1 //===- RewriteStatepointsForGC.cpp - Make GC relocations explicit ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Rewrite an existing set of gc.statepoints such that they make potential
11 // relocations performed by the garbage collector explicit in the IR.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Pass.h"
16 #include "llvm/Analysis/CFG.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/TargetTransformInfo.h"
19 #include "llvm/ADT/SetOperations.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/ADT/DenseSet.h"
22 #include "llvm/ADT/SetVector.h"
23 #include "llvm/ADT/StringRef.h"
24 #include "llvm/ADT/MapVector.h"
25 #include "llvm/IR/BasicBlock.h"
26 #include "llvm/IR/CallSite.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/InstIterator.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Statepoint.h"
37 #include "llvm/IR/Value.h"
38 #include "llvm/IR/Verifier.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Transforms/Scalar.h"
42 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 #include "llvm/Transforms/Utils/Cloning.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
47 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
51 // Print the liveset found at the insert location
52 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
54 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden,
56 // Print out the base pointers for debugging
57 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden,
60 // Cost threshold measuring when it is profitable to rematerialize value instead
62 static cl::opt<unsigned>
63 RematerializationThreshold("spp-rematerialization-threshold", cl::Hidden,
67 static bool ClobberNonLive = true;
69 static bool ClobberNonLive = false;
71 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
72 cl::location(ClobberNonLive),
75 static cl::opt<bool> UseDeoptBundles("rs4gc-use-deopt-bundles", cl::Hidden,
78 AllowStatepointWithNoDeoptInfo("rs4gc-allow-statepoint-with-no-deopt-info",
79 cl::Hidden, cl::init(true));
82 struct RewriteStatepointsForGC : public ModulePass {
83 static char ID; // Pass identification, replacement for typeid
85 RewriteStatepointsForGC() : ModulePass(ID) {
86 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
88 bool runOnFunction(Function &F);
89 bool runOnModule(Module &M) override {
92 Changed |= runOnFunction(F);
95 // stripNonValidAttributes asserts that shouldRewriteStatepointsIn
96 // returns true for at least one function in the module. Since at least
97 // one function changed, we know that the precondition is satisfied.
98 stripNonValidAttributes(M);
104 void getAnalysisUsage(AnalysisUsage &AU) const override {
105 // We add and rewrite a bunch of instructions, but don't really do much
106 // else. We could in theory preserve a lot more analyses here.
107 AU.addRequired<DominatorTreeWrapperPass>();
108 AU.addRequired<TargetTransformInfoWrapperPass>();
111 /// The IR fed into RewriteStatepointsForGC may have had attributes implying
112 /// dereferenceability that are no longer valid/correct after
113 /// RewriteStatepointsForGC has run. This is because semantically, after
114 /// RewriteStatepointsForGC runs, all calls to gc.statepoint "free" the entire
115 /// heap. stripNonValidAttributes (conservatively) restores correctness
116 /// by erasing all attributes in the module that externally imply
117 /// dereferenceability.
118 /// Similar reasoning also applies to the noalias attributes. gc.statepoint
119 /// can touch the entire heap including noalias objects.
120 void stripNonValidAttributes(Module &M);
122 // Helpers for stripNonValidAttributes
123 void stripNonValidAttributesFromBody(Function &F);
124 void stripNonValidAttributesFromPrototype(Function &F);
128 char RewriteStatepointsForGC::ID = 0;
130 ModulePass *llvm::createRewriteStatepointsForGCPass() {
131 return new RewriteStatepointsForGC();
134 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
135 "Make relocations explicit at statepoints", false, false)
136 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
137 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
138 "Make relocations explicit at statepoints", false, false)
141 struct GCPtrLivenessData {
142 /// Values defined in this block.
143 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet;
144 /// Values used in this block (and thus live); does not included values
145 /// killed within this block.
146 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet;
148 /// Values live into this basic block (i.e. used by any
149 /// instruction in this basic block or ones reachable from here)
150 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn;
152 /// Values live out of this basic block (i.e. live into
153 /// any successor block)
154 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut;
157 // The type of the internal cache used inside the findBasePointers family
158 // of functions. From the callers perspective, this is an opaque type and
159 // should not be inspected.
161 // In the actual implementation this caches two relations:
162 // - The base relation itself (i.e. this pointer is based on that one)
163 // - The base defining value relation (i.e. before base_phi insertion)
164 // Generally, after the execution of a full findBasePointer call, only the
165 // base relation will remain. Internally, we add a mixture of the two
166 // types, then update all the second type to the first type
167 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
168 typedef DenseSet<Value *> StatepointLiveSetTy;
169 typedef DenseMap<AssertingVH<Instruction>, AssertingVH<Value>>
170 RematerializedValueMapTy;
172 struct PartiallyConstructedSafepointRecord {
173 /// The set of values known to be live across this safepoint
174 StatepointLiveSetTy LiveSet;
176 /// Mapping from live pointers to a base-defining-value
177 DenseMap<Value *, Value *> PointerToBase;
179 /// The *new* gc.statepoint instruction itself. This produces the token
180 /// that normal path gc.relocates and the gc.result are tied to.
181 Instruction *StatepointToken;
183 /// Instruction to which exceptional gc relocates are attached
184 /// Makes it easier to iterate through them during relocationViaAlloca.
185 Instruction *UnwindToken;
187 /// Record live values we are rematerialized instead of relocating.
188 /// They are not included into 'LiveSet' field.
189 /// Maps rematerialized copy to it's original value.
190 RematerializedValueMapTy RematerializedValues;
194 static ArrayRef<Use> GetDeoptBundleOperands(ImmutableCallSite CS) {
195 assert(UseDeoptBundles && "Should not be called otherwise!");
197 Optional<OperandBundleUse> DeoptBundle = CS.getOperandBundle("deopt");
199 if (!DeoptBundle.hasValue()) {
200 assert(AllowStatepointWithNoDeoptInfo &&
201 "Found non-leaf call without deopt info!");
205 return DeoptBundle.getValue().Inputs;
208 /// Compute the live-in set for every basic block in the function
209 static void computeLiveInValues(DominatorTree &DT, Function &F,
210 GCPtrLivenessData &Data);
212 /// Given results from the dataflow liveness computation, find the set of live
213 /// Values at a particular instruction.
214 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
215 StatepointLiveSetTy &out);
217 // TODO: Once we can get to the GCStrategy, this becomes
218 // Optional<bool> isGCManagedPointer(const Type *Ty) const override {
220 static bool isGCPointerType(Type *T) {
221 if (auto *PT = dyn_cast<PointerType>(T))
222 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
223 // GC managed heap. We know that a pointer into this heap needs to be
224 // updated and that no other pointer does.
225 return (1 == PT->getAddressSpace());
229 // Return true if this type is one which a) is a gc pointer or contains a GC
230 // pointer and b) is of a type this code expects to encounter as a live value.
231 // (The insertion code will assert that a type which matches (a) and not (b)
232 // is not encountered.)
233 static bool isHandledGCPointerType(Type *T) {
234 // We fully support gc pointers
235 if (isGCPointerType(T))
237 // We partially support vectors of gc pointers. The code will assert if it
238 // can't handle something.
239 if (auto VT = dyn_cast<VectorType>(T))
240 if (isGCPointerType(VT->getElementType()))
246 /// Returns true if this type contains a gc pointer whether we know how to
247 /// handle that type or not.
248 static bool containsGCPtrType(Type *Ty) {
249 if (isGCPointerType(Ty))
251 if (VectorType *VT = dyn_cast<VectorType>(Ty))
252 return isGCPointerType(VT->getScalarType());
253 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
254 return containsGCPtrType(AT->getElementType());
255 if (StructType *ST = dyn_cast<StructType>(Ty))
256 return std::any_of(ST->subtypes().begin(), ST->subtypes().end(),
261 // Returns true if this is a type which a) is a gc pointer or contains a GC
262 // pointer and b) is of a type which the code doesn't expect (i.e. first class
263 // aggregates). Used to trip assertions.
264 static bool isUnhandledGCPointerType(Type *Ty) {
265 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
269 static bool order_by_name(Value *a, Value *b) {
270 if (a->hasName() && b->hasName()) {
271 return -1 == a->getName().compare(b->getName());
272 } else if (a->hasName() && !b->hasName()) {
274 } else if (!a->hasName() && b->hasName()) {
277 // Better than nothing, but not stable
282 // Return the name of the value suffixed with the provided value, or if the
283 // value didn't have a name, the default value specified.
284 static std::string suffixed_name_or(Value *V, StringRef Suffix,
285 StringRef DefaultName) {
286 return V->hasName() ? (V->getName() + Suffix).str() : DefaultName.str();
289 // Conservatively identifies any definitions which might be live at the
290 // given instruction. The analysis is performed immediately before the
291 // given instruction. Values defined by that instruction are not considered
292 // live. Values used by that instruction are considered live.
293 static void analyzeParsePointLiveness(
294 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
295 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
296 Instruction *inst = CS.getInstruction();
298 StatepointLiveSetTy LiveSet;
299 findLiveSetAtInst(inst, OriginalLivenessData, LiveSet);
302 // Note: This output is used by several of the test cases
303 // The order of elements in a set is not stable, put them in a vec and sort
305 SmallVector<Value *, 64> Temp;
306 Temp.insert(Temp.end(), LiveSet.begin(), LiveSet.end());
307 std::sort(Temp.begin(), Temp.end(), order_by_name);
308 errs() << "Live Variables:\n";
309 for (Value *V : Temp)
310 dbgs() << " " << V->getName() << " " << *V << "\n";
312 if (PrintLiveSetSize) {
313 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
314 errs() << "Number live values: " << LiveSet.size() << "\n";
316 result.LiveSet = LiveSet;
319 static bool isKnownBaseResult(Value *V);
321 /// A single base defining value - An immediate base defining value for an
322 /// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'.
323 /// For instructions which have multiple pointer [vector] inputs or that
324 /// transition between vector and scalar types, there is no immediate base
325 /// defining value. The 'base defining value' for 'Def' is the transitive
326 /// closure of this relation stopping at the first instruction which has no
327 /// immediate base defining value. The b.d.v. might itself be a base pointer,
328 /// but it can also be an arbitrary derived pointer.
329 struct BaseDefiningValueResult {
330 /// Contains the value which is the base defining value.
332 /// True if the base defining value is also known to be an actual base
334 const bool IsKnownBase;
335 BaseDefiningValueResult(Value *BDV, bool IsKnownBase)
336 : BDV(BDV), IsKnownBase(IsKnownBase) {
338 // Check consistency between new and old means of checking whether a BDV is
340 bool MustBeBase = isKnownBaseResult(BDV);
341 assert(!MustBeBase || MustBeBase == IsKnownBase);
347 static BaseDefiningValueResult findBaseDefiningValue(Value *I);
349 /// Return a base defining value for the 'Index' element of the given vector
350 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
351 /// 'I'. As an optimization, this method will try to determine when the
352 /// element is known to already be a base pointer. If this can be established,
353 /// the second value in the returned pair will be true. Note that either a
354 /// vector or a pointer typed value can be returned. For the former, the
355 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
356 /// If the later, the return pointer is a BDV (or possibly a base) for the
357 /// particular element in 'I'.
358 static BaseDefiningValueResult
359 findBaseDefiningValueOfVector(Value *I) {
360 assert(I->getType()->isVectorTy() &&
361 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
362 "Illegal to ask for the base pointer of a non-pointer type");
364 // Each case parallels findBaseDefiningValue below, see that code for
365 // detailed motivation.
367 if (isa<Argument>(I))
368 // An incoming argument to the function is a base pointer
369 return BaseDefiningValueResult(I, true);
371 // We shouldn't see the address of a global as a vector value?
372 assert(!isa<GlobalVariable>(I) &&
373 "unexpected global variable found in base of vector");
375 // inlining could possibly introduce phi node that contains
376 // undef if callee has multiple returns
377 if (isa<UndefValue>(I))
378 // utterly meaningless, but useful for dealing with partially optimized
380 return BaseDefiningValueResult(I, true);
382 // Due to inheritance, this must be _after_ the global variable and undef
384 if (Constant *Con = dyn_cast<Constant>(I)) {
385 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
386 "order of checks wrong!");
387 assert(Con->isNullValue() && "null is the only case which makes sense");
388 return BaseDefiningValueResult(Con, true);
391 if (isa<LoadInst>(I))
392 return BaseDefiningValueResult(I, true);
394 if (isa<InsertElementInst>(I))
395 // We don't know whether this vector contains entirely base pointers or
396 // not. To be conservatively correct, we treat it as a BDV and will
397 // duplicate code as needed to construct a parallel vector of bases.
398 return BaseDefiningValueResult(I, false);
400 if (isa<ShuffleVectorInst>(I))
401 // We don't know whether this vector contains entirely base pointers or
402 // not. To be conservatively correct, we treat it as a BDV and will
403 // duplicate code as needed to construct a parallel vector of bases.
404 // TODO: There a number of local optimizations which could be applied here
405 // for particular sufflevector patterns.
406 return BaseDefiningValueResult(I, false);
408 // A PHI or Select is a base defining value. The outer findBasePointer
409 // algorithm is responsible for constructing a base value for this BDV.
410 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
411 "unknown vector instruction - no base found for vector element");
412 return BaseDefiningValueResult(I, false);
415 /// Helper function for findBasePointer - Will return a value which either a)
416 /// defines the base pointer for the input, b) blocks the simple search
417 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
418 /// from pointer to vector type or back.
419 static BaseDefiningValueResult findBaseDefiningValue(Value *I) {
420 if (I->getType()->isVectorTy())
421 return findBaseDefiningValueOfVector(I);
423 assert(I->getType()->isPointerTy() &&
424 "Illegal to ask for the base pointer of a non-pointer type");
426 if (isa<Argument>(I))
427 // An incoming argument to the function is a base pointer
428 // We should have never reached here if this argument isn't an gc value
429 return BaseDefiningValueResult(I, true);
431 if (isa<Constant>(I))
432 // We assume that objects with a constant base (e.g. a global) can't move
433 // and don't need to be reported to the collector because they are always
434 // live. All constants have constant bases. Besides global references, all
435 // kinds of constants (e.g. undef, constant expressions, null pointers) can
436 // be introduced by the inliner or the optimizer, especially on dynamically
437 // dead paths. See e.g. test4 in constants.ll.
438 return BaseDefiningValueResult(I, true);
440 if (CastInst *CI = dyn_cast<CastInst>(I)) {
441 Value *Def = CI->stripPointerCasts();
442 // If stripping pointer casts changes the address space there is an
443 // addrspacecast in between.
444 assert(cast<PointerType>(Def->getType())->getAddressSpace() ==
445 cast<PointerType>(CI->getType())->getAddressSpace() &&
446 "unsupported addrspacecast");
447 // If we find a cast instruction here, it means we've found a cast which is
448 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
449 // handle int->ptr conversion.
450 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
451 return findBaseDefiningValue(Def);
454 if (isa<LoadInst>(I))
455 // The value loaded is an gc base itself
456 return BaseDefiningValueResult(I, true);
459 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
460 // The base of this GEP is the base
461 return findBaseDefiningValue(GEP->getPointerOperand());
463 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
464 switch (II->getIntrinsicID()) {
466 // fall through to general call handling
468 case Intrinsic::experimental_gc_statepoint:
469 llvm_unreachable("statepoints don't produce pointers");
470 case Intrinsic::experimental_gc_relocate: {
471 // Rerunning safepoint insertion after safepoints are already
472 // inserted is not supported. It could probably be made to work,
473 // but why are you doing this? There's no good reason.
474 llvm_unreachable("repeat safepoint insertion is not supported");
476 case Intrinsic::gcroot:
477 // Currently, this mechanism hasn't been extended to work with gcroot.
478 // There's no reason it couldn't be, but I haven't thought about the
479 // implications much.
481 "interaction with the gcroot mechanism is not supported");
484 // We assume that functions in the source language only return base
485 // pointers. This should probably be generalized via attributes to support
486 // both source language and internal functions.
487 if (isa<CallInst>(I) || isa<InvokeInst>(I))
488 return BaseDefiningValueResult(I, true);
490 // I have absolutely no idea how to implement this part yet. It's not
491 // necessarily hard, I just haven't really looked at it yet.
492 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
494 if (isa<AtomicCmpXchgInst>(I))
495 // A CAS is effectively a atomic store and load combined under a
496 // predicate. From the perspective of base pointers, we just treat it
498 return BaseDefiningValueResult(I, true);
500 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
501 "binary ops which don't apply to pointers");
503 // The aggregate ops. Aggregates can either be in the heap or on the
504 // stack, but in either case, this is simply a field load. As a result,
505 // this is a defining definition of the base just like a load is.
506 if (isa<ExtractValueInst>(I))
507 return BaseDefiningValueResult(I, true);
509 // We should never see an insert vector since that would require we be
510 // tracing back a struct value not a pointer value.
511 assert(!isa<InsertValueInst>(I) &&
512 "Base pointer for a struct is meaningless");
514 // An extractelement produces a base result exactly when it's input does.
515 // We may need to insert a parallel instruction to extract the appropriate
516 // element out of the base vector corresponding to the input. Given this,
517 // it's analogous to the phi and select case even though it's not a merge.
518 if (isa<ExtractElementInst>(I))
519 // Note: There a lot of obvious peephole cases here. This are deliberately
520 // handled after the main base pointer inference algorithm to make writing
521 // test cases to exercise that code easier.
522 return BaseDefiningValueResult(I, false);
524 // The last two cases here don't return a base pointer. Instead, they
525 // return a value which dynamically selects from among several base
526 // derived pointers (each with it's own base potentially). It's the job of
527 // the caller to resolve these.
528 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
529 "missing instruction case in findBaseDefiningValing");
530 return BaseDefiningValueResult(I, false);
533 /// Returns the base defining value for this value.
534 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
535 Value *&Cached = Cache[I];
537 Cached = findBaseDefiningValue(I).BDV;
538 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
539 << Cached->getName() << "\n");
541 assert(Cache[I] != nullptr);
545 /// Return a base pointer for this value if known. Otherwise, return it's
546 /// base defining value.
547 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
548 Value *Def = findBaseDefiningValueCached(I, Cache);
549 auto Found = Cache.find(Def);
550 if (Found != Cache.end()) {
551 // Either a base-of relation, or a self reference. Caller must check.
552 return Found->second;
554 // Only a BDV available
558 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
559 /// is it known to be a base pointer? Or do we need to continue searching.
560 static bool isKnownBaseResult(Value *V) {
561 if (!isa<PHINode>(V) && !isa<SelectInst>(V) &&
562 !isa<ExtractElementInst>(V) && !isa<InsertElementInst>(V) &&
563 !isa<ShuffleVectorInst>(V)) {
564 // no recursion possible
567 if (isa<Instruction>(V) &&
568 cast<Instruction>(V)->getMetadata("is_base_value")) {
569 // This is a previously inserted base phi or select. We know
570 // that this is a base value.
574 // We need to keep searching
579 /// Models the state of a single base defining value in the findBasePointer
580 /// algorithm for determining where a new instruction is needed to propagate
581 /// the base of this BDV.
584 enum Status { Unknown, Base, Conflict };
586 BDVState(Status s, Value *b = nullptr) : status(s), base(b) {
587 assert(status != Base || b);
589 explicit BDVState(Value *b) : status(Base), base(b) {}
590 BDVState() : status(Unknown), base(nullptr) {}
592 Status getStatus() const { return status; }
593 Value *getBase() const { return base; }
595 bool isBase() const { return getStatus() == Base; }
596 bool isUnknown() const { return getStatus() == Unknown; }
597 bool isConflict() const { return getStatus() == Conflict; }
599 bool operator==(const BDVState &other) const {
600 return base == other.base && status == other.status;
603 bool operator!=(const BDVState &other) const { return !(*this == other); }
606 void dump() const { print(dbgs()); dbgs() << '\n'; }
608 void print(raw_ostream &OS) const {
620 OS << " (" << base << " - "
621 << (base ? base->getName() : "nullptr") << "): ";
626 AssertingVH<Value> base; // non null only if status == base
631 static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
638 // Values of type BDVState form a lattice, and this is a helper
639 // class that implementes the meet operation. The meat of the meet
640 // operation is implemented in MeetBDVStates::pureMeet
641 class MeetBDVStates {
643 /// Initializes the currentResult to the TOP state so that if can be met with
644 /// any other state to produce that state.
647 // Destructively meet the current result with the given BDVState
648 void meetWith(BDVState otherState) {
649 currentResult = meet(otherState, currentResult);
652 BDVState getResult() const { return currentResult; }
655 BDVState currentResult;
657 /// Perform a meet operation on two elements of the BDVState lattice.
658 static BDVState meet(BDVState LHS, BDVState RHS) {
659 assert((pureMeet(LHS, RHS) == pureMeet(RHS, LHS)) &&
660 "math is wrong: meet does not commute!");
661 BDVState Result = pureMeet(LHS, RHS);
662 DEBUG(dbgs() << "meet of " << LHS << " with " << RHS
663 << " produced " << Result << "\n");
667 static BDVState pureMeet(const BDVState &stateA, const BDVState &stateB) {
668 switch (stateA.getStatus()) {
669 case BDVState::Unknown:
673 assert(stateA.getBase() && "can't be null");
674 if (stateB.isUnknown())
677 if (stateB.isBase()) {
678 if (stateA.getBase() == stateB.getBase()) {
679 assert(stateA == stateB && "equality broken!");
682 return BDVState(BDVState::Conflict);
684 assert(stateB.isConflict() && "only three states!");
685 return BDVState(BDVState::Conflict);
687 case BDVState::Conflict:
690 llvm_unreachable("only three states!");
696 /// For a given value or instruction, figure out what base ptr it's derived
697 /// from. For gc objects, this is simply itself. On success, returns a value
698 /// which is the base pointer. (This is reliable and can be used for
699 /// relocation.) On failure, returns nullptr.
700 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
701 Value *def = findBaseOrBDV(I, cache);
703 if (isKnownBaseResult(def)) {
707 // Here's the rough algorithm:
708 // - For every SSA value, construct a mapping to either an actual base
709 // pointer or a PHI which obscures the base pointer.
710 // - Construct a mapping from PHI to unknown TOP state. Use an
711 // optimistic algorithm to propagate base pointer information. Lattice
716 // When algorithm terminates, all PHIs will either have a single concrete
717 // base or be in a conflict state.
718 // - For every conflict, insert a dummy PHI node without arguments. Add
719 // these to the base[Instruction] = BasePtr mapping. For every
720 // non-conflict, add the actual base.
721 // - For every conflict, add arguments for the base[a] of each input
724 // Note: A simpler form of this would be to add the conflict form of all
725 // PHIs without running the optimistic algorithm. This would be
726 // analogous to pessimistic data flow and would likely lead to an
727 // overall worse solution.
730 auto isExpectedBDVType = [](Value *BDV) {
731 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) ||
732 isa<ExtractElementInst>(BDV) || isa<InsertElementInst>(BDV);
736 // Once populated, will contain a mapping from each potentially non-base BDV
737 // to a lattice value (described above) which corresponds to that BDV.
738 // We use the order of insertion (DFS over the def/use graph) to provide a
739 // stable deterministic ordering for visiting DenseMaps (which are unordered)
740 // below. This is important for deterministic compilation.
741 MapVector<Value *, BDVState> States;
743 // Recursively fill in all base defining values reachable from the initial
744 // one for which we don't already know a definite base value for
746 SmallVector<Value*, 16> Worklist;
747 Worklist.push_back(def);
748 States.insert(std::make_pair(def, BDVState()));
749 while (!Worklist.empty()) {
750 Value *Current = Worklist.pop_back_val();
751 assert(!isKnownBaseResult(Current) && "why did it get added?");
753 auto visitIncomingValue = [&](Value *InVal) {
754 Value *Base = findBaseOrBDV(InVal, cache);
755 if (isKnownBaseResult(Base))
756 // Known bases won't need new instructions introduced and can be
759 assert(isExpectedBDVType(Base) && "the only non-base values "
760 "we see should be base defining values");
761 if (States.insert(std::make_pair(Base, BDVState())).second)
762 Worklist.push_back(Base);
764 if (PHINode *Phi = dyn_cast<PHINode>(Current)) {
765 for (Value *InVal : Phi->incoming_values())
766 visitIncomingValue(InVal);
767 } else if (SelectInst *Sel = dyn_cast<SelectInst>(Current)) {
768 visitIncomingValue(Sel->getTrueValue());
769 visitIncomingValue(Sel->getFalseValue());
770 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
771 visitIncomingValue(EE->getVectorOperand());
772 } else if (auto *IE = dyn_cast<InsertElementInst>(Current)) {
773 visitIncomingValue(IE->getOperand(0)); // vector operand
774 visitIncomingValue(IE->getOperand(1)); // scalar operand
776 // There is one known class of instructions we know we don't handle.
777 assert(isa<ShuffleVectorInst>(Current));
778 llvm_unreachable("unimplemented instruction case");
784 DEBUG(dbgs() << "States after initialization:\n");
785 for (auto Pair : States) {
786 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
790 // Return a phi state for a base defining value. We'll generate a new
791 // base state for known bases and expect to find a cached state otherwise.
792 auto getStateForBDV = [&](Value *baseValue) {
793 if (isKnownBaseResult(baseValue))
794 return BDVState(baseValue);
795 auto I = States.find(baseValue);
796 assert(I != States.end() && "lookup failed!");
800 bool progress = true;
803 const size_t oldSize = States.size();
806 // We're only changing values in this loop, thus safe to keep iterators.
807 // Since this is computing a fixed point, the order of visit does not
808 // effect the result. TODO: We could use a worklist here and make this run
810 for (auto Pair : States) {
811 Value *BDV = Pair.first;
812 assert(!isKnownBaseResult(BDV) && "why did it get added?");
814 // Given an input value for the current instruction, return a BDVState
815 // instance which represents the BDV of that value.
816 auto getStateForInput = [&](Value *V) mutable {
817 Value *BDV = findBaseOrBDV(V, cache);
818 return getStateForBDV(BDV);
821 MeetBDVStates calculateMeet;
822 if (SelectInst *select = dyn_cast<SelectInst>(BDV)) {
823 calculateMeet.meetWith(getStateForInput(select->getTrueValue()));
824 calculateMeet.meetWith(getStateForInput(select->getFalseValue()));
825 } else if (PHINode *Phi = dyn_cast<PHINode>(BDV)) {
826 for (Value *Val : Phi->incoming_values())
827 calculateMeet.meetWith(getStateForInput(Val));
828 } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) {
829 // The 'meet' for an extractelement is slightly trivial, but it's still
830 // useful in that it drives us to conflict if our input is.
831 calculateMeet.meetWith(getStateForInput(EE->getVectorOperand()));
833 // Given there's a inherent type mismatch between the operands, will
834 // *always* produce Conflict.
835 auto *IE = cast<InsertElementInst>(BDV);
836 calculateMeet.meetWith(getStateForInput(IE->getOperand(0)));
837 calculateMeet.meetWith(getStateForInput(IE->getOperand(1)));
840 BDVState oldState = States[BDV];
841 BDVState newState = calculateMeet.getResult();
842 if (oldState != newState) {
844 States[BDV] = newState;
848 assert(oldSize == States.size() &&
849 "fixed point shouldn't be adding any new nodes to state");
853 DEBUG(dbgs() << "States after meet iteration:\n");
854 for (auto Pair : States) {
855 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
859 // Insert Phis for all conflicts
860 // TODO: adjust naming patterns to avoid this order of iteration dependency
861 for (auto Pair : States) {
862 Instruction *I = cast<Instruction>(Pair.first);
863 BDVState State = Pair.second;
864 assert(!isKnownBaseResult(I) && "why did it get added?");
865 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
867 // extractelement instructions are a bit special in that we may need to
868 // insert an extract even when we know an exact base for the instruction.
869 // The problem is that we need to convert from a vector base to a scalar
870 // base for the particular indice we're interested in.
871 if (State.isBase() && isa<ExtractElementInst>(I) &&
872 isa<VectorType>(State.getBase()->getType())) {
873 auto *EE = cast<ExtractElementInst>(I);
874 // TODO: In many cases, the new instruction is just EE itself. We should
875 // exploit this, but can't do it here since it would break the invariant
876 // about the BDV not being known to be a base.
877 auto *BaseInst = ExtractElementInst::Create(State.getBase(),
878 EE->getIndexOperand(),
880 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
881 States[I] = BDVState(BDVState::Base, BaseInst);
884 // Since we're joining a vector and scalar base, they can never be the
885 // same. As a result, we should always see insert element having reached
886 // the conflict state.
887 if (isa<InsertElementInst>(I)) {
888 assert(State.isConflict());
891 if (!State.isConflict())
894 /// Create and insert a new instruction which will represent the base of
895 /// the given instruction 'I'.
896 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
897 if (isa<PHINode>(I)) {
898 BasicBlock *BB = I->getParent();
899 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
900 assert(NumPreds > 0 && "how did we reach here");
901 std::string Name = suffixed_name_or(I, ".base", "base_phi");
902 return PHINode::Create(I->getType(), NumPreds, Name, I);
903 } else if (SelectInst *Sel = dyn_cast<SelectInst>(I)) {
904 // The undef will be replaced later
905 UndefValue *Undef = UndefValue::get(Sel->getType());
906 std::string Name = suffixed_name_or(I, ".base", "base_select");
907 return SelectInst::Create(Sel->getCondition(), Undef,
909 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
910 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
911 std::string Name = suffixed_name_or(I, ".base", "base_ee");
912 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
915 auto *IE = cast<InsertElementInst>(I);
916 UndefValue *VecUndef = UndefValue::get(IE->getOperand(0)->getType());
917 UndefValue *ScalarUndef = UndefValue::get(IE->getOperand(1)->getType());
918 std::string Name = suffixed_name_or(I, ".base", "base_ie");
919 return InsertElementInst::Create(VecUndef, ScalarUndef,
920 IE->getOperand(2), Name, IE);
924 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
925 // Add metadata marking this as a base value
926 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
927 States[I] = BDVState(BDVState::Conflict, BaseInst);
930 // Returns a instruction which produces the base pointer for a given
931 // instruction. The instruction is assumed to be an input to one of the BDVs
932 // seen in the inference algorithm above. As such, we must either already
933 // know it's base defining value is a base, or have inserted a new
934 // instruction to propagate the base of it's BDV and have entered that newly
935 // introduced instruction into the state table. In either case, we are
936 // assured to be able to determine an instruction which produces it's base
938 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
939 Value *BDV = findBaseOrBDV(Input, cache);
940 Value *Base = nullptr;
941 if (isKnownBaseResult(BDV)) {
944 // Either conflict or base.
945 assert(States.count(BDV));
946 Base = States[BDV].getBase();
948 assert(Base && "can't be null");
949 // The cast is needed since base traversal may strip away bitcasts
950 if (Base->getType() != Input->getType() &&
952 Base = new BitCastInst(Base, Input->getType(), "cast",
958 // Fixup all the inputs of the new PHIs. Visit order needs to be
959 // deterministic and predictable because we're naming newly created
961 for (auto Pair : States) {
962 Instruction *BDV = cast<Instruction>(Pair.first);
963 BDVState State = Pair.second;
965 assert(!isKnownBaseResult(BDV) && "why did it get added?");
966 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
967 if (!State.isConflict())
970 if (PHINode *basephi = dyn_cast<PHINode>(State.getBase())) {
971 PHINode *phi = cast<PHINode>(BDV);
972 unsigned NumPHIValues = phi->getNumIncomingValues();
973 for (unsigned i = 0; i < NumPHIValues; i++) {
974 Value *InVal = phi->getIncomingValue(i);
975 BasicBlock *InBB = phi->getIncomingBlock(i);
977 // If we've already seen InBB, add the same incoming value
978 // we added for it earlier. The IR verifier requires phi
979 // nodes with multiple entries from the same basic block
980 // to have the same incoming value for each of those
981 // entries. If we don't do this check here and basephi
982 // has a different type than base, we'll end up adding two
983 // bitcasts (and hence two distinct values) as incoming
984 // values for the same basic block.
986 int blockIndex = basephi->getBasicBlockIndex(InBB);
987 if (blockIndex != -1) {
988 Value *oldBase = basephi->getIncomingValue(blockIndex);
989 basephi->addIncoming(oldBase, InBB);
992 Value *Base = getBaseForInput(InVal, nullptr);
993 // In essence this assert states: the only way two
994 // values incoming from the same basic block may be
995 // different is by being different bitcasts of the same
996 // value. A cleanup that remains TODO is changing
997 // findBaseOrBDV to return an llvm::Value of the correct
998 // type (and still remain pure). This will remove the
999 // need to add bitcasts.
1000 assert(Base->stripPointerCasts() == oldBase->stripPointerCasts() &&
1001 "sanity -- findBaseOrBDV should be pure!");
1006 // Find the instruction which produces the base for each input. We may
1007 // need to insert a bitcast in the incoming block.
1008 // TODO: Need to split critical edges if insertion is needed
1009 Value *Base = getBaseForInput(InVal, InBB->getTerminator());
1010 basephi->addIncoming(Base, InBB);
1012 assert(basephi->getNumIncomingValues() == NumPHIValues);
1013 } else if (SelectInst *BaseSel = dyn_cast<SelectInst>(State.getBase())) {
1014 SelectInst *Sel = cast<SelectInst>(BDV);
1015 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
1016 // something more safe and less hacky.
1017 for (int i = 1; i <= 2; i++) {
1018 Value *InVal = Sel->getOperand(i);
1019 // Find the instruction which produces the base for each input. We may
1020 // need to insert a bitcast.
1021 Value *Base = getBaseForInput(InVal, BaseSel);
1022 BaseSel->setOperand(i, Base);
1024 } else if (auto *BaseEE = dyn_cast<ExtractElementInst>(State.getBase())) {
1025 Value *InVal = cast<ExtractElementInst>(BDV)->getVectorOperand();
1026 // Find the instruction which produces the base for each input. We may
1027 // need to insert a bitcast.
1028 Value *Base = getBaseForInput(InVal, BaseEE);
1029 BaseEE->setOperand(0, Base);
1031 auto *BaseIE = cast<InsertElementInst>(State.getBase());
1032 auto *BdvIE = cast<InsertElementInst>(BDV);
1033 auto UpdateOperand = [&](int OperandIdx) {
1034 Value *InVal = BdvIE->getOperand(OperandIdx);
1035 Value *Base = getBaseForInput(InVal, BaseIE);
1036 BaseIE->setOperand(OperandIdx, Base);
1038 UpdateOperand(0); // vector operand
1039 UpdateOperand(1); // scalar operand
1044 // Now that we're done with the algorithm, see if we can optimize the
1045 // results slightly by reducing the number of new instructions needed.
1046 // Arguably, this should be integrated into the algorithm above, but
1047 // doing as a post process step is easier to reason about for the moment.
1048 DenseMap<Value *, Value *> ReverseMap;
1049 SmallPtrSet<Instruction *, 16> NewInsts;
1050 SmallSetVector<AssertingVH<Instruction>, 16> Worklist;
1051 // Note: We need to visit the states in a deterministic order. We uses the
1052 // Keys we sorted above for this purpose. Note that we are papering over a
1053 // bigger problem with the algorithm above - it's visit order is not
1054 // deterministic. A larger change is needed to fix this.
1055 for (auto Pair : States) {
1056 auto *BDV = Pair.first;
1057 auto State = Pair.second;
1058 Value *Base = State.getBase();
1059 assert(BDV && Base);
1060 assert(!isKnownBaseResult(BDV) && "why did it get added?");
1061 assert(isKnownBaseResult(Base) &&
1062 "must be something we 'know' is a base pointer");
1063 if (!State.isConflict())
1066 ReverseMap[Base] = BDV;
1067 if (auto *BaseI = dyn_cast<Instruction>(Base)) {
1068 NewInsts.insert(BaseI);
1069 Worklist.insert(BaseI);
1072 auto ReplaceBaseInstWith = [&](Value *BDV, Instruction *BaseI,
1073 Value *Replacement) {
1074 // Add users which are new instructions (excluding self references)
1075 for (User *U : BaseI->users())
1076 if (auto *UI = dyn_cast<Instruction>(U))
1077 if (NewInsts.count(UI) && UI != BaseI)
1078 Worklist.insert(UI);
1079 // Then do the actual replacement
1080 NewInsts.erase(BaseI);
1081 ReverseMap.erase(BaseI);
1082 BaseI->replaceAllUsesWith(Replacement);
1083 assert(States.count(BDV));
1084 assert(States[BDV].isConflict() && States[BDV].getBase() == BaseI);
1085 States[BDV] = BDVState(BDVState::Conflict, Replacement);
1086 BaseI->eraseFromParent();
1088 const DataLayout &DL = cast<Instruction>(def)->getModule()->getDataLayout();
1089 while (!Worklist.empty()) {
1090 Instruction *BaseI = Worklist.pop_back_val();
1091 assert(NewInsts.count(BaseI));
1092 Value *Bdv = ReverseMap[BaseI];
1093 if (auto *BdvI = dyn_cast<Instruction>(Bdv))
1094 if (BaseI->isIdenticalTo(BdvI)) {
1095 DEBUG(dbgs() << "Identical Base: " << *BaseI << "\n");
1096 ReplaceBaseInstWith(Bdv, BaseI, Bdv);
1099 if (Value *V = SimplifyInstruction(BaseI, DL)) {
1100 DEBUG(dbgs() << "Base " << *BaseI << " simplified to " << *V << "\n");
1101 ReplaceBaseInstWith(Bdv, BaseI, V);
1106 // Cache all of our results so we can cheaply reuse them
1107 // NOTE: This is actually two caches: one of the base defining value
1108 // relation and one of the base pointer relation! FIXME
1109 for (auto Pair : States) {
1110 auto *BDV = Pair.first;
1111 Value *base = Pair.second.getBase();
1112 assert(BDV && base);
1114 std::string fromstr = cache.count(BDV) ? cache[BDV]->getName() : "none";
1115 DEBUG(dbgs() << "Updating base value cache"
1116 << " for: " << BDV->getName()
1117 << " from: " << fromstr
1118 << " to: " << base->getName() << "\n");
1120 if (cache.count(BDV)) {
1121 // Once we transition from the BDV relation being store in the cache to
1122 // the base relation being stored, it must be stable
1123 assert((!isKnownBaseResult(cache[BDV]) || cache[BDV] == base) &&
1124 "base relation should be stable");
1128 assert(cache.count(def));
1132 // For a set of live pointers (base and/or derived), identify the base
1133 // pointer of the object which they are derived from. This routine will
1134 // mutate the IR graph as needed to make the 'base' pointer live at the
1135 // definition site of 'derived'. This ensures that any use of 'derived' can
1136 // also use 'base'. This may involve the insertion of a number of
1137 // additional PHI nodes.
1139 // preconditions: live is a set of pointer type Values
1141 // side effects: may insert PHI nodes into the existing CFG, will preserve
1142 // CFG, will not remove or mutate any existing nodes
1144 // post condition: PointerToBase contains one (derived, base) pair for every
1145 // pointer in live. Note that derived can be equal to base if the original
1146 // pointer was a base pointer.
1148 findBasePointers(const StatepointLiveSetTy &live,
1149 DenseMap<Value *, Value *> &PointerToBase,
1150 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1151 // For the naming of values inserted to be deterministic - which makes for
1152 // much cleaner and more stable tests - we need to assign an order to the
1153 // live values. DenseSets do not provide a deterministic order across runs.
1154 SmallVector<Value *, 64> Temp;
1155 Temp.insert(Temp.end(), live.begin(), live.end());
1156 std::sort(Temp.begin(), Temp.end(), order_by_name);
1157 for (Value *ptr : Temp) {
1158 Value *base = findBasePointer(ptr, DVCache);
1159 assert(base && "failed to find base pointer");
1160 PointerToBase[ptr] = base;
1161 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1162 DT->dominates(cast<Instruction>(base)->getParent(),
1163 cast<Instruction>(ptr)->getParent())) &&
1164 "The base we found better dominate the derived pointer");
1166 // If you see this trip and like to live really dangerously, the code should
1167 // be correct, just with idioms the verifier can't handle. You can try
1168 // disabling the verifier at your own substantial risk.
1169 assert(!isa<ConstantPointerNull>(base) &&
1170 "the relocation code needs adjustment to handle the relocation of "
1171 "a null pointer constant without causing false positives in the "
1172 "safepoint ir verifier.");
1176 /// Find the required based pointers (and adjust the live set) for the given
1178 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1180 PartiallyConstructedSafepointRecord &result) {
1181 DenseMap<Value *, Value *> PointerToBase;
1182 findBasePointers(result.LiveSet, PointerToBase, &DT, DVCache);
1184 if (PrintBasePointers) {
1185 // Note: Need to print these in a stable order since this is checked in
1187 errs() << "Base Pairs (w/o Relocation):\n";
1188 SmallVector<Value *, 64> Temp;
1189 Temp.reserve(PointerToBase.size());
1190 for (auto Pair : PointerToBase) {
1191 Temp.push_back(Pair.first);
1193 std::sort(Temp.begin(), Temp.end(), order_by_name);
1194 for (Value *Ptr : Temp) {
1195 Value *Base = PointerToBase[Ptr];
1196 errs() << " derived ";
1197 Ptr->printAsOperand(errs(), false);
1199 Base->printAsOperand(errs(), false);
1204 result.PointerToBase = PointerToBase;
1207 /// Given an updated version of the dataflow liveness results, update the
1208 /// liveset and base pointer maps for the call site CS.
1209 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1211 PartiallyConstructedSafepointRecord &result);
1213 static void recomputeLiveInValues(
1214 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1215 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1216 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1217 // again. The old values are still live and will help it stabilize quickly.
1218 GCPtrLivenessData RevisedLivenessData;
1219 computeLiveInValues(DT, F, RevisedLivenessData);
1220 for (size_t i = 0; i < records.size(); i++) {
1221 struct PartiallyConstructedSafepointRecord &info = records[i];
1222 const CallSite &CS = toUpdate[i];
1223 recomputeLiveInValues(RevisedLivenessData, CS, info);
1227 // When inserting gc.relocate and gc.result calls, we need to ensure there are
1228 // no uses of the original value / return value between the gc.statepoint and
1229 // the gc.relocate / gc.result call. One case which can arise is a phi node
1230 // starting one of the successor blocks. We also need to be able to insert the
1231 // gc.relocates only on the path which goes through the statepoint. We might
1232 // need to split an edge to make this possible.
1234 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1235 DominatorTree &DT) {
1236 BasicBlock *Ret = BB;
1237 if (!BB->getUniquePredecessor())
1238 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1240 // Now that 'Ret' has unique predecessor we can safely remove all phi nodes
1242 FoldSingleEntryPHINodes(Ret);
1243 assert(!isa<PHINode>(Ret->begin()) &&
1244 "All PHI nodes should have been removed!");
1246 // At this point, we can safely insert a gc.relocate or gc.result as the first
1247 // instruction in Ret if needed.
1251 // Create new attribute set containing only attributes which can be transferred
1252 // from original call to the safepoint.
1253 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1256 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1257 unsigned Index = AS.getSlotIndex(Slot);
1259 if (Index == AttributeSet::ReturnIndex ||
1260 Index == AttributeSet::FunctionIndex) {
1262 for (Attribute Attr : make_range(AS.begin(Slot), AS.end(Slot))) {
1264 // Do not allow certain attributes - just skip them
1265 // Safepoint can not be read only or read none.
1266 if (Attr.hasAttribute(Attribute::ReadNone) ||
1267 Attr.hasAttribute(Attribute::ReadOnly))
1270 // These attributes control the generation of the gc.statepoint call /
1271 // invoke itself; and once the gc.statepoint is in place, they're of no
1273 if (Attr.hasAttribute("statepoint-num-patch-bytes") ||
1274 Attr.hasAttribute("statepoint-id"))
1277 Ret = Ret.addAttributes(
1278 AS.getContext(), Index,
1279 AttributeSet::get(AS.getContext(), Index, AttrBuilder(Attr)));
1283 // Just skip parameter attributes for now
1289 /// Helper function to place all gc relocates necessary for the given
1292 /// liveVariables - list of variables to be relocated.
1293 /// liveStart - index of the first live variable.
1294 /// basePtrs - base pointers.
1295 /// statepointToken - statepoint instruction to which relocates should be
1297 /// Builder - Llvm IR builder to be used to construct new calls.
1298 static void CreateGCRelocates(ArrayRef<Value *> LiveVariables,
1299 const int LiveStart,
1300 ArrayRef<Value *> BasePtrs,
1301 Instruction *StatepointToken,
1302 IRBuilder<> Builder) {
1303 if (LiveVariables.empty())
1306 auto FindIndex = [](ArrayRef<Value *> LiveVec, Value *Val) {
1307 auto ValIt = std::find(LiveVec.begin(), LiveVec.end(), Val);
1308 assert(ValIt != LiveVec.end() && "Val not found in LiveVec!");
1309 size_t Index = std::distance(LiveVec.begin(), ValIt);
1310 assert(Index < LiveVec.size() && "Bug in std::find?");
1314 // All gc_relocate are set to i8 addrspace(1)* type. We originally generated
1315 // unique declarations for each pointer type, but this proved problematic
1316 // because the intrinsic mangling code is incomplete and fragile. Since
1317 // we're moving towards a single unified pointer type anyways, we can just
1318 // cast everything to an i8* of the right address space. A bitcast is added
1319 // later to convert gc_relocate to the actual value's type.
1320 Module *M = StatepointToken->getModule();
1321 auto AS = cast<PointerType>(LiveVariables[0]->getType())->getAddressSpace();
1322 Type *Types[] = {Type::getInt8PtrTy(M->getContext(), AS)};
1323 Value *GCRelocateDecl =
1324 Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, Types);
1326 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1327 // Generate the gc.relocate call and save the result
1329 Builder.getInt32(LiveStart + FindIndex(LiveVariables, BasePtrs[i]));
1330 Value *LiveIdx = Builder.getInt32(LiveStart + i);
1332 // only specify a debug name if we can give a useful one
1333 CallInst *Reloc = Builder.CreateCall(
1334 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1335 suffixed_name_or(LiveVariables[i], ".relocated", ""));
1336 // Trick CodeGen into thinking there are lots of free registers at this
1338 Reloc->setCallingConv(CallingConv::Cold);
1344 /// This struct is used to defer RAUWs and `eraseFromParent` s. Using this
1345 /// avoids having to worry about keeping around dangling pointers to Values.
1346 class DeferredReplacement {
1347 AssertingVH<Instruction> Old;
1348 AssertingVH<Instruction> New;
1351 explicit DeferredReplacement(Instruction *Old, Instruction *New) :
1352 Old(Old), New(New) {
1353 assert(Old != New && "Not allowed!");
1356 /// Does the task represented by this instance.
1357 void doReplacement() {
1358 Instruction *OldI = Old;
1359 Instruction *NewI = New;
1361 assert(OldI != NewI && "Disallowed at construction?!");
1367 OldI->replaceAllUsesWith(NewI);
1368 OldI->eraseFromParent();
1374 makeStatepointExplicitImpl(const CallSite CS, /* to replace */
1375 const SmallVectorImpl<Value *> &BasePtrs,
1376 const SmallVectorImpl<Value *> &LiveVariables,
1377 PartiallyConstructedSafepointRecord &Result,
1378 std::vector<DeferredReplacement> &Replacements) {
1379 assert(BasePtrs.size() == LiveVariables.size());
1380 assert((UseDeoptBundles || isStatepoint(CS)) &&
1381 "This method expects to be rewriting a statepoint");
1383 // Then go ahead and use the builder do actually do the inserts. We insert
1384 // immediately before the previous instruction under the assumption that all
1385 // arguments will be available here. We can't insert afterwards since we may
1386 // be replacing a terminator.
1387 Instruction *InsertBefore = CS.getInstruction();
1388 IRBuilder<> Builder(InsertBefore);
1390 ArrayRef<Value *> GCArgs(LiveVariables);
1391 uint64_t StatepointID = 0xABCDEF00;
1392 uint32_t NumPatchBytes = 0;
1393 uint32_t Flags = uint32_t(StatepointFlags::None);
1395 ArrayRef<Use> CallArgs;
1396 ArrayRef<Use> DeoptArgs;
1397 ArrayRef<Use> TransitionArgs;
1399 Value *CallTarget = nullptr;
1401 if (UseDeoptBundles) {
1402 CallArgs = {CS.arg_begin(), CS.arg_end()};
1403 DeoptArgs = GetDeoptBundleOperands(CS);
1404 // TODO: we don't fill in TransitionArgs or Flags in this branch, but we
1405 // could have an operand bundle for that too.
1406 AttributeSet OriginalAttrs = CS.getAttributes();
1408 Attribute AttrID = OriginalAttrs.getAttribute(AttributeSet::FunctionIndex,
1410 if (AttrID.isStringAttribute())
1411 AttrID.getValueAsString().getAsInteger(10, StatepointID);
1413 Attribute AttrNumPatchBytes = OriginalAttrs.getAttribute(
1414 AttributeSet::FunctionIndex, "statepoint-num-patch-bytes");
1415 if (AttrNumPatchBytes.isStringAttribute())
1416 AttrNumPatchBytes.getValueAsString().getAsInteger(10, NumPatchBytes);
1418 CallTarget = CS.getCalledValue();
1420 // This branch will be gone soon, and we will soon only support the
1421 // UseDeoptBundles == true configuration.
1422 Statepoint OldSP(CS);
1423 StatepointID = OldSP.getID();
1424 NumPatchBytes = OldSP.getNumPatchBytes();
1425 Flags = OldSP.getFlags();
1427 CallArgs = {OldSP.arg_begin(), OldSP.arg_end()};
1428 DeoptArgs = {OldSP.vm_state_begin(), OldSP.vm_state_end()};
1429 TransitionArgs = {OldSP.gc_transition_args_begin(),
1430 OldSP.gc_transition_args_end()};
1431 CallTarget = OldSP.getCalledValue();
1434 // Create the statepoint given all the arguments
1435 Instruction *Token = nullptr;
1436 AttributeSet ReturnAttrs;
1438 CallInst *ToReplace = cast<CallInst>(CS.getInstruction());
1439 CallInst *Call = Builder.CreateGCStatepointCall(
1440 StatepointID, NumPatchBytes, CallTarget, Flags, CallArgs,
1441 TransitionArgs, DeoptArgs, GCArgs, "safepoint_token");
1443 Call->setTailCall(ToReplace->isTailCall());
1444 Call->setCallingConv(ToReplace->getCallingConv());
1446 // Currently we will fail on parameter attributes and on certain
1447 // function attributes.
1448 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes());
1449 // In case if we can handle this set of attributes - set up function attrs
1450 // directly on statepoint and return attrs later for gc_result intrinsic.
1451 Call->setAttributes(NewAttrs.getFnAttributes());
1452 ReturnAttrs = NewAttrs.getRetAttributes();
1456 // Put the following gc_result and gc_relocate calls immediately after the
1457 // the old call (which we're about to delete)
1458 assert(ToReplace->getNextNode() && "Not a terminator, must have next!");
1459 Builder.SetInsertPoint(ToReplace->getNextNode());
1460 Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc());
1462 InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction());
1464 // Insert the new invoke into the old block. We'll remove the old one in a
1465 // moment at which point this will become the new terminator for the
1467 InvokeInst *Invoke = Builder.CreateGCStatepointInvoke(
1468 StatepointID, NumPatchBytes, CallTarget, ToReplace->getNormalDest(),
1469 ToReplace->getUnwindDest(), Flags, CallArgs, TransitionArgs, DeoptArgs,
1470 GCArgs, "statepoint_token");
1472 Invoke->setCallingConv(ToReplace->getCallingConv());
1474 // Currently we will fail on parameter attributes and on certain
1475 // function attributes.
1476 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes());
1477 // In case if we can handle this set of attributes - set up function attrs
1478 // directly on statepoint and return attrs later for gc_result intrinsic.
1479 Invoke->setAttributes(NewAttrs.getFnAttributes());
1480 ReturnAttrs = NewAttrs.getRetAttributes();
1484 // Generate gc relocates in exceptional path
1485 BasicBlock *UnwindBlock = ToReplace->getUnwindDest();
1486 assert(!isa<PHINode>(UnwindBlock->begin()) &&
1487 UnwindBlock->getUniquePredecessor() &&
1488 "can't safely insert in this block!");
1490 Builder.SetInsertPoint(&*UnwindBlock->getFirstInsertionPt());
1491 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1493 // Attach exceptional gc relocates to the landingpad.
1494 Instruction *ExceptionalToken = UnwindBlock->getLandingPadInst();
1495 Result.UnwindToken = ExceptionalToken;
1497 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1498 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, ExceptionalToken,
1501 // Generate gc relocates and returns for normal block
1502 BasicBlock *NormalDest = ToReplace->getNormalDest();
1503 assert(!isa<PHINode>(NormalDest->begin()) &&
1504 NormalDest->getUniquePredecessor() &&
1505 "can't safely insert in this block!");
1507 Builder.SetInsertPoint(&*NormalDest->getFirstInsertionPt());
1509 // gc relocates will be generated later as if it were regular call
1512 assert(Token && "Should be set in one of the above branches!");
1514 if (UseDeoptBundles) {
1515 Token->setName("statepoint_token");
1516 if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
1518 CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "";
1519 CallInst *GCResult = Builder.CreateGCResult(Token, CS.getType(), Name);
1520 GCResult->setAttributes(CS.getAttributes().getRetAttributes());
1522 // We cannot RAUW or delete CS.getInstruction() because it could be in the
1523 // live set of some other safepoint, in which case that safepoint's
1524 // PartiallyConstructedSafepointRecord will hold a raw pointer to this
1525 // llvm::Instruction. Instead, we defer the replacement and deletion to
1526 // after the live sets have been made explicit in the IR, and we no longer
1527 // have raw pointers to worry about.
1528 Replacements.emplace_back(CS.getInstruction(), GCResult);
1530 Replacements.emplace_back(CS.getInstruction(), nullptr);
1533 assert(!CS.getInstruction()->hasNUsesOrMore(2) &&
1534 "only valid use before rewrite is gc.result");
1535 assert(!CS.getInstruction()->hasOneUse() ||
1536 isGCResult(cast<Instruction>(*CS.getInstruction()->user_begin())));
1538 // Take the name of the original statepoint token if there was one.
1539 Token->takeName(CS.getInstruction());
1541 // Update the gc.result of the original statepoint (if any) to use the newly
1542 // inserted statepoint. This is safe to do here since the token can't be
1543 // considered a live reference.
1544 CS.getInstruction()->replaceAllUsesWith(Token);
1545 CS.getInstruction()->eraseFromParent();
1548 Result.StatepointToken = Token;
1550 // Second, create a gc.relocate for every live variable
1551 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1552 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, Token, Builder);
1556 struct NameOrdering {
1560 bool operator()(NameOrdering const &a, NameOrdering const &b) {
1561 return -1 == a.Derived->getName().compare(b.Derived->getName());
1566 static void StabilizeOrder(SmallVectorImpl<Value *> &BaseVec,
1567 SmallVectorImpl<Value *> &LiveVec) {
1568 assert(BaseVec.size() == LiveVec.size());
1570 SmallVector<NameOrdering, 64> Temp;
1571 for (size_t i = 0; i < BaseVec.size(); i++) {
1573 v.Base = BaseVec[i];
1574 v.Derived = LiveVec[i];
1578 std::sort(Temp.begin(), Temp.end(), NameOrdering());
1579 for (size_t i = 0; i < BaseVec.size(); i++) {
1580 BaseVec[i] = Temp[i].Base;
1581 LiveVec[i] = Temp[i].Derived;
1585 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1586 // which make the relocations happening at this safepoint explicit.
1588 // WARNING: Does not do any fixup to adjust users of the original live
1589 // values. That's the callers responsibility.
1591 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS,
1592 PartiallyConstructedSafepointRecord &Result,
1593 std::vector<DeferredReplacement> &Replacements) {
1594 const auto &LiveSet = Result.LiveSet;
1595 const auto &PointerToBase = Result.PointerToBase;
1597 // Convert to vector for efficient cross referencing.
1598 SmallVector<Value *, 64> BaseVec, LiveVec;
1599 LiveVec.reserve(LiveSet.size());
1600 BaseVec.reserve(LiveSet.size());
1601 for (Value *L : LiveSet) {
1602 LiveVec.push_back(L);
1603 assert(PointerToBase.count(L));
1604 Value *Base = PointerToBase.find(L)->second;
1605 BaseVec.push_back(Base);
1607 assert(LiveVec.size() == BaseVec.size());
1609 // To make the output IR slightly more stable (for use in diffs), ensure a
1610 // fixed order of the values in the safepoint (by sorting the value name).
1611 // The order is otherwise meaningless.
1612 StabilizeOrder(BaseVec, LiveVec);
1614 // Do the actual rewriting and delete the old statepoint
1615 makeStatepointExplicitImpl(CS, BaseVec, LiveVec, Result, Replacements);
1618 // Helper function for the relocationViaAlloca.
1620 // It receives iterator to the statepoint gc relocates and emits a store to the
1621 // assigned location (via allocaMap) for the each one of them. It adds the
1622 // visited values into the visitedLiveValues set, which we will later use them
1623 // for sanity checking.
1625 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1626 DenseMap<Value *, Value *> &AllocaMap,
1627 DenseSet<Value *> &VisitedLiveValues) {
1629 for (User *U : GCRelocs) {
1630 GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U);
1634 Value *OriginalValue = const_cast<Value *>(Relocate->getDerivedPtr());
1635 assert(AllocaMap.count(OriginalValue));
1636 Value *Alloca = AllocaMap[OriginalValue];
1638 // Emit store into the related alloca
1639 // All gc_relocates are i8 addrspace(1)* typed, and it must be bitcasted to
1640 // the correct type according to alloca.
1641 assert(Relocate->getNextNode() &&
1642 "Should always have one since it's not a terminator");
1643 IRBuilder<> Builder(Relocate->getNextNode());
1644 Value *CastedRelocatedValue =
1645 Builder.CreateBitCast(Relocate,
1646 cast<AllocaInst>(Alloca)->getAllocatedType(),
1647 suffixed_name_or(Relocate, ".casted", ""));
1649 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1650 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1653 VisitedLiveValues.insert(OriginalValue);
1658 // Helper function for the "relocationViaAlloca". Similar to the
1659 // "insertRelocationStores" but works for rematerialized values.
1661 insertRematerializationStores(
1662 RematerializedValueMapTy RematerializedValues,
1663 DenseMap<Value *, Value *> &AllocaMap,
1664 DenseSet<Value *> &VisitedLiveValues) {
1666 for (auto RematerializedValuePair: RematerializedValues) {
1667 Instruction *RematerializedValue = RematerializedValuePair.first;
1668 Value *OriginalValue = RematerializedValuePair.second;
1670 assert(AllocaMap.count(OriginalValue) &&
1671 "Can not find alloca for rematerialized value");
1672 Value *Alloca = AllocaMap[OriginalValue];
1674 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1675 Store->insertAfter(RematerializedValue);
1678 VisitedLiveValues.insert(OriginalValue);
1683 /// Do all the relocation update via allocas and mem2reg
1684 static void relocationViaAlloca(
1685 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1686 ArrayRef<PartiallyConstructedSafepointRecord> Records) {
1688 // record initial number of (static) allocas; we'll check we have the same
1689 // number when we get done.
1690 int InitialAllocaNum = 0;
1691 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1693 if (isa<AllocaInst>(*I))
1697 // TODO-PERF: change data structures, reserve
1698 DenseMap<Value *, Value *> AllocaMap;
1699 SmallVector<AllocaInst *, 200> PromotableAllocas;
1700 // Used later to chack that we have enough allocas to store all values
1701 std::size_t NumRematerializedValues = 0;
1702 PromotableAllocas.reserve(Live.size());
1704 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1705 // "PromotableAllocas"
1706 auto emitAllocaFor = [&](Value *LiveValue) {
1707 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1708 F.getEntryBlock().getFirstNonPHI());
1709 AllocaMap[LiveValue] = Alloca;
1710 PromotableAllocas.push_back(Alloca);
1713 // Emit alloca for each live gc pointer
1714 for (Value *V : Live)
1717 // Emit allocas for rematerialized values
1718 for (const auto &Info : Records)
1719 for (auto RematerializedValuePair : Info.RematerializedValues) {
1720 Value *OriginalValue = RematerializedValuePair.second;
1721 if (AllocaMap.count(OriginalValue) != 0)
1724 emitAllocaFor(OriginalValue);
1725 ++NumRematerializedValues;
1728 // The next two loops are part of the same conceptual operation. We need to
1729 // insert a store to the alloca after the original def and at each
1730 // redefinition. We need to insert a load before each use. These are split
1731 // into distinct loops for performance reasons.
1733 // Update gc pointer after each statepoint: either store a relocated value or
1734 // null (if no relocated value was found for this gc pointer and it is not a
1735 // gc_result). This must happen before we update the statepoint with load of
1736 // alloca otherwise we lose the link between statepoint and old def.
1737 for (const auto &Info : Records) {
1738 Value *Statepoint = Info.StatepointToken;
1740 // This will be used for consistency check
1741 DenseSet<Value *> VisitedLiveValues;
1743 // Insert stores for normal statepoint gc relocates
1744 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1746 // In case if it was invoke statepoint
1747 // we will insert stores for exceptional path gc relocates.
1748 if (isa<InvokeInst>(Statepoint)) {
1749 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1753 // Do similar thing with rematerialized values
1754 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1757 if (ClobberNonLive) {
1758 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1759 // the gc.statepoint. This will turn some subtle GC problems into
1760 // slightly easier to debug SEGVs. Note that on large IR files with
1761 // lots of gc.statepoints this is extremely costly both memory and time
1763 SmallVector<AllocaInst *, 64> ToClobber;
1764 for (auto Pair : AllocaMap) {
1765 Value *Def = Pair.first;
1766 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1768 // This value was relocated
1769 if (VisitedLiveValues.count(Def)) {
1772 ToClobber.push_back(Alloca);
1775 auto InsertClobbersAt = [&](Instruction *IP) {
1776 for (auto *AI : ToClobber) {
1777 auto AIType = cast<PointerType>(AI->getType());
1778 auto PT = cast<PointerType>(AIType->getElementType());
1779 Constant *CPN = ConstantPointerNull::get(PT);
1780 StoreInst *Store = new StoreInst(CPN, AI);
1781 Store->insertBefore(IP);
1785 // Insert the clobbering stores. These may get intermixed with the
1786 // gc.results and gc.relocates, but that's fine.
1787 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1788 InsertClobbersAt(&*II->getNormalDest()->getFirstInsertionPt());
1789 InsertClobbersAt(&*II->getUnwindDest()->getFirstInsertionPt());
1791 InsertClobbersAt(cast<Instruction>(Statepoint)->getNextNode());
1796 // Update use with load allocas and add store for gc_relocated.
1797 for (auto Pair : AllocaMap) {
1798 Value *Def = Pair.first;
1799 Value *Alloca = Pair.second;
1801 // We pre-record the uses of allocas so that we dont have to worry about
1802 // later update that changes the user information..
1804 SmallVector<Instruction *, 20> Uses;
1805 // PERF: trade a linear scan for repeated reallocation
1806 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1807 for (User *U : Def->users()) {
1808 if (!isa<ConstantExpr>(U)) {
1809 // If the def has a ConstantExpr use, then the def is either a
1810 // ConstantExpr use itself or null. In either case
1811 // (recursively in the first, directly in the second), the oop
1812 // it is ultimately dependent on is null and this particular
1813 // use does not need to be fixed up.
1814 Uses.push_back(cast<Instruction>(U));
1818 std::sort(Uses.begin(), Uses.end());
1819 auto Last = std::unique(Uses.begin(), Uses.end());
1820 Uses.erase(Last, Uses.end());
1822 for (Instruction *Use : Uses) {
1823 if (isa<PHINode>(Use)) {
1824 PHINode *Phi = cast<PHINode>(Use);
1825 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1826 if (Def == Phi->getIncomingValue(i)) {
1827 LoadInst *Load = new LoadInst(
1828 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1829 Phi->setIncomingValue(i, Load);
1833 LoadInst *Load = new LoadInst(Alloca, "", Use);
1834 Use->replaceUsesOfWith(Def, Load);
1838 // Emit store for the initial gc value. Store must be inserted after load,
1839 // otherwise store will be in alloca's use list and an extra load will be
1840 // inserted before it.
1841 StoreInst *Store = new StoreInst(Def, Alloca);
1842 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1843 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1844 // InvokeInst is a TerminatorInst so the store need to be inserted
1845 // into its normal destination block.
1846 BasicBlock *NormalDest = Invoke->getNormalDest();
1847 Store->insertBefore(NormalDest->getFirstNonPHI());
1849 assert(!Inst->isTerminator() &&
1850 "The only TerminatorInst that can produce a value is "
1851 "InvokeInst which is handled above.");
1852 Store->insertAfter(Inst);
1855 assert(isa<Argument>(Def));
1856 Store->insertAfter(cast<Instruction>(Alloca));
1860 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1861 "we must have the same allocas with lives");
1862 if (!PromotableAllocas.empty()) {
1863 // Apply mem2reg to promote alloca to SSA
1864 PromoteMemToReg(PromotableAllocas, DT);
1868 for (auto &I : F.getEntryBlock())
1869 if (isa<AllocaInst>(I))
1871 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1875 /// Implement a unique function which doesn't require we sort the input
1876 /// vector. Doing so has the effect of changing the output of a couple of
1877 /// tests in ways which make them less useful in testing fused safepoints.
1878 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1879 SmallSet<T, 8> Seen;
1880 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1881 return !Seen.insert(V).second;
1885 /// Insert holders so that each Value is obviously live through the entire
1886 /// lifetime of the call.
1887 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1888 SmallVectorImpl<CallInst *> &Holders) {
1890 // No values to hold live, might as well not insert the empty holder
1893 Module *M = CS.getInstruction()->getModule();
1894 // Use a dummy vararg function to actually hold the values live
1895 Function *Func = cast<Function>(M->getOrInsertFunction(
1896 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1898 // For call safepoints insert dummy calls right after safepoint
1899 Holders.push_back(CallInst::Create(Func, Values, "",
1900 &*++CS.getInstruction()->getIterator()));
1903 // For invoke safepooints insert dummy calls both in normal and
1904 // exceptional destination blocks
1905 auto *II = cast<InvokeInst>(CS.getInstruction());
1906 Holders.push_back(CallInst::Create(
1907 Func, Values, "", &*II->getNormalDest()->getFirstInsertionPt()));
1908 Holders.push_back(CallInst::Create(
1909 Func, Values, "", &*II->getUnwindDest()->getFirstInsertionPt()));
1912 static void findLiveReferences(
1913 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1914 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1915 GCPtrLivenessData OriginalLivenessData;
1916 computeLiveInValues(DT, F, OriginalLivenessData);
1917 for (size_t i = 0; i < records.size(); i++) {
1918 struct PartiallyConstructedSafepointRecord &info = records[i];
1919 const CallSite &CS = toUpdate[i];
1920 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1924 /// Remove any vector of pointers from the live set by scalarizing them over the
1925 /// statepoint instruction. Adds the scalarized pieces to the live set. It
1926 /// would be preferable to include the vector in the statepoint itself, but
1927 /// the lowering code currently does not handle that. Extending it would be
1928 /// slightly non-trivial since it requires a format change. Given how rare
1929 /// such cases are (for the moment?) scalarizing is an acceptable compromise.
1930 static void splitVectorValues(Instruction *StatepointInst,
1931 StatepointLiveSetTy &LiveSet,
1932 DenseMap<Value *, Value *>& PointerToBase,
1933 DominatorTree &DT) {
1934 SmallVector<Value *, 16> ToSplit;
1935 for (Value *V : LiveSet)
1936 if (isa<VectorType>(V->getType()))
1937 ToSplit.push_back(V);
1939 if (ToSplit.empty())
1942 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1944 Function &F = *(StatepointInst->getParent()->getParent());
1946 DenseMap<Value *, AllocaInst *> AllocaMap;
1947 // First is normal return, second is exceptional return (invoke only)
1948 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1949 for (Value *V : ToSplit) {
1950 AllocaInst *Alloca =
1951 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1952 AllocaMap[V] = Alloca;
1954 VectorType *VT = cast<VectorType>(V->getType());
1955 IRBuilder<> Builder(StatepointInst);
1956 SmallVector<Value *, 16> Elements;
1957 for (unsigned i = 0; i < VT->getNumElements(); i++)
1958 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1959 ElementMapping[V] = Elements;
1961 auto InsertVectorReform = [&](Instruction *IP) {
1962 Builder.SetInsertPoint(IP);
1963 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1964 Value *ResultVec = UndefValue::get(VT);
1965 for (unsigned i = 0; i < VT->getNumElements(); i++)
1966 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1967 Builder.getInt32(i));
1971 if (isa<CallInst>(StatepointInst)) {
1972 BasicBlock::iterator Next(StatepointInst);
1974 Instruction *IP = &*(Next);
1975 Replacements[V].first = InsertVectorReform(IP);
1976 Replacements[V].second = nullptr;
1978 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1979 // We've already normalized - check that we don't have shared destination
1981 BasicBlock *NormalDest = Invoke->getNormalDest();
1982 assert(!isa<PHINode>(NormalDest->begin()));
1983 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1984 assert(!isa<PHINode>(UnwindDest->begin()));
1985 // Insert insert element sequences in both successors
1986 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1987 Replacements[V].first = InsertVectorReform(IP);
1988 IP = &*(UnwindDest->getFirstInsertionPt());
1989 Replacements[V].second = InsertVectorReform(IP);
1993 for (Value *V : ToSplit) {
1994 AllocaInst *Alloca = AllocaMap[V];
1996 // Capture all users before we start mutating use lists
1997 SmallVector<Instruction *, 16> Users;
1998 for (User *U : V->users())
1999 Users.push_back(cast<Instruction>(U));
2001 for (Instruction *I : Users) {
2002 if (auto Phi = dyn_cast<PHINode>(I)) {
2003 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
2004 if (V == Phi->getIncomingValue(i)) {
2005 LoadInst *Load = new LoadInst(
2006 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
2007 Phi->setIncomingValue(i, Load);
2010 LoadInst *Load = new LoadInst(Alloca, "", I);
2011 I->replaceUsesOfWith(V, Load);
2015 // Store the original value and the replacement value into the alloca
2016 StoreInst *Store = new StoreInst(V, Alloca);
2017 if (auto I = dyn_cast<Instruction>(V))
2018 Store->insertAfter(I);
2020 Store->insertAfter(Alloca);
2022 // Normal return for invoke, or call return
2023 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
2024 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
2025 // Unwind return for invoke only
2026 Replacement = cast_or_null<Instruction>(Replacements[V].second);
2028 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
2031 // apply mem2reg to promote alloca to SSA
2032 SmallVector<AllocaInst *, 16> Allocas;
2033 for (Value *V : ToSplit)
2034 Allocas.push_back(AllocaMap[V]);
2035 PromoteMemToReg(Allocas, DT);
2037 // Update our tracking of live pointers and base mappings to account for the
2038 // changes we just made.
2039 for (Value *V : ToSplit) {
2040 auto &Elements = ElementMapping[V];
2043 LiveSet.insert(Elements.begin(), Elements.end());
2044 // We need to update the base mapping as well.
2045 assert(PointerToBase.count(V));
2046 Value *OldBase = PointerToBase[V];
2047 auto &BaseElements = ElementMapping[OldBase];
2048 PointerToBase.erase(V);
2049 assert(Elements.size() == BaseElements.size());
2050 for (unsigned i = 0; i < Elements.size(); i++) {
2051 Value *Elem = Elements[i];
2052 PointerToBase[Elem] = BaseElements[i];
2057 // Helper function for the "rematerializeLiveValues". It walks use chain
2058 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
2059 // values are visited (currently it is GEP's and casts). Returns true if it
2060 // successfully reached "BaseValue" and false otherwise.
2061 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
2063 static bool findRematerializableChainToBasePointer(
2064 SmallVectorImpl<Instruction*> &ChainToBase,
2065 Value *CurrentValue, Value *BaseValue) {
2067 // We have found a base value
2068 if (CurrentValue == BaseValue) {
2072 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
2073 ChainToBase.push_back(GEP);
2074 return findRematerializableChainToBasePointer(ChainToBase,
2075 GEP->getPointerOperand(),
2079 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
2080 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
2083 ChainToBase.push_back(CI);
2084 return findRematerializableChainToBasePointer(ChainToBase,
2085 CI->getOperand(0), BaseValue);
2088 // Not supported instruction in the chain
2092 // Helper function for the "rematerializeLiveValues". Compute cost of the use
2093 // chain we are going to rematerialize.
2095 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
2096 TargetTransformInfo &TTI) {
2099 for (Instruction *Instr : Chain) {
2100 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
2101 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
2102 "non noop cast is found during rematerialization");
2104 Type *SrcTy = CI->getOperand(0)->getType();
2105 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
2107 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
2108 // Cost of the address calculation
2109 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
2110 Cost += TTI.getAddressComputationCost(ValTy);
2112 // And cost of the GEP itself
2113 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
2114 // allowed for the external usage)
2115 if (!GEP->hasAllConstantIndices())
2119 llvm_unreachable("unsupported instruciton type during rematerialization");
2126 // From the statepoint live set pick values that are cheaper to recompute then
2127 // to relocate. Remove this values from the live set, rematerialize them after
2128 // statepoint and record them in "Info" structure. Note that similar to
2129 // relocated values we don't do any user adjustments here.
2130 static void rematerializeLiveValues(CallSite CS,
2131 PartiallyConstructedSafepointRecord &Info,
2132 TargetTransformInfo &TTI) {
2133 const unsigned int ChainLengthThreshold = 10;
2135 // Record values we are going to delete from this statepoint live set.
2136 // We can not di this in following loop due to iterator invalidation.
2137 SmallVector<Value *, 32> LiveValuesToBeDeleted;
2139 for (Value *LiveValue: Info.LiveSet) {
2140 // For each live pointer find it's defining chain
2141 SmallVector<Instruction *, 3> ChainToBase;
2142 assert(Info.PointerToBase.count(LiveValue));
2144 findRematerializableChainToBasePointer(ChainToBase,
2146 Info.PointerToBase[LiveValue]);
2147 // Nothing to do, or chain is too long
2149 ChainToBase.size() == 0 ||
2150 ChainToBase.size() > ChainLengthThreshold)
2153 // Compute cost of this chain
2154 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
2155 // TODO: We can also account for cases when we will be able to remove some
2156 // of the rematerialized values by later optimization passes. I.e if
2157 // we rematerialized several intersecting chains. Or if original values
2158 // don't have any uses besides this statepoint.
2160 // For invokes we need to rematerialize each chain twice - for normal and
2161 // for unwind basic blocks. Model this by multiplying cost by two.
2162 if (CS.isInvoke()) {
2165 // If it's too expensive - skip it
2166 if (Cost >= RematerializationThreshold)
2169 // Remove value from the live set
2170 LiveValuesToBeDeleted.push_back(LiveValue);
2172 // Clone instructions and record them inside "Info" structure
2174 // Walk backwards to visit top-most instructions first
2175 std::reverse(ChainToBase.begin(), ChainToBase.end());
2177 // Utility function which clones all instructions from "ChainToBase"
2178 // and inserts them before "InsertBefore". Returns rematerialized value
2179 // which should be used after statepoint.
2180 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
2181 Instruction *LastClonedValue = nullptr;
2182 Instruction *LastValue = nullptr;
2183 for (Instruction *Instr: ChainToBase) {
2184 // Only GEP's and casts are suported as we need to be careful to not
2185 // introduce any new uses of pointers not in the liveset.
2186 // Note that it's fine to introduce new uses of pointers which were
2187 // otherwise not used after this statepoint.
2188 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2190 Instruction *ClonedValue = Instr->clone();
2191 ClonedValue->insertBefore(InsertBefore);
2192 ClonedValue->setName(Instr->getName() + ".remat");
2194 // If it is not first instruction in the chain then it uses previously
2195 // cloned value. We should update it to use cloned value.
2196 if (LastClonedValue) {
2198 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2200 // Assert that cloned instruction does not use any instructions from
2201 // this chain other than LastClonedValue
2202 for (auto OpValue : ClonedValue->operand_values()) {
2203 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2204 ChainToBase.end() &&
2205 "incorrect use in rematerialization chain");
2210 LastClonedValue = ClonedValue;
2213 assert(LastClonedValue);
2214 return LastClonedValue;
2217 // Different cases for calls and invokes. For invokes we need to clone
2218 // instructions both on normal and unwind path.
2220 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2221 assert(InsertBefore);
2222 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2223 Info.RematerializedValues[RematerializedValue] = LiveValue;
2225 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2227 Instruction *NormalInsertBefore =
2228 &*Invoke->getNormalDest()->getFirstInsertionPt();
2229 Instruction *UnwindInsertBefore =
2230 &*Invoke->getUnwindDest()->getFirstInsertionPt();
2232 Instruction *NormalRematerializedValue =
2233 rematerializeChain(NormalInsertBefore);
2234 Instruction *UnwindRematerializedValue =
2235 rematerializeChain(UnwindInsertBefore);
2237 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2238 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2242 // Remove rematerializaed values from the live set
2243 for (auto LiveValue: LiveValuesToBeDeleted) {
2244 Info.LiveSet.erase(LiveValue);
2248 static bool insertParsePoints(Function &F, DominatorTree &DT,
2249 TargetTransformInfo &TTI,
2250 SmallVectorImpl<CallSite> &ToUpdate) {
2252 // sanity check the input
2253 std::set<CallSite> Uniqued;
2254 Uniqued.insert(ToUpdate.begin(), ToUpdate.end());
2255 assert(Uniqued.size() == ToUpdate.size() && "no duplicates please!");
2257 for (CallSite CS : ToUpdate) {
2258 assert(CS.getInstruction()->getParent()->getParent() == &F);
2259 assert((UseDeoptBundles || isStatepoint(CS)) &&
2260 "expected to already be a deopt statepoint");
2264 // When inserting gc.relocates for invokes, we need to be able to insert at
2265 // the top of the successor blocks. See the comment on
2266 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2267 // may restructure the CFG.
2268 for (CallSite CS : ToUpdate) {
2271 auto *II = cast<InvokeInst>(CS.getInstruction());
2272 normalizeForInvokeSafepoint(II->getNormalDest(), II->getParent(), DT);
2273 normalizeForInvokeSafepoint(II->getUnwindDest(), II->getParent(), DT);
2276 // A list of dummy calls added to the IR to keep various values obviously
2277 // live in the IR. We'll remove all of these when done.
2278 SmallVector<CallInst *, 64> Holders;
2280 // Insert a dummy call with all of the arguments to the vm_state we'll need
2281 // for the actual safepoint insertion. This ensures reference arguments in
2282 // the deopt argument list are considered live through the safepoint (and
2283 // thus makes sure they get relocated.)
2284 for (CallSite CS : ToUpdate) {
2285 SmallVector<Value *, 64> DeoptValues;
2287 iterator_range<const Use *> DeoptStateRange =
2289 ? iterator_range<const Use *>(GetDeoptBundleOperands(CS))
2290 : iterator_range<const Use *>(Statepoint(CS).vm_state_args());
2292 for (Value *Arg : DeoptStateRange) {
2293 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2294 "support for FCA unimplemented");
2295 if (isHandledGCPointerType(Arg->getType()))
2296 DeoptValues.push_back(Arg);
2299 insertUseHolderAfter(CS, DeoptValues, Holders);
2302 SmallVector<PartiallyConstructedSafepointRecord, 64> Records(ToUpdate.size());
2304 // A) Identify all gc pointers which are statically live at the given call
2306 findLiveReferences(F, DT, ToUpdate, Records);
2308 // B) Find the base pointers for each live pointer
2309 /* scope for caching */ {
2310 // Cache the 'defining value' relation used in the computation and
2311 // insertion of base phis and selects. This ensures that we don't insert
2312 // large numbers of duplicate base_phis.
2313 DefiningValueMapTy DVCache;
2315 for (size_t i = 0; i < Records.size(); i++) {
2316 PartiallyConstructedSafepointRecord &info = Records[i];
2317 findBasePointers(DT, DVCache, ToUpdate[i], info);
2319 } // end of cache scope
2321 // The base phi insertion logic (for any safepoint) may have inserted new
2322 // instructions which are now live at some safepoint. The simplest such
2325 // phi a <-- will be a new base_phi here
2326 // safepoint 1 <-- that needs to be live here
2330 // We insert some dummy calls after each safepoint to definitely hold live
2331 // the base pointers which were identified for that safepoint. We'll then
2332 // ask liveness for _every_ base inserted to see what is now live. Then we
2333 // remove the dummy calls.
2334 Holders.reserve(Holders.size() + Records.size());
2335 for (size_t i = 0; i < Records.size(); i++) {
2336 PartiallyConstructedSafepointRecord &Info = Records[i];
2338 SmallVector<Value *, 128> Bases;
2339 for (auto Pair : Info.PointerToBase)
2340 Bases.push_back(Pair.second);
2342 insertUseHolderAfter(ToUpdate[i], Bases, Holders);
2345 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2346 // need to rerun liveness. We may *also* have inserted new defs, but that's
2347 // not the key issue.
2348 recomputeLiveInValues(F, DT, ToUpdate, Records);
2350 if (PrintBasePointers) {
2351 for (auto &Info : Records) {
2352 errs() << "Base Pairs: (w/Relocation)\n";
2353 for (auto Pair : Info.PointerToBase) {
2354 errs() << " derived ";
2355 Pair.first->printAsOperand(errs(), false);
2357 Pair.second->printAsOperand(errs(), false);
2363 // It is possible that non-constant live variables have a constant base. For
2364 // example, a GEP with a variable offset from a global. In this case we can
2365 // remove it from the liveset. We already don't add constants to the liveset
2366 // because we assume they won't move at runtime and the GC doesn't need to be
2367 // informed about them. The same reasoning applies if the base is constant.
2368 // Note that the relocation placement code relies on this filtering for
2369 // correctness as it expects the base to be in the liveset, which isn't true
2370 // if the base is constant.
2371 for (auto &Info : Records)
2372 for (auto &BasePair : Info.PointerToBase)
2373 if (isa<Constant>(BasePair.second))
2374 Info.LiveSet.erase(BasePair.first);
2376 for (CallInst *CI : Holders)
2377 CI->eraseFromParent();
2381 // Do a limited scalarization of any live at safepoint vector values which
2382 // contain pointers. This enables this pass to run after vectorization at
2383 // the cost of some possible performance loss. TODO: it would be nice to
2384 // natively support vectors all the way through the backend so we don't need
2385 // to scalarize here.
2386 for (size_t i = 0; i < Records.size(); i++) {
2387 PartiallyConstructedSafepointRecord &Info = Records[i];
2388 Instruction *Statepoint = ToUpdate[i].getInstruction();
2389 splitVectorValues(cast<Instruction>(Statepoint), Info.LiveSet,
2390 Info.PointerToBase, DT);
2393 // In order to reduce live set of statepoint we might choose to rematerialize
2394 // some values instead of relocating them. This is purely an optimization and
2395 // does not influence correctness.
2396 for (size_t i = 0; i < Records.size(); i++)
2397 rematerializeLiveValues(ToUpdate[i], Records[i], TTI);
2399 // We need this to safely RAUW and delete call or invoke return values that
2400 // may themselves be live over a statepoint. For details, please see usage in
2401 // makeStatepointExplicitImpl.
2402 std::vector<DeferredReplacement> Replacements;
2404 // Now run through and replace the existing statepoints with new ones with
2405 // the live variables listed. We do not yet update uses of the values being
2406 // relocated. We have references to live variables that need to
2407 // survive to the last iteration of this loop. (By construction, the
2408 // previous statepoint can not be a live variable, thus we can and remove
2409 // the old statepoint calls as we go.)
2410 for (size_t i = 0; i < Records.size(); i++)
2411 makeStatepointExplicit(DT, ToUpdate[i], Records[i], Replacements);
2413 ToUpdate.clear(); // prevent accident use of invalid CallSites
2415 for (auto &PR : Replacements)
2418 Replacements.clear();
2420 for (auto &Info : Records) {
2421 // These live sets may contain state Value pointers, since we replaced calls
2422 // with operand bundles with calls wrapped in gc.statepoint, and some of
2423 // those calls may have been def'ing live gc pointers. Clear these out to
2424 // avoid accidentally using them.
2426 // TODO: We should create a separate data structure that does not contain
2427 // these live sets, and migrate to using that data structure from this point
2429 Info.LiveSet.clear();
2430 Info.PointerToBase.clear();
2433 // Do all the fixups of the original live variables to their relocated selves
2434 SmallVector<Value *, 128> Live;
2435 for (size_t i = 0; i < Records.size(); i++) {
2436 PartiallyConstructedSafepointRecord &Info = Records[i];
2438 // We can't simply save the live set from the original insertion. One of
2439 // the live values might be the result of a call which needs a safepoint.
2440 // That Value* no longer exists and we need to use the new gc_result.
2441 // Thankfully, the live set is embedded in the statepoint (and updated), so
2442 // we just grab that.
2443 Statepoint Statepoint(Info.StatepointToken);
2444 Live.insert(Live.end(), Statepoint.gc_args_begin(),
2445 Statepoint.gc_args_end());
2447 // Do some basic sanity checks on our liveness results before performing
2448 // relocation. Relocation can and will turn mistakes in liveness results
2449 // into non-sensical code which is must harder to debug.
2450 // TODO: It would be nice to test consistency as well
2451 assert(DT.isReachableFromEntry(Info.StatepointToken->getParent()) &&
2452 "statepoint must be reachable or liveness is meaningless");
2453 for (Value *V : Statepoint.gc_args()) {
2454 if (!isa<Instruction>(V))
2455 // Non-instruction values trivial dominate all possible uses
2457 auto *LiveInst = cast<Instruction>(V);
2458 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2459 "unreachable values should never be live");
2460 assert(DT.dominates(LiveInst, Info.StatepointToken) &&
2461 "basic SSA liveness expectation violated by liveness analysis");
2465 unique_unsorted(Live);
2469 for (auto *Ptr : Live)
2470 assert(isGCPointerType(Ptr->getType()) && "must be a gc pointer type");
2473 relocationViaAlloca(F, DT, Live, Records);
2474 return !Records.empty();
2477 // Handles both return values and arguments for Functions and CallSites.
2478 template <typename AttrHolder>
2479 static void RemoveNonValidAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2482 if (AH.getDereferenceableBytes(Index))
2483 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2484 AH.getDereferenceableBytes(Index)));
2485 if (AH.getDereferenceableOrNullBytes(Index))
2486 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2487 AH.getDereferenceableOrNullBytes(Index)));
2488 if (AH.doesNotAlias(Index))
2489 R.addAttribute(Attribute::NoAlias);
2492 AH.setAttributes(AH.getAttributes().removeAttributes(
2493 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2497 RewriteStatepointsForGC::stripNonValidAttributesFromPrototype(Function &F) {
2498 LLVMContext &Ctx = F.getContext();
2500 for (Argument &A : F.args())
2501 if (isa<PointerType>(A.getType()))
2502 RemoveNonValidAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2504 if (isa<PointerType>(F.getReturnType()))
2505 RemoveNonValidAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2508 void RewriteStatepointsForGC::stripNonValidAttributesFromBody(Function &F) {
2512 LLVMContext &Ctx = F.getContext();
2513 MDBuilder Builder(Ctx);
2515 for (Instruction &I : instructions(F)) {
2516 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2517 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2518 bool IsImmutableTBAA =
2519 MD->getNumOperands() == 4 &&
2520 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2522 if (!IsImmutableTBAA)
2523 continue; // no work to do, MD_tbaa is already marked mutable
2525 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2526 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2528 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2530 MDNode *MutableTBAA =
2531 Builder.createTBAAStructTagNode(Base, Access, Offset);
2532 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2535 if (CallSite CS = CallSite(&I)) {
2536 for (int i = 0, e = CS.arg_size(); i != e; i++)
2537 if (isa<PointerType>(CS.getArgument(i)->getType()))
2538 RemoveNonValidAttrAtIndex(Ctx, CS, i + 1);
2539 if (isa<PointerType>(CS.getType()))
2540 RemoveNonValidAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2545 /// Returns true if this function should be rewritten by this pass. The main
2546 /// point of this function is as an extension point for custom logic.
2547 static bool shouldRewriteStatepointsIn(Function &F) {
2548 // TODO: This should check the GCStrategy
2550 const char *FunctionGCName = F.getGC();
2551 const StringRef StatepointExampleName("statepoint-example");
2552 const StringRef CoreCLRName("coreclr");
2553 return (StatepointExampleName == FunctionGCName) ||
2554 (CoreCLRName == FunctionGCName);
2559 void RewriteStatepointsForGC::stripNonValidAttributes(Module &M) {
2561 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2565 for (Function &F : M)
2566 stripNonValidAttributesFromPrototype(F);
2568 for (Function &F : M)
2569 stripNonValidAttributesFromBody(F);
2572 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2573 // Nothing to do for declarations.
2574 if (F.isDeclaration() || F.empty())
2577 // Policy choice says not to rewrite - the most common reason is that we're
2578 // compiling code without a GCStrategy.
2579 if (!shouldRewriteStatepointsIn(F))
2582 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2583 TargetTransformInfo &TTI =
2584 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2586 auto NeedsRewrite = [](Instruction &I) {
2587 if (UseDeoptBundles) {
2588 if (ImmutableCallSite CS = ImmutableCallSite(&I))
2589 return !callsGCLeafFunction(CS);
2593 return isStatepoint(I);
2596 // Gather all the statepoints which need rewritten. Be careful to only
2597 // consider those in reachable code since we need to ask dominance queries
2598 // when rewriting. We'll delete the unreachable ones in a moment.
2599 SmallVector<CallSite, 64> ParsePointNeeded;
2600 bool HasUnreachableStatepoint = false;
2601 for (Instruction &I : instructions(F)) {
2602 // TODO: only the ones with the flag set!
2603 if (NeedsRewrite(I)) {
2604 if (DT.isReachableFromEntry(I.getParent()))
2605 ParsePointNeeded.push_back(CallSite(&I));
2607 HasUnreachableStatepoint = true;
2611 bool MadeChange = false;
2613 // Delete any unreachable statepoints so that we don't have unrewritten
2614 // statepoints surviving this pass. This makes testing easier and the
2615 // resulting IR less confusing to human readers. Rather than be fancy, we
2616 // just reuse a utility function which removes the unreachable blocks.
2617 if (HasUnreachableStatepoint)
2618 MadeChange |= removeUnreachableBlocks(F);
2620 // Return early if no work to do.
2621 if (ParsePointNeeded.empty())
2624 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2625 // These are created by LCSSA. They have the effect of increasing the size
2626 // of liveness sets for no good reason. It may be harder to do this post
2627 // insertion since relocations and base phis can confuse things.
2628 for (BasicBlock &BB : F)
2629 if (BB.getUniquePredecessor()) {
2631 FoldSingleEntryPHINodes(&BB);
2634 // Before we start introducing relocations, we want to tweak the IR a bit to
2635 // avoid unfortunate code generation effects. The main example is that we
2636 // want to try to make sure the comparison feeding a branch is after any
2637 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2638 // values feeding a branch after relocation. This is semantically correct,
2639 // but results in extra register pressure since both the pre-relocation and
2640 // post-relocation copies must be available in registers. For code without
2641 // relocations this is handled elsewhere, but teaching the scheduler to
2642 // reverse the transform we're about to do would be slightly complex.
2643 // Note: This may extend the live range of the inputs to the icmp and thus
2644 // increase the liveset of any statepoint we move over. This is profitable
2645 // as long as all statepoints are in rare blocks. If we had in-register
2646 // lowering for live values this would be a much safer transform.
2647 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2648 if (auto *BI = dyn_cast<BranchInst>(TI))
2649 if (BI->isConditional())
2650 return dyn_cast<Instruction>(BI->getCondition());
2651 // TODO: Extend this to handle switches
2654 for (BasicBlock &BB : F) {
2655 TerminatorInst *TI = BB.getTerminator();
2656 if (auto *Cond = getConditionInst(TI))
2657 // TODO: Handle more than just ICmps here. We should be able to move
2658 // most instructions without side effects or memory access.
2659 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2661 Cond->moveBefore(TI);
2665 MadeChange |= insertParsePoints(F, DT, TTI, ParsePointNeeded);
2669 // liveness computation via standard dataflow
2670 // -------------------------------------------------------------------
2672 // TODO: Consider using bitvectors for liveness, the set of potentially
2673 // interesting values should be small and easy to pre-compute.
2675 /// Compute the live-in set for the location rbegin starting from
2676 /// the live-out set of the basic block
2677 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2678 BasicBlock::reverse_iterator rend,
2679 DenseSet<Value *> &LiveTmp) {
2681 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2682 Instruction *I = &*ritr;
2684 // KILL/Def - Remove this definition from LiveIn
2687 // Don't consider *uses* in PHI nodes, we handle their contribution to
2688 // predecessor blocks when we seed the LiveOut sets
2689 if (isa<PHINode>(I))
2692 // USE - Add to the LiveIn set for this instruction
2693 for (Value *V : I->operands()) {
2694 assert(!isUnhandledGCPointerType(V->getType()) &&
2695 "support for FCA unimplemented");
2696 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2697 // The choice to exclude all things constant here is slightly subtle.
2698 // There are two independent reasons:
2699 // - We assume that things which are constant (from LLVM's definition)
2700 // do not move at runtime. For example, the address of a global
2701 // variable is fixed, even though it's contents may not be.
2702 // - Second, we can't disallow arbitrary inttoptr constants even
2703 // if the language frontend does. Optimization passes are free to
2704 // locally exploit facts without respect to global reachability. This
2705 // can create sections of code which are dynamically unreachable and
2706 // contain just about anything. (see constants.ll in tests)
2713 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2715 for (BasicBlock *Succ : successors(BB)) {
2716 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2717 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2718 PHINode *Phi = cast<PHINode>(&*I);
2719 Value *V = Phi->getIncomingValueForBlock(BB);
2720 assert(!isUnhandledGCPointerType(V->getType()) &&
2721 "support for FCA unimplemented");
2722 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2729 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2730 DenseSet<Value *> KillSet;
2731 for (Instruction &I : *BB)
2732 if (isHandledGCPointerType(I.getType()))
2738 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2739 /// sanity check for the liveness computation.
2740 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2741 TerminatorInst *TI, bool TermOkay = false) {
2742 for (Value *V : Live) {
2743 if (auto *I = dyn_cast<Instruction>(V)) {
2744 // The terminator can be a member of the LiveOut set. LLVM's definition
2745 // of instruction dominance states that V does not dominate itself. As
2746 // such, we need to special case this to allow it.
2747 if (TermOkay && TI == I)
2749 assert(DT.dominates(I, TI) &&
2750 "basic SSA liveness expectation violated by liveness analysis");
2755 /// Check that all the liveness sets used during the computation of liveness
2756 /// obey basic SSA properties. This is useful for finding cases where we miss
2758 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2760 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2761 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2762 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2766 static void computeLiveInValues(DominatorTree &DT, Function &F,
2767 GCPtrLivenessData &Data) {
2769 SmallSetVector<BasicBlock *, 200> Worklist;
2770 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2771 // We use a SetVector so that we don't have duplicates in the worklist.
2772 Worklist.insert(pred_begin(BB), pred_end(BB));
2774 auto NextItem = [&]() {
2775 BasicBlock *BB = Worklist.back();
2776 Worklist.pop_back();
2780 // Seed the liveness for each individual block
2781 for (BasicBlock &BB : F) {
2782 Data.KillSet[&BB] = computeKillSet(&BB);
2783 Data.LiveSet[&BB].clear();
2784 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2787 for (Value *Kill : Data.KillSet[&BB])
2788 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2791 Data.LiveOut[&BB] = DenseSet<Value *>();
2792 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2793 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2794 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2795 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2796 if (!Data.LiveIn[&BB].empty())
2797 AddPredsToWorklist(&BB);
2800 // Propagate that liveness until stable
2801 while (!Worklist.empty()) {
2802 BasicBlock *BB = NextItem();
2804 // Compute our new liveout set, then exit early if it hasn't changed
2805 // despite the contribution of our successor.
2806 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2807 const auto OldLiveOutSize = LiveOut.size();
2808 for (BasicBlock *Succ : successors(BB)) {
2809 assert(Data.LiveIn.count(Succ));
2810 set_union(LiveOut, Data.LiveIn[Succ]);
2812 // assert OutLiveOut is a subset of LiveOut
2813 if (OldLiveOutSize == LiveOut.size()) {
2814 // If the sets are the same size, then we didn't actually add anything
2815 // when unioning our successors LiveIn Thus, the LiveIn of this block
2819 Data.LiveOut[BB] = LiveOut;
2821 // Apply the effects of this basic block
2822 DenseSet<Value *> LiveTmp = LiveOut;
2823 set_union(LiveTmp, Data.LiveSet[BB]);
2824 set_subtract(LiveTmp, Data.KillSet[BB]);
2826 assert(Data.LiveIn.count(BB));
2827 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2828 // assert: OldLiveIn is a subset of LiveTmp
2829 if (OldLiveIn.size() != LiveTmp.size()) {
2830 Data.LiveIn[BB] = LiveTmp;
2831 AddPredsToWorklist(BB);
2833 } // while( !worklist.empty() )
2836 // Sanity check our output against SSA properties. This helps catch any
2837 // missing kills during the above iteration.
2838 for (BasicBlock &BB : F) {
2839 checkBasicSSA(DT, Data, BB);
2844 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2845 StatepointLiveSetTy &Out) {
2847 BasicBlock *BB = Inst->getParent();
2849 // Note: The copy is intentional and required
2850 assert(Data.LiveOut.count(BB));
2851 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2853 // We want to handle the statepoint itself oddly. It's
2854 // call result is not live (normal), nor are it's arguments
2855 // (unless they're used again later). This adjustment is
2856 // specifically what we need to relocate
2857 BasicBlock::reverse_iterator rend(Inst->getIterator());
2858 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2859 LiveOut.erase(Inst);
2860 Out.insert(LiveOut.begin(), LiveOut.end());
2863 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2865 PartiallyConstructedSafepointRecord &Info) {
2866 Instruction *Inst = CS.getInstruction();
2867 StatepointLiveSetTy Updated;
2868 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2871 DenseSet<Value *> Bases;
2872 for (auto KVPair : Info.PointerToBase) {
2873 Bases.insert(KVPair.second);
2876 // We may have base pointers which are now live that weren't before. We need
2877 // to update the PointerToBase structure to reflect this.
2878 for (auto V : Updated)
2879 if (!Info.PointerToBase.count(V)) {
2880 assert(Bases.count(V) && "can't find base for unexpected live value");
2881 Info.PointerToBase[V] = V;
2886 for (auto V : Updated) {
2887 assert(Info.PointerToBase.count(V) &&
2888 "must be able to find base for live value");
2892 // Remove any stale base mappings - this can happen since our liveness is
2893 // more precise then the one inherent in the base pointer analysis
2894 DenseSet<Value *> ToErase;
2895 for (auto KVPair : Info.PointerToBase)
2896 if (!Updated.count(KVPair.first))
2897 ToErase.insert(KVPair.first);
2898 for (auto V : ToErase)
2899 Info.PointerToBase.erase(V);
2902 for (auto KVPair : Info.PointerToBase)
2903 assert(Updated.count(KVPair.first) && "record for non-live value");
2906 Info.LiveSet = Updated;