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),
76 struct RewriteStatepointsForGC : public ModulePass {
77 static char ID; // Pass identification, replacement for typeid
79 RewriteStatepointsForGC() : ModulePass(ID) {
80 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
82 bool runOnFunction(Function &F);
83 bool runOnModule(Module &M) override {
86 Changed |= runOnFunction(F);
89 // stripDereferenceabilityInfo asserts that shouldRewriteStatepointsIn
90 // returns true for at least one function in the module. Since at least
91 // one function changed, we know that the precondition is satisfied.
92 stripDereferenceabilityInfo(M);
98 void getAnalysisUsage(AnalysisUsage &AU) const override {
99 // We add and rewrite a bunch of instructions, but don't really do much
100 // else. We could in theory preserve a lot more analyses here.
101 AU.addRequired<DominatorTreeWrapperPass>();
102 AU.addRequired<TargetTransformInfoWrapperPass>();
105 /// The IR fed into RewriteStatepointsForGC may have had attributes implying
106 /// dereferenceability that are no longer valid/correct after
107 /// RewriteStatepointsForGC has run. This is because semantically, after
108 /// RewriteStatepointsForGC runs, all calls to gc.statepoint "free" the entire
109 /// heap. stripDereferenceabilityInfo (conservatively) restores correctness
110 /// by erasing all attributes in the module that externally imply
111 /// dereferenceability.
113 void stripDereferenceabilityInfo(Module &M);
115 // Helpers for stripDereferenceabilityInfo
116 void stripDereferenceabilityInfoFromBody(Function &F);
117 void stripDereferenceabilityInfoFromPrototype(Function &F);
121 char RewriteStatepointsForGC::ID = 0;
123 ModulePass *llvm::createRewriteStatepointsForGCPass() {
124 return new RewriteStatepointsForGC();
127 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
128 "Make relocations explicit at statepoints", false, false)
129 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
130 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
131 "Make relocations explicit at statepoints", false, false)
134 struct GCPtrLivenessData {
135 /// Values defined in this block.
136 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet;
137 /// Values used in this block (and thus live); does not included values
138 /// killed within this block.
139 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet;
141 /// Values live into this basic block (i.e. used by any
142 /// instruction in this basic block or ones reachable from here)
143 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn;
145 /// Values live out of this basic block (i.e. live into
146 /// any successor block)
147 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut;
150 // The type of the internal cache used inside the findBasePointers family
151 // of functions. From the callers perspective, this is an opaque type and
152 // should not be inspected.
154 // In the actual implementation this caches two relations:
155 // - The base relation itself (i.e. this pointer is based on that one)
156 // - The base defining value relation (i.e. before base_phi insertion)
157 // Generally, after the execution of a full findBasePointer call, only the
158 // base relation will remain. Internally, we add a mixture of the two
159 // types, then update all the second type to the first type
160 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
161 typedef DenseSet<Value *> StatepointLiveSetTy;
162 typedef DenseMap<AssertingVH<Instruction>, AssertingVH<Value>>
163 RematerializedValueMapTy;
165 struct PartiallyConstructedSafepointRecord {
166 /// The set of values known to be live across this safepoint
167 StatepointLiveSetTy LiveSet;
169 /// Mapping from live pointers to a base-defining-value
170 DenseMap<Value *, Value *> PointerToBase;
172 /// The *new* gc.statepoint instruction itself. This produces the token
173 /// that normal path gc.relocates and the gc.result are tied to.
174 Instruction *StatepointToken;
176 /// Instruction to which exceptional gc relocates are attached
177 /// Makes it easier to iterate through them during relocationViaAlloca.
178 Instruction *UnwindToken;
180 /// Record live values we are rematerialized instead of relocating.
181 /// They are not included into 'LiveSet' field.
182 /// Maps rematerialized copy to it's original value.
183 RematerializedValueMapTy RematerializedValues;
187 /// Compute the live-in set for every basic block in the function
188 static void computeLiveInValues(DominatorTree &DT, Function &F,
189 GCPtrLivenessData &Data);
191 /// Given results from the dataflow liveness computation, find the set of live
192 /// Values at a particular instruction.
193 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
194 StatepointLiveSetTy &out);
196 // TODO: Once we can get to the GCStrategy, this becomes
197 // Optional<bool> isGCManagedPointer(const Value *V) const override {
199 static bool isGCPointerType(Type *T) {
200 if (auto *PT = dyn_cast<PointerType>(T))
201 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
202 // GC managed heap. We know that a pointer into this heap needs to be
203 // updated and that no other pointer does.
204 return (1 == PT->getAddressSpace());
208 // Return true if this type is one which a) is a gc pointer or contains a GC
209 // pointer and b) is of a type this code expects to encounter as a live value.
210 // (The insertion code will assert that a type which matches (a) and not (b)
211 // is not encountered.)
212 static bool isHandledGCPointerType(Type *T) {
213 // We fully support gc pointers
214 if (isGCPointerType(T))
216 // We partially support vectors of gc pointers. The code will assert if it
217 // can't handle something.
218 if (auto VT = dyn_cast<VectorType>(T))
219 if (isGCPointerType(VT->getElementType()))
225 /// Returns true if this type contains a gc pointer whether we know how to
226 /// handle that type or not.
227 static bool containsGCPtrType(Type *Ty) {
228 if (isGCPointerType(Ty))
230 if (VectorType *VT = dyn_cast<VectorType>(Ty))
231 return isGCPointerType(VT->getScalarType());
232 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
233 return containsGCPtrType(AT->getElementType());
234 if (StructType *ST = dyn_cast<StructType>(Ty))
236 ST->subtypes().begin(), ST->subtypes().end(),
237 [](Type *SubType) { return containsGCPtrType(SubType); });
241 // Returns true if this is a type which a) is a gc pointer or contains a GC
242 // pointer and b) is of a type which the code doesn't expect (i.e. first class
243 // aggregates). Used to trip assertions.
244 static bool isUnhandledGCPointerType(Type *Ty) {
245 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
249 static bool order_by_name(Value *a, Value *b) {
250 if (a->hasName() && b->hasName()) {
251 return -1 == a->getName().compare(b->getName());
252 } else if (a->hasName() && !b->hasName()) {
254 } else if (!a->hasName() && b->hasName()) {
257 // Better than nothing, but not stable
262 // Return the name of the value suffixed with the provided value, or if the
263 // value didn't have a name, the default value specified.
264 static std::string suffixed_name_or(Value *V, StringRef Suffix,
265 StringRef DefaultName) {
266 return V->hasName() ? (V->getName() + Suffix).str() : DefaultName.str();
269 // Conservatively identifies any definitions which might be live at the
270 // given instruction. The analysis is performed immediately before the
271 // given instruction. Values defined by that instruction are not considered
272 // live. Values used by that instruction are considered live.
273 static void analyzeParsePointLiveness(
274 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
275 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
276 Instruction *inst = CS.getInstruction();
278 StatepointLiveSetTy LiveSet;
279 findLiveSetAtInst(inst, OriginalLivenessData, LiveSet);
282 // Note: This output is used by several of the test cases
283 // The order of elements in a set is not stable, put them in a vec and sort
285 SmallVector<Value *, 64> Temp;
286 Temp.insert(Temp.end(), LiveSet.begin(), LiveSet.end());
287 std::sort(Temp.begin(), Temp.end(), order_by_name);
288 errs() << "Live Variables:\n";
289 for (Value *V : Temp)
290 dbgs() << " " << V->getName() << " " << *V << "\n";
292 if (PrintLiveSetSize) {
293 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
294 errs() << "Number live values: " << LiveSet.size() << "\n";
296 result.LiveSet = LiveSet;
299 static bool isKnownBaseResult(Value *V);
301 /// A single base defining value - An immediate base defining value for an
302 /// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'.
303 /// For instructions which have multiple pointer [vector] inputs or that
304 /// transition between vector and scalar types, there is no immediate base
305 /// defining value. The 'base defining value' for 'Def' is the transitive
306 /// closure of this relation stopping at the first instruction which has no
307 /// immediate base defining value. The b.d.v. might itself be a base pointer,
308 /// but it can also be an arbitrary derived pointer.
309 struct BaseDefiningValueResult {
310 /// Contains the value which is the base defining value.
312 /// True if the base defining value is also known to be an actual base
314 const bool IsKnownBase;
315 BaseDefiningValueResult(Value *BDV, bool IsKnownBase)
316 : BDV(BDV), IsKnownBase(IsKnownBase) {
318 // Check consistency between new and old means of checking whether a BDV is
320 bool MustBeBase = isKnownBaseResult(BDV);
321 assert(!MustBeBase || MustBeBase == IsKnownBase);
327 static BaseDefiningValueResult findBaseDefiningValue(Value *I);
329 /// Return a base defining value for the 'Index' element of the given vector
330 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
331 /// 'I'. As an optimization, this method will try to determine when the
332 /// element is known to already be a base pointer. If this can be established,
333 /// the second value in the returned pair will be true. Note that either a
334 /// vector or a pointer typed value can be returned. For the former, the
335 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
336 /// If the later, the return pointer is a BDV (or possibly a base) for the
337 /// particular element in 'I'.
338 static BaseDefiningValueResult
339 findBaseDefiningValueOfVector(Value *I) {
340 assert(I->getType()->isVectorTy() &&
341 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
342 "Illegal to ask for the base pointer of a non-pointer type");
344 // Each case parallels findBaseDefiningValue below, see that code for
345 // detailed motivation.
347 if (isa<Argument>(I))
348 // An incoming argument to the function is a base pointer
349 return BaseDefiningValueResult(I, true);
351 // We shouldn't see the address of a global as a vector value?
352 assert(!isa<GlobalVariable>(I) &&
353 "unexpected global variable found in base of vector");
355 // inlining could possibly introduce phi node that contains
356 // undef if callee has multiple returns
357 if (isa<UndefValue>(I))
358 // utterly meaningless, but useful for dealing with partially optimized
360 return BaseDefiningValueResult(I, true);
362 // Due to inheritance, this must be _after_ the global variable and undef
364 if (Constant *Con = dyn_cast<Constant>(I)) {
365 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
366 "order of checks wrong!");
367 assert(Con->isNullValue() && "null is the only case which makes sense");
368 return BaseDefiningValueResult(Con, true);
371 if (isa<LoadInst>(I))
372 return BaseDefiningValueResult(I, true);
374 if (isa<InsertElementInst>(I))
375 // We don't know whether this vector contains entirely base pointers or
376 // not. To be conservatively correct, we treat it as a BDV and will
377 // duplicate code as needed to construct a parallel vector of bases.
378 return BaseDefiningValueResult(I, false);
380 if (isa<ShuffleVectorInst>(I))
381 // We don't know whether this vector contains entirely base pointers or
382 // not. To be conservatively correct, we treat it as a BDV and will
383 // duplicate code as needed to construct a parallel vector of bases.
384 // TODO: There a number of local optimizations which could be applied here
385 // for particular sufflevector patterns.
386 return BaseDefiningValueResult(I, false);
388 // A PHI or Select is a base defining value. The outer findBasePointer
389 // algorithm is responsible for constructing a base value for this BDV.
390 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
391 "unknown vector instruction - no base found for vector element");
392 return BaseDefiningValueResult(I, false);
395 /// Helper function for findBasePointer - Will return a value which either a)
396 /// defines the base pointer for the input, b) blocks the simple search
397 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
398 /// from pointer to vector type or back.
399 static BaseDefiningValueResult findBaseDefiningValue(Value *I) {
400 if (I->getType()->isVectorTy())
401 return findBaseDefiningValueOfVector(I);
403 assert(I->getType()->isPointerTy() &&
404 "Illegal to ask for the base pointer of a non-pointer type");
406 if (isa<Argument>(I))
407 // An incoming argument to the function is a base pointer
408 // We should have never reached here if this argument isn't an gc value
409 return BaseDefiningValueResult(I, true);
411 if (isa<GlobalVariable>(I))
413 return BaseDefiningValueResult(I, true);
415 // inlining could possibly introduce phi node that contains
416 // undef if callee has multiple returns
417 if (isa<UndefValue>(I))
418 // utterly meaningless, but useful for dealing with
419 // partially optimized code.
420 return BaseDefiningValueResult(I, true);
422 // Due to inheritance, this must be _after_ the global variable and undef
424 if (isa<Constant>(I)) {
425 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
426 "order of checks wrong!");
427 // Note: Finding a constant base for something marked for relocation
428 // doesn't really make sense. The most likely case is either a) some
429 // screwed up the address space usage or b) your validating against
430 // compiled C++ code w/o the proper separation. The only real exception
431 // is a null pointer. You could have generic code written to index of
432 // off a potentially null value and have proven it null. We also use
433 // null pointers in dead paths of relocation phis (which we might later
434 // want to find a base pointer for).
435 assert(isa<ConstantPointerNull>(I) &&
436 "null is the only case which makes sense");
437 return BaseDefiningValueResult(I, true);
440 if (CastInst *CI = dyn_cast<CastInst>(I)) {
441 Value *Def = CI->stripPointerCasts();
442 // If we find a cast instruction here, it means we've found a cast which is
443 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
444 // handle int->ptr conversion.
445 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
446 return findBaseDefiningValue(Def);
449 if (isa<LoadInst>(I))
450 // The value loaded is an gc base itself
451 return BaseDefiningValueResult(I, true);
454 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
455 // The base of this GEP is the base
456 return findBaseDefiningValue(GEP->getPointerOperand());
458 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
459 switch (II->getIntrinsicID()) {
460 case Intrinsic::experimental_gc_result_ptr:
462 // fall through to general call handling
464 case Intrinsic::experimental_gc_statepoint:
465 case Intrinsic::experimental_gc_result_float:
466 case Intrinsic::experimental_gc_result_int:
467 llvm_unreachable("these don't produce pointers");
468 case Intrinsic::experimental_gc_relocate: {
469 // Rerunning safepoint insertion after safepoints are already
470 // inserted is not supported. It could probably be made to work,
471 // but why are you doing this? There's no good reason.
472 llvm_unreachable("repeat safepoint insertion is not supported");
474 case Intrinsic::gcroot:
475 // Currently, this mechanism hasn't been extended to work with gcroot.
476 // There's no reason it couldn't be, but I haven't thought about the
477 // implications much.
479 "interaction with the gcroot mechanism is not supported");
482 // We assume that functions in the source language only return base
483 // pointers. This should probably be generalized via attributes to support
484 // both source language and internal functions.
485 if (isa<CallInst>(I) || isa<InvokeInst>(I))
486 return BaseDefiningValueResult(I, true);
488 // I have absolutely no idea how to implement this part yet. It's not
489 // necessarily hard, I just haven't really looked at it yet.
490 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
492 if (isa<AtomicCmpXchgInst>(I))
493 // A CAS is effectively a atomic store and load combined under a
494 // predicate. From the perspective of base pointers, we just treat it
496 return BaseDefiningValueResult(I, true);
498 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
499 "binary ops which don't apply to pointers");
501 // The aggregate ops. Aggregates can either be in the heap or on the
502 // stack, but in either case, this is simply a field load. As a result,
503 // this is a defining definition of the base just like a load is.
504 if (isa<ExtractValueInst>(I))
505 return BaseDefiningValueResult(I, true);
507 // We should never see an insert vector since that would require we be
508 // tracing back a struct value not a pointer value.
509 assert(!isa<InsertValueInst>(I) &&
510 "Base pointer for a struct is meaningless");
512 // An extractelement produces a base result exactly when it's input does.
513 // We may need to insert a parallel instruction to extract the appropriate
514 // element out of the base vector corresponding to the input. Given this,
515 // it's analogous to the phi and select case even though it's not a merge.
516 if (isa<ExtractElementInst>(I))
517 // Note: There a lot of obvious peephole cases here. This are deliberately
518 // handled after the main base pointer inference algorithm to make writing
519 // test cases to exercise that code easier.
520 return BaseDefiningValueResult(I, false);
522 // The last two cases here don't return a base pointer. Instead, they
523 // return a value which dynamically selects from among several base
524 // derived pointers (each with it's own base potentially). It's the job of
525 // the caller to resolve these.
526 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
527 "missing instruction case in findBaseDefiningValing");
528 return BaseDefiningValueResult(I, false);
531 /// Returns the base defining value for this value.
532 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
533 Value *&Cached = Cache[I];
535 Cached = findBaseDefiningValue(I).BDV;
536 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
537 << Cached->getName() << "\n");
539 assert(Cache[I] != nullptr);
543 /// Return a base pointer for this value if known. Otherwise, return it's
544 /// base defining value.
545 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
546 Value *Def = findBaseDefiningValueCached(I, Cache);
547 auto Found = Cache.find(Def);
548 if (Found != Cache.end()) {
549 // Either a base-of relation, or a self reference. Caller must check.
550 return Found->second;
552 // Only a BDV available
556 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
557 /// is it known to be a base pointer? Or do we need to continue searching.
558 static bool isKnownBaseResult(Value *V) {
559 if (!isa<PHINode>(V) && !isa<SelectInst>(V) &&
560 !isa<ExtractElementInst>(V) && !isa<InsertElementInst>(V) &&
561 !isa<ShuffleVectorInst>(V)) {
562 // no recursion possible
565 if (isa<Instruction>(V) &&
566 cast<Instruction>(V)->getMetadata("is_base_value")) {
567 // This is a previously inserted base phi or select. We know
568 // that this is a base value.
572 // We need to keep searching
577 /// Models the state of a single base defining value in the findBasePointer
578 /// algorithm for determining where a new instruction is needed to propagate
579 /// the base of this BDV.
582 enum Status { Unknown, Base, Conflict };
584 BDVState(Status s, Value *b = nullptr) : status(s), base(b) {
585 assert(status != Base || b);
587 explicit BDVState(Value *b) : status(Base), base(b) {}
588 BDVState() : status(Unknown), base(nullptr) {}
590 Status getStatus() const { return status; }
591 Value *getBase() const { return base; }
593 bool isBase() const { return getStatus() == Base; }
594 bool isUnknown() const { return getStatus() == Unknown; }
595 bool isConflict() const { return getStatus() == Conflict; }
597 bool operator==(const BDVState &other) const {
598 return base == other.base && status == other.status;
601 bool operator!=(const BDVState &other) const { return !(*this == other); }
604 void dump() const { print(dbgs()); dbgs() << '\n'; }
606 void print(raw_ostream &OS) const {
618 OS << " (" << base << " - "
619 << (base ? base->getName() : "nullptr") << "): ";
624 Value *base; // non null only if status == base
629 static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
636 // Values of type BDVState form a lattice, and this is a helper
637 // class that implementes the meet operation. The meat of the meet
638 // operation is implemented in MeetBDVStates::pureMeet
639 class MeetBDVStates {
641 /// Initializes the currentResult to the TOP state so that if can be met with
642 /// any other state to produce that state.
645 // Destructively meet the current result with the given BDVState
646 void meetWith(BDVState otherState) {
647 currentResult = meet(otherState, currentResult);
650 BDVState getResult() const { return currentResult; }
653 BDVState currentResult;
655 /// Perform a meet operation on two elements of the BDVState lattice.
656 static BDVState meet(BDVState LHS, BDVState RHS) {
657 assert((pureMeet(LHS, RHS) == pureMeet(RHS, LHS)) &&
658 "math is wrong: meet does not commute!");
659 BDVState Result = pureMeet(LHS, RHS);
660 DEBUG(dbgs() << "meet of " << LHS << " with " << RHS
661 << " produced " << Result << "\n");
665 static BDVState pureMeet(const BDVState &stateA, const BDVState &stateB) {
666 switch (stateA.getStatus()) {
667 case BDVState::Unknown:
671 assert(stateA.getBase() && "can't be null");
672 if (stateB.isUnknown())
675 if (stateB.isBase()) {
676 if (stateA.getBase() == stateB.getBase()) {
677 assert(stateA == stateB && "equality broken!");
680 return BDVState(BDVState::Conflict);
682 assert(stateB.isConflict() && "only three states!");
683 return BDVState(BDVState::Conflict);
685 case BDVState::Conflict:
688 llvm_unreachable("only three states!");
694 /// For a given value or instruction, figure out what base ptr it's derived
695 /// from. For gc objects, this is simply itself. On success, returns a value
696 /// which is the base pointer. (This is reliable and can be used for
697 /// relocation.) On failure, returns nullptr.
698 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
699 Value *def = findBaseOrBDV(I, cache);
701 if (isKnownBaseResult(def)) {
705 // Here's the rough algorithm:
706 // - For every SSA value, construct a mapping to either an actual base
707 // pointer or a PHI which obscures the base pointer.
708 // - Construct a mapping from PHI to unknown TOP state. Use an
709 // optimistic algorithm to propagate base pointer information. Lattice
714 // When algorithm terminates, all PHIs will either have a single concrete
715 // base or be in a conflict state.
716 // - For every conflict, insert a dummy PHI node without arguments. Add
717 // these to the base[Instruction] = BasePtr mapping. For every
718 // non-conflict, add the actual base.
719 // - For every conflict, add arguments for the base[a] of each input
722 // Note: A simpler form of this would be to add the conflict form of all
723 // PHIs without running the optimistic algorithm. This would be
724 // analogous to pessimistic data flow and would likely lead to an
725 // overall worse solution.
728 auto isExpectedBDVType = [](Value *BDV) {
729 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) ||
730 isa<ExtractElementInst>(BDV) || isa<InsertElementInst>(BDV);
734 // Once populated, will contain a mapping from each potentially non-base BDV
735 // to a lattice value (described above) which corresponds to that BDV.
736 // We use the order of insertion (DFS over the def/use graph) to provide a
737 // stable deterministic ordering for visiting DenseMaps (which are unordered)
738 // below. This is important for deterministic compilation.
739 MapVector<Value *, BDVState> States;
741 // Recursively fill in all base defining values reachable from the initial
742 // one for which we don't already know a definite base value for
744 SmallVector<Value*, 16> Worklist;
745 Worklist.push_back(def);
746 States.insert(std::make_pair(def, BDVState()));
747 while (!Worklist.empty()) {
748 Value *Current = Worklist.pop_back_val();
749 assert(!isKnownBaseResult(Current) && "why did it get added?");
751 auto visitIncomingValue = [&](Value *InVal) {
752 Value *Base = findBaseOrBDV(InVal, cache);
753 if (isKnownBaseResult(Base))
754 // Known bases won't need new instructions introduced and can be
757 assert(isExpectedBDVType(Base) && "the only non-base values "
758 "we see should be base defining values");
759 if (States.insert(std::make_pair(Base, BDVState())).second)
760 Worklist.push_back(Base);
762 if (PHINode *Phi = dyn_cast<PHINode>(Current)) {
763 for (Value *InVal : Phi->incoming_values())
764 visitIncomingValue(InVal);
765 } else if (SelectInst *Sel = dyn_cast<SelectInst>(Current)) {
766 visitIncomingValue(Sel->getTrueValue());
767 visitIncomingValue(Sel->getFalseValue());
768 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
769 visitIncomingValue(EE->getVectorOperand());
770 } else if (auto *IE = dyn_cast<InsertElementInst>(Current)) {
771 visitIncomingValue(IE->getOperand(0)); // vector operand
772 visitIncomingValue(IE->getOperand(1)); // scalar operand
774 // There is one known class of instructions we know we don't handle.
775 assert(isa<ShuffleVectorInst>(Current));
776 llvm_unreachable("unimplemented instruction case");
782 DEBUG(dbgs() << "States after initialization:\n");
783 for (auto Pair : States) {
784 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
788 // Return a phi state for a base defining value. We'll generate a new
789 // base state for known bases and expect to find a cached state otherwise.
790 auto getStateForBDV = [&](Value *baseValue) {
791 if (isKnownBaseResult(baseValue))
792 return BDVState(baseValue);
793 auto I = States.find(baseValue);
794 assert(I != States.end() && "lookup failed!");
798 bool progress = true;
801 const size_t oldSize = States.size();
804 // We're only changing values in this loop, thus safe to keep iterators.
805 // Since this is computing a fixed point, the order of visit does not
806 // effect the result. TODO: We could use a worklist here and make this run
808 for (auto Pair : States) {
809 Value *BDV = Pair.first;
810 assert(!isKnownBaseResult(BDV) && "why did it get added?");
812 // Given an input value for the current instruction, return a BDVState
813 // instance which represents the BDV of that value.
814 auto getStateForInput = [&](Value *V) mutable {
815 Value *BDV = findBaseOrBDV(V, cache);
816 return getStateForBDV(BDV);
819 MeetBDVStates calculateMeet;
820 if (SelectInst *select = dyn_cast<SelectInst>(BDV)) {
821 calculateMeet.meetWith(getStateForInput(select->getTrueValue()));
822 calculateMeet.meetWith(getStateForInput(select->getFalseValue()));
823 } else if (PHINode *Phi = dyn_cast<PHINode>(BDV)) {
824 for (Value *Val : Phi->incoming_values())
825 calculateMeet.meetWith(getStateForInput(Val));
826 } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) {
827 // The 'meet' for an extractelement is slightly trivial, but it's still
828 // useful in that it drives us to conflict if our input is.
829 calculateMeet.meetWith(getStateForInput(EE->getVectorOperand()));
831 // Given there's a inherent type mismatch between the operands, will
832 // *always* produce Conflict.
833 auto *IE = cast<InsertElementInst>(BDV);
834 calculateMeet.meetWith(getStateForInput(IE->getOperand(0)));
835 calculateMeet.meetWith(getStateForInput(IE->getOperand(1)));
838 BDVState oldState = States[BDV];
839 BDVState newState = calculateMeet.getResult();
840 if (oldState != newState) {
842 States[BDV] = newState;
846 assert(oldSize == States.size() &&
847 "fixed point shouldn't be adding any new nodes to state");
851 DEBUG(dbgs() << "States after meet iteration:\n");
852 for (auto Pair : States) {
853 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
857 // Insert Phis for all conflicts
858 // TODO: adjust naming patterns to avoid this order of iteration dependency
859 for (auto Pair : States) {
860 Instruction *I = cast<Instruction>(Pair.first);
861 BDVState State = Pair.second;
862 assert(!isKnownBaseResult(I) && "why did it get added?");
863 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
865 // extractelement instructions are a bit special in that we may need to
866 // insert an extract even when we know an exact base for the instruction.
867 // The problem is that we need to convert from a vector base to a scalar
868 // base for the particular indice we're interested in.
869 if (State.isBase() && isa<ExtractElementInst>(I) &&
870 isa<VectorType>(State.getBase()->getType())) {
871 auto *EE = cast<ExtractElementInst>(I);
872 // TODO: In many cases, the new instruction is just EE itself. We should
873 // exploit this, but can't do it here since it would break the invariant
874 // about the BDV not being known to be a base.
875 auto *BaseInst = ExtractElementInst::Create(State.getBase(),
876 EE->getIndexOperand(),
878 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
879 States[I] = BDVState(BDVState::Base, BaseInst);
882 // Since we're joining a vector and scalar base, they can never be the
883 // same. As a result, we should always see insert element having reached
884 // the conflict state.
885 if (isa<InsertElementInst>(I)) {
886 assert(State.isConflict());
889 if (!State.isConflict())
892 /// Create and insert a new instruction which will represent the base of
893 /// the given instruction 'I'.
894 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
895 if (isa<PHINode>(I)) {
896 BasicBlock *BB = I->getParent();
897 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
898 assert(NumPreds > 0 && "how did we reach here");
899 std::string Name = suffixed_name_or(I, ".base", "base_phi");
900 return PHINode::Create(I->getType(), NumPreds, Name, I);
901 } else if (SelectInst *Sel = dyn_cast<SelectInst>(I)) {
902 // The undef will be replaced later
903 UndefValue *Undef = UndefValue::get(Sel->getType());
904 std::string Name = suffixed_name_or(I, ".base", "base_select");
905 return SelectInst::Create(Sel->getCondition(), Undef,
907 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
908 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
909 std::string Name = suffixed_name_or(I, ".base", "base_ee");
910 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
913 auto *IE = cast<InsertElementInst>(I);
914 UndefValue *VecUndef = UndefValue::get(IE->getOperand(0)->getType());
915 UndefValue *ScalarUndef = UndefValue::get(IE->getOperand(1)->getType());
916 std::string Name = suffixed_name_or(I, ".base", "base_ie");
917 return InsertElementInst::Create(VecUndef, ScalarUndef,
918 IE->getOperand(2), Name, IE);
922 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
923 // Add metadata marking this as a base value
924 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
925 States[I] = BDVState(BDVState::Conflict, BaseInst);
928 // Returns a instruction which produces the base pointer for a given
929 // instruction. The instruction is assumed to be an input to one of the BDVs
930 // seen in the inference algorithm above. As such, we must either already
931 // know it's base defining value is a base, or have inserted a new
932 // instruction to propagate the base of it's BDV and have entered that newly
933 // introduced instruction into the state table. In either case, we are
934 // assured to be able to determine an instruction which produces it's base
936 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
937 Value *BDV = findBaseOrBDV(Input, cache);
938 Value *Base = nullptr;
939 if (isKnownBaseResult(BDV)) {
942 // Either conflict or base.
943 assert(States.count(BDV));
944 Base = States[BDV].getBase();
946 assert(Base && "can't be null");
947 // The cast is needed since base traversal may strip away bitcasts
948 if (Base->getType() != Input->getType() &&
950 Base = new BitCastInst(Base, Input->getType(), "cast",
956 // Fixup all the inputs of the new PHIs. Visit order needs to be
957 // deterministic and predictable because we're naming newly created
959 for (auto Pair : States) {
960 Instruction *BDV = cast<Instruction>(Pair.first);
961 BDVState State = Pair.second;
963 assert(!isKnownBaseResult(BDV) && "why did it get added?");
964 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
965 if (!State.isConflict())
968 if (PHINode *basephi = dyn_cast<PHINode>(State.getBase())) {
969 PHINode *phi = cast<PHINode>(BDV);
970 unsigned NumPHIValues = phi->getNumIncomingValues();
971 for (unsigned i = 0; i < NumPHIValues; i++) {
972 Value *InVal = phi->getIncomingValue(i);
973 BasicBlock *InBB = phi->getIncomingBlock(i);
975 // If we've already seen InBB, add the same incoming value
976 // we added for it earlier. The IR verifier requires phi
977 // nodes with multiple entries from the same basic block
978 // to have the same incoming value for each of those
979 // entries. If we don't do this check here and basephi
980 // has a different type than base, we'll end up adding two
981 // bitcasts (and hence two distinct values) as incoming
982 // values for the same basic block.
984 int blockIndex = basephi->getBasicBlockIndex(InBB);
985 if (blockIndex != -1) {
986 Value *oldBase = basephi->getIncomingValue(blockIndex);
987 basephi->addIncoming(oldBase, InBB);
990 Value *Base = getBaseForInput(InVal, nullptr);
991 // In essence this assert states: the only way two
992 // values incoming from the same basic block may be
993 // different is by being different bitcasts of the same
994 // value. A cleanup that remains TODO is changing
995 // findBaseOrBDV to return an llvm::Value of the correct
996 // type (and still remain pure). This will remove the
997 // need to add bitcasts.
998 assert(Base->stripPointerCasts() == oldBase->stripPointerCasts() &&
999 "sanity -- findBaseOrBDV should be pure!");
1004 // Find the instruction which produces the base for each input. We may
1005 // need to insert a bitcast in the incoming block.
1006 // TODO: Need to split critical edges if insertion is needed
1007 Value *Base = getBaseForInput(InVal, InBB->getTerminator());
1008 basephi->addIncoming(Base, InBB);
1010 assert(basephi->getNumIncomingValues() == NumPHIValues);
1011 } else if (SelectInst *BaseSel = dyn_cast<SelectInst>(State.getBase())) {
1012 SelectInst *Sel = cast<SelectInst>(BDV);
1013 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
1014 // something more safe and less hacky.
1015 for (int i = 1; i <= 2; i++) {
1016 Value *InVal = Sel->getOperand(i);
1017 // Find the instruction which produces the base for each input. We may
1018 // need to insert a bitcast.
1019 Value *Base = getBaseForInput(InVal, BaseSel);
1020 BaseSel->setOperand(i, Base);
1022 } else if (auto *BaseEE = dyn_cast<ExtractElementInst>(State.getBase())) {
1023 Value *InVal = cast<ExtractElementInst>(BDV)->getVectorOperand();
1024 // Find the instruction which produces the base for each input. We may
1025 // need to insert a bitcast.
1026 Value *Base = getBaseForInput(InVal, BaseEE);
1027 BaseEE->setOperand(0, Base);
1029 auto *BaseIE = cast<InsertElementInst>(State.getBase());
1030 auto *BdvIE = cast<InsertElementInst>(BDV);
1031 auto UpdateOperand = [&](int OperandIdx) {
1032 Value *InVal = BdvIE->getOperand(OperandIdx);
1033 Value *Base = getBaseForInput(InVal, BaseIE);
1034 BaseIE->setOperand(OperandIdx, Base);
1036 UpdateOperand(0); // vector operand
1037 UpdateOperand(1); // scalar operand
1042 // Now that we're done with the algorithm, see if we can optimize the
1043 // results slightly by reducing the number of new instructions needed.
1044 // Arguably, this should be integrated into the algorithm above, but
1045 // doing as a post process step is easier to reason about for the moment.
1046 DenseMap<Value *, Value *> ReverseMap;
1047 SmallPtrSet<Instruction *, 16> NewInsts;
1048 SmallSetVector<AssertingVH<Instruction>, 16> Worklist;
1049 // Note: We need to visit the states in a deterministic order. We uses the
1050 // Keys we sorted above for this purpose. Note that we are papering over a
1051 // bigger problem with the algorithm above - it's visit order is not
1052 // deterministic. A larger change is needed to fix this.
1053 for (auto Pair : States) {
1054 auto *BDV = Pair.first;
1055 auto State = Pair.second;
1056 Value *Base = State.getBase();
1057 assert(BDV && Base);
1058 assert(!isKnownBaseResult(BDV) && "why did it get added?");
1059 assert(isKnownBaseResult(Base) &&
1060 "must be something we 'know' is a base pointer");
1061 if (!State.isConflict())
1064 ReverseMap[Base] = BDV;
1065 if (auto *BaseI = dyn_cast<Instruction>(Base)) {
1066 NewInsts.insert(BaseI);
1067 Worklist.insert(BaseI);
1070 auto ReplaceBaseInstWith = [&](Value *BDV, Instruction *BaseI,
1071 Value *Replacement) {
1072 // Add users which are new instructions (excluding self references)
1073 for (User *U : BaseI->users())
1074 if (auto *UI = dyn_cast<Instruction>(U))
1075 if (NewInsts.count(UI) && UI != BaseI)
1076 Worklist.insert(UI);
1077 // Then do the actual replacement
1078 NewInsts.erase(BaseI);
1079 ReverseMap.erase(BaseI);
1080 BaseI->replaceAllUsesWith(Replacement);
1081 BaseI->eraseFromParent();
1082 assert(States.count(BDV));
1083 assert(States[BDV].isConflict() && States[BDV].getBase() == BaseI);
1084 States[BDV] = BDVState(BDVState::Conflict, Replacement);
1086 const DataLayout &DL = cast<Instruction>(def)->getModule()->getDataLayout();
1087 while (!Worklist.empty()) {
1088 Instruction *BaseI = Worklist.pop_back_val();
1089 assert(NewInsts.count(BaseI));
1090 Value *Bdv = ReverseMap[BaseI];
1091 if (auto *BdvI = dyn_cast<Instruction>(Bdv))
1092 if (BaseI->isIdenticalTo(BdvI)) {
1093 DEBUG(dbgs() << "Identical Base: " << *BaseI << "\n");
1094 ReplaceBaseInstWith(Bdv, BaseI, Bdv);
1097 if (Value *V = SimplifyInstruction(BaseI, DL)) {
1098 DEBUG(dbgs() << "Base " << *BaseI << " simplified to " << *V << "\n");
1099 ReplaceBaseInstWith(Bdv, BaseI, V);
1104 // Cache all of our results so we can cheaply reuse them
1105 // NOTE: This is actually two caches: one of the base defining value
1106 // relation and one of the base pointer relation! FIXME
1107 for (auto Pair : States) {
1108 auto *BDV = Pair.first;
1109 Value *base = Pair.second.getBase();
1110 assert(BDV && base);
1112 std::string fromstr = cache.count(BDV) ? cache[BDV]->getName() : "none";
1113 DEBUG(dbgs() << "Updating base value cache"
1114 << " for: " << BDV->getName()
1115 << " from: " << fromstr
1116 << " to: " << base->getName() << "\n");
1118 if (cache.count(BDV)) {
1119 // Once we transition from the BDV relation being store in the cache to
1120 // the base relation being stored, it must be stable
1121 assert((!isKnownBaseResult(cache[BDV]) || cache[BDV] == base) &&
1122 "base relation should be stable");
1126 assert(cache.find(def) != cache.end());
1130 // For a set of live pointers (base and/or derived), identify the base
1131 // pointer of the object which they are derived from. This routine will
1132 // mutate the IR graph as needed to make the 'base' pointer live at the
1133 // definition site of 'derived'. This ensures that any use of 'derived' can
1134 // also use 'base'. This may involve the insertion of a number of
1135 // additional PHI nodes.
1137 // preconditions: live is a set of pointer type Values
1139 // side effects: may insert PHI nodes into the existing CFG, will preserve
1140 // CFG, will not remove or mutate any existing nodes
1142 // post condition: PointerToBase contains one (derived, base) pair for every
1143 // pointer in live. Note that derived can be equal to base if the original
1144 // pointer was a base pointer.
1146 findBasePointers(const StatepointLiveSetTy &live,
1147 DenseMap<Value *, Value *> &PointerToBase,
1148 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1149 // For the naming of values inserted to be deterministic - which makes for
1150 // much cleaner and more stable tests - we need to assign an order to the
1151 // live values. DenseSets do not provide a deterministic order across runs.
1152 SmallVector<Value *, 64> Temp;
1153 Temp.insert(Temp.end(), live.begin(), live.end());
1154 std::sort(Temp.begin(), Temp.end(), order_by_name);
1155 for (Value *ptr : Temp) {
1156 Value *base = findBasePointer(ptr, DVCache);
1157 assert(base && "failed to find base pointer");
1158 PointerToBase[ptr] = base;
1159 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1160 DT->dominates(cast<Instruction>(base)->getParent(),
1161 cast<Instruction>(ptr)->getParent())) &&
1162 "The base we found better dominate the derived pointer");
1164 // If you see this trip and like to live really dangerously, the code should
1165 // be correct, just with idioms the verifier can't handle. You can try
1166 // disabling the verifier at your own substantial risk.
1167 assert(!isa<ConstantPointerNull>(base) &&
1168 "the relocation code needs adjustment to handle the relocation of "
1169 "a null pointer constant without causing false positives in the "
1170 "safepoint ir verifier.");
1174 /// Find the required based pointers (and adjust the live set) for the given
1176 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1178 PartiallyConstructedSafepointRecord &result) {
1179 DenseMap<Value *, Value *> PointerToBase;
1180 findBasePointers(result.LiveSet, PointerToBase, &DT, DVCache);
1182 if (PrintBasePointers) {
1183 // Note: Need to print these in a stable order since this is checked in
1185 errs() << "Base Pairs (w/o Relocation):\n";
1186 SmallVector<Value *, 64> Temp;
1187 Temp.reserve(PointerToBase.size());
1188 for (auto Pair : PointerToBase) {
1189 Temp.push_back(Pair.first);
1191 std::sort(Temp.begin(), Temp.end(), order_by_name);
1192 for (Value *Ptr : Temp) {
1193 Value *Base = PointerToBase[Ptr];
1194 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
1199 result.PointerToBase = PointerToBase;
1202 /// Given an updated version of the dataflow liveness results, update the
1203 /// liveset and base pointer maps for the call site CS.
1204 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1206 PartiallyConstructedSafepointRecord &result);
1208 static void recomputeLiveInValues(
1209 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1210 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1211 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1212 // again. The old values are still live and will help it stabilize quickly.
1213 GCPtrLivenessData RevisedLivenessData;
1214 computeLiveInValues(DT, F, RevisedLivenessData);
1215 for (size_t i = 0; i < records.size(); i++) {
1216 struct PartiallyConstructedSafepointRecord &info = records[i];
1217 const CallSite &CS = toUpdate[i];
1218 recomputeLiveInValues(RevisedLivenessData, CS, info);
1222 // When inserting gc.relocate calls, we need to ensure there are no uses
1223 // of the original value between the gc.statepoint and the gc.relocate call.
1224 // One case which can arise is a phi node starting one of the successor blocks.
1225 // We also need to be able to insert the gc.relocates only on the path which
1226 // goes through the statepoint. We might need to split an edge to make this
1229 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1230 DominatorTree &DT) {
1231 BasicBlock *Ret = BB;
1232 if (!BB->getUniquePredecessor()) {
1233 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1236 // Now that 'ret' has unique predecessor we can safely remove all phi nodes
1238 FoldSingleEntryPHINodes(Ret);
1239 assert(!isa<PHINode>(Ret->begin()));
1241 // At this point, we can safely insert a gc.relocate as the first instruction
1242 // in Ret if needed.
1246 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1247 auto itr = std::find(livevec.begin(), livevec.end(), val);
1248 assert(livevec.end() != itr);
1249 size_t index = std::distance(livevec.begin(), itr);
1250 assert(index < livevec.size());
1254 // Create new attribute set containing only attributes which can be transferred
1255 // from original call to the safepoint.
1256 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1259 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1260 unsigned index = AS.getSlotIndex(Slot);
1262 if (index == AttributeSet::ReturnIndex ||
1263 index == AttributeSet::FunctionIndex) {
1265 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1267 Attribute attr = *it;
1269 // Do not allow certain attributes - just skip them
1270 // Safepoint can not be read only or read none.
1271 if (attr.hasAttribute(Attribute::ReadNone) ||
1272 attr.hasAttribute(Attribute::ReadOnly))
1275 ret = ret.addAttributes(
1276 AS.getContext(), index,
1277 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1281 // Just skip parameter attributes for now
1287 /// Helper function to place all gc relocates necessary for the given
1290 /// liveVariables - list of variables to be relocated.
1291 /// liveStart - index of the first live variable.
1292 /// basePtrs - base pointers.
1293 /// statepointToken - statepoint instruction to which relocates should be
1295 /// Builder - Llvm IR builder to be used to construct new calls.
1296 static void CreateGCRelocates(ArrayRef<Value *> LiveVariables,
1297 const int LiveStart,
1298 ArrayRef<Value *> BasePtrs,
1299 Instruction *StatepointToken,
1300 IRBuilder<> Builder) {
1301 if (LiveVariables.empty())
1304 // All gc_relocate are set to i8 addrspace(1)* type. We originally generated
1305 // unique declarations for each pointer type, but this proved problematic
1306 // because the intrinsic mangling code is incomplete and fragile. Since
1307 // we're moving towards a single unified pointer type anyways, we can just
1308 // cast everything to an i8* of the right address space. A bitcast is added
1309 // later to convert gc_relocate to the actual value's type.
1310 Module *M = StatepointToken->getModule();
1311 auto AS = cast<PointerType>(LiveVariables[0]->getType())->getAddressSpace();
1312 Type *Types[] = {Type::getInt8PtrTy(M->getContext(), AS)};
1313 Value *GCRelocateDecl =
1314 Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, Types);
1316 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1317 // Generate the gc.relocate call and save the result
1319 Builder.getInt32(LiveStart + find_index(LiveVariables, BasePtrs[i]));
1321 Builder.getInt32(LiveStart + find_index(LiveVariables, LiveVariables[i]));
1323 // only specify a debug name if we can give a useful one
1324 CallInst *Reloc = Builder.CreateCall(
1325 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1326 suffixed_name_or(LiveVariables[i], ".relocated", ""));
1327 // Trick CodeGen into thinking there are lots of free registers at this
1329 Reloc->setCallingConv(CallingConv::Cold);
1334 makeStatepointExplicitImpl(const CallSite CS, /* to replace */
1335 const SmallVectorImpl<Value *> &BasePtrs,
1336 const SmallVectorImpl<Value *> &LiveVariables,
1337 PartiallyConstructedSafepointRecord &Result) {
1338 assert(BasePtrs.size() == LiveVariables.size());
1339 assert(isStatepoint(CS) &&
1340 "This method expects to be rewriting a statepoint");
1342 // We're not changing the function signature of the statepoint since the gc
1343 // arguments go into the var args section.
1344 Function *GCStatepointDecl = CS.getCalledFunction();
1346 // Then go ahead and use the builder do actually do the inserts. We insert
1347 // immediately before the previous instruction under the assumption that all
1348 // arguments will be available here. We can't insert afterwards since we may
1349 // be replacing a terminator.
1350 Instruction *InsertBefore = CS.getInstruction();
1351 IRBuilder<> Builder(InsertBefore);
1353 // Copy all of the arguments from the original statepoint - this includes the
1354 // target, call args, and deopt args
1355 SmallVector<llvm::Value *, 64> Args;
1356 Args.insert(Args.end(), CS.arg_begin(), CS.arg_end());
1357 // TODO: Clear the 'needs rewrite' flag
1359 // Add all the pointers to be relocated (gc arguments) and capture the start
1360 // of the live variable list for use in the gc_relocates
1361 const int LiveStartIdx = Args.size();
1362 Args.insert(Args.end(), LiveVariables.begin(), LiveVariables.end());
1364 // Create the statepoint given all the arguments
1365 Instruction *Token = nullptr;
1366 AttributeSet ReturnAttrs;
1368 CallInst *ToReplace = cast<CallInst>(CS.getInstruction());
1370 Builder.CreateCall(GCStatepointDecl, Args, "safepoint_token");
1371 Call->setTailCall(ToReplace->isTailCall());
1372 Call->setCallingConv(ToReplace->getCallingConv());
1374 // Currently we will fail on parameter attributes and on certain
1375 // function attributes.
1376 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes());
1377 // In case if we can handle this set of attributes - set up function attrs
1378 // directly on statepoint and return attrs later for gc_result intrinsic.
1379 Call->setAttributes(NewAttrs.getFnAttributes());
1380 ReturnAttrs = NewAttrs.getRetAttributes();
1384 // Put the following gc_result and gc_relocate calls immediately after the
1385 // the old call (which we're about to delete)
1386 assert(ToReplace->getNextNode() && "Not a terminator, must have next!");
1387 Builder.SetInsertPoint(ToReplace->getNextNode());
1388 Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc());
1390 InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction());
1392 // Insert the new invoke into the old block. We'll remove the old one in a
1393 // moment at which point this will become the new terminator for the
1395 InvokeInst *Invoke =
1396 InvokeInst::Create(GCStatepointDecl, ToReplace->getNormalDest(),
1397 ToReplace->getUnwindDest(), Args, "statepoint_token",
1398 ToReplace->getParent());
1399 Invoke->setCallingConv(ToReplace->getCallingConv());
1401 // Currently we will fail on parameter attributes and on certain
1402 // function attributes.
1403 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes());
1404 // In case if we can handle this set of attributes - set up function attrs
1405 // directly on statepoint and return attrs later for gc_result intrinsic.
1406 Invoke->setAttributes(NewAttrs.getFnAttributes());
1407 ReturnAttrs = NewAttrs.getRetAttributes();
1411 // Generate gc relocates in exceptional path
1412 BasicBlock *UnwindBlock = ToReplace->getUnwindDest();
1413 assert(!isa<PHINode>(UnwindBlock->begin()) &&
1414 UnwindBlock->getUniquePredecessor() &&
1415 "can't safely insert in this block!");
1417 Builder.SetInsertPoint(UnwindBlock->getFirstInsertionPt());
1418 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1420 // Extract second element from landingpad return value. We will attach
1421 // exceptional gc relocates to it.
1422 Instruction *ExceptionalToken =
1423 cast<Instruction>(Builder.CreateExtractValue(
1424 UnwindBlock->getLandingPadInst(), 1, "relocate_token"));
1425 Result.UnwindToken = ExceptionalToken;
1427 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, ExceptionalToken,
1430 // Generate gc relocates and returns for normal block
1431 BasicBlock *NormalDest = ToReplace->getNormalDest();
1432 assert(!isa<PHINode>(NormalDest->begin()) &&
1433 NormalDest->getUniquePredecessor() &&
1434 "can't safely insert in this block!");
1436 Builder.SetInsertPoint(NormalDest->getFirstInsertionPt());
1438 // gc relocates will be generated later as if it were regular call
1441 assert(Token && "Should be set in one of the above branches!");
1443 // Take the name of the original value call if it had one.
1444 Token->takeName(CS.getInstruction());
1446 // The GCResult is already inserted, we just need to find it
1448 Instruction *ToReplace = CS.getInstruction();
1449 assert(!ToReplace->hasNUsesOrMore(2) &&
1450 "only valid use before rewrite is gc.result");
1451 assert(!ToReplace->hasOneUse() ||
1452 isGCResult(cast<Instruction>(*ToReplace->user_begin())));
1455 // Update the gc.result of the original statepoint (if any) to use the newly
1456 // inserted statepoint. This is safe to do here since the token can't be
1457 // considered a live reference.
1458 CS.getInstruction()->replaceAllUsesWith(Token);
1460 Result.StatepointToken = Token;
1462 // Second, create a gc.relocate for every live variable
1463 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, Token, Builder);
1467 struct NameOrdering {
1471 bool operator()(NameOrdering const &a, NameOrdering const &b) {
1472 return -1 == a.Derived->getName().compare(b.Derived->getName());
1477 static void StabilizeOrder(SmallVectorImpl<Value *> &BaseVec,
1478 SmallVectorImpl<Value *> &LiveVec) {
1479 assert(BaseVec.size() == LiveVec.size());
1481 SmallVector<NameOrdering, 64> Temp;
1482 for (size_t i = 0; i < BaseVec.size(); i++) {
1484 v.Base = BaseVec[i];
1485 v.Derived = LiveVec[i];
1489 std::sort(Temp.begin(), Temp.end(), NameOrdering());
1490 for (size_t i = 0; i < BaseVec.size(); i++) {
1491 BaseVec[i] = Temp[i].Base;
1492 LiveVec[i] = Temp[i].Derived;
1496 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1497 // which make the relocations happening at this safepoint explicit.
1499 // WARNING: Does not do any fixup to adjust users of the original live
1500 // values. That's the callers responsibility.
1502 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS,
1503 PartiallyConstructedSafepointRecord &Result) {
1504 const auto &LiveSet = Result.LiveSet;
1505 const auto &PointerToBase = Result.PointerToBase;
1507 // Convert to vector for efficient cross referencing.
1508 SmallVector<Value *, 64> BaseVec, LiveVec;
1509 LiveVec.reserve(LiveSet.size());
1510 BaseVec.reserve(LiveSet.size());
1511 for (Value *L : LiveSet) {
1512 LiveVec.push_back(L);
1513 assert(PointerToBase.count(L));
1514 Value *Base = PointerToBase.find(L)->second;
1515 BaseVec.push_back(Base);
1517 assert(LiveVec.size() == BaseVec.size());
1519 // To make the output IR slightly more stable (for use in diffs), ensure a
1520 // fixed order of the values in the safepoint (by sorting the value name).
1521 // The order is otherwise meaningless.
1522 StabilizeOrder(BaseVec, LiveVec);
1524 // Do the actual rewriting and delete the old statepoint
1525 makeStatepointExplicitImpl(CS, BaseVec, LiveVec, Result);
1526 CS.getInstruction()->eraseFromParent();
1529 // Helper function for the relocationViaAlloca.
1531 // It receives iterator to the statepoint gc relocates and emits a store to the
1532 // assigned location (via allocaMap) for the each one of them. It adds the
1533 // visited values into the visitedLiveValues set, which we will later use them
1534 // for sanity checking.
1536 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1537 DenseMap<Value *, Value *> &AllocaMap,
1538 DenseSet<Value *> &VisitedLiveValues) {
1540 for (User *U : GCRelocs) {
1541 if (!isa<IntrinsicInst>(U))
1544 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1546 // We only care about relocates
1547 if (RelocatedValue->getIntrinsicID() !=
1548 Intrinsic::experimental_gc_relocate) {
1552 GCRelocateOperands RelocateOperands(RelocatedValue);
1553 Value *OriginalValue =
1554 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1555 assert(AllocaMap.count(OriginalValue));
1556 Value *Alloca = AllocaMap[OriginalValue];
1558 // Emit store into the related alloca
1559 // All gc_relocates are i8 addrspace(1)* typed, and it must be bitcasted to
1560 // the correct type according to alloca.
1561 assert(RelocatedValue->getNextNode() &&
1562 "Should always have one since it's not a terminator");
1563 IRBuilder<> Builder(RelocatedValue->getNextNode());
1564 Value *CastedRelocatedValue =
1565 Builder.CreateBitCast(RelocatedValue,
1566 cast<AllocaInst>(Alloca)->getAllocatedType(),
1567 suffixed_name_or(RelocatedValue, ".casted", ""));
1569 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1570 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1573 VisitedLiveValues.insert(OriginalValue);
1578 // Helper function for the "relocationViaAlloca". Similar to the
1579 // "insertRelocationStores" but works for rematerialized values.
1581 insertRematerializationStores(
1582 RematerializedValueMapTy RematerializedValues,
1583 DenseMap<Value *, Value *> &AllocaMap,
1584 DenseSet<Value *> &VisitedLiveValues) {
1586 for (auto RematerializedValuePair: RematerializedValues) {
1587 Instruction *RematerializedValue = RematerializedValuePair.first;
1588 Value *OriginalValue = RematerializedValuePair.second;
1590 assert(AllocaMap.count(OriginalValue) &&
1591 "Can not find alloca for rematerialized value");
1592 Value *Alloca = AllocaMap[OriginalValue];
1594 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1595 Store->insertAfter(RematerializedValue);
1598 VisitedLiveValues.insert(OriginalValue);
1603 /// Do all the relocation update via allocas and mem2reg
1604 static void relocationViaAlloca(
1605 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1606 ArrayRef<PartiallyConstructedSafepointRecord> Records) {
1608 // record initial number of (static) allocas; we'll check we have the same
1609 // number when we get done.
1610 int InitialAllocaNum = 0;
1611 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1613 if (isa<AllocaInst>(*I))
1617 // TODO-PERF: change data structures, reserve
1618 DenseMap<Value *, Value *> AllocaMap;
1619 SmallVector<AllocaInst *, 200> PromotableAllocas;
1620 // Used later to chack that we have enough allocas to store all values
1621 std::size_t NumRematerializedValues = 0;
1622 PromotableAllocas.reserve(Live.size());
1624 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1625 // "PromotableAllocas"
1626 auto emitAllocaFor = [&](Value *LiveValue) {
1627 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1628 F.getEntryBlock().getFirstNonPHI());
1629 AllocaMap[LiveValue] = Alloca;
1630 PromotableAllocas.push_back(Alloca);
1633 // Emit alloca for each live gc pointer
1634 for (Value *V : Live)
1637 // Emit allocas for rematerialized values
1638 for (const auto &Info : Records)
1639 for (auto RematerializedValuePair : Info.RematerializedValues) {
1640 Value *OriginalValue = RematerializedValuePair.second;
1641 if (AllocaMap.count(OriginalValue) != 0)
1644 emitAllocaFor(OriginalValue);
1645 ++NumRematerializedValues;
1648 // The next two loops are part of the same conceptual operation. We need to
1649 // insert a store to the alloca after the original def and at each
1650 // redefinition. We need to insert a load before each use. These are split
1651 // into distinct loops for performance reasons.
1653 // Update gc pointer after each statepoint: either store a relocated value or
1654 // null (if no relocated value was found for this gc pointer and it is not a
1655 // gc_result). This must happen before we update the statepoint with load of
1656 // alloca otherwise we lose the link between statepoint and old def.
1657 for (const auto &Info : Records) {
1658 Value *Statepoint = Info.StatepointToken;
1660 // This will be used for consistency check
1661 DenseSet<Value *> VisitedLiveValues;
1663 // Insert stores for normal statepoint gc relocates
1664 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1666 // In case if it was invoke statepoint
1667 // we will insert stores for exceptional path gc relocates.
1668 if (isa<InvokeInst>(Statepoint)) {
1669 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1673 // Do similar thing with rematerialized values
1674 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1677 if (ClobberNonLive) {
1678 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1679 // the gc.statepoint. This will turn some subtle GC problems into
1680 // slightly easier to debug SEGVs. Note that on large IR files with
1681 // lots of gc.statepoints this is extremely costly both memory and time
1683 SmallVector<AllocaInst *, 64> ToClobber;
1684 for (auto Pair : AllocaMap) {
1685 Value *Def = Pair.first;
1686 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1688 // This value was relocated
1689 if (VisitedLiveValues.count(Def)) {
1692 ToClobber.push_back(Alloca);
1695 auto InsertClobbersAt = [&](Instruction *IP) {
1696 for (auto *AI : ToClobber) {
1697 auto AIType = cast<PointerType>(AI->getType());
1698 auto PT = cast<PointerType>(AIType->getElementType());
1699 Constant *CPN = ConstantPointerNull::get(PT);
1700 StoreInst *Store = new StoreInst(CPN, AI);
1701 Store->insertBefore(IP);
1705 // Insert the clobbering stores. These may get intermixed with the
1706 // gc.results and gc.relocates, but that's fine.
1707 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1708 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1709 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1711 InsertClobbersAt(cast<Instruction>(Statepoint)->getNextNode());
1716 // Update use with load allocas and add store for gc_relocated.
1717 for (auto Pair : AllocaMap) {
1718 Value *Def = Pair.first;
1719 Value *Alloca = Pair.second;
1721 // We pre-record the uses of allocas so that we dont have to worry about
1722 // later update that changes the user information..
1724 SmallVector<Instruction *, 20> Uses;
1725 // PERF: trade a linear scan for repeated reallocation
1726 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1727 for (User *U : Def->users()) {
1728 if (!isa<ConstantExpr>(U)) {
1729 // If the def has a ConstantExpr use, then the def is either a
1730 // ConstantExpr use itself or null. In either case
1731 // (recursively in the first, directly in the second), the oop
1732 // it is ultimately dependent on is null and this particular
1733 // use does not need to be fixed up.
1734 Uses.push_back(cast<Instruction>(U));
1738 std::sort(Uses.begin(), Uses.end());
1739 auto Last = std::unique(Uses.begin(), Uses.end());
1740 Uses.erase(Last, Uses.end());
1742 for (Instruction *Use : Uses) {
1743 if (isa<PHINode>(Use)) {
1744 PHINode *Phi = cast<PHINode>(Use);
1745 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1746 if (Def == Phi->getIncomingValue(i)) {
1747 LoadInst *Load = new LoadInst(
1748 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1749 Phi->setIncomingValue(i, Load);
1753 LoadInst *Load = new LoadInst(Alloca, "", Use);
1754 Use->replaceUsesOfWith(Def, Load);
1758 // Emit store for the initial gc value. Store must be inserted after load,
1759 // otherwise store will be in alloca's use list and an extra load will be
1760 // inserted before it.
1761 StoreInst *Store = new StoreInst(Def, Alloca);
1762 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1763 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1764 // InvokeInst is a TerminatorInst so the store need to be inserted
1765 // into its normal destination block.
1766 BasicBlock *NormalDest = Invoke->getNormalDest();
1767 Store->insertBefore(NormalDest->getFirstNonPHI());
1769 assert(!Inst->isTerminator() &&
1770 "The only TerminatorInst that can produce a value is "
1771 "InvokeInst which is handled above.");
1772 Store->insertAfter(Inst);
1775 assert(isa<Argument>(Def));
1776 Store->insertAfter(cast<Instruction>(Alloca));
1780 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1781 "we must have the same allocas with lives");
1782 if (!PromotableAllocas.empty()) {
1783 // Apply mem2reg to promote alloca to SSA
1784 PromoteMemToReg(PromotableAllocas, DT);
1788 for (auto &I : F.getEntryBlock())
1789 if (isa<AllocaInst>(I))
1791 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1795 /// Implement a unique function which doesn't require we sort the input
1796 /// vector. Doing so has the effect of changing the output of a couple of
1797 /// tests in ways which make them less useful in testing fused safepoints.
1798 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1799 SmallSet<T, 8> Seen;
1800 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1801 return !Seen.insert(V).second;
1805 /// Insert holders so that each Value is obviously live through the entire
1806 /// lifetime of the call.
1807 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1808 SmallVectorImpl<CallInst *> &Holders) {
1810 // No values to hold live, might as well not insert the empty holder
1813 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1814 // Use a dummy vararg function to actually hold the values live
1815 Function *Func = cast<Function>(M->getOrInsertFunction(
1816 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1818 // For call safepoints insert dummy calls right after safepoint
1819 BasicBlock::iterator Next(CS.getInstruction());
1821 Holders.push_back(CallInst::Create(Func, Values, "", Next));
1824 // For invoke safepooints insert dummy calls both in normal and
1825 // exceptional destination blocks
1826 auto *II = cast<InvokeInst>(CS.getInstruction());
1827 Holders.push_back(CallInst::Create(
1828 Func, Values, "", II->getNormalDest()->getFirstInsertionPt()));
1829 Holders.push_back(CallInst::Create(
1830 Func, Values, "", II->getUnwindDest()->getFirstInsertionPt()));
1833 static void findLiveReferences(
1834 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1835 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1836 GCPtrLivenessData OriginalLivenessData;
1837 computeLiveInValues(DT, F, OriginalLivenessData);
1838 for (size_t i = 0; i < records.size(); i++) {
1839 struct PartiallyConstructedSafepointRecord &info = records[i];
1840 const CallSite &CS = toUpdate[i];
1841 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1845 /// Remove any vector of pointers from the live set by scalarizing them over the
1846 /// statepoint instruction. Adds the scalarized pieces to the live set. It
1847 /// would be preferable to include the vector in the statepoint itself, but
1848 /// the lowering code currently does not handle that. Extending it would be
1849 /// slightly non-trivial since it requires a format change. Given how rare
1850 /// such cases are (for the moment?) scalarizing is an acceptable compromise.
1851 static void splitVectorValues(Instruction *StatepointInst,
1852 StatepointLiveSetTy &LiveSet,
1853 DenseMap<Value *, Value *>& PointerToBase,
1854 DominatorTree &DT) {
1855 SmallVector<Value *, 16> ToSplit;
1856 for (Value *V : LiveSet)
1857 if (isa<VectorType>(V->getType()))
1858 ToSplit.push_back(V);
1860 if (ToSplit.empty())
1863 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1865 Function &F = *(StatepointInst->getParent()->getParent());
1867 DenseMap<Value *, AllocaInst *> AllocaMap;
1868 // First is normal return, second is exceptional return (invoke only)
1869 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1870 for (Value *V : ToSplit) {
1871 AllocaInst *Alloca =
1872 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1873 AllocaMap[V] = Alloca;
1875 VectorType *VT = cast<VectorType>(V->getType());
1876 IRBuilder<> Builder(StatepointInst);
1877 SmallVector<Value *, 16> Elements;
1878 for (unsigned i = 0; i < VT->getNumElements(); i++)
1879 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1880 ElementMapping[V] = Elements;
1882 auto InsertVectorReform = [&](Instruction *IP) {
1883 Builder.SetInsertPoint(IP);
1884 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1885 Value *ResultVec = UndefValue::get(VT);
1886 for (unsigned i = 0; i < VT->getNumElements(); i++)
1887 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1888 Builder.getInt32(i));
1892 if (isa<CallInst>(StatepointInst)) {
1893 BasicBlock::iterator Next(StatepointInst);
1895 Instruction *IP = &*(Next);
1896 Replacements[V].first = InsertVectorReform(IP);
1897 Replacements[V].second = nullptr;
1899 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1900 // We've already normalized - check that we don't have shared destination
1902 BasicBlock *NormalDest = Invoke->getNormalDest();
1903 assert(!isa<PHINode>(NormalDest->begin()));
1904 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1905 assert(!isa<PHINode>(UnwindDest->begin()));
1906 // Insert insert element sequences in both successors
1907 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1908 Replacements[V].first = InsertVectorReform(IP);
1909 IP = &*(UnwindDest->getFirstInsertionPt());
1910 Replacements[V].second = InsertVectorReform(IP);
1914 for (Value *V : ToSplit) {
1915 AllocaInst *Alloca = AllocaMap[V];
1917 // Capture all users before we start mutating use lists
1918 SmallVector<Instruction *, 16> Users;
1919 for (User *U : V->users())
1920 Users.push_back(cast<Instruction>(U));
1922 for (Instruction *I : Users) {
1923 if (auto Phi = dyn_cast<PHINode>(I)) {
1924 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1925 if (V == Phi->getIncomingValue(i)) {
1926 LoadInst *Load = new LoadInst(
1927 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1928 Phi->setIncomingValue(i, Load);
1931 LoadInst *Load = new LoadInst(Alloca, "", I);
1932 I->replaceUsesOfWith(V, Load);
1936 // Store the original value and the replacement value into the alloca
1937 StoreInst *Store = new StoreInst(V, Alloca);
1938 if (auto I = dyn_cast<Instruction>(V))
1939 Store->insertAfter(I);
1941 Store->insertAfter(Alloca);
1943 // Normal return for invoke, or call return
1944 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1945 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1946 // Unwind return for invoke only
1947 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1949 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1952 // apply mem2reg to promote alloca to SSA
1953 SmallVector<AllocaInst *, 16> Allocas;
1954 for (Value *V : ToSplit)
1955 Allocas.push_back(AllocaMap[V]);
1956 PromoteMemToReg(Allocas, DT);
1958 // Update our tracking of live pointers and base mappings to account for the
1959 // changes we just made.
1960 for (Value *V : ToSplit) {
1961 auto &Elements = ElementMapping[V];
1964 LiveSet.insert(Elements.begin(), Elements.end());
1965 // We need to update the base mapping as well.
1966 assert(PointerToBase.count(V));
1967 Value *OldBase = PointerToBase[V];
1968 auto &BaseElements = ElementMapping[OldBase];
1969 PointerToBase.erase(V);
1970 assert(Elements.size() == BaseElements.size());
1971 for (unsigned i = 0; i < Elements.size(); i++) {
1972 Value *Elem = Elements[i];
1973 PointerToBase[Elem] = BaseElements[i];
1978 // Helper function for the "rematerializeLiveValues". It walks use chain
1979 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
1980 // values are visited (currently it is GEP's and casts). Returns true if it
1981 // successfully reached "BaseValue" and false otherwise.
1982 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
1984 static bool findRematerializableChainToBasePointer(
1985 SmallVectorImpl<Instruction*> &ChainToBase,
1986 Value *CurrentValue, Value *BaseValue) {
1988 // We have found a base value
1989 if (CurrentValue == BaseValue) {
1993 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
1994 ChainToBase.push_back(GEP);
1995 return findRematerializableChainToBasePointer(ChainToBase,
1996 GEP->getPointerOperand(),
2000 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
2001 Value *Def = CI->stripPointerCasts();
2003 // This two checks are basically similar. First one is here for the
2004 // consistency with findBasePointers logic.
2005 assert(!isa<CastInst>(Def) && "not a pointer cast found");
2006 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
2009 ChainToBase.push_back(CI);
2010 return findRematerializableChainToBasePointer(ChainToBase, Def, BaseValue);
2013 // Not supported instruction in the chain
2017 // Helper function for the "rematerializeLiveValues". Compute cost of the use
2018 // chain we are going to rematerialize.
2020 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
2021 TargetTransformInfo &TTI) {
2024 for (Instruction *Instr : Chain) {
2025 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
2026 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
2027 "non noop cast is found during rematerialization");
2029 Type *SrcTy = CI->getOperand(0)->getType();
2030 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
2032 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
2033 // Cost of the address calculation
2034 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
2035 Cost += TTI.getAddressComputationCost(ValTy);
2037 // And cost of the GEP itself
2038 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
2039 // allowed for the external usage)
2040 if (!GEP->hasAllConstantIndices())
2044 llvm_unreachable("unsupported instruciton type during rematerialization");
2051 // From the statepoint live set pick values that are cheaper to recompute then
2052 // to relocate. Remove this values from the live set, rematerialize them after
2053 // statepoint and record them in "Info" structure. Note that similar to
2054 // relocated values we don't do any user adjustments here.
2055 static void rematerializeLiveValues(CallSite CS,
2056 PartiallyConstructedSafepointRecord &Info,
2057 TargetTransformInfo &TTI) {
2058 const unsigned int ChainLengthThreshold = 10;
2060 // Record values we are going to delete from this statepoint live set.
2061 // We can not di this in following loop due to iterator invalidation.
2062 SmallVector<Value *, 32> LiveValuesToBeDeleted;
2064 for (Value *LiveValue: Info.LiveSet) {
2065 // For each live pointer find it's defining chain
2066 SmallVector<Instruction *, 3> ChainToBase;
2067 assert(Info.PointerToBase.count(LiveValue));
2069 findRematerializableChainToBasePointer(ChainToBase,
2071 Info.PointerToBase[LiveValue]);
2072 // Nothing to do, or chain is too long
2074 ChainToBase.size() == 0 ||
2075 ChainToBase.size() > ChainLengthThreshold)
2078 // Compute cost of this chain
2079 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
2080 // TODO: We can also account for cases when we will be able to remove some
2081 // of the rematerialized values by later optimization passes. I.e if
2082 // we rematerialized several intersecting chains. Or if original values
2083 // don't have any uses besides this statepoint.
2085 // For invokes we need to rematerialize each chain twice - for normal and
2086 // for unwind basic blocks. Model this by multiplying cost by two.
2087 if (CS.isInvoke()) {
2090 // If it's too expensive - skip it
2091 if (Cost >= RematerializationThreshold)
2094 // Remove value from the live set
2095 LiveValuesToBeDeleted.push_back(LiveValue);
2097 // Clone instructions and record them inside "Info" structure
2099 // Walk backwards to visit top-most instructions first
2100 std::reverse(ChainToBase.begin(), ChainToBase.end());
2102 // Utility function which clones all instructions from "ChainToBase"
2103 // and inserts them before "InsertBefore". Returns rematerialized value
2104 // which should be used after statepoint.
2105 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
2106 Instruction *LastClonedValue = nullptr;
2107 Instruction *LastValue = nullptr;
2108 for (Instruction *Instr: ChainToBase) {
2109 // Only GEP's and casts are suported as we need to be careful to not
2110 // introduce any new uses of pointers not in the liveset.
2111 // Note that it's fine to introduce new uses of pointers which were
2112 // otherwise not used after this statepoint.
2113 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2115 Instruction *ClonedValue = Instr->clone();
2116 ClonedValue->insertBefore(InsertBefore);
2117 ClonedValue->setName(Instr->getName() + ".remat");
2119 // If it is not first instruction in the chain then it uses previously
2120 // cloned value. We should update it to use cloned value.
2121 if (LastClonedValue) {
2123 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2125 // Assert that cloned instruction does not use any instructions from
2126 // this chain other than LastClonedValue
2127 for (auto OpValue : ClonedValue->operand_values()) {
2128 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2129 ChainToBase.end() &&
2130 "incorrect use in rematerialization chain");
2135 LastClonedValue = ClonedValue;
2138 assert(LastClonedValue);
2139 return LastClonedValue;
2142 // Different cases for calls and invokes. For invokes we need to clone
2143 // instructions both on normal and unwind path.
2145 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2146 assert(InsertBefore);
2147 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2148 Info.RematerializedValues[RematerializedValue] = LiveValue;
2150 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2152 Instruction *NormalInsertBefore =
2153 Invoke->getNormalDest()->getFirstInsertionPt();
2154 Instruction *UnwindInsertBefore =
2155 Invoke->getUnwindDest()->getFirstInsertionPt();
2157 Instruction *NormalRematerializedValue =
2158 rematerializeChain(NormalInsertBefore);
2159 Instruction *UnwindRematerializedValue =
2160 rematerializeChain(UnwindInsertBefore);
2162 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2163 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2167 // Remove rematerializaed values from the live set
2168 for (auto LiveValue: LiveValuesToBeDeleted) {
2169 Info.LiveSet.erase(LiveValue);
2173 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
2174 SmallVectorImpl<CallSite> &ToUpdate) {
2176 // sanity check the input
2177 std::set<CallSite> Uniqued;
2178 Uniqued.insert(ToUpdate.begin(), ToUpdate.end());
2179 assert(Uniqued.size() == ToUpdate.size() && "no duplicates please!");
2181 for (CallSite CS : ToUpdate) {
2182 assert(CS.getInstruction()->getParent()->getParent() == &F);
2183 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
2187 // When inserting gc.relocates for invokes, we need to be able to insert at
2188 // the top of the successor blocks. See the comment on
2189 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2190 // may restructure the CFG.
2191 for (CallSite CS : ToUpdate) {
2194 auto *II = cast<InvokeInst>(CS.getInstruction());
2195 normalizeForInvokeSafepoint(II->getNormalDest(), II->getParent(), DT);
2196 normalizeForInvokeSafepoint(II->getUnwindDest(), II->getParent(), DT);
2199 // A list of dummy calls added to the IR to keep various values obviously
2200 // live in the IR. We'll remove all of these when done.
2201 SmallVector<CallInst *, 64> Holders;
2203 // Insert a dummy call with all of the arguments to the vm_state we'll need
2204 // for the actual safepoint insertion. This ensures reference arguments in
2205 // the deopt argument list are considered live through the safepoint (and
2206 // thus makes sure they get relocated.)
2207 for (CallSite CS : ToUpdate) {
2208 Statepoint StatepointCS(CS);
2210 SmallVector<Value *, 64> DeoptValues;
2211 for (Use &U : StatepointCS.vm_state_args()) {
2212 Value *Arg = cast<Value>(&U);
2213 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2214 "support for FCA unimplemented");
2215 if (isHandledGCPointerType(Arg->getType()))
2216 DeoptValues.push_back(Arg);
2218 insertUseHolderAfter(CS, DeoptValues, Holders);
2221 SmallVector<PartiallyConstructedSafepointRecord, 64> Records(ToUpdate.size());
2223 // A) Identify all gc pointers which are statically live at the given call
2225 findLiveReferences(F, DT, P, ToUpdate, Records);
2227 // B) Find the base pointers for each live pointer
2228 /* scope for caching */ {
2229 // Cache the 'defining value' relation used in the computation and
2230 // insertion of base phis and selects. This ensures that we don't insert
2231 // large numbers of duplicate base_phis.
2232 DefiningValueMapTy DVCache;
2234 for (size_t i = 0; i < Records.size(); i++) {
2235 PartiallyConstructedSafepointRecord &info = Records[i];
2236 findBasePointers(DT, DVCache, ToUpdate[i], info);
2238 } // end of cache scope
2240 // The base phi insertion logic (for any safepoint) may have inserted new
2241 // instructions which are now live at some safepoint. The simplest such
2244 // phi a <-- will be a new base_phi here
2245 // safepoint 1 <-- that needs to be live here
2249 // We insert some dummy calls after each safepoint to definitely hold live
2250 // the base pointers which were identified for that safepoint. We'll then
2251 // ask liveness for _every_ base inserted to see what is now live. Then we
2252 // remove the dummy calls.
2253 Holders.reserve(Holders.size() + Records.size());
2254 for (size_t i = 0; i < Records.size(); i++) {
2255 PartiallyConstructedSafepointRecord &Info = Records[i];
2257 SmallVector<Value *, 128> Bases;
2258 for (auto Pair : Info.PointerToBase)
2259 Bases.push_back(Pair.second);
2261 insertUseHolderAfter(ToUpdate[i], Bases, Holders);
2264 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2265 // need to rerun liveness. We may *also* have inserted new defs, but that's
2266 // not the key issue.
2267 recomputeLiveInValues(F, DT, P, ToUpdate, Records);
2269 if (PrintBasePointers) {
2270 for (auto &Info : Records) {
2271 errs() << "Base Pairs: (w/Relocation)\n";
2272 for (auto Pair : Info.PointerToBase)
2273 errs() << " derived %" << Pair.first->getName() << " base %"
2274 << Pair.second->getName() << "\n";
2278 for (CallInst *CI : Holders)
2279 CI->eraseFromParent();
2283 // Do a limited scalarization of any live at safepoint vector values which
2284 // contain pointers. This enables this pass to run after vectorization at
2285 // the cost of some possible performance loss. TODO: it would be nice to
2286 // natively support vectors all the way through the backend so we don't need
2287 // to scalarize here.
2288 for (size_t i = 0; i < Records.size(); i++) {
2289 PartiallyConstructedSafepointRecord &Info = Records[i];
2290 Instruction *Statepoint = ToUpdate[i].getInstruction();
2291 splitVectorValues(cast<Instruction>(Statepoint), Info.LiveSet,
2292 Info.PointerToBase, DT);
2295 // In order to reduce live set of statepoint we might choose to rematerialize
2296 // some values instead of relocating them. This is purely an optimization and
2297 // does not influence correctness.
2298 TargetTransformInfo &TTI =
2299 P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2301 for (size_t i = 0; i < Records.size(); i++)
2302 rematerializeLiveValues(ToUpdate[i], Records[i], TTI);
2304 // Now run through and replace the existing statepoints with new ones with
2305 // the live variables listed. We do not yet update uses of the values being
2306 // relocated. We have references to live variables that need to
2307 // survive to the last iteration of this loop. (By construction, the
2308 // previous statepoint can not be a live variable, thus we can and remove
2309 // the old statepoint calls as we go.)
2310 for (size_t i = 0; i < Records.size(); i++)
2311 makeStatepointExplicit(DT, ToUpdate[i], Records[i]);
2313 ToUpdate.clear(); // prevent accident use of invalid CallSites
2315 // Do all the fixups of the original live variables to their relocated selves
2316 SmallVector<Value *, 128> Live;
2317 for (size_t i = 0; i < Records.size(); i++) {
2318 PartiallyConstructedSafepointRecord &Info = Records[i];
2319 // We can't simply save the live set from the original insertion. One of
2320 // the live values might be the result of a call which needs a safepoint.
2321 // That Value* no longer exists and we need to use the new gc_result.
2322 // Thankfully, the live set is embedded in the statepoint (and updated), so
2323 // we just grab that.
2324 Statepoint Statepoint(Info.StatepointToken);
2325 Live.insert(Live.end(), Statepoint.gc_args_begin(),
2326 Statepoint.gc_args_end());
2328 // Do some basic sanity checks on our liveness results before performing
2329 // relocation. Relocation can and will turn mistakes in liveness results
2330 // into non-sensical code which is must harder to debug.
2331 // TODO: It would be nice to test consistency as well
2332 assert(DT.isReachableFromEntry(Info.StatepointToken->getParent()) &&
2333 "statepoint must be reachable or liveness is meaningless");
2334 for (Value *V : Statepoint.gc_args()) {
2335 if (!isa<Instruction>(V))
2336 // Non-instruction values trivial dominate all possible uses
2338 auto *LiveInst = cast<Instruction>(V);
2339 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2340 "unreachable values should never be live");
2341 assert(DT.dominates(LiveInst, Info.StatepointToken) &&
2342 "basic SSA liveness expectation violated by liveness analysis");
2346 unique_unsorted(Live);
2350 for (auto *Ptr : Live)
2351 assert(isGCPointerType(Ptr->getType()) && "must be a gc pointer type");
2354 relocationViaAlloca(F, DT, Live, Records);
2355 return !Records.empty();
2358 // Handles both return values and arguments for Functions and CallSites.
2359 template <typename AttrHolder>
2360 static void RemoveDerefAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2363 if (AH.getDereferenceableBytes(Index))
2364 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2365 AH.getDereferenceableBytes(Index)));
2366 if (AH.getDereferenceableOrNullBytes(Index))
2367 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2368 AH.getDereferenceableOrNullBytes(Index)));
2371 AH.setAttributes(AH.getAttributes().removeAttributes(
2372 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2376 RewriteStatepointsForGC::stripDereferenceabilityInfoFromPrototype(Function &F) {
2377 LLVMContext &Ctx = F.getContext();
2379 for (Argument &A : F.args())
2380 if (isa<PointerType>(A.getType()))
2381 RemoveDerefAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2383 if (isa<PointerType>(F.getReturnType()))
2384 RemoveDerefAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2387 void RewriteStatepointsForGC::stripDereferenceabilityInfoFromBody(Function &F) {
2391 LLVMContext &Ctx = F.getContext();
2392 MDBuilder Builder(Ctx);
2394 for (Instruction &I : instructions(F)) {
2395 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2396 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2397 bool IsImmutableTBAA =
2398 MD->getNumOperands() == 4 &&
2399 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2401 if (!IsImmutableTBAA)
2402 continue; // no work to do, MD_tbaa is already marked mutable
2404 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2405 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2407 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2409 MDNode *MutableTBAA =
2410 Builder.createTBAAStructTagNode(Base, Access, Offset);
2411 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2414 if (CallSite CS = CallSite(&I)) {
2415 for (int i = 0, e = CS.arg_size(); i != e; i++)
2416 if (isa<PointerType>(CS.getArgument(i)->getType()))
2417 RemoveDerefAttrAtIndex(Ctx, CS, i + 1);
2418 if (isa<PointerType>(CS.getType()))
2419 RemoveDerefAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2424 /// Returns true if this function should be rewritten by this pass. The main
2425 /// point of this function is as an extension point for custom logic.
2426 static bool shouldRewriteStatepointsIn(Function &F) {
2427 // TODO: This should check the GCStrategy
2429 const char *FunctionGCName = F.getGC();
2430 const StringRef StatepointExampleName("statepoint-example");
2431 const StringRef CoreCLRName("coreclr");
2432 return (StatepointExampleName == FunctionGCName) ||
2433 (CoreCLRName == FunctionGCName);
2438 void RewriteStatepointsForGC::stripDereferenceabilityInfo(Module &M) {
2440 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2444 for (Function &F : M)
2445 stripDereferenceabilityInfoFromPrototype(F);
2447 for (Function &F : M)
2448 stripDereferenceabilityInfoFromBody(F);
2451 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2452 // Nothing to do for declarations.
2453 if (F.isDeclaration() || F.empty())
2456 // Policy choice says not to rewrite - the most common reason is that we're
2457 // compiling code without a GCStrategy.
2458 if (!shouldRewriteStatepointsIn(F))
2461 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2463 // Gather all the statepoints which need rewritten. Be careful to only
2464 // consider those in reachable code since we need to ask dominance queries
2465 // when rewriting. We'll delete the unreachable ones in a moment.
2466 SmallVector<CallSite, 64> ParsePointNeeded;
2467 bool HasUnreachableStatepoint = false;
2468 for (Instruction &I : instructions(F)) {
2469 // TODO: only the ones with the flag set!
2470 if (isStatepoint(I)) {
2471 if (DT.isReachableFromEntry(I.getParent()))
2472 ParsePointNeeded.push_back(CallSite(&I));
2474 HasUnreachableStatepoint = true;
2478 bool MadeChange = false;
2480 // Delete any unreachable statepoints so that we don't have unrewritten
2481 // statepoints surviving this pass. This makes testing easier and the
2482 // resulting IR less confusing to human readers. Rather than be fancy, we
2483 // just reuse a utility function which removes the unreachable blocks.
2484 if (HasUnreachableStatepoint)
2485 MadeChange |= removeUnreachableBlocks(F);
2487 // Return early if no work to do.
2488 if (ParsePointNeeded.empty())
2491 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2492 // These are created by LCSSA. They have the effect of increasing the size
2493 // of liveness sets for no good reason. It may be harder to do this post
2494 // insertion since relocations and base phis can confuse things.
2495 for (BasicBlock &BB : F)
2496 if (BB.getUniquePredecessor()) {
2498 FoldSingleEntryPHINodes(&BB);
2501 // Before we start introducing relocations, we want to tweak the IR a bit to
2502 // avoid unfortunate code generation effects. The main example is that we
2503 // want to try to make sure the comparison feeding a branch is after any
2504 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2505 // values feeding a branch after relocation. This is semantically correct,
2506 // but results in extra register pressure since both the pre-relocation and
2507 // post-relocation copies must be available in registers. For code without
2508 // relocations this is handled elsewhere, but teaching the scheduler to
2509 // reverse the transform we're about to do would be slightly complex.
2510 // Note: This may extend the live range of the inputs to the icmp and thus
2511 // increase the liveset of any statepoint we move over. This is profitable
2512 // as long as all statepoints are in rare blocks. If we had in-register
2513 // lowering for live values this would be a much safer transform.
2514 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2515 if (auto *BI = dyn_cast<BranchInst>(TI))
2516 if (BI->isConditional())
2517 return dyn_cast<Instruction>(BI->getCondition());
2518 // TODO: Extend this to handle switches
2521 for (BasicBlock &BB : F) {
2522 TerminatorInst *TI = BB.getTerminator();
2523 if (auto *Cond = getConditionInst(TI))
2524 // TODO: Handle more than just ICmps here. We should be able to move
2525 // most instructions without side effects or memory access.
2526 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2528 Cond->moveBefore(TI);
2532 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
2536 // liveness computation via standard dataflow
2537 // -------------------------------------------------------------------
2539 // TODO: Consider using bitvectors for liveness, the set of potentially
2540 // interesting values should be small and easy to pre-compute.
2542 /// Compute the live-in set for the location rbegin starting from
2543 /// the live-out set of the basic block
2544 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2545 BasicBlock::reverse_iterator rend,
2546 DenseSet<Value *> &LiveTmp) {
2548 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2549 Instruction *I = &*ritr;
2551 // KILL/Def - Remove this definition from LiveIn
2554 // Don't consider *uses* in PHI nodes, we handle their contribution to
2555 // predecessor blocks when we seed the LiveOut sets
2556 if (isa<PHINode>(I))
2559 // USE - Add to the LiveIn set for this instruction
2560 for (Value *V : I->operands()) {
2561 assert(!isUnhandledGCPointerType(V->getType()) &&
2562 "support for FCA unimplemented");
2563 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2564 // The choice to exclude all things constant here is slightly subtle.
2565 // There are two independent reasons:
2566 // - We assume that things which are constant (from LLVM's definition)
2567 // do not move at runtime. For example, the address of a global
2568 // variable is fixed, even though it's contents may not be.
2569 // - Second, we can't disallow arbitrary inttoptr constants even
2570 // if the language frontend does. Optimization passes are free to
2571 // locally exploit facts without respect to global reachability. This
2572 // can create sections of code which are dynamically unreachable and
2573 // contain just about anything. (see constants.ll in tests)
2580 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2582 for (BasicBlock *Succ : successors(BB)) {
2583 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2584 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2585 PHINode *Phi = cast<PHINode>(&*I);
2586 Value *V = Phi->getIncomingValueForBlock(BB);
2587 assert(!isUnhandledGCPointerType(V->getType()) &&
2588 "support for FCA unimplemented");
2589 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2596 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2597 DenseSet<Value *> KillSet;
2598 for (Instruction &I : *BB)
2599 if (isHandledGCPointerType(I.getType()))
2605 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2606 /// sanity check for the liveness computation.
2607 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2608 TerminatorInst *TI, bool TermOkay = false) {
2609 for (Value *V : Live) {
2610 if (auto *I = dyn_cast<Instruction>(V)) {
2611 // The terminator can be a member of the LiveOut set. LLVM's definition
2612 // of instruction dominance states that V does not dominate itself. As
2613 // such, we need to special case this to allow it.
2614 if (TermOkay && TI == I)
2616 assert(DT.dominates(I, TI) &&
2617 "basic SSA liveness expectation violated by liveness analysis");
2622 /// Check that all the liveness sets used during the computation of liveness
2623 /// obey basic SSA properties. This is useful for finding cases where we miss
2625 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2627 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2628 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2629 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2633 static void computeLiveInValues(DominatorTree &DT, Function &F,
2634 GCPtrLivenessData &Data) {
2636 SmallSetVector<BasicBlock *, 200> Worklist;
2637 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2638 // We use a SetVector so that we don't have duplicates in the worklist.
2639 Worklist.insert(pred_begin(BB), pred_end(BB));
2641 auto NextItem = [&]() {
2642 BasicBlock *BB = Worklist.back();
2643 Worklist.pop_back();
2647 // Seed the liveness for each individual block
2648 for (BasicBlock &BB : F) {
2649 Data.KillSet[&BB] = computeKillSet(&BB);
2650 Data.LiveSet[&BB].clear();
2651 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2654 for (Value *Kill : Data.KillSet[&BB])
2655 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2658 Data.LiveOut[&BB] = DenseSet<Value *>();
2659 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2660 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2661 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2662 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2663 if (!Data.LiveIn[&BB].empty())
2664 AddPredsToWorklist(&BB);
2667 // Propagate that liveness until stable
2668 while (!Worklist.empty()) {
2669 BasicBlock *BB = NextItem();
2671 // Compute our new liveout set, then exit early if it hasn't changed
2672 // despite the contribution of our successor.
2673 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2674 const auto OldLiveOutSize = LiveOut.size();
2675 for (BasicBlock *Succ : successors(BB)) {
2676 assert(Data.LiveIn.count(Succ));
2677 set_union(LiveOut, Data.LiveIn[Succ]);
2679 // assert OutLiveOut is a subset of LiveOut
2680 if (OldLiveOutSize == LiveOut.size()) {
2681 // If the sets are the same size, then we didn't actually add anything
2682 // when unioning our successors LiveIn Thus, the LiveIn of this block
2686 Data.LiveOut[BB] = LiveOut;
2688 // Apply the effects of this basic block
2689 DenseSet<Value *> LiveTmp = LiveOut;
2690 set_union(LiveTmp, Data.LiveSet[BB]);
2691 set_subtract(LiveTmp, Data.KillSet[BB]);
2693 assert(Data.LiveIn.count(BB));
2694 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2695 // assert: OldLiveIn is a subset of LiveTmp
2696 if (OldLiveIn.size() != LiveTmp.size()) {
2697 Data.LiveIn[BB] = LiveTmp;
2698 AddPredsToWorklist(BB);
2700 } // while( !worklist.empty() )
2703 // Sanity check our output against SSA properties. This helps catch any
2704 // missing kills during the above iteration.
2705 for (BasicBlock &BB : F) {
2706 checkBasicSSA(DT, Data, BB);
2711 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2712 StatepointLiveSetTy &Out) {
2714 BasicBlock *BB = Inst->getParent();
2716 // Note: The copy is intentional and required
2717 assert(Data.LiveOut.count(BB));
2718 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2720 // We want to handle the statepoint itself oddly. It's
2721 // call result is not live (normal), nor are it's arguments
2722 // (unless they're used again later). This adjustment is
2723 // specifically what we need to relocate
2724 BasicBlock::reverse_iterator rend(Inst);
2725 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2726 LiveOut.erase(Inst);
2727 Out.insert(LiveOut.begin(), LiveOut.end());
2730 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2732 PartiallyConstructedSafepointRecord &Info) {
2733 Instruction *Inst = CS.getInstruction();
2734 StatepointLiveSetTy Updated;
2735 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2738 DenseSet<Value *> Bases;
2739 for (auto KVPair : Info.PointerToBase) {
2740 Bases.insert(KVPair.second);
2743 // We may have base pointers which are now live that weren't before. We need
2744 // to update the PointerToBase structure to reflect this.
2745 for (auto V : Updated)
2746 if (!Info.PointerToBase.count(V)) {
2747 assert(Bases.count(V) && "can't find base for unexpected live value");
2748 Info.PointerToBase[V] = V;
2753 for (auto V : Updated) {
2754 assert(Info.PointerToBase.count(V) &&
2755 "must be able to find base for live value");
2759 // Remove any stale base mappings - this can happen since our liveness is
2760 // more precise then the one inherent in the base pointer analysis
2761 DenseSet<Value *> ToErase;
2762 for (auto KVPair : Info.PointerToBase)
2763 if (!Updated.count(KVPair.first))
2764 ToErase.insert(KVPair.first);
2765 for (auto V : ToErase)
2766 Info.PointerToBase.erase(V);
2769 for (auto KVPair : Info.PointerToBase)
2770 assert(Updated.count(KVPair.first) && "record for non-live value");
2773 Info.LiveSet = Updated;