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<llvm::Value *> StatepointLiveSetTy;
162 typedef DenseMap<Instruction *, Value *> RematerializedValueMapTy;
164 struct PartiallyConstructedSafepointRecord {
165 /// The set of values known to be live across this safepoint
166 StatepointLiveSetTy liveset;
168 /// Mapping from live pointers to a base-defining-value
169 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
171 /// The *new* gc.statepoint instruction itself. This produces the token
172 /// that normal path gc.relocates and the gc.result are tied to.
173 Instruction *StatepointToken;
175 /// Instruction to which exceptional gc relocates are attached
176 /// Makes it easier to iterate through them during relocationViaAlloca.
177 Instruction *UnwindToken;
179 /// Record live values we are rematerialized instead of relocating.
180 /// They are not included into 'liveset' field.
181 /// Maps rematerialized copy to it's original value.
182 RematerializedValueMapTy RematerializedValues;
186 /// Compute the live-in set for every basic block in the function
187 static void computeLiveInValues(DominatorTree &DT, Function &F,
188 GCPtrLivenessData &Data);
190 /// Given results from the dataflow liveness computation, find the set of live
191 /// Values at a particular instruction.
192 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
193 StatepointLiveSetTy &out);
195 // TODO: Once we can get to the GCStrategy, this becomes
196 // Optional<bool> isGCManagedPointer(const Value *V) const override {
198 static bool isGCPointerType(Type *T) {
199 if (auto *PT = dyn_cast<PointerType>(T))
200 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
201 // GC managed heap. We know that a pointer into this heap needs to be
202 // updated and that no other pointer does.
203 return (1 == PT->getAddressSpace());
207 // Return true if this type is one which a) is a gc pointer or contains a GC
208 // pointer and b) is of a type this code expects to encounter as a live value.
209 // (The insertion code will assert that a type which matches (a) and not (b)
210 // is not encountered.)
211 static bool isHandledGCPointerType(Type *T) {
212 // We fully support gc pointers
213 if (isGCPointerType(T))
215 // We partially support vectors of gc pointers. The code will assert if it
216 // can't handle something.
217 if (auto VT = dyn_cast<VectorType>(T))
218 if (isGCPointerType(VT->getElementType()))
224 /// Returns true if this type contains a gc pointer whether we know how to
225 /// handle that type or not.
226 static bool containsGCPtrType(Type *Ty) {
227 if (isGCPointerType(Ty))
229 if (VectorType *VT = dyn_cast<VectorType>(Ty))
230 return isGCPointerType(VT->getScalarType());
231 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
232 return containsGCPtrType(AT->getElementType());
233 if (StructType *ST = dyn_cast<StructType>(Ty))
235 ST->subtypes().begin(), ST->subtypes().end(),
236 [](Type *SubType) { return containsGCPtrType(SubType); });
240 // Returns true if this is a type which a) is a gc pointer or contains a GC
241 // pointer and b) is of a type which the code doesn't expect (i.e. first class
242 // aggregates). Used to trip assertions.
243 static bool isUnhandledGCPointerType(Type *Ty) {
244 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
248 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
249 if (a->hasName() && b->hasName()) {
250 return -1 == a->getName().compare(b->getName());
251 } else if (a->hasName() && !b->hasName()) {
253 } else if (!a->hasName() && b->hasName()) {
256 // Better than nothing, but not stable
261 // Return the name of the value suffixed with the provided value, or if the
262 // value didn't have a name, the default value specified.
263 static std::string suffixed_name_or(Value *V, StringRef Suffix,
264 StringRef DefaultName) {
265 return V->hasName() ? (V->getName() + Suffix).str() : DefaultName.str();
268 // Conservatively identifies any definitions which might be live at the
269 // given instruction. The analysis is performed immediately before the
270 // given instruction. Values defined by that instruction are not considered
271 // live. Values used by that instruction are considered live.
272 static void analyzeParsePointLiveness(
273 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
274 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
275 Instruction *inst = CS.getInstruction();
277 StatepointLiveSetTy liveset;
278 findLiveSetAtInst(inst, OriginalLivenessData, liveset);
281 // Note: This output is used by several of the test cases
282 // The order of elements in a set is not stable, put them in a vec and sort
284 SmallVector<Value *, 64> Temp;
285 Temp.insert(Temp.end(), liveset.begin(), liveset.end());
286 std::sort(Temp.begin(), Temp.end(), order_by_name);
287 errs() << "Live Variables:\n";
288 for (Value *V : Temp)
289 dbgs() << " " << V->getName() << " " << *V << "\n";
291 if (PrintLiveSetSize) {
292 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
293 errs() << "Number live values: " << liveset.size() << "\n";
295 result.liveset = liveset;
298 static bool isKnownBaseResult(Value *V);
300 /// A single base defining value - An immediate base defining value for an
301 /// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'.
302 /// For instructions which have multiple pointer [vector] inputs or that
303 /// transition between vector and scalar types, there is no immediate base
304 /// defining value. The 'base defining value' for 'Def' is the transitive
305 /// closure of this relation stopping at the first instruction which has no
306 /// immediate base defining value. The b.d.v. might itself be a base pointer,
307 /// but it can also be an arbitrary derived pointer.
308 struct BaseDefiningValueResult {
309 /// Contains the value which is the base defining value.
311 /// True if the base defining value is also known to be an actual base
313 const bool IsKnownBase;
314 BaseDefiningValueResult(Value *BDV, bool IsKnownBase)
315 : BDV(BDV), IsKnownBase(IsKnownBase) {
317 // Check consistency between new and old means of checking whether a BDV is
319 bool MustBeBase = isKnownBaseResult(BDV);
320 assert(!MustBeBase || MustBeBase == IsKnownBase);
326 static BaseDefiningValueResult findBaseDefiningValue(Value *I);
328 /// Return a base defining value for the 'Index' element of the given vector
329 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
330 /// 'I'. As an optimization, this method will try to determine when the
331 /// element is known to already be a base pointer. If this can be established,
332 /// the second value in the returned pair will be true. Note that either a
333 /// vector or a pointer typed value can be returned. For the former, the
334 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
335 /// If the later, the return pointer is a BDV (or possibly a base) for the
336 /// particular element in 'I'.
337 static BaseDefiningValueResult
338 findBaseDefiningValueOfVector(Value *I) {
339 assert(I->getType()->isVectorTy() &&
340 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
341 "Illegal to ask for the base pointer of a non-pointer type");
343 // Each case parallels findBaseDefiningValue below, see that code for
344 // detailed motivation.
346 if (isa<Argument>(I))
347 // An incoming argument to the function is a base pointer
348 return BaseDefiningValueResult(I, true);
350 // We shouldn't see the address of a global as a vector value?
351 assert(!isa<GlobalVariable>(I) &&
352 "unexpected global variable found in base of vector");
354 // inlining could possibly introduce phi node that contains
355 // undef if callee has multiple returns
356 if (isa<UndefValue>(I))
357 // utterly meaningless, but useful for dealing with partially optimized
359 return BaseDefiningValueResult(I, true);
361 // Due to inheritance, this must be _after_ the global variable and undef
363 if (Constant *Con = dyn_cast<Constant>(I)) {
364 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
365 "order of checks wrong!");
366 assert(Con->isNullValue() && "null is the only case which makes sense");
367 return BaseDefiningValueResult(Con, true);
370 if (isa<LoadInst>(I))
371 return BaseDefiningValueResult(I, true);
373 if (isa<InsertElementInst>(I))
374 // We don't know whether this vector contains entirely base pointers or
375 // not. To be conservatively correct, we treat it as a BDV and will
376 // duplicate code as needed to construct a parallel vector of bases.
377 return BaseDefiningValueResult(I, false);
379 if (isa<ShuffleVectorInst>(I))
380 // We don't know whether this vector contains entirely base pointers or
381 // not. To be conservatively correct, we treat it as a BDV and will
382 // duplicate code as needed to construct a parallel vector of bases.
383 // TODO: There a number of local optimizations which could be applied here
384 // for particular sufflevector patterns.
385 return BaseDefiningValueResult(I, false);
387 // A PHI or Select is a base defining value. The outer findBasePointer
388 // algorithm is responsible for constructing a base value for this BDV.
389 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
390 "unknown vector instruction - no base found for vector element");
391 return BaseDefiningValueResult(I, false);
394 /// Helper function for findBasePointer - Will return a value which either a)
395 /// defines the base pointer for the input, b) blocks the simple search
396 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
397 /// from pointer to vector type or back.
398 static BaseDefiningValueResult findBaseDefiningValue(Value *I) {
399 if (I->getType()->isVectorTy())
400 return findBaseDefiningValueOfVector(I);
402 assert(I->getType()->isPointerTy() &&
403 "Illegal to ask for the base pointer of a non-pointer type");
405 if (isa<Argument>(I))
406 // An incoming argument to the function is a base pointer
407 // We should have never reached here if this argument isn't an gc value
408 return BaseDefiningValueResult(I, true);
410 if (isa<GlobalVariable>(I))
412 return BaseDefiningValueResult(I, true);
414 // inlining could possibly introduce phi node that contains
415 // undef if callee has multiple returns
416 if (isa<UndefValue>(I))
417 // utterly meaningless, but useful for dealing with
418 // partially optimized code.
419 return BaseDefiningValueResult(I, true);
421 // Due to inheritance, this must be _after_ the global variable and undef
423 if (isa<Constant>(I)) {
424 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
425 "order of checks wrong!");
426 // Note: Finding a constant base for something marked for relocation
427 // doesn't really make sense. The most likely case is either a) some
428 // screwed up the address space usage or b) your validating against
429 // compiled C++ code w/o the proper separation. The only real exception
430 // is a null pointer. You could have generic code written to index of
431 // off a potentially null value and have proven it null. We also use
432 // null pointers in dead paths of relocation phis (which we might later
433 // want to find a base pointer for).
434 assert(isa<ConstantPointerNull>(I) &&
435 "null is the only case which makes sense");
436 return BaseDefiningValueResult(I, true);
439 if (CastInst *CI = dyn_cast<CastInst>(I)) {
440 Value *Def = CI->stripPointerCasts();
441 // If we find a cast instruction here, it means we've found a cast which is
442 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
443 // handle int->ptr conversion.
444 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
445 return findBaseDefiningValue(Def);
448 if (isa<LoadInst>(I))
449 // The value loaded is an gc base itself
450 return BaseDefiningValueResult(I, true);
453 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
454 // The base of this GEP is the base
455 return findBaseDefiningValue(GEP->getPointerOperand());
457 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
458 switch (II->getIntrinsicID()) {
459 case Intrinsic::experimental_gc_result_ptr:
461 // fall through to general call handling
463 case Intrinsic::experimental_gc_statepoint:
464 case Intrinsic::experimental_gc_result_float:
465 case Intrinsic::experimental_gc_result_int:
466 llvm_unreachable("these don't produce pointers");
467 case Intrinsic::experimental_gc_relocate: {
468 // Rerunning safepoint insertion after safepoints are already
469 // inserted is not supported. It could probably be made to work,
470 // but why are you doing this? There's no good reason.
471 llvm_unreachable("repeat safepoint insertion is not supported");
473 case Intrinsic::gcroot:
474 // Currently, this mechanism hasn't been extended to work with gcroot.
475 // There's no reason it couldn't be, but I haven't thought about the
476 // implications much.
478 "interaction with the gcroot mechanism is not supported");
481 // We assume that functions in the source language only return base
482 // pointers. This should probably be generalized via attributes to support
483 // both source language and internal functions.
484 if (isa<CallInst>(I) || isa<InvokeInst>(I))
485 return BaseDefiningValueResult(I, true);
487 // I have absolutely no idea how to implement this part yet. It's not
488 // necessarily hard, I just haven't really looked at it yet.
489 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
491 if (isa<AtomicCmpXchgInst>(I))
492 // A CAS is effectively a atomic store and load combined under a
493 // predicate. From the perspective of base pointers, we just treat it
495 return BaseDefiningValueResult(I, true);
497 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
498 "binary ops which don't apply to pointers");
500 // The aggregate ops. Aggregates can either be in the heap or on the
501 // stack, but in either case, this is simply a field load. As a result,
502 // this is a defining definition of the base just like a load is.
503 if (isa<ExtractValueInst>(I))
504 return BaseDefiningValueResult(I, true);
506 // We should never see an insert vector since that would require we be
507 // tracing back a struct value not a pointer value.
508 assert(!isa<InsertValueInst>(I) &&
509 "Base pointer for a struct is meaningless");
511 // An extractelement produces a base result exactly when it's input does.
512 // We may need to insert a parallel instruction to extract the appropriate
513 // element out of the base vector corresponding to the input. Given this,
514 // it's analogous to the phi and select case even though it's not a merge.
515 if (isa<ExtractElementInst>(I))
516 // Note: There a lot of obvious peephole cases here. This are deliberately
517 // handled after the main base pointer inference algorithm to make writing
518 // test cases to exercise that code easier.
519 return BaseDefiningValueResult(I, false);
521 // The last two cases here don't return a base pointer. Instead, they
522 // return a value which dynamically selects from among several base
523 // derived pointers (each with it's own base potentially). It's the job of
524 // the caller to resolve these.
525 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
526 "missing instruction case in findBaseDefiningValing");
527 return BaseDefiningValueResult(I, false);
530 /// Returns the base defining value for this value.
531 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
532 Value *&Cached = Cache[I];
534 Cached = findBaseDefiningValue(I).BDV;
535 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
536 << Cached->getName() << "\n");
538 assert(Cache[I] != nullptr);
542 /// Return a base pointer for this value if known. Otherwise, return it's
543 /// base defining value.
544 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
545 Value *Def = findBaseDefiningValueCached(I, Cache);
546 auto Found = Cache.find(Def);
547 if (Found != Cache.end()) {
548 // Either a base-of relation, or a self reference. Caller must check.
549 return Found->second;
551 // Only a BDV available
555 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
556 /// is it known to be a base pointer? Or do we need to continue searching.
557 static bool isKnownBaseResult(Value *V) {
558 if (!isa<PHINode>(V) && !isa<SelectInst>(V) &&
559 !isa<ExtractElementInst>(V) && !isa<InsertElementInst>(V) &&
560 !isa<ShuffleVectorInst>(V)) {
561 // no recursion possible
564 if (isa<Instruction>(V) &&
565 cast<Instruction>(V)->getMetadata("is_base_value")) {
566 // This is a previously inserted base phi or select. We know
567 // that this is a base value.
571 // We need to keep searching
576 /// Models the state of a single base defining value in the findBasePointer
577 /// algorithm for determining where a new instruction is needed to propagate
578 /// the base of this BDV.
581 enum Status { Unknown, Base, Conflict };
583 BDVState(Status s, Value *b = nullptr) : status(s), base(b) {
584 assert(status != Base || b);
586 explicit BDVState(Value *b) : status(Base), base(b) {}
587 BDVState() : status(Unknown), base(nullptr) {}
589 Status getStatus() const { return status; }
590 Value *getBase() const { return base; }
592 bool isBase() const { return getStatus() == Base; }
593 bool isUnknown() const { return getStatus() == Unknown; }
594 bool isConflict() const { return getStatus() == Conflict; }
596 bool operator==(const BDVState &other) const {
597 return base == other.base && status == other.status;
600 bool operator!=(const BDVState &other) const { return !(*this == other); }
603 void dump() const { print(dbgs()); dbgs() << '\n'; }
605 void print(raw_ostream &OS) const {
617 OS << " (" << base << " - "
618 << (base ? base->getName() : "nullptr") << "): ";
623 Value *base; // non null only if status == base
628 static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
635 // Values of type BDVState form a lattice, and this is a helper
636 // class that implementes the meet operation. The meat of the meet
637 // operation is implemented in MeetBDVStates::pureMeet
638 class MeetBDVStates {
640 /// Initializes the currentResult to the TOP state so that if can be met with
641 /// any other state to produce that state.
644 // Destructively meet the current result with the given BDVState
645 void meetWith(BDVState otherState) {
646 currentResult = meet(otherState, currentResult);
649 BDVState getResult() const { return currentResult; }
652 BDVState currentResult;
654 /// Perform a meet operation on two elements of the BDVState lattice.
655 static BDVState meet(BDVState LHS, BDVState RHS) {
656 assert((pureMeet(LHS, RHS) == pureMeet(RHS, LHS)) &&
657 "math is wrong: meet does not commute!");
658 BDVState Result = pureMeet(LHS, RHS);
659 DEBUG(dbgs() << "meet of " << LHS << " with " << RHS
660 << " produced " << Result << "\n");
664 static BDVState pureMeet(const BDVState &stateA, const BDVState &stateB) {
665 switch (stateA.getStatus()) {
666 case BDVState::Unknown:
670 assert(stateA.getBase() && "can't be null");
671 if (stateB.isUnknown())
674 if (stateB.isBase()) {
675 if (stateA.getBase() == stateB.getBase()) {
676 assert(stateA == stateB && "equality broken!");
679 return BDVState(BDVState::Conflict);
681 assert(stateB.isConflict() && "only three states!");
682 return BDVState(BDVState::Conflict);
684 case BDVState::Conflict:
687 llvm_unreachable("only three states!");
693 /// For a given value or instruction, figure out what base ptr it's derived
694 /// from. For gc objects, this is simply itself. On success, returns a value
695 /// which is the base pointer. (This is reliable and can be used for
696 /// relocation.) On failure, returns nullptr.
697 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
698 Value *def = findBaseOrBDV(I, cache);
700 if (isKnownBaseResult(def)) {
704 // Here's the rough algorithm:
705 // - For every SSA value, construct a mapping to either an actual base
706 // pointer or a PHI which obscures the base pointer.
707 // - Construct a mapping from PHI to unknown TOP state. Use an
708 // optimistic algorithm to propagate base pointer information. Lattice
713 // When algorithm terminates, all PHIs will either have a single concrete
714 // base or be in a conflict state.
715 // - For every conflict, insert a dummy PHI node without arguments. Add
716 // these to the base[Instruction] = BasePtr mapping. For every
717 // non-conflict, add the actual base.
718 // - For every conflict, add arguments for the base[a] of each input
721 // Note: A simpler form of this would be to add the conflict form of all
722 // PHIs without running the optimistic algorithm. This would be
723 // analogous to pessimistic data flow and would likely lead to an
724 // overall worse solution.
727 auto isExpectedBDVType = [](Value *BDV) {
728 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) ||
729 isa<ExtractElementInst>(BDV) || isa<InsertElementInst>(BDV);
733 // Once populated, will contain a mapping from each potentially non-base BDV
734 // to a lattice value (described above) which corresponds to that BDV.
735 // We use the order of insertion (DFS over the def/use graph) to provide a
736 // stable deterministic ordering for visiting DenseMaps (which are unordered)
737 // below. This is important for deterministic compilation.
738 MapVector<Value *, BDVState> States;
740 // Recursively fill in all base defining values reachable from the initial
741 // one for which we don't already know a definite base value for
743 SmallVector<Value*, 16> Worklist;
744 Worklist.push_back(def);
745 States.insert(std::make_pair(def, BDVState()));
746 while (!Worklist.empty()) {
747 Value *Current = Worklist.pop_back_val();
748 assert(!isKnownBaseResult(Current) && "why did it get added?");
750 auto visitIncomingValue = [&](Value *InVal) {
751 Value *Base = findBaseOrBDV(InVal, cache);
752 if (isKnownBaseResult(Base))
753 // Known bases won't need new instructions introduced and can be
756 assert(isExpectedBDVType(Base) && "the only non-base values "
757 "we see should be base defining values");
758 if (States.insert(std::make_pair(Base, BDVState())).second)
759 Worklist.push_back(Base);
761 if (PHINode *Phi = dyn_cast<PHINode>(Current)) {
762 for (Value *InVal : Phi->incoming_values())
763 visitIncomingValue(InVal);
764 } else if (SelectInst *Sel = dyn_cast<SelectInst>(Current)) {
765 visitIncomingValue(Sel->getTrueValue());
766 visitIncomingValue(Sel->getFalseValue());
767 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
768 visitIncomingValue(EE->getVectorOperand());
769 } else if (auto *IE = dyn_cast<InsertElementInst>(Current)) {
770 visitIncomingValue(IE->getOperand(0)); // vector operand
771 visitIncomingValue(IE->getOperand(1)); // scalar operand
773 // There is one known class of instructions we know we don't handle.
774 assert(isa<ShuffleVectorInst>(Current));
775 llvm_unreachable("unimplemented instruction case");
781 DEBUG(dbgs() << "States after initialization:\n");
782 for (auto Pair : States) {
783 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
787 // Return a phi state for a base defining value. We'll generate a new
788 // base state for known bases and expect to find a cached state otherwise.
789 auto getStateForBDV = [&](Value *baseValue) {
790 if (isKnownBaseResult(baseValue))
791 return BDVState(baseValue);
792 auto I = States.find(baseValue);
793 assert(I != States.end() && "lookup failed!");
797 bool progress = true;
800 const size_t oldSize = States.size();
803 // We're only changing values in this loop, thus safe to keep iterators.
804 // Since this is computing a fixed point, the order of visit does not
805 // effect the result. TODO: We could use a worklist here and make this run
807 for (auto Pair : States) {
808 Value *BDV = Pair.first;
809 assert(!isKnownBaseResult(BDV) && "why did it get added?");
811 // Given an input value for the current instruction, return a BDVState
812 // instance which represents the BDV of that value.
813 auto getStateForInput = [&](Value *V) mutable {
814 Value *BDV = findBaseOrBDV(V, cache);
815 return getStateForBDV(BDV);
818 MeetBDVStates calculateMeet;
819 if (SelectInst *select = dyn_cast<SelectInst>(BDV)) {
820 calculateMeet.meetWith(getStateForInput(select->getTrueValue()));
821 calculateMeet.meetWith(getStateForInput(select->getFalseValue()));
822 } else if (PHINode *Phi = dyn_cast<PHINode>(BDV)) {
823 for (Value *Val : Phi->incoming_values())
824 calculateMeet.meetWith(getStateForInput(Val));
825 } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) {
826 // The 'meet' for an extractelement is slightly trivial, but it's still
827 // useful in that it drives us to conflict if our input is.
828 calculateMeet.meetWith(getStateForInput(EE->getVectorOperand()));
830 // Given there's a inherent type mismatch between the operands, will
831 // *always* produce Conflict.
832 auto *IE = cast<InsertElementInst>(BDV);
833 calculateMeet.meetWith(getStateForInput(IE->getOperand(0)));
834 calculateMeet.meetWith(getStateForInput(IE->getOperand(1)));
837 BDVState oldState = States[BDV];
838 BDVState newState = calculateMeet.getResult();
839 if (oldState != newState) {
841 States[BDV] = newState;
845 assert(oldSize == States.size() &&
846 "fixed point shouldn't be adding any new nodes to state");
850 DEBUG(dbgs() << "States after meet iteration:\n");
851 for (auto Pair : States) {
852 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
856 // Insert Phis for all conflicts
857 // TODO: adjust naming patterns to avoid this order of iteration dependency
858 for (auto Pair : States) {
859 Instruction *I = cast<Instruction>(Pair.first);
860 BDVState State = Pair.second;
861 assert(!isKnownBaseResult(I) && "why did it get added?");
862 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
864 // extractelement instructions are a bit special in that we may need to
865 // insert an extract even when we know an exact base for the instruction.
866 // The problem is that we need to convert from a vector base to a scalar
867 // base for the particular indice we're interested in.
868 if (State.isBase() && isa<ExtractElementInst>(I) &&
869 isa<VectorType>(State.getBase()->getType())) {
870 auto *EE = cast<ExtractElementInst>(I);
871 // TODO: In many cases, the new instruction is just EE itself. We should
872 // exploit this, but can't do it here since it would break the invariant
873 // about the BDV not being known to be a base.
874 auto *BaseInst = ExtractElementInst::Create(State.getBase(),
875 EE->getIndexOperand(),
877 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
878 States[I] = BDVState(BDVState::Base, BaseInst);
881 // Since we're joining a vector and scalar base, they can never be the
882 // same. As a result, we should always see insert element having reached
883 // the conflict state.
884 if (isa<InsertElementInst>(I)) {
885 assert(State.isConflict());
888 if (!State.isConflict())
891 /// Create and insert a new instruction which will represent the base of
892 /// the given instruction 'I'.
893 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
894 if (isa<PHINode>(I)) {
895 BasicBlock *BB = I->getParent();
896 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
897 assert(NumPreds > 0 && "how did we reach here");
898 std::string Name = suffixed_name_or(I, ".base", "base_phi");
899 return PHINode::Create(I->getType(), NumPreds, Name, I);
900 } else if (SelectInst *Sel = dyn_cast<SelectInst>(I)) {
901 // The undef will be replaced later
902 UndefValue *Undef = UndefValue::get(Sel->getType());
903 std::string Name = suffixed_name_or(I, ".base", "base_select");
904 return SelectInst::Create(Sel->getCondition(), Undef,
906 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
907 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
908 std::string Name = suffixed_name_or(I, ".base", "base_ee");
909 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
912 auto *IE = cast<InsertElementInst>(I);
913 UndefValue *VecUndef = UndefValue::get(IE->getOperand(0)->getType());
914 UndefValue *ScalarUndef = UndefValue::get(IE->getOperand(1)->getType());
915 std::string Name = suffixed_name_or(I, ".base", "base_ie");
916 return InsertElementInst::Create(VecUndef, ScalarUndef,
917 IE->getOperand(2), Name, IE);
921 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
922 // Add metadata marking this as a base value
923 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
924 States[I] = BDVState(BDVState::Conflict, BaseInst);
927 // Returns a instruction which produces the base pointer for a given
928 // instruction. The instruction is assumed to be an input to one of the BDVs
929 // seen in the inference algorithm above. As such, we must either already
930 // know it's base defining value is a base, or have inserted a new
931 // instruction to propagate the base of it's BDV and have entered that newly
932 // introduced instruction into the state table. In either case, we are
933 // assured to be able to determine an instruction which produces it's base
935 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
936 Value *BDV = findBaseOrBDV(Input, cache);
937 Value *Base = nullptr;
938 if (isKnownBaseResult(BDV)) {
941 // Either conflict or base.
942 assert(States.count(BDV));
943 Base = States[BDV].getBase();
945 assert(Base && "can't be null");
946 // The cast is needed since base traversal may strip away bitcasts
947 if (Base->getType() != Input->getType() &&
949 Base = new BitCastInst(Base, Input->getType(), "cast",
955 // Fixup all the inputs of the new PHIs. Visit order needs to be
956 // deterministic and predictable because we're naming newly created
958 for (auto Pair : States) {
959 Instruction *BDV = cast<Instruction>(Pair.first);
960 BDVState State = Pair.second;
962 assert(!isKnownBaseResult(BDV) && "why did it get added?");
963 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
964 if (!State.isConflict())
967 if (PHINode *basephi = dyn_cast<PHINode>(State.getBase())) {
968 PHINode *phi = cast<PHINode>(BDV);
969 unsigned NumPHIValues = phi->getNumIncomingValues();
970 for (unsigned i = 0; i < NumPHIValues; i++) {
971 Value *InVal = phi->getIncomingValue(i);
972 BasicBlock *InBB = phi->getIncomingBlock(i);
974 // If we've already seen InBB, add the same incoming value
975 // we added for it earlier. The IR verifier requires phi
976 // nodes with multiple entries from the same basic block
977 // to have the same incoming value for each of those
978 // entries. If we don't do this check here and basephi
979 // has a different type than base, we'll end up adding two
980 // bitcasts (and hence two distinct values) as incoming
981 // values for the same basic block.
983 int blockIndex = basephi->getBasicBlockIndex(InBB);
984 if (blockIndex != -1) {
985 Value *oldBase = basephi->getIncomingValue(blockIndex);
986 basephi->addIncoming(oldBase, InBB);
989 Value *Base = getBaseForInput(InVal, nullptr);
990 // In essence this assert states: the only way two
991 // values incoming from the same basic block may be
992 // different is by being different bitcasts of the same
993 // value. A cleanup that remains TODO is changing
994 // findBaseOrBDV to return an llvm::Value of the correct
995 // type (and still remain pure). This will remove the
996 // need to add bitcasts.
997 assert(Base->stripPointerCasts() == oldBase->stripPointerCasts() &&
998 "sanity -- findBaseOrBDV should be pure!");
1003 // Find the instruction which produces the base for each input. We may
1004 // need to insert a bitcast in the incoming block.
1005 // TODO: Need to split critical edges if insertion is needed
1006 Value *Base = getBaseForInput(InVal, InBB->getTerminator());
1007 basephi->addIncoming(Base, InBB);
1009 assert(basephi->getNumIncomingValues() == NumPHIValues);
1010 } else if (SelectInst *BaseSel = dyn_cast<SelectInst>(State.getBase())) {
1011 SelectInst *Sel = cast<SelectInst>(BDV);
1012 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
1013 // something more safe and less hacky.
1014 for (int i = 1; i <= 2; i++) {
1015 Value *InVal = Sel->getOperand(i);
1016 // Find the instruction which produces the base for each input. We may
1017 // need to insert a bitcast.
1018 Value *Base = getBaseForInput(InVal, BaseSel);
1019 BaseSel->setOperand(i, Base);
1021 } else if (auto *BaseEE = dyn_cast<ExtractElementInst>(State.getBase())) {
1022 Value *InVal = cast<ExtractElementInst>(BDV)->getVectorOperand();
1023 // Find the instruction which produces the base for each input. We may
1024 // need to insert a bitcast.
1025 Value *Base = getBaseForInput(InVal, BaseEE);
1026 BaseEE->setOperand(0, Base);
1028 auto *BaseIE = cast<InsertElementInst>(State.getBase());
1029 auto *BdvIE = cast<InsertElementInst>(BDV);
1030 auto UpdateOperand = [&](int OperandIdx) {
1031 Value *InVal = BdvIE->getOperand(OperandIdx);
1032 Value *Base = getBaseForInput(InVal, BaseIE);
1033 BaseIE->setOperand(OperandIdx, Base);
1035 UpdateOperand(0); // vector operand
1036 UpdateOperand(1); // scalar operand
1041 // Now that we're done with the algorithm, see if we can optimize the
1042 // results slightly by reducing the number of new instructions needed.
1043 // Arguably, this should be integrated into the algorithm above, but
1044 // doing as a post process step is easier to reason about for the moment.
1045 DenseMap<Value *, Value *> ReverseMap;
1046 SmallPtrSet<Instruction *, 16> NewInsts;
1047 SmallSetVector<AssertingVH<Instruction>, 16> Worklist;
1048 // Note: We need to visit the states in a deterministic order. We uses the
1049 // Keys we sorted above for this purpose. Note that we are papering over a
1050 // bigger problem with the algorithm above - it's visit order is not
1051 // deterministic. A larger change is needed to fix this.
1052 for (auto Pair : States) {
1053 auto *BDV = Pair.first;
1054 auto State = Pair.second;
1055 Value *Base = State.getBase();
1056 assert(BDV && Base);
1057 assert(!isKnownBaseResult(BDV) && "why did it get added?");
1058 assert(isKnownBaseResult(Base) &&
1059 "must be something we 'know' is a base pointer");
1060 if (!State.isConflict())
1063 ReverseMap[Base] = BDV;
1064 if (auto *BaseI = dyn_cast<Instruction>(Base)) {
1065 NewInsts.insert(BaseI);
1066 Worklist.insert(BaseI);
1069 auto ReplaceBaseInstWith = [&](Value *BDV, Instruction *BaseI,
1070 Value *Replacement) {
1071 // Add users which are new instructions (excluding self references)
1072 for (User *U : BaseI->users())
1073 if (auto *UI = dyn_cast<Instruction>(U))
1074 if (NewInsts.count(UI) && UI != BaseI)
1075 Worklist.insert(UI);
1076 // Then do the actual replacement
1077 NewInsts.erase(BaseI);
1078 ReverseMap.erase(BaseI);
1079 BaseI->replaceAllUsesWith(Replacement);
1080 BaseI->eraseFromParent();
1081 assert(States.count(BDV));
1082 assert(States[BDV].isConflict() && States[BDV].getBase() == BaseI);
1083 States[BDV] = BDVState(BDVState::Conflict, Replacement);
1085 const DataLayout &DL = cast<Instruction>(def)->getModule()->getDataLayout();
1086 while (!Worklist.empty()) {
1087 Instruction *BaseI = Worklist.pop_back_val();
1088 assert(NewInsts.count(BaseI));
1089 Value *Bdv = ReverseMap[BaseI];
1090 if (auto *BdvI = dyn_cast<Instruction>(Bdv))
1091 if (BaseI->isIdenticalTo(BdvI)) {
1092 DEBUG(dbgs() << "Identical Base: " << *BaseI << "\n");
1093 ReplaceBaseInstWith(Bdv, BaseI, Bdv);
1096 if (Value *V = SimplifyInstruction(BaseI, DL)) {
1097 DEBUG(dbgs() << "Base " << *BaseI << " simplified to " << *V << "\n");
1098 ReplaceBaseInstWith(Bdv, BaseI, V);
1103 // Cache all of our results so we can cheaply reuse them
1104 // NOTE: This is actually two caches: one of the base defining value
1105 // relation and one of the base pointer relation! FIXME
1106 for (auto Pair : States) {
1107 auto *BDV = Pair.first;
1108 Value *base = Pair.second.getBase();
1109 assert(BDV && base);
1111 std::string fromstr = cache.count(BDV) ? cache[BDV]->getName() : "none";
1112 DEBUG(dbgs() << "Updating base value cache"
1113 << " for: " << BDV->getName()
1114 << " from: " << fromstr
1115 << " to: " << base->getName() << "\n");
1117 if (cache.count(BDV)) {
1118 // Once we transition from the BDV relation being store in the cache to
1119 // the base relation being stored, it must be stable
1120 assert((!isKnownBaseResult(cache[BDV]) || cache[BDV] == base) &&
1121 "base relation should be stable");
1125 assert(cache.find(def) != cache.end());
1129 // For a set of live pointers (base and/or derived), identify the base
1130 // pointer of the object which they are derived from. This routine will
1131 // mutate the IR graph as needed to make the 'base' pointer live at the
1132 // definition site of 'derived'. This ensures that any use of 'derived' can
1133 // also use 'base'. This may involve the insertion of a number of
1134 // additional PHI nodes.
1136 // preconditions: live is a set of pointer type Values
1138 // side effects: may insert PHI nodes into the existing CFG, will preserve
1139 // CFG, will not remove or mutate any existing nodes
1141 // post condition: PointerToBase contains one (derived, base) pair for every
1142 // pointer in live. Note that derived can be equal to base if the original
1143 // pointer was a base pointer.
1145 findBasePointers(const StatepointLiveSetTy &live,
1146 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
1147 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1148 // For the naming of values inserted to be deterministic - which makes for
1149 // much cleaner and more stable tests - we need to assign an order to the
1150 // live values. DenseSets do not provide a deterministic order across runs.
1151 SmallVector<Value *, 64> Temp;
1152 Temp.insert(Temp.end(), live.begin(), live.end());
1153 std::sort(Temp.begin(), Temp.end(), order_by_name);
1154 for (Value *ptr : Temp) {
1155 Value *base = findBasePointer(ptr, DVCache);
1156 assert(base && "failed to find base pointer");
1157 PointerToBase[ptr] = base;
1158 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1159 DT->dominates(cast<Instruction>(base)->getParent(),
1160 cast<Instruction>(ptr)->getParent())) &&
1161 "The base we found better dominate the derived pointer");
1163 // If you see this trip and like to live really dangerously, the code should
1164 // be correct, just with idioms the verifier can't handle. You can try
1165 // disabling the verifier at your own substantial risk.
1166 assert(!isa<ConstantPointerNull>(base) &&
1167 "the relocation code needs adjustment to handle the relocation of "
1168 "a null pointer constant without causing false positives in the "
1169 "safepoint ir verifier.");
1173 /// Find the required based pointers (and adjust the live set) for the given
1175 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1177 PartiallyConstructedSafepointRecord &result) {
1178 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
1179 findBasePointers(result.liveset, PointerToBase, &DT, DVCache);
1181 if (PrintBasePointers) {
1182 // Note: Need to print these in a stable order since this is checked in
1184 errs() << "Base Pairs (w/o Relocation):\n";
1185 SmallVector<Value *, 64> Temp;
1186 Temp.reserve(PointerToBase.size());
1187 for (auto Pair : PointerToBase) {
1188 Temp.push_back(Pair.first);
1190 std::sort(Temp.begin(), Temp.end(), order_by_name);
1191 for (Value *Ptr : Temp) {
1192 Value *Base = PointerToBase[Ptr];
1193 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
1198 result.PointerToBase = PointerToBase;
1201 /// Given an updated version of the dataflow liveness results, update the
1202 /// liveset and base pointer maps for the call site CS.
1203 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1205 PartiallyConstructedSafepointRecord &result);
1207 static void recomputeLiveInValues(
1208 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1209 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1210 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1211 // again. The old values are still live and will help it stabilize quickly.
1212 GCPtrLivenessData RevisedLivenessData;
1213 computeLiveInValues(DT, F, RevisedLivenessData);
1214 for (size_t i = 0; i < records.size(); i++) {
1215 struct PartiallyConstructedSafepointRecord &info = records[i];
1216 const CallSite &CS = toUpdate[i];
1217 recomputeLiveInValues(RevisedLivenessData, CS, info);
1221 // When inserting gc.relocate calls, we need to ensure there are no uses
1222 // of the original value between the gc.statepoint and the gc.relocate call.
1223 // One case which can arise is a phi node starting one of the successor blocks.
1224 // We also need to be able to insert the gc.relocates only on the path which
1225 // goes through the statepoint. We might need to split an edge to make this
1228 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1229 DominatorTree &DT) {
1230 BasicBlock *Ret = BB;
1231 if (!BB->getUniquePredecessor()) {
1232 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1235 // Now that 'ret' has unique predecessor we can safely remove all phi nodes
1237 FoldSingleEntryPHINodes(Ret);
1238 assert(!isa<PHINode>(Ret->begin()));
1240 // At this point, we can safely insert a gc.relocate as the first instruction
1241 // in Ret if needed.
1245 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1246 auto itr = std::find(livevec.begin(), livevec.end(), val);
1247 assert(livevec.end() != itr);
1248 size_t index = std::distance(livevec.begin(), itr);
1249 assert(index < livevec.size());
1253 // Create new attribute set containing only attributes which can be transferred
1254 // from original call to the safepoint.
1255 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1258 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1259 unsigned index = AS.getSlotIndex(Slot);
1261 if (index == AttributeSet::ReturnIndex ||
1262 index == AttributeSet::FunctionIndex) {
1264 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1266 Attribute attr = *it;
1268 // Do not allow certain attributes - just skip them
1269 // Safepoint can not be read only or read none.
1270 if (attr.hasAttribute(Attribute::ReadNone) ||
1271 attr.hasAttribute(Attribute::ReadOnly))
1274 ret = ret.addAttributes(
1275 AS.getContext(), index,
1276 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1280 // Just skip parameter attributes for now
1286 /// Helper function to place all gc relocates necessary for the given
1289 /// liveVariables - list of variables to be relocated.
1290 /// liveStart - index of the first live variable.
1291 /// basePtrs - base pointers.
1292 /// statepointToken - statepoint instruction to which relocates should be
1294 /// Builder - Llvm IR builder to be used to construct new calls.
1295 static void CreateGCRelocates(ArrayRef<llvm::Value *> LiveVariables,
1296 const int LiveStart,
1297 ArrayRef<llvm::Value *> BasePtrs,
1298 Instruction *StatepointToken,
1299 IRBuilder<> Builder) {
1300 if (LiveVariables.empty())
1303 // All gc_relocate are set to i8 addrspace(1)* type. We originally generated
1304 // unique declarations for each pointer type, but this proved problematic
1305 // because the intrinsic mangling code is incomplete and fragile. Since
1306 // we're moving towards a single unified pointer type anyways, we can just
1307 // cast everything to an i8* of the right address space. A bitcast is added
1308 // later to convert gc_relocate to the actual value's type.
1309 Module *M = StatepointToken->getModule();
1310 auto AS = cast<PointerType>(LiveVariables[0]->getType())->getAddressSpace();
1311 Type *Types[] = {Type::getInt8PtrTy(M->getContext(), AS)};
1312 Value *GCRelocateDecl =
1313 Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, Types);
1315 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1316 // Generate the gc.relocate call and save the result
1318 Builder.getInt32(LiveStart + find_index(LiveVariables, BasePtrs[i]));
1320 Builder.getInt32(LiveStart + find_index(LiveVariables, LiveVariables[i]));
1322 // only specify a debug name if we can give a useful one
1323 CallInst *Reloc = Builder.CreateCall(
1324 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1325 suffixed_name_or(LiveVariables[i], ".relocated", ""));
1326 // Trick CodeGen into thinking there are lots of free registers at this
1328 Reloc->setCallingConv(CallingConv::Cold);
1333 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1334 const SmallVectorImpl<llvm::Value *> &basePtrs,
1335 const SmallVectorImpl<llvm::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 BasicBlock *BB = CS.getInstruction()->getParent();
1344 Function *F = BB->getParent();
1345 assert(F && "must be set");
1346 Module *M = F->getParent();
1348 assert(M && "must be set");
1350 // We're not changing the function signature of the statepoint since the gc
1351 // arguments go into the var args section.
1352 Function *gc_statepoint_decl = CS.getCalledFunction();
1354 // Then go ahead and use the builder do actually do the inserts. We insert
1355 // immediately before the previous instruction under the assumption that all
1356 // arguments will be available here. We can't insert afterwards since we may
1357 // be replacing a terminator.
1358 Instruction *insertBefore = CS.getInstruction();
1359 IRBuilder<> Builder(insertBefore);
1360 // Copy all of the arguments from the original statepoint - this includes the
1361 // target, call args, and deopt args
1362 SmallVector<llvm::Value *, 64> args;
1363 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1364 // TODO: Clear the 'needs rewrite' flag
1366 // add all the pointers to be relocated (gc arguments)
1367 // Capture the start of the live variable list for use in the gc_relocates
1368 const int live_start = args.size();
1369 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1371 // Create the statepoint given all the arguments
1372 Instruction *token = nullptr;
1373 AttributeSet return_attributes;
1375 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1377 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1378 call->setTailCall(toReplace->isTailCall());
1379 call->setCallingConv(toReplace->getCallingConv());
1381 // Currently we will fail on parameter attributes and on certain
1382 // function attributes.
1383 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1384 // In case if we can handle this set of attributes - set up function attrs
1385 // directly on statepoint and return attrs later for gc_result intrinsic.
1386 call->setAttributes(new_attrs.getFnAttributes());
1387 return_attributes = new_attrs.getRetAttributes();
1391 // Put the following gc_result and gc_relocate calls immediately after the
1392 // the old call (which we're about to delete)
1393 BasicBlock::iterator next(toReplace);
1394 assert(BB->end() != next && "not a terminator, must have next");
1396 Instruction *IP = &*(next);
1397 Builder.SetInsertPoint(IP);
1398 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1401 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1403 // Insert the new invoke into the old block. We'll remove the old one in a
1404 // moment at which point this will become the new terminator for the
1406 InvokeInst *invoke = InvokeInst::Create(
1407 gc_statepoint_decl, toReplace->getNormalDest(),
1408 toReplace->getUnwindDest(), args, "statepoint_token", toReplace->getParent());
1409 invoke->setCallingConv(toReplace->getCallingConv());
1411 // Currently we will fail on parameter attributes and on certain
1412 // function attributes.
1413 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1414 // In case if we can handle this set of attributes - set up function attrs
1415 // directly on statepoint and return attrs later for gc_result intrinsic.
1416 invoke->setAttributes(new_attrs.getFnAttributes());
1417 return_attributes = new_attrs.getRetAttributes();
1421 // Generate gc relocates in exceptional path
1422 BasicBlock *unwindBlock = toReplace->getUnwindDest();
1423 assert(!isa<PHINode>(unwindBlock->begin()) &&
1424 unwindBlock->getUniquePredecessor() &&
1425 "can't safely insert in this block!");
1427 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1428 Builder.SetInsertPoint(IP);
1429 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1431 // Extract second element from landingpad return value. We will attach
1432 // exceptional gc relocates to it.
1433 const unsigned idx = 1;
1434 Instruction *exceptional_token =
1435 cast<Instruction>(Builder.CreateExtractValue(
1436 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1437 result.UnwindToken = exceptional_token;
1439 CreateGCRelocates(liveVariables, live_start, basePtrs,
1440 exceptional_token, Builder);
1442 // Generate gc relocates and returns for normal block
1443 BasicBlock *normalDest = toReplace->getNormalDest();
1444 assert(!isa<PHINode>(normalDest->begin()) &&
1445 normalDest->getUniquePredecessor() &&
1446 "can't safely insert in this block!");
1448 IP = &*(normalDest->getFirstInsertionPt());
1449 Builder.SetInsertPoint(IP);
1451 // gc relocates will be generated later as if it were regular call
1456 // Take the name of the original value call if it had one.
1457 token->takeName(CS.getInstruction());
1459 // The GCResult is already inserted, we just need to find it
1461 Instruction *toReplace = CS.getInstruction();
1462 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1463 "only valid use before rewrite is gc.result");
1464 assert(!toReplace->hasOneUse() ||
1465 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1468 // Update the gc.result of the original statepoint (if any) to use the newly
1469 // inserted statepoint. This is safe to do here since the token can't be
1470 // considered a live reference.
1471 CS.getInstruction()->replaceAllUsesWith(token);
1473 result.StatepointToken = token;
1475 // Second, create a gc.relocate for every live variable
1476 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1480 struct name_ordering {
1483 bool operator()(name_ordering const &a, name_ordering const &b) {
1484 return -1 == a.derived->getName().compare(b.derived->getName());
1488 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1489 SmallVectorImpl<Value *> &livevec) {
1490 assert(basevec.size() == livevec.size());
1492 SmallVector<name_ordering, 64> temp;
1493 for (size_t i = 0; i < basevec.size(); i++) {
1495 v.base = basevec[i];
1496 v.derived = livevec[i];
1499 std::sort(temp.begin(), temp.end(), name_ordering());
1500 for (size_t i = 0; i < basevec.size(); i++) {
1501 basevec[i] = temp[i].base;
1502 livevec[i] = temp[i].derived;
1506 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1507 // which make the relocations happening at this safepoint explicit.
1509 // WARNING: Does not do any fixup to adjust users of the original live
1510 // values. That's the callers responsibility.
1512 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1513 PartiallyConstructedSafepointRecord &result) {
1514 auto liveset = result.liveset;
1515 auto PointerToBase = result.PointerToBase;
1517 // Convert to vector for efficient cross referencing.
1518 SmallVector<Value *, 64> basevec, livevec;
1519 livevec.reserve(liveset.size());
1520 basevec.reserve(liveset.size());
1521 for (Value *L : liveset) {
1522 livevec.push_back(L);
1523 assert(PointerToBase.count(L));
1524 Value *base = PointerToBase[L];
1525 basevec.push_back(base);
1527 assert(livevec.size() == basevec.size());
1529 // To make the output IR slightly more stable (for use in diffs), ensure a
1530 // fixed order of the values in the safepoint (by sorting the value name).
1531 // The order is otherwise meaningless.
1532 stablize_order(basevec, livevec);
1534 // Do the actual rewriting and delete the old statepoint
1535 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1536 CS.getInstruction()->eraseFromParent();
1539 // Helper function for the relocationViaAlloca.
1540 // It receives iterator to the statepoint gc relocates and emits store to the
1542 // location (via allocaMap) for the each one of them.
1543 // Add visited values into the visitedLiveValues set we will later use them
1544 // for sanity check.
1546 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1547 DenseMap<Value *, Value *> &AllocaMap,
1548 DenseSet<Value *> &VisitedLiveValues) {
1550 for (User *U : GCRelocs) {
1551 if (!isa<IntrinsicInst>(U))
1554 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1556 // We only care about relocates
1557 if (RelocatedValue->getIntrinsicID() !=
1558 Intrinsic::experimental_gc_relocate) {
1562 GCRelocateOperands RelocateOperands(RelocatedValue);
1563 Value *OriginalValue =
1564 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1565 assert(AllocaMap.count(OriginalValue));
1566 Value *Alloca = AllocaMap[OriginalValue];
1568 // Emit store into the related alloca
1569 // All gc_relocate are i8 addrspace(1)* typed, and it must be bitcasted to
1570 // the correct type according to alloca.
1571 assert(RelocatedValue->getNextNode() && "Should always have one since it's not a terminator");
1572 IRBuilder<> Builder(RelocatedValue->getNextNode());
1573 Value *CastedRelocatedValue =
1574 Builder.CreateBitCast(RelocatedValue,
1575 cast<AllocaInst>(Alloca)->getAllocatedType(),
1576 suffixed_name_or(RelocatedValue, ".casted", ""));
1578 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1579 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1582 VisitedLiveValues.insert(OriginalValue);
1587 // Helper function for the "relocationViaAlloca". Similar to the
1588 // "insertRelocationStores" but works for rematerialized values.
1590 insertRematerializationStores(
1591 RematerializedValueMapTy RematerializedValues,
1592 DenseMap<Value *, Value *> &AllocaMap,
1593 DenseSet<Value *> &VisitedLiveValues) {
1595 for (auto RematerializedValuePair: RematerializedValues) {
1596 Instruction *RematerializedValue = RematerializedValuePair.first;
1597 Value *OriginalValue = RematerializedValuePair.second;
1599 assert(AllocaMap.count(OriginalValue) &&
1600 "Can not find alloca for rematerialized value");
1601 Value *Alloca = AllocaMap[OriginalValue];
1603 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1604 Store->insertAfter(RematerializedValue);
1607 VisitedLiveValues.insert(OriginalValue);
1612 /// do all the relocation update via allocas and mem2reg
1613 static void relocationViaAlloca(
1614 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1615 ArrayRef<struct PartiallyConstructedSafepointRecord> Records) {
1617 // record initial number of (static) allocas; we'll check we have the same
1618 // number when we get done.
1619 int InitialAllocaNum = 0;
1620 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1622 if (isa<AllocaInst>(*I))
1626 // TODO-PERF: change data structures, reserve
1627 DenseMap<Value *, Value *> AllocaMap;
1628 SmallVector<AllocaInst *, 200> PromotableAllocas;
1629 // Used later to chack that we have enough allocas to store all values
1630 std::size_t NumRematerializedValues = 0;
1631 PromotableAllocas.reserve(Live.size());
1633 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1634 // "PromotableAllocas"
1635 auto emitAllocaFor = [&](Value *LiveValue) {
1636 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1637 F.getEntryBlock().getFirstNonPHI());
1638 AllocaMap[LiveValue] = Alloca;
1639 PromotableAllocas.push_back(Alloca);
1642 // emit alloca for each live gc pointer
1643 for (unsigned i = 0; i < Live.size(); i++) {
1644 emitAllocaFor(Live[i]);
1647 // emit allocas for rematerialized values
1648 for (size_t i = 0; i < Records.size(); i++) {
1649 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1651 for (auto RematerializedValuePair : Info.RematerializedValues) {
1652 Value *OriginalValue = RematerializedValuePair.second;
1653 if (AllocaMap.count(OriginalValue) != 0)
1656 emitAllocaFor(OriginalValue);
1657 ++NumRematerializedValues;
1661 // The next two loops are part of the same conceptual operation. We need to
1662 // insert a store to the alloca after the original def and at each
1663 // redefinition. We need to insert a load before each use. These are split
1664 // into distinct loops for performance reasons.
1666 // update gc pointer after each statepoint
1667 // either store a relocated value or null (if no relocated value found for
1668 // this gc pointer and it is not a gc_result)
1669 // this must happen before we update the statepoint with load of alloca
1670 // otherwise we lose the link between statepoint and old def
1671 for (size_t i = 0; i < Records.size(); i++) {
1672 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1673 Value *Statepoint = Info.StatepointToken;
1675 // This will be used for consistency check
1676 DenseSet<Value *> VisitedLiveValues;
1678 // Insert stores for normal statepoint gc relocates
1679 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1681 // In case if it was invoke statepoint
1682 // we will insert stores for exceptional path gc relocates.
1683 if (isa<InvokeInst>(Statepoint)) {
1684 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1688 // Do similar thing with rematerialized values
1689 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1692 if (ClobberNonLive) {
1693 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1694 // the gc.statepoint. This will turn some subtle GC problems into
1695 // slightly easier to debug SEGVs. Note that on large IR files with
1696 // lots of gc.statepoints this is extremely costly both memory and time
1698 SmallVector<AllocaInst *, 64> ToClobber;
1699 for (auto Pair : AllocaMap) {
1700 Value *Def = Pair.first;
1701 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1703 // This value was relocated
1704 if (VisitedLiveValues.count(Def)) {
1707 ToClobber.push_back(Alloca);
1710 auto InsertClobbersAt = [&](Instruction *IP) {
1711 for (auto *AI : ToClobber) {
1712 auto AIType = cast<PointerType>(AI->getType());
1713 auto PT = cast<PointerType>(AIType->getElementType());
1714 Constant *CPN = ConstantPointerNull::get(PT);
1715 StoreInst *Store = new StoreInst(CPN, AI);
1716 Store->insertBefore(IP);
1720 // Insert the clobbering stores. These may get intermixed with the
1721 // gc.results and gc.relocates, but that's fine.
1722 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1723 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1724 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1726 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1728 InsertClobbersAt(Next);
1732 // update use with load allocas and add store for gc_relocated
1733 for (auto Pair : AllocaMap) {
1734 Value *Def = Pair.first;
1735 Value *Alloca = Pair.second;
1737 // we pre-record the uses of allocas so that we dont have to worry about
1739 // that change the user information.
1740 SmallVector<Instruction *, 20> Uses;
1741 // PERF: trade a linear scan for repeated reallocation
1742 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1743 for (User *U : Def->users()) {
1744 if (!isa<ConstantExpr>(U)) {
1745 // If the def has a ConstantExpr use, then the def is either a
1746 // ConstantExpr use itself or null. In either case
1747 // (recursively in the first, directly in the second), the oop
1748 // it is ultimately dependent on is null and this particular
1749 // use does not need to be fixed up.
1750 Uses.push_back(cast<Instruction>(U));
1754 std::sort(Uses.begin(), Uses.end());
1755 auto Last = std::unique(Uses.begin(), Uses.end());
1756 Uses.erase(Last, Uses.end());
1758 for (Instruction *Use : Uses) {
1759 if (isa<PHINode>(Use)) {
1760 PHINode *Phi = cast<PHINode>(Use);
1761 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1762 if (Def == Phi->getIncomingValue(i)) {
1763 LoadInst *Load = new LoadInst(
1764 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1765 Phi->setIncomingValue(i, Load);
1769 LoadInst *Load = new LoadInst(Alloca, "", Use);
1770 Use->replaceUsesOfWith(Def, Load);
1774 // emit store for the initial gc value
1775 // store must be inserted after load, otherwise store will be in alloca's
1776 // use list and an extra load will be inserted before it
1777 StoreInst *Store = new StoreInst(Def, Alloca);
1778 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1779 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1780 // InvokeInst is a TerminatorInst so the store need to be inserted
1781 // into its normal destination block.
1782 BasicBlock *NormalDest = Invoke->getNormalDest();
1783 Store->insertBefore(NormalDest->getFirstNonPHI());
1785 assert(!Inst->isTerminator() &&
1786 "The only TerminatorInst that can produce a value is "
1787 "InvokeInst which is handled above.");
1788 Store->insertAfter(Inst);
1791 assert(isa<Argument>(Def));
1792 Store->insertAfter(cast<Instruction>(Alloca));
1796 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1797 "we must have the same allocas with lives");
1798 if (!PromotableAllocas.empty()) {
1799 // apply mem2reg to promote alloca to SSA
1800 PromoteMemToReg(PromotableAllocas, DT);
1804 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1806 if (isa<AllocaInst>(*I))
1808 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1812 /// Implement a unique function which doesn't require we sort the input
1813 /// vector. Doing so has the effect of changing the output of a couple of
1814 /// tests in ways which make them less useful in testing fused safepoints.
1815 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1816 SmallSet<T, 8> Seen;
1817 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1818 return !Seen.insert(V).second;
1822 /// Insert holders so that each Value is obviously live through the entire
1823 /// lifetime of the call.
1824 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1825 SmallVectorImpl<CallInst *> &Holders) {
1827 // No values to hold live, might as well not insert the empty holder
1830 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1831 // Use a dummy vararg function to actually hold the values live
1832 Function *Func = cast<Function>(M->getOrInsertFunction(
1833 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1835 // For call safepoints insert dummy calls right after safepoint
1836 BasicBlock::iterator Next(CS.getInstruction());
1838 Holders.push_back(CallInst::Create(Func, Values, "", Next));
1841 // For invoke safepooints insert dummy calls both in normal and
1842 // exceptional destination blocks
1843 auto *II = cast<InvokeInst>(CS.getInstruction());
1844 Holders.push_back(CallInst::Create(
1845 Func, Values, "", II->getNormalDest()->getFirstInsertionPt()));
1846 Holders.push_back(CallInst::Create(
1847 Func, Values, "", II->getUnwindDest()->getFirstInsertionPt()));
1850 static void findLiveReferences(
1851 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1852 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1853 GCPtrLivenessData OriginalLivenessData;
1854 computeLiveInValues(DT, F, OriginalLivenessData);
1855 for (size_t i = 0; i < records.size(); i++) {
1856 struct PartiallyConstructedSafepointRecord &info = records[i];
1857 const CallSite &CS = toUpdate[i];
1858 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1862 /// Remove any vector of pointers from the liveset by scalarizing them over the
1863 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1864 /// would be preferable to include the vector in the statepoint itself, but
1865 /// the lowering code currently does not handle that. Extending it would be
1866 /// slightly non-trivial since it requires a format change. Given how rare
1867 /// such cases are (for the moment?) scalarizing is an acceptable compromise.
1868 static void splitVectorValues(Instruction *StatepointInst,
1869 StatepointLiveSetTy &LiveSet,
1870 DenseMap<Value *, Value *>& PointerToBase,
1871 DominatorTree &DT) {
1872 SmallVector<Value *, 16> ToSplit;
1873 for (Value *V : LiveSet)
1874 if (isa<VectorType>(V->getType()))
1875 ToSplit.push_back(V);
1877 if (ToSplit.empty())
1880 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1882 Function &F = *(StatepointInst->getParent()->getParent());
1884 DenseMap<Value *, AllocaInst *> AllocaMap;
1885 // First is normal return, second is exceptional return (invoke only)
1886 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1887 for (Value *V : ToSplit) {
1888 AllocaInst *Alloca =
1889 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1890 AllocaMap[V] = Alloca;
1892 VectorType *VT = cast<VectorType>(V->getType());
1893 IRBuilder<> Builder(StatepointInst);
1894 SmallVector<Value *, 16> Elements;
1895 for (unsigned i = 0; i < VT->getNumElements(); i++)
1896 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1897 ElementMapping[V] = Elements;
1899 auto InsertVectorReform = [&](Instruction *IP) {
1900 Builder.SetInsertPoint(IP);
1901 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1902 Value *ResultVec = UndefValue::get(VT);
1903 for (unsigned i = 0; i < VT->getNumElements(); i++)
1904 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1905 Builder.getInt32(i));
1909 if (isa<CallInst>(StatepointInst)) {
1910 BasicBlock::iterator Next(StatepointInst);
1912 Instruction *IP = &*(Next);
1913 Replacements[V].first = InsertVectorReform(IP);
1914 Replacements[V].second = nullptr;
1916 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1917 // We've already normalized - check that we don't have shared destination
1919 BasicBlock *NormalDest = Invoke->getNormalDest();
1920 assert(!isa<PHINode>(NormalDest->begin()));
1921 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1922 assert(!isa<PHINode>(UnwindDest->begin()));
1923 // Insert insert element sequences in both successors
1924 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1925 Replacements[V].first = InsertVectorReform(IP);
1926 IP = &*(UnwindDest->getFirstInsertionPt());
1927 Replacements[V].second = InsertVectorReform(IP);
1931 for (Value *V : ToSplit) {
1932 AllocaInst *Alloca = AllocaMap[V];
1934 // Capture all users before we start mutating use lists
1935 SmallVector<Instruction *, 16> Users;
1936 for (User *U : V->users())
1937 Users.push_back(cast<Instruction>(U));
1939 for (Instruction *I : Users) {
1940 if (auto Phi = dyn_cast<PHINode>(I)) {
1941 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1942 if (V == Phi->getIncomingValue(i)) {
1943 LoadInst *Load = new LoadInst(
1944 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1945 Phi->setIncomingValue(i, Load);
1948 LoadInst *Load = new LoadInst(Alloca, "", I);
1949 I->replaceUsesOfWith(V, Load);
1953 // Store the original value and the replacement value into the alloca
1954 StoreInst *Store = new StoreInst(V, Alloca);
1955 if (auto I = dyn_cast<Instruction>(V))
1956 Store->insertAfter(I);
1958 Store->insertAfter(Alloca);
1960 // Normal return for invoke, or call return
1961 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1962 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1963 // Unwind return for invoke only
1964 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1966 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1969 // apply mem2reg to promote alloca to SSA
1970 SmallVector<AllocaInst *, 16> Allocas;
1971 for (Value *V : ToSplit)
1972 Allocas.push_back(AllocaMap[V]);
1973 PromoteMemToReg(Allocas, DT);
1975 // Update our tracking of live pointers and base mappings to account for the
1976 // changes we just made.
1977 for (Value *V : ToSplit) {
1978 auto &Elements = ElementMapping[V];
1981 LiveSet.insert(Elements.begin(), Elements.end());
1982 // We need to update the base mapping as well.
1983 assert(PointerToBase.count(V));
1984 Value *OldBase = PointerToBase[V];
1985 auto &BaseElements = ElementMapping[OldBase];
1986 PointerToBase.erase(V);
1987 assert(Elements.size() == BaseElements.size());
1988 for (unsigned i = 0; i < Elements.size(); i++) {
1989 Value *Elem = Elements[i];
1990 PointerToBase[Elem] = BaseElements[i];
1995 // Helper function for the "rematerializeLiveValues". It walks use chain
1996 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
1997 // values are visited (currently it is GEP's and casts). Returns true if it
1998 // successfully reached "BaseValue" and false otherwise.
1999 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
2001 static bool findRematerializableChainToBasePointer(
2002 SmallVectorImpl<Instruction*> &ChainToBase,
2003 Value *CurrentValue, Value *BaseValue) {
2005 // We have found a base value
2006 if (CurrentValue == BaseValue) {
2010 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
2011 ChainToBase.push_back(GEP);
2012 return findRematerializableChainToBasePointer(ChainToBase,
2013 GEP->getPointerOperand(),
2017 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
2018 Value *Def = CI->stripPointerCasts();
2020 // This two checks are basically similar. First one is here for the
2021 // consistency with findBasePointers logic.
2022 assert(!isa<CastInst>(Def) && "not a pointer cast found");
2023 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
2026 ChainToBase.push_back(CI);
2027 return findRematerializableChainToBasePointer(ChainToBase, Def, BaseValue);
2030 // Not supported instruction in the chain
2034 // Helper function for the "rematerializeLiveValues". Compute cost of the use
2035 // chain we are going to rematerialize.
2037 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
2038 TargetTransformInfo &TTI) {
2041 for (Instruction *Instr : Chain) {
2042 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
2043 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
2044 "non noop cast is found during rematerialization");
2046 Type *SrcTy = CI->getOperand(0)->getType();
2047 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
2049 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
2050 // Cost of the address calculation
2051 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
2052 Cost += TTI.getAddressComputationCost(ValTy);
2054 // And cost of the GEP itself
2055 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
2056 // allowed for the external usage)
2057 if (!GEP->hasAllConstantIndices())
2061 llvm_unreachable("unsupported instruciton type during rematerialization");
2068 // From the statepoint liveset pick values that are cheaper to recompute then to
2069 // relocate. Remove this values from the liveset, rematerialize them after
2070 // statepoint and record them in "Info" structure. Note that similar to
2071 // relocated values we don't do any user adjustments here.
2072 static void rematerializeLiveValues(CallSite CS,
2073 PartiallyConstructedSafepointRecord &Info,
2074 TargetTransformInfo &TTI) {
2075 const unsigned int ChainLengthThreshold = 10;
2077 // Record values we are going to delete from this statepoint live set.
2078 // We can not di this in following loop due to iterator invalidation.
2079 SmallVector<Value *, 32> LiveValuesToBeDeleted;
2081 for (Value *LiveValue: Info.liveset) {
2082 // For each live pointer find it's defining chain
2083 SmallVector<Instruction *, 3> ChainToBase;
2084 assert(Info.PointerToBase.count(LiveValue));
2086 findRematerializableChainToBasePointer(ChainToBase,
2088 Info.PointerToBase[LiveValue]);
2089 // Nothing to do, or chain is too long
2091 ChainToBase.size() == 0 ||
2092 ChainToBase.size() > ChainLengthThreshold)
2095 // Compute cost of this chain
2096 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
2097 // TODO: We can also account for cases when we will be able to remove some
2098 // of the rematerialized values by later optimization passes. I.e if
2099 // we rematerialized several intersecting chains. Or if original values
2100 // don't have any uses besides this statepoint.
2102 // For invokes we need to rematerialize each chain twice - for normal and
2103 // for unwind basic blocks. Model this by multiplying cost by two.
2104 if (CS.isInvoke()) {
2107 // If it's too expensive - skip it
2108 if (Cost >= RematerializationThreshold)
2111 // Remove value from the live set
2112 LiveValuesToBeDeleted.push_back(LiveValue);
2114 // Clone instructions and record them inside "Info" structure
2116 // Walk backwards to visit top-most instructions first
2117 std::reverse(ChainToBase.begin(), ChainToBase.end());
2119 // Utility function which clones all instructions from "ChainToBase"
2120 // and inserts them before "InsertBefore". Returns rematerialized value
2121 // which should be used after statepoint.
2122 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
2123 Instruction *LastClonedValue = nullptr;
2124 Instruction *LastValue = nullptr;
2125 for (Instruction *Instr: ChainToBase) {
2126 // Only GEP's and casts are suported as we need to be careful to not
2127 // introduce any new uses of pointers not in the liveset.
2128 // Note that it's fine to introduce new uses of pointers which were
2129 // otherwise not used after this statepoint.
2130 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2132 Instruction *ClonedValue = Instr->clone();
2133 ClonedValue->insertBefore(InsertBefore);
2134 ClonedValue->setName(Instr->getName() + ".remat");
2136 // If it is not first instruction in the chain then it uses previously
2137 // cloned value. We should update it to use cloned value.
2138 if (LastClonedValue) {
2140 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2142 // Assert that cloned instruction does not use any instructions from
2143 // this chain other than LastClonedValue
2144 for (auto OpValue : ClonedValue->operand_values()) {
2145 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2146 ChainToBase.end() &&
2147 "incorrect use in rematerialization chain");
2152 LastClonedValue = ClonedValue;
2155 assert(LastClonedValue);
2156 return LastClonedValue;
2159 // Different cases for calls and invokes. For invokes we need to clone
2160 // instructions both on normal and unwind path.
2162 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2163 assert(InsertBefore);
2164 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2165 Info.RematerializedValues[RematerializedValue] = LiveValue;
2167 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2169 Instruction *NormalInsertBefore =
2170 Invoke->getNormalDest()->getFirstInsertionPt();
2171 Instruction *UnwindInsertBefore =
2172 Invoke->getUnwindDest()->getFirstInsertionPt();
2174 Instruction *NormalRematerializedValue =
2175 rematerializeChain(NormalInsertBefore);
2176 Instruction *UnwindRematerializedValue =
2177 rematerializeChain(UnwindInsertBefore);
2179 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2180 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2184 // Remove rematerializaed values from the live set
2185 for (auto LiveValue: LiveValuesToBeDeleted) {
2186 Info.liveset.erase(LiveValue);
2190 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
2191 SmallVectorImpl<CallSite> &toUpdate) {
2193 // sanity check the input
2194 std::set<CallSite> uniqued;
2195 uniqued.insert(toUpdate.begin(), toUpdate.end());
2196 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
2198 for (size_t i = 0; i < toUpdate.size(); i++) {
2199 CallSite &CS = toUpdate[i];
2200 assert(CS.getInstruction()->getParent()->getParent() == &F);
2201 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
2205 // When inserting gc.relocates for invokes, we need to be able to insert at
2206 // the top of the successor blocks. See the comment on
2207 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2208 // may restructure the CFG.
2209 for (CallSite CS : toUpdate) {
2212 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
2213 normalizeForInvokeSafepoint(invoke->getNormalDest(), invoke->getParent(),
2215 normalizeForInvokeSafepoint(invoke->getUnwindDest(), invoke->getParent(),
2219 // A list of dummy calls added to the IR to keep various values obviously
2220 // live in the IR. We'll remove all of these when done.
2221 SmallVector<CallInst *, 64> holders;
2223 // Insert a dummy call with all of the arguments to the vm_state we'll need
2224 // for the actual safepoint insertion. This ensures reference arguments in
2225 // the deopt argument list are considered live through the safepoint (and
2226 // thus makes sure they get relocated.)
2227 for (size_t i = 0; i < toUpdate.size(); i++) {
2228 CallSite &CS = toUpdate[i];
2229 Statepoint StatepointCS(CS);
2231 SmallVector<Value *, 64> DeoptValues;
2232 for (Use &U : StatepointCS.vm_state_args()) {
2233 Value *Arg = cast<Value>(&U);
2234 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2235 "support for FCA unimplemented");
2236 if (isHandledGCPointerType(Arg->getType()))
2237 DeoptValues.push_back(Arg);
2239 insertUseHolderAfter(CS, DeoptValues, holders);
2242 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
2243 records.reserve(toUpdate.size());
2244 for (size_t i = 0; i < toUpdate.size(); i++) {
2245 struct PartiallyConstructedSafepointRecord info;
2246 records.push_back(info);
2248 assert(records.size() == toUpdate.size());
2250 // A) Identify all gc pointers which are statically live at the given call
2252 findLiveReferences(F, DT, P, toUpdate, records);
2254 // B) Find the base pointers for each live pointer
2255 /* scope for caching */ {
2256 // Cache the 'defining value' relation used in the computation and
2257 // insertion of base phis and selects. This ensures that we don't insert
2258 // large numbers of duplicate base_phis.
2259 DefiningValueMapTy DVCache;
2261 for (size_t i = 0; i < records.size(); i++) {
2262 struct PartiallyConstructedSafepointRecord &info = records[i];
2263 CallSite &CS = toUpdate[i];
2264 findBasePointers(DT, DVCache, CS, info);
2266 } // end of cache scope
2268 // The base phi insertion logic (for any safepoint) may have inserted new
2269 // instructions which are now live at some safepoint. The simplest such
2272 // phi a <-- will be a new base_phi here
2273 // safepoint 1 <-- that needs to be live here
2277 // We insert some dummy calls after each safepoint to definitely hold live
2278 // the base pointers which were identified for that safepoint. We'll then
2279 // ask liveness for _every_ base inserted to see what is now live. Then we
2280 // remove the dummy calls.
2281 holders.reserve(holders.size() + records.size());
2282 for (size_t i = 0; i < records.size(); i++) {
2283 struct PartiallyConstructedSafepointRecord &info = records[i];
2284 CallSite &CS = toUpdate[i];
2286 SmallVector<Value *, 128> Bases;
2287 for (auto Pair : info.PointerToBase) {
2288 Bases.push_back(Pair.second);
2290 insertUseHolderAfter(CS, Bases, holders);
2293 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2294 // need to rerun liveness. We may *also* have inserted new defs, but that's
2295 // not the key issue.
2296 recomputeLiveInValues(F, DT, P, toUpdate, records);
2298 if (PrintBasePointers) {
2299 for (size_t i = 0; i < records.size(); i++) {
2300 struct PartiallyConstructedSafepointRecord &info = records[i];
2301 errs() << "Base Pairs: (w/Relocation)\n";
2302 for (auto Pair : info.PointerToBase) {
2303 errs() << " derived %" << Pair.first->getName() << " base %"
2304 << Pair.second->getName() << "\n";
2308 for (size_t i = 0; i < holders.size(); i++) {
2309 holders[i]->eraseFromParent();
2310 holders[i] = nullptr;
2314 // Do a limited scalarization of any live at safepoint vector values which
2315 // contain pointers. This enables this pass to run after vectorization at
2316 // the cost of some possible performance loss. TODO: it would be nice to
2317 // natively support vectors all the way through the backend so we don't need
2318 // to scalarize here.
2319 for (size_t i = 0; i < records.size(); i++) {
2320 struct PartiallyConstructedSafepointRecord &info = records[i];
2321 Instruction *statepoint = toUpdate[i].getInstruction();
2322 splitVectorValues(cast<Instruction>(statepoint), info.liveset,
2323 info.PointerToBase, DT);
2326 // In order to reduce live set of statepoint we might choose to rematerialize
2327 // some values instead of relocating them. This is purely an optimization and
2328 // does not influence correctness.
2329 TargetTransformInfo &TTI =
2330 P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2332 for (size_t i = 0; i < records.size(); i++) {
2333 struct PartiallyConstructedSafepointRecord &info = records[i];
2334 CallSite &CS = toUpdate[i];
2336 rematerializeLiveValues(CS, info, TTI);
2339 // Now run through and replace the existing statepoints with new ones with
2340 // the live variables listed. We do not yet update uses of the values being
2341 // relocated. We have references to live variables that need to
2342 // survive to the last iteration of this loop. (By construction, the
2343 // previous statepoint can not be a live variable, thus we can and remove
2344 // the old statepoint calls as we go.)
2345 for (size_t i = 0; i < records.size(); i++) {
2346 struct PartiallyConstructedSafepointRecord &info = records[i];
2347 CallSite &CS = toUpdate[i];
2348 makeStatepointExplicit(DT, CS, P, info);
2350 toUpdate.clear(); // prevent accident use of invalid CallSites
2352 // Do all the fixups of the original live variables to their relocated selves
2353 SmallVector<Value *, 128> live;
2354 for (size_t i = 0; i < records.size(); i++) {
2355 struct PartiallyConstructedSafepointRecord &info = records[i];
2356 // We can't simply save the live set from the original insertion. One of
2357 // the live values might be the result of a call which needs a safepoint.
2358 // That Value* no longer exists and we need to use the new gc_result.
2359 // Thankfully, the liveset is embedded in the statepoint (and updated), so
2360 // we just grab that.
2361 Statepoint statepoint(info.StatepointToken);
2362 live.insert(live.end(), statepoint.gc_args_begin(),
2363 statepoint.gc_args_end());
2365 // Do some basic sanity checks on our liveness results before performing
2366 // relocation. Relocation can and will turn mistakes in liveness results
2367 // into non-sensical code which is must harder to debug.
2368 // TODO: It would be nice to test consistency as well
2369 assert(DT.isReachableFromEntry(info.StatepointToken->getParent()) &&
2370 "statepoint must be reachable or liveness is meaningless");
2371 for (Value *V : statepoint.gc_args()) {
2372 if (!isa<Instruction>(V))
2373 // Non-instruction values trivial dominate all possible uses
2375 auto LiveInst = cast<Instruction>(V);
2376 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2377 "unreachable values should never be live");
2378 assert(DT.dominates(LiveInst, info.StatepointToken) &&
2379 "basic SSA liveness expectation violated by liveness analysis");
2383 unique_unsorted(live);
2387 for (auto ptr : live) {
2388 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
2392 relocationViaAlloca(F, DT, live, records);
2393 return !records.empty();
2396 // Handles both return values and arguments for Functions and CallSites.
2397 template <typename AttrHolder>
2398 static void RemoveDerefAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2401 if (AH.getDereferenceableBytes(Index))
2402 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2403 AH.getDereferenceableBytes(Index)));
2404 if (AH.getDereferenceableOrNullBytes(Index))
2405 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2406 AH.getDereferenceableOrNullBytes(Index)));
2409 AH.setAttributes(AH.getAttributes().removeAttributes(
2410 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2414 RewriteStatepointsForGC::stripDereferenceabilityInfoFromPrototype(Function &F) {
2415 LLVMContext &Ctx = F.getContext();
2417 for (Argument &A : F.args())
2418 if (isa<PointerType>(A.getType()))
2419 RemoveDerefAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2421 if (isa<PointerType>(F.getReturnType()))
2422 RemoveDerefAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2425 void RewriteStatepointsForGC::stripDereferenceabilityInfoFromBody(Function &F) {
2429 LLVMContext &Ctx = F.getContext();
2430 MDBuilder Builder(Ctx);
2432 for (Instruction &I : instructions(F)) {
2433 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2434 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2435 bool IsImmutableTBAA =
2436 MD->getNumOperands() == 4 &&
2437 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2439 if (!IsImmutableTBAA)
2440 continue; // no work to do, MD_tbaa is already marked mutable
2442 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2443 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2445 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2447 MDNode *MutableTBAA =
2448 Builder.createTBAAStructTagNode(Base, Access, Offset);
2449 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2452 if (CallSite CS = CallSite(&I)) {
2453 for (int i = 0, e = CS.arg_size(); i != e; i++)
2454 if (isa<PointerType>(CS.getArgument(i)->getType()))
2455 RemoveDerefAttrAtIndex(Ctx, CS, i + 1);
2456 if (isa<PointerType>(CS.getType()))
2457 RemoveDerefAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2462 /// Returns true if this function should be rewritten by this pass. The main
2463 /// point of this function is as an extension point for custom logic.
2464 static bool shouldRewriteStatepointsIn(Function &F) {
2465 // TODO: This should check the GCStrategy
2467 const char *FunctionGCName = F.getGC();
2468 const StringRef StatepointExampleName("statepoint-example");
2469 const StringRef CoreCLRName("coreclr");
2470 return (StatepointExampleName == FunctionGCName) ||
2471 (CoreCLRName == FunctionGCName);
2476 void RewriteStatepointsForGC::stripDereferenceabilityInfo(Module &M) {
2478 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2482 for (Function &F : M)
2483 stripDereferenceabilityInfoFromPrototype(F);
2485 for (Function &F : M)
2486 stripDereferenceabilityInfoFromBody(F);
2489 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2490 // Nothing to do for declarations.
2491 if (F.isDeclaration() || F.empty())
2494 // Policy choice says not to rewrite - the most common reason is that we're
2495 // compiling code without a GCStrategy.
2496 if (!shouldRewriteStatepointsIn(F))
2499 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2501 // Gather all the statepoints which need rewritten. Be careful to only
2502 // consider those in reachable code since we need to ask dominance queries
2503 // when rewriting. We'll delete the unreachable ones in a moment.
2504 SmallVector<CallSite, 64> ParsePointNeeded;
2505 bool HasUnreachableStatepoint = false;
2506 for (Instruction &I : instructions(F)) {
2507 // TODO: only the ones with the flag set!
2508 if (isStatepoint(I)) {
2509 if (DT.isReachableFromEntry(I.getParent()))
2510 ParsePointNeeded.push_back(CallSite(&I));
2512 HasUnreachableStatepoint = true;
2516 bool MadeChange = false;
2518 // Delete any unreachable statepoints so that we don't have unrewritten
2519 // statepoints surviving this pass. This makes testing easier and the
2520 // resulting IR less confusing to human readers. Rather than be fancy, we
2521 // just reuse a utility function which removes the unreachable blocks.
2522 if (HasUnreachableStatepoint)
2523 MadeChange |= removeUnreachableBlocks(F);
2525 // Return early if no work to do.
2526 if (ParsePointNeeded.empty())
2529 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2530 // These are created by LCSSA. They have the effect of increasing the size
2531 // of liveness sets for no good reason. It may be harder to do this post
2532 // insertion since relocations and base phis can confuse things.
2533 for (BasicBlock &BB : F)
2534 if (BB.getUniquePredecessor()) {
2536 FoldSingleEntryPHINodes(&BB);
2539 // Before we start introducing relocations, we want to tweak the IR a bit to
2540 // avoid unfortunate code generation effects. The main example is that we
2541 // want to try to make sure the comparison feeding a branch is after any
2542 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2543 // values feeding a branch after relocation. This is semantically correct,
2544 // but results in extra register pressure since both the pre-relocation and
2545 // post-relocation copies must be available in registers. For code without
2546 // relocations this is handled elsewhere, but teaching the scheduler to
2547 // reverse the transform we're about to do would be slightly complex.
2548 // Note: This may extend the live range of the inputs to the icmp and thus
2549 // increase the liveset of any statepoint we move over. This is profitable
2550 // as long as all statepoints are in rare blocks. If we had in-register
2551 // lowering for live values this would be a much safer transform.
2552 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2553 if (auto *BI = dyn_cast<BranchInst>(TI))
2554 if (BI->isConditional())
2555 return dyn_cast<Instruction>(BI->getCondition());
2556 // TODO: Extend this to handle switches
2559 for (BasicBlock &BB : F) {
2560 TerminatorInst *TI = BB.getTerminator();
2561 if (auto *Cond = getConditionInst(TI))
2562 // TODO: Handle more than just ICmps here. We should be able to move
2563 // most instructions without side effects or memory access.
2564 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2566 Cond->moveBefore(TI);
2570 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
2574 // liveness computation via standard dataflow
2575 // -------------------------------------------------------------------
2577 // TODO: Consider using bitvectors for liveness, the set of potentially
2578 // interesting values should be small and easy to pre-compute.
2580 /// Compute the live-in set for the location rbegin starting from
2581 /// the live-out set of the basic block
2582 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2583 BasicBlock::reverse_iterator rend,
2584 DenseSet<Value *> &LiveTmp) {
2586 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2587 Instruction *I = &*ritr;
2589 // KILL/Def - Remove this definition from LiveIn
2592 // Don't consider *uses* in PHI nodes, we handle their contribution to
2593 // predecessor blocks when we seed the LiveOut sets
2594 if (isa<PHINode>(I))
2597 // USE - Add to the LiveIn set for this instruction
2598 for (Value *V : I->operands()) {
2599 assert(!isUnhandledGCPointerType(V->getType()) &&
2600 "support for FCA unimplemented");
2601 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2602 // The choice to exclude all things constant here is slightly subtle.
2603 // There are two independent reasons:
2604 // - We assume that things which are constant (from LLVM's definition)
2605 // do not move at runtime. For example, the address of a global
2606 // variable is fixed, even though it's contents may not be.
2607 // - Second, we can't disallow arbitrary inttoptr constants even
2608 // if the language frontend does. Optimization passes are free to
2609 // locally exploit facts without respect to global reachability. This
2610 // can create sections of code which are dynamically unreachable and
2611 // contain just about anything. (see constants.ll in tests)
2618 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2620 for (BasicBlock *Succ : successors(BB)) {
2621 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2622 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2623 PHINode *Phi = cast<PHINode>(&*I);
2624 Value *V = Phi->getIncomingValueForBlock(BB);
2625 assert(!isUnhandledGCPointerType(V->getType()) &&
2626 "support for FCA unimplemented");
2627 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2634 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2635 DenseSet<Value *> KillSet;
2636 for (Instruction &I : *BB)
2637 if (isHandledGCPointerType(I.getType()))
2643 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2644 /// sanity check for the liveness computation.
2645 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2646 TerminatorInst *TI, bool TermOkay = false) {
2647 for (Value *V : Live) {
2648 if (auto *I = dyn_cast<Instruction>(V)) {
2649 // The terminator can be a member of the LiveOut set. LLVM's definition
2650 // of instruction dominance states that V does not dominate itself. As
2651 // such, we need to special case this to allow it.
2652 if (TermOkay && TI == I)
2654 assert(DT.dominates(I, TI) &&
2655 "basic SSA liveness expectation violated by liveness analysis");
2660 /// Check that all the liveness sets used during the computation of liveness
2661 /// obey basic SSA properties. This is useful for finding cases where we miss
2663 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2665 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2666 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2667 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2671 static void computeLiveInValues(DominatorTree &DT, Function &F,
2672 GCPtrLivenessData &Data) {
2674 SmallSetVector<BasicBlock *, 200> Worklist;
2675 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2676 // We use a SetVector so that we don't have duplicates in the worklist.
2677 Worklist.insert(pred_begin(BB), pred_end(BB));
2679 auto NextItem = [&]() {
2680 BasicBlock *BB = Worklist.back();
2681 Worklist.pop_back();
2685 // Seed the liveness for each individual block
2686 for (BasicBlock &BB : F) {
2687 Data.KillSet[&BB] = computeKillSet(&BB);
2688 Data.LiveSet[&BB].clear();
2689 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2692 for (Value *Kill : Data.KillSet[&BB])
2693 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2696 Data.LiveOut[&BB] = DenseSet<Value *>();
2697 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2698 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2699 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2700 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2701 if (!Data.LiveIn[&BB].empty())
2702 AddPredsToWorklist(&BB);
2705 // Propagate that liveness until stable
2706 while (!Worklist.empty()) {
2707 BasicBlock *BB = NextItem();
2709 // Compute our new liveout set, then exit early if it hasn't changed
2710 // despite the contribution of our successor.
2711 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2712 const auto OldLiveOutSize = LiveOut.size();
2713 for (BasicBlock *Succ : successors(BB)) {
2714 assert(Data.LiveIn.count(Succ));
2715 set_union(LiveOut, Data.LiveIn[Succ]);
2717 // assert OutLiveOut is a subset of LiveOut
2718 if (OldLiveOutSize == LiveOut.size()) {
2719 // If the sets are the same size, then we didn't actually add anything
2720 // when unioning our successors LiveIn Thus, the LiveIn of this block
2724 Data.LiveOut[BB] = LiveOut;
2726 // Apply the effects of this basic block
2727 DenseSet<Value *> LiveTmp = LiveOut;
2728 set_union(LiveTmp, Data.LiveSet[BB]);
2729 set_subtract(LiveTmp, Data.KillSet[BB]);
2731 assert(Data.LiveIn.count(BB));
2732 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2733 // assert: OldLiveIn is a subset of LiveTmp
2734 if (OldLiveIn.size() != LiveTmp.size()) {
2735 Data.LiveIn[BB] = LiveTmp;
2736 AddPredsToWorklist(BB);
2738 } // while( !worklist.empty() )
2741 // Sanity check our output against SSA properties. This helps catch any
2742 // missing kills during the above iteration.
2743 for (BasicBlock &BB : F) {
2744 checkBasicSSA(DT, Data, BB);
2749 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2750 StatepointLiveSetTy &Out) {
2752 BasicBlock *BB = Inst->getParent();
2754 // Note: The copy is intentional and required
2755 assert(Data.LiveOut.count(BB));
2756 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2758 // We want to handle the statepoint itself oddly. It's
2759 // call result is not live (normal), nor are it's arguments
2760 // (unless they're used again later). This adjustment is
2761 // specifically what we need to relocate
2762 BasicBlock::reverse_iterator rend(Inst);
2763 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2764 LiveOut.erase(Inst);
2765 Out.insert(LiveOut.begin(), LiveOut.end());
2768 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2770 PartiallyConstructedSafepointRecord &Info) {
2771 Instruction *Inst = CS.getInstruction();
2772 StatepointLiveSetTy Updated;
2773 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2776 DenseSet<Value *> Bases;
2777 for (auto KVPair : Info.PointerToBase) {
2778 Bases.insert(KVPair.second);
2781 // We may have base pointers which are now live that weren't before. We need
2782 // to update the PointerToBase structure to reflect this.
2783 for (auto V : Updated)
2784 if (!Info.PointerToBase.count(V)) {
2785 assert(Bases.count(V) && "can't find base for unexpected live value");
2786 Info.PointerToBase[V] = V;
2791 for (auto V : Updated) {
2792 assert(Info.PointerToBase.count(V) &&
2793 "must be able to find base for live value");
2797 // Remove any stale base mappings - this can happen since our liveness is
2798 // more precise then the one inherent in the base pointer analysis
2799 DenseSet<Value *> ToErase;
2800 for (auto KVPair : Info.PointerToBase)
2801 if (!Updated.count(KVPair.first))
2802 ToErase.insert(KVPair.first);
2803 for (auto V : ToErase)
2804 Info.PointerToBase.erase(V);
2807 for (auto KVPair : Info.PointerToBase)
2808 assert(Updated.count(KVPair.first) && "record for non-live value");
2811 Info.liveset = Updated;