1 //===- RewriteStatepointsForGC.cpp - Make GC relocations explicit ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Rewrite an existing set of gc.statepoints such that they make potential
11 // relocations performed by the garbage collector explicit in the IR.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Pass.h"
16 #include "llvm/Analysis/CFG.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/TargetTransformInfo.h"
19 #include "llvm/ADT/SetOperations.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/ADT/DenseSet.h"
22 #include "llvm/ADT/SetVector.h"
23 #include "llvm/ADT/StringRef.h"
24 #include "llvm/ADT/MapVector.h"
25 #include "llvm/IR/BasicBlock.h"
26 #include "llvm/IR/CallSite.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/InstIterator.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Statepoint.h"
37 #include "llvm/IR/Value.h"
38 #include "llvm/IR/Verifier.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Transforms/Scalar.h"
42 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 #include "llvm/Transforms/Utils/Cloning.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
47 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
51 // Print the liveset found at the insert location
52 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
54 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden,
56 // Print out the base pointers for debugging
57 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden,
60 // Cost threshold measuring when it is profitable to rematerialize value instead
62 static cl::opt<unsigned>
63 RematerializationThreshold("spp-rematerialization-threshold", cl::Hidden,
67 static bool ClobberNonLive = true;
69 static bool ClobberNonLive = false;
71 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
72 cl::location(ClobberNonLive),
75 static cl::opt<bool> UseDeoptBundles("rs4gc-use-deopt-bundles", cl::Hidden,
78 AllowStatepointWithNoDeoptInfo("rs4gc-allow-statepoint-with-no-deopt-info",
79 cl::Hidden, cl::init(true));
82 struct RewriteStatepointsForGC : public ModulePass {
83 static char ID; // Pass identification, replacement for typeid
85 RewriteStatepointsForGC() : ModulePass(ID) {
86 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
88 bool runOnFunction(Function &F);
89 bool runOnModule(Module &M) override {
92 Changed |= runOnFunction(F);
95 // stripNonValidAttributes asserts that shouldRewriteStatepointsIn
96 // returns true for at least one function in the module. Since at least
97 // one function changed, we know that the precondition is satisfied.
98 stripNonValidAttributes(M);
104 void getAnalysisUsage(AnalysisUsage &AU) const override {
105 // We add and rewrite a bunch of instructions, but don't really do much
106 // else. We could in theory preserve a lot more analyses here.
107 AU.addRequired<DominatorTreeWrapperPass>();
108 AU.addRequired<TargetTransformInfoWrapperPass>();
111 /// The IR fed into RewriteStatepointsForGC may have had attributes implying
112 /// dereferenceability that are no longer valid/correct after
113 /// RewriteStatepointsForGC has run. This is because semantically, after
114 /// RewriteStatepointsForGC runs, all calls to gc.statepoint "free" the entire
115 /// heap. stripNonValidAttributes (conservatively) restores correctness
116 /// by erasing all attributes in the module that externally imply
117 /// dereferenceability.
118 /// Similar reasoning also applies to the noalias attributes. gc.statepoint
119 /// can touch the entire heap including noalias objects.
120 void stripNonValidAttributes(Module &M);
122 // Helpers for stripNonValidAttributes
123 void stripNonValidAttributesFromBody(Function &F);
124 void stripNonValidAttributesFromPrototype(Function &F);
128 char RewriteStatepointsForGC::ID = 0;
130 ModulePass *llvm::createRewriteStatepointsForGCPass() {
131 return new RewriteStatepointsForGC();
134 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
135 "Make relocations explicit at statepoints", false, false)
136 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
137 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
138 "Make relocations explicit at statepoints", false, false)
141 struct GCPtrLivenessData {
142 /// Values defined in this block.
143 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet;
144 /// Values used in this block (and thus live); does not included values
145 /// killed within this block.
146 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet;
148 /// Values live into this basic block (i.e. used by any
149 /// instruction in this basic block or ones reachable from here)
150 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn;
152 /// Values live out of this basic block (i.e. live into
153 /// any successor block)
154 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut;
157 // The type of the internal cache used inside the findBasePointers family
158 // of functions. From the callers perspective, this is an opaque type and
159 // should not be inspected.
161 // In the actual implementation this caches two relations:
162 // - The base relation itself (i.e. this pointer is based on that one)
163 // - The base defining value relation (i.e. before base_phi insertion)
164 // Generally, after the execution of a full findBasePointer call, only the
165 // base relation will remain. Internally, we add a mixture of the two
166 // types, then update all the second type to the first type
167 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
168 typedef DenseSet<Value *> StatepointLiveSetTy;
169 typedef DenseMap<AssertingVH<Instruction>, AssertingVH<Value>>
170 RematerializedValueMapTy;
172 struct PartiallyConstructedSafepointRecord {
173 /// The set of values known to be live across this safepoint
174 StatepointLiveSetTy LiveSet;
176 /// Mapping from live pointers to a base-defining-value
177 DenseMap<Value *, Value *> PointerToBase;
179 /// The *new* gc.statepoint instruction itself. This produces the token
180 /// that normal path gc.relocates and the gc.result are tied to.
181 Instruction *StatepointToken;
183 /// Instruction to which exceptional gc relocates are attached
184 /// Makes it easier to iterate through them during relocationViaAlloca.
185 Instruction *UnwindToken;
187 /// Record live values we are rematerialized instead of relocating.
188 /// They are not included into 'LiveSet' field.
189 /// Maps rematerialized copy to it's original value.
190 RematerializedValueMapTy RematerializedValues;
194 static ArrayRef<Use> GetDeoptBundleOperands(ImmutableCallSite CS) {
195 assert(UseDeoptBundles && "Should not be called otherwise!");
197 Optional<OperandBundleUse> DeoptBundle = CS.getOperandBundle("deopt");
199 if (!DeoptBundle.hasValue()) {
200 assert(AllowStatepointWithNoDeoptInfo &&
201 "Found non-leaf call without deopt info!");
205 return DeoptBundle.getValue().Inputs;
208 /// Compute the live-in set for every basic block in the function
209 static void computeLiveInValues(DominatorTree &DT, Function &F,
210 GCPtrLivenessData &Data);
212 /// Given results from the dataflow liveness computation, find the set of live
213 /// Values at a particular instruction.
214 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
215 StatepointLiveSetTy &out);
217 // TODO: Once we can get to the GCStrategy, this becomes
218 // Optional<bool> isGCManagedPointer(const Value *V) const override {
220 static bool isGCPointerType(Type *T) {
221 if (auto *PT = dyn_cast<PointerType>(T))
222 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
223 // GC managed heap. We know that a pointer into this heap needs to be
224 // updated and that no other pointer does.
225 return (1 == PT->getAddressSpace());
229 // Return true if this type is one which a) is a gc pointer or contains a GC
230 // pointer and b) is of a type this code expects to encounter as a live value.
231 // (The insertion code will assert that a type which matches (a) and not (b)
232 // is not encountered.)
233 static bool isHandledGCPointerType(Type *T) {
234 // We fully support gc pointers
235 if (isGCPointerType(T))
237 // We partially support vectors of gc pointers. The code will assert if it
238 // can't handle something.
239 if (auto VT = dyn_cast<VectorType>(T))
240 if (isGCPointerType(VT->getElementType()))
246 /// Returns true if this type contains a gc pointer whether we know how to
247 /// handle that type or not.
248 static bool containsGCPtrType(Type *Ty) {
249 if (isGCPointerType(Ty))
251 if (VectorType *VT = dyn_cast<VectorType>(Ty))
252 return isGCPointerType(VT->getScalarType());
253 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
254 return containsGCPtrType(AT->getElementType());
255 if (StructType *ST = dyn_cast<StructType>(Ty))
256 return std::any_of(ST->subtypes().begin(), ST->subtypes().end(),
261 // Returns true if this is a type which a) is a gc pointer or contains a GC
262 // pointer and b) is of a type which the code doesn't expect (i.e. first class
263 // aggregates). Used to trip assertions.
264 static bool isUnhandledGCPointerType(Type *Ty) {
265 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
269 static bool order_by_name(Value *a, Value *b) {
270 if (a->hasName() && b->hasName()) {
271 return -1 == a->getName().compare(b->getName());
272 } else if (a->hasName() && !b->hasName()) {
274 } else if (!a->hasName() && b->hasName()) {
277 // Better than nothing, but not stable
282 // Return the name of the value suffixed with the provided value, or if the
283 // value didn't have a name, the default value specified.
284 static std::string suffixed_name_or(Value *V, StringRef Suffix,
285 StringRef DefaultName) {
286 return V->hasName() ? (V->getName() + Suffix).str() : DefaultName.str();
289 // Conservatively identifies any definitions which might be live at the
290 // given instruction. The analysis is performed immediately before the
291 // given instruction. Values defined by that instruction are not considered
292 // live. Values used by that instruction are considered live.
293 static void analyzeParsePointLiveness(
294 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
295 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
296 Instruction *inst = CS.getInstruction();
298 StatepointLiveSetTy LiveSet;
299 findLiveSetAtInst(inst, OriginalLivenessData, LiveSet);
302 // Note: This output is used by several of the test cases
303 // The order of elements in a set is not stable, put them in a vec and sort
305 SmallVector<Value *, 64> Temp;
306 Temp.insert(Temp.end(), LiveSet.begin(), LiveSet.end());
307 std::sort(Temp.begin(), Temp.end(), order_by_name);
308 errs() << "Live Variables:\n";
309 for (Value *V : Temp)
310 dbgs() << " " << V->getName() << " " << *V << "\n";
312 if (PrintLiveSetSize) {
313 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
314 errs() << "Number live values: " << LiveSet.size() << "\n";
316 result.LiveSet = LiveSet;
319 static bool isKnownBaseResult(Value *V);
321 /// A single base defining value - An immediate base defining value for an
322 /// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'.
323 /// For instructions which have multiple pointer [vector] inputs or that
324 /// transition between vector and scalar types, there is no immediate base
325 /// defining value. The 'base defining value' for 'Def' is the transitive
326 /// closure of this relation stopping at the first instruction which has no
327 /// immediate base defining value. The b.d.v. might itself be a base pointer,
328 /// but it can also be an arbitrary derived pointer.
329 struct BaseDefiningValueResult {
330 /// Contains the value which is the base defining value.
332 /// True if the base defining value is also known to be an actual base
334 const bool IsKnownBase;
335 BaseDefiningValueResult(Value *BDV, bool IsKnownBase)
336 : BDV(BDV), IsKnownBase(IsKnownBase) {
338 // Check consistency between new and old means of checking whether a BDV is
340 bool MustBeBase = isKnownBaseResult(BDV);
341 assert(!MustBeBase || MustBeBase == IsKnownBase);
347 static BaseDefiningValueResult findBaseDefiningValue(Value *I);
349 /// Return a base defining value for the 'Index' element of the given vector
350 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
351 /// 'I'. As an optimization, this method will try to determine when the
352 /// element is known to already be a base pointer. If this can be established,
353 /// the second value in the returned pair will be true. Note that either a
354 /// vector or a pointer typed value can be returned. For the former, the
355 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
356 /// If the later, the return pointer is a BDV (or possibly a base) for the
357 /// particular element in 'I'.
358 static BaseDefiningValueResult
359 findBaseDefiningValueOfVector(Value *I) {
360 assert(I->getType()->isVectorTy() &&
361 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
362 "Illegal to ask for the base pointer of a non-pointer type");
364 // Each case parallels findBaseDefiningValue below, see that code for
365 // detailed motivation.
367 if (isa<Argument>(I))
368 // An incoming argument to the function is a base pointer
369 return BaseDefiningValueResult(I, true);
371 // We shouldn't see the address of a global as a vector value?
372 assert(!isa<GlobalVariable>(I) &&
373 "unexpected global variable found in base of vector");
375 // inlining could possibly introduce phi node that contains
376 // undef if callee has multiple returns
377 if (isa<UndefValue>(I))
378 // utterly meaningless, but useful for dealing with partially optimized
380 return BaseDefiningValueResult(I, true);
382 // Due to inheritance, this must be _after_ the global variable and undef
384 if (Constant *Con = dyn_cast<Constant>(I)) {
385 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
386 "order of checks wrong!");
387 assert(Con->isNullValue() && "null is the only case which makes sense");
388 return BaseDefiningValueResult(Con, true);
391 if (isa<LoadInst>(I))
392 return BaseDefiningValueResult(I, true);
394 if (isa<InsertElementInst>(I))
395 // We don't know whether this vector contains entirely base pointers or
396 // not. To be conservatively correct, we treat it as a BDV and will
397 // duplicate code as needed to construct a parallel vector of bases.
398 return BaseDefiningValueResult(I, false);
400 if (isa<ShuffleVectorInst>(I))
401 // We don't know whether this vector contains entirely base pointers or
402 // not. To be conservatively correct, we treat it as a BDV and will
403 // duplicate code as needed to construct a parallel vector of bases.
404 // TODO: There a number of local optimizations which could be applied here
405 // for particular sufflevector patterns.
406 return BaseDefiningValueResult(I, false);
408 // A PHI or Select is a base defining value. The outer findBasePointer
409 // algorithm is responsible for constructing a base value for this BDV.
410 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
411 "unknown vector instruction - no base found for vector element");
412 return BaseDefiningValueResult(I, false);
415 /// Helper function for findBasePointer - Will return a value which either a)
416 /// defines the base pointer for the input, b) blocks the simple search
417 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
418 /// from pointer to vector type or back.
419 static BaseDefiningValueResult findBaseDefiningValue(Value *I) {
420 if (I->getType()->isVectorTy())
421 return findBaseDefiningValueOfVector(I);
423 assert(I->getType()->isPointerTy() &&
424 "Illegal to ask for the base pointer of a non-pointer type");
426 if (isa<Argument>(I))
427 // An incoming argument to the function is a base pointer
428 // We should have never reached here if this argument isn't an gc value
429 return BaseDefiningValueResult(I, true);
431 if (isa<GlobalVariable>(I))
433 return BaseDefiningValueResult(I, true);
435 // inlining could possibly introduce phi node that contains
436 // undef if callee has multiple returns
437 if (isa<UndefValue>(I))
438 // utterly meaningless, but useful for dealing with
439 // partially optimized code.
440 return BaseDefiningValueResult(I, true);
442 // Due to inheritance, this must be _after_ the global variable and undef
444 if (isa<Constant>(I)) {
445 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
446 "order of checks wrong!");
447 // Note: Finding a constant base for something marked for relocation
448 // doesn't really make sense. The most likely case is either a) some
449 // screwed up the address space usage or b) your validating against
450 // compiled C++ code w/o the proper separation. The only real exception
451 // is a null pointer. You could have generic code written to index of
452 // off a potentially null value and have proven it null. We also use
453 // null pointers in dead paths of relocation phis (which we might later
454 // want to find a base pointer for).
455 assert(isa<ConstantPointerNull>(I) &&
456 "null is the only case which makes sense");
457 return BaseDefiningValueResult(I, true);
460 if (CastInst *CI = dyn_cast<CastInst>(I)) {
461 Value *Def = CI->stripPointerCasts();
462 // If we find a cast instruction here, it means we've found a cast which is
463 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
464 // handle int->ptr conversion.
465 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
466 return findBaseDefiningValue(Def);
469 if (isa<LoadInst>(I))
470 // The value loaded is an gc base itself
471 return BaseDefiningValueResult(I, true);
474 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
475 // The base of this GEP is the base
476 return findBaseDefiningValue(GEP->getPointerOperand());
478 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
479 switch (II->getIntrinsicID()) {
480 case Intrinsic::experimental_gc_result_ptr:
482 // fall through to general call handling
484 case Intrinsic::experimental_gc_statepoint:
485 case Intrinsic::experimental_gc_result_float:
486 case Intrinsic::experimental_gc_result_int:
487 llvm_unreachable("these don't produce pointers");
488 case Intrinsic::experimental_gc_relocate: {
489 // Rerunning safepoint insertion after safepoints are already
490 // inserted is not supported. It could probably be made to work,
491 // but why are you doing this? There's no good reason.
492 llvm_unreachable("repeat safepoint insertion is not supported");
494 case Intrinsic::gcroot:
495 // Currently, this mechanism hasn't been extended to work with gcroot.
496 // There's no reason it couldn't be, but I haven't thought about the
497 // implications much.
499 "interaction with the gcroot mechanism is not supported");
502 // We assume that functions in the source language only return base
503 // pointers. This should probably be generalized via attributes to support
504 // both source language and internal functions.
505 if (isa<CallInst>(I) || isa<InvokeInst>(I))
506 return BaseDefiningValueResult(I, true);
508 // I have absolutely no idea how to implement this part yet. It's not
509 // necessarily hard, I just haven't really looked at it yet.
510 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
512 if (isa<AtomicCmpXchgInst>(I))
513 // A CAS is effectively a atomic store and load combined under a
514 // predicate. From the perspective of base pointers, we just treat it
516 return BaseDefiningValueResult(I, true);
518 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
519 "binary ops which don't apply to pointers");
521 // The aggregate ops. Aggregates can either be in the heap or on the
522 // stack, but in either case, this is simply a field load. As a result,
523 // this is a defining definition of the base just like a load is.
524 if (isa<ExtractValueInst>(I))
525 return BaseDefiningValueResult(I, true);
527 // We should never see an insert vector since that would require we be
528 // tracing back a struct value not a pointer value.
529 assert(!isa<InsertValueInst>(I) &&
530 "Base pointer for a struct is meaningless");
532 // An extractelement produces a base result exactly when it's input does.
533 // We may need to insert a parallel instruction to extract the appropriate
534 // element out of the base vector corresponding to the input. Given this,
535 // it's analogous to the phi and select case even though it's not a merge.
536 if (isa<ExtractElementInst>(I))
537 // Note: There a lot of obvious peephole cases here. This are deliberately
538 // handled after the main base pointer inference algorithm to make writing
539 // test cases to exercise that code easier.
540 return BaseDefiningValueResult(I, false);
542 // The last two cases here don't return a base pointer. Instead, they
543 // return a value which dynamically selects from among several base
544 // derived pointers (each with it's own base potentially). It's the job of
545 // the caller to resolve these.
546 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
547 "missing instruction case in findBaseDefiningValing");
548 return BaseDefiningValueResult(I, false);
551 /// Returns the base defining value for this value.
552 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
553 Value *&Cached = Cache[I];
555 Cached = findBaseDefiningValue(I).BDV;
556 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
557 << Cached->getName() << "\n");
559 assert(Cache[I] != nullptr);
563 /// Return a base pointer for this value if known. Otherwise, return it's
564 /// base defining value.
565 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
566 Value *Def = findBaseDefiningValueCached(I, Cache);
567 auto Found = Cache.find(Def);
568 if (Found != Cache.end()) {
569 // Either a base-of relation, or a self reference. Caller must check.
570 return Found->second;
572 // Only a BDV available
576 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
577 /// is it known to be a base pointer? Or do we need to continue searching.
578 static bool isKnownBaseResult(Value *V) {
579 if (!isa<PHINode>(V) && !isa<SelectInst>(V) &&
580 !isa<ExtractElementInst>(V) && !isa<InsertElementInst>(V) &&
581 !isa<ShuffleVectorInst>(V)) {
582 // no recursion possible
585 if (isa<Instruction>(V) &&
586 cast<Instruction>(V)->getMetadata("is_base_value")) {
587 // This is a previously inserted base phi or select. We know
588 // that this is a base value.
592 // We need to keep searching
597 /// Models the state of a single base defining value in the findBasePointer
598 /// algorithm for determining where a new instruction is needed to propagate
599 /// the base of this BDV.
602 enum Status { Unknown, Base, Conflict };
604 BDVState(Status s, Value *b = nullptr) : status(s), base(b) {
605 assert(status != Base || b);
607 explicit BDVState(Value *b) : status(Base), base(b) {}
608 BDVState() : status(Unknown), base(nullptr) {}
610 Status getStatus() const { return status; }
611 Value *getBase() const { return base; }
613 bool isBase() const { return getStatus() == Base; }
614 bool isUnknown() const { return getStatus() == Unknown; }
615 bool isConflict() const { return getStatus() == Conflict; }
617 bool operator==(const BDVState &other) const {
618 return base == other.base && status == other.status;
621 bool operator!=(const BDVState &other) const { return !(*this == other); }
624 void dump() const { print(dbgs()); dbgs() << '\n'; }
626 void print(raw_ostream &OS) const {
638 OS << " (" << base << " - "
639 << (base ? base->getName() : "nullptr") << "): ";
644 AssertingVH<Value> base; // non null only if status == base
649 static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
656 // Values of type BDVState form a lattice, and this is a helper
657 // class that implementes the meet operation. The meat of the meet
658 // operation is implemented in MeetBDVStates::pureMeet
659 class MeetBDVStates {
661 /// Initializes the currentResult to the TOP state so that if can be met with
662 /// any other state to produce that state.
665 // Destructively meet the current result with the given BDVState
666 void meetWith(BDVState otherState) {
667 currentResult = meet(otherState, currentResult);
670 BDVState getResult() const { return currentResult; }
673 BDVState currentResult;
675 /// Perform a meet operation on two elements of the BDVState lattice.
676 static BDVState meet(BDVState LHS, BDVState RHS) {
677 assert((pureMeet(LHS, RHS) == pureMeet(RHS, LHS)) &&
678 "math is wrong: meet does not commute!");
679 BDVState Result = pureMeet(LHS, RHS);
680 DEBUG(dbgs() << "meet of " << LHS << " with " << RHS
681 << " produced " << Result << "\n");
685 static BDVState pureMeet(const BDVState &stateA, const BDVState &stateB) {
686 switch (stateA.getStatus()) {
687 case BDVState::Unknown:
691 assert(stateA.getBase() && "can't be null");
692 if (stateB.isUnknown())
695 if (stateB.isBase()) {
696 if (stateA.getBase() == stateB.getBase()) {
697 assert(stateA == stateB && "equality broken!");
700 return BDVState(BDVState::Conflict);
702 assert(stateB.isConflict() && "only three states!");
703 return BDVState(BDVState::Conflict);
705 case BDVState::Conflict:
708 llvm_unreachable("only three states!");
714 /// For a given value or instruction, figure out what base ptr it's derived
715 /// from. For gc objects, this is simply itself. On success, returns a value
716 /// which is the base pointer. (This is reliable and can be used for
717 /// relocation.) On failure, returns nullptr.
718 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
719 Value *def = findBaseOrBDV(I, cache);
721 if (isKnownBaseResult(def)) {
725 // Here's the rough algorithm:
726 // - For every SSA value, construct a mapping to either an actual base
727 // pointer or a PHI which obscures the base pointer.
728 // - Construct a mapping from PHI to unknown TOP state. Use an
729 // optimistic algorithm to propagate base pointer information. Lattice
734 // When algorithm terminates, all PHIs will either have a single concrete
735 // base or be in a conflict state.
736 // - For every conflict, insert a dummy PHI node without arguments. Add
737 // these to the base[Instruction] = BasePtr mapping. For every
738 // non-conflict, add the actual base.
739 // - For every conflict, add arguments for the base[a] of each input
742 // Note: A simpler form of this would be to add the conflict form of all
743 // PHIs without running the optimistic algorithm. This would be
744 // analogous to pessimistic data flow and would likely lead to an
745 // overall worse solution.
748 auto isExpectedBDVType = [](Value *BDV) {
749 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) ||
750 isa<ExtractElementInst>(BDV) || isa<InsertElementInst>(BDV);
754 // Once populated, will contain a mapping from each potentially non-base BDV
755 // to a lattice value (described above) which corresponds to that BDV.
756 // We use the order of insertion (DFS over the def/use graph) to provide a
757 // stable deterministic ordering for visiting DenseMaps (which are unordered)
758 // below. This is important for deterministic compilation.
759 MapVector<Value *, BDVState> States;
761 // Recursively fill in all base defining values reachable from the initial
762 // one for which we don't already know a definite base value for
764 SmallVector<Value*, 16> Worklist;
765 Worklist.push_back(def);
766 States.insert(std::make_pair(def, BDVState()));
767 while (!Worklist.empty()) {
768 Value *Current = Worklist.pop_back_val();
769 assert(!isKnownBaseResult(Current) && "why did it get added?");
771 auto visitIncomingValue = [&](Value *InVal) {
772 Value *Base = findBaseOrBDV(InVal, cache);
773 if (isKnownBaseResult(Base))
774 // Known bases won't need new instructions introduced and can be
777 assert(isExpectedBDVType(Base) && "the only non-base values "
778 "we see should be base defining values");
779 if (States.insert(std::make_pair(Base, BDVState())).second)
780 Worklist.push_back(Base);
782 if (PHINode *Phi = dyn_cast<PHINode>(Current)) {
783 for (Value *InVal : Phi->incoming_values())
784 visitIncomingValue(InVal);
785 } else if (SelectInst *Sel = dyn_cast<SelectInst>(Current)) {
786 visitIncomingValue(Sel->getTrueValue());
787 visitIncomingValue(Sel->getFalseValue());
788 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
789 visitIncomingValue(EE->getVectorOperand());
790 } else if (auto *IE = dyn_cast<InsertElementInst>(Current)) {
791 visitIncomingValue(IE->getOperand(0)); // vector operand
792 visitIncomingValue(IE->getOperand(1)); // scalar operand
794 // There is one known class of instructions we know we don't handle.
795 assert(isa<ShuffleVectorInst>(Current));
796 llvm_unreachable("unimplemented instruction case");
802 DEBUG(dbgs() << "States after initialization:\n");
803 for (auto Pair : States) {
804 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
808 // Return a phi state for a base defining value. We'll generate a new
809 // base state for known bases and expect to find a cached state otherwise.
810 auto getStateForBDV = [&](Value *baseValue) {
811 if (isKnownBaseResult(baseValue))
812 return BDVState(baseValue);
813 auto I = States.find(baseValue);
814 assert(I != States.end() && "lookup failed!");
818 bool progress = true;
821 const size_t oldSize = States.size();
824 // We're only changing values in this loop, thus safe to keep iterators.
825 // Since this is computing a fixed point, the order of visit does not
826 // effect the result. TODO: We could use a worklist here and make this run
828 for (auto Pair : States) {
829 Value *BDV = Pair.first;
830 assert(!isKnownBaseResult(BDV) && "why did it get added?");
832 // Given an input value for the current instruction, return a BDVState
833 // instance which represents the BDV of that value.
834 auto getStateForInput = [&](Value *V) mutable {
835 Value *BDV = findBaseOrBDV(V, cache);
836 return getStateForBDV(BDV);
839 MeetBDVStates calculateMeet;
840 if (SelectInst *select = dyn_cast<SelectInst>(BDV)) {
841 calculateMeet.meetWith(getStateForInput(select->getTrueValue()));
842 calculateMeet.meetWith(getStateForInput(select->getFalseValue()));
843 } else if (PHINode *Phi = dyn_cast<PHINode>(BDV)) {
844 for (Value *Val : Phi->incoming_values())
845 calculateMeet.meetWith(getStateForInput(Val));
846 } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) {
847 // The 'meet' for an extractelement is slightly trivial, but it's still
848 // useful in that it drives us to conflict if our input is.
849 calculateMeet.meetWith(getStateForInput(EE->getVectorOperand()));
851 // Given there's a inherent type mismatch between the operands, will
852 // *always* produce Conflict.
853 auto *IE = cast<InsertElementInst>(BDV);
854 calculateMeet.meetWith(getStateForInput(IE->getOperand(0)));
855 calculateMeet.meetWith(getStateForInput(IE->getOperand(1)));
858 BDVState oldState = States[BDV];
859 BDVState newState = calculateMeet.getResult();
860 if (oldState != newState) {
862 States[BDV] = newState;
866 assert(oldSize == States.size() &&
867 "fixed point shouldn't be adding any new nodes to state");
871 DEBUG(dbgs() << "States after meet iteration:\n");
872 for (auto Pair : States) {
873 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
877 // Insert Phis for all conflicts
878 // TODO: adjust naming patterns to avoid this order of iteration dependency
879 for (auto Pair : States) {
880 Instruction *I = cast<Instruction>(Pair.first);
881 BDVState State = Pair.second;
882 assert(!isKnownBaseResult(I) && "why did it get added?");
883 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
885 // extractelement instructions are a bit special in that we may need to
886 // insert an extract even when we know an exact base for the instruction.
887 // The problem is that we need to convert from a vector base to a scalar
888 // base for the particular indice we're interested in.
889 if (State.isBase() && isa<ExtractElementInst>(I) &&
890 isa<VectorType>(State.getBase()->getType())) {
891 auto *EE = cast<ExtractElementInst>(I);
892 // TODO: In many cases, the new instruction is just EE itself. We should
893 // exploit this, but can't do it here since it would break the invariant
894 // about the BDV not being known to be a base.
895 auto *BaseInst = ExtractElementInst::Create(State.getBase(),
896 EE->getIndexOperand(),
898 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
899 States[I] = BDVState(BDVState::Base, BaseInst);
902 // Since we're joining a vector and scalar base, they can never be the
903 // same. As a result, we should always see insert element having reached
904 // the conflict state.
905 if (isa<InsertElementInst>(I)) {
906 assert(State.isConflict());
909 if (!State.isConflict())
912 /// Create and insert a new instruction which will represent the base of
913 /// the given instruction 'I'.
914 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
915 if (isa<PHINode>(I)) {
916 BasicBlock *BB = I->getParent();
917 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
918 assert(NumPreds > 0 && "how did we reach here");
919 std::string Name = suffixed_name_or(I, ".base", "base_phi");
920 return PHINode::Create(I->getType(), NumPreds, Name, I);
921 } else if (SelectInst *Sel = dyn_cast<SelectInst>(I)) {
922 // The undef will be replaced later
923 UndefValue *Undef = UndefValue::get(Sel->getType());
924 std::string Name = suffixed_name_or(I, ".base", "base_select");
925 return SelectInst::Create(Sel->getCondition(), Undef,
927 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
928 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
929 std::string Name = suffixed_name_or(I, ".base", "base_ee");
930 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
933 auto *IE = cast<InsertElementInst>(I);
934 UndefValue *VecUndef = UndefValue::get(IE->getOperand(0)->getType());
935 UndefValue *ScalarUndef = UndefValue::get(IE->getOperand(1)->getType());
936 std::string Name = suffixed_name_or(I, ".base", "base_ie");
937 return InsertElementInst::Create(VecUndef, ScalarUndef,
938 IE->getOperand(2), Name, IE);
942 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
943 // Add metadata marking this as a base value
944 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
945 States[I] = BDVState(BDVState::Conflict, BaseInst);
948 // Returns a instruction which produces the base pointer for a given
949 // instruction. The instruction is assumed to be an input to one of the BDVs
950 // seen in the inference algorithm above. As such, we must either already
951 // know it's base defining value is a base, or have inserted a new
952 // instruction to propagate the base of it's BDV and have entered that newly
953 // introduced instruction into the state table. In either case, we are
954 // assured to be able to determine an instruction which produces it's base
956 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
957 Value *BDV = findBaseOrBDV(Input, cache);
958 Value *Base = nullptr;
959 if (isKnownBaseResult(BDV)) {
962 // Either conflict or base.
963 assert(States.count(BDV));
964 Base = States[BDV].getBase();
966 assert(Base && "can't be null");
967 // The cast is needed since base traversal may strip away bitcasts
968 if (Base->getType() != Input->getType() &&
970 Base = new BitCastInst(Base, Input->getType(), "cast",
976 // Fixup all the inputs of the new PHIs. Visit order needs to be
977 // deterministic and predictable because we're naming newly created
979 for (auto Pair : States) {
980 Instruction *BDV = cast<Instruction>(Pair.first);
981 BDVState State = Pair.second;
983 assert(!isKnownBaseResult(BDV) && "why did it get added?");
984 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
985 if (!State.isConflict())
988 if (PHINode *basephi = dyn_cast<PHINode>(State.getBase())) {
989 PHINode *phi = cast<PHINode>(BDV);
990 unsigned NumPHIValues = phi->getNumIncomingValues();
991 for (unsigned i = 0; i < NumPHIValues; i++) {
992 Value *InVal = phi->getIncomingValue(i);
993 BasicBlock *InBB = phi->getIncomingBlock(i);
995 // If we've already seen InBB, add the same incoming value
996 // we added for it earlier. The IR verifier requires phi
997 // nodes with multiple entries from the same basic block
998 // to have the same incoming value for each of those
999 // entries. If we don't do this check here and basephi
1000 // has a different type than base, we'll end up adding two
1001 // bitcasts (and hence two distinct values) as incoming
1002 // values for the same basic block.
1004 int blockIndex = basephi->getBasicBlockIndex(InBB);
1005 if (blockIndex != -1) {
1006 Value *oldBase = basephi->getIncomingValue(blockIndex);
1007 basephi->addIncoming(oldBase, InBB);
1010 Value *Base = getBaseForInput(InVal, nullptr);
1011 // In essence this assert states: the only way two
1012 // values incoming from the same basic block may be
1013 // different is by being different bitcasts of the same
1014 // value. A cleanup that remains TODO is changing
1015 // findBaseOrBDV to return an llvm::Value of the correct
1016 // type (and still remain pure). This will remove the
1017 // need to add bitcasts.
1018 assert(Base->stripPointerCasts() == oldBase->stripPointerCasts() &&
1019 "sanity -- findBaseOrBDV should be pure!");
1024 // Find the instruction which produces the base for each input. We may
1025 // need to insert a bitcast in the incoming block.
1026 // TODO: Need to split critical edges if insertion is needed
1027 Value *Base = getBaseForInput(InVal, InBB->getTerminator());
1028 basephi->addIncoming(Base, InBB);
1030 assert(basephi->getNumIncomingValues() == NumPHIValues);
1031 } else if (SelectInst *BaseSel = dyn_cast<SelectInst>(State.getBase())) {
1032 SelectInst *Sel = cast<SelectInst>(BDV);
1033 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
1034 // something more safe and less hacky.
1035 for (int i = 1; i <= 2; i++) {
1036 Value *InVal = Sel->getOperand(i);
1037 // Find the instruction which produces the base for each input. We may
1038 // need to insert a bitcast.
1039 Value *Base = getBaseForInput(InVal, BaseSel);
1040 BaseSel->setOperand(i, Base);
1042 } else if (auto *BaseEE = dyn_cast<ExtractElementInst>(State.getBase())) {
1043 Value *InVal = cast<ExtractElementInst>(BDV)->getVectorOperand();
1044 // Find the instruction which produces the base for each input. We may
1045 // need to insert a bitcast.
1046 Value *Base = getBaseForInput(InVal, BaseEE);
1047 BaseEE->setOperand(0, Base);
1049 auto *BaseIE = cast<InsertElementInst>(State.getBase());
1050 auto *BdvIE = cast<InsertElementInst>(BDV);
1051 auto UpdateOperand = [&](int OperandIdx) {
1052 Value *InVal = BdvIE->getOperand(OperandIdx);
1053 Value *Base = getBaseForInput(InVal, BaseIE);
1054 BaseIE->setOperand(OperandIdx, Base);
1056 UpdateOperand(0); // vector operand
1057 UpdateOperand(1); // scalar operand
1062 // Now that we're done with the algorithm, see if we can optimize the
1063 // results slightly by reducing the number of new instructions needed.
1064 // Arguably, this should be integrated into the algorithm above, but
1065 // doing as a post process step is easier to reason about for the moment.
1066 DenseMap<Value *, Value *> ReverseMap;
1067 SmallPtrSet<Instruction *, 16> NewInsts;
1068 SmallSetVector<AssertingVH<Instruction>, 16> Worklist;
1069 // Note: We need to visit the states in a deterministic order. We uses the
1070 // Keys we sorted above for this purpose. Note that we are papering over a
1071 // bigger problem with the algorithm above - it's visit order is not
1072 // deterministic. A larger change is needed to fix this.
1073 for (auto Pair : States) {
1074 auto *BDV = Pair.first;
1075 auto State = Pair.second;
1076 Value *Base = State.getBase();
1077 assert(BDV && Base);
1078 assert(!isKnownBaseResult(BDV) && "why did it get added?");
1079 assert(isKnownBaseResult(Base) &&
1080 "must be something we 'know' is a base pointer");
1081 if (!State.isConflict())
1084 ReverseMap[Base] = BDV;
1085 if (auto *BaseI = dyn_cast<Instruction>(Base)) {
1086 NewInsts.insert(BaseI);
1087 Worklist.insert(BaseI);
1090 auto ReplaceBaseInstWith = [&](Value *BDV, Instruction *BaseI,
1091 Value *Replacement) {
1092 // Add users which are new instructions (excluding self references)
1093 for (User *U : BaseI->users())
1094 if (auto *UI = dyn_cast<Instruction>(U))
1095 if (NewInsts.count(UI) && UI != BaseI)
1096 Worklist.insert(UI);
1097 // Then do the actual replacement
1098 NewInsts.erase(BaseI);
1099 ReverseMap.erase(BaseI);
1100 BaseI->replaceAllUsesWith(Replacement);
1101 assert(States.count(BDV));
1102 assert(States[BDV].isConflict() && States[BDV].getBase() == BaseI);
1103 States[BDV] = BDVState(BDVState::Conflict, Replacement);
1104 BaseI->eraseFromParent();
1106 const DataLayout &DL = cast<Instruction>(def)->getModule()->getDataLayout();
1107 while (!Worklist.empty()) {
1108 Instruction *BaseI = Worklist.pop_back_val();
1109 assert(NewInsts.count(BaseI));
1110 Value *Bdv = ReverseMap[BaseI];
1111 if (auto *BdvI = dyn_cast<Instruction>(Bdv))
1112 if (BaseI->isIdenticalTo(BdvI)) {
1113 DEBUG(dbgs() << "Identical Base: " << *BaseI << "\n");
1114 ReplaceBaseInstWith(Bdv, BaseI, Bdv);
1117 if (Value *V = SimplifyInstruction(BaseI, DL)) {
1118 DEBUG(dbgs() << "Base " << *BaseI << " simplified to " << *V << "\n");
1119 ReplaceBaseInstWith(Bdv, BaseI, V);
1124 // Cache all of our results so we can cheaply reuse them
1125 // NOTE: This is actually two caches: one of the base defining value
1126 // relation and one of the base pointer relation! FIXME
1127 for (auto Pair : States) {
1128 auto *BDV = Pair.first;
1129 Value *base = Pair.second.getBase();
1130 assert(BDV && base);
1132 std::string fromstr = cache.count(BDV) ? cache[BDV]->getName() : "none";
1133 DEBUG(dbgs() << "Updating base value cache"
1134 << " for: " << BDV->getName()
1135 << " from: " << fromstr
1136 << " to: " << base->getName() << "\n");
1138 if (cache.count(BDV)) {
1139 // Once we transition from the BDV relation being store in the cache to
1140 // the base relation being stored, it must be stable
1141 assert((!isKnownBaseResult(cache[BDV]) || cache[BDV] == base) &&
1142 "base relation should be stable");
1146 assert(cache.find(def) != cache.end());
1150 // For a set of live pointers (base and/or derived), identify the base
1151 // pointer of the object which they are derived from. This routine will
1152 // mutate the IR graph as needed to make the 'base' pointer live at the
1153 // definition site of 'derived'. This ensures that any use of 'derived' can
1154 // also use 'base'. This may involve the insertion of a number of
1155 // additional PHI nodes.
1157 // preconditions: live is a set of pointer type Values
1159 // side effects: may insert PHI nodes into the existing CFG, will preserve
1160 // CFG, will not remove or mutate any existing nodes
1162 // post condition: PointerToBase contains one (derived, base) pair for every
1163 // pointer in live. Note that derived can be equal to base if the original
1164 // pointer was a base pointer.
1166 findBasePointers(const StatepointLiveSetTy &live,
1167 DenseMap<Value *, Value *> &PointerToBase,
1168 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1169 // For the naming of values inserted to be deterministic - which makes for
1170 // much cleaner and more stable tests - we need to assign an order to the
1171 // live values. DenseSets do not provide a deterministic order across runs.
1172 SmallVector<Value *, 64> Temp;
1173 Temp.insert(Temp.end(), live.begin(), live.end());
1174 std::sort(Temp.begin(), Temp.end(), order_by_name);
1175 for (Value *ptr : Temp) {
1176 Value *base = findBasePointer(ptr, DVCache);
1177 assert(base && "failed to find base pointer");
1178 PointerToBase[ptr] = base;
1179 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1180 DT->dominates(cast<Instruction>(base)->getParent(),
1181 cast<Instruction>(ptr)->getParent())) &&
1182 "The base we found better dominate the derived pointer");
1184 // If you see this trip and like to live really dangerously, the code should
1185 // be correct, just with idioms the verifier can't handle. You can try
1186 // disabling the verifier at your own substantial risk.
1187 assert(!isa<ConstantPointerNull>(base) &&
1188 "the relocation code needs adjustment to handle the relocation of "
1189 "a null pointer constant without causing false positives in the "
1190 "safepoint ir verifier.");
1194 /// Find the required based pointers (and adjust the live set) for the given
1196 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1198 PartiallyConstructedSafepointRecord &result) {
1199 DenseMap<Value *, Value *> PointerToBase;
1200 findBasePointers(result.LiveSet, PointerToBase, &DT, DVCache);
1202 if (PrintBasePointers) {
1203 // Note: Need to print these in a stable order since this is checked in
1205 errs() << "Base Pairs (w/o Relocation):\n";
1206 SmallVector<Value *, 64> Temp;
1207 Temp.reserve(PointerToBase.size());
1208 for (auto Pair : PointerToBase) {
1209 Temp.push_back(Pair.first);
1211 std::sort(Temp.begin(), Temp.end(), order_by_name);
1212 for (Value *Ptr : Temp) {
1213 Value *Base = PointerToBase[Ptr];
1214 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
1219 result.PointerToBase = PointerToBase;
1222 /// Given an updated version of the dataflow liveness results, update the
1223 /// liveset and base pointer maps for the call site CS.
1224 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1226 PartiallyConstructedSafepointRecord &result);
1228 static void recomputeLiveInValues(
1229 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1230 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1231 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1232 // again. The old values are still live and will help it stabilize quickly.
1233 GCPtrLivenessData RevisedLivenessData;
1234 computeLiveInValues(DT, F, RevisedLivenessData);
1235 for (size_t i = 0; i < records.size(); i++) {
1236 struct PartiallyConstructedSafepointRecord &info = records[i];
1237 const CallSite &CS = toUpdate[i];
1238 recomputeLiveInValues(RevisedLivenessData, CS, info);
1242 // When inserting gc.relocate and gc.result calls, we need to ensure there are
1243 // no uses of the original value / return value between the gc.statepoint and
1244 // the gc.relocate / gc.result call. One case which can arise is a phi node
1245 // starting one of the successor blocks. We also need to be able to insert the
1246 // gc.relocates only on the path which goes through the statepoint. We might
1247 // need to split an edge to make this possible.
1249 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1250 DominatorTree &DT) {
1251 BasicBlock *Ret = BB;
1252 if (!BB->getUniquePredecessor())
1253 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1255 // Now that 'Ret' has unique predecessor we can safely remove all phi nodes
1257 FoldSingleEntryPHINodes(Ret);
1258 assert(!isa<PHINode>(Ret->begin()) &&
1259 "All PHI nodes should have been removed!");
1261 // At this point, we can safely insert a gc.relocate or gc.result as the first
1262 // instruction in Ret if needed.
1266 // Create new attribute set containing only attributes which can be transferred
1267 // from original call to the safepoint.
1268 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1271 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1272 unsigned Index = AS.getSlotIndex(Slot);
1274 if (Index == AttributeSet::ReturnIndex ||
1275 Index == AttributeSet::FunctionIndex) {
1277 for (Attribute Attr : make_range(AS.begin(Slot), AS.end(Slot))) {
1279 // Do not allow certain attributes - just skip them
1280 // Safepoint can not be read only or read none.
1281 if (Attr.hasAttribute(Attribute::ReadNone) ||
1282 Attr.hasAttribute(Attribute::ReadOnly))
1285 // These attributes control the generation of the gc.statepoint call /
1286 // invoke itself; and once the gc.statepoint is in place, they're of no
1288 if (Attr.hasAttribute("statepoint-num-patch-bytes") ||
1289 Attr.hasAttribute("statepoint-id"))
1292 Ret = Ret.addAttributes(
1293 AS.getContext(), Index,
1294 AttributeSet::get(AS.getContext(), Index, AttrBuilder(Attr)));
1298 // Just skip parameter attributes for now
1304 /// Helper function to place all gc relocates necessary for the given
1307 /// liveVariables - list of variables to be relocated.
1308 /// liveStart - index of the first live variable.
1309 /// basePtrs - base pointers.
1310 /// statepointToken - statepoint instruction to which relocates should be
1312 /// Builder - Llvm IR builder to be used to construct new calls.
1313 static void CreateGCRelocates(ArrayRef<Value *> LiveVariables,
1314 const int LiveStart,
1315 ArrayRef<Value *> BasePtrs,
1316 Instruction *StatepointToken,
1317 IRBuilder<> Builder) {
1318 if (LiveVariables.empty())
1321 auto FindIndex = [](ArrayRef<Value *> LiveVec, Value *Val) {
1322 auto ValIt = std::find(LiveVec.begin(), LiveVec.end(), Val);
1323 assert(ValIt != LiveVec.end() && "Val not found in LiveVec!");
1324 size_t Index = std::distance(LiveVec.begin(), ValIt);
1325 assert(Index < LiveVec.size() && "Bug in std::find?");
1329 // All gc_relocate are set to i8 addrspace(1)* type. We originally generated
1330 // unique declarations for each pointer type, but this proved problematic
1331 // because the intrinsic mangling code is incomplete and fragile. Since
1332 // we're moving towards a single unified pointer type anyways, we can just
1333 // cast everything to an i8* of the right address space. A bitcast is added
1334 // later to convert gc_relocate to the actual value's type.
1335 Module *M = StatepointToken->getModule();
1336 auto AS = cast<PointerType>(LiveVariables[0]->getType())->getAddressSpace();
1337 Type *Types[] = {Type::getInt8PtrTy(M->getContext(), AS)};
1338 Value *GCRelocateDecl =
1339 Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, Types);
1341 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1342 // Generate the gc.relocate call and save the result
1344 Builder.getInt32(LiveStart + FindIndex(LiveVariables, BasePtrs[i]));
1345 Value *LiveIdx = Builder.getInt32(LiveStart + i);
1347 // only specify a debug name if we can give a useful one
1348 CallInst *Reloc = Builder.CreateCall(
1349 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1350 suffixed_name_or(LiveVariables[i], ".relocated", ""));
1351 // Trick CodeGen into thinking there are lots of free registers at this
1353 Reloc->setCallingConv(CallingConv::Cold);
1359 /// This struct is used to defer RAUWs and `eraseFromParent` s. Using this
1360 /// avoids having to worry about keeping around dangling pointers to Values.
1361 class DeferredReplacement {
1362 AssertingVH<Instruction> Old;
1363 AssertingVH<Instruction> New;
1366 explicit DeferredReplacement(Instruction *Old, Instruction *New) :
1367 Old(Old), New(New) {
1368 assert(Old != New && "Not allowed!");
1371 /// Does the task represented by this instance.
1372 void doReplacement() {
1373 Instruction *OldI = Old;
1374 Instruction *NewI = New;
1376 assert(OldI != NewI && "Disallowed at construction?!");
1382 OldI->replaceAllUsesWith(NewI);
1383 OldI->eraseFromParent();
1389 makeStatepointExplicitImpl(const CallSite CS, /* to replace */
1390 const SmallVectorImpl<Value *> &BasePtrs,
1391 const SmallVectorImpl<Value *> &LiveVariables,
1392 PartiallyConstructedSafepointRecord &Result,
1393 std::vector<DeferredReplacement> &Replacements) {
1394 assert(BasePtrs.size() == LiveVariables.size());
1395 assert((UseDeoptBundles || isStatepoint(CS)) &&
1396 "This method expects to be rewriting a statepoint");
1398 // Then go ahead and use the builder do actually do the inserts. We insert
1399 // immediately before the previous instruction under the assumption that all
1400 // arguments will be available here. We can't insert afterwards since we may
1401 // be replacing a terminator.
1402 Instruction *InsertBefore = CS.getInstruction();
1403 IRBuilder<> Builder(InsertBefore);
1405 ArrayRef<Value *> GCArgs(LiveVariables);
1406 uint64_t StatepointID = 0xABCDEF00;
1407 uint32_t NumPatchBytes = 0;
1408 uint32_t Flags = uint32_t(StatepointFlags::None);
1410 ArrayRef<Use> CallArgs;
1411 ArrayRef<Use> DeoptArgs;
1412 ArrayRef<Use> TransitionArgs;
1414 Value *CallTarget = nullptr;
1416 if (UseDeoptBundles) {
1417 CallArgs = {CS.arg_begin(), CS.arg_end()};
1418 DeoptArgs = GetDeoptBundleOperands(CS);
1419 // TODO: we don't fill in TransitionArgs or Flags in this branch, but we
1420 // could have an operand bundle for that too.
1421 AttributeSet OriginalAttrs = CS.getAttributes();
1423 Attribute AttrID = OriginalAttrs.getAttribute(AttributeSet::FunctionIndex,
1425 if (AttrID.isStringAttribute())
1426 AttrID.getValueAsString().getAsInteger(10, StatepointID);
1428 Attribute AttrNumPatchBytes = OriginalAttrs.getAttribute(
1429 AttributeSet::FunctionIndex, "statepoint-num-patch-bytes");
1430 if (AttrNumPatchBytes.isStringAttribute())
1431 AttrNumPatchBytes.getValueAsString().getAsInteger(10, NumPatchBytes);
1433 CallTarget = CS.getCalledValue();
1435 // This branch will be gone soon, and we will soon only support the
1436 // UseDeoptBundles == true configuration.
1437 Statepoint OldSP(CS);
1438 StatepointID = OldSP.getID();
1439 NumPatchBytes = OldSP.getNumPatchBytes();
1440 Flags = OldSP.getFlags();
1442 CallArgs = {OldSP.arg_begin(), OldSP.arg_end()};
1443 DeoptArgs = {OldSP.vm_state_begin(), OldSP.vm_state_end()};
1444 TransitionArgs = {OldSP.gc_transition_args_begin(),
1445 OldSP.gc_transition_args_end()};
1446 CallTarget = OldSP.getCalledValue();
1449 // Create the statepoint given all the arguments
1450 Instruction *Token = nullptr;
1451 AttributeSet ReturnAttrs;
1453 CallInst *ToReplace = cast<CallInst>(CS.getInstruction());
1454 CallInst *Call = Builder.CreateGCStatepointCall(
1455 StatepointID, NumPatchBytes, CallTarget, Flags, CallArgs,
1456 TransitionArgs, DeoptArgs, GCArgs, "safepoint_token");
1458 Call->setTailCall(ToReplace->isTailCall());
1459 Call->setCallingConv(ToReplace->getCallingConv());
1461 // Currently we will fail on parameter attributes and on certain
1462 // function attributes.
1463 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes());
1464 // In case if we can handle this set of attributes - set up function attrs
1465 // directly on statepoint and return attrs later for gc_result intrinsic.
1466 Call->setAttributes(NewAttrs.getFnAttributes());
1467 ReturnAttrs = NewAttrs.getRetAttributes();
1471 // Put the following gc_result and gc_relocate calls immediately after the
1472 // the old call (which we're about to delete)
1473 assert(ToReplace->getNextNode() && "Not a terminator, must have next!");
1474 Builder.SetInsertPoint(ToReplace->getNextNode());
1475 Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc());
1477 InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction());
1479 // Insert the new invoke into the old block. We'll remove the old one in a
1480 // moment at which point this will become the new terminator for the
1482 InvokeInst *Invoke = Builder.CreateGCStatepointInvoke(
1483 StatepointID, NumPatchBytes, CallTarget, ToReplace->getNormalDest(),
1484 ToReplace->getUnwindDest(), Flags, CallArgs, TransitionArgs, DeoptArgs,
1485 GCArgs, "statepoint_token");
1487 Invoke->setCallingConv(ToReplace->getCallingConv());
1489 // Currently we will fail on parameter attributes and on certain
1490 // function attributes.
1491 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes());
1492 // In case if we can handle this set of attributes - set up function attrs
1493 // directly on statepoint and return attrs later for gc_result intrinsic.
1494 Invoke->setAttributes(NewAttrs.getFnAttributes());
1495 ReturnAttrs = NewAttrs.getRetAttributes();
1499 // Generate gc relocates in exceptional path
1500 BasicBlock *UnwindBlock = ToReplace->getUnwindDest();
1501 assert(!isa<PHINode>(UnwindBlock->begin()) &&
1502 UnwindBlock->getUniquePredecessor() &&
1503 "can't safely insert in this block!");
1505 Builder.SetInsertPoint(&*UnwindBlock->getFirstInsertionPt());
1506 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1508 // Extract second element from landingpad return value. We will attach
1509 // exceptional gc relocates to it.
1510 Instruction *ExceptionalToken =
1511 cast<Instruction>(Builder.CreateExtractValue(
1512 UnwindBlock->getLandingPadInst(), 1, "relocate_token"));
1513 Result.UnwindToken = ExceptionalToken;
1515 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1516 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, ExceptionalToken,
1519 // Generate gc relocates and returns for normal block
1520 BasicBlock *NormalDest = ToReplace->getNormalDest();
1521 assert(!isa<PHINode>(NormalDest->begin()) &&
1522 NormalDest->getUniquePredecessor() &&
1523 "can't safely insert in this block!");
1525 Builder.SetInsertPoint(&*NormalDest->getFirstInsertionPt());
1527 // gc relocates will be generated later as if it were regular call
1530 assert(Token && "Should be set in one of the above branches!");
1532 if (UseDeoptBundles) {
1533 Token->setName("statepoint_token");
1534 if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
1536 CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "";
1537 CallInst *GCResult = Builder.CreateGCResult(Token, CS.getType(), Name);
1538 GCResult->setAttributes(CS.getAttributes().getRetAttributes());
1540 // We cannot RAUW or delete CS.getInstruction() because it could be in the
1541 // live set of some other safepoint, in which case that safepoint's
1542 // PartiallyConstructedSafepointRecord will hold a raw pointer to this
1543 // llvm::Instruction. Instead, we defer the replacement and deletion to
1544 // after the live sets have been made explicit in the IR, and we no longer
1545 // have raw pointers to worry about.
1546 Replacements.emplace_back(CS.getInstruction(), GCResult);
1548 Replacements.emplace_back(CS.getInstruction(), nullptr);
1551 assert(!CS.getInstruction()->hasNUsesOrMore(2) &&
1552 "only valid use before rewrite is gc.result");
1553 assert(!CS.getInstruction()->hasOneUse() ||
1554 isGCResult(cast<Instruction>(*CS.getInstruction()->user_begin())));
1556 // Take the name of the original statepoint token if there was one.
1557 Token->takeName(CS.getInstruction());
1559 // Update the gc.result of the original statepoint (if any) to use the newly
1560 // inserted statepoint. This is safe to do here since the token can't be
1561 // considered a live reference.
1562 CS.getInstruction()->replaceAllUsesWith(Token);
1563 CS.getInstruction()->eraseFromParent();
1566 Result.StatepointToken = Token;
1568 // Second, create a gc.relocate for every live variable
1569 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1570 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, Token, Builder);
1574 struct NameOrdering {
1578 bool operator()(NameOrdering const &a, NameOrdering const &b) {
1579 return -1 == a.Derived->getName().compare(b.Derived->getName());
1584 static void StabilizeOrder(SmallVectorImpl<Value *> &BaseVec,
1585 SmallVectorImpl<Value *> &LiveVec) {
1586 assert(BaseVec.size() == LiveVec.size());
1588 SmallVector<NameOrdering, 64> Temp;
1589 for (size_t i = 0; i < BaseVec.size(); i++) {
1591 v.Base = BaseVec[i];
1592 v.Derived = LiveVec[i];
1596 std::sort(Temp.begin(), Temp.end(), NameOrdering());
1597 for (size_t i = 0; i < BaseVec.size(); i++) {
1598 BaseVec[i] = Temp[i].Base;
1599 LiveVec[i] = Temp[i].Derived;
1603 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1604 // which make the relocations happening at this safepoint explicit.
1606 // WARNING: Does not do any fixup to adjust users of the original live
1607 // values. That's the callers responsibility.
1609 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS,
1610 PartiallyConstructedSafepointRecord &Result,
1611 std::vector<DeferredReplacement> &Replacements) {
1612 const auto &LiveSet = Result.LiveSet;
1613 const auto &PointerToBase = Result.PointerToBase;
1615 // Convert to vector for efficient cross referencing.
1616 SmallVector<Value *, 64> BaseVec, LiveVec;
1617 LiveVec.reserve(LiveSet.size());
1618 BaseVec.reserve(LiveSet.size());
1619 for (Value *L : LiveSet) {
1620 LiveVec.push_back(L);
1621 assert(PointerToBase.count(L));
1622 Value *Base = PointerToBase.find(L)->second;
1623 BaseVec.push_back(Base);
1625 assert(LiveVec.size() == BaseVec.size());
1627 // To make the output IR slightly more stable (for use in diffs), ensure a
1628 // fixed order of the values in the safepoint (by sorting the value name).
1629 // The order is otherwise meaningless.
1630 StabilizeOrder(BaseVec, LiveVec);
1632 // Do the actual rewriting and delete the old statepoint
1633 makeStatepointExplicitImpl(CS, BaseVec, LiveVec, Result, Replacements);
1636 // Helper function for the relocationViaAlloca.
1638 // It receives iterator to the statepoint gc relocates and emits a store to the
1639 // assigned location (via allocaMap) for the each one of them. It adds the
1640 // visited values into the visitedLiveValues set, which we will later use them
1641 // for sanity checking.
1643 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1644 DenseMap<Value *, Value *> &AllocaMap,
1645 DenseSet<Value *> &VisitedLiveValues) {
1647 for (User *U : GCRelocs) {
1648 if (!isa<IntrinsicInst>(U))
1651 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1653 // We only care about relocates
1654 if (RelocatedValue->getIntrinsicID() !=
1655 Intrinsic::experimental_gc_relocate) {
1659 GCRelocateOperands RelocateOperands(RelocatedValue);
1660 Value *OriginalValue =
1661 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1662 assert(AllocaMap.count(OriginalValue));
1663 Value *Alloca = AllocaMap[OriginalValue];
1665 // Emit store into the related alloca
1666 // All gc_relocates are i8 addrspace(1)* typed, and it must be bitcasted to
1667 // the correct type according to alloca.
1668 assert(RelocatedValue->getNextNode() &&
1669 "Should always have one since it's not a terminator");
1670 IRBuilder<> Builder(RelocatedValue->getNextNode());
1671 Value *CastedRelocatedValue =
1672 Builder.CreateBitCast(RelocatedValue,
1673 cast<AllocaInst>(Alloca)->getAllocatedType(),
1674 suffixed_name_or(RelocatedValue, ".casted", ""));
1676 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1677 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1680 VisitedLiveValues.insert(OriginalValue);
1685 // Helper function for the "relocationViaAlloca". Similar to the
1686 // "insertRelocationStores" but works for rematerialized values.
1688 insertRematerializationStores(
1689 RematerializedValueMapTy RematerializedValues,
1690 DenseMap<Value *, Value *> &AllocaMap,
1691 DenseSet<Value *> &VisitedLiveValues) {
1693 for (auto RematerializedValuePair: RematerializedValues) {
1694 Instruction *RematerializedValue = RematerializedValuePair.first;
1695 Value *OriginalValue = RematerializedValuePair.second;
1697 assert(AllocaMap.count(OriginalValue) &&
1698 "Can not find alloca for rematerialized value");
1699 Value *Alloca = AllocaMap[OriginalValue];
1701 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1702 Store->insertAfter(RematerializedValue);
1705 VisitedLiveValues.insert(OriginalValue);
1710 /// Do all the relocation update via allocas and mem2reg
1711 static void relocationViaAlloca(
1712 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1713 ArrayRef<PartiallyConstructedSafepointRecord> Records) {
1715 // record initial number of (static) allocas; we'll check we have the same
1716 // number when we get done.
1717 int InitialAllocaNum = 0;
1718 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1720 if (isa<AllocaInst>(*I))
1724 // TODO-PERF: change data structures, reserve
1725 DenseMap<Value *, Value *> AllocaMap;
1726 SmallVector<AllocaInst *, 200> PromotableAllocas;
1727 // Used later to chack that we have enough allocas to store all values
1728 std::size_t NumRematerializedValues = 0;
1729 PromotableAllocas.reserve(Live.size());
1731 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1732 // "PromotableAllocas"
1733 auto emitAllocaFor = [&](Value *LiveValue) {
1734 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1735 F.getEntryBlock().getFirstNonPHI());
1736 AllocaMap[LiveValue] = Alloca;
1737 PromotableAllocas.push_back(Alloca);
1740 // Emit alloca for each live gc pointer
1741 for (Value *V : Live)
1744 // Emit allocas for rematerialized values
1745 for (const auto &Info : Records)
1746 for (auto RematerializedValuePair : Info.RematerializedValues) {
1747 Value *OriginalValue = RematerializedValuePair.second;
1748 if (AllocaMap.count(OriginalValue) != 0)
1751 emitAllocaFor(OriginalValue);
1752 ++NumRematerializedValues;
1755 // The next two loops are part of the same conceptual operation. We need to
1756 // insert a store to the alloca after the original def and at each
1757 // redefinition. We need to insert a load before each use. These are split
1758 // into distinct loops for performance reasons.
1760 // Update gc pointer after each statepoint: either store a relocated value or
1761 // null (if no relocated value was found for this gc pointer and it is not a
1762 // gc_result). This must happen before we update the statepoint with load of
1763 // alloca otherwise we lose the link between statepoint and old def.
1764 for (const auto &Info : Records) {
1765 Value *Statepoint = Info.StatepointToken;
1767 // This will be used for consistency check
1768 DenseSet<Value *> VisitedLiveValues;
1770 // Insert stores for normal statepoint gc relocates
1771 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1773 // In case if it was invoke statepoint
1774 // we will insert stores for exceptional path gc relocates.
1775 if (isa<InvokeInst>(Statepoint)) {
1776 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1780 // Do similar thing with rematerialized values
1781 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1784 if (ClobberNonLive) {
1785 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1786 // the gc.statepoint. This will turn some subtle GC problems into
1787 // slightly easier to debug SEGVs. Note that on large IR files with
1788 // lots of gc.statepoints this is extremely costly both memory and time
1790 SmallVector<AllocaInst *, 64> ToClobber;
1791 for (auto Pair : AllocaMap) {
1792 Value *Def = Pair.first;
1793 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1795 // This value was relocated
1796 if (VisitedLiveValues.count(Def)) {
1799 ToClobber.push_back(Alloca);
1802 auto InsertClobbersAt = [&](Instruction *IP) {
1803 for (auto *AI : ToClobber) {
1804 auto AIType = cast<PointerType>(AI->getType());
1805 auto PT = cast<PointerType>(AIType->getElementType());
1806 Constant *CPN = ConstantPointerNull::get(PT);
1807 StoreInst *Store = new StoreInst(CPN, AI);
1808 Store->insertBefore(IP);
1812 // Insert the clobbering stores. These may get intermixed with the
1813 // gc.results and gc.relocates, but that's fine.
1814 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1815 InsertClobbersAt(&*II->getNormalDest()->getFirstInsertionPt());
1816 InsertClobbersAt(&*II->getUnwindDest()->getFirstInsertionPt());
1818 InsertClobbersAt(cast<Instruction>(Statepoint)->getNextNode());
1823 // Update use with load allocas and add store for gc_relocated.
1824 for (auto Pair : AllocaMap) {
1825 Value *Def = Pair.first;
1826 Value *Alloca = Pair.second;
1828 // We pre-record the uses of allocas so that we dont have to worry about
1829 // later update that changes the user information..
1831 SmallVector<Instruction *, 20> Uses;
1832 // PERF: trade a linear scan for repeated reallocation
1833 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1834 for (User *U : Def->users()) {
1835 if (!isa<ConstantExpr>(U)) {
1836 // If the def has a ConstantExpr use, then the def is either a
1837 // ConstantExpr use itself or null. In either case
1838 // (recursively in the first, directly in the second), the oop
1839 // it is ultimately dependent on is null and this particular
1840 // use does not need to be fixed up.
1841 Uses.push_back(cast<Instruction>(U));
1845 std::sort(Uses.begin(), Uses.end());
1846 auto Last = std::unique(Uses.begin(), Uses.end());
1847 Uses.erase(Last, Uses.end());
1849 for (Instruction *Use : Uses) {
1850 if (isa<PHINode>(Use)) {
1851 PHINode *Phi = cast<PHINode>(Use);
1852 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1853 if (Def == Phi->getIncomingValue(i)) {
1854 LoadInst *Load = new LoadInst(
1855 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1856 Phi->setIncomingValue(i, Load);
1860 LoadInst *Load = new LoadInst(Alloca, "", Use);
1861 Use->replaceUsesOfWith(Def, Load);
1865 // Emit store for the initial gc value. Store must be inserted after load,
1866 // otherwise store will be in alloca's use list and an extra load will be
1867 // inserted before it.
1868 StoreInst *Store = new StoreInst(Def, Alloca);
1869 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1870 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1871 // InvokeInst is a TerminatorInst so the store need to be inserted
1872 // into its normal destination block.
1873 BasicBlock *NormalDest = Invoke->getNormalDest();
1874 Store->insertBefore(NormalDest->getFirstNonPHI());
1876 assert(!Inst->isTerminator() &&
1877 "The only TerminatorInst that can produce a value is "
1878 "InvokeInst which is handled above.");
1879 Store->insertAfter(Inst);
1882 assert(isa<Argument>(Def));
1883 Store->insertAfter(cast<Instruction>(Alloca));
1887 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1888 "we must have the same allocas with lives");
1889 if (!PromotableAllocas.empty()) {
1890 // Apply mem2reg to promote alloca to SSA
1891 PromoteMemToReg(PromotableAllocas, DT);
1895 for (auto &I : F.getEntryBlock())
1896 if (isa<AllocaInst>(I))
1898 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1902 /// Implement a unique function which doesn't require we sort the input
1903 /// vector. Doing so has the effect of changing the output of a couple of
1904 /// tests in ways which make them less useful in testing fused safepoints.
1905 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1906 SmallSet<T, 8> Seen;
1907 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1908 return !Seen.insert(V).second;
1912 /// Insert holders so that each Value is obviously live through the entire
1913 /// lifetime of the call.
1914 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1915 SmallVectorImpl<CallInst *> &Holders) {
1917 // No values to hold live, might as well not insert the empty holder
1920 Module *M = CS.getInstruction()->getModule();
1921 // Use a dummy vararg function to actually hold the values live
1922 Function *Func = cast<Function>(M->getOrInsertFunction(
1923 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1925 // For call safepoints insert dummy calls right after safepoint
1926 Holders.push_back(CallInst::Create(Func, Values, "",
1927 &*++CS.getInstruction()->getIterator()));
1930 // For invoke safepooints insert dummy calls both in normal and
1931 // exceptional destination blocks
1932 auto *II = cast<InvokeInst>(CS.getInstruction());
1933 Holders.push_back(CallInst::Create(
1934 Func, Values, "", &*II->getNormalDest()->getFirstInsertionPt()));
1935 Holders.push_back(CallInst::Create(
1936 Func, Values, "", &*II->getUnwindDest()->getFirstInsertionPt()));
1939 static void findLiveReferences(
1940 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1941 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1942 GCPtrLivenessData OriginalLivenessData;
1943 computeLiveInValues(DT, F, OriginalLivenessData);
1944 for (size_t i = 0; i < records.size(); i++) {
1945 struct PartiallyConstructedSafepointRecord &info = records[i];
1946 const CallSite &CS = toUpdate[i];
1947 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1951 /// Remove any vector of pointers from the live set by scalarizing them over the
1952 /// statepoint instruction. Adds the scalarized pieces to the live set. It
1953 /// would be preferable to include the vector in the statepoint itself, but
1954 /// the lowering code currently does not handle that. Extending it would be
1955 /// slightly non-trivial since it requires a format change. Given how rare
1956 /// such cases are (for the moment?) scalarizing is an acceptable compromise.
1957 static void splitVectorValues(Instruction *StatepointInst,
1958 StatepointLiveSetTy &LiveSet,
1959 DenseMap<Value *, Value *>& PointerToBase,
1960 DominatorTree &DT) {
1961 SmallVector<Value *, 16> ToSplit;
1962 for (Value *V : LiveSet)
1963 if (isa<VectorType>(V->getType()))
1964 ToSplit.push_back(V);
1966 if (ToSplit.empty())
1969 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1971 Function &F = *(StatepointInst->getParent()->getParent());
1973 DenseMap<Value *, AllocaInst *> AllocaMap;
1974 // First is normal return, second is exceptional return (invoke only)
1975 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1976 for (Value *V : ToSplit) {
1977 AllocaInst *Alloca =
1978 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1979 AllocaMap[V] = Alloca;
1981 VectorType *VT = cast<VectorType>(V->getType());
1982 IRBuilder<> Builder(StatepointInst);
1983 SmallVector<Value *, 16> Elements;
1984 for (unsigned i = 0; i < VT->getNumElements(); i++)
1985 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1986 ElementMapping[V] = Elements;
1988 auto InsertVectorReform = [&](Instruction *IP) {
1989 Builder.SetInsertPoint(IP);
1990 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1991 Value *ResultVec = UndefValue::get(VT);
1992 for (unsigned i = 0; i < VT->getNumElements(); i++)
1993 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1994 Builder.getInt32(i));
1998 if (isa<CallInst>(StatepointInst)) {
1999 BasicBlock::iterator Next(StatepointInst);
2001 Instruction *IP = &*(Next);
2002 Replacements[V].first = InsertVectorReform(IP);
2003 Replacements[V].second = nullptr;
2005 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
2006 // We've already normalized - check that we don't have shared destination
2008 BasicBlock *NormalDest = Invoke->getNormalDest();
2009 assert(!isa<PHINode>(NormalDest->begin()));
2010 BasicBlock *UnwindDest = Invoke->getUnwindDest();
2011 assert(!isa<PHINode>(UnwindDest->begin()));
2012 // Insert insert element sequences in both successors
2013 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
2014 Replacements[V].first = InsertVectorReform(IP);
2015 IP = &*(UnwindDest->getFirstInsertionPt());
2016 Replacements[V].second = InsertVectorReform(IP);
2020 for (Value *V : ToSplit) {
2021 AllocaInst *Alloca = AllocaMap[V];
2023 // Capture all users before we start mutating use lists
2024 SmallVector<Instruction *, 16> Users;
2025 for (User *U : V->users())
2026 Users.push_back(cast<Instruction>(U));
2028 for (Instruction *I : Users) {
2029 if (auto Phi = dyn_cast<PHINode>(I)) {
2030 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
2031 if (V == Phi->getIncomingValue(i)) {
2032 LoadInst *Load = new LoadInst(
2033 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
2034 Phi->setIncomingValue(i, Load);
2037 LoadInst *Load = new LoadInst(Alloca, "", I);
2038 I->replaceUsesOfWith(V, Load);
2042 // Store the original value and the replacement value into the alloca
2043 StoreInst *Store = new StoreInst(V, Alloca);
2044 if (auto I = dyn_cast<Instruction>(V))
2045 Store->insertAfter(I);
2047 Store->insertAfter(Alloca);
2049 // Normal return for invoke, or call return
2050 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
2051 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
2052 // Unwind return for invoke only
2053 Replacement = cast_or_null<Instruction>(Replacements[V].second);
2055 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
2058 // apply mem2reg to promote alloca to SSA
2059 SmallVector<AllocaInst *, 16> Allocas;
2060 for (Value *V : ToSplit)
2061 Allocas.push_back(AllocaMap[V]);
2062 PromoteMemToReg(Allocas, DT);
2064 // Update our tracking of live pointers and base mappings to account for the
2065 // changes we just made.
2066 for (Value *V : ToSplit) {
2067 auto &Elements = ElementMapping[V];
2070 LiveSet.insert(Elements.begin(), Elements.end());
2071 // We need to update the base mapping as well.
2072 assert(PointerToBase.count(V));
2073 Value *OldBase = PointerToBase[V];
2074 auto &BaseElements = ElementMapping[OldBase];
2075 PointerToBase.erase(V);
2076 assert(Elements.size() == BaseElements.size());
2077 for (unsigned i = 0; i < Elements.size(); i++) {
2078 Value *Elem = Elements[i];
2079 PointerToBase[Elem] = BaseElements[i];
2084 // Helper function for the "rematerializeLiveValues". It walks use chain
2085 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
2086 // values are visited (currently it is GEP's and casts). Returns true if it
2087 // successfully reached "BaseValue" and false otherwise.
2088 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
2090 static bool findRematerializableChainToBasePointer(
2091 SmallVectorImpl<Instruction*> &ChainToBase,
2092 Value *CurrentValue, Value *BaseValue) {
2094 // We have found a base value
2095 if (CurrentValue == BaseValue) {
2099 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
2100 ChainToBase.push_back(GEP);
2101 return findRematerializableChainToBasePointer(ChainToBase,
2102 GEP->getPointerOperand(),
2106 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
2107 Value *Def = CI->stripPointerCasts();
2109 // This two checks are basically similar. First one is here for the
2110 // consistency with findBasePointers logic.
2111 assert(!isa<CastInst>(Def) && "not a pointer cast found");
2112 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
2115 ChainToBase.push_back(CI);
2116 return findRematerializableChainToBasePointer(ChainToBase, Def, BaseValue);
2119 // Not supported instruction in the chain
2123 // Helper function for the "rematerializeLiveValues". Compute cost of the use
2124 // chain we are going to rematerialize.
2126 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
2127 TargetTransformInfo &TTI) {
2130 for (Instruction *Instr : Chain) {
2131 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
2132 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
2133 "non noop cast is found during rematerialization");
2135 Type *SrcTy = CI->getOperand(0)->getType();
2136 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
2138 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
2139 // Cost of the address calculation
2140 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
2141 Cost += TTI.getAddressComputationCost(ValTy);
2143 // And cost of the GEP itself
2144 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
2145 // allowed for the external usage)
2146 if (!GEP->hasAllConstantIndices())
2150 llvm_unreachable("unsupported instruciton type during rematerialization");
2157 // From the statepoint live set pick values that are cheaper to recompute then
2158 // to relocate. Remove this values from the live set, rematerialize them after
2159 // statepoint and record them in "Info" structure. Note that similar to
2160 // relocated values we don't do any user adjustments here.
2161 static void rematerializeLiveValues(CallSite CS,
2162 PartiallyConstructedSafepointRecord &Info,
2163 TargetTransformInfo &TTI) {
2164 const unsigned int ChainLengthThreshold = 10;
2166 // Record values we are going to delete from this statepoint live set.
2167 // We can not di this in following loop due to iterator invalidation.
2168 SmallVector<Value *, 32> LiveValuesToBeDeleted;
2170 for (Value *LiveValue: Info.LiveSet) {
2171 // For each live pointer find it's defining chain
2172 SmallVector<Instruction *, 3> ChainToBase;
2173 assert(Info.PointerToBase.count(LiveValue));
2175 findRematerializableChainToBasePointer(ChainToBase,
2177 Info.PointerToBase[LiveValue]);
2178 // Nothing to do, or chain is too long
2180 ChainToBase.size() == 0 ||
2181 ChainToBase.size() > ChainLengthThreshold)
2184 // Compute cost of this chain
2185 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
2186 // TODO: We can also account for cases when we will be able to remove some
2187 // of the rematerialized values by later optimization passes. I.e if
2188 // we rematerialized several intersecting chains. Or if original values
2189 // don't have any uses besides this statepoint.
2191 // For invokes we need to rematerialize each chain twice - for normal and
2192 // for unwind basic blocks. Model this by multiplying cost by two.
2193 if (CS.isInvoke()) {
2196 // If it's too expensive - skip it
2197 if (Cost >= RematerializationThreshold)
2200 // Remove value from the live set
2201 LiveValuesToBeDeleted.push_back(LiveValue);
2203 // Clone instructions and record them inside "Info" structure
2205 // Walk backwards to visit top-most instructions first
2206 std::reverse(ChainToBase.begin(), ChainToBase.end());
2208 // Utility function which clones all instructions from "ChainToBase"
2209 // and inserts them before "InsertBefore". Returns rematerialized value
2210 // which should be used after statepoint.
2211 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
2212 Instruction *LastClonedValue = nullptr;
2213 Instruction *LastValue = nullptr;
2214 for (Instruction *Instr: ChainToBase) {
2215 // Only GEP's and casts are suported as we need to be careful to not
2216 // introduce any new uses of pointers not in the liveset.
2217 // Note that it's fine to introduce new uses of pointers which were
2218 // otherwise not used after this statepoint.
2219 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2221 Instruction *ClonedValue = Instr->clone();
2222 ClonedValue->insertBefore(InsertBefore);
2223 ClonedValue->setName(Instr->getName() + ".remat");
2225 // If it is not first instruction in the chain then it uses previously
2226 // cloned value. We should update it to use cloned value.
2227 if (LastClonedValue) {
2229 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2231 // Assert that cloned instruction does not use any instructions from
2232 // this chain other than LastClonedValue
2233 for (auto OpValue : ClonedValue->operand_values()) {
2234 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2235 ChainToBase.end() &&
2236 "incorrect use in rematerialization chain");
2241 LastClonedValue = ClonedValue;
2244 assert(LastClonedValue);
2245 return LastClonedValue;
2248 // Different cases for calls and invokes. For invokes we need to clone
2249 // instructions both on normal and unwind path.
2251 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2252 assert(InsertBefore);
2253 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2254 Info.RematerializedValues[RematerializedValue] = LiveValue;
2256 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2258 Instruction *NormalInsertBefore =
2259 &*Invoke->getNormalDest()->getFirstInsertionPt();
2260 Instruction *UnwindInsertBefore =
2261 &*Invoke->getUnwindDest()->getFirstInsertionPt();
2263 Instruction *NormalRematerializedValue =
2264 rematerializeChain(NormalInsertBefore);
2265 Instruction *UnwindRematerializedValue =
2266 rematerializeChain(UnwindInsertBefore);
2268 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2269 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2273 // Remove rematerializaed values from the live set
2274 for (auto LiveValue: LiveValuesToBeDeleted) {
2275 Info.LiveSet.erase(LiveValue);
2279 static bool insertParsePoints(Function &F, DominatorTree &DT,
2280 TargetTransformInfo &TTI,
2281 SmallVectorImpl<CallSite> &ToUpdate) {
2283 // sanity check the input
2284 std::set<CallSite> Uniqued;
2285 Uniqued.insert(ToUpdate.begin(), ToUpdate.end());
2286 assert(Uniqued.size() == ToUpdate.size() && "no duplicates please!");
2288 for (CallSite CS : ToUpdate) {
2289 assert(CS.getInstruction()->getParent()->getParent() == &F);
2290 assert((UseDeoptBundles || isStatepoint(CS)) &&
2291 "expected to already be a deopt statepoint");
2295 // When inserting gc.relocates for invokes, we need to be able to insert at
2296 // the top of the successor blocks. See the comment on
2297 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2298 // may restructure the CFG.
2299 for (CallSite CS : ToUpdate) {
2302 auto *II = cast<InvokeInst>(CS.getInstruction());
2303 normalizeForInvokeSafepoint(II->getNormalDest(), II->getParent(), DT);
2304 normalizeForInvokeSafepoint(II->getUnwindDest(), II->getParent(), DT);
2307 // A list of dummy calls added to the IR to keep various values obviously
2308 // live in the IR. We'll remove all of these when done.
2309 SmallVector<CallInst *, 64> Holders;
2311 // Insert a dummy call with all of the arguments to the vm_state we'll need
2312 // for the actual safepoint insertion. This ensures reference arguments in
2313 // the deopt argument list are considered live through the safepoint (and
2314 // thus makes sure they get relocated.)
2315 for (CallSite CS : ToUpdate) {
2316 SmallVector<Value *, 64> DeoptValues;
2318 iterator_range<const Use *> DeoptStateRange =
2320 ? iterator_range<const Use *>(GetDeoptBundleOperands(CS))
2321 : iterator_range<const Use *>(Statepoint(CS).vm_state_args());
2323 for (Value *Arg : DeoptStateRange) {
2324 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2325 "support for FCA unimplemented");
2326 if (isHandledGCPointerType(Arg->getType()))
2327 DeoptValues.push_back(Arg);
2330 insertUseHolderAfter(CS, DeoptValues, Holders);
2333 SmallVector<PartiallyConstructedSafepointRecord, 64> Records(ToUpdate.size());
2335 // A) Identify all gc pointers which are statically live at the given call
2337 findLiveReferences(F, DT, ToUpdate, Records);
2339 // B) Find the base pointers for each live pointer
2340 /* scope for caching */ {
2341 // Cache the 'defining value' relation used in the computation and
2342 // insertion of base phis and selects. This ensures that we don't insert
2343 // large numbers of duplicate base_phis.
2344 DefiningValueMapTy DVCache;
2346 for (size_t i = 0; i < Records.size(); i++) {
2347 PartiallyConstructedSafepointRecord &info = Records[i];
2348 findBasePointers(DT, DVCache, ToUpdate[i], info);
2350 } // end of cache scope
2352 // The base phi insertion logic (for any safepoint) may have inserted new
2353 // instructions which are now live at some safepoint. The simplest such
2356 // phi a <-- will be a new base_phi here
2357 // safepoint 1 <-- that needs to be live here
2361 // We insert some dummy calls after each safepoint to definitely hold live
2362 // the base pointers which were identified for that safepoint. We'll then
2363 // ask liveness for _every_ base inserted to see what is now live. Then we
2364 // remove the dummy calls.
2365 Holders.reserve(Holders.size() + Records.size());
2366 for (size_t i = 0; i < Records.size(); i++) {
2367 PartiallyConstructedSafepointRecord &Info = Records[i];
2369 SmallVector<Value *, 128> Bases;
2370 for (auto Pair : Info.PointerToBase)
2371 Bases.push_back(Pair.second);
2373 insertUseHolderAfter(ToUpdate[i], Bases, Holders);
2376 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2377 // need to rerun liveness. We may *also* have inserted new defs, but that's
2378 // not the key issue.
2379 recomputeLiveInValues(F, DT, ToUpdate, Records);
2381 if (PrintBasePointers) {
2382 for (auto &Info : Records) {
2383 errs() << "Base Pairs: (w/Relocation)\n";
2384 for (auto Pair : Info.PointerToBase)
2385 errs() << " derived %" << Pair.first->getName() << " base %"
2386 << Pair.second->getName() << "\n";
2390 for (CallInst *CI : Holders)
2391 CI->eraseFromParent();
2395 // Do a limited scalarization of any live at safepoint vector values which
2396 // contain pointers. This enables this pass to run after vectorization at
2397 // the cost of some possible performance loss. TODO: it would be nice to
2398 // natively support vectors all the way through the backend so we don't need
2399 // to scalarize here.
2400 for (size_t i = 0; i < Records.size(); i++) {
2401 PartiallyConstructedSafepointRecord &Info = Records[i];
2402 Instruction *Statepoint = ToUpdate[i].getInstruction();
2403 splitVectorValues(cast<Instruction>(Statepoint), Info.LiveSet,
2404 Info.PointerToBase, DT);
2407 // In order to reduce live set of statepoint we might choose to rematerialize
2408 // some values instead of relocating them. This is purely an optimization and
2409 // does not influence correctness.
2410 for (size_t i = 0; i < Records.size(); i++)
2411 rematerializeLiveValues(ToUpdate[i], Records[i], TTI);
2413 // We need this to safely RAUW and delete call or invoke return values that
2414 // may themselves be live over a statepoint. For details, please see usage in
2415 // makeStatepointExplicitImpl.
2416 std::vector<DeferredReplacement> Replacements;
2418 // Now run through and replace the existing statepoints with new ones with
2419 // the live variables listed. We do not yet update uses of the values being
2420 // relocated. We have references to live variables that need to
2421 // survive to the last iteration of this loop. (By construction, the
2422 // previous statepoint can not be a live variable, thus we can and remove
2423 // the old statepoint calls as we go.)
2424 for (size_t i = 0; i < Records.size(); i++)
2425 makeStatepointExplicit(DT, ToUpdate[i], Records[i], Replacements);
2427 ToUpdate.clear(); // prevent accident use of invalid CallSites
2429 for (auto &PR : Replacements)
2432 Replacements.clear();
2434 for (auto &Info : Records) {
2435 // These live sets may contain state Value pointers, since we replaced calls
2436 // with operand bundles with calls wrapped in gc.statepoint, and some of
2437 // those calls may have been def'ing live gc pointers. Clear these out to
2438 // avoid accidentally using them.
2440 // TODO: We should create a separate data structure that does not contain
2441 // these live sets, and migrate to using that data structure from this point
2443 Info.LiveSet.clear();
2444 Info.PointerToBase.clear();
2447 // Do all the fixups of the original live variables to their relocated selves
2448 SmallVector<Value *, 128> Live;
2449 for (size_t i = 0; i < Records.size(); i++) {
2450 PartiallyConstructedSafepointRecord &Info = Records[i];
2452 // We can't simply save the live set from the original insertion. One of
2453 // the live values might be the result of a call which needs a safepoint.
2454 // That Value* no longer exists and we need to use the new gc_result.
2455 // Thankfully, the live set is embedded in the statepoint (and updated), so
2456 // we just grab that.
2457 Statepoint Statepoint(Info.StatepointToken);
2458 Live.insert(Live.end(), Statepoint.gc_args_begin(),
2459 Statepoint.gc_args_end());
2461 // Do some basic sanity checks on our liveness results before performing
2462 // relocation. Relocation can and will turn mistakes in liveness results
2463 // into non-sensical code which is must harder to debug.
2464 // TODO: It would be nice to test consistency as well
2465 assert(DT.isReachableFromEntry(Info.StatepointToken->getParent()) &&
2466 "statepoint must be reachable or liveness is meaningless");
2467 for (Value *V : Statepoint.gc_args()) {
2468 if (!isa<Instruction>(V))
2469 // Non-instruction values trivial dominate all possible uses
2471 auto *LiveInst = cast<Instruction>(V);
2472 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2473 "unreachable values should never be live");
2474 assert(DT.dominates(LiveInst, Info.StatepointToken) &&
2475 "basic SSA liveness expectation violated by liveness analysis");
2479 unique_unsorted(Live);
2483 for (auto *Ptr : Live)
2484 assert(isGCPointerType(Ptr->getType()) && "must be a gc pointer type");
2487 relocationViaAlloca(F, DT, Live, Records);
2488 return !Records.empty();
2491 // Handles both return values and arguments for Functions and CallSites.
2492 template <typename AttrHolder>
2493 static void RemoveNonValidAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2496 if (AH.getDereferenceableBytes(Index))
2497 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2498 AH.getDereferenceableBytes(Index)));
2499 if (AH.getDereferenceableOrNullBytes(Index))
2500 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2501 AH.getDereferenceableOrNullBytes(Index)));
2502 if (AH.doesNotAlias(Index))
2503 R.addAttribute(Attribute::NoAlias);
2506 AH.setAttributes(AH.getAttributes().removeAttributes(
2507 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2511 RewriteStatepointsForGC::stripNonValidAttributesFromPrototype(Function &F) {
2512 LLVMContext &Ctx = F.getContext();
2514 for (Argument &A : F.args())
2515 if (isa<PointerType>(A.getType()))
2516 RemoveNonValidAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2518 if (isa<PointerType>(F.getReturnType()))
2519 RemoveNonValidAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2522 void RewriteStatepointsForGC::stripNonValidAttributesFromBody(Function &F) {
2526 LLVMContext &Ctx = F.getContext();
2527 MDBuilder Builder(Ctx);
2529 for (Instruction &I : instructions(F)) {
2530 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2531 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2532 bool IsImmutableTBAA =
2533 MD->getNumOperands() == 4 &&
2534 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2536 if (!IsImmutableTBAA)
2537 continue; // no work to do, MD_tbaa is already marked mutable
2539 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2540 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2542 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2544 MDNode *MutableTBAA =
2545 Builder.createTBAAStructTagNode(Base, Access, Offset);
2546 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2549 if (CallSite CS = CallSite(&I)) {
2550 for (int i = 0, e = CS.arg_size(); i != e; i++)
2551 if (isa<PointerType>(CS.getArgument(i)->getType()))
2552 RemoveNonValidAttrAtIndex(Ctx, CS, i + 1);
2553 if (isa<PointerType>(CS.getType()))
2554 RemoveNonValidAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2559 /// Returns true if this function should be rewritten by this pass. The main
2560 /// point of this function is as an extension point for custom logic.
2561 static bool shouldRewriteStatepointsIn(Function &F) {
2562 // TODO: This should check the GCStrategy
2564 const char *FunctionGCName = F.getGC();
2565 const StringRef StatepointExampleName("statepoint-example");
2566 const StringRef CoreCLRName("coreclr");
2567 return (StatepointExampleName == FunctionGCName) ||
2568 (CoreCLRName == FunctionGCName);
2573 void RewriteStatepointsForGC::stripNonValidAttributes(Module &M) {
2575 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2579 for (Function &F : M)
2580 stripNonValidAttributesFromPrototype(F);
2582 for (Function &F : M)
2583 stripNonValidAttributesFromBody(F);
2586 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2587 // Nothing to do for declarations.
2588 if (F.isDeclaration() || F.empty())
2591 // Policy choice says not to rewrite - the most common reason is that we're
2592 // compiling code without a GCStrategy.
2593 if (!shouldRewriteStatepointsIn(F))
2596 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2597 TargetTransformInfo &TTI =
2598 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2600 auto NeedsRewrite = [](Instruction &I) {
2601 if (UseDeoptBundles) {
2602 if (ImmutableCallSite CS = ImmutableCallSite(&I))
2603 return !callsGCLeafFunction(CS);
2607 return isStatepoint(I);
2610 // Gather all the statepoints which need rewritten. Be careful to only
2611 // consider those in reachable code since we need to ask dominance queries
2612 // when rewriting. We'll delete the unreachable ones in a moment.
2613 SmallVector<CallSite, 64> ParsePointNeeded;
2614 bool HasUnreachableStatepoint = false;
2615 for (Instruction &I : instructions(F)) {
2616 // TODO: only the ones with the flag set!
2617 if (NeedsRewrite(I)) {
2618 if (DT.isReachableFromEntry(I.getParent()))
2619 ParsePointNeeded.push_back(CallSite(&I));
2621 HasUnreachableStatepoint = true;
2625 bool MadeChange = false;
2627 // Delete any unreachable statepoints so that we don't have unrewritten
2628 // statepoints surviving this pass. This makes testing easier and the
2629 // resulting IR less confusing to human readers. Rather than be fancy, we
2630 // just reuse a utility function which removes the unreachable blocks.
2631 if (HasUnreachableStatepoint)
2632 MadeChange |= removeUnreachableBlocks(F);
2634 // Return early if no work to do.
2635 if (ParsePointNeeded.empty())
2638 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2639 // These are created by LCSSA. They have the effect of increasing the size
2640 // of liveness sets for no good reason. It may be harder to do this post
2641 // insertion since relocations and base phis can confuse things.
2642 for (BasicBlock &BB : F)
2643 if (BB.getUniquePredecessor()) {
2645 FoldSingleEntryPHINodes(&BB);
2648 // Before we start introducing relocations, we want to tweak the IR a bit to
2649 // avoid unfortunate code generation effects. The main example is that we
2650 // want to try to make sure the comparison feeding a branch is after any
2651 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2652 // values feeding a branch after relocation. This is semantically correct,
2653 // but results in extra register pressure since both the pre-relocation and
2654 // post-relocation copies must be available in registers. For code without
2655 // relocations this is handled elsewhere, but teaching the scheduler to
2656 // reverse the transform we're about to do would be slightly complex.
2657 // Note: This may extend the live range of the inputs to the icmp and thus
2658 // increase the liveset of any statepoint we move over. This is profitable
2659 // as long as all statepoints are in rare blocks. If we had in-register
2660 // lowering for live values this would be a much safer transform.
2661 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2662 if (auto *BI = dyn_cast<BranchInst>(TI))
2663 if (BI->isConditional())
2664 return dyn_cast<Instruction>(BI->getCondition());
2665 // TODO: Extend this to handle switches
2668 for (BasicBlock &BB : F) {
2669 TerminatorInst *TI = BB.getTerminator();
2670 if (auto *Cond = getConditionInst(TI))
2671 // TODO: Handle more than just ICmps here. We should be able to move
2672 // most instructions without side effects or memory access.
2673 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2675 Cond->moveBefore(TI);
2679 MadeChange |= insertParsePoints(F, DT, TTI, ParsePointNeeded);
2683 // liveness computation via standard dataflow
2684 // -------------------------------------------------------------------
2686 // TODO: Consider using bitvectors for liveness, the set of potentially
2687 // interesting values should be small and easy to pre-compute.
2689 /// Compute the live-in set for the location rbegin starting from
2690 /// the live-out set of the basic block
2691 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2692 BasicBlock::reverse_iterator rend,
2693 DenseSet<Value *> &LiveTmp) {
2695 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2696 Instruction *I = &*ritr;
2698 // KILL/Def - Remove this definition from LiveIn
2701 // Don't consider *uses* in PHI nodes, we handle their contribution to
2702 // predecessor blocks when we seed the LiveOut sets
2703 if (isa<PHINode>(I))
2706 // USE - Add to the LiveIn set for this instruction
2707 for (Value *V : I->operands()) {
2708 assert(!isUnhandledGCPointerType(V->getType()) &&
2709 "support for FCA unimplemented");
2710 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2711 // The choice to exclude all things constant here is slightly subtle.
2712 // There are two independent reasons:
2713 // - We assume that things which are constant (from LLVM's definition)
2714 // do not move at runtime. For example, the address of a global
2715 // variable is fixed, even though it's contents may not be.
2716 // - Second, we can't disallow arbitrary inttoptr constants even
2717 // if the language frontend does. Optimization passes are free to
2718 // locally exploit facts without respect to global reachability. This
2719 // can create sections of code which are dynamically unreachable and
2720 // contain just about anything. (see constants.ll in tests)
2727 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2729 for (BasicBlock *Succ : successors(BB)) {
2730 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2731 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2732 PHINode *Phi = cast<PHINode>(&*I);
2733 Value *V = Phi->getIncomingValueForBlock(BB);
2734 assert(!isUnhandledGCPointerType(V->getType()) &&
2735 "support for FCA unimplemented");
2736 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2743 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2744 DenseSet<Value *> KillSet;
2745 for (Instruction &I : *BB)
2746 if (isHandledGCPointerType(I.getType()))
2752 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2753 /// sanity check for the liveness computation.
2754 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2755 TerminatorInst *TI, bool TermOkay = false) {
2756 for (Value *V : Live) {
2757 if (auto *I = dyn_cast<Instruction>(V)) {
2758 // The terminator can be a member of the LiveOut set. LLVM's definition
2759 // of instruction dominance states that V does not dominate itself. As
2760 // such, we need to special case this to allow it.
2761 if (TermOkay && TI == I)
2763 assert(DT.dominates(I, TI) &&
2764 "basic SSA liveness expectation violated by liveness analysis");
2769 /// Check that all the liveness sets used during the computation of liveness
2770 /// obey basic SSA properties. This is useful for finding cases where we miss
2772 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2774 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2775 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2776 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2780 static void computeLiveInValues(DominatorTree &DT, Function &F,
2781 GCPtrLivenessData &Data) {
2783 SmallSetVector<BasicBlock *, 200> Worklist;
2784 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2785 // We use a SetVector so that we don't have duplicates in the worklist.
2786 Worklist.insert(pred_begin(BB), pred_end(BB));
2788 auto NextItem = [&]() {
2789 BasicBlock *BB = Worklist.back();
2790 Worklist.pop_back();
2794 // Seed the liveness for each individual block
2795 for (BasicBlock &BB : F) {
2796 Data.KillSet[&BB] = computeKillSet(&BB);
2797 Data.LiveSet[&BB].clear();
2798 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2801 for (Value *Kill : Data.KillSet[&BB])
2802 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2805 Data.LiveOut[&BB] = DenseSet<Value *>();
2806 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2807 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2808 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2809 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2810 if (!Data.LiveIn[&BB].empty())
2811 AddPredsToWorklist(&BB);
2814 // Propagate that liveness until stable
2815 while (!Worklist.empty()) {
2816 BasicBlock *BB = NextItem();
2818 // Compute our new liveout set, then exit early if it hasn't changed
2819 // despite the contribution of our successor.
2820 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2821 const auto OldLiveOutSize = LiveOut.size();
2822 for (BasicBlock *Succ : successors(BB)) {
2823 assert(Data.LiveIn.count(Succ));
2824 set_union(LiveOut, Data.LiveIn[Succ]);
2826 // assert OutLiveOut is a subset of LiveOut
2827 if (OldLiveOutSize == LiveOut.size()) {
2828 // If the sets are the same size, then we didn't actually add anything
2829 // when unioning our successors LiveIn Thus, the LiveIn of this block
2833 Data.LiveOut[BB] = LiveOut;
2835 // Apply the effects of this basic block
2836 DenseSet<Value *> LiveTmp = LiveOut;
2837 set_union(LiveTmp, Data.LiveSet[BB]);
2838 set_subtract(LiveTmp, Data.KillSet[BB]);
2840 assert(Data.LiveIn.count(BB));
2841 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2842 // assert: OldLiveIn is a subset of LiveTmp
2843 if (OldLiveIn.size() != LiveTmp.size()) {
2844 Data.LiveIn[BB] = LiveTmp;
2845 AddPredsToWorklist(BB);
2847 } // while( !worklist.empty() )
2850 // Sanity check our output against SSA properties. This helps catch any
2851 // missing kills during the above iteration.
2852 for (BasicBlock &BB : F) {
2853 checkBasicSSA(DT, Data, BB);
2858 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2859 StatepointLiveSetTy &Out) {
2861 BasicBlock *BB = Inst->getParent();
2863 // Note: The copy is intentional and required
2864 assert(Data.LiveOut.count(BB));
2865 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2867 // We want to handle the statepoint itself oddly. It's
2868 // call result is not live (normal), nor are it's arguments
2869 // (unless they're used again later). This adjustment is
2870 // specifically what we need to relocate
2871 BasicBlock::reverse_iterator rend(Inst->getIterator());
2872 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2873 LiveOut.erase(Inst);
2874 Out.insert(LiveOut.begin(), LiveOut.end());
2877 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2879 PartiallyConstructedSafepointRecord &Info) {
2880 Instruction *Inst = CS.getInstruction();
2881 StatepointLiveSetTy Updated;
2882 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2885 DenseSet<Value *> Bases;
2886 for (auto KVPair : Info.PointerToBase) {
2887 Bases.insert(KVPair.second);
2890 // We may have base pointers which are now live that weren't before. We need
2891 // to update the PointerToBase structure to reflect this.
2892 for (auto V : Updated)
2893 if (!Info.PointerToBase.count(V)) {
2894 assert(Bases.count(V) && "can't find base for unexpected live value");
2895 Info.PointerToBase[V] = V;
2900 for (auto V : Updated) {
2901 assert(Info.PointerToBase.count(V) &&
2902 "must be able to find base for live value");
2906 // Remove any stale base mappings - this can happen since our liveness is
2907 // more precise then the one inherent in the base pointer analysis
2908 DenseSet<Value *> ToErase;
2909 for (auto KVPair : Info.PointerToBase)
2910 if (!Updated.count(KVPair.first))
2911 ToErase.insert(KVPair.first);
2912 for (auto V : ToErase)
2913 Info.PointerToBase.erase(V);
2916 for (auto KVPair : Info.PointerToBase)
2917 assert(Updated.count(KVPair.first) && "record for non-live value");
2920 Info.LiveSet = Updated;