h"
17 #include "llvm/ADT/SetOperations.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/ADT/StringRef.h"
21 #include "llvm/Analysis/CFG.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/IR/BasicBlock.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/IRBuilder.h"
28 #include "llvm/IR/InstIterator.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/Intrinsics.h"
32 #include "llvm/IR/MDBuilder.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/IR/Statepoint.h"
35 #include "llvm/IR/Value.h"
36 #include "llvm/IR/Verifier.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.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 // stripNonValidAttributesAndMetadata 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 stripNonValidAttributesAndMetadata(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 and
109 /// metadata implying 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. stripNonValidAttributesAndMetadata (conservatively) restores
113 /// correctness by erasing all attributes in the module that externally imply
114 /// dereferenceability. Similar reasoning also applies to the noalias
115 /// attributes and metadata. gc.statepoint can touch the entire heap including
117 void stripNonValidAttributesAndMetadata(Module &M);
119 // Helpers for stripNonValidAttributesAndMetadata
120 void stripNonValidAttributesAndMetadataFromBody(Function &F);
121 void stripNonValidAttributesFromPrototype(Function &F);
122 // Certain metadata on instructions are invalid after running RS4GC.
123 // Optimizations that run after RS4GC can incorrectly use this metadata to
124 // optimize functions. We drop such metadata on the instruction.
125 void stripInvalidMetadataFromInstruction(Instruction &I);
129 char RewriteStatepointsForGC::ID = 0;
131 ModulePass *llvm::createRewriteStatepointsForGCPass() {
132 return new RewriteStatepointsForGC();
135 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
136 "Make relocations explicit at statepoints", false, false)
137 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
138 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
139 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
140 "Make relocations explicit at statepoints", false, false)
143 struct GCPtrLivenessData {
144 /// Values defined in this block.
145 MapVector<BasicBlock *, SetVector<Value *>> KillSet;
146 /// Values used in this block (and thus live); does not included values
147 /// killed within this block.
148 MapVector<BasicBlock *, SetVector<Value *>> LiveSet;
150 /// Values live into this basic block (i.e. used by any
151 /// instruction in this basic block or ones reachable from here)
152 MapVector<BasicBlock *, SetVector<Value *>> LiveIn;
154 /// Values live out of this basic block (i.e. live into
155 /// any successor block)
156 MapVector<BasicBlock *, SetVector<Value *>> LiveOut;
159 // The type of the internal cache used inside the findBasePointers family
160 // of functions. From the callers perspective, this is an opaque type and
161 // should not be inspected.
163 // In the actual implementation this caches two relations:
164 // - The base relation itself (i.e. this pointer is based on that one)
165 // - The base defining value relation (i.e. before base_phi insertion)
166 // Generally, after the execution of a full findBasePointer call, only the
167 // base relation will remain. Internally, we add a mixture of the two
168 // types, then update all the second type to the first type
169 typedef MapVector<Value *, Value *> DefiningValueMapTy;
170 typedef SetVector<Value *> StatepointLiveSetTy;
171 typedef MapVector<AssertingVH<Instruction>, AssertingVH<Value>>
172 RematerializedValueMapTy;
174 struct PartiallyConstructedSafepointRecord {
175 /// The set of values known to be live across this safepoint
176 StatepointLiveSetTy LiveSet;
178 /// Mapping from live pointers to a base-defining-value
179 MapVector<Value *, Value *> PointerToBase;
181 /// The *new* gc.statepoint instruction itself. This produces the token
182 /// that normal path gc.relocates and the gc.result are tied to.
183 Instruction *StatepointToken;
185 /// Instruction to which exceptional gc relocates are attached
186 /// Makes it easier to iterate through them during relocationViaAlloca.
187 Instruction *UnwindToken;
189 /// Record live values we are rematerialized instead of relocating.
190 /// They are not included into 'LiveSet' field.
191 /// Maps rematerialized copy to it's original value.
192 RematerializedValueMapTy RematerializedValues;
196 static ArrayRef<Use> GetDeoptBundleOperands(ImmutableCallSite CS) {
197 Optional<OperandBundleUse> DeoptBundle =
198 CS.getOperandBundle(LLVMContext::OB_deopt);
200 if (!DeoptBundle.hasValue()) {
201 assert(AllowStatepointWithNoDeoptInfo &&
202 "Found non-leaf call without deopt info!");
206 return DeoptBundle.getValue().Inputs;
209 /// Compute the live-in set for every basic block in the function
210 static void computeLiveInValues(DominatorTree &DT, Function &F,
211 GCPtrLivenessData &Data);
213 /// Given results from the dataflow liveness computation, find the set of live
214 /// Values at a particular instruction.
215 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
216 StatepointLiveSetTy &out);
218 // TODO: Once we can get to the GCStrategy, this becomes
219 // Optional<bool> isGCManagedPointer(const Type *Ty) const override {
221 static bool isGCPointerType(Type *T) {
222 if (auto *PT = dyn_cast<PointerType>(T))
223 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
224 // GC managed heap. We know that a pointer into this heap needs to be
225 // updated and that no other pointer does.
226 return PT->getAddressSpace() == 1;
230 // Return true if this type is one which a) is a gc pointer or contains a GC
231 // pointer and b) is of a type this code expects to encounter as a live value.
232 // (The insertion code will assert that a type which matches (a) and not (b)
233 // is not encountered.)
234 static bool isHandledGCPointerType(Type *T) {
235 // We fully support gc pointers
236 if (isGCPointerType(T))
238 // We partially support vectors of gc pointers. The code will assert if it
239 // can't handle something.
240 if (auto VT = dyn_cast<VectorType>(T))
241 if (isGCPointerType(VT->getElementType()))
247 /// Returns true if this type contains a gc pointer whether we know how to
248 /// handle that type or not.
249 static bool containsGCPtrType(Type *Ty) {
250 if (isGCPointerType(Ty))
252 if (VectorType *VT = dyn_cast<VectorType>(Ty))
253 return isGCPointerType(VT->getScalarType());
254 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
255 return containsGCPtrType(AT->getElementType());
256 if (StructType *ST = dyn_cast<StructType>(Ty))
257 return any_of(ST->subtypes(), containsGCPtrType);
261 // Returns true if this is a type which a) is a gc pointer or contains a GC
262 // pointer and b) is of a type which the code doesn't expect (i.e. first class
263 // aggregates). Used to trip assertions.
264 static bool isUnhandledGCPointerType(Type *Ty) {
265 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
269 // Return the name of the value suffixed with the provided value, or if the
270 // value didn't have a name, the default value specified.
271 static std::string suffixed_name_or(Value *V, StringRef Suffix,
272 StringRef DefaultName) {
273 return V->hasName() ? (V->getName() + Suffix).str() : DefaultName.str();
276 // Conservatively identifies any definitions which might be live at the
277 // given instruction. The analysis is performed immediately before the
278 // given instruction. Values defined by that instruction are not considered
279 // live. Values used by that instruction are considered live.
281 analyzeParsePointLiveness(DominatorTree &DT,
282 GCPtrLivenessData &OriginalLivenessData, CallSite CS,
283 PartiallyConstructedSafepointRecord &Result) {
284 Instruction *Inst = CS.getInstruction();
286 StatepointLiveSetTy LiveSet;
287 findLiveSetAtInst(Inst, OriginalLivenessData, LiveSet);
290 dbgs() << "Live Variables:\n";
291 for (Value *V : LiveSet)
292 dbgs() << " " << V->getName() << " " << *V << "\n";
294 if (PrintLiveSetSize) {
295 dbgs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
296 dbgs() << "Number live values: " << LiveSet.size() << "\n";
298 Result.LiveSet = LiveSet;
301 static bool isKnownBaseResult(Value *V);
303 /// A single base defining value - An immediate base defining value for an
304 /// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'.
305 /// For instructions which have multiple pointer [vector] inputs or that
306 /// transition between vector and scalar types, there is no immediate base
307 /// defining value. The 'base defining value' for 'Def' is the transitive
308 /// closure of this relation stopping at the first instruction which has no
309 /// immediate base defining value. The b.d.v. might itself be a base pointer,
310 /// but it can also be an arbitrary derived pointer.
311 struct BaseDefiningValueResult {
312 /// Contains the value which is the base defining value.
314 /// True if the base defining value is also known to be an actual base
316 const bool IsKnownBase;
317 BaseDefiningValueResult(Value *BDV, bool IsKnownBase)
318 : BDV(BDV), IsKnownBase(IsKnownBase) {
320 // Check consistency between new and old means of checking whether a BDV is
322 bool MustBeBase = isKnownBaseResult(BDV);
323 assert(!MustBeBase || MustBeBase == IsKnownBase);
329 static BaseDefiningValueResult findBaseDefiningValue(Value *I);
331 /// Return a base defining value for the 'Index' element of the given vector
332 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
333 /// 'I'. As an optimization, this method will try to determine when the
334 /// element is known to already be a base pointer. If this can be established,
335 /// the second value in the returned pair will be true. Note that either a
336 /// vector or a pointer typed value can be returned. For the former, the
337 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
338 /// If the later, the return pointer is a BDV (or possibly a base) for the
339 /// particular element in 'I'.
340 static BaseDefiningValueResult
341 findBaseDefiningValueOfVector(Value *I) {
342 // Each case parallels findBaseDefiningValue below, see that code for
343 // detailed motivation.
345 if (isa<Argument>(I))
346 // An incoming argument to the function is a base pointer
347 return BaseDefiningValueResult(I, true);
349 if (isa<Constant>(I))
350 // Base of constant vector consists only of constant null pointers.
351 // For reasoning see similar case inside 'findBaseDefiningValue' function.
352 return BaseDefiningValueResult(ConstantAggregateZero::get(I->getType()),
355 if (isa<LoadInst>(I))
356 return BaseDefiningValueResult(I, true);
358 if (isa<InsertElementInst>(I))
359 // We don't know whether this vector contains entirely base pointers or
360 // not. To be conservatively correct, we treat it as a BDV and will
361 // duplicate code as needed to construct a parallel vector of bases.
362 return BaseDefiningValueResult(I, false);
364 if (isa<ShuffleVectorInst>(I))
365 // We don't know whether this vector contains entirely base pointers or
366 // not. To be conservatively correct, we treat it as a BDV and will
367 // duplicate code as needed to construct a parallel vector of bases.
368 // TODO: There a number of local optimizations which could be applied here
369 // for particular sufflevector patterns.
370 return BaseDefiningValueResult(I, false);
372 // The behavior of getelementptr instructions is the same for vector and
373 // non-vector data types.
374 if (auto *GEP = dyn_cast<GetElementPtrInst>(I))
375 return findBaseDefiningValue(GEP->getPointerOperand());
377 // A PHI or Select is a base defining value. The outer findBasePointer
378 // algorithm is responsible for constructing a base value for this BDV.
379 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
380 "unknown vector instruction - no base found for vector element");
381 return BaseDefiningValueResult(I, false);
384 /// Helper function for findBasePointer - Will return a value which either a)
385 /// defines the base pointer for the input, b) blocks the simple search
386 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
387 /// from pointer to vector type or back.
388 static BaseDefiningValueResult findBaseDefiningValue(Value *I) {
389 assert(I->getType()->isPtrOrPtrVectorTy() &&
390 "Illegal to ask for the base pointer of a non-pointer type");
392 if (I->getType()->isVectorTy())
393 return findBaseDefiningValueOfVector(I);
395 if (isa<Argument>(I))
396 // An incoming argument to the function is a base pointer
397 // We should have never reached here if this argument isn't an gc value
398 return BaseDefiningValueResult(I, true);
400 if (isa<Constant>(I)) {
401 // We assume that objects with a constant base (e.g. a global) can't move
402 // and don't need to be reported to the collector because they are always
403 // live. Besides global references, all kinds of constants (e.g. undef,
404 // constant expressions, null pointers) can be introduced by the inliner or
405 // the optimizer, especially on dynamically dead paths.
406 // Here we treat all of them as having single null base. By doing this we
407 // trying to avoid problems reporting various conflicts in a form of
408 // "phi (const1, const2)" or "phi (const, regular gc ptr)".
409 // See constant.ll file for relevant test cases.
411 return BaseDefiningValueResult(
412 ConstantPointerNull::get(cast<PointerType>(I->getType())), true);
415 if (CastInst *CI = dyn_cast<CastInst>(I)) {
416 Value *Def = CI->stripPointerCasts();
417 // If stripping pointer casts changes the address space there is an
418 // addrspacecast in between.
419 assert(cast<PointerType>(Def->getType())->getAddressSpace() ==
420 cast<PointerType>(CI->getType())->getAddressSpace() &&
421 "unsupported addrspacecast");
422 // If we find a cast instruction here, it means we've found a cast which is
423 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
424 // handle int->ptr conversion.
425 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
426 return findBaseDefiningValue(Def);
429 if (isa<LoadInst>(I))
430 // The value loaded is an gc base itself
431 return BaseDefiningValueResult(I, true);
434 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
435 // The base of this GEP is the base
436 return findBaseDefiningValue(GEP->getPointerOperand());
438 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
439 switch (II->getIntrinsicID()) {
441 // fall through to general call handling
443 case Intrinsic::experimental_gc_statepoint:
444 llvm_unreachable("statepoints don't produce pointers");
445 case Intrinsic::experimental_gc_relocate: {
446 // Rerunning safepoint insertion after safepoints are already
447 // inserted is not supported. It could probably be made to work,
448 // but why are you doing this? There's no good reason.
449 llvm_unreachable("repeat safepoint insertion is not supported");
451 case Intrinsic::gcroot:
452 // Currently, this mechanism hasn't been extended to work with gcroot.
453 // There's no reason it couldn't be, but I haven't thought about the
454 // implications much.
456 "interaction with the gcroot mechanism is not supported");
459 // We assume that functions in the source language only return base
460 // pointers. This should probably be generalized via attributes to support
461 // both source language and internal functions.
462 if (isa<CallInst>(I) || isa<InvokeInst>(I))
463 return BaseDefiningValueResult(I, true);
465 // TODO: I have absolutely no idea how to implement this part yet. It's not
466 // necessarily hard, I just haven't really looked at it yet.
467 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
469 if (isa<AtomicCmpXchgInst>(I))
470 // A CAS is effectively a atomic store and load combined under a
471 // predicate. From the perspective of base pointers, we just treat it
473 return BaseDefiningValueResult(I, true);
475 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
476 "binary ops which don't apply to pointers");
478 // The aggregate ops. Aggregates can either be in the heap or on the
479 // stack, but in either case, this is simply a field load. As a result,
480 // this is a defining definition of the base just like a load is.
481 if (isa<ExtractValueInst>(I))
482 return BaseDefiningValueResult(I, true);
484 // We should never see an insert vector since that would require we be
485 // tracing back a struct value not a pointer value.
486 assert(!isa<InsertValueInst>(I) &&
487 "Base pointer for a struct is meaningless");
489 // An extractelement produces a base result exactly when it's input does.
490 // We may need to insert a parallel instruction to extract the appropriate
491 // element out of the base vector corresponding to the input. Given this,
492 // it's analogous to the phi and select case even though it's not a merge.
493 if (isa<ExtractElementInst>(I))
494 // Note: There a lot of obvious peephole cases here. This are deliberately
495 // handled after the main base pointer inference algorithm to make writing
496 // test cases to exercise that code easier.
497 return BaseDefiningValueResult(I, false);
499 // The last two cases here don't return a base pointer. Instead, they
500 // return a value which dynamically selects from among several base
501 // derived pointers (each with it's own base potentially). It's the job of
502 // the caller to resolve these.
503 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
504 "missing instruction case in findBaseDefiningValing");
505 return BaseDefiningValueResult(I, false);
508 /// Returns the base defining value for this value.
509 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
510 Value *&Cached = Cache[I];
512 Cached = findBaseDefiningValue(I).BDV;
513 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
514 << Cached->getName() << "\n");
516 assert(Cache[I] != nullptr);
520 /// Return a base pointer for this value if known. Otherwise, return it's
521 /// base defining value.
522 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
523 Value *Def = findBaseDefiningValueCached(I, Cache);
524 auto Found = Cache.find(Def);
525 if (Found != Cache.end()) {
526 // Either a base-of relation, or a self reference. Caller must check.
527 return Found->second;
529 // Only a BDV available
533 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
534 /// is it known to be a base pointer? Or do we need to continue searching.
535 static bool isKnownBaseResult(Value *V) {
536 if (!isa<PHINode>(V) && !isa<SelectInst>(V) &&
537 !isa<ExtractElementInst>(V) && !isa<InsertElementInst>(V) &&
538 !isa<ShuffleVectorInst>(V)) {
539 // no recursion possible
542 if (isa<Instruction>(V) &&
543 cast<Instruction>(V)->getMetadata("is_base_value")) {
544 // This is a previously inserted base phi or select. We know
545 // that this is a base value.
549 // We need to keep searching
554 /// Models the state of a single base defining value in the findBasePointer
555 /// algorithm for determining where a new instruction is needed to propagate
556 /// the base of this BDV.
559 enum Status { Unknown, Base, Conflict };
561 BDVState() : Status(Unknown), BaseValue(nullptr) {}
563 explicit BDVState(Status Status, Value *BaseValue = nullptr)
564 : Status(Status), BaseValue(BaseValue) {
565 assert(Status != Base || BaseValue);
568 explicit BDVState(Value *BaseValue) : Status(Base), BaseValue(BaseValue) {}
570 Status getStatus() const { return Status; }
571 Value *getBaseValue() const { return BaseValue; }
573 bool isBase() const { return getStatus() == Base; }
574 bool isUnknown() const { return getStatus() == Unknown; }
575 bool isConflict() const { return getStatus() == Conflict; }
577 bool operator==(const BDVState &Other) const {
578 return BaseValue == Other.BaseValue && Status == Other.Status;
581 bool operator!=(const BDVState &other) const { return !(*this == other); }
589 void print(raw_ostream &OS) const {
590 switch (getStatus()) {
601 OS << " (" << getBaseValue() << " - "
602 << (getBaseValue() ? getBaseValue()->getName() : "nullptr") << "): ";
607 AssertingVH<Value> BaseValue; // Non-null only if Status == Base.
612 static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
618 static BDVState meetBDVStateImpl(const BDVState &LHS, const BDVState &RHS) {
619 switch (LHS.getStatus()) {
620 case BDVState::Unknown:
624 assert(LHS.getBaseValue() && "can't be null");
629 if (LHS.getBaseValue() == RHS.getBaseValue()) {
630 assert(LHS == RHS && "equality broken!");
633 return BDVState(BDVState::Conflict);
635 assert(RHS.isConflict() && "only three states!");
636 return BDVState(BDVState::Conflict);
638 case BDVState::Conflict:
641 llvm_unreachable("only three states!");
644 // Values of type BDVState form a lattice, and this function implements the meet
646 static BDVState meetBDVState(const BDVState &LHS, const BDVState &RHS) {
647 BDVState Result = meetBDVStateImpl(LHS, RHS);
648 assert(Result == meetBDVStateImpl(RHS, LHS) &&
649 "Math is wrong: meet does not commute!");
653 /// For a given value or instruction, figure out what base ptr its derived from.
654 /// For gc objects, this is simply itself. On success, returns a value which is
655 /// the base pointer. (This is reliable and can be used for relocation.) On
656 /// failure, returns nullptr.
657 static Value *findBasePointer(Value *I, DefiningValueMapTy &Cache) {
658 Value *Def = findBaseOrBDV(I, Cache);
660 if (isKnownBaseResult(Def))
663 // Here's the rough algorithm:
664 // - For every SSA value, construct a mapping to either an actual base
665 // pointer or a PHI which obscures the base pointer.
666 // - Construct a mapping from PHI to unknown TOP state. Use an
667 // optimistic algorithm to propagate base pointer information. Lattice
672 // When algorithm terminates, all PHIs will either have a single concrete
673 // base or be in a conflict state.
674 // - For every conflict, insert a dummy PHI node without arguments. Add
675 // these to the base[Instruction] = BasePtr mapping. For every
676 // non-conflict, add the actual base.
677 // - For every conflict, add arguments for the base[a] of each input
680 // Note: A simpler form of this would be to add the conflict form of all
681 // PHIs without running the optimistic algorithm. This would be
682 // analogous to pessimistic data flow and would likely lead to an
683 // overall worse solution.
686 auto isExpectedBDVType = [](Value *BDV) {
687 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) ||
688 isa<ExtractElementInst>(BDV) || isa<InsertElementInst>(BDV) ||
689 isa<ShuffleVectorInst>(BDV);
693 // Once populated, will contain a mapping from each potentially non-base BDV
694 // to a lattice value (described above) which corresponds to that BDV.
695 // We use the order of insertion (DFS over the def/use graph) to provide a
696 // stable deterministic ordering for visiting DenseMaps (which are unordered)
697 // below. This is important for deterministic compilation.
698 MapVector<Value *, BDVState> States;
700 // Recursively fill in all base defining values reachable from the initial
701 // one for which we don't already know a definite base value for
703 SmallVector<Value*, 16> Worklist;
704 Worklist.push_back(Def);
705 States.insert({Def, BDVState()});
706 while (!Worklist.empty()) {
707 Value *Current = Worklist.pop_back_val();
708 assert(!isKnownBaseResult(Current) && "why did it get added?");
710 auto visitIncomingValue = [&](Value *InVal) {
711 Value *Base = findBaseOrBDV(InVal, Cache);
712 if (isKnownBaseResult(Base))
713 // Known bases won't need new instructions introduced and can be
716 assert(isExpectedBDVType(Base) && "the only non-base values "
717 "we see should be base defining values");
718 if (States.insert(std::make_pair(Base, BDVState())).second)
719 Worklist.push_back(Base);
721 if (PHINode *PN = dyn_cast<PHINode>(Current)) {
722 for (Value *InVal : PN->incoming_values())
723 visitIncomingValue(InVal);
724 } else if (SelectInst *SI = dyn_cast<SelectInst>(Current)) {
725 visitIncomingValue(SI->getTrueValue());
726 visitIncomingValue(SI->getFalseValue());
727 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
728 visitIncomingValue(EE->getVectorOperand());
729 } else if (auto *IE = dyn_cast<InsertElementInst>(Current)) {
730 visitIncomingValue(IE->getOperand(0)); // vector operand
731 visitIncomingValue(IE->getOperand(1)); // scalar operand
732 } else if (auto *SV = dyn_cast<ShuffleVectorInst>(Current)) {
733 visitIncomingValue(SV->getOperand(0));
734 visitIncomingValue(SV->getOperand(1));
737 llvm_unreachable("Unimplemented instruction case");
743 DEBUG(dbgs() << "States after initialization:\n");
744 for (auto Pair : States) {
745 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
749 // Return a phi state for a base defining value. We'll generate a new
750 // base state for known bases and expect to find a cached state otherwise.
751 auto getStateForBDV = [&](Value *baseValue) {
752 if (isKnownBaseResult(baseValue))
753 return BDVState(baseValue);
754 auto I = States.find(baseValue);
755 assert(I != States.end() && "lookup failed!");
759 bool Progress = true;
762 const size_t OldSize = States.size();
765 // We're only changing values in this loop, thus safe to keep iterators.
766 // Since this is computing a fixed point, the order of visit does not
767 // effect the result. TODO: We could use a worklist here and make this run
769 for (auto Pair : States) {
770 Value *BDV = Pair.first;
771 assert(!isKnownBaseResult(BDV) && "why did it get added?");
773 // Given an input value for the current instruction, return a BDVState
774 // instance which represents the BDV of that value.
775 auto getStateForInput = [&](Value *V) mutable {
776 Value *BDV = findBaseOrBDV(V, Cache);
777 return getStateForBDV(BDV);
781 if (SelectInst *SI = dyn_cast<SelectInst>(BDV)) {
782 NewState = meetBDVState(NewState, getStateForInput(SI->getTrueValue()));
784 meetBDVState(NewState, getStateForInput(SI->getFalseValue()));
785 } else if (PHINode *PN = dyn_cast<PHINode>(BDV)) {
786 for (Value *Val : PN->incoming_values())
787 NewState = meetBDVState(NewState, getStateForInput(Val));
788 } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) {
789 // The 'meet' for an extractelement is slightly trivial, but it's still
790 // useful in that it drives us to conflict if our input is.
792 meetBDVState(NewState, getStateForInput(EE->getVectorOperand()));
793 } else if (auto *IE = dyn_cast<InsertElementInst>(BDV)){
794 // Given there's a inherent type mismatch between the operands, will
795 // *always* produce Conflict.
796 NewState = meetBDVState(NewState, getStateForInput(IE->getOperand(0)));
797 NewState = meetBDVState(NewState, getStateForInput(IE->getOperand(1)));
799 // The only instance this does not return a Conflict is when both the
800 // vector operands are the same vector.
801 auto *SV = cast<ShuffleVectorInst>(BDV);
802 NewState = meetBDVState(NewState, getStateForInput(SV->getOperand(0)));
803 NewState = meetBDVState(NewState, getStateForInput(SV->getOperand(1)));
806 BDVState OldState = States[BDV];
807 if (OldState != NewState) {
809 States[BDV] = NewState;
813 assert(OldSize == States.size() &&
814 "fixed point shouldn't be adding any new nodes to state");
818 DEBUG(dbgs() << "States after meet iteration:\n");
819 for (auto Pair : States) {
820 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
824 // Insert Phis for all conflicts
825 // TODO: adjust naming patterns to avoid this order of iteration dependency
826 for (auto Pair : States) {
827 Instruction *I = cast<Instruction>(Pair.first);
828 BDVState State = Pair.second;
829 assert(!isKnownBaseResult(I) && "why did it get added?");
830 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
832 // extractelement instructions are a bit special in that we may need to
833 // insert an extract even when we know an exact base for the instruction.
834 // The problem is that we need to convert from a vector base to a scalar
835 // base for the particular indice we're interested in.
836 if (State.isBase() && isa<ExtractElementInst>(I) &&
837 isa<VectorType>(State.getBaseValue()->getType())) {
838 auto *EE = cast<ExtractElementInst>(I);
839 // TODO: In many cases, the new instruction is just EE itself. We should
840 // exploit this, but can't do it here since it would break the invariant
841 // about the BDV not being known to be a base.
842 auto *BaseInst = ExtractElementInst::Create(
843 State.getBaseValue(), EE->getIndexOperand(), "base_ee", EE);
844 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
845 States[I] = BDVState(BDVState::Base, BaseInst);
848 // Since we're joining a vector and scalar base, they can never be the
849 // same. As a result, we should always see insert element having reached
850 // the conflict state.
851 assert(!isa<InsertElementInst>(I) || State.isConflict());
853 if (!State.isConflict())
856 /// Create and insert a new instruction which will represent the base of
857 /// the given instruction 'I'.
858 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
859 if (isa<PHINode>(I)) {
860 BasicBlock *BB = I->getParent();
861 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
862 assert(NumPreds > 0 && "how did we reach here");
863 std::string Name = suffixed_name_or(I, ".base", "base_phi");
864 return PHINode::Create(I->getType(), NumPreds, Name, I);
865 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
866 // The undef will be replaced later
867 UndefValue *Undef = UndefValue::get(SI->getType());
868 std::string Name = suffixed_name_or(I, ".base", "base_select");
869 return SelectInst::Create(SI->getCondition(), Undef, Undef, Name, SI);
870 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
871 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
872 std::string Name = suffixed_name_or(I, ".base", "base_ee");
873 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
875 } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
876 UndefValue *VecUndef = UndefValue::get(IE->getOperand(0)->getType());
877 UndefValue *ScalarUndef = UndefValue::get(IE->getOperand(1)->getType());
878 std::string Name = suffixed_name_or(I, ".base", "base_ie");
879 return InsertElementInst::Create(VecUndef, ScalarUndef,
880 IE->getOperand(2), Name, IE);
882 auto *SV = cast<ShuffleVectorInst>(I);
883 UndefValue *VecUndef = UndefValue::get(SV->getOperand(0)->getType());
884 std::string Name = suffixed_name_or(I, ".base", "base_sv");
885 return new ShuffleVectorInst(VecUndef, VecUndef, SV->getOperand(2),
889 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
890 // Add metadata marking this as a base value
891 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
892 States[I] = BDVState(BDVState::Conflict, BaseInst);
895 // Returns a instruction which produces the base pointer for a given
896 // instruction. The instruction is assumed to be an input to one of the BDVs
897 // seen in the inference algorithm above. As such, we must either already
898 // know it's base defining value is a base, or have inserted a new
899 // instruction to propagate the base of it's BDV and have entered that newly
900 // introduced instruction into the state table. In either case, we are
901 // assured to be able to determine an instruction which produces it's base
903 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
904 Value *BDV = findBaseOrBDV(Input, Cache);
905 Value *Base = nullptr;
906 if (isKnownBaseResult(BDV)) {
909 // Either conflict or base.
910 assert(States.count(BDV));
911 Base = States[BDV].getBaseValue();
913 assert(Base && "Can't be null");
914 // The cast is needed since base traversal may strip away bitcasts
915 if (Base->getType() != Input->getType() && InsertPt)
916 Base = new BitCastInst(Base, Input->getType(), "cast", InsertPt);
920 // Fixup all the inputs of the new PHIs. Visit order needs to be
921 // deterministic and predictable because we're naming newly created
923 for (auto Pair : States) {
924 Instruction *BDV = cast<Instruction>(Pair.first);
925 BDVState State = Pair.second;
927 assert(!isKnownBaseResult(BDV) && "why did it get added?");
928 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
929 if (!State.isConflict())
932 if (PHINode *BasePHI = dyn_cast<PHINode>(State.getBaseValue())) {
933 PHINode *PN = cast<PHINode>(BDV);
934 unsigned NumPHIValues = PN->getNumIncomingValues();
935 for (unsigned i = 0; i < NumPHIValues; i++) {
936 Value *InVal = PN->getIncomingValue(i);
937 BasicBlock *InBB = PN->getIncomingBlock(i);
939 // If we've already seen InBB, add the same incoming value
940 // we added for it earlier. The IR verifier requires phi
941 // nodes with multiple entries from the same basic block
942 // to have the same incoming value for each of those
943 // entries. If we don't do this check here and basephi
944 // has a different type than base, we'll end up adding two
945 // bitcasts (and hence two distinct values) as incoming
946 // values for the same basic block.
948 int BlockIndex = BasePHI->getBasicBlockIndex(InBB);
949 if (BlockIndex != -1) {
950 Value *OldBase = BasePHI->getIncomingValue(BlockIndex);
951 BasePHI->addIncoming(OldBase, InBB);
954 Value *Base = getBaseForInput(InVal, nullptr);
955 // In essence this assert states: the only way two values
956 // incoming from the same basic block may be different is by
957 // being different bitcasts of the same value. A cleanup
958 // that remains TODO is changing findBaseOrBDV to return an
959 // llvm::Value of the correct type (and still remain pure).
960 // This will remove the need to add bitcasts.
961 assert(Base->stripPointerCasts() == OldBase->stripPointerCasts() &&
962 "Sanity -- findBaseOrBDV should be pure!");
967 // Find the instruction which produces the base for each input. We may
968 // need to insert a bitcast in the incoming block.
969 // TODO: Need to split critical edges if insertion is needed
970 Value *Base = getBaseForInput(InVal, InBB->getTerminator());
971 BasePHI->addIncoming(Base, InBB);
973 assert(BasePHI->getNumIncomingValues() == NumPHIValues);
974 } else if (SelectInst *BaseSI =
975 dyn_cast<SelectInst>(State.getBaseValue())) {
976 SelectInst *SI = cast<SelectInst>(BDV);
978 // Find the instruction which produces the base for each input.
979 // We may need to insert a bitcast.
980 BaseSI->setTrueValue(getBaseForInput(SI->getTrueValue(), BaseSI));
981 BaseSI->setFalseValue(getBaseForInput(SI->getFalseValue(), BaseSI));
982 } else if (auto *BaseEE =
983 dyn_cast<ExtractElementInst>(State.getBaseValue())) {
984 Value *InVal = cast<ExtractElementInst>(BDV)->getVectorOperand();
985 // Find the instruction which produces the base for each input. We may
986 // need to insert a bitcast.
987 BaseEE->setOperand(0, getBaseForInput(InVal, BaseEE));
988 } else if (auto *BaseIE = dyn_cast<InsertElementInst>(State.getBaseValue())){
989 auto *BdvIE = cast<InsertElementInst>(BDV);
990 auto UpdateOperand = [&](int OperandIdx) {
991 Value *InVal = BdvIE->getOperand(OperandIdx);
992 Value *Base = getBaseForInput(InVal, BaseIE);
993 BaseIE->setOperand(OperandIdx, Base);
995 UpdateOperand(0); // vector operand
996 UpdateOperand(1); // scalar operand
998 auto *BaseSV = cast<ShuffleVectorInst>(State.getBaseValue());
999 auto *BdvSV = cast<ShuffleVectorInst>(BDV);
1000 auto UpdateOperand = [&](int OperandIdx) {
1001 Value *InVal = BdvSV->getOperand(OperandIdx);
1002 Value *Base = getBaseForInput(InVal, BaseSV);
1003 BaseSV->setOperand(OperandIdx, Base);
1005 UpdateOperand(0); // vector operand
1006 UpdateOperand(1); // vector operand
1010 // Cache all of our results so we can cheaply reuse them
1011 // NOTE: This is actually two caches: one of the base defining value
1012 // relation and one of the base pointer relation! FIXME
1013 for (auto Pair : States) {
1014 auto *BDV = Pair.first;
1015 Value *Base = Pair.second.getBaseValue();
1016 assert(BDV && Base);
1017 assert(!isKnownBaseResult(BDV) && "why did it get added?");
1019 DEBUG(dbgs() << "Updating base value cache"
1020 << " for: " << BDV->getName() << " from: "
1021 << (Cache.count(BDV) ? Cache[BDV]->getName().str() : "none")
1022 << " to: " << Base->getName() << "\n");
1024 if (Cache.count(BDV)) {
1025 assert(isKnownBaseResult(Base) &&
1026 "must be something we 'know' is a base pointer");
1027 // Once we transition from the BDV relation being store in the Cache to
1028 // the base relation being stored, it must be stable
1029 assert((!isKnownBaseResult(Cache[BDV]) || Cache[BDV] == Base) &&
1030 "base relation should be stable");
1034 assert(Cache.count(Def));
1038 // For a set of live pointers (base and/or derived), identify the base
1039 // pointer of the object which they are derived from. This routine will
1040 // mutate the IR graph as needed to make the 'base' pointer live at the
1041 // definition site of 'derived'. This ensures that any use of 'derived' can
1042 // also use 'base'. This may involve the insertion of a number of
1043 // additional PHI nodes.
1045 // preconditions: live is a set of pointer type Values
1047 // side effects: may insert PHI nodes into the existing CFG, will preserve
1048 // CFG, will not remove or mutate any existing nodes
1050 // post condition: PointerToBase contains one (derived, base) pair for every
1051 // pointer in live. Note that derived can be equal to base if the original
1052 // pointer was a base pointer.
1054 findBasePointers(const StatepointLiveSetTy &live,
1055 MapVector<Value *, Value *> &PointerToBase,
1056 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1057 for (Value *ptr : live) {
1058 Value *base = findBasePointer(ptr, DVCache);
1059 assert(base && "failed to find base pointer");
1060 PointerToBase[ptr] = base;
1061 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1062 DT->dominates(cast<Instruction>(base)->getParent(),
1063 cast<Instruction>(ptr)->getParent())) &&
1064 "The base we found better dominate the derived pointer");
1068 /// Find the required based pointers (and adjust the live set) for the given
1070 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1072 PartiallyConstructedSafepointRecord &result) {
1073 MapVector<Value *, Value *> PointerToBase;
1074 findBasePointers(result.LiveSet, PointerToBase, &DT, DVCache);
1076 if (PrintBasePointers) {
1077 errs() << "Base Pairs (w/o Relocation):\n";
1078 for (auto &Pair : PointerToBase) {
1079 errs() << " derived ";
1080 Pair.first->printAsOperand(errs(), false);
1082 Pair.second->printAsOperand(errs(), false);
1087 result.PointerToBase = PointerToBase;
1090 /// Given an updated version of the dataflow liveness results, update the
1091 /// liveset and base pointer maps for the call site CS.
1092 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1094 PartiallyConstructedSafepointRecord &result);
1096 static void recomputeLiveInValues(
1097 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1098 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1099 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1100 // again. The old values are still live and will help it stabilize quickly.
1101 GCPtrLivenessData RevisedLivenessData;
1102 computeLiveInValues(DT, F, RevisedLivenessData);
1103 for (size_t i = 0; i < records.size(); i++) {
1104 struct PartiallyConstructedSafepointRecord &info = records[i];
1105 recomputeLiveInValues(RevisedLivenessData, toUpdate[i], info);
1109 // When inserting gc.relocate and gc.result calls, we need to ensure there are
1110 // no uses of the original value / return value between the gc.statepoint and
1111 // the gc.relocate / gc.result call. One case which can arise is a phi node
1112 // starting one of the successor blocks. We also need to be able to insert the
1113 // gc.relocates only on the path which goes through the statepoint. We might
1114 // need to split an edge to make this possible.
1116 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1117 DominatorTree &DT) {
1118 BasicBlock *Ret = BB;
1119 if (!BB->getUniquePredecessor())
1120 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1122 // Now that 'Ret' has unique predecessor we can safely remove all phi nodes
1124 FoldSingleEntryPHINodes(Ret);
1125 assert(!isa<PHINode>(Ret->begin()) &&
1126 "All PHI nodes should have been removed!");
1128 // At this point, we can safely insert a gc.relocate or gc.result as the first
1129 // instruction in Ret if needed.
1133 // Create new attribute set containing only attributes which can be transferred
1134 // from original call to the safepoint.
1135 static AttributeList legalizeCallAttributes(AttributeList AL) {
1139 // Remove the readonly, readnone, and statepoint function attributes.
1140 AttrBuilder FnAttrs = AL.getFnAttributes();
1141 FnAttrs.removeAttribute(Attribute::ReadNone);
1142 FnAttrs.removeAttribute(Attribute::ReadOnly);
1143 for (Attribute A : AL.getFnAttributes()) {
1144 if (isStatepointDirectiveAttr(A))
1148 // Just skip parameter and return attributes for now
1149 LLVMContext &Ctx = AL.getContext();
1150 return AttributeList::get(Ctx, AttributeList::FunctionIndex,
1151 AttributeSet::get(Ctx, FnAttrs));
1154 /// Helper function to place all gc relocates necessary for the given
1157 /// liveVariables - list of variables to be relocated.
1158 /// liveStart - index of the first live variable.
1159 /// basePtrs - base pointers.
1160 /// statepointToken - statepoint instruction to which relocates should be
1162 /// Builder - Llvm IR builder to be used to construct new calls.
1163 static void CreateGCRelocates(ArrayRef<Value *> LiveVariables,
1164 const int LiveStart,
1165 ArrayRef<Value *> BasePtrs,
1166 Instruction *StatepointToken,
1167 IRBuilder<> Builder) {
1168 if (LiveVariables.empty())
1171 auto FindIndex = [](ArrayRef<Value *> LiveVec, Value *Val) {
1172 auto ValIt = find(LiveVec, Val);
1173 assert(ValIt != LiveVec.end() && "Val not found in LiveVec!");
1174 size_t Index = std::distance(LiveVec.begin(), ValIt);
1175 assert(Index < LiveVec.size() && "Bug in std::find?");
1178 Module *M = StatepointToken->getModule();
1180 // All gc_relocate are generated as i8 addrspace(1)* (or a vector type whose
1181 // element type is i8 addrspace(1)*). We originally generated unique
1182 // declarations for each pointer type, but this proved problematic because
1183 // the intrinsic mangling code is incomplete and fragile. Since we're moving
1184 // towards a single unified pointer type anyways, we can just cast everything
1185 // to an i8* of the right address space. A bitcast is added later to convert
1186 // gc_relocate to the actual value's type.
1187 auto getGCRelocateDecl = [&] (Type *Ty) {
1188 assert(isHandledGCPointerType(Ty));
1189 auto AS = Ty->getScalarType()->getPointerAddressSpace();
1190 Type *NewTy = Type::getInt8PtrTy(M->getContext(), AS);
1191 if (auto *VT = dyn_cast<VectorType>(Ty))
1192 NewTy = VectorType::get(NewTy, VT->getNumElements());
1193 return Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate,
1197 // Lazily populated map from input types to the canonicalized form mentioned
1198 // in the comment above. This should probably be cached somewhere more
1200 DenseMap<Type*, Value*> TypeToDeclMap;
1202 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1203 // Generate the gc.relocate call and save the result
1205 Builder.getInt32(LiveStart + FindIndex(LiveVariables, BasePtrs[i]));
1206 Value *LiveIdx = Builder.getInt32(LiveStart + i);
1208 Type *Ty = LiveVariables[i]->getType();
1209 if (!TypeToDeclMap.count(Ty))
1210 TypeToDeclMap[Ty] = getGCRelocateDecl(Ty);
1211 Value *GCRelocateDecl = TypeToDeclMap[Ty];
1213 // only specify a debug name if we can give a useful one
1214 CallInst *Reloc = Builder.CreateCall(
1215 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1216 suffixed_name_or(LiveVariables[i], ".relocated", ""));
1217 // Trick CodeGen into thinking there are lots of free registers at this
1219 Reloc->setCallingConv(CallingConv::Cold);
1225 /// This struct is used to defer RAUWs and `eraseFromParent` s. Using this
1226 /// avoids having to worry about keeping around dangling pointers to Values.
1227 class DeferredReplacement {
1228 AssertingVH<Instruction> Old;
1229 AssertingVH<Instruction> New;
1230 bool IsDeoptimize = false;
1232 DeferredReplacement() {}
1235 static DeferredReplacement createRAUW(Instruction *Old, Instruction *New) {
1236 assert(Old != New && Old && New &&
1237 "Cannot RAUW equal values or to / from null!");
1239 DeferredReplacement D;
1245 static DeferredReplacement createDelete(Instruction *ToErase) {
1246 DeferredReplacement D;
1251 static DeferredReplacement createDeoptimizeReplacement(Instruction *Old) {
1253 auto *F = cast<CallInst>(Old)->getCalledFunction();
1254 assert(F && F->getIntrinsicID() == Intrinsic::experimental_deoptimize &&
1255 "Only way to construct a deoptimize deferred replacement");
1257 DeferredReplacement D;
1259 D.IsDeoptimize = true;
1263 /// Does the task represented by this instance.
1264 void doReplacement() {
1265 Instruction *OldI = Old;
1266 Instruction *NewI = New;
1268 assert(OldI != NewI && "Disallowed at construction?!");
1269 assert((!IsDeoptimize || !New) &&
1270 "Deoptimize instrinsics are not replaced!");
1276 OldI->replaceAllUsesWith(NewI);
1279 // Note: we've inserted instructions, so the call to llvm.deoptimize may
1280 // not necessarilly be followed by the matching return.
1281 auto *RI = cast<ReturnInst>(OldI->getParent()->getTerminator());
1282 new UnreachableInst(RI->getContext(), RI);
1283 RI->eraseFromParent();
1286 OldI->eraseFromParent();
1291 static StringRef getDeoptLowering(CallSite CS) {
1292 const char *DeoptLowering = "deopt-lowering";
1293 if (CS.hasFnAttr(DeoptLowering)) {
1294 // FIXME: CallSite has a *really* confusing interface around attributes
1296 const AttributeList &CSAS = CS.getAttributes();
1297 if (CSAS.hasAttribute(AttributeList::FunctionIndex, DeoptLowering))
1298 return CSAS.getAttribute(AttributeList::FunctionIndex, DeoptLowering)
1299 .getValueAsString();
1300 Function *F = CS.getCalledFunction();
1301 assert(F && F->hasFnAttribute(DeoptLowering));
1302 return F->getFnAttribute(DeoptLowering).getValueAsString();
1304 return "live-through";
1309 makeStatepointExplicitImpl(const CallSite CS, /* to replace */
1310 const SmallVectorImpl<Value *> &BasePtrs,
1311 const SmallVectorImpl<Value *> &LiveVariables,
1312 PartiallyConstructedSafepointRecord &Result,
1313 std::vector<DeferredReplacement> &Replacements) {
1314 assert(BasePtrs.size() == LiveVariables.size());
1316 // Then go ahead and use the builder do actually do the inserts. We insert
1317 // immediately before the previous instruction under the assumption that all
1318 // arguments will be available here. We can't insert afterwards since we may
1319 // be replacing a terminator.
1320 Instruction *InsertBefore = CS.getInstruction();
1321 IRBuilder<> Builder(InsertBefore);
1323 ArrayRef<Value *> GCArgs(LiveVariables);
1324 uint64_t StatepointID = StatepointDirectives::DefaultStatepointID;
1325 uint32_t NumPatchBytes = 0;
1326 uint32_t Flags = uint32_t(StatepointFlags::None);
1328 ArrayRef<Use> CallArgs(CS.arg_begin(), CS.arg_end());
1329 ArrayRef<Use> DeoptArgs = GetDeoptBundleOperands(CS);
1330 ArrayRef<Use> TransitionArgs;
1331 if (auto TransitionBundle =
1332 CS.getOperandBundle(LLVMContext::OB_gc_transition)) {
1333 Flags |= uint32_t(StatepointFlags::GCTransition);
1334 TransitionArgs = TransitionBundle->Inputs;
1337 // Instead of lowering calls to @llvm.experimental.deoptimize as normal calls
1338 // with a return value, we lower then as never returning calls to
1339 // __llvm_deoptimize that are followed by unreachable to get better codegen.
1340 bool IsDeoptimize = false;
1342 StatepointDirectives SD =
1343 parseStatepointDirectivesFromAttrs(CS.getAttributes());
1344 if (SD.NumPatchBytes)
1345 NumPatchBytes = *SD.NumPatchBytes;
1346 if (SD.StatepointID)
1347 StatepointID = *SD.StatepointID;
1349 // Pass through the requested lowering if any. The default is live-through.
1350 StringRef DeoptLowering = getDeoptLowering(CS);
1351 if (DeoptLowering.equals("live-in"))
1352 Flags |= uint32_t(StatepointFlags::DeoptLiveIn);
1354 assert(DeoptLowering.equals("live-through") && "Unsupported value!");
1357 Value *CallTarget = CS.getCalledValue();
1358 if (Function *F = dyn_cast<Function>(CallTarget)) {
1359 if (F->getIntrinsicID() == Intrinsic::experimental_deoptimize) {
1360 // Calls to llvm.experimental.deoptimize are lowered to calls to the
1361 // __llvm_deoptimize symbol. We want to resolve this now, since the
1362 // verifier does not allow taking the address of an intrinsic function.
1364 SmallVector<Type *, 8> DomainTy;
1365 for (Value *Arg : CallArgs)
1366 DomainTy.push_back(Arg->getType());
1367 auto *FTy = FunctionType::get(Type::getVoidTy(F->getContext()), DomainTy,
1368 /* isVarArg = */ false);
1370 // Note: CallTarget can be a bitcast instruction of a symbol if there are
1371 // calls to @llvm.experimental.deoptimize with different argument types in
1372 // the same module. This is fine -- we assume the frontend knew what it
1373 // was doing when generating this kind of IR.
1375 F->getParent()->getOrInsertFunction("__llvm_deoptimize", FTy);
1377 IsDeoptimize = true;
1381 // Create the statepoint given all the arguments
1382 Instruction *Token = nullptr;
1384 CallInst *ToReplace = cast<CallInst>(CS.getInstruction());
1385 CallInst *Call = Builder.CreateGCStatepointCall(
1386 StatepointID, NumPatchBytes, CallTarget, Flags, CallArgs,
1387 TransitionArgs, DeoptArgs, GCArgs, "safepoint_token");
1389 Call->setTailCallKind(ToReplace->getTailCallKind());
1390 Call->setCallingConv(ToReplace->getCallingConv());
1392 // Currently we will fail on parameter attributes and on certain
1393 // function attributes. In case if we can handle this set of attributes -
1394 // set up function attrs directly on statepoint and return attrs later for
1395 // gc_result intrinsic.
1396 Call->setAttributes(legalizeCallAttributes(ToReplace->getAttributes()));
1400 // Put the following gc_result and gc_relocate calls immediately after the
1401 // the old call (which we're about to delete)
1402 assert(ToReplace->getNextNode() && "Not a terminator, must have next!");
1403 Builder.SetInsertPoint(ToReplace->getNextNode());
1404 Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc());
1406 InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction());
1408 // Insert the new invoke into the old block. We'll remove the old one in a
1409 // moment at which point this will become the new terminator for the
1411 InvokeInst *Invoke = Builder.CreateGCStatepointInvoke(
1412 StatepointID, NumPatchBytes, CallTarget, ToReplace->getNormalDest(),
1413 ToReplace->getUnwindDest(), Flags, CallArgs, TransitionArgs, DeoptArgs,
1414 GCArgs, "statepoint_token");
1416 Invoke->setCallingConv(ToReplace->getCallingConv());
1418 // Currently we will fail on parameter attributes and on certain
1419 // function attributes. In case if we can handle this set of attributes -
1420 // set up function attrs directly on statepoint and return attrs later for
1421 // gc_result intrinsic.
1422 Invoke->setAttributes(legalizeCallAttributes(ToReplace->getAttributes()));
1426 // Generate gc relocates in exceptional path
1427 BasicBlock *UnwindBlock = ToReplace->getUnwindDest();
1428 assert(!isa<PHINode>(UnwindBlock->begin()) &&
1429 UnwindBlock->getUniquePredecessor() &&
1430 "can't safely insert in this block!");
1432 Builder.SetInsertPoint(&*UnwindBlock->getFirstInsertionPt());
1433 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1435 // Attach exceptional gc relocates to the landingpad.
1436 Instruction *ExceptionalToken = UnwindBlock->getLandingPadInst();
1437 Result.UnwindToken = ExceptionalToken;
1439 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1440 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, ExceptionalToken,
1443 // Generate gc relocates and returns for normal block
1444 BasicBlock *NormalDest = ToReplace->getNormalDest();
1445 assert(!isa<PHINode>(NormalDest->begin()) &&
1446 NormalDest->getUniquePredecessor() &&
1447 "can't safely insert in this block!");
1449 Builder.SetInsertPoint(&*NormalDest->getFirstInsertionPt());
1451 // gc relocates will be generated later as if it were regular call
1454 assert(Token && "Should be set in one of the above branches!");
1457 // If we're wrapping an @llvm.experimental.deoptimize in a statepoint, we
1458 // transform the tail-call like structure to a call to a void function
1459 // followed by unreachable to get better codegen.
1460 Replacements.push_back(
1461 DeferredReplacement::createDeoptimizeReplacement(CS.getInstruction()));
1463 Token->setName("statepoint_token");
1464 if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
1466 CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "";
1467 CallInst *GCResult = Builder.CreateGCResult(Token, CS.getType(), Name);
1468 GCResult->setAttributes(
1469 AttributeList::get(GCResult->getContext(), AttributeList::ReturnIndex,
1470 CS.getAttributes().getRetAttributes()));
1472 // We cannot RAUW or delete CS.getInstruction() because it could be in the
1473 // live set of some other safepoint, in which case that safepoint's
1474 // PartiallyConstructedSafepointRecord will hold a raw pointer to this
1475 // llvm::Instruction. Instead, we defer the replacement and deletion to
1476 // after the live sets have been made explicit in the IR, and we no longer
1477 // have raw pointers to worry about.
1478 Replacements.emplace_back(
1479 DeferredReplacement::createRAUW(CS.getInstruction(), GCResult));
1481 Replacements.emplace_back(
1482 DeferredReplacement::createDelete(CS.getInstruction()));
1486 Result.StatepointToken = Token;
1488 // Second, create a gc.relocate for every live variable
1489 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1490 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, Token, Builder);
1493 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1494 // which make the relocations happening at this safepoint explicit.
1496 // WARNING: Does not do any fixup to adjust users of the original live
1497 // values. That's the callers responsibility.
1499 makeStatepointExplicit(DominatorTree &DT, CallSite CS,
1500 PartiallyConstructedSafepointRecord &Result,
1501 std::vector<DeferredReplacement> &Replacements) {
1502 const auto &LiveSet = Result.LiveSet;
1503 const auto &PointerToBase = Result.PointerToBase;
1505 // Convert to vector for efficient cross referencing.
1506 SmallVector<Value *, 64> BaseVec, LiveVec;
1507 LiveVec.reserve(LiveSet.size());
1508 BaseVec.reserve(LiveSet.size());
1509 for (Value *L : LiveSet) {
1510 LiveVec.push_back(L);
1511 assert(PointerToBase.count(L));
1512 Value *Base = PointerToBase.find(L)->second;
1513 BaseVec.push_back(Base);
1515 assert(LiveVec.size() == BaseVec.size());
1517 // Do the actual rewriting and delete the old statepoint
1518 makeStatepointExplicitImpl(CS, BaseVec, LiveVec, Result, Replacements);
1521 // Helper function for the relocationViaAlloca.
1523 // It receives iterator to the statepoint gc relocates and emits a store to the
1524 // assigned location (via allocaMap) for the each one of them. It adds the
1525 // visited values into the visitedLiveValues set, which we will later use them
1526 // for sanity checking.
1528 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1529 DenseMap<Value *, Value *> &AllocaMap,
1530 DenseSet<Value *> &VisitedLiveValues) {
1532 for (User *U : GCRelocs) {
1533 GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U);
1537 Value *OriginalValue = Relocate->getDerivedPtr();
1538 assert(AllocaMap.count(OriginalValue));
1539 Value *Alloca = AllocaMap[OriginalValue];
1541 // Emit store into the related alloca
1542 // All gc_relocates are i8 addrspace(1)* typed, and it must be bitcasted to
1543 // the correct type according to alloca.
1544 assert(Relocate->getNextNode() &&
1545 "Should always have one since it's not a terminator");
1546 IRBuilder<> Builder(Relocate->getNextNode());
1547 Value *CastedRelocatedValue =
1548 Builder.CreateBitCast(Relocate,
1549 cast<AllocaInst>(Alloca)->getAllocatedType(),
1550 suffixed_name_or(Relocate, ".casted", ""));
1552 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1553 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1556 VisitedLiveValues.insert(OriginalValue);
1561 // Helper function for the "relocationViaAlloca". Similar to the
1562 // "insertRelocationStores" but works for rematerialized values.
1563 static void insertRematerializationStores(
1564 const RematerializedValueMapTy &RematerializedValues,
1565 DenseMap<Value *, Value *> &AllocaMap,
1566 DenseSet<Value *> &VisitedLiveValues) {
1568 for (auto RematerializedValuePair: RematerializedValues) {
1569 Instruction *RematerializedValue = RematerializedValuePair.first;
1570 Value *OriginalValue = RematerializedValuePair.second;
1572 assert(AllocaMap.count(OriginalValue) &&
1573 "Can not find alloca for rematerialized value");
1574 Value *Alloca = AllocaMap[OriginalValue];
1576 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1577 Store->insertAfter(RematerializedValue);
1580 VisitedLiveValues.insert(OriginalValue);
1585 /// Do all the relocation update via allocas and mem2reg
1586 static void relocationViaAlloca(
1587 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1588 ArrayRef<PartiallyConstructedSafepointRecord> Records) {
1590 // record initial number of (static) allocas; we'll check we have the same
1591 // number when we get done.
1592 int InitialAllocaNum = 0;
1593 for (Instruction &I : F.getEntryBlock())
1594 if (isa<AllocaInst>(I))
1598 // TODO-PERF: change data structures, reserve
1599 DenseMap<Value *, Value *> AllocaMap;
1600 SmallVector<AllocaInst *, 200> PromotableAllocas;
1601 // Used later to chack that we have enough allocas to store all values
1602 std::size_t NumRematerializedValues = 0;
1603 PromotableAllocas.reserve(Live.size());
1605 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1606 // "PromotableAllocas"
1607 const DataLayout &DL = F.getParent()->getDataLayout();
1608 auto emitAllocaFor = [&](Value *LiveValue) {
1609 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(),
1610 DL.getAllocaAddrSpace(), "",
1611 F.getEntryBlock().getFirstNonPHI());
1612 AllocaMap[LiveValue] = Alloca;
1613 PromotableAllocas.push_back(Alloca);
1616 // Emit alloca for each live gc pointer
1617 for (Value *V : Live)
1620 // Emit allocas for rematerialized values
1621 for (const auto &Info : Records)
1622 for (auto RematerializedValuePair : Info.RematerializedValues) {
1623 Value *OriginalValue = RematerializedValuePair.second;
1624 if (AllocaMap.count(OriginalValue) != 0)
1627 emitAllocaFor(OriginalValue);
1628 ++NumRematerializedValues;
1631 // The next two loops are part of the same conceptual operation. We need to
1632 // insert a store to the alloca after the original def and at each
1633 // redefinition. We need to insert a load before each use. These are split
1634 // into distinct loops for performance reasons.
1636 // Update gc pointer after each statepoint: either store a relocated value or
1637 // null (if no relocated value was found for this gc pointer and it is not a
1638 // gc_result). This must happen before we update the statepoint with load of
1639 // alloca otherwise we lose the link between statepoint and old def.
1640 for (const auto &Info : Records) {
1641 Value *Statepoint = Info.StatepointToken;
1643 // This will be used for consistency check
1644 DenseSet<Value *> VisitedLiveValues;
1646 // Insert stores for normal statepoint gc relocates
1647 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1649 // In case if it was invoke statepoint
1650 // we will insert stores for exceptional path gc relocates.
1651 if (isa<InvokeInst>(Statepoint)) {
1652 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1656 // Do similar thing with rematerialized values
1657 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1660 if (ClobberNonLive) {
1661 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1662 // the gc.statepoint. This will turn some subtle GC problems into
1663 // slightly easier to debug SEGVs. Note that on large IR files with
1664 // lots of gc.statepoints this is extremely costly both memory and time
1666 SmallVector<AllocaInst *, 64> ToClobber;
1667 for (auto Pair : AllocaMap) {
1668 Value *Def = Pair.first;
1669 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1671 // This value was relocated
1672 if (VisitedLiveValues.count(Def)) {
1675 ToClobber.push_back(Alloca);
1678 auto InsertClobbersAt = [&](Instruction *IP) {
1679 for (auto *AI : ToClobber) {
1680 auto PT = cast<PointerType>(AI->getAllocatedType());
1681 Constant *CPN = ConstantPointerNull::get(PT);
1682 StoreInst *Store = new StoreInst(CPN, AI);
1683 Store->insertBefore(IP);
1687 // Insert the clobbering stores. These may get intermixed with the
1688 // gc.results and gc.relocates, but that's fine.
1689 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1690 InsertClobbersAt(&*II->getNormalDest()->getFirstInsertionPt());
1691 InsertClobbersAt(&*II->getUnwindDest()->getFirstInsertionPt());
1693 InsertClobbersAt(cast<Instruction>(Statepoint)->getNextNode());
1698 // Update use with load allocas and add store for gc_relocated.
1699 for (auto Pair : AllocaMap) {
1700 Value *Def = Pair.first;
1701 Value *Alloca = Pair.second;
1703 // We pre-record the uses of allocas so that we dont have to worry about
1704 // later update that changes the user information..
1706 SmallVector<Instruction *, 20> Uses;
1707 // PERF: trade a linear scan for repeated reallocation
1708 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1709 for (User *U : Def->users()) {
1710 if (!isa<ConstantExpr>(U)) {
1711 // If the def has a ConstantExpr use, then the def is either a
1712 // ConstantExpr use itself or null. In either case
1713 // (recursively in the first, directly in the second), the oop
1714 // it is ultimately dependent on is null and this particular
1715 // use does not need to be fixed up.
1716 Uses.push_back(cast<Instruction>(U));
1720 std::sort(Uses.begin(), Uses.end());
1721 auto Last = std::unique(Uses.begin(), Uses.end());
1722 Uses.erase(Last, Uses.end());
1724 for (Instruction *Use : Uses) {
1725 if (isa<PHINode>(Use)) {
1726 PHINode *Phi = cast<PHINode>(Use);
1727 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1728 if (Def == Phi->getIncomingValue(i)) {
1729 LoadInst *Load = new LoadInst(
1730 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1731 Phi->setIncomingValue(i, Load);
1735 LoadInst *Load = new LoadInst(Alloca, "", Use);
1736 Use->replaceUsesOfWith(Def, Load);
1740 // Emit store for the initial gc value. Store must be inserted after load,
1741 // otherwise store will be in alloca's use list and an extra load will be
1742 // inserted before it.
1743 StoreInst *Store = new StoreInst(Def, Alloca);
1744 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1745 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1746 // InvokeInst is a TerminatorInst so the store need to be inserted
1747 // into its normal destination block.
1748 BasicBlock *NormalDest = Invoke->getNormalDest();
1749 Store->insertBefore(NormalDest->getFirstNonPHI());
1751 assert(!Inst->isTerminator() &&
1752 "The only TerminatorInst that can produce a value is "
1753 "InvokeInst which is handled above.");
1754 Store->insertAfter(Inst);
1757 assert(isa<Argument>(Def));
1758 Store->insertAfter(cast<Instruction>(Alloca));
1762 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1763 "we must have the same allocas with lives");
1764 if (!PromotableAllocas.empty()) {
1765 // Apply mem2reg to promote alloca to SSA
1766 PromoteMemToReg(PromotableAllocas, DT);
1770 for (auto &I : F.getEntryBlock())
1771 if (isa<AllocaInst>(I))
1773 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1777 /// Implement a unique function which doesn't require we sort the input
1778 /// vector. Doing so has the effect of changing the output of a couple of
1779 /// tests in ways which make them less useful in testing fused safepoints.
1780 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1781 SmallSet<T, 8> Seen;
1782 Vec.erase(remove_if(Vec, [&](const T &V) { return !Seen.insert(V).second; }),
1786 /// Insert holders so that each Value is obviously live through the entire
1787 /// lifetime of the call.
1788 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1789 SmallVectorImpl<CallInst *> &Holders) {
1791 // No values to hold live, might as well not insert the empty holder
1794 Module *M = CS.getInstruction()->getModule();
1795 // Use a dummy vararg function to actually hold the values live
1796 Function *Func = cast<Function>(M->getOrInsertFunction(
1797 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1799 // For call safepoints insert dummy calls right after safepoint
1800 Holders.push_back(CallInst::Create(Func, Values, "",
1801 &*++CS.getInstruction()->getIterator()));
1804 // For invoke safepooints insert dummy calls both in normal and
1805 // exceptional destination blocks
1806 auto *II = cast<InvokeInst>(CS.getInstruction());
1807 Holders.push_back(CallInst::Create(
1808 Func, Values, "", &*II->getNormalDest()->getFirstInsertionPt()));
1809 Holders.push_back(CallInst::Create(
1810 Func, Values, "", &*II->getUnwindDest()->getFirstInsertionPt()));
1813 static void findLiveReferences(
1814 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1815 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1816 GCPtrLivenessData OriginalLivenessData;
1817 computeLiveInValues(DT, F, OriginalLivenessData);
1818 for (size_t i = 0; i < records.size(); i++) {
1819 struct PartiallyConstructedSafepointRecord &info = records[i];
1820 analyzeParsePointLiveness(DT, OriginalLivenessData, toUpdate[i], info);
1824 // Helper function for the "rematerializeLiveValues". It walks use chain
1825 // starting from the "CurrentValue" until it reaches the root of the chain, i.e.
1826 // the base or a value it cannot process. Only "simple" values are processed
1827 // (currently it is GEP's and casts). The returned root is examined by the
1828 // callers of findRematerializableChainToBasePointer. Fills "ChainToBase" array
1829 // with all visited values.
1830 static Value* findRematerializableChainToBasePointer(
1831 SmallVectorImpl<Instruction*> &ChainToBase,
1832 Value *CurrentValue) {
1834 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
1835 ChainToBase.push_back(GEP);
1836 return findRematerializableChainToBasePointer(ChainToBase,
1837 GEP->getPointerOperand());
1840 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
1841 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
1844 ChainToBase.push_back(CI);
1845 return findRematerializableChainToBasePointer(ChainToBase,
1849 // We have reached the root of the chain, which is either equal to the base or
1850 // is the first unsupported value along the use chain.
1851 return CurrentValue;
1854 // Helper function for the "rematerializeLiveValues". Compute cost of the use
1855 // chain we are going to rematerialize.
1857 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
1858 TargetTransformInfo &TTI) {
1861 for (Instruction *Instr : Chain) {
1862 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
1863 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
1864 "non noop cast is found during rematerialization");
1866 Type *SrcTy = CI->getOperand(0)->getType();
1867 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy, CI);
1869 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
1870 // Cost of the address calculation
1871 Type *ValTy = GEP->getSourceElementType();
1872 Cost += TTI.getAddressComputationCost(ValTy);
1874 // And cost of the GEP itself
1875 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
1876 // allowed for the external usage)
1877 if (!GEP->hasAllConstantIndices())
1881 llvm_unreachable("unsupported instruciton type during rematerialization");
1888 static bool AreEquivalentPhiNodes(PHINode &OrigRootPhi, PHINode &AlternateRootPhi) {
1890 unsigned PhiNum = OrigRootPhi.getNumIncomingValues();
1891 if (PhiNum != AlternateRootPhi.getNumIncomingValues() ||
1892 OrigRootPhi.getParent() != AlternateRootPhi.getParent())
1894 // Map of incoming values and their corresponding basic blocks of
1896 SmallDenseMap<Value *, BasicBlock *, 8> CurrentIncomingValues;
1897 for (unsigned i = 0; i < PhiNum; i++)
1898 CurrentIncomingValues[OrigRootPhi.getIncomingValue(i)] =
1899 OrigRootPhi.getIncomingBlock(i);
1901 // Both current and base PHIs should have same incoming values and
1902 // the same basic blocks corresponding to the incoming values.
1903 for (unsigned i = 0; i < PhiNum; i++) {
1905 CurrentIncomingValues.find(AlternateRootPhi.getIncomingValue(i));
1906 if (CIVI == CurrentIncomingValues.end())
1908 BasicBlock *CurrentIncomingBB = CIVI->second;
1909 if (CurrentIncomingBB != AlternateRootPhi.getIncomingBlock(i))
1916 // From the statepoint live set pick values that are cheaper to recompute then
1917 // to relocate. Remove this values from the live set, rematerialize them after
1918 // statepoint and record them in "Info" structure. Note that similar to
1919 // relocated values we don't do any user adjustments here.
1920 static void rematerializeLiveValues(CallSite CS,
1921 PartiallyConstructedSafepointRecord &Info,
1922 TargetTransformInfo &TTI) {
1923 const unsigned int ChainLengthThreshold = 10;
1925 // Record values we are going to delete from this statepoint live set.
1926 // We can not di this in following loop due to iterator invalidation.
1927 SmallVector<Value *, 32> LiveValuesToBeDeleted;
1929 for (Value *LiveValue: Info.LiveSet) {
1930 // For each live pointer find it's defining chain
1931 SmallVector<Instruction *, 3> ChainToBase;
1932 assert(Info.PointerToBase.count(LiveValue));
1933 Value *RootOfChain =
1934 findRematerializableChainToBasePointer(ChainToBase,
1937 // Nothing to do, or chain is too long
1938 if ( ChainToBase.size() == 0 ||
1939 ChainToBase.size() > ChainLengthThreshold)
1942 // Handle the scenario where the RootOfChain is not equal to the
1943 // Base Value, but they are essentially the same phi values.
1944 if (RootOfChain != Info.PointerToBase[LiveValue]) {
1945 PHINode *OrigRootPhi = dyn_cast<PHINode>(RootOfChain);
1946 PHINode *AlternateRootPhi = dyn_cast<PHINode>(Info.PointerToBase[LiveValue]);
1947 if (!OrigRootPhi || !AlternateRootPhi)
1949 // PHI nodes that have the same incoming values, and belonging to the same
1950 // basic blocks are essentially the same SSA value. When the original phi
1951 // has incoming values with different base pointers, the original phi is
1952 // marked as conflict, and an additional `AlternateRootPhi` with the same
1953 // incoming values get generated by the findBasePointer function. We need
1954 // to identify the newly generated AlternateRootPhi (.base version of phi)
1955 // and RootOfChain (the original phi node itself) are the same, so that we
1956 // can rematerialize the gep and casts. This is a workaround for the
1957 // deficiency in the findBasePointer algorithm.
1958 if (!AreEquivalentPhiNodes(*OrigRootPhi, *AlternateRootPhi))
1960 // Now that the phi nodes are proved to be the same, assert that
1961 // findBasePointer's newly generated AlternateRootPhi is present in the
1962 // liveset of the call.
1963 assert(Info.LiveSet.count(AlternateRootPhi));
1965 // Compute cost of this chain
1966 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
1967 // TODO: We can also account for cases when we will be able to remove some
1968 // of the rematerialized values by later optimization passes. I.e if
1969 // we rematerialized several intersecting chains. Or if original values
1970 // don't have any uses besides this statepoint.
1972 // For invokes we need to rematerialize each chain twice - for normal and
1973 // for unwind basic blocks. Model this by multiplying cost by two.
1974 if (CS.isInvoke()) {
1977 // If it's too expensive - skip it
1978 if (Cost >= RematerializationThreshold)
1981 // Remove value from the live set
1982 LiveValuesToBeDeleted.push_back(LiveValue);
1984 // Clone instructions and record them inside "Info" structure
1986 // Walk backwards to visit top-most instructions first
1987 std::reverse(ChainToBase.begin(), ChainToBase.end());
1989 // Utility function which clones all instructions from "ChainToBase"
1990 // and inserts them before "InsertBefore". Returns rematerialized value
1991 // which should be used after statepoint.
1992 auto rematerializeChain = [&ChainToBase](
1993 Instruction *InsertBefore, Value *RootOfChain, Value *AlternateLiveBase) {
1994 Instruction *LastClonedValue = nullptr;
1995 Instruction *LastValue = nullptr;
1996 for (Instruction *Instr: ChainToBase) {
1997 // Only GEP's and casts are supported as we need to be careful to not
1998 // introduce any new uses of pointers not in the liveset.
1999 // Note that it's fine to introduce new uses of pointers which were
2000 // otherwise not used after this statepoint.
2001 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2003 Instruction *ClonedValue = Instr->clone();
2004 ClonedValue->insertBefore(InsertBefore);
2005 ClonedValue->setName(Instr->getName() + ".remat");
2007 // If it is not first instruction in the chain then it uses previously
2008 // cloned value. We should update it to use cloned value.
2009 if (LastClonedValue) {
2011 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2013 for (auto OpValue : ClonedValue->operand_values()) {
2014 // Assert that cloned instruction does not use any instructions from
2015 // this chain other than LastClonedValue
2016 assert(!is_contained(ChainToBase, OpValue) &&
2017 "incorrect use in rematerialization chain");
2018 // Assert that the cloned instruction does not use the RootOfChain
2019 // or the AlternateLiveBase.
2020 assert(OpValue != RootOfChain && OpValue != AlternateLiveBase);
2024 // For the first instruction, replace the use of unrelocated base i.e.
2025 // RootOfChain/OrigRootPhi, with the corresponding PHI present in the
2026 // live set. They have been proved to be the same PHI nodes. Note
2027 // that the *only* use of the RootOfChain in the ChainToBase list is
2028 // the first Value in the list.
2029 if (RootOfChain != AlternateLiveBase)
2030 ClonedValue->replaceUsesOfWith(RootOfChain, AlternateLiveBase);
2033 LastClonedValue = ClonedValue;
2036 assert(LastClonedValue);
2037 return LastClonedValue;
2040 // Different cases for calls and invokes. For invokes we need to clone
2041 // instructions both on normal and unwind path.
2043 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2044 assert(InsertBefore);
2045 Instruction *RematerializedValue = rematerializeChain(
2046 InsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2047 Info.RematerializedValues[RematerializedValue] = LiveValue;
2049 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2051 Instruction *NormalInsertBefore =
2052 &*Invoke->getNormalDest()->getFirstInsertionPt();
2053 Instruction *UnwindInsertBefore =
2054 &*Invoke->getUnwindDest()->getFirstInsertionPt();
2056 Instruction *NormalRematerializedValue = rematerializeChain(
2057 NormalInsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2058 Instruction *UnwindRematerializedValue = rematerializeChain(
2059 UnwindInsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2061 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2062 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2066 // Remove rematerializaed values from the live set
2067 for (auto LiveValue: LiveValuesToBeDeleted) {
2068 Info.LiveSet.remove(LiveValue);
2072 static bool insertParsePoints(Function &F, DominatorTree &DT,
2073 TargetTransformInfo &TTI,
2074 SmallVectorImpl<CallSite> &ToUpdate) {
2076 // sanity check the input
2077 std::set<CallSite> Uniqued;
2078 Uniqued.insert(ToUpdate.begin(), ToUpdate.end());
2079 assert(Uniqued.size() == ToUpdate.size() && "no duplicates please!");
2081 for (CallSite CS : ToUpdate)
2082 assert(CS.getInstruction()->getFunction() == &F);
2085 // When inserting gc.relocates for invokes, we need to be able to insert at
2086 // the top of the successor blocks. See the comment on
2087 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2088 // may restructure the CFG.
2089 for (CallSite CS : ToUpdate) {
2092 auto *II = cast<InvokeInst>(CS.getInstruction());
2093 normalizeForInvokeSafepoint(II->getNormalDest(), II->getParent(), DT);
2094 normalizeForInvokeSafepoint(II->getUnwindDest(), II->getParent(), DT);
2097 // A list of dummy calls added to the IR to keep various values obviously
2098 // live in the IR. We'll remove all of these when done.
2099 SmallVector<CallInst *, 64> Holders;
2101 // Insert a dummy call with all of the deopt operands we'll need for the
2102 // actual safepoint insertion as arguments. This ensures reference operands
2103 // in the deopt argument list are considered live through the safepoint (and
2104 // thus makes sure they get relocated.)
2105 for (CallSite CS : ToUpdate) {
2106 SmallVector<Value *, 64> DeoptValues;
2108 for (Value *Arg : GetDeoptBundleOperands(CS)) {
2109 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2110 "support for FCA unimplemented");
2111 if (isHandledGCPointerType(Arg->getType()))
2112 DeoptValues.push_back(Arg);
2115 insertUseHolderAfter(CS, DeoptValues, Holders);
2118 SmallVector<PartiallyConstructedSafepointRecord, 64> Records(ToUpdate.size());
2120 // A) Identify all gc pointers which are statically live at the given call
2122 findLiveReferences(F, DT, ToUpdate, Records);
2124 // B) Find the base pointers for each live pointer
2125 /* scope for caching */ {
2126 // Cache the 'defining value' relation used in the computation and
2127 // insertion of base phis and selects. This ensures that we don't insert
2128 // large numbers of duplicate base_phis.
2129 DefiningValueMapTy DVCache;
2131 for (size_t i = 0; i < Records.size(); i++) {
2132 PartiallyConstructedSafepointRecord &info = Records[i];
2133 findBasePointers(DT, DVCache, ToUpdate[i], info);
2135 } // end of cache scope
2137 // The base phi insertion logic (for any safepoint) may have inserted new
2138 // instructions which are now live at some safepoint. The simplest such
2141 // phi a <-- will be a new base_phi here
2142 // safepoint 1 <-- that needs to be live here
2146 // We insert some dummy calls after each safepoint to definitely hold live
2147 // the base pointers which were identified for that safepoint. We'll then
2148 // ask liveness for _every_ base inserted to see what is now live. Then we
2149 // remove the dummy calls.
2150 Holders.reserve(Holders.size() + Records.size());
2151 for (size_t i = 0; i < Records.size(); i++) {
2152 PartiallyConstructedSafepointRecord &Info = Records[i];
2154 SmallVector<Value *, 128> Bases;
2155 for (auto Pair : Info.PointerToBase)
2156 Bases.push_back(Pair.second);
2158 insertUseHolderAfter(ToUpdate[i], Bases, Holders);
2161 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2162 // need to rerun liveness. We may *also* have inserted new defs, but that's
2163 // not the key issue.
2164 recomputeLiveInValues(F, DT, ToUpdate, Records);
2166 if (PrintBasePointers) {
2167 for (auto &Info : Records) {
2168 errs() << "Base Pairs: (w/Relocation)\n";
2169 for (auto Pair : Info.PointerToBase) {
2170 errs() << " derived ";
2171 Pair.first->printAsOperand(errs(), false);
2173 Pair.second->printAsOperand(errs(), false);
2179 // It is possible that non-constant live variables have a constant base. For
2180 // example, a GEP with a variable offset from a global. In this case we can
2181 // remove it from the liveset. We already don't add constants to the liveset
2182 // because we assume they won't move at runtime and the GC doesn't need to be
2183 // informed about them. The same reasoning applies if the base is constant.
2184 // Note that the relocation placement code relies on this filtering for
2185 // correctness as it expects the base to be in the liveset, which isn't true
2186 // if the base is constant.
2187 for (auto &Info : Records)
2188 for (auto &BasePair : Info.PointerToBase)
2189 if (isa<Constant>(BasePair.second))
2190 Info.LiveSet.remove(BasePair.first);
2192 for (CallInst *CI : Holders)
2193 CI->eraseFromParent();
2197 // In order to reduce live set of statepoint we might choose to rematerialize
2198 // some values instead of relocating them. This is purely an optimization and
2199 // does not influence correctness.
2200 for (size_t i = 0; i < Records.size(); i++)
2201 rematerializeLiveValues(ToUpdate[i], Records[i], TTI);
2203 // We need this to safely RAUW and delete call or invoke return values that
2204 // may themselves be live over a statepoint. For details, please see usage in
2205 // makeStatepointExplicitImpl.
2206 std::vector<DeferredReplacement> Replacements;
2208 // Now run through and replace the existing statepoints with new ones with
2209 // the live variables listed. We do not yet update uses of the values being
2210 // relocated. We have references to live variables that need to
2211 // survive to the last iteration of this loop. (By construction, the
2212 // previous statepoint can not be a live variable, thus we can and remove
2213 // the old statepoint calls as we go.)
2214 for (size_t i = 0; i < Records.size(); i++)
2215 makeStatepointExplicit(DT, ToUpdate[i], Records[i], Replacements);
2217 ToUpdate.clear(); // prevent accident use of invalid CallSites
2219 for (auto &PR : Replacements)
2222 Replacements.clear();
2224 for (auto &Info : Records) {
2225 // These live sets may contain state Value pointers, since we replaced calls
2226 // with operand bundles with calls wrapped in gc.statepoint, and some of
2227 // those calls may have been def'ing live gc pointers. Clear these out to
2228 // avoid accidentally using them.
2230 // TODO: We should create a separate data structure that does not contain
2231 // these live sets, and migrate to using that data structure from this point
2233 Info.LiveSet.clear();
2234 Info.PointerToBase.clear();
2237 // Do all the fixups of the original live variables to their relocated selves
2238 SmallVector<Value *, 128> Live;
2239 for (size_t i = 0; i < Records.size(); i++) {
2240 PartiallyConstructedSafepointRecord &Info = Records[i];
2242 // We can't simply save the live set from the original insertion. One of
2243 // the live values might be the result of a call which needs a safepoint.
2244 // That Value* no longer exists and we need to use the new gc_result.
2245 // Thankfully, the live set is embedded in the statepoint (and updated), so
2246 // we just grab that.
2247 Statepoint Statepoint(Info.StatepointToken);
2248 Live.insert(Live.end(), Statepoint.gc_args_begin(),
2249 Statepoint.gc_args_end());
2251 // Do some basic sanity checks on our liveness results before performing
2252 // relocation. Relocation can and will turn mistakes in liveness results
2253 // into non-sensical code which is must harder to debug.
2254 // TODO: It would be nice to test consistency as well
2255 assert(DT.isReachableFromEntry(Info.StatepointToken->getParent()) &&
2256 "statepoint must be reachable or liveness is meaningless");
2257 for (Value *V : Statepoint.gc_args()) {
2258 if (!isa<Instruction>(V))
2259 // Non-instruction values trivial dominate all possible uses
2261 auto *LiveInst = cast<Instruction>(V);
2262 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2263 "unreachable values should never be live");
2264 assert(DT.dominates(LiveInst, Info.StatepointToken) &&
2265 "basic SSA liveness expectation violated by liveness analysis");
2269 unique_unsorted(Live);
2273 for (auto *Ptr : Live)
2274 assert(isHandledGCPointerType(Ptr->getType()) &&
2275 "must be a gc pointer type");
2278 relocationViaAlloca(F, DT, Live, Records);
2279 return !Records.empty();
2282 // Handles both return values and arguments for Functions and CallSites.
2283 template <typename AttrHolder>
2284 static void RemoveNonValidAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2287 if (AH.getDereferenceableBytes(Index))
2288 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2289 AH.getDereferenceableBytes(Index)));
2290 if (AH.getDereferenceableOrNullBytes(Index))
2291 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2292 AH.getDereferenceableOrNullBytes(Index)));
2293 if (AH.getAttributes().hasAttribute(Index, Attribute::NoAlias))
2294 R.addAttribute(Attribute::NoAlias);
2297 AH.setAttributes(AH.getAttributes().removeAttributes(Ctx, Index, R));
2301 RewriteStatepointsForGC::stripNonValidAttributesFromPrototype(Function &F) {
2302 LLVMContext &Ctx = F.getContext();
2304 for (Argument &A : F.args())
2305 if (isa<PointerType>(A.getType()))
2306 RemoveNonValidAttrAtIndex(Ctx, F,
2307 A.getArgNo() + AttributeList::FirstArgIndex);
2309 if (isa<PointerType>(F.getReturnType()))
2310 RemoveNonValidAttrAtIndex(Ctx, F, AttributeList::ReturnIndex);
2313 void RewriteStatepointsForGC::stripInvalidMetadataFromInstruction(Instruction &I) {
2315 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
2317 // These are the attributes that are still valid on loads and stores after
2319 // The metadata implying dereferenceability and noalias are (conservatively)
2320 // dropped. This is because semantically, after RewriteStatepointsForGC runs,
2321 // all calls to gc.statepoint "free" the entire heap. Also, gc.statepoint can
2322 // touch the entire heap including noalias objects. Note: The reasoning is
2323 // same as stripping the dereferenceability and noalias attributes that are
2324 // analogous to the metadata counterparts.
2325 // We also drop the invariant.load metadata on the load because that metadata
2326 // implies the address operand to the load points to memory that is never
2327 // changed once it became dereferenceable. This is no longer true after RS4GC.
2328 // Similar reasoning applies to invariant.group metadata, which applies to
2329 // loads within a group.
2330 unsigned ValidMetadataAfterRS4GC[] = {LLVMContext::MD_tbaa,
2331 LLVMContext::MD_range,
2332 LLVMContext::MD_alias_scope,
2333 LLVMContext::MD_nontemporal,
2334 LLVMContext::MD_nonnull,
2335 LLVMContext::MD_align,
2336 LLVMContext::MD_type};
2338 // Drops all metadata on the instruction other than ValidMetadataAfterRS4GC.
2339 I.dropUnknownNonDebugMetadata(ValidMetadataAfterRS4GC);
2343 void RewriteStatepointsForGC::stripNonValidAttributesAndMetadataFromBody(Function &F) {
2347 LLVMContext &Ctx = F.getContext();
2348 MDBuilder Builder(Ctx);
2351 for (Instruction &I : instructions(F)) {
2352 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2353 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2354 bool IsImmutableTBAA =
2355 MD->getNumOperands() == 4 &&
2356 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2358 if (!IsImmutableTBAA)
2359 continue; // no work to do, MD_tbaa is already marked mutable
2361 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2362 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2364 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2366 MDNode *MutableTBAA =
2367 Builder.createTBAAStructTagNode(Base, Access, Offset);
2368 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2371 stripInvalidMetadataFromInstruction(I);
2373 if (CallSite CS = CallSite(&I)) {
2374 for (int i = 0, e = CS.arg_size(); i != e; i++)
2375 if (isa<PointerType>(CS.getArgument(i)->getType()))
2376 RemoveNonValidAttrAtIndex(Ctx, CS, i + AttributeList::FirstArgIndex);
2377 if (isa<PointerType>(CS.getType()))
2378 RemoveNonValidAttrAtIndex(Ctx, CS, AttributeList::ReturnIndex);
2383 /// Returns true if this function should be rewritten by this pass. The main
2384 /// point of this function is as an extension point for custom logic.
2385 static bool shouldRewriteStatepointsIn(Function &F) {
2386 // TODO: This should check the GCStrategy
2388 const auto &FunctionGCName = F.getGC();
2389 const StringRef StatepointExampleName("statepoint-example");
2390 const StringRef CoreCLRName("coreclr");
2391 return (StatepointExampleName == FunctionGCName) ||
2392 (CoreCLRName == FunctionGCName);
2397 void RewriteStatepointsForGC::stripNonValidAttributesAndMetadata(Module &M) {
2399 assert(any_of(M, shouldRewriteStatepointsIn) && "precondition!");
2402 for (Function &F : M)
2403 stripNonValidAttributesFromPrototype(F);
2405 for (Function &F : M)
2406 stripNonValidAttributesAndMetadataFromBody(F);
2409 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2410 // Nothing to do for declarations.
2411 if (F.isDeclaration() || F.empty())
2414 // Policy choice says not to rewrite - the most common reason is that we're
2415 // compiling code without a GCStrategy.
2416 if (!shouldRewriteStatepointsIn(F))
2419 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2420 TargetTransformInfo &TTI =
2421 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2423 auto NeedsRewrite = [](Instruction &I) {
2424 if (ImmutableCallSite CS = ImmutableCallSite(&I))
2425 return !callsGCLeafFunction(CS) && !isStatepoint(CS);
2429 // Gather all the statepoints which need rewritten. Be careful to only
2430 // consider those in reachable code since we need to ask dominance queries
2431 // when rewriting. We'll delete the unreachable ones in a moment.
2432 SmallVector<CallSite, 64> ParsePointNeeded;
2433 bool HasUnreachableStatepoint = false;
2434 for (Instruction &I : instructions(F)) {
2435 // TODO: only the ones with the flag set!
2436 if (NeedsRewrite(I)) {
2437 if (DT.isReachableFromEntry(I.getParent()))
2438 ParsePointNeeded.push_back(CallSite(&I));
2440 HasUnreachableStatepoint = true;
2444 bool MadeChange = false;
2446 // Delete any unreachable statepoints so that we don't have unrewritten
2447 // statepoints surviving this pass. This makes testing easier and the
2448 // resulting IR less confusing to human readers. Rather than be fancy, we
2449 // just reuse a utility function which removes the unreachable blocks.
2450 if (HasUnreachableStatepoint)
2451 MadeChange |= removeUnreachableBlocks(F);
2453 // Return early if no work to do.
2454 if (ParsePointNeeded.empty())
2457 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2458 // These are created by LCSSA. They have the effect of increasing the size
2459 // of liveness sets for no good reason. It may be harder to do this post
2460 // insertion since relocations and base phis can confuse things.
2461 for (BasicBlock &BB : F)
2462 if (BB.getUniquePredecessor()) {
2464 FoldSingleEntryPHINodes(&BB);
2467 // Before we start introducing relocations, we want to tweak the IR a bit to
2468 // avoid unfortunate code generation effects. The main example is that we
2469 // want to try to make sure the comparison feeding a branch is after any
2470 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2471 // values feeding a branch after relocation. This is semantically correct,
2472 // but results in extra register pressure since both the pre-relocation and
2473 // post-relocation copies must be available in registers. For code without
2474 // relocations this is handled elsewhere, but teaching the scheduler to
2475 // reverse the transform we're about to do would be slightly complex.
2476 // Note: This may extend the live range of the inputs to the icmp and thus
2477 // increase the liveset of any statepoint we move over. This is profitable
2478 // as long as all statepoints are in rare blocks. If we had in-register
2479 // lowering for live values this would be a much safer transform.
2480 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2481 if (auto *BI = dyn_cast<BranchInst>(TI))
2482 if (BI->isConditional())
2483 return dyn_cast<Instruction>(BI->getCondition());
2484 // TODO: Extend this to handle switches
2487 for (BasicBlock &BB : F) {
2488 TerminatorInst *TI = BB.getTerminator();
2489 if (auto *Cond = getConditionInst(TI))
2490 // TODO: Handle more than just ICmps here. We should be able to move
2491 // most instructions without side effects or memory access.
2492 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2494 Cond->moveBefore(TI);
2498 MadeChange |= insertParsePoints(F, DT, TTI, ParsePointNeeded);
2502 // liveness computation via standard dataflow
2503 // -------------------------------------------------------------------
2505 // TODO: Consider using bitvectors for liveness, the set of potentially
2506 // interesting values should be small and easy to pre-compute.
2508 /// Compute the live-in set for the location rbegin starting from
2509 /// the live-out set of the basic block
2510 static void computeLiveInValues(BasicBlock::reverse_iterator Begin,
2511 BasicBlock::reverse_iterator End,
2512 SetVector<Value *> &LiveTmp) {
2513 for (auto &I : make_range(Begin, End)) {
2514 // KILL/Def - Remove this definition from LiveIn
2517 // Don't consider *uses* in PHI nodes, we handle their contribution to
2518 // predecessor blocks when we seed the LiveOut sets
2519 if (isa<PHINode>(I))
2522 // USE - Add to the LiveIn set for this instruction
2523 for (Value *V : I.operands()) {
2524 assert(!isUnhandledGCPointerType(V->getType()) &&
2525 "support for FCA unimplemented");
2526 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2527 // The choice to exclude all things constant here is slightly subtle.
2528 // There are two independent reasons:
2529 // - We assume that things which are constant (from LLVM's definition)
2530 // do not move at runtime. For example, the address of a global
2531 // variable is fixed, even though it's contents may not be.
2532 // - Second, we can't disallow arbitrary inttoptr constants even
2533 // if the language frontend does. Optimization passes are free to
2534 // locally exploit facts without respect to global reachability. This
2535 // can create sections of code which are dynamically unreachable and
2536 // contain just about anything. (see constants.ll in tests)
2543 static void computeLiveOutSeed(BasicBlock *BB, SetVector<Value *> &LiveTmp) {
2544 for (BasicBlock *Succ : successors(BB)) {
2545 for (auto &I : *Succ) {
2546 PHINode *PN = dyn_cast<PHINode>(&I);
2550 Value *V = PN->getIncomingValueForBlock(BB);
2551 assert(!isUnhandledGCPointerType(V->getType()) &&
2552 "support for FCA unimplemented");
2553 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V))
2559 static SetVector<Value *> computeKillSet(BasicBlock *BB) {
2560 SetVector<Value *> KillSet;
2561 for (Instruction &I : *BB)
2562 if (isHandledGCPointerType(I.getType()))
2568 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2569 /// sanity check for the liveness computation.
2570 static void checkBasicSSA(DominatorTree &DT, SetVector<Value *> &Live,
2571 TerminatorInst *TI, bool TermOkay = false) {
2572 for (Value *V : Live) {
2573 if (auto *I = dyn_cast<Instruction>(V)) {
2574 // The terminator can be a member of the LiveOut set. LLVM's definition
2575 // of instruction dominance states that V does not dominate itself. As
2576 // such, we need to special case this to allow it.
2577 if (TermOkay && TI == I)
2579 assert(DT.dominates(I, TI) &&
2580 "basic SSA liveness expectation violated by liveness analysis");
2585 /// Check that all the liveness sets used during the computation of liveness
2586 /// obey basic SSA properties. This is useful for finding cases where we miss
2588 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2590 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2591 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2592 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2596 static void computeLiveInValues(DominatorTree &DT, Function &F,
2597 GCPtrLivenessData &Data) {
2598 SmallSetVector<BasicBlock *, 32> Worklist;
2600 // Seed the liveness for each individual block
2601 for (BasicBlock &BB : F) {
2602 Data.KillSet[&BB] = computeKillSet(&BB);
2603 Data.LiveSet[&BB].clear();
2604 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2607 for (Value *Kill : Data.KillSet[&BB])
2608 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2611 Data.LiveOut[&BB] = SetVector<Value *>();
2612 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2613 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2614 Data.LiveIn[&BB].set_union(Data.LiveOut[&BB]);
2615 Data.LiveIn[&BB].set_subtract(Data.KillSet[&BB]);
2616 if (!Data.LiveIn[&BB].empty())
2617 Worklist.insert(pred_begin(&BB), pred_end(&BB));
2620 // Propagate that liveness until stable
2621 while (!Worklist.empty()) {
2622 BasicBlock *BB = Worklist.pop_back_val();
2624 // Compute our new liveout set, then exit early if it hasn't changed despite
2625 // the contribution of our successor.
2626 SetVector<Value *> LiveOut = Data.LiveOut[BB];
2627 const auto OldLiveOutSize = LiveOut.size();
2628 for (BasicBlock *Succ : successors(BB)) {
2629 assert(Data.LiveIn.count(Succ));
2630 LiveOut.set_union(Data.LiveIn[Succ]);
2632 // assert OutLiveOut is a subset of LiveOut
2633 if (OldLiveOutSize == LiveOut.size()) {
2634 // If the sets are the same size, then we didn't actually add anything
2635 // when unioning our successors LiveIn. Thus, the LiveIn of this block
2639 Data.LiveOut[BB] = LiveOut;
2641 // Apply the effects of this basic block
2642 SetVector<Value *> LiveTmp = LiveOut;
2643 LiveTmp.set_union(Data.LiveSet[BB]);
2644 LiveTmp.set_subtract(Data.KillSet[BB]);
2646 assert(Data.LiveIn.count(BB));
2647 const SetVector<Value *> &OldLiveIn = Data.LiveIn[BB];
2648 // assert: OldLiveIn is a subset of LiveTmp
2649 if (OldLiveIn.size() != LiveTmp.size()) {
2650 Data.LiveIn[BB] = LiveTmp;
2651 Worklist.insert(pred_begin(BB), pred_end(BB));
2653 } // while (!Worklist.empty())
2656 // Sanity check our output against SSA properties. This helps catch any
2657 // missing kills during the above iteration.
2658 for (BasicBlock &BB : F)
2659 checkBasicSSA(DT, Data, BB);
2663 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2664 StatepointLiveSetTy &Out) {
2666 BasicBlock *BB = Inst->getParent();
2668 // Note: The copy is intentional and required
2669 assert(Data.LiveOut.count(BB));
2670 SetVector<Value *> LiveOut = Data.LiveOut[BB];
2672 // We want to handle the statepoint itself oddly. It's
2673 // call result is not live (normal), nor are it's arguments
2674 // (unless they're used again later). This adjustment is
2675 // specifically what we need to relocate
2676 computeLiveInValues(BB->rbegin(), ++Inst->getIterator().getReverse(),
2678 LiveOut.remove(Inst);
2679 Out.insert(LiveOut.begin(), LiveOut.end());
2682 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2684 PartiallyConstructedSafepointRecord &Info) {
2685 Instruction *Inst = CS.getInstruction();
2686 StatepointLiveSetTy Updated;
2687 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2690 DenseSet<Value *> Bases;
2691 for (auto KVPair : Info.PointerToBase)
2692 Bases.insert(KVPair.second);
2695 // We may have base pointers which are now live that weren't before. We need
2696 // to update the PointerToBase structure to reflect this.
2697 for (auto V : Updated)
2698 if (Info.PointerToBase.insert({V, V}).second) {
2699 assert(Bases.count(V) && "Can't find base for unexpected live value!");
2704 for (auto V : Updated)
2705 assert(Info.PointerToBase.count(V) &&
2706 "Must be able to find base for live value!");
2709 // Remove any stale base mappings - this can happen since our liveness is
2710 // more precise then the one inherent in the base pointer analysis.
2711 DenseSet<Value *> ToErase;
2712 for (auto KVPair : Info.PointerToBase)
2713 if (!Updated.count(KVPair.first))
2714 ToErase.insert(KVPair.first);
2716 for (auto *V : ToErase)
2717 Info.PointerToBase.erase(V);
2720 for (auto KVPair : Info.PointerToBase)
2721 assert(Updated.count(KVPair.first) && "record for non-live value");
2724 Info.LiveSet = Updated;