1 //===- RewriteStatepointsForGC.cpp - Make GC relocations explicit ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Rewrite an existing set of gc.statepoints such that they make potential
11 // relocations performed by the garbage collector explicit in the IR.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Pass.h"
16 #include "llvm/Analysis/CFG.h"
17 #include "llvm/Analysis/TargetTransformInfo.h"
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/ADT/DenseSet.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/StringRef.h"
23 #include "llvm/ADT/MapVector.h"
24 #include "llvm/IR/BasicBlock.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/InstIterator.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/Intrinsics.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/Statepoint.h"
36 #include "llvm/IR/Value.h"
37 #include "llvm/IR/Verifier.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Transforms/Scalar.h"
41 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
42 #include "llvm/Transforms/Utils/Cloning.h"
43 #include "llvm/Transforms/Utils/Local.h"
44 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
46 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
50 // Print the liveset found at the insert location
51 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
53 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden,
55 // Print out the base pointers for debugging
56 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden,
59 // Cost threshold measuring when it is profitable to rematerialize value instead
61 static cl::opt<unsigned>
62 RematerializationThreshold("spp-rematerialization-threshold", cl::Hidden,
65 #ifdef EXPENSIVE_CHECKS
66 static bool ClobberNonLive = true;
68 static bool ClobberNonLive = false;
70 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
71 cl::location(ClobberNonLive),
75 AllowStatepointWithNoDeoptInfo("rs4gc-allow-statepoint-with-no-deopt-info",
76 cl::Hidden, cl::init(true));
79 struct RewriteStatepointsForGC : public ModulePass {
80 static char ID; // Pass identification, replacement for typeid
82 RewriteStatepointsForGC() : ModulePass(ID) {
83 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
85 bool runOnFunction(Function &F);
86 bool runOnModule(Module &M) override {
89 Changed |= runOnFunction(F);
92 // stripNonValidAttributes asserts that shouldRewriteStatepointsIn
93 // returns true for at least one function in the module. Since at least
94 // one function changed, we know that the precondition is satisfied.
95 stripNonValidAttributes(M);
101 void getAnalysisUsage(AnalysisUsage &AU) const override {
102 // We add and rewrite a bunch of instructions, but don't really do much
103 // else. We could in theory preserve a lot more analyses here.
104 AU.addRequired<DominatorTreeWrapperPass>();
105 AU.addRequired<TargetTransformInfoWrapperPass>();
108 /// The IR fed into RewriteStatepointsForGC may have had attributes implying
109 /// dereferenceability that are no longer valid/correct after
110 /// RewriteStatepointsForGC has run. This is because semantically, after
111 /// RewriteStatepointsForGC runs, all calls to gc.statepoint "free" the entire
112 /// heap. stripNonValidAttributes (conservatively) restores correctness
113 /// by erasing all attributes in the module that externally imply
114 /// dereferenceability.
115 /// Similar reasoning also applies to the noalias attributes. gc.statepoint
116 /// can touch the entire heap including noalias objects.
117 void stripNonValidAttributes(Module &M);
119 // Helpers for stripNonValidAttributes
120 void stripNonValidAttributesFromBody(Function &F);
121 void stripNonValidAttributesFromPrototype(Function &F);
125 char RewriteStatepointsForGC::ID = 0;
127 ModulePass *llvm::createRewriteStatepointsForGCPass() {
128 return new RewriteStatepointsForGC();
131 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
132 "Make relocations explicit at statepoints", false, false)
133 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
134 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
135 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
136 "Make relocations explicit at statepoints", false, false)
139 struct GCPtrLivenessData {
140 /// Values defined in this block.
141 MapVector<BasicBlock *, SetVector<Value *>> KillSet;
142 /// Values used in this block (and thus live); does not included values
143 /// killed within this block.
144 MapVector<BasicBlock *, SetVector<Value *>> LiveSet;
146 /// Values live into this basic block (i.e. used by any
147 /// instruction in this basic block or ones reachable from here)
148 MapVector<BasicBlock *, SetVector<Value *>> LiveIn;
150 /// Values live out of this basic block (i.e. live into
151 /// any successor block)
152 MapVector<BasicBlock *, SetVector<Value *>> LiveOut;
155 // The type of the internal cache used inside the findBasePointers family
156 // of functions. From the callers perspective, this is an opaque type and
157 // should not be inspected.
159 // In the actual implementation this caches two relations:
160 // - The base relation itself (i.e. this pointer is based on that one)
161 // - The base defining value relation (i.e. before base_phi insertion)
162 // Generally, after the execution of a full findBasePointer call, only the
163 // base relation will remain. Internally, we add a mixture of the two
164 // types, then update all the second type to the first type
165 typedef MapVector<Value *, Value *> DefiningValueMapTy;
166 typedef SetVector<Value *> StatepointLiveSetTy;
167 typedef MapVector<AssertingVH<Instruction>, AssertingVH<Value>>
168 RematerializedValueMapTy;
170 struct PartiallyConstructedSafepointRecord {
171 /// The set of values known to be live across this safepoint
172 StatepointLiveSetTy LiveSet;
174 /// Mapping from live pointers to a base-defining-value
175 MapVector<Value *, Value *> PointerToBase;
177 /// The *new* gc.statepoint instruction itself. This produces the token
178 /// that normal path gc.relocates and the gc.result are tied to.
179 Instruction *StatepointToken;
181 /// Instruction to which exceptional gc relocates are attached
182 /// Makes it easier to iterate through them during relocationViaAlloca.
183 Instruction *UnwindToken;
185 /// Record live values we are rematerialized instead of relocating.
186 /// They are not included into 'LiveSet' field.
187 /// Maps rematerialized copy to it's original value.
188 RematerializedValueMapTy RematerializedValues;
192 static ArrayRef<Use> GetDeoptBundleOperands(ImmutableCallSite CS) {
193 Optional<OperandBundleUse> DeoptBundle =
194 CS.getOperandBundle(LLVMContext::OB_deopt);
196 if (!DeoptBundle.hasValue()) {
197 assert(AllowStatepointWithNoDeoptInfo &&
198 "Found non-leaf call without deopt info!");
202 return DeoptBundle.getValue().Inputs;
205 /// Compute the live-in set for every basic block in the function
206 static void computeLiveInValues(DominatorTree &DT, Function &F,
207 GCPtrLivenessData &Data);
209 /// Given results from the dataflow liveness computation, find the set of live
210 /// Values at a particular instruction.
211 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
212 StatepointLiveSetTy &out);
214 // TODO: Once we can get to the GCStrategy, this becomes
215 // Optional<bool> isGCManagedPointer(const Type *Ty) const override {
217 static bool isGCPointerType(Type *T) {
218 if (auto *PT = dyn_cast<PointerType>(T))
219 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
220 // GC managed heap. We know that a pointer into this heap needs to be
221 // updated and that no other pointer does.
222 return PT->getAddressSpace() == 1;
226 // Return true if this type is one which a) is a gc pointer or contains a GC
227 // pointer and b) is of a type this code expects to encounter as a live value.
228 // (The insertion code will assert that a type which matches (a) and not (b)
229 // is not encountered.)
230 static bool isHandledGCPointerType(Type *T) {
231 // We fully support gc pointers
232 if (isGCPointerType(T))
234 // We partially support vectors of gc pointers. The code will assert if it
235 // can't handle something.
236 if (auto VT = dyn_cast<VectorType>(T))
237 if (isGCPointerType(VT->getElementType()))
243 /// Returns true if this type contains a gc pointer whether we know how to
244 /// handle that type or not.
245 static bool containsGCPtrType(Type *Ty) {
246 if (isGCPointerType(Ty))
248 if (VectorType *VT = dyn_cast<VectorType>(Ty))
249 return isGCPointerType(VT->getScalarType());
250 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
251 return containsGCPtrType(AT->getElementType());
252 if (StructType *ST = dyn_cast<StructType>(Ty))
253 return any_of(ST->subtypes(), containsGCPtrType);
257 // Returns true if this is a type which a) is a gc pointer or contains a GC
258 // pointer and b) is of a type which the code doesn't expect (i.e. first class
259 // aggregates). Used to trip assertions.
260 static bool isUnhandledGCPointerType(Type *Ty) {
261 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
265 // Return the name of the value suffixed with the provided value, or if the
266 // value didn't have a name, the default value specified.
267 static std::string suffixed_name_or(Value *V, StringRef Suffix,
268 StringRef DefaultName) {
269 return V->hasName() ? (V->getName() + Suffix).str() : DefaultName.str();
272 // Conservatively identifies any definitions which might be live at the
273 // given instruction. The analysis is performed immediately before the
274 // given instruction. Values defined by that instruction are not considered
275 // live. Values used by that instruction are considered live.
277 analyzeParsePointLiveness(DominatorTree &DT,
278 GCPtrLivenessData &OriginalLivenessData, CallSite CS,
279 PartiallyConstructedSafepointRecord &Result) {
280 Instruction *Inst = CS.getInstruction();
282 StatepointLiveSetTy LiveSet;
283 findLiveSetAtInst(Inst, OriginalLivenessData, LiveSet);
286 dbgs() << "Live Variables:\n";
287 for (Value *V : LiveSet)
288 dbgs() << " " << V->getName() << " " << *V << "\n";
290 if (PrintLiveSetSize) {
291 dbgs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
292 dbgs() << "Number live values: " << LiveSet.size() << "\n";
294 Result.LiveSet = LiveSet;
297 static bool isKnownBaseResult(Value *V);
299 /// A single base defining value - An immediate base defining value for an
300 /// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'.
301 /// For instructions which have multiple pointer [vector] inputs or that
302 /// transition between vector and scalar types, there is no immediate base
303 /// defining value. The 'base defining value' for 'Def' is the transitive
304 /// closure of this relation stopping at the first instruction which has no
305 /// immediate base defining value. The b.d.v. might itself be a base pointer,
306 /// but it can also be an arbitrary derived pointer.
307 struct BaseDefiningValueResult {
308 /// Contains the value which is the base defining value.
310 /// True if the base defining value is also known to be an actual base
312 const bool IsKnownBase;
313 BaseDefiningValueResult(Value *BDV, bool IsKnownBase)
314 : BDV(BDV), IsKnownBase(IsKnownBase) {
316 // Check consistency between new and old means of checking whether a BDV is
318 bool MustBeBase = isKnownBaseResult(BDV);
319 assert(!MustBeBase || MustBeBase == IsKnownBase);
325 static BaseDefiningValueResult findBaseDefiningValue(Value *I);
327 /// Return a base defining value for the 'Index' element of the given vector
328 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
329 /// 'I'. As an optimization, this method will try to determine when the
330 /// element is known to already be a base pointer. If this can be established,
331 /// the second value in the returned pair will be true. Note that either a
332 /// vector or a pointer typed value can be returned. For the former, the
333 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
334 /// If the later, the return pointer is a BDV (or possibly a base) for the
335 /// particular element in 'I'.
336 static BaseDefiningValueResult
337 findBaseDefiningValueOfVector(Value *I) {
338 // Each case parallels findBaseDefiningValue below, see that code for
339 // detailed motivation.
341 if (isa<Argument>(I))
342 // An incoming argument to the function is a base pointer
343 return BaseDefiningValueResult(I, true);
345 if (isa<Constant>(I))
346 // Base of constant vector consists only of constant null pointers.
347 // For reasoning see similar case inside 'findBaseDefiningValue' function.
348 return BaseDefiningValueResult(ConstantAggregateZero::get(I->getType()),
351 if (isa<LoadInst>(I))
352 return BaseDefiningValueResult(I, true);
354 if (isa<InsertElementInst>(I))
355 // We don't know whether this vector contains entirely base pointers or
356 // not. To be conservatively correct, we treat it as a BDV and will
357 // duplicate code as needed to construct a parallel vector of bases.
358 return BaseDefiningValueResult(I, false);
360 if (isa<ShuffleVectorInst>(I))
361 // We don't know whether this vector contains entirely base pointers or
362 // not. To be conservatively correct, we treat it as a BDV and will
363 // duplicate code as needed to construct a parallel vector of bases.
364 // TODO: There a number of local optimizations which could be applied here
365 // for particular sufflevector patterns.
366 return BaseDefiningValueResult(I, false);
368 // A PHI or Select is a base defining value. The outer findBasePointer
369 // algorithm is responsible for constructing a base value for this BDV.
370 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
371 "unknown vector instruction - no base found for vector element");
372 return BaseDefiningValueResult(I, false);
375 /// Helper function for findBasePointer - Will return a value which either a)
376 /// defines the base pointer for the input, b) blocks the simple search
377 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
378 /// from pointer to vector type or back.
379 static BaseDefiningValueResult findBaseDefiningValue(Value *I) {
380 assert(I->getType()->isPtrOrPtrVectorTy() &&
381 "Illegal to ask for the base pointer of a non-pointer type");
383 if (I->getType()->isVectorTy())
384 return findBaseDefiningValueOfVector(I);
386 if (isa<Argument>(I))
387 // An incoming argument to the function is a base pointer
388 // We should have never reached here if this argument isn't an gc value
389 return BaseDefiningValueResult(I, true);
391 if (isa<Constant>(I)) {
392 // We assume that objects with a constant base (e.g. a global) can't move
393 // and don't need to be reported to the collector because they are always
394 // live. Besides global references, all kinds of constants (e.g. undef,
395 // constant expressions, null pointers) can be introduced by the inliner or
396 // the optimizer, especially on dynamically dead paths.
397 // Here we treat all of them as having single null base. By doing this we
398 // trying to avoid problems reporting various conflicts in a form of
399 // "phi (const1, const2)" or "phi (const, regular gc ptr)".
400 // See constant.ll file for relevant test cases.
402 return BaseDefiningValueResult(
403 ConstantPointerNull::get(cast<PointerType>(I->getType())), true);
406 if (CastInst *CI = dyn_cast<CastInst>(I)) {
407 Value *Def = CI->stripPointerCasts();
408 // If stripping pointer casts changes the address space there is an
409 // addrspacecast in between.
410 assert(cast<PointerType>(Def->getType())->getAddressSpace() ==
411 cast<PointerType>(CI->getType())->getAddressSpace() &&
412 "unsupported addrspacecast");
413 // If we find a cast instruction here, it means we've found a cast which is
414 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
415 // handle int->ptr conversion.
416 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
417 return findBaseDefiningValue(Def);
420 if (isa<LoadInst>(I))
421 // The value loaded is an gc base itself
422 return BaseDefiningValueResult(I, true);
425 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
426 // The base of this GEP is the base
427 return findBaseDefiningValue(GEP->getPointerOperand());
429 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
430 switch (II->getIntrinsicID()) {
432 // fall through to general call handling
434 case Intrinsic::experimental_gc_statepoint:
435 llvm_unreachable("statepoints don't produce pointers");
436 case Intrinsic::experimental_gc_relocate: {
437 // Rerunning safepoint insertion after safepoints are already
438 // inserted is not supported. It could probably be made to work,
439 // but why are you doing this? There's no good reason.
440 llvm_unreachable("repeat safepoint insertion is not supported");
442 case Intrinsic::gcroot:
443 // Currently, this mechanism hasn't been extended to work with gcroot.
444 // There's no reason it couldn't be, but I haven't thought about the
445 // implications much.
447 "interaction with the gcroot mechanism is not supported");
450 // We assume that functions in the source language only return base
451 // pointers. This should probably be generalized via attributes to support
452 // both source language and internal functions.
453 if (isa<CallInst>(I) || isa<InvokeInst>(I))
454 return BaseDefiningValueResult(I, true);
456 // TODO: I have absolutely no idea how to implement this part yet. It's not
457 // necessarily hard, I just haven't really looked at it yet.
458 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
460 if (isa<AtomicCmpXchgInst>(I))
461 // A CAS is effectively a atomic store and load combined under a
462 // predicate. From the perspective of base pointers, we just treat it
464 return BaseDefiningValueResult(I, true);
466 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
467 "binary ops which don't apply to pointers");
469 // The aggregate ops. Aggregates can either be in the heap or on the
470 // stack, but in either case, this is simply a field load. As a result,
471 // this is a defining definition of the base just like a load is.
472 if (isa<ExtractValueInst>(I))
473 return BaseDefiningValueResult(I, true);
475 // We should never see an insert vector since that would require we be
476 // tracing back a struct value not a pointer value.
477 assert(!isa<InsertValueInst>(I) &&
478 "Base pointer for a struct is meaningless");
480 // An extractelement produces a base result exactly when it's input does.
481 // We may need to insert a parallel instruction to extract the appropriate
482 // element out of the base vector corresponding to the input. Given this,
483 // it's analogous to the phi and select case even though it's not a merge.
484 if (isa<ExtractElementInst>(I))
485 // Note: There a lot of obvious peephole cases here. This are deliberately
486 // handled after the main base pointer inference algorithm to make writing
487 // test cases to exercise that code easier.
488 return BaseDefiningValueResult(I, false);
490 // The last two cases here don't return a base pointer. Instead, they
491 // return a value which dynamically selects from among several base
492 // derived pointers (each with it's own base potentially). It's the job of
493 // the caller to resolve these.
494 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
495 "missing instruction case in findBaseDefiningValing");
496 return BaseDefiningValueResult(I, false);
499 /// Returns the base defining value for this value.
500 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
501 Value *&Cached = Cache[I];
503 Cached = findBaseDefiningValue(I).BDV;
504 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
505 << Cached->getName() << "\n");
507 assert(Cache[I] != nullptr);
511 /// Return a base pointer for this value if known. Otherwise, return it's
512 /// base defining value.
513 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
514 Value *Def = findBaseDefiningValueCached(I, Cache);
515 auto Found = Cache.find(Def);
516 if (Found != Cache.end()) {
517 // Either a base-of relation, or a self reference. Caller must check.
518 return Found->second;
520 // Only a BDV available
524 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
525 /// is it known to be a base pointer? Or do we need to continue searching.
526 static bool isKnownBaseResult(Value *V) {
527 if (!isa<PHINode>(V) && !isa<SelectInst>(V) &&
528 !isa<ExtractElementInst>(V) && !isa<InsertElementInst>(V) &&
529 !isa<ShuffleVectorInst>(V)) {
530 // no recursion possible
533 if (isa<Instruction>(V) &&
534 cast<Instruction>(V)->getMetadata("is_base_value")) {
535 // This is a previously inserted base phi or select. We know
536 // that this is a base value.
540 // We need to keep searching
545 /// Models the state of a single base defining value in the findBasePointer
546 /// algorithm for determining where a new instruction is needed to propagate
547 /// the base of this BDV.
550 enum Status { Unknown, Base, Conflict };
552 BDVState() : Status(Unknown), BaseValue(nullptr) {}
554 explicit BDVState(Status Status, Value *BaseValue = nullptr)
555 : Status(Status), BaseValue(BaseValue) {
556 assert(Status != Base || BaseValue);
559 explicit BDVState(Value *BaseValue) : Status(Base), BaseValue(BaseValue) {}
561 Status getStatus() const { return Status; }
562 Value *getBaseValue() const { return BaseValue; }
564 bool isBase() const { return getStatus() == Base; }
565 bool isUnknown() const { return getStatus() == Unknown; }
566 bool isConflict() const { return getStatus() == Conflict; }
568 bool operator==(const BDVState &Other) const {
569 return BaseValue == Other.BaseValue && Status == Other.Status;
572 bool operator!=(const BDVState &other) const { return !(*this == other); }
580 void print(raw_ostream &OS) const {
581 switch (getStatus()) {
592 OS << " (" << getBaseValue() << " - "
593 << (getBaseValue() ? getBaseValue()->getName() : "nullptr") << "): ";
598 AssertingVH<Value> BaseValue; // Non-null only if Status == Base.
603 static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
609 static BDVState meetBDVStateImpl(const BDVState &LHS, const BDVState &RHS) {
610 switch (LHS.getStatus()) {
611 case BDVState::Unknown:
615 assert(LHS.getBaseValue() && "can't be null");
620 if (LHS.getBaseValue() == RHS.getBaseValue()) {
621 assert(LHS == RHS && "equality broken!");
624 return BDVState(BDVState::Conflict);
626 assert(RHS.isConflict() && "only three states!");
627 return BDVState(BDVState::Conflict);
629 case BDVState::Conflict:
632 llvm_unreachable("only three states!");
635 // Values of type BDVState form a lattice, and this function implements the meet
637 static BDVState meetBDVState(BDVState LHS, BDVState RHS) {
638 BDVState Result = meetBDVStateImpl(LHS, RHS);
639 assert(Result == meetBDVStateImpl(RHS, LHS) &&
640 "Math is wrong: meet does not commute!");
644 /// For a given value or instruction, figure out what base ptr its derived from.
645 /// For gc objects, this is simply itself. On success, returns a value which is
646 /// the base pointer. (This is reliable and can be used for relocation.) On
647 /// failure, returns nullptr.
648 static Value *findBasePointer(Value *I, DefiningValueMapTy &Cache) {
649 Value *Def = findBaseOrBDV(I, Cache);
651 if (isKnownBaseResult(Def))
654 // Here's the rough algorithm:
655 // - For every SSA value, construct a mapping to either an actual base
656 // pointer or a PHI which obscures the base pointer.
657 // - Construct a mapping from PHI to unknown TOP state. Use an
658 // optimistic algorithm to propagate base pointer information. Lattice
663 // When algorithm terminates, all PHIs will either have a single concrete
664 // base or be in a conflict state.
665 // - For every conflict, insert a dummy PHI node without arguments. Add
666 // these to the base[Instruction] = BasePtr mapping. For every
667 // non-conflict, add the actual base.
668 // - For every conflict, add arguments for the base[a] of each input
671 // Note: A simpler form of this would be to add the conflict form of all
672 // PHIs without running the optimistic algorithm. This would be
673 // analogous to pessimistic data flow and would likely lead to an
674 // overall worse solution.
677 auto isExpectedBDVType = [](Value *BDV) {
678 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) ||
679 isa<ExtractElementInst>(BDV) || isa<InsertElementInst>(BDV) ||
680 isa<ShuffleVectorInst>(BDV);
684 // Once populated, will contain a mapping from each potentially non-base BDV
685 // to a lattice value (described above) which corresponds to that BDV.
686 // We use the order of insertion (DFS over the def/use graph) to provide a
687 // stable deterministic ordering for visiting DenseMaps (which are unordered)
688 // below. This is important for deterministic compilation.
689 MapVector<Value *, BDVState> States;
691 // Recursively fill in all base defining values reachable from the initial
692 // one for which we don't already know a definite base value for
694 SmallVector<Value*, 16> Worklist;
695 Worklist.push_back(Def);
696 States.insert({Def, BDVState()});
697 while (!Worklist.empty()) {
698 Value *Current = Worklist.pop_back_val();
699 assert(!isKnownBaseResult(Current) && "why did it get added?");
701 auto visitIncomingValue = [&](Value *InVal) {
702 Value *Base = findBaseOrBDV(InVal, Cache);
703 if (isKnownBaseResult(Base))
704 // Known bases won't need new instructions introduced and can be
707 assert(isExpectedBDVType(Base) && "the only non-base values "
708 "we see should be base defining values");
709 if (States.insert(std::make_pair(Base, BDVState())).second)
710 Worklist.push_back(Base);
712 if (PHINode *PN = dyn_cast<PHINode>(Current)) {
713 for (Value *InVal : PN->incoming_values())
714 visitIncomingValue(InVal);
715 } else if (SelectInst *SI = dyn_cast<SelectInst>(Current)) {
716 visitIncomingValue(SI->getTrueValue());
717 visitIncomingValue(SI->getFalseValue());
718 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
719 visitIncomingValue(EE->getVectorOperand());
720 } else if (auto *IE = dyn_cast<InsertElementInst>(Current)) {
721 visitIncomingValue(IE->getOperand(0)); // vector operand
722 visitIncomingValue(IE->getOperand(1)); // scalar operand
723 } else if (auto *SV = dyn_cast<ShuffleVectorInst>(Current)) {
724 visitIncomingValue(SV->getOperand(0));
725 visitIncomingValue(SV->getOperand(1));
728 llvm_unreachable("Unimplemented instruction case");
734 DEBUG(dbgs() << "States after initialization:\n");
735 for (auto Pair : States) {
736 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
740 // Return a phi state for a base defining value. We'll generate a new
741 // base state for known bases and expect to find a cached state otherwise.
742 auto getStateForBDV = [&](Value *baseValue) {
743 if (isKnownBaseResult(baseValue))
744 return BDVState(baseValue);
745 auto I = States.find(baseValue);
746 assert(I != States.end() && "lookup failed!");
750 bool Progress = true;
753 const size_t OldSize = States.size();
756 // We're only changing values in this loop, thus safe to keep iterators.
757 // Since this is computing a fixed point, the order of visit does not
758 // effect the result. TODO: We could use a worklist here and make this run
760 for (auto Pair : States) {
761 Value *BDV = Pair.first;
762 assert(!isKnownBaseResult(BDV) && "why did it get added?");
764 // Given an input value for the current instruction, return a BDVState
765 // instance which represents the BDV of that value.
766 auto getStateForInput = [&](Value *V) mutable {
767 Value *BDV = findBaseOrBDV(V, Cache);
768 return getStateForBDV(BDV);
772 if (SelectInst *SI = dyn_cast<SelectInst>(BDV)) {
773 NewState = meetBDVState(NewState, getStateForInput(SI->getTrueValue()));
775 meetBDVState(NewState, getStateForInput(SI->getFalseValue()));
776 } else if (PHINode *PN = dyn_cast<PHINode>(BDV)) {
777 for (Value *Val : PN->incoming_values())
778 NewState = meetBDVState(NewState, getStateForInput(Val));
779 } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) {
780 // The 'meet' for an extractelement is slightly trivial, but it's still
781 // useful in that it drives us to conflict if our input is.
783 meetBDVState(NewState, getStateForInput(EE->getVectorOperand()));
784 } else if (auto *IE = dyn_cast<InsertElementInst>(BDV)){
785 // Given there's a inherent type mismatch between the operands, will
786 // *always* produce Conflict.
787 NewState = meetBDVState(NewState, getStateForInput(IE->getOperand(0)));
788 NewState = meetBDVState(NewState, getStateForInput(IE->getOperand(1)));
790 // The only instance this does not return a Conflict is when both the
791 // vector operands are the same vector.
792 auto *SV = cast<ShuffleVectorInst>(BDV);
793 NewState = meetBDVState(NewState, getStateForInput(SV->getOperand(0)));
794 NewState = meetBDVState(NewState, getStateForInput(SV->getOperand(1)));
797 BDVState OldState = States[BDV];
798 if (OldState != NewState) {
800 States[BDV] = NewState;
804 assert(OldSize == States.size() &&
805 "fixed point shouldn't be adding any new nodes to state");
809 DEBUG(dbgs() << "States after meet iteration:\n");
810 for (auto Pair : States) {
811 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
815 // Insert Phis for all conflicts
816 // TODO: adjust naming patterns to avoid this order of iteration dependency
817 for (auto Pair : States) {
818 Instruction *I = cast<Instruction>(Pair.first);
819 BDVState State = Pair.second;
820 assert(!isKnownBaseResult(I) && "why did it get added?");
821 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
823 // extractelement instructions are a bit special in that we may need to
824 // insert an extract even when we know an exact base for the instruction.
825 // The problem is that we need to convert from a vector base to a scalar
826 // base for the particular indice we're interested in.
827 if (State.isBase() && isa<ExtractElementInst>(I) &&
828 isa<VectorType>(State.getBaseValue()->getType())) {
829 auto *EE = cast<ExtractElementInst>(I);
830 // TODO: In many cases, the new instruction is just EE itself. We should
831 // exploit this, but can't do it here since it would break the invariant
832 // about the BDV not being known to be a base.
833 auto *BaseInst = ExtractElementInst::Create(
834 State.getBaseValue(), EE->getIndexOperand(), "base_ee", EE);
835 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
836 States[I] = BDVState(BDVState::Base, BaseInst);
839 // Since we're joining a vector and scalar base, they can never be the
840 // same. As a result, we should always see insert element having reached
841 // the conflict state.
842 assert(!isa<InsertElementInst>(I) || State.isConflict());
844 if (!State.isConflict())
847 /// Create and insert a new instruction which will represent the base of
848 /// the given instruction 'I'.
849 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
850 if (isa<PHINode>(I)) {
851 BasicBlock *BB = I->getParent();
852 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
853 assert(NumPreds > 0 && "how did we reach here");
854 std::string Name = suffixed_name_or(I, ".base", "base_phi");
855 return PHINode::Create(I->getType(), NumPreds, Name, I);
856 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
857 // The undef will be replaced later
858 UndefValue *Undef = UndefValue::get(SI->getType());
859 std::string Name = suffixed_name_or(I, ".base", "base_select");
860 return SelectInst::Create(SI->getCondition(), Undef, Undef, Name, SI);
861 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
862 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
863 std::string Name = suffixed_name_or(I, ".base", "base_ee");
864 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
866 } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
867 UndefValue *VecUndef = UndefValue::get(IE->getOperand(0)->getType());
868 UndefValue *ScalarUndef = UndefValue::get(IE->getOperand(1)->getType());
869 std::string Name = suffixed_name_or(I, ".base", "base_ie");
870 return InsertElementInst::Create(VecUndef, ScalarUndef,
871 IE->getOperand(2), Name, IE);
873 auto *SV = cast<ShuffleVectorInst>(I);
874 UndefValue *VecUndef = UndefValue::get(SV->getOperand(0)->getType());
875 std::string Name = suffixed_name_or(I, ".base", "base_sv");
876 return new ShuffleVectorInst(VecUndef, VecUndef, SV->getOperand(2),
880 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
881 // Add metadata marking this as a base value
882 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
883 States[I] = BDVState(BDVState::Conflict, BaseInst);
886 // Returns a instruction which produces the base pointer for a given
887 // instruction. The instruction is assumed to be an input to one of the BDVs
888 // seen in the inference algorithm above. As such, we must either already
889 // know it's base defining value is a base, or have inserted a new
890 // instruction to propagate the base of it's BDV and have entered that newly
891 // introduced instruction into the state table. In either case, we are
892 // assured to be able to determine an instruction which produces it's base
894 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
895 Value *BDV = findBaseOrBDV(Input, Cache);
896 Value *Base = nullptr;
897 if (isKnownBaseResult(BDV)) {
900 // Either conflict or base.
901 assert(States.count(BDV));
902 Base = States[BDV].getBaseValue();
904 assert(Base && "Can't be null");
905 // The cast is needed since base traversal may strip away bitcasts
906 if (Base->getType() != Input->getType() && InsertPt)
907 Base = new BitCastInst(Base, Input->getType(), "cast", InsertPt);
911 // Fixup all the inputs of the new PHIs. Visit order needs to be
912 // deterministic and predictable because we're naming newly created
914 for (auto Pair : States) {
915 Instruction *BDV = cast<Instruction>(Pair.first);
916 BDVState State = Pair.second;
918 assert(!isKnownBaseResult(BDV) && "why did it get added?");
919 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
920 if (!State.isConflict())
923 if (PHINode *BasePHI = dyn_cast<PHINode>(State.getBaseValue())) {
924 PHINode *PN = cast<PHINode>(BDV);
925 unsigned NumPHIValues = PN->getNumIncomingValues();
926 for (unsigned i = 0; i < NumPHIValues; i++) {
927 Value *InVal = PN->getIncomingValue(i);
928 BasicBlock *InBB = PN->getIncomingBlock(i);
930 // If we've already seen InBB, add the same incoming value
931 // we added for it earlier. The IR verifier requires phi
932 // nodes with multiple entries from the same basic block
933 // to have the same incoming value for each of those
934 // entries. If we don't do this check here and basephi
935 // has a different type than base, we'll end up adding two
936 // bitcasts (and hence two distinct values) as incoming
937 // values for the same basic block.
939 int BlockIndex = BasePHI->getBasicBlockIndex(InBB);
940 if (BlockIndex != -1) {
941 Value *OldBase = BasePHI->getIncomingValue(BlockIndex);
942 BasePHI->addIncoming(OldBase, InBB);
945 Value *Base = getBaseForInput(InVal, nullptr);
946 // In essence this assert states: the only way two values
947 // incoming from the same basic block may be different is by
948 // being different bitcasts of the same value. A cleanup
949 // that remains TODO is changing findBaseOrBDV to return an
950 // llvm::Value of the correct type (and still remain pure).
951 // This will remove the need to add bitcasts.
952 assert(Base->stripPointerCasts() == OldBase->stripPointerCasts() &&
953 "Sanity -- findBaseOrBDV should be pure!");
958 // Find the instruction which produces the base for each input. We may
959 // need to insert a bitcast in the incoming block.
960 // TODO: Need to split critical edges if insertion is needed
961 Value *Base = getBaseForInput(InVal, InBB->getTerminator());
962 BasePHI->addIncoming(Base, InBB);
964 assert(BasePHI->getNumIncomingValues() == NumPHIValues);
965 } else if (SelectInst *BaseSI =
966 dyn_cast<SelectInst>(State.getBaseValue())) {
967 SelectInst *SI = cast<SelectInst>(BDV);
969 // Find the instruction which produces the base for each input.
970 // We may need to insert a bitcast.
971 BaseSI->setTrueValue(getBaseForInput(SI->getTrueValue(), BaseSI));
972 BaseSI->setFalseValue(getBaseForInput(SI->getFalseValue(), BaseSI));
973 } else if (auto *BaseEE =
974 dyn_cast<ExtractElementInst>(State.getBaseValue())) {
975 Value *InVal = cast<ExtractElementInst>(BDV)->getVectorOperand();
976 // Find the instruction which produces the base for each input. We may
977 // need to insert a bitcast.
978 BaseEE->setOperand(0, getBaseForInput(InVal, BaseEE));
979 } else if (auto *BaseIE = dyn_cast<InsertElementInst>(State.getBaseValue())){
980 auto *BdvIE = cast<InsertElementInst>(BDV);
981 auto UpdateOperand = [&](int OperandIdx) {
982 Value *InVal = BdvIE->getOperand(OperandIdx);
983 Value *Base = getBaseForInput(InVal, BaseIE);
984 BaseIE->setOperand(OperandIdx, Base);
986 UpdateOperand(0); // vector operand
987 UpdateOperand(1); // scalar operand
989 auto *BaseSV = cast<ShuffleVectorInst>(State.getBaseValue());
990 auto *BdvSV = cast<ShuffleVectorInst>(BDV);
991 auto UpdateOperand = [&](int OperandIdx) {
992 Value *InVal = BdvSV->getOperand(OperandIdx);
993 Value *Base = getBaseForInput(InVal, BaseSV);
994 BaseSV->setOperand(OperandIdx, Base);
996 UpdateOperand(0); // vector operand
997 UpdateOperand(1); // vector operand
1001 // Cache all of our results so we can cheaply reuse them
1002 // NOTE: This is actually two caches: one of the base defining value
1003 // relation and one of the base pointer relation! FIXME
1004 for (auto Pair : States) {
1005 auto *BDV = Pair.first;
1006 Value *Base = Pair.second.getBaseValue();
1007 assert(BDV && Base);
1008 assert(!isKnownBaseResult(BDV) && "why did it get added?");
1010 DEBUG(dbgs() << "Updating base value cache"
1011 << " for: " << BDV->getName() << " from: "
1012 << (Cache.count(BDV) ? Cache[BDV]->getName().str() : "none")
1013 << " to: " << Base->getName() << "\n");
1015 if (Cache.count(BDV)) {
1016 assert(isKnownBaseResult(Base) &&
1017 "must be something we 'know' is a base pointer");
1018 // Once we transition from the BDV relation being store in the Cache to
1019 // the base relation being stored, it must be stable
1020 assert((!isKnownBaseResult(Cache[BDV]) || Cache[BDV] == Base) &&
1021 "base relation should be stable");
1025 assert(Cache.count(Def));
1029 // For a set of live pointers (base and/or derived), identify the base
1030 // pointer of the object which they are derived from. This routine will
1031 // mutate the IR graph as needed to make the 'base' pointer live at the
1032 // definition site of 'derived'. This ensures that any use of 'derived' can
1033 // also use 'base'. This may involve the insertion of a number of
1034 // additional PHI nodes.
1036 // preconditions: live is a set of pointer type Values
1038 // side effects: may insert PHI nodes into the existing CFG, will preserve
1039 // CFG, will not remove or mutate any existing nodes
1041 // post condition: PointerToBase contains one (derived, base) pair for every
1042 // pointer in live. Note that derived can be equal to base if the original
1043 // pointer was a base pointer.
1045 findBasePointers(const StatepointLiveSetTy &live,
1046 MapVector<Value *, Value *> &PointerToBase,
1047 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1048 for (Value *ptr : live) {
1049 Value *base = findBasePointer(ptr, DVCache);
1050 assert(base && "failed to find base pointer");
1051 PointerToBase[ptr] = base;
1052 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1053 DT->dominates(cast<Instruction>(base)->getParent(),
1054 cast<Instruction>(ptr)->getParent())) &&
1055 "The base we found better dominate the derived pointer");
1059 /// Find the required based pointers (and adjust the live set) for the given
1061 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1063 PartiallyConstructedSafepointRecord &result) {
1064 MapVector<Value *, Value *> PointerToBase;
1065 findBasePointers(result.LiveSet, PointerToBase, &DT, DVCache);
1067 if (PrintBasePointers) {
1068 errs() << "Base Pairs (w/o Relocation):\n";
1069 for (auto &Pair : PointerToBase) {
1070 errs() << " derived ";
1071 Pair.first->printAsOperand(errs(), false);
1073 Pair.second->printAsOperand(errs(), false);
1078 result.PointerToBase = PointerToBase;
1081 /// Given an updated version of the dataflow liveness results, update the
1082 /// liveset and base pointer maps for the call site CS.
1083 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1085 PartiallyConstructedSafepointRecord &result);
1087 static void recomputeLiveInValues(
1088 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1089 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1090 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1091 // again. The old values are still live and will help it stabilize quickly.
1092 GCPtrLivenessData RevisedLivenessData;
1093 computeLiveInValues(DT, F, RevisedLivenessData);
1094 for (size_t i = 0; i < records.size(); i++) {
1095 struct PartiallyConstructedSafepointRecord &info = records[i];
1096 recomputeLiveInValues(RevisedLivenessData, toUpdate[i], info);
1100 // When inserting gc.relocate and gc.result calls, we need to ensure there are
1101 // no uses of the original value / return value between the gc.statepoint and
1102 // the gc.relocate / gc.result call. One case which can arise is a phi node
1103 // starting one of the successor blocks. We also need to be able to insert the
1104 // gc.relocates only on the path which goes through the statepoint. We might
1105 // need to split an edge to make this possible.
1107 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1108 DominatorTree &DT) {
1109 BasicBlock *Ret = BB;
1110 if (!BB->getUniquePredecessor())
1111 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1113 // Now that 'Ret' has unique predecessor we can safely remove all phi nodes
1115 FoldSingleEntryPHINodes(Ret);
1116 assert(!isa<PHINode>(Ret->begin()) &&
1117 "All PHI nodes should have been removed!");
1119 // At this point, we can safely insert a gc.relocate or gc.result as the first
1120 // instruction in Ret if needed.
1124 // Create new attribute set containing only attributes which can be transferred
1125 // from original call to the safepoint.
1126 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1129 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1130 unsigned Index = AS.getSlotIndex(Slot);
1132 if (Index == AttributeSet::ReturnIndex ||
1133 Index == AttributeSet::FunctionIndex) {
1135 for (Attribute Attr : make_range(AS.begin(Slot), AS.end(Slot))) {
1137 // Do not allow certain attributes - just skip them
1138 // Safepoint can not be read only or read none.
1139 if (Attr.hasAttribute(Attribute::ReadNone) ||
1140 Attr.hasAttribute(Attribute::ReadOnly))
1143 // These attributes control the generation of the gc.statepoint call /
1144 // invoke itself; and once the gc.statepoint is in place, they're of no
1146 if (isStatepointDirectiveAttr(Attr))
1149 Ret = Ret.addAttributes(
1150 AS.getContext(), Index,
1151 AttributeSet::get(AS.getContext(), Index, AttrBuilder(Attr)));
1155 // Just skip parameter attributes for now
1161 /// Helper function to place all gc relocates necessary for the given
1164 /// liveVariables - list of variables to be relocated.
1165 /// liveStart - index of the first live variable.
1166 /// basePtrs - base pointers.
1167 /// statepointToken - statepoint instruction to which relocates should be
1169 /// Builder - Llvm IR builder to be used to construct new calls.
1170 static void CreateGCRelocates(ArrayRef<Value *> LiveVariables,
1171 const int LiveStart,
1172 ArrayRef<Value *> BasePtrs,
1173 Instruction *StatepointToken,
1174 IRBuilder<> Builder) {
1175 if (LiveVariables.empty())
1178 auto FindIndex = [](ArrayRef<Value *> LiveVec, Value *Val) {
1179 auto ValIt = find(LiveVec, Val);
1180 assert(ValIt != LiveVec.end() && "Val not found in LiveVec!");
1181 size_t Index = std::distance(LiveVec.begin(), ValIt);
1182 assert(Index < LiveVec.size() && "Bug in std::find?");
1185 Module *M = StatepointToken->getModule();
1187 // All gc_relocate are generated as i8 addrspace(1)* (or a vector type whose
1188 // element type is i8 addrspace(1)*). We originally generated unique
1189 // declarations for each pointer type, but this proved problematic because
1190 // the intrinsic mangling code is incomplete and fragile. Since we're moving
1191 // towards a single unified pointer type anyways, we can just cast everything
1192 // to an i8* of the right address space. A bitcast is added later to convert
1193 // gc_relocate to the actual value's type.
1194 auto getGCRelocateDecl = [&] (Type *Ty) {
1195 assert(isHandledGCPointerType(Ty));
1196 auto AS = Ty->getScalarType()->getPointerAddressSpace();
1197 Type *NewTy = Type::getInt8PtrTy(M->getContext(), AS);
1198 if (auto *VT = dyn_cast<VectorType>(Ty))
1199 NewTy = VectorType::get(NewTy, VT->getNumElements());
1200 return Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate,
1204 // Lazily populated map from input types to the canonicalized form mentioned
1205 // in the comment above. This should probably be cached somewhere more
1207 DenseMap<Type*, Value*> TypeToDeclMap;
1209 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1210 // Generate the gc.relocate call and save the result
1212 Builder.getInt32(LiveStart + FindIndex(LiveVariables, BasePtrs[i]));
1213 Value *LiveIdx = Builder.getInt32(LiveStart + i);
1215 Type *Ty = LiveVariables[i]->getType();
1216 if (!TypeToDeclMap.count(Ty))
1217 TypeToDeclMap[Ty] = getGCRelocateDecl(Ty);
1218 Value *GCRelocateDecl = TypeToDeclMap[Ty];
1220 // only specify a debug name if we can give a useful one
1221 CallInst *Reloc = Builder.CreateCall(
1222 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1223 suffixed_name_or(LiveVariables[i], ".relocated", ""));
1224 // Trick CodeGen into thinking there are lots of free registers at this
1226 Reloc->setCallingConv(CallingConv::Cold);
1232 /// This struct is used to defer RAUWs and `eraseFromParent` s. Using this
1233 /// avoids having to worry about keeping around dangling pointers to Values.
1234 class DeferredReplacement {
1235 AssertingVH<Instruction> Old;
1236 AssertingVH<Instruction> New;
1237 bool IsDeoptimize = false;
1239 DeferredReplacement() {}
1242 static DeferredReplacement createRAUW(Instruction *Old, Instruction *New) {
1243 assert(Old != New && Old && New &&
1244 "Cannot RAUW equal values or to / from null!");
1246 DeferredReplacement D;
1252 static DeferredReplacement createDelete(Instruction *ToErase) {
1253 DeferredReplacement D;
1258 static DeferredReplacement createDeoptimizeReplacement(Instruction *Old) {
1260 auto *F = cast<CallInst>(Old)->getCalledFunction();
1261 assert(F && F->getIntrinsicID() == Intrinsic::experimental_deoptimize &&
1262 "Only way to construct a deoptimize deferred replacement");
1264 DeferredReplacement D;
1266 D.IsDeoptimize = true;
1270 /// Does the task represented by this instance.
1271 void doReplacement() {
1272 Instruction *OldI = Old;
1273 Instruction *NewI = New;
1275 assert(OldI != NewI && "Disallowed at construction?!");
1276 assert((!IsDeoptimize || !New) &&
1277 "Deoptimize instrinsics are not replaced!");
1283 OldI->replaceAllUsesWith(NewI);
1286 // Note: we've inserted instructions, so the call to llvm.deoptimize may
1287 // not necessarilly be followed by the matching return.
1288 auto *RI = cast<ReturnInst>(OldI->getParent()->getTerminator());
1289 new UnreachableInst(RI->getContext(), RI);
1290 RI->eraseFromParent();
1293 OldI->eraseFromParent();
1298 static StringRef getDeoptLowering(CallSite CS) {
1299 const char *DeoptLowering = "deopt-lowering";
1300 if (CS.hasFnAttr(DeoptLowering)) {
1301 // FIXME: CallSite has a *really* confusing interface around attributes
1303 const AttributeSet &CSAS = CS.getAttributes();
1304 if (CSAS.hasAttribute(AttributeSet::FunctionIndex,
1306 return CSAS.getAttribute(AttributeSet::FunctionIndex,
1307 DeoptLowering).getValueAsString();
1308 Function *F = CS.getCalledFunction();
1309 assert(F && F->hasFnAttribute(DeoptLowering));
1310 return F->getFnAttribute(DeoptLowering).getValueAsString();
1312 return "live-through";
1317 makeStatepointExplicitImpl(const CallSite CS, /* to replace */
1318 const SmallVectorImpl<Value *> &BasePtrs,
1319 const SmallVectorImpl<Value *> &LiveVariables,
1320 PartiallyConstructedSafepointRecord &Result,
1321 std::vector<DeferredReplacement> &Replacements) {
1322 assert(BasePtrs.size() == LiveVariables.size());
1324 // Then go ahead and use the builder do actually do the inserts. We insert
1325 // immediately before the previous instruction under the assumption that all
1326 // arguments will be available here. We can't insert afterwards since we may
1327 // be replacing a terminator.
1328 Instruction *InsertBefore = CS.getInstruction();
1329 IRBuilder<> Builder(InsertBefore);
1331 ArrayRef<Value *> GCArgs(LiveVariables);
1332 uint64_t StatepointID = StatepointDirectives::DefaultStatepointID;
1333 uint32_t NumPatchBytes = 0;
1334 uint32_t Flags = uint32_t(StatepointFlags::None);
1336 ArrayRef<Use> CallArgs(CS.arg_begin(), CS.arg_end());
1337 ArrayRef<Use> DeoptArgs = GetDeoptBundleOperands(CS);
1338 ArrayRef<Use> TransitionArgs;
1339 if (auto TransitionBundle =
1340 CS.getOperandBundle(LLVMContext::OB_gc_transition)) {
1341 Flags |= uint32_t(StatepointFlags::GCTransition);
1342 TransitionArgs = TransitionBundle->Inputs;
1345 // Instead of lowering calls to @llvm.experimental.deoptimize as normal calls
1346 // with a return value, we lower then as never returning calls to
1347 // __llvm_deoptimize that are followed by unreachable to get better codegen.
1348 bool IsDeoptimize = false;
1350 StatepointDirectives SD =
1351 parseStatepointDirectivesFromAttrs(CS.getAttributes());
1352 if (SD.NumPatchBytes)
1353 NumPatchBytes = *SD.NumPatchBytes;
1354 if (SD.StatepointID)
1355 StatepointID = *SD.StatepointID;
1357 // Pass through the requested lowering if any. The default is live-through.
1358 StringRef DeoptLowering = getDeoptLowering(CS);
1359 if (DeoptLowering.equals("live-in"))
1360 Flags |= uint32_t(StatepointFlags::DeoptLiveIn);
1362 assert(DeoptLowering.equals("live-through") && "Unsupported value!");
1365 Value *CallTarget = CS.getCalledValue();
1366 if (Function *F = dyn_cast<Function>(CallTarget)) {
1367 if (F->getIntrinsicID() == Intrinsic::experimental_deoptimize) {
1368 // Calls to llvm.experimental.deoptimize are lowered to calls to the
1369 // __llvm_deoptimize symbol. We want to resolve this now, since the
1370 // verifier does not allow taking the address of an intrinsic function.
1372 SmallVector<Type *, 8> DomainTy;
1373 for (Value *Arg : CallArgs)
1374 DomainTy.push_back(Arg->getType());
1375 auto *FTy = FunctionType::get(Type::getVoidTy(F->getContext()), DomainTy,
1376 /* isVarArg = */ false);
1378 // Note: CallTarget can be a bitcast instruction of a symbol if there are
1379 // calls to @llvm.experimental.deoptimize with different argument types in
1380 // the same module. This is fine -- we assume the frontend knew what it
1381 // was doing when generating this kind of IR.
1383 F->getParent()->getOrInsertFunction("__llvm_deoptimize", FTy);
1385 IsDeoptimize = true;
1389 // Create the statepoint given all the arguments
1390 Instruction *Token = nullptr;
1391 AttributeSet ReturnAttrs;
1393 CallInst *ToReplace = cast<CallInst>(CS.getInstruction());
1394 CallInst *Call = Builder.CreateGCStatepointCall(
1395 StatepointID, NumPatchBytes, CallTarget, Flags, CallArgs,
1396 TransitionArgs, DeoptArgs, GCArgs, "safepoint_token");
1398 Call->setTailCallKind(ToReplace->getTailCallKind());
1399 Call->setCallingConv(ToReplace->getCallingConv());
1401 // Currently we will fail on parameter attributes and on certain
1402 // function attributes.
1403 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes());
1404 // In case if we can handle this set of attributes - set up function attrs
1405 // directly on statepoint and return attrs later for gc_result intrinsic.
1406 Call->setAttributes(NewAttrs.getFnAttributes());
1407 ReturnAttrs = NewAttrs.getRetAttributes();
1411 // Put the following gc_result and gc_relocate calls immediately after the
1412 // the old call (which we're about to delete)
1413 assert(ToReplace->getNextNode() && "Not a terminator, must have next!");
1414 Builder.SetInsertPoint(ToReplace->getNextNode());
1415 Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc());
1417 InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction());
1419 // Insert the new invoke into the old block. We'll remove the old one in a
1420 // moment at which point this will become the new terminator for the
1422 InvokeInst *Invoke = Builder.CreateGCStatepointInvoke(
1423 StatepointID, NumPatchBytes, CallTarget, ToReplace->getNormalDest(),
1424 ToReplace->getUnwindDest(), Flags, CallArgs, TransitionArgs, DeoptArgs,
1425 GCArgs, "statepoint_token");
1427 Invoke->setCallingConv(ToReplace->getCallingConv());
1429 // Currently we will fail on parameter attributes and on certain
1430 // function attributes.
1431 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes());
1432 // In case if we can handle this set of attributes - set up function attrs
1433 // directly on statepoint and return attrs later for gc_result intrinsic.
1434 Invoke->setAttributes(NewAttrs.getFnAttributes());
1435 ReturnAttrs = NewAttrs.getRetAttributes();
1439 // Generate gc relocates in exceptional path
1440 BasicBlock *UnwindBlock = ToReplace->getUnwindDest();
1441 assert(!isa<PHINode>(UnwindBlock->begin()) &&
1442 UnwindBlock->getUniquePredecessor() &&
1443 "can't safely insert in this block!");
1445 Builder.SetInsertPoint(&*UnwindBlock->getFirstInsertionPt());
1446 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1448 // Attach exceptional gc relocates to the landingpad.
1449 Instruction *ExceptionalToken = UnwindBlock->getLandingPadInst();
1450 Result.UnwindToken = ExceptionalToken;
1452 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1453 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, ExceptionalToken,
1456 // Generate gc relocates and returns for normal block
1457 BasicBlock *NormalDest = ToReplace->getNormalDest();
1458 assert(!isa<PHINode>(NormalDest->begin()) &&
1459 NormalDest->getUniquePredecessor() &&
1460 "can't safely insert in this block!");
1462 Builder.SetInsertPoint(&*NormalDest->getFirstInsertionPt());
1464 // gc relocates will be generated later as if it were regular call
1467 assert(Token && "Should be set in one of the above branches!");
1470 // If we're wrapping an @llvm.experimental.deoptimize in a statepoint, we
1471 // transform the tail-call like structure to a call to a void function
1472 // followed by unreachable to get better codegen.
1473 Replacements.push_back(
1474 DeferredReplacement::createDeoptimizeReplacement(CS.getInstruction()));
1476 Token->setName("statepoint_token");
1477 if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
1479 CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "";
1480 CallInst *GCResult = Builder.CreateGCResult(Token, CS.getType(), Name);
1481 GCResult->setAttributes(CS.getAttributes().getRetAttributes());
1483 // We cannot RAUW or delete CS.getInstruction() because it could be in the
1484 // live set of some other safepoint, in which case that safepoint's
1485 // PartiallyConstructedSafepointRecord will hold a raw pointer to this
1486 // llvm::Instruction. Instead, we defer the replacement and deletion to
1487 // after the live sets have been made explicit in the IR, and we no longer
1488 // have raw pointers to worry about.
1489 Replacements.emplace_back(
1490 DeferredReplacement::createRAUW(CS.getInstruction(), GCResult));
1492 Replacements.emplace_back(
1493 DeferredReplacement::createDelete(CS.getInstruction()));
1497 Result.StatepointToken = Token;
1499 // Second, create a gc.relocate for every live variable
1500 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1501 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, Token, Builder);
1504 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1505 // which make the relocations happening at this safepoint explicit.
1507 // WARNING: Does not do any fixup to adjust users of the original live
1508 // values. That's the callers responsibility.
1510 makeStatepointExplicit(DominatorTree &DT, CallSite CS,
1511 PartiallyConstructedSafepointRecord &Result,
1512 std::vector<DeferredReplacement> &Replacements) {
1513 const auto &LiveSet = Result.LiveSet;
1514 const auto &PointerToBase = Result.PointerToBase;
1516 // Convert to vector for efficient cross referencing.
1517 SmallVector<Value *, 64> BaseVec, LiveVec;
1518 LiveVec.reserve(LiveSet.size());
1519 BaseVec.reserve(LiveSet.size());
1520 for (Value *L : LiveSet) {
1521 LiveVec.push_back(L);
1522 assert(PointerToBase.count(L));
1523 Value *Base = PointerToBase.find(L)->second;
1524 BaseVec.push_back(Base);
1526 assert(LiveVec.size() == BaseVec.size());
1528 // Do the actual rewriting and delete the old statepoint
1529 makeStatepointExplicitImpl(CS, BaseVec, LiveVec, Result, Replacements);
1532 // Helper function for the relocationViaAlloca.
1534 // It receives iterator to the statepoint gc relocates and emits a store to the
1535 // assigned location (via allocaMap) for the each one of them. It adds the
1536 // visited values into the visitedLiveValues set, which we will later use them
1537 // for sanity checking.
1539 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1540 DenseMap<Value *, Value *> &AllocaMap,
1541 DenseSet<Value *> &VisitedLiveValues) {
1543 for (User *U : GCRelocs) {
1544 GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U);
1548 Value *OriginalValue = Relocate->getDerivedPtr();
1549 assert(AllocaMap.count(OriginalValue));
1550 Value *Alloca = AllocaMap[OriginalValue];
1552 // Emit store into the related alloca
1553 // All gc_relocates are i8 addrspace(1)* typed, and it must be bitcasted to
1554 // the correct type according to alloca.
1555 assert(Relocate->getNextNode() &&
1556 "Should always have one since it's not a terminator");
1557 IRBuilder<> Builder(Relocate->getNextNode());
1558 Value *CastedRelocatedValue =
1559 Builder.CreateBitCast(Relocate,
1560 cast<AllocaInst>(Alloca)->getAllocatedType(),
1561 suffixed_name_or(Relocate, ".casted", ""));
1563 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1564 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1567 VisitedLiveValues.insert(OriginalValue);
1572 // Helper function for the "relocationViaAlloca". Similar to the
1573 // "insertRelocationStores" but works for rematerialized values.
1574 static void insertRematerializationStores(
1575 const RematerializedValueMapTy &RematerializedValues,
1576 DenseMap<Value *, Value *> &AllocaMap,
1577 DenseSet<Value *> &VisitedLiveValues) {
1579 for (auto RematerializedValuePair: RematerializedValues) {
1580 Instruction *RematerializedValue = RematerializedValuePair.first;
1581 Value *OriginalValue = RematerializedValuePair.second;
1583 assert(AllocaMap.count(OriginalValue) &&
1584 "Can not find alloca for rematerialized value");
1585 Value *Alloca = AllocaMap[OriginalValue];
1587 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1588 Store->insertAfter(RematerializedValue);
1591 VisitedLiveValues.insert(OriginalValue);
1596 /// Do all the relocation update via allocas and mem2reg
1597 static void relocationViaAlloca(
1598 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1599 ArrayRef<PartiallyConstructedSafepointRecord> Records) {
1601 // record initial number of (static) allocas; we'll check we have the same
1602 // number when we get done.
1603 int InitialAllocaNum = 0;
1604 for (Instruction &I : F.getEntryBlock())
1605 if (isa<AllocaInst>(I))
1609 // TODO-PERF: change data structures, reserve
1610 DenseMap<Value *, Value *> AllocaMap;
1611 SmallVector<AllocaInst *, 200> PromotableAllocas;
1612 // Used later to chack that we have enough allocas to store all values
1613 std::size_t NumRematerializedValues = 0;
1614 PromotableAllocas.reserve(Live.size());
1616 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1617 // "PromotableAllocas"
1618 auto emitAllocaFor = [&](Value *LiveValue) {
1619 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1620 F.getEntryBlock().getFirstNonPHI());
1621 AllocaMap[LiveValue] = Alloca;
1622 PromotableAllocas.push_back(Alloca);
1625 // Emit alloca for each live gc pointer
1626 for (Value *V : Live)
1629 // Emit allocas for rematerialized values
1630 for (const auto &Info : Records)
1631 for (auto RematerializedValuePair : Info.RematerializedValues) {
1632 Value *OriginalValue = RematerializedValuePair.second;
1633 if (AllocaMap.count(OriginalValue) != 0)
1636 emitAllocaFor(OriginalValue);
1637 ++NumRematerializedValues;
1640 // The next two loops are part of the same conceptual operation. We need to
1641 // insert a store to the alloca after the original def and at each
1642 // redefinition. We need to insert a load before each use. These are split
1643 // into distinct loops for performance reasons.
1645 // Update gc pointer after each statepoint: either store a relocated value or
1646 // null (if no relocated value was found for this gc pointer and it is not a
1647 // gc_result). This must happen before we update the statepoint with load of
1648 // alloca otherwise we lose the link between statepoint and old def.
1649 for (const auto &Info : Records) {
1650 Value *Statepoint = Info.StatepointToken;
1652 // This will be used for consistency check
1653 DenseSet<Value *> VisitedLiveValues;
1655 // Insert stores for normal statepoint gc relocates
1656 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1658 // In case if it was invoke statepoint
1659 // we will insert stores for exceptional path gc relocates.
1660 if (isa<InvokeInst>(Statepoint)) {
1661 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1665 // Do similar thing with rematerialized values
1666 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1669 if (ClobberNonLive) {
1670 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1671 // the gc.statepoint. This will turn some subtle GC problems into
1672 // slightly easier to debug SEGVs. Note that on large IR files with
1673 // lots of gc.statepoints this is extremely costly both memory and time
1675 SmallVector<AllocaInst *, 64> ToClobber;
1676 for (auto Pair : AllocaMap) {
1677 Value *Def = Pair.first;
1678 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1680 // This value was relocated
1681 if (VisitedLiveValues.count(Def)) {
1684 ToClobber.push_back(Alloca);
1687 auto InsertClobbersAt = [&](Instruction *IP) {
1688 for (auto *AI : ToClobber) {
1689 auto PT = cast<PointerType>(AI->getAllocatedType());
1690 Constant *CPN = ConstantPointerNull::get(PT);
1691 StoreInst *Store = new StoreInst(CPN, AI);
1692 Store->insertBefore(IP);
1696 // Insert the clobbering stores. These may get intermixed with the
1697 // gc.results and gc.relocates, but that's fine.
1698 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1699 InsertClobbersAt(&*II->getNormalDest()->getFirstInsertionPt());
1700 InsertClobbersAt(&*II->getUnwindDest()->getFirstInsertionPt());
1702 InsertClobbersAt(cast<Instruction>(Statepoint)->getNextNode());
1707 // Update use with load allocas and add store for gc_relocated.
1708 for (auto Pair : AllocaMap) {
1709 Value *Def = Pair.first;
1710 Value *Alloca = Pair.second;
1712 // We pre-record the uses of allocas so that we dont have to worry about
1713 // later update that changes the user information..
1715 SmallVector<Instruction *, 20> Uses;
1716 // PERF: trade a linear scan for repeated reallocation
1717 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1718 for (User *U : Def->users()) {
1719 if (!isa<ConstantExpr>(U)) {
1720 // If the def has a ConstantExpr use, then the def is either a
1721 // ConstantExpr use itself or null. In either case
1722 // (recursively in the first, directly in the second), the oop
1723 // it is ultimately dependent on is null and this particular
1724 // use does not need to be fixed up.
1725 Uses.push_back(cast<Instruction>(U));
1729 std::sort(Uses.begin(), Uses.end());
1730 auto Last = std::unique(Uses.begin(), Uses.end());
1731 Uses.erase(Last, Uses.end());
1733 for (Instruction *Use : Uses) {
1734 if (isa<PHINode>(Use)) {
1735 PHINode *Phi = cast<PHINode>(Use);
1736 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1737 if (Def == Phi->getIncomingValue(i)) {
1738 LoadInst *Load = new LoadInst(
1739 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1740 Phi->setIncomingValue(i, Load);
1744 LoadInst *Load = new LoadInst(Alloca, "", Use);
1745 Use->replaceUsesOfWith(Def, Load);
1749 // Emit store for the initial gc value. Store must be inserted after load,
1750 // otherwise store will be in alloca's use list and an extra load will be
1751 // inserted before it.
1752 StoreInst *Store = new StoreInst(Def, Alloca);
1753 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1754 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1755 // InvokeInst is a TerminatorInst so the store need to be inserted
1756 // into its normal destination block.
1757 BasicBlock *NormalDest = Invoke->getNormalDest();
1758 Store->insertBefore(NormalDest->getFirstNonPHI());
1760 assert(!Inst->isTerminator() &&
1761 "The only TerminatorInst that can produce a value is "
1762 "InvokeInst which is handled above.");
1763 Store->insertAfter(Inst);
1766 assert(isa<Argument>(Def));
1767 Store->insertAfter(cast<Instruction>(Alloca));
1771 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1772 "we must have the same allocas with lives");
1773 if (!PromotableAllocas.empty()) {
1774 // Apply mem2reg to promote alloca to SSA
1775 PromoteMemToReg(PromotableAllocas, DT);
1779 for (auto &I : F.getEntryBlock())
1780 if (isa<AllocaInst>(I))
1782 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1786 /// Implement a unique function which doesn't require we sort the input
1787 /// vector. Doing so has the effect of changing the output of a couple of
1788 /// tests in ways which make them less useful in testing fused safepoints.
1789 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1790 SmallSet<T, 8> Seen;
1791 Vec.erase(remove_if(Vec, [&](const T &V) { return !Seen.insert(V).second; }),
1795 /// Insert holders so that each Value is obviously live through the entire
1796 /// lifetime of the call.
1797 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1798 SmallVectorImpl<CallInst *> &Holders) {
1800 // No values to hold live, might as well not insert the empty holder
1803 Module *M = CS.getInstruction()->getModule();
1804 // Use a dummy vararg function to actually hold the values live
1805 Function *Func = cast<Function>(M->getOrInsertFunction(
1806 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1808 // For call safepoints insert dummy calls right after safepoint
1809 Holders.push_back(CallInst::Create(Func, Values, "",
1810 &*++CS.getInstruction()->getIterator()));
1813 // For invoke safepooints insert dummy calls both in normal and
1814 // exceptional destination blocks
1815 auto *II = cast<InvokeInst>(CS.getInstruction());
1816 Holders.push_back(CallInst::Create(
1817 Func, Values, "", &*II->getNormalDest()->getFirstInsertionPt()));
1818 Holders.push_back(CallInst::Create(
1819 Func, Values, "", &*II->getUnwindDest()->getFirstInsertionPt()));
1822 static void findLiveReferences(
1823 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1824 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1825 GCPtrLivenessData OriginalLivenessData;
1826 computeLiveInValues(DT, F, OriginalLivenessData);
1827 for (size_t i = 0; i < records.size(); i++) {
1828 struct PartiallyConstructedSafepointRecord &info = records[i];
1829 analyzeParsePointLiveness(DT, OriginalLivenessData, toUpdate[i], info);
1833 // Helper function for the "rematerializeLiveValues". It walks use chain
1834 // starting from the "CurrentValue" until it reaches the root of the chain, i.e.
1835 // the base or a value it cannot process. Only "simple" values are processed
1836 // (currently it is GEP's and casts). The returned root is examined by the
1837 // callers of findRematerializableChainToBasePointer. Fills "ChainToBase" array
1838 // with all visited values.
1839 static Value* findRematerializableChainToBasePointer(
1840 SmallVectorImpl<Instruction*> &ChainToBase,
1841 Value *CurrentValue) {
1843 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
1844 ChainToBase.push_back(GEP);
1845 return findRematerializableChainToBasePointer(ChainToBase,
1846 GEP->getPointerOperand());
1849 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
1850 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
1853 ChainToBase.push_back(CI);
1854 return findRematerializableChainToBasePointer(ChainToBase,
1858 // We have reached the root of the chain, which is either equal to the base or
1859 // is the first unsupported value along the use chain.
1860 return CurrentValue;
1863 // Helper function for the "rematerializeLiveValues". Compute cost of the use
1864 // chain we are going to rematerialize.
1866 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
1867 TargetTransformInfo &TTI) {
1870 for (Instruction *Instr : Chain) {
1871 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
1872 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
1873 "non noop cast is found during rematerialization");
1875 Type *SrcTy = CI->getOperand(0)->getType();
1876 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
1878 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
1879 // Cost of the address calculation
1880 Type *ValTy = GEP->getSourceElementType();
1881 Cost += TTI.getAddressComputationCost(ValTy);
1883 // And cost of the GEP itself
1884 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
1885 // allowed for the external usage)
1886 if (!GEP->hasAllConstantIndices())
1890 llvm_unreachable("unsupported instruciton type during rematerialization");
1897 static bool AreEquivalentPhiNodes(PHINode &OrigRootPhi, PHINode &AlternateRootPhi) {
1899 unsigned PhiNum = OrigRootPhi.getNumIncomingValues();
1900 if (PhiNum != AlternateRootPhi.getNumIncomingValues() ||
1901 OrigRootPhi.getParent() != AlternateRootPhi.getParent())
1903 // Map of incoming values and their corresponding basic blocks of
1905 SmallDenseMap<Value *, BasicBlock *, 8> CurrentIncomingValues;
1906 for (unsigned i = 0; i < PhiNum; i++)
1907 CurrentIncomingValues[OrigRootPhi.getIncomingValue(i)] =
1908 OrigRootPhi.getIncomingBlock(i);
1910 // Both current and base PHIs should have same incoming values and
1911 // the same basic blocks corresponding to the incoming values.
1912 for (unsigned i = 0; i < PhiNum; i++) {
1914 CurrentIncomingValues.find(AlternateRootPhi.getIncomingValue(i));
1915 if (CIVI == CurrentIncomingValues.end())
1917 BasicBlock *CurrentIncomingBB = CIVI->second;
1918 if (CurrentIncomingBB != AlternateRootPhi.getIncomingBlock(i))
1925 // From the statepoint live set pick values that are cheaper to recompute then
1926 // to relocate. Remove this values from the live set, rematerialize them after
1927 // statepoint and record them in "Info" structure. Note that similar to
1928 // relocated values we don't do any user adjustments here.
1929 static void rematerializeLiveValues(CallSite CS,
1930 PartiallyConstructedSafepointRecord &Info,
1931 TargetTransformInfo &TTI) {
1932 const unsigned int ChainLengthThreshold = 10;
1934 // Record values we are going to delete from this statepoint live set.
1935 // We can not di this in following loop due to iterator invalidation.
1936 SmallVector<Value *, 32> LiveValuesToBeDeleted;
1938 for (Value *LiveValue: Info.LiveSet) {
1939 // For each live pointer find it's defining chain
1940 SmallVector<Instruction *, 3> ChainToBase;
1941 assert(Info.PointerToBase.count(LiveValue));
1942 Value *RootOfChain =
1943 findRematerializableChainToBasePointer(ChainToBase,
1946 // Nothing to do, or chain is too long
1947 if ( ChainToBase.size() == 0 ||
1948 ChainToBase.size() > ChainLengthThreshold)
1951 // Handle the scenario where the RootOfChain is not equal to the
1952 // Base Value, but they are essentially the same phi values.
1953 if (RootOfChain != Info.PointerToBase[LiveValue]) {
1954 PHINode *OrigRootPhi = dyn_cast<PHINode>(RootOfChain);
1955 PHINode *AlternateRootPhi = dyn_cast<PHINode>(Info.PointerToBase[LiveValue]);
1956 if (!OrigRootPhi || !AlternateRootPhi)
1958 // PHI nodes that have the same incoming values, and belonging to the same
1959 // basic blocks are essentially the same SSA value. When the original phi
1960 // has incoming values with different base pointers, the original phi is
1961 // marked as conflict, and an additional `AlternateRootPhi` with the same
1962 // incoming values get generated by the findBasePointer function. We need
1963 // to identify the newly generated AlternateRootPhi (.base version of phi)
1964 // and RootOfChain (the original phi node itself) are the same, so that we
1965 // can rematerialize the gep and casts. This is a workaround for the
1966 // deficieny in the findBasePointer algorithm.
1967 if (!AreEquivalentPhiNodes(*OrigRootPhi, *AlternateRootPhi))
1969 // Now that the phi nodes are proved to be the same, assert that
1970 // findBasePointer's newly generated AlternateRootPhi is present in the
1971 // liveset of the call.
1972 assert(Info.LiveSet.count(AlternateRootPhi));
1974 // Compute cost of this chain
1975 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
1976 // TODO: We can also account for cases when we will be able to remove some
1977 // of the rematerialized values by later optimization passes. I.e if
1978 // we rematerialized several intersecting chains. Or if original values
1979 // don't have any uses besides this statepoint.
1981 // For invokes we need to rematerialize each chain twice - for normal and
1982 // for unwind basic blocks. Model this by multiplying cost by two.
1983 if (CS.isInvoke()) {
1986 // If it's too expensive - skip it
1987 if (Cost >= RematerializationThreshold)
1990 // Remove value from the live set
1991 LiveValuesToBeDeleted.push_back(LiveValue);
1993 // Clone instructions and record them inside "Info" structure
1995 // Walk backwards to visit top-most instructions first
1996 std::reverse(ChainToBase.begin(), ChainToBase.end());
1998 // Utility function which clones all instructions from "ChainToBase"
1999 // and inserts them before "InsertBefore". Returns rematerialized value
2000 // which should be used after statepoint.
2001 auto rematerializeChain = [&ChainToBase](
2002 Instruction *InsertBefore, Value *RootOfChain, Value *AlternateLiveBase) {
2003 Instruction *LastClonedValue = nullptr;
2004 Instruction *LastValue = nullptr;
2005 for (Instruction *Instr: ChainToBase) {
2006 // Only GEP's and casts are suported as we need to be careful to not
2007 // introduce any new uses of pointers not in the liveset.
2008 // Note that it's fine to introduce new uses of pointers which were
2009 // otherwise not used after this statepoint.
2010 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2012 Instruction *ClonedValue = Instr->clone();
2013 ClonedValue->insertBefore(InsertBefore);
2014 ClonedValue->setName(Instr->getName() + ".remat");
2016 // If it is not first instruction in the chain then it uses previously
2017 // cloned value. We should update it to use cloned value.
2018 if (LastClonedValue) {
2020 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2022 for (auto OpValue : ClonedValue->operand_values()) {
2023 // Assert that cloned instruction does not use any instructions from
2024 // this chain other than LastClonedValue
2025 assert(!is_contained(ChainToBase, OpValue) &&
2026 "incorrect use in rematerialization chain");
2027 // Assert that the cloned instruction does not use the RootOfChain
2028 // or the AlternateLiveBase.
2029 assert(OpValue != RootOfChain && OpValue != AlternateLiveBase);
2033 // For the first instruction, replace the use of unrelocated base i.e.
2034 // RootOfChain/OrigRootPhi, with the corresponding PHI present in the
2035 // live set. They have been proved to be the same PHI nodes. Note
2036 // that the *only* use of the RootOfChain in the ChainToBase list is
2037 // the first Value in the list.
2038 if (RootOfChain != AlternateLiveBase)
2039 ClonedValue->replaceUsesOfWith(RootOfChain, AlternateLiveBase);
2042 LastClonedValue = ClonedValue;
2045 assert(LastClonedValue);
2046 return LastClonedValue;
2049 // Different cases for calls and invokes. For invokes we need to clone
2050 // instructions both on normal and unwind path.
2052 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2053 assert(InsertBefore);
2054 Instruction *RematerializedValue = rematerializeChain(
2055 InsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2056 Info.RematerializedValues[RematerializedValue] = LiveValue;
2058 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2060 Instruction *NormalInsertBefore =
2061 &*Invoke->getNormalDest()->getFirstInsertionPt();
2062 Instruction *UnwindInsertBefore =
2063 &*Invoke->getUnwindDest()->getFirstInsertionPt();
2065 Instruction *NormalRematerializedValue = rematerializeChain(
2066 NormalInsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2067 Instruction *UnwindRematerializedValue = rematerializeChain(
2068 UnwindInsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2070 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2071 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2075 // Remove rematerializaed values from the live set
2076 for (auto LiveValue: LiveValuesToBeDeleted) {
2077 Info.LiveSet.remove(LiveValue);
2081 static bool insertParsePoints(Function &F, DominatorTree &DT,
2082 TargetTransformInfo &TTI,
2083 SmallVectorImpl<CallSite> &ToUpdate) {
2085 // sanity check the input
2086 std::set<CallSite> Uniqued;
2087 Uniqued.insert(ToUpdate.begin(), ToUpdate.end());
2088 assert(Uniqued.size() == ToUpdate.size() && "no duplicates please!");
2090 for (CallSite CS : ToUpdate)
2091 assert(CS.getInstruction()->getFunction() == &F);
2094 // When inserting gc.relocates for invokes, we need to be able to insert at
2095 // the top of the successor blocks. See the comment on
2096 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2097 // may restructure the CFG.
2098 for (CallSite CS : ToUpdate) {
2101 auto *II = cast<InvokeInst>(CS.getInstruction());
2102 normalizeForInvokeSafepoint(II->getNormalDest(), II->getParent(), DT);
2103 normalizeForInvokeSafepoint(II->getUnwindDest(), II->getParent(), DT);
2106 // A list of dummy calls added to the IR to keep various values obviously
2107 // live in the IR. We'll remove all of these when done.
2108 SmallVector<CallInst *, 64> Holders;
2110 // Insert a dummy call with all of the arguments to the vm_state we'll need
2111 // for the actual safepoint insertion. This ensures reference arguments in
2112 // the deopt argument list are considered live through the safepoint (and
2113 // thus makes sure they get relocated.)
2114 for (CallSite CS : ToUpdate) {
2115 SmallVector<Value *, 64> DeoptValues;
2117 for (Value *Arg : GetDeoptBundleOperands(CS)) {
2118 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2119 "support for FCA unimplemented");
2120 if (isHandledGCPointerType(Arg->getType()))
2121 DeoptValues.push_back(Arg);
2124 insertUseHolderAfter(CS, DeoptValues, Holders);
2127 SmallVector<PartiallyConstructedSafepointRecord, 64> Records(ToUpdate.size());
2129 // A) Identify all gc pointers which are statically live at the given call
2131 findLiveReferences(F, DT, ToUpdate, Records);
2133 // B) Find the base pointers for each live pointer
2134 /* scope for caching */ {
2135 // Cache the 'defining value' relation used in the computation and
2136 // insertion of base phis and selects. This ensures that we don't insert
2137 // large numbers of duplicate base_phis.
2138 DefiningValueMapTy DVCache;
2140 for (size_t i = 0; i < Records.size(); i++) {
2141 PartiallyConstructedSafepointRecord &info = Records[i];
2142 findBasePointers(DT, DVCache, ToUpdate[i], info);
2144 } // end of cache scope
2146 // The base phi insertion logic (for any safepoint) may have inserted new
2147 // instructions which are now live at some safepoint. The simplest such
2150 // phi a <-- will be a new base_phi here
2151 // safepoint 1 <-- that needs to be live here
2155 // We insert some dummy calls after each safepoint to definitely hold live
2156 // the base pointers which were identified for that safepoint. We'll then
2157 // ask liveness for _every_ base inserted to see what is now live. Then we
2158 // remove the dummy calls.
2159 Holders.reserve(Holders.size() + Records.size());
2160 for (size_t i = 0; i < Records.size(); i++) {
2161 PartiallyConstructedSafepointRecord &Info = Records[i];
2163 SmallVector<Value *, 128> Bases;
2164 for (auto Pair : Info.PointerToBase)
2165 Bases.push_back(Pair.second);
2167 insertUseHolderAfter(ToUpdate[i], Bases, Holders);
2170 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2171 // need to rerun liveness. We may *also* have inserted new defs, but that's
2172 // not the key issue.
2173 recomputeLiveInValues(F, DT, ToUpdate, Records);
2175 if (PrintBasePointers) {
2176 for (auto &Info : Records) {
2177 errs() << "Base Pairs: (w/Relocation)\n";
2178 for (auto Pair : Info.PointerToBase) {
2179 errs() << " derived ";
2180 Pair.first->printAsOperand(errs(), false);
2182 Pair.second->printAsOperand(errs(), false);
2188 // It is possible that non-constant live variables have a constant base. For
2189 // example, a GEP with a variable offset from a global. In this case we can
2190 // remove it from the liveset. We already don't add constants to the liveset
2191 // because we assume they won't move at runtime and the GC doesn't need to be
2192 // informed about them. The same reasoning applies if the base is constant.
2193 // Note that the relocation placement code relies on this filtering for
2194 // correctness as it expects the base to be in the liveset, which isn't true
2195 // if the base is constant.
2196 for (auto &Info : Records)
2197 for (auto &BasePair : Info.PointerToBase)
2198 if (isa<Constant>(BasePair.second))
2199 Info.LiveSet.remove(BasePair.first);
2201 for (CallInst *CI : Holders)
2202 CI->eraseFromParent();
2206 // In order to reduce live set of statepoint we might choose to rematerialize
2207 // some values instead of relocating them. This is purely an optimization and
2208 // does not influence correctness.
2209 for (size_t i = 0; i < Records.size(); i++)
2210 rematerializeLiveValues(ToUpdate[i], Records[i], TTI);
2212 // We need this to safely RAUW and delete call or invoke return values that
2213 // may themselves be live over a statepoint. For details, please see usage in
2214 // makeStatepointExplicitImpl.
2215 std::vector<DeferredReplacement> Replacements;
2217 // Now run through and replace the existing statepoints with new ones with
2218 // the live variables listed. We do not yet update uses of the values being
2219 // relocated. We have references to live variables that need to
2220 // survive to the last iteration of this loop. (By construction, the
2221 // previous statepoint can not be a live variable, thus we can and remove
2222 // the old statepoint calls as we go.)
2223 for (size_t i = 0; i < Records.size(); i++)
2224 makeStatepointExplicit(DT, ToUpdate[i], Records[i], Replacements);
2226 ToUpdate.clear(); // prevent accident use of invalid CallSites
2228 for (auto &PR : Replacements)
2231 Replacements.clear();
2233 for (auto &Info : Records) {
2234 // These live sets may contain state Value pointers, since we replaced calls
2235 // with operand bundles with calls wrapped in gc.statepoint, and some of
2236 // those calls may have been def'ing live gc pointers. Clear these out to
2237 // avoid accidentally using them.
2239 // TODO: We should create a separate data structure that does not contain
2240 // these live sets, and migrate to using that data structure from this point
2242 Info.LiveSet.clear();
2243 Info.PointerToBase.clear();
2246 // Do all the fixups of the original live variables to their relocated selves
2247 SmallVector<Value *, 128> Live;
2248 for (size_t i = 0; i < Records.size(); i++) {
2249 PartiallyConstructedSafepointRecord &Info = Records[i];
2251 // We can't simply save the live set from the original insertion. One of
2252 // the live values might be the result of a call which needs a safepoint.
2253 // That Value* no longer exists and we need to use the new gc_result.
2254 // Thankfully, the live set is embedded in the statepoint (and updated), so
2255 // we just grab that.
2256 Statepoint Statepoint(Info.StatepointToken);
2257 Live.insert(Live.end(), Statepoint.gc_args_begin(),
2258 Statepoint.gc_args_end());
2260 // Do some basic sanity checks on our liveness results before performing
2261 // relocation. Relocation can and will turn mistakes in liveness results
2262 // into non-sensical code which is must harder to debug.
2263 // TODO: It would be nice to test consistency as well
2264 assert(DT.isReachableFromEntry(Info.StatepointToken->getParent()) &&
2265 "statepoint must be reachable or liveness is meaningless");
2266 for (Value *V : Statepoint.gc_args()) {
2267 if (!isa<Instruction>(V))
2268 // Non-instruction values trivial dominate all possible uses
2270 auto *LiveInst = cast<Instruction>(V);
2271 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2272 "unreachable values should never be live");
2273 assert(DT.dominates(LiveInst, Info.StatepointToken) &&
2274 "basic SSA liveness expectation violated by liveness analysis");
2278 unique_unsorted(Live);
2282 for (auto *Ptr : Live)
2283 assert(isHandledGCPointerType(Ptr->getType()) &&
2284 "must be a gc pointer type");
2287 relocationViaAlloca(F, DT, Live, Records);
2288 return !Records.empty();
2291 // Handles both return values and arguments for Functions and CallSites.
2292 template <typename AttrHolder>
2293 static void RemoveNonValidAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2296 if (AH.getDereferenceableBytes(Index))
2297 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2298 AH.getDereferenceableBytes(Index)));
2299 if (AH.getDereferenceableOrNullBytes(Index))
2300 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2301 AH.getDereferenceableOrNullBytes(Index)));
2302 if (AH.doesNotAlias(Index))
2303 R.addAttribute(Attribute::NoAlias);
2306 AH.setAttributes(AH.getAttributes().removeAttributes(
2307 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2311 RewriteStatepointsForGC::stripNonValidAttributesFromPrototype(Function &F) {
2312 LLVMContext &Ctx = F.getContext();
2314 for (Argument &A : F.args())
2315 if (isa<PointerType>(A.getType()))
2316 RemoveNonValidAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2318 if (isa<PointerType>(F.getReturnType()))
2319 RemoveNonValidAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2322 void RewriteStatepointsForGC::stripNonValidAttributesFromBody(Function &F) {
2326 LLVMContext &Ctx = F.getContext();
2327 MDBuilder Builder(Ctx);
2329 for (Instruction &I : instructions(F)) {
2330 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2331 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2332 bool IsImmutableTBAA =
2333 MD->getNumOperands() == 4 &&
2334 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2336 if (!IsImmutableTBAA)
2337 continue; // no work to do, MD_tbaa is already marked mutable
2339 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2340 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2342 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2344 MDNode *MutableTBAA =
2345 Builder.createTBAAStructTagNode(Base, Access, Offset);
2346 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2349 if (CallSite CS = CallSite(&I)) {
2350 for (int i = 0, e = CS.arg_size(); i != e; i++)
2351 if (isa<PointerType>(CS.getArgument(i)->getType()))
2352 RemoveNonValidAttrAtIndex(Ctx, CS, i + 1);
2353 if (isa<PointerType>(CS.getType()))
2354 RemoveNonValidAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2359 /// Returns true if this function should be rewritten by this pass. The main
2360 /// point of this function is as an extension point for custom logic.
2361 static bool shouldRewriteStatepointsIn(Function &F) {
2362 // TODO: This should check the GCStrategy
2364 const auto &FunctionGCName = F.getGC();
2365 const StringRef StatepointExampleName("statepoint-example");
2366 const StringRef CoreCLRName("coreclr");
2367 return (StatepointExampleName == FunctionGCName) ||
2368 (CoreCLRName == FunctionGCName);
2373 void RewriteStatepointsForGC::stripNonValidAttributes(Module &M) {
2375 assert(any_of(M, shouldRewriteStatepointsIn) && "precondition!");
2378 for (Function &F : M)
2379 stripNonValidAttributesFromPrototype(F);
2381 for (Function &F : M)
2382 stripNonValidAttributesFromBody(F);
2385 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2386 // Nothing to do for declarations.
2387 if (F.isDeclaration() || F.empty())
2390 // Policy choice says not to rewrite - the most common reason is that we're
2391 // compiling code without a GCStrategy.
2392 if (!shouldRewriteStatepointsIn(F))
2395 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2396 TargetTransformInfo &TTI =
2397 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2399 auto NeedsRewrite = [](Instruction &I) {
2400 if (ImmutableCallSite CS = ImmutableCallSite(&I))
2401 return !callsGCLeafFunction(CS) && !isStatepoint(CS);
2405 // Gather all the statepoints which need rewritten. Be careful to only
2406 // consider those in reachable code since we need to ask dominance queries
2407 // when rewriting. We'll delete the unreachable ones in a moment.
2408 SmallVector<CallSite, 64> ParsePointNeeded;
2409 bool HasUnreachableStatepoint = false;
2410 for (Instruction &I : instructions(F)) {
2411 // TODO: only the ones with the flag set!
2412 if (NeedsRewrite(I)) {
2413 if (DT.isReachableFromEntry(I.getParent()))
2414 ParsePointNeeded.push_back(CallSite(&I));
2416 HasUnreachableStatepoint = true;
2420 bool MadeChange = false;
2422 // Delete any unreachable statepoints so that we don't have unrewritten
2423 // statepoints surviving this pass. This makes testing easier and the
2424 // resulting IR less confusing to human readers. Rather than be fancy, we
2425 // just reuse a utility function which removes the unreachable blocks.
2426 if (HasUnreachableStatepoint)
2427 MadeChange |= removeUnreachableBlocks(F);
2429 // Return early if no work to do.
2430 if (ParsePointNeeded.empty())
2433 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2434 // These are created by LCSSA. They have the effect of increasing the size
2435 // of liveness sets for no good reason. It may be harder to do this post
2436 // insertion since relocations and base phis can confuse things.
2437 for (BasicBlock &BB : F)
2438 if (BB.getUniquePredecessor()) {
2440 FoldSingleEntryPHINodes(&BB);
2443 // Before we start introducing relocations, we want to tweak the IR a bit to
2444 // avoid unfortunate code generation effects. The main example is that we
2445 // want to try to make sure the comparison feeding a branch is after any
2446 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2447 // values feeding a branch after relocation. This is semantically correct,
2448 // but results in extra register pressure since both the pre-relocation and
2449 // post-relocation copies must be available in registers. For code without
2450 // relocations this is handled elsewhere, but teaching the scheduler to
2451 // reverse the transform we're about to do would be slightly complex.
2452 // Note: This may extend the live range of the inputs to the icmp and thus
2453 // increase the liveset of any statepoint we move over. This is profitable
2454 // as long as all statepoints are in rare blocks. If we had in-register
2455 // lowering for live values this would be a much safer transform.
2456 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2457 if (auto *BI = dyn_cast<BranchInst>(TI))
2458 if (BI->isConditional())
2459 return dyn_cast<Instruction>(BI->getCondition());
2460 // TODO: Extend this to handle switches
2463 for (BasicBlock &BB : F) {
2464 TerminatorInst *TI = BB.getTerminator();
2465 if (auto *Cond = getConditionInst(TI))
2466 // TODO: Handle more than just ICmps here. We should be able to move
2467 // most instructions without side effects or memory access.
2468 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2470 Cond->moveBefore(TI);
2474 MadeChange |= insertParsePoints(F, DT, TTI, ParsePointNeeded);
2478 // liveness computation via standard dataflow
2479 // -------------------------------------------------------------------
2481 // TODO: Consider using bitvectors for liveness, the set of potentially
2482 // interesting values should be small and easy to pre-compute.
2484 /// Compute the live-in set for the location rbegin starting from
2485 /// the live-out set of the basic block
2486 static void computeLiveInValues(BasicBlock::reverse_iterator Begin,
2487 BasicBlock::reverse_iterator End,
2488 SetVector<Value *> &LiveTmp) {
2489 for (auto &I : make_range(Begin, End)) {
2490 // KILL/Def - Remove this definition from LiveIn
2493 // Don't consider *uses* in PHI nodes, we handle their contribution to
2494 // predecessor blocks when we seed the LiveOut sets
2495 if (isa<PHINode>(I))
2498 // USE - Add to the LiveIn set for this instruction
2499 for (Value *V : I.operands()) {
2500 assert(!isUnhandledGCPointerType(V->getType()) &&
2501 "support for FCA unimplemented");
2502 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2503 // The choice to exclude all things constant here is slightly subtle.
2504 // There are two independent reasons:
2505 // - We assume that things which are constant (from LLVM's definition)
2506 // do not move at runtime. For example, the address of a global
2507 // variable is fixed, even though it's contents may not be.
2508 // - Second, we can't disallow arbitrary inttoptr constants even
2509 // if the language frontend does. Optimization passes are free to
2510 // locally exploit facts without respect to global reachability. This
2511 // can create sections of code which are dynamically unreachable and
2512 // contain just about anything. (see constants.ll in tests)
2519 static void computeLiveOutSeed(BasicBlock *BB, SetVector<Value *> &LiveTmp) {
2520 for (BasicBlock *Succ : successors(BB)) {
2521 for (auto &I : *Succ) {
2522 PHINode *PN = dyn_cast<PHINode>(&I);
2526 Value *V = PN->getIncomingValueForBlock(BB);
2527 assert(!isUnhandledGCPointerType(V->getType()) &&
2528 "support for FCA unimplemented");
2529 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V))
2535 static SetVector<Value *> computeKillSet(BasicBlock *BB) {
2536 SetVector<Value *> KillSet;
2537 for (Instruction &I : *BB)
2538 if (isHandledGCPointerType(I.getType()))
2544 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2545 /// sanity check for the liveness computation.
2546 static void checkBasicSSA(DominatorTree &DT, SetVector<Value *> &Live,
2547 TerminatorInst *TI, bool TermOkay = false) {
2548 for (Value *V : Live) {
2549 if (auto *I = dyn_cast<Instruction>(V)) {
2550 // The terminator can be a member of the LiveOut set. LLVM's definition
2551 // of instruction dominance states that V does not dominate itself. As
2552 // such, we need to special case this to allow it.
2553 if (TermOkay && TI == I)
2555 assert(DT.dominates(I, TI) &&
2556 "basic SSA liveness expectation violated by liveness analysis");
2561 /// Check that all the liveness sets used during the computation of liveness
2562 /// obey basic SSA properties. This is useful for finding cases where we miss
2564 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2566 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2567 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2568 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2572 static void computeLiveInValues(DominatorTree &DT, Function &F,
2573 GCPtrLivenessData &Data) {
2574 SmallSetVector<BasicBlock *, 32> Worklist;
2576 // Seed the liveness for each individual block
2577 for (BasicBlock &BB : F) {
2578 Data.KillSet[&BB] = computeKillSet(&BB);
2579 Data.LiveSet[&BB].clear();
2580 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2583 for (Value *Kill : Data.KillSet[&BB])
2584 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2587 Data.LiveOut[&BB] = SetVector<Value *>();
2588 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2589 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2590 Data.LiveIn[&BB].set_union(Data.LiveOut[&BB]);
2591 Data.LiveIn[&BB].set_subtract(Data.KillSet[&BB]);
2592 if (!Data.LiveIn[&BB].empty())
2593 Worklist.insert(pred_begin(&BB), pred_end(&BB));
2596 // Propagate that liveness until stable
2597 while (!Worklist.empty()) {
2598 BasicBlock *BB = Worklist.pop_back_val();
2600 // Compute our new liveout set, then exit early if it hasn't changed despite
2601 // the contribution of our successor.
2602 SetVector<Value *> LiveOut = Data.LiveOut[BB];
2603 const auto OldLiveOutSize = LiveOut.size();
2604 for (BasicBlock *Succ : successors(BB)) {
2605 assert(Data.LiveIn.count(Succ));
2606 LiveOut.set_union(Data.LiveIn[Succ]);
2608 // assert OutLiveOut is a subset of LiveOut
2609 if (OldLiveOutSize == LiveOut.size()) {
2610 // If the sets are the same size, then we didn't actually add anything
2611 // when unioning our successors LiveIn. Thus, the LiveIn of this block
2615 Data.LiveOut[BB] = LiveOut;
2617 // Apply the effects of this basic block
2618 SetVector<Value *> LiveTmp = LiveOut;
2619 LiveTmp.set_union(Data.LiveSet[BB]);
2620 LiveTmp.set_subtract(Data.KillSet[BB]);
2622 assert(Data.LiveIn.count(BB));
2623 const SetVector<Value *> &OldLiveIn = Data.LiveIn[BB];
2624 // assert: OldLiveIn is a subset of LiveTmp
2625 if (OldLiveIn.size() != LiveTmp.size()) {
2626 Data.LiveIn[BB] = LiveTmp;
2627 Worklist.insert(pred_begin(BB), pred_end(BB));
2629 } // while (!Worklist.empty())
2632 // Sanity check our output against SSA properties. This helps catch any
2633 // missing kills during the above iteration.
2634 for (BasicBlock &BB : F)
2635 checkBasicSSA(DT, Data, BB);
2639 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2640 StatepointLiveSetTy &Out) {
2642 BasicBlock *BB = Inst->getParent();
2644 // Note: The copy is intentional and required
2645 assert(Data.LiveOut.count(BB));
2646 SetVector<Value *> LiveOut = Data.LiveOut[BB];
2648 // We want to handle the statepoint itself oddly. It's
2649 // call result is not live (normal), nor are it's arguments
2650 // (unless they're used again later). This adjustment is
2651 // specifically what we need to relocate
2652 computeLiveInValues(BB->rbegin(), ++Inst->getIterator().getReverse(),
2654 LiveOut.remove(Inst);
2655 Out.insert(LiveOut.begin(), LiveOut.end());
2658 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2660 PartiallyConstructedSafepointRecord &Info) {
2661 Instruction *Inst = CS.getInstruction();
2662 StatepointLiveSetTy Updated;
2663 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2666 DenseSet<Value *> Bases;
2667 for (auto KVPair : Info.PointerToBase)
2668 Bases.insert(KVPair.second);
2671 // We may have base pointers which are now live that weren't before. We need
2672 // to update the PointerToBase structure to reflect this.
2673 for (auto V : Updated)
2674 if (Info.PointerToBase.insert({V, V}).second) {
2675 assert(Bases.count(V) && "Can't find base for unexpected live value!");
2680 for (auto V : Updated)
2681 assert(Info.PointerToBase.count(V) &&
2682 "Must be able to find base for live value!");
2685 // Remove any stale base mappings - this can happen since our liveness is
2686 // more precise then the one inherent in the base pointer analysis.
2687 DenseSet<Value *> ToErase;
2688 for (auto KVPair : Info.PointerToBase)
2689 if (!Updated.count(KVPair.first))
2690 ToErase.insert(KVPair.first);
2692 for (auto *V : ToErase)
2693 Info.PointerToBase.erase(V);
2696 for (auto KVPair : Info.PointerToBase)
2697 assert(Updated.count(KVPair.first) && "record for non-live value");
2700 Info.LiveSet = Updated;