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/IR/BasicBlock.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/InstIterator.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/Intrinsics.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/IR/MDBuilder.h"
35 #include "llvm/IR/Statepoint.h"
36 #include "llvm/IR/Value.h"
37 #include "llvm/IR/Verifier.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Transforms/Scalar.h"
41 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
42 #include "llvm/Transforms/Utils/Cloning.h"
43 #include "llvm/Transforms/Utils/Local.h"
44 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
46 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
50 // Print the liveset found at the insert location
51 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
53 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden,
55 // Print out the base pointers for debugging
56 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden,
59 // Cost threshold measuring when it is profitable to rematerialize value instead
61 static cl::opt<unsigned>
62 RematerializationThreshold("spp-rematerialization-threshold", cl::Hidden,
66 static bool ClobberNonLive = true;
68 static bool ClobberNonLive = false;
70 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
71 cl::location(ClobberNonLive),
75 struct RewriteStatepointsForGC : public ModulePass {
76 static char ID; // Pass identification, replacement for typeid
78 RewriteStatepointsForGC() : ModulePass(ID) {
79 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
81 bool runOnFunction(Function &F);
82 bool runOnModule(Module &M) override {
85 Changed |= runOnFunction(F);
88 // stripDereferenceabilityInfo asserts that shouldRewriteStatepointsIn
89 // returns true for at least one function in the module. Since at least
90 // one function changed, we know that the precondition is satisfied.
91 stripDereferenceabilityInfo(M);
97 void getAnalysisUsage(AnalysisUsage &AU) const override {
98 // We add and rewrite a bunch of instructions, but don't really do much
99 // else. We could in theory preserve a lot more analyses here.
100 AU.addRequired<DominatorTreeWrapperPass>();
101 AU.addRequired<TargetTransformInfoWrapperPass>();
104 /// The IR fed into RewriteStatepointsForGC may have had attributes implying
105 /// dereferenceability that are no longer valid/correct after
106 /// RewriteStatepointsForGC has run. This is because semantically, after
107 /// RewriteStatepointsForGC runs, all calls to gc.statepoint "free" the entire
108 /// heap. stripDereferenceabilityInfo (conservatively) restores correctness
109 /// by erasing all attributes in the module that externally imply
110 /// dereferenceability.
112 void stripDereferenceabilityInfo(Module &M);
114 // Helpers for stripDereferenceabilityInfo
115 void stripDereferenceabilityInfoFromBody(Function &F);
116 void stripDereferenceabilityInfoFromPrototype(Function &F);
120 char RewriteStatepointsForGC::ID = 0;
122 ModulePass *llvm::createRewriteStatepointsForGCPass() {
123 return new RewriteStatepointsForGC();
126 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
127 "Make relocations explicit at statepoints", false, false)
128 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
129 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
130 "Make relocations explicit at statepoints", false, false)
133 struct GCPtrLivenessData {
134 /// Values defined in this block.
135 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet;
136 /// Values used in this block (and thus live); does not included values
137 /// killed within this block.
138 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet;
140 /// Values live into this basic block (i.e. used by any
141 /// instruction in this basic block or ones reachable from here)
142 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn;
144 /// Values live out of this basic block (i.e. live into
145 /// any successor block)
146 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut;
149 // The type of the internal cache used inside the findBasePointers family
150 // of functions. From the callers perspective, this is an opaque type and
151 // should not be inspected.
153 // In the actual implementation this caches two relations:
154 // - The base relation itself (i.e. this pointer is based on that one)
155 // - The base defining value relation (i.e. before base_phi insertion)
156 // Generally, after the execution of a full findBasePointer call, only the
157 // base relation will remain. Internally, we add a mixture of the two
158 // types, then update all the second type to the first type
159 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
160 typedef DenseSet<llvm::Value *> StatepointLiveSetTy;
161 typedef DenseMap<Instruction *, Value *> RematerializedValueMapTy;
163 struct PartiallyConstructedSafepointRecord {
164 /// The set of values known to be live across this safepoint
165 StatepointLiveSetTy liveset;
167 /// Mapping from live pointers to a base-defining-value
168 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
170 /// The *new* gc.statepoint instruction itself. This produces the token
171 /// that normal path gc.relocates and the gc.result are tied to.
172 Instruction *StatepointToken;
174 /// Instruction to which exceptional gc relocates are attached
175 /// Makes it easier to iterate through them during relocationViaAlloca.
176 Instruction *UnwindToken;
178 /// Record live values we are rematerialized instead of relocating.
179 /// They are not included into 'liveset' field.
180 /// Maps rematerialized copy to it's original value.
181 RematerializedValueMapTy RematerializedValues;
185 /// Compute the live-in set for every basic block in the function
186 static void computeLiveInValues(DominatorTree &DT, Function &F,
187 GCPtrLivenessData &Data);
189 /// Given results from the dataflow liveness computation, find the set of live
190 /// Values at a particular instruction.
191 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
192 StatepointLiveSetTy &out);
194 // TODO: Once we can get to the GCStrategy, this becomes
195 // Optional<bool> isGCManagedPointer(const Value *V) const override {
197 static bool isGCPointerType(Type *T) {
198 if (auto *PT = dyn_cast<PointerType>(T))
199 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
200 // GC managed heap. We know that a pointer into this heap needs to be
201 // updated and that no other pointer does.
202 return (1 == PT->getAddressSpace());
206 // Return true if this type is one which a) is a gc pointer or contains a GC
207 // pointer and b) is of a type this code expects to encounter as a live value.
208 // (The insertion code will assert that a type which matches (a) and not (b)
209 // is not encountered.)
210 static bool isHandledGCPointerType(Type *T) {
211 // We fully support gc pointers
212 if (isGCPointerType(T))
214 // We partially support vectors of gc pointers. The code will assert if it
215 // can't handle something.
216 if (auto VT = dyn_cast<VectorType>(T))
217 if (isGCPointerType(VT->getElementType()))
223 /// Returns true if this type contains a gc pointer whether we know how to
224 /// handle that type or not.
225 static bool containsGCPtrType(Type *Ty) {
226 if (isGCPointerType(Ty))
228 if (VectorType *VT = dyn_cast<VectorType>(Ty))
229 return isGCPointerType(VT->getScalarType());
230 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
231 return containsGCPtrType(AT->getElementType());
232 if (StructType *ST = dyn_cast<StructType>(Ty))
234 ST->subtypes().begin(), ST->subtypes().end(),
235 [](Type *SubType) { return containsGCPtrType(SubType); });
239 // Returns true if this is a type which a) is a gc pointer or contains a GC
240 // pointer and b) is of a type which the code doesn't expect (i.e. first class
241 // aggregates). Used to trip assertions.
242 static bool isUnhandledGCPointerType(Type *Ty) {
243 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
247 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
248 if (a->hasName() && b->hasName()) {
249 return -1 == a->getName().compare(b->getName());
250 } else if (a->hasName() && !b->hasName()) {
252 } else if (!a->hasName() && b->hasName()) {
255 // Better than nothing, but not stable
260 // Conservatively identifies any definitions which might be live at the
261 // given instruction. The analysis is performed immediately before the
262 // given instruction. Values defined by that instruction are not considered
263 // live. Values used by that instruction are considered live.
264 static void analyzeParsePointLiveness(
265 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
266 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
267 Instruction *inst = CS.getInstruction();
269 StatepointLiveSetTy liveset;
270 findLiveSetAtInst(inst, OriginalLivenessData, liveset);
273 // Note: This output is used by several of the test cases
274 // The order of elements in a set is not stable, put them in a vec and sort
276 SmallVector<Value *, 64> Temp;
277 Temp.insert(Temp.end(), liveset.begin(), liveset.end());
278 std::sort(Temp.begin(), Temp.end(), order_by_name);
279 errs() << "Live Variables:\n";
280 for (Value *V : Temp)
281 dbgs() << " " << V->getName() << " " << *V << "\n";
283 if (PrintLiveSetSize) {
284 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
285 errs() << "Number live values: " << liveset.size() << "\n";
287 result.liveset = liveset;
290 static Value *findBaseDefiningValue(Value *I);
292 /// Return a base defining value for the 'Index' element of the given vector
293 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
294 /// 'I'. As an optimization, this method will try to determine when the
295 /// element is known to already be a base pointer. If this can be established,
296 /// the second value in the returned pair will be true. Note that either a
297 /// vector or a pointer typed value can be returned. For the former, the
298 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
299 /// If the later, the return pointer is a BDV (or possibly a base) for the
300 /// particular element in 'I'.
301 static std::pair<Value *, bool>
302 findBaseDefiningValueOfVector(Value *I, Value *Index = nullptr) {
303 assert(I->getType()->isVectorTy() &&
304 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
305 "Illegal to ask for the base pointer of a non-pointer type");
307 // Each case parallels findBaseDefiningValue below, see that code for
308 // detailed motivation.
310 if (isa<Argument>(I))
311 // An incoming argument to the function is a base pointer
312 return std::make_pair(I, true);
314 // We shouldn't see the address of a global as a vector value?
315 assert(!isa<GlobalVariable>(I) &&
316 "unexpected global variable found in base of vector");
318 // inlining could possibly introduce phi node that contains
319 // undef if callee has multiple returns
320 if (isa<UndefValue>(I))
321 // utterly meaningless, but useful for dealing with partially optimized
323 return std::make_pair(I, true);
325 // Due to inheritance, this must be _after_ the global variable and undef
327 if (Constant *Con = dyn_cast<Constant>(I)) {
328 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
329 "order of checks wrong!");
330 assert(Con->isNullValue() && "null is the only case which makes sense");
331 return std::make_pair(Con, true);
334 if (isa<LoadInst>(I))
335 return std::make_pair(I, true);
337 // For an insert element, we might be able to look through it if we know
338 // something about the indexes.
339 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(I)) {
341 Value *InsertIndex = IEI->getOperand(2);
342 // This index is inserting the value, look for its BDV
343 if (InsertIndex == Index)
344 return std::make_pair(findBaseDefiningValue(IEI->getOperand(1)), false);
345 // Both constant, and can't be equal per above. This insert is definitely
346 // not relevant, look back at the rest of the vector and keep trying.
347 if (isa<ConstantInt>(Index) && isa<ConstantInt>(InsertIndex))
348 return findBaseDefiningValueOfVector(IEI->getOperand(0), Index);
351 // We don't know whether this vector contains entirely base pointers or
352 // not. To be conservatively correct, we treat it as a BDV and will
353 // duplicate code as needed to construct a parallel vector of bases.
354 return std::make_pair(IEI, false);
357 if (isa<ShuffleVectorInst>(I))
358 // We don't know whether this vector contains entirely base pointers or
359 // not. To be conservatively correct, we treat it as a BDV and will
360 // duplicate code as needed to construct a parallel vector of bases.
361 // TODO: There a number of local optimizations which could be applied here
362 // for particular sufflevector patterns.
363 return std::make_pair(I, false);
365 // A PHI or Select is a base defining value. The outer findBasePointer
366 // algorithm is responsible for constructing a base value for this BDV.
367 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
368 "unknown vector instruction - no base found for vector element");
369 return std::make_pair(I, false);
372 static bool isKnownBaseResult(Value *V);
374 /// Helper function for findBasePointer - Will return a value which either a)
375 /// defines the base pointer for the input, b) blocks the simple search
376 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
377 /// from pointer to vector type or back.
378 static Value *findBaseDefiningValue(Value *I) {
379 if (I->getType()->isVectorTy())
380 return findBaseDefiningValueOfVector(I).first;
382 assert(I->getType()->isPointerTy() &&
383 "Illegal to ask for the base pointer of a non-pointer type");
385 if (isa<Argument>(I))
386 // An incoming argument to the function is a base pointer
387 // We should have never reached here if this argument isn't an gc value
390 if (isa<GlobalVariable>(I))
394 // inlining could possibly introduce phi node that contains
395 // undef if callee has multiple returns
396 if (isa<UndefValue>(I))
397 // utterly meaningless, but useful for dealing with
398 // partially optimized code.
401 // Due to inheritance, this must be _after_ the global variable and undef
403 if (Constant *Con = dyn_cast<Constant>(I)) {
404 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
405 "order of checks wrong!");
406 // Note: Finding a constant base for something marked for relocation
407 // doesn't really make sense. The most likely case is either a) some
408 // screwed up the address space usage or b) your validating against
409 // compiled C++ code w/o the proper separation. The only real exception
410 // is a null pointer. You could have generic code written to index of
411 // off a potentially null value and have proven it null. We also use
412 // null pointers in dead paths of relocation phis (which we might later
413 // want to find a base pointer for).
414 assert(isa<ConstantPointerNull>(Con) &&
415 "null is the only case which makes sense");
419 if (CastInst *CI = dyn_cast<CastInst>(I)) {
420 Value *Def = CI->stripPointerCasts();
421 // If we find a cast instruction here, it means we've found a cast which is
422 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
423 // handle int->ptr conversion.
424 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
425 return findBaseDefiningValue(Def);
428 if (isa<LoadInst>(I))
429 return I; // The value loaded is an gc base itself
431 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
432 // The base of this GEP is the base
433 return findBaseDefiningValue(GEP->getPointerOperand());
435 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
436 switch (II->getIntrinsicID()) {
437 case Intrinsic::experimental_gc_result_ptr:
439 // fall through to general call handling
441 case Intrinsic::experimental_gc_statepoint:
442 case Intrinsic::experimental_gc_result_float:
443 case Intrinsic::experimental_gc_result_int:
444 llvm_unreachable("these don't produce pointers");
445 case Intrinsic::experimental_gc_relocate: {
446 // Rerunning safepoint insertion after safepoints are already
447 // inserted is not supported. It could probably be made to work,
448 // but why are you doing this? There's no good reason.
449 llvm_unreachable("repeat safepoint insertion is not supported");
451 case Intrinsic::gcroot:
452 // Currently, this mechanism hasn't been extended to work with gcroot.
453 // There's no reason it couldn't be, but I haven't thought about the
454 // implications much.
456 "interaction with the gcroot mechanism is not supported");
459 // We assume that functions in the source language only return base
460 // pointers. This should probably be generalized via attributes to support
461 // both source language and internal functions.
462 if (isa<CallInst>(I) || isa<InvokeInst>(I))
465 // I have absolutely no idea how to implement this part yet. It's not
466 // necessarily hard, I just haven't really looked at it yet.
467 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
469 if (isa<AtomicCmpXchgInst>(I))
470 // A CAS is effectively a atomic store and load combined under a
471 // predicate. From the perspective of base pointers, we just treat it
475 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
476 "binary ops which don't apply to pointers");
478 // The aggregate ops. Aggregates can either be in the heap or on the
479 // stack, but in either case, this is simply a field load. As a result,
480 // this is a defining definition of the base just like a load is.
481 if (isa<ExtractValueInst>(I))
484 // We should never see an insert vector since that would require we be
485 // tracing back a struct value not a pointer value.
486 assert(!isa<InsertValueInst>(I) &&
487 "Base pointer for a struct is meaningless");
489 // An extractelement produces a base result exactly when it's input does.
490 // We may need to insert a parallel instruction to extract the appropriate
491 // element out of the base vector corresponding to the input. Given this,
492 // it's analogous to the phi and select case even though it's not a merge.
493 if (auto *EEI = dyn_cast<ExtractElementInst>(I)) {
494 Value *VectorOperand = EEI->getVectorOperand();
495 Value *Index = EEI->getIndexOperand();
496 std::pair<Value *, bool> pair =
497 findBaseDefiningValueOfVector(VectorOperand, Index);
498 Value *VectorBase = pair.first;
499 if (VectorBase->getType()->isPointerTy())
500 // We found a BDV for this specific element with the vector. This is an
501 // optimization, but in practice it covers most of the useful cases
502 // created via scalarization. Note: The peephole optimization here is
503 // currently needed for correctness since the general algorithm doesn't
504 // yet handle insertelements. That will change shortly.
507 assert(VectorBase->getType()->isVectorTy());
508 // Otherwise, we have an instruction which potentially produces a
509 // derived pointer and we need findBasePointers to clone code for us
510 // such that we can create an instruction which produces the
511 // accompanying base pointer.
516 // The last two cases here don't return a base pointer. Instead, they
517 // return a value which dynamically selects from among several base
518 // derived pointers (each with it's own base potentially). It's the job of
519 // the caller to resolve these.
520 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
521 "missing instruction case in findBaseDefiningValing");
525 /// Returns the base defining value for this value.
526 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
527 Value *&Cached = Cache[I];
529 Cached = findBaseDefiningValue(I);
530 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
531 << Cached->getName() << "\n");
533 assert(Cache[I] != nullptr);
537 /// Return a base pointer for this value if known. Otherwise, return it's
538 /// base defining value.
539 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
540 Value *Def = findBaseDefiningValueCached(I, Cache);
541 auto Found = Cache.find(Def);
542 if (Found != Cache.end()) {
543 // Either a base-of relation, or a self reference. Caller must check.
544 return Found->second;
546 // Only a BDV available
550 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
551 /// is it known to be a base pointer? Or do we need to continue searching.
552 static bool isKnownBaseResult(Value *V) {
553 if (!isa<PHINode>(V) && !isa<SelectInst>(V) && !isa<ExtractElementInst>(V)) {
554 // no recursion possible
557 if (isa<Instruction>(V) &&
558 cast<Instruction>(V)->getMetadata("is_base_value")) {
559 // This is a previously inserted base phi or select. We know
560 // that this is a base value.
564 // We need to keep searching
569 /// Models the state of a single base defining value in the findBasePointer
570 /// algorithm for determining where a new instruction is needed to propagate
571 /// the base of this BDV.
574 enum Status { Unknown, Base, Conflict };
576 BDVState(Status s, Value *b = nullptr) : status(s), base(b) {
577 assert(status != Base || b);
579 explicit BDVState(Value *b) : status(Base), base(b) {}
580 BDVState() : status(Unknown), base(nullptr) {}
582 Status getStatus() const { return status; }
583 Value *getBase() const { return base; }
585 bool isBase() const { return getStatus() == Base; }
586 bool isUnknown() const { return getStatus() == Unknown; }
587 bool isConflict() const { return getStatus() == Conflict; }
589 bool operator==(const BDVState &other) const {
590 return base == other.base && status == other.status;
593 bool operator!=(const BDVState &other) const { return !(*this == other); }
596 void dump() const { print(dbgs()); dbgs() << '\n'; }
598 void print(raw_ostream &OS) const {
610 OS << " (" << base << " - "
611 << (base ? base->getName() : "nullptr") << "): ";
616 Value *base; // non null only if status == base
620 inline raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
626 typedef DenseMap<Value *, BDVState> ConflictStateMapTy;
627 // Values of type BDVState form a lattice, and this is a helper
628 // class that implementes the meet operation. The meat of the meet
629 // operation is implemented in MeetBDVStates::pureMeet
630 class MeetBDVStates {
632 /// Initializes the currentResult to the TOP state so that if can be met with
633 /// any other state to produce that state.
636 // Destructively meet the current result with the given BDVState
637 void meetWith(BDVState otherState) {
638 currentResult = meet(otherState, currentResult);
641 BDVState getResult() const { return currentResult; }
644 BDVState currentResult;
646 /// Perform a meet operation on two elements of the BDVState lattice.
647 static BDVState meet(BDVState LHS, BDVState RHS) {
648 assert((pureMeet(LHS, RHS) == pureMeet(RHS, LHS)) &&
649 "math is wrong: meet does not commute!");
650 BDVState Result = pureMeet(LHS, RHS);
651 DEBUG(dbgs() << "meet of " << LHS << " with " << RHS
652 << " produced " << Result << "\n");
656 static BDVState pureMeet(const BDVState &stateA, const BDVState &stateB) {
657 switch (stateA.getStatus()) {
658 case BDVState::Unknown:
662 assert(stateA.getBase() && "can't be null");
663 if (stateB.isUnknown())
666 if (stateB.isBase()) {
667 if (stateA.getBase() == stateB.getBase()) {
668 assert(stateA == stateB && "equality broken!");
671 return BDVState(BDVState::Conflict);
673 assert(stateB.isConflict() && "only three states!");
674 return BDVState(BDVState::Conflict);
676 case BDVState::Conflict:
679 llvm_unreachable("only three states!");
683 /// For a given value or instruction, figure out what base ptr it's derived
684 /// from. For gc objects, this is simply itself. On success, returns a value
685 /// which is the base pointer. (This is reliable and can be used for
686 /// relocation.) On failure, returns nullptr.
687 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
688 Value *def = findBaseOrBDV(I, cache);
690 if (isKnownBaseResult(def)) {
694 // Here's the rough algorithm:
695 // - For every SSA value, construct a mapping to either an actual base
696 // pointer or a PHI which obscures the base pointer.
697 // - Construct a mapping from PHI to unknown TOP state. Use an
698 // optimistic algorithm to propagate base pointer information. Lattice
703 // When algorithm terminates, all PHIs will either have a single concrete
704 // base or be in a conflict state.
705 // - For every conflict, insert a dummy PHI node without arguments. Add
706 // these to the base[Instruction] = BasePtr mapping. For every
707 // non-conflict, add the actual base.
708 // - For every conflict, add arguments for the base[a] of each input
711 // Note: A simpler form of this would be to add the conflict form of all
712 // PHIs without running the optimistic algorithm. This would be
713 // analogous to pessimistic data flow and would likely lead to an
714 // overall worse solution.
717 auto isExpectedBDVType = [](Value *BDV) {
718 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) || isa<ExtractElementInst>(BDV);
722 // Once populated, will contain a mapping from each potentially non-base BDV
723 // to a lattice value (described above) which corresponds to that BDV.
724 ConflictStateMapTy states;
725 // Recursively fill in all phis & selects reachable from the initial one
726 // for which we don't already know a definite base value for
728 DenseSet<Value *> Visited;
729 SmallVector<Value*, 16> Worklist;
730 Worklist.push_back(def);
732 while (!Worklist.empty()) {
733 Value *Current = Worklist.pop_back_val();
734 assert(!isKnownBaseResult(Current) && "why did it get added?");
736 auto visitIncomingValue = [&](Value *InVal) {
737 Value *Base = findBaseOrBDV(InVal, cache);
738 if (isKnownBaseResult(Base))
739 // Known bases won't need new instructions introduced and can be
742 assert(isExpectedBDVType(Base) && "the only non-base values "
743 "we see should be base defining values");
744 if (Visited.insert(Base).second)
745 Worklist.push_back(Base);
747 if (PHINode *Phi = dyn_cast<PHINode>(Current)) {
748 for (Value *InVal : Phi->incoming_values())
749 visitIncomingValue(InVal);
750 } else if (SelectInst *Sel = dyn_cast<SelectInst>(Current)) {
751 visitIncomingValue(Sel->getTrueValue());
752 visitIncomingValue(Sel->getFalseValue());
753 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
754 visitIncomingValue(EE->getVectorOperand());
756 // There are two classes of instructions we know we don't handle.
757 assert(isa<ShuffleVectorInst>(Current) ||
758 isa<InsertElementInst>(Current));
759 llvm_unreachable("unimplemented instruction case");
762 // The frontier of visited instructions are the ones we might need to
763 // duplicate, so fill in the starting state for the optimistic algorithm
765 for (Value *BDV : Visited) {
766 states[BDV] = BDVState();
771 DEBUG(dbgs() << "States after initialization:\n");
772 for (auto Pair : states) {
773 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
777 // TODO: come back and revisit the state transitions around inputs which
778 // have reached conflict state. The current version seems too conservative.
780 // Return a phi state for a base defining value. We'll generate a new
781 // base state for known bases and expect to find a cached state otherwise.
782 auto getStateForBDV = [&](Value *baseValue) {
783 if (isKnownBaseResult(baseValue))
784 return BDVState(baseValue);
785 auto I = states.find(baseValue);
786 assert(I != states.end() && "lookup failed!");
790 bool progress = true;
793 size_t oldSize = states.size();
796 // We're only changing keys in this loop, thus safe to keep iterators
797 for (auto Pair : states) {
798 Value *v = Pair.first;
799 assert(!isKnownBaseResult(v) && "why did it get added?");
801 // Given an input value for the current instruction, return a BDVState
802 // instance which represents the BDV of that value.
803 auto getStateForInput = [&](Value *V) mutable {
804 Value *BDV = findBaseOrBDV(V, cache);
805 return getStateForBDV(BDV);
808 MeetBDVStates calculateMeet;
809 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
810 calculateMeet.meetWith(getStateForInput(select->getTrueValue()));
811 calculateMeet.meetWith(getStateForInput(select->getFalseValue()));
812 } else if (PHINode *Phi = dyn_cast<PHINode>(v)) {
813 for (Value *Val : Phi->incoming_values())
814 calculateMeet.meetWith(getStateForInput(Val));
816 // The 'meet' for an extractelement is slightly trivial, but it's still
817 // useful in that it drives us to conflict if our input is.
818 auto *EE = cast<ExtractElementInst>(v);
819 calculateMeet.meetWith(getStateForInput(EE->getVectorOperand()));
823 BDVState oldState = states[v];
824 BDVState newState = calculateMeet.getResult();
825 if (oldState != newState) {
827 states[v] = newState;
831 assert(oldSize <= states.size());
832 assert(oldSize == states.size() || progress);
836 DEBUG(dbgs() << "States after meet iteration:\n");
837 for (auto Pair : states) {
838 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
842 // Insert Phis for all conflicts
843 // We want to keep naming deterministic in the loop that follows, so
844 // sort the keys before iteration. This is useful in allowing us to
845 // write stable tests. Note that there is no invalidation issue here.
846 SmallVector<Value *, 16> Keys;
847 Keys.reserve(states.size());
848 for (auto Pair : states) {
849 Value *V = Pair.first;
852 std::sort(Keys.begin(), Keys.end(), order_by_name);
853 // TODO: adjust naming patterns to avoid this order of iteration dependency
854 for (Value *V : Keys) {
855 Instruction *I = cast<Instruction>(V);
856 BDVState State = states[I];
857 assert(!isKnownBaseResult(I) && "why did it get added?");
858 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
860 // extractelement instructions are a bit special in that we may need to
861 // insert an extract even when we know an exact base for the instruction.
862 // The problem is that we need to convert from a vector base to a scalar
863 // base for the particular indice we're interested in.
864 if (State.isBase() && isa<ExtractElementInst>(I) &&
865 isa<VectorType>(State.getBase()->getType())) {
866 auto *EE = cast<ExtractElementInst>(I);
867 // TODO: In many cases, the new instruction is just EE itself. We should
868 // exploit this, but can't do it here since it would break the invariant
869 // about the BDV not being known to be a base.
870 auto *BaseInst = ExtractElementInst::Create(State.getBase(),
871 EE->getIndexOperand(),
873 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
874 states[I] = BDVState(BDVState::Base, BaseInst);
877 if (!State.isConflict())
880 /// Create and insert a new instruction which will represent the base of
881 /// the given instruction 'I'.
882 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
883 if (isa<PHINode>(I)) {
884 BasicBlock *BB = I->getParent();
885 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
886 assert(NumPreds > 0 && "how did we reach here");
887 std::string Name = I->hasName() ?
888 (I->getName() + ".base").str() : "base_phi";
889 return PHINode::Create(I->getType(), NumPreds, Name, I);
890 } else if (SelectInst *Sel = dyn_cast<SelectInst>(I)) {
891 // The undef will be replaced later
892 UndefValue *Undef = UndefValue::get(Sel->getType());
893 std::string Name = I->hasName() ?
894 (I->getName() + ".base").str() : "base_select";
895 return SelectInst::Create(Sel->getCondition(), Undef,
898 auto *EE = cast<ExtractElementInst>(I);
899 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
900 std::string Name = I->hasName() ?
901 (I->getName() + ".base").str() : "base_ee";
902 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
906 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
907 // Add metadata marking this as a base value
908 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
909 states[I] = BDVState(BDVState::Conflict, BaseInst);
912 // Fixup all the inputs of the new PHIs
913 for (auto Pair : states) {
914 Instruction *v = cast<Instruction>(Pair.first);
915 BDVState state = Pair.second;
917 assert(!isKnownBaseResult(v) && "why did it get added?");
918 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
919 if (!state.isConflict())
922 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
923 PHINode *phi = cast<PHINode>(v);
924 unsigned NumPHIValues = phi->getNumIncomingValues();
925 for (unsigned i = 0; i < NumPHIValues; i++) {
926 Value *InVal = phi->getIncomingValue(i);
927 BasicBlock *InBB = phi->getIncomingBlock(i);
929 // If we've already seen InBB, add the same incoming value
930 // we added for it earlier. The IR verifier requires phi
931 // nodes with multiple entries from the same basic block
932 // to have the same incoming value for each of those
933 // entries. If we don't do this check here and basephi
934 // has a different type than base, we'll end up adding two
935 // bitcasts (and hence two distinct values) as incoming
936 // values for the same basic block.
938 int blockIndex = basephi->getBasicBlockIndex(InBB);
939 if (blockIndex != -1) {
940 Value *oldBase = basephi->getIncomingValue(blockIndex);
941 basephi->addIncoming(oldBase, InBB);
943 Value *base = findBaseOrBDV(InVal, cache);
944 if (!isKnownBaseResult(base)) {
945 // Either conflict or base.
946 assert(states.count(base));
947 base = states[base].getBase();
948 assert(base != nullptr && "unknown BDVState!");
951 // In essence this assert states: the only way two
952 // values incoming from the same basic block may be
953 // different is by being different bitcasts of the same
954 // value. A cleanup that remains TODO is changing
955 // findBaseOrBDV to return an llvm::Value of the correct
956 // type (and still remain pure). This will remove the
957 // need to add bitcasts.
958 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
959 "sanity -- findBaseOrBDV should be pure!");
964 // Find either the defining value for the PHI or the normal base for
966 Value *base = findBaseOrBDV(InVal, cache);
967 if (!isKnownBaseResult(base)) {
968 // Either conflict or base.
969 assert(states.count(base));
970 base = states[base].getBase();
971 assert(base != nullptr && "unknown BDVState!");
973 assert(base && "can't be null");
974 // Must use original input BB since base may not be Instruction
975 // The cast is needed since base traversal may strip away bitcasts
976 if (base->getType() != basephi->getType()) {
977 base = new BitCastInst(base, basephi->getType(), "cast",
978 InBB->getTerminator());
980 basephi->addIncoming(base, InBB);
982 assert(basephi->getNumIncomingValues() == NumPHIValues);
983 } else if (SelectInst *basesel = dyn_cast<SelectInst>(state.getBase())) {
984 SelectInst *sel = cast<SelectInst>(v);
985 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
986 // something more safe and less hacky.
987 for (int i = 1; i <= 2; i++) {
988 Value *InVal = sel->getOperand(i);
989 // Find either the defining value for the PHI or the normal base for
991 Value *base = findBaseOrBDV(InVal, cache);
992 if (!isKnownBaseResult(base)) {
993 // Either conflict or base.
994 assert(states.count(base));
995 base = states[base].getBase();
996 assert(base != nullptr && "unknown BDVState!");
998 assert(base && "can't be null");
999 // Must use original input BB since base may not be Instruction
1000 // The cast is needed since base traversal may strip away bitcasts
1001 if (base->getType() != basesel->getType()) {
1002 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
1004 basesel->setOperand(i, base);
1007 auto *BaseEE = cast<ExtractElementInst>(state.getBase());
1008 Value *InVal = cast<ExtractElementInst>(v)->getVectorOperand();
1009 Value *Base = findBaseOrBDV(InVal, cache);
1010 if (!isKnownBaseResult(Base)) {
1011 // Either conflict or base.
1012 assert(states.count(Base));
1013 Base = states[Base].getBase();
1014 assert(Base != nullptr && "unknown BDVState!");
1016 assert(Base && "can't be null");
1017 BaseEE->setOperand(0, Base);
1021 // Now that we're done with the algorithm, see if we can optimize the
1022 // results slightly by reducing the number of new instructions needed.
1023 // Arguably, this should be integrated into the algorithm above, but
1024 // doing as a post process step is easier to reason about for the moment.
1025 DenseMap<Value *, Value *> ReverseMap;
1026 SmallPtrSet<Instruction *, 16> NewInsts;
1027 SmallSetVector<Instruction *, 16> Worklist;
1028 for (auto Item : states) {
1029 Value *V = Item.first;
1030 Value *Base = Item.second.getBase();
1032 assert(!isKnownBaseResult(V) && "why did it get added?");
1033 assert(isKnownBaseResult(Base) &&
1034 "must be something we 'know' is a base pointer");
1035 if (!Item.second.isConflict())
1038 ReverseMap[Base] = V;
1039 if (auto *BaseI = dyn_cast<Instruction>(Base)) {
1040 NewInsts.insert(BaseI);
1041 Worklist.insert(BaseI);
1044 auto PushNewUsers = [&](Instruction *I) {
1045 for (User *U : I->users())
1046 if (auto *UI = dyn_cast<Instruction>(U))
1047 if (NewInsts.count(UI))
1048 Worklist.insert(UI);
1050 const DataLayout &DL = cast<Instruction>(def)->getModule()->getDataLayout();
1051 while (!Worklist.empty()) {
1052 Instruction *BaseI = Worklist.pop_back_val();
1053 assert(NewInsts.count(BaseI));
1054 Value *Bdv = ReverseMap[BaseI];
1055 if (auto *BdvI = dyn_cast<Instruction>(Bdv))
1056 if (BaseI->isIdenticalTo(BdvI)) {
1057 DEBUG(dbgs() << "Identical Base: " << *BaseI << "\n");
1058 PushNewUsers(BaseI);
1059 BaseI->replaceAllUsesWith(Bdv);
1060 BaseI->eraseFromParent();
1061 states[Bdv] = BDVState(BDVState::Conflict, Bdv);
1062 NewInsts.erase(BaseI);
1063 ReverseMap.erase(BaseI);
1066 if (Value *V = SimplifyInstruction(BaseI, DL)) {
1067 DEBUG(dbgs() << "Base " << *BaseI << " simplified to " << *V << "\n");
1068 PushNewUsers(BaseI);
1069 BaseI->replaceAllUsesWith(V);
1070 BaseI->eraseFromParent();
1071 states[Bdv] = BDVState(BDVState::Conflict, V);
1072 NewInsts.erase(BaseI);
1073 ReverseMap.erase(BaseI);
1078 // Cache all of our results so we can cheaply reuse them
1079 // NOTE: This is actually two caches: one of the base defining value
1080 // relation and one of the base pointer relation! FIXME
1081 for (auto item : states) {
1082 Value *v = item.first;
1083 Value *base = item.second.getBase();
1086 std::string fromstr =
1087 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
1089 DEBUG(dbgs() << "Updating base value cache"
1090 << " for: " << (v->hasName() ? v->getName() : "")
1091 << " from: " << fromstr
1092 << " to: " << (base->hasName() ? base->getName() : "") << "\n");
1094 if (cache.count(v)) {
1095 // Once we transition from the BDV relation being store in the cache to
1096 // the base relation being stored, it must be stable
1097 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
1098 "base relation should be stable");
1102 assert(cache.find(def) != cache.end());
1106 // For a set of live pointers (base and/or derived), identify the base
1107 // pointer of the object which they are derived from. This routine will
1108 // mutate the IR graph as needed to make the 'base' pointer live at the
1109 // definition site of 'derived'. This ensures that any use of 'derived' can
1110 // also use 'base'. This may involve the insertion of a number of
1111 // additional PHI nodes.
1113 // preconditions: live is a set of pointer type Values
1115 // side effects: may insert PHI nodes into the existing CFG, will preserve
1116 // CFG, will not remove or mutate any existing nodes
1118 // post condition: PointerToBase contains one (derived, base) pair for every
1119 // pointer in live. Note that derived can be equal to base if the original
1120 // pointer was a base pointer.
1122 findBasePointers(const StatepointLiveSetTy &live,
1123 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
1124 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1125 // For the naming of values inserted to be deterministic - which makes for
1126 // much cleaner and more stable tests - we need to assign an order to the
1127 // live values. DenseSets do not provide a deterministic order across runs.
1128 SmallVector<Value *, 64> Temp;
1129 Temp.insert(Temp.end(), live.begin(), live.end());
1130 std::sort(Temp.begin(), Temp.end(), order_by_name);
1131 for (Value *ptr : Temp) {
1132 Value *base = findBasePointer(ptr, DVCache);
1133 assert(base && "failed to find base pointer");
1134 PointerToBase[ptr] = base;
1135 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1136 DT->dominates(cast<Instruction>(base)->getParent(),
1137 cast<Instruction>(ptr)->getParent())) &&
1138 "The base we found better dominate the derived pointer");
1140 // If you see this trip and like to live really dangerously, the code should
1141 // be correct, just with idioms the verifier can't handle. You can try
1142 // disabling the verifier at your own substantial risk.
1143 assert(!isa<ConstantPointerNull>(base) &&
1144 "the relocation code needs adjustment to handle the relocation of "
1145 "a null pointer constant without causing false positives in the "
1146 "safepoint ir verifier.");
1150 /// Find the required based pointers (and adjust the live set) for the given
1152 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1154 PartiallyConstructedSafepointRecord &result) {
1155 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
1156 findBasePointers(result.liveset, PointerToBase, &DT, DVCache);
1158 if (PrintBasePointers) {
1159 // Note: Need to print these in a stable order since this is checked in
1161 errs() << "Base Pairs (w/o Relocation):\n";
1162 SmallVector<Value *, 64> Temp;
1163 Temp.reserve(PointerToBase.size());
1164 for (auto Pair : PointerToBase) {
1165 Temp.push_back(Pair.first);
1167 std::sort(Temp.begin(), Temp.end(), order_by_name);
1168 for (Value *Ptr : Temp) {
1169 Value *Base = PointerToBase[Ptr];
1170 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
1175 result.PointerToBase = PointerToBase;
1178 /// Given an updated version of the dataflow liveness results, update the
1179 /// liveset and base pointer maps for the call site CS.
1180 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1182 PartiallyConstructedSafepointRecord &result);
1184 static void recomputeLiveInValues(
1185 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1186 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1187 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1188 // again. The old values are still live and will help it stabilize quickly.
1189 GCPtrLivenessData RevisedLivenessData;
1190 computeLiveInValues(DT, F, RevisedLivenessData);
1191 for (size_t i = 0; i < records.size(); i++) {
1192 struct PartiallyConstructedSafepointRecord &info = records[i];
1193 const CallSite &CS = toUpdate[i];
1194 recomputeLiveInValues(RevisedLivenessData, CS, info);
1198 // When inserting gc.relocate calls, we need to ensure there are no uses
1199 // of the original value between the gc.statepoint and the gc.relocate call.
1200 // One case which can arise is a phi node starting one of the successor blocks.
1201 // We also need to be able to insert the gc.relocates only on the path which
1202 // goes through the statepoint. We might need to split an edge to make this
1205 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1206 DominatorTree &DT) {
1207 BasicBlock *Ret = BB;
1208 if (!BB->getUniquePredecessor()) {
1209 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1212 // Now that 'ret' has unique predecessor we can safely remove all phi nodes
1214 FoldSingleEntryPHINodes(Ret);
1215 assert(!isa<PHINode>(Ret->begin()));
1217 // At this point, we can safely insert a gc.relocate as the first instruction
1218 // in Ret if needed.
1222 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1223 auto itr = std::find(livevec.begin(), livevec.end(), val);
1224 assert(livevec.end() != itr);
1225 size_t index = std::distance(livevec.begin(), itr);
1226 assert(index < livevec.size());
1230 // Create new attribute set containing only attributes which can be transferred
1231 // from original call to the safepoint.
1232 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1235 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1236 unsigned index = AS.getSlotIndex(Slot);
1238 if (index == AttributeSet::ReturnIndex ||
1239 index == AttributeSet::FunctionIndex) {
1241 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1243 Attribute attr = *it;
1245 // Do not allow certain attributes - just skip them
1246 // Safepoint can not be read only or read none.
1247 if (attr.hasAttribute(Attribute::ReadNone) ||
1248 attr.hasAttribute(Attribute::ReadOnly))
1251 ret = ret.addAttributes(
1252 AS.getContext(), index,
1253 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1257 // Just skip parameter attributes for now
1263 /// Helper function to place all gc relocates necessary for the given
1266 /// liveVariables - list of variables to be relocated.
1267 /// liveStart - index of the first live variable.
1268 /// basePtrs - base pointers.
1269 /// statepointToken - statepoint instruction to which relocates should be
1271 /// Builder - Llvm IR builder to be used to construct new calls.
1272 static void CreateGCRelocates(ArrayRef<llvm::Value *> LiveVariables,
1273 const int LiveStart,
1274 ArrayRef<llvm::Value *> BasePtrs,
1275 Instruction *StatepointToken,
1276 IRBuilder<> Builder) {
1277 if (LiveVariables.empty())
1280 // All gc_relocate are set to i8 addrspace(1)* type. We originally generated
1281 // unique declarations for each pointer type, but this proved problematic
1282 // because the intrinsic mangling code is incomplete and fragile. Since
1283 // we're moving towards a single unified pointer type anyways, we can just
1284 // cast everything to an i8* of the right address space. A bitcast is added
1285 // later to convert gc_relocate to the actual value's type.
1286 Module *M = StatepointToken->getModule();
1287 auto AS = cast<PointerType>(LiveVariables[0]->getType())->getAddressSpace();
1288 Type *Types[] = {Type::getInt8PtrTy(M->getContext(), AS)};
1289 Value *GCRelocateDecl =
1290 Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, Types);
1292 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1293 // Generate the gc.relocate call and save the result
1295 Builder.getInt32(LiveStart + find_index(LiveVariables, BasePtrs[i]));
1297 Builder.getInt32(LiveStart + find_index(LiveVariables, LiveVariables[i]));
1299 // only specify a debug name if we can give a useful one
1300 CallInst *Reloc = Builder.CreateCall(
1301 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1302 LiveVariables[i]->hasName() ? LiveVariables[i]->getName() + ".relocated"
1304 // Trick CodeGen into thinking there are lots of free registers at this
1306 Reloc->setCallingConv(CallingConv::Cold);
1311 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1312 const SmallVectorImpl<llvm::Value *> &basePtrs,
1313 const SmallVectorImpl<llvm::Value *> &liveVariables,
1315 PartiallyConstructedSafepointRecord &result) {
1316 assert(basePtrs.size() == liveVariables.size());
1317 assert(isStatepoint(CS) &&
1318 "This method expects to be rewriting a statepoint");
1320 BasicBlock *BB = CS.getInstruction()->getParent();
1322 Function *F = BB->getParent();
1323 assert(F && "must be set");
1324 Module *M = F->getParent();
1326 assert(M && "must be set");
1328 // We're not changing the function signature of the statepoint since the gc
1329 // arguments go into the var args section.
1330 Function *gc_statepoint_decl = CS.getCalledFunction();
1332 // Then go ahead and use the builder do actually do the inserts. We insert
1333 // immediately before the previous instruction under the assumption that all
1334 // arguments will be available here. We can't insert afterwards since we may
1335 // be replacing a terminator.
1336 Instruction *insertBefore = CS.getInstruction();
1337 IRBuilder<> Builder(insertBefore);
1338 // Copy all of the arguments from the original statepoint - this includes the
1339 // target, call args, and deopt args
1340 SmallVector<llvm::Value *, 64> args;
1341 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1342 // TODO: Clear the 'needs rewrite' flag
1344 // add all the pointers to be relocated (gc arguments)
1345 // Capture the start of the live variable list for use in the gc_relocates
1346 const int live_start = args.size();
1347 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1349 // Create the statepoint given all the arguments
1350 Instruction *token = nullptr;
1351 AttributeSet return_attributes;
1353 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1355 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1356 call->setTailCall(toReplace->isTailCall());
1357 call->setCallingConv(toReplace->getCallingConv());
1359 // Currently we will fail on parameter attributes and on certain
1360 // function attributes.
1361 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1362 // In case if we can handle this set of attributes - set up function attrs
1363 // directly on statepoint and return attrs later for gc_result intrinsic.
1364 call->setAttributes(new_attrs.getFnAttributes());
1365 return_attributes = new_attrs.getRetAttributes();
1369 // Put the following gc_result and gc_relocate calls immediately after the
1370 // the old call (which we're about to delete)
1371 BasicBlock::iterator next(toReplace);
1372 assert(BB->end() != next && "not a terminator, must have next");
1374 Instruction *IP = &*(next);
1375 Builder.SetInsertPoint(IP);
1376 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1379 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1381 // Insert the new invoke into the old block. We'll remove the old one in a
1382 // moment at which point this will become the new terminator for the
1384 InvokeInst *invoke = InvokeInst::Create(
1385 gc_statepoint_decl, toReplace->getNormalDest(),
1386 toReplace->getUnwindDest(), args, "statepoint_token", toReplace->getParent());
1387 invoke->setCallingConv(toReplace->getCallingConv());
1389 // Currently we will fail on parameter attributes and on certain
1390 // function attributes.
1391 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1392 // In case if we can handle this set of attributes - set up function attrs
1393 // directly on statepoint and return attrs later for gc_result intrinsic.
1394 invoke->setAttributes(new_attrs.getFnAttributes());
1395 return_attributes = new_attrs.getRetAttributes();
1399 // Generate gc relocates in exceptional path
1400 BasicBlock *unwindBlock = toReplace->getUnwindDest();
1401 assert(!isa<PHINode>(unwindBlock->begin()) &&
1402 unwindBlock->getUniquePredecessor() &&
1403 "can't safely insert in this block!");
1405 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1406 Builder.SetInsertPoint(IP);
1407 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1409 // Extract second element from landingpad return value. We will attach
1410 // exceptional gc relocates to it.
1411 const unsigned idx = 1;
1412 Instruction *exceptional_token =
1413 cast<Instruction>(Builder.CreateExtractValue(
1414 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1415 result.UnwindToken = exceptional_token;
1417 CreateGCRelocates(liveVariables, live_start, basePtrs,
1418 exceptional_token, Builder);
1420 // Generate gc relocates and returns for normal block
1421 BasicBlock *normalDest = toReplace->getNormalDest();
1422 assert(!isa<PHINode>(normalDest->begin()) &&
1423 normalDest->getUniquePredecessor() &&
1424 "can't safely insert in this block!");
1426 IP = &*(normalDest->getFirstInsertionPt());
1427 Builder.SetInsertPoint(IP);
1429 // gc relocates will be generated later as if it were regular call
1434 // Take the name of the original value call if it had one.
1435 token->takeName(CS.getInstruction());
1437 // The GCResult is already inserted, we just need to find it
1439 Instruction *toReplace = CS.getInstruction();
1440 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1441 "only valid use before rewrite is gc.result");
1442 assert(!toReplace->hasOneUse() ||
1443 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1446 // Update the gc.result of the original statepoint (if any) to use the newly
1447 // inserted statepoint. This is safe to do here since the token can't be
1448 // considered a live reference.
1449 CS.getInstruction()->replaceAllUsesWith(token);
1451 result.StatepointToken = token;
1453 // Second, create a gc.relocate for every live variable
1454 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1458 struct name_ordering {
1461 bool operator()(name_ordering const &a, name_ordering const &b) {
1462 return -1 == a.derived->getName().compare(b.derived->getName());
1466 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1467 SmallVectorImpl<Value *> &livevec) {
1468 assert(basevec.size() == livevec.size());
1470 SmallVector<name_ordering, 64> temp;
1471 for (size_t i = 0; i < basevec.size(); i++) {
1473 v.base = basevec[i];
1474 v.derived = livevec[i];
1477 std::sort(temp.begin(), temp.end(), name_ordering());
1478 for (size_t i = 0; i < basevec.size(); i++) {
1479 basevec[i] = temp[i].base;
1480 livevec[i] = temp[i].derived;
1484 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1485 // which make the relocations happening at this safepoint explicit.
1487 // WARNING: Does not do any fixup to adjust users of the original live
1488 // values. That's the callers responsibility.
1490 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1491 PartiallyConstructedSafepointRecord &result) {
1492 auto liveset = result.liveset;
1493 auto PointerToBase = result.PointerToBase;
1495 // Convert to vector for efficient cross referencing.
1496 SmallVector<Value *, 64> basevec, livevec;
1497 livevec.reserve(liveset.size());
1498 basevec.reserve(liveset.size());
1499 for (Value *L : liveset) {
1500 livevec.push_back(L);
1501 assert(PointerToBase.count(L));
1502 Value *base = PointerToBase[L];
1503 basevec.push_back(base);
1505 assert(livevec.size() == basevec.size());
1507 // To make the output IR slightly more stable (for use in diffs), ensure a
1508 // fixed order of the values in the safepoint (by sorting the value name).
1509 // The order is otherwise meaningless.
1510 stablize_order(basevec, livevec);
1512 // Do the actual rewriting and delete the old statepoint
1513 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1514 CS.getInstruction()->eraseFromParent();
1517 // Helper function for the relocationViaAlloca.
1518 // It receives iterator to the statepoint gc relocates and emits store to the
1520 // location (via allocaMap) for the each one of them.
1521 // Add visited values into the visitedLiveValues set we will later use them
1522 // for sanity check.
1524 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1525 DenseMap<Value *, Value *> &AllocaMap,
1526 DenseSet<Value *> &VisitedLiveValues) {
1528 for (User *U : GCRelocs) {
1529 if (!isa<IntrinsicInst>(U))
1532 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1534 // We only care about relocates
1535 if (RelocatedValue->getIntrinsicID() !=
1536 Intrinsic::experimental_gc_relocate) {
1540 GCRelocateOperands RelocateOperands(RelocatedValue);
1541 Value *OriginalValue =
1542 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1543 assert(AllocaMap.count(OriginalValue));
1544 Value *Alloca = AllocaMap[OriginalValue];
1546 // Emit store into the related alloca
1547 // All gc_relocate are i8 addrspace(1)* typed, and it must be bitcasted to
1548 // the correct type according to alloca.
1549 assert(RelocatedValue->getNextNode() && "Should always have one since it's not a terminator");
1550 IRBuilder<> Builder(RelocatedValue->getNextNode());
1551 Value *CastedRelocatedValue =
1552 Builder.CreateBitCast(RelocatedValue, cast<AllocaInst>(Alloca)->getAllocatedType(),
1553 RelocatedValue->hasName() ? RelocatedValue->getName() + ".casted" : "");
1555 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1556 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1559 VisitedLiveValues.insert(OriginalValue);
1564 // Helper function for the "relocationViaAlloca". Similar to the
1565 // "insertRelocationStores" but works for rematerialized values.
1567 insertRematerializationStores(
1568 RematerializedValueMapTy RematerializedValues,
1569 DenseMap<Value *, Value *> &AllocaMap,
1570 DenseSet<Value *> &VisitedLiveValues) {
1572 for (auto RematerializedValuePair: RematerializedValues) {
1573 Instruction *RematerializedValue = RematerializedValuePair.first;
1574 Value *OriginalValue = RematerializedValuePair.second;
1576 assert(AllocaMap.count(OriginalValue) &&
1577 "Can not find alloca for rematerialized value");
1578 Value *Alloca = AllocaMap[OriginalValue];
1580 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1581 Store->insertAfter(RematerializedValue);
1584 VisitedLiveValues.insert(OriginalValue);
1589 /// do all the relocation update via allocas and mem2reg
1590 static void relocationViaAlloca(
1591 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1592 ArrayRef<struct PartiallyConstructedSafepointRecord> Records) {
1594 // record initial number of (static) allocas; we'll check we have the same
1595 // number when we get done.
1596 int InitialAllocaNum = 0;
1597 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1599 if (isa<AllocaInst>(*I))
1603 // TODO-PERF: change data structures, reserve
1604 DenseMap<Value *, Value *> AllocaMap;
1605 SmallVector<AllocaInst *, 200> PromotableAllocas;
1606 // Used later to chack that we have enough allocas to store all values
1607 std::size_t NumRematerializedValues = 0;
1608 PromotableAllocas.reserve(Live.size());
1610 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1611 // "PromotableAllocas"
1612 auto emitAllocaFor = [&](Value *LiveValue) {
1613 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1614 F.getEntryBlock().getFirstNonPHI());
1615 AllocaMap[LiveValue] = Alloca;
1616 PromotableAllocas.push_back(Alloca);
1619 // emit alloca for each live gc pointer
1620 for (unsigned i = 0; i < Live.size(); i++) {
1621 emitAllocaFor(Live[i]);
1624 // emit allocas for rematerialized values
1625 for (size_t i = 0; i < Records.size(); i++) {
1626 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1628 for (auto RematerializedValuePair : Info.RematerializedValues) {
1629 Value *OriginalValue = RematerializedValuePair.second;
1630 if (AllocaMap.count(OriginalValue) != 0)
1633 emitAllocaFor(OriginalValue);
1634 ++NumRematerializedValues;
1638 // The next two loops are part of the same conceptual operation. We need to
1639 // insert a store to the alloca after the original def and at each
1640 // redefinition. We need to insert a load before each use. These are split
1641 // into distinct loops for performance reasons.
1643 // update gc pointer after each statepoint
1644 // either store a relocated value or null (if no relocated value found for
1645 // this gc pointer and it is not a gc_result)
1646 // this must happen before we update the statepoint with load of alloca
1647 // otherwise we lose the link between statepoint and old def
1648 for (size_t i = 0; i < Records.size(); i++) {
1649 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1650 Value *Statepoint = Info.StatepointToken;
1652 // This will be used for consistency check
1653 DenseSet<Value *> VisitedLiveValues;
1655 // Insert stores for normal statepoint gc relocates
1656 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1658 // In case if it was invoke statepoint
1659 // we will insert stores for exceptional path gc relocates.
1660 if (isa<InvokeInst>(Statepoint)) {
1661 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1665 // Do similar thing with rematerialized values
1666 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1669 if (ClobberNonLive) {
1670 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1671 // the gc.statepoint. This will turn some subtle GC problems into
1672 // slightly easier to debug SEGVs. Note that on large IR files with
1673 // lots of gc.statepoints this is extremely costly both memory and time
1675 SmallVector<AllocaInst *, 64> ToClobber;
1676 for (auto Pair : AllocaMap) {
1677 Value *Def = Pair.first;
1678 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1680 // This value was relocated
1681 if (VisitedLiveValues.count(Def)) {
1684 ToClobber.push_back(Alloca);
1687 auto InsertClobbersAt = [&](Instruction *IP) {
1688 for (auto *AI : ToClobber) {
1689 auto AIType = cast<PointerType>(AI->getType());
1690 auto PT = cast<PointerType>(AIType->getElementType());
1691 Constant *CPN = ConstantPointerNull::get(PT);
1692 StoreInst *Store = new StoreInst(CPN, AI);
1693 Store->insertBefore(IP);
1697 // Insert the clobbering stores. These may get intermixed with the
1698 // gc.results and gc.relocates, but that's fine.
1699 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1700 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1701 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1703 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1705 InsertClobbersAt(Next);
1709 // update use with load allocas and add store for gc_relocated
1710 for (auto Pair : AllocaMap) {
1711 Value *Def = Pair.first;
1712 Value *Alloca = Pair.second;
1714 // we pre-record the uses of allocas so that we dont have to worry about
1716 // that change the user information.
1717 SmallVector<Instruction *, 20> Uses;
1718 // PERF: trade a linear scan for repeated reallocation
1719 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1720 for (User *U : Def->users()) {
1721 if (!isa<ConstantExpr>(U)) {
1722 // If the def has a ConstantExpr use, then the def is either a
1723 // ConstantExpr use itself or null. In either case
1724 // (recursively in the first, directly in the second), the oop
1725 // it is ultimately dependent on is null and this particular
1726 // use does not need to be fixed up.
1727 Uses.push_back(cast<Instruction>(U));
1731 std::sort(Uses.begin(), Uses.end());
1732 auto Last = std::unique(Uses.begin(), Uses.end());
1733 Uses.erase(Last, Uses.end());
1735 for (Instruction *Use : Uses) {
1736 if (isa<PHINode>(Use)) {
1737 PHINode *Phi = cast<PHINode>(Use);
1738 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1739 if (Def == Phi->getIncomingValue(i)) {
1740 LoadInst *Load = new LoadInst(
1741 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1742 Phi->setIncomingValue(i, Load);
1746 LoadInst *Load = new LoadInst(Alloca, "", Use);
1747 Use->replaceUsesOfWith(Def, Load);
1751 // emit store for the initial gc value
1752 // store must be inserted after load, otherwise store will be in alloca's
1753 // use list and an extra load will be inserted before it
1754 StoreInst *Store = new StoreInst(Def, Alloca);
1755 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1756 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1757 // InvokeInst is a TerminatorInst so the store need to be inserted
1758 // into its normal destination block.
1759 BasicBlock *NormalDest = Invoke->getNormalDest();
1760 Store->insertBefore(NormalDest->getFirstNonPHI());
1762 assert(!Inst->isTerminator() &&
1763 "The only TerminatorInst that can produce a value is "
1764 "InvokeInst which is handled above.");
1765 Store->insertAfter(Inst);
1768 assert(isa<Argument>(Def));
1769 Store->insertAfter(cast<Instruction>(Alloca));
1773 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1774 "we must have the same allocas with lives");
1775 if (!PromotableAllocas.empty()) {
1776 // apply mem2reg to promote alloca to SSA
1777 PromoteMemToReg(PromotableAllocas, DT);
1781 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1783 if (isa<AllocaInst>(*I))
1785 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1789 /// Implement a unique function which doesn't require we sort the input
1790 /// vector. Doing so has the effect of changing the output of a couple of
1791 /// tests in ways which make them less useful in testing fused safepoints.
1792 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1793 SmallSet<T, 8> Seen;
1794 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1795 return !Seen.insert(V).second;
1799 /// Insert holders so that each Value is obviously live through the entire
1800 /// lifetime of the call.
1801 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1802 SmallVectorImpl<CallInst *> &Holders) {
1804 // No values to hold live, might as well not insert the empty holder
1807 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1808 // Use a dummy vararg function to actually hold the values live
1809 Function *Func = cast<Function>(M->getOrInsertFunction(
1810 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1812 // For call safepoints insert dummy calls right after safepoint
1813 BasicBlock::iterator Next(CS.getInstruction());
1815 Holders.push_back(CallInst::Create(Func, Values, "", Next));
1818 // For invoke safepooints insert dummy calls both in normal and
1819 // exceptional destination blocks
1820 auto *II = cast<InvokeInst>(CS.getInstruction());
1821 Holders.push_back(CallInst::Create(
1822 Func, Values, "", II->getNormalDest()->getFirstInsertionPt()));
1823 Holders.push_back(CallInst::Create(
1824 Func, Values, "", II->getUnwindDest()->getFirstInsertionPt()));
1827 static void findLiveReferences(
1828 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1829 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1830 GCPtrLivenessData OriginalLivenessData;
1831 computeLiveInValues(DT, F, OriginalLivenessData);
1832 for (size_t i = 0; i < records.size(); i++) {
1833 struct PartiallyConstructedSafepointRecord &info = records[i];
1834 const CallSite &CS = toUpdate[i];
1835 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1839 /// Remove any vector of pointers from the liveset by scalarizing them over the
1840 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1841 /// would be preferable to include the vector in the statepoint itself, but
1842 /// the lowering code currently does not handle that. Extending it would be
1843 /// slightly non-trivial since it requires a format change. Given how rare
1844 /// such cases are (for the moment?) scalarizing is an acceptable compromise.
1845 static void splitVectorValues(Instruction *StatepointInst,
1846 StatepointLiveSetTy &LiveSet,
1847 DenseMap<Value *, Value *>& PointerToBase,
1848 DominatorTree &DT) {
1849 SmallVector<Value *, 16> ToSplit;
1850 for (Value *V : LiveSet)
1851 if (isa<VectorType>(V->getType()))
1852 ToSplit.push_back(V);
1854 if (ToSplit.empty())
1857 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1859 Function &F = *(StatepointInst->getParent()->getParent());
1861 DenseMap<Value *, AllocaInst *> AllocaMap;
1862 // First is normal return, second is exceptional return (invoke only)
1863 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1864 for (Value *V : ToSplit) {
1865 AllocaInst *Alloca =
1866 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1867 AllocaMap[V] = Alloca;
1869 VectorType *VT = cast<VectorType>(V->getType());
1870 IRBuilder<> Builder(StatepointInst);
1871 SmallVector<Value *, 16> Elements;
1872 for (unsigned i = 0; i < VT->getNumElements(); i++)
1873 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1874 ElementMapping[V] = Elements;
1876 auto InsertVectorReform = [&](Instruction *IP) {
1877 Builder.SetInsertPoint(IP);
1878 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1879 Value *ResultVec = UndefValue::get(VT);
1880 for (unsigned i = 0; i < VT->getNumElements(); i++)
1881 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1882 Builder.getInt32(i));
1886 if (isa<CallInst>(StatepointInst)) {
1887 BasicBlock::iterator Next(StatepointInst);
1889 Instruction *IP = &*(Next);
1890 Replacements[V].first = InsertVectorReform(IP);
1891 Replacements[V].second = nullptr;
1893 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1894 // We've already normalized - check that we don't have shared destination
1896 BasicBlock *NormalDest = Invoke->getNormalDest();
1897 assert(!isa<PHINode>(NormalDest->begin()));
1898 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1899 assert(!isa<PHINode>(UnwindDest->begin()));
1900 // Insert insert element sequences in both successors
1901 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1902 Replacements[V].first = InsertVectorReform(IP);
1903 IP = &*(UnwindDest->getFirstInsertionPt());
1904 Replacements[V].second = InsertVectorReform(IP);
1908 for (Value *V : ToSplit) {
1909 AllocaInst *Alloca = AllocaMap[V];
1911 // Capture all users before we start mutating use lists
1912 SmallVector<Instruction *, 16> Users;
1913 for (User *U : V->users())
1914 Users.push_back(cast<Instruction>(U));
1916 for (Instruction *I : Users) {
1917 if (auto Phi = dyn_cast<PHINode>(I)) {
1918 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1919 if (V == Phi->getIncomingValue(i)) {
1920 LoadInst *Load = new LoadInst(
1921 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1922 Phi->setIncomingValue(i, Load);
1925 LoadInst *Load = new LoadInst(Alloca, "", I);
1926 I->replaceUsesOfWith(V, Load);
1930 // Store the original value and the replacement value into the alloca
1931 StoreInst *Store = new StoreInst(V, Alloca);
1932 if (auto I = dyn_cast<Instruction>(V))
1933 Store->insertAfter(I);
1935 Store->insertAfter(Alloca);
1937 // Normal return for invoke, or call return
1938 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1939 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1940 // Unwind return for invoke only
1941 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1943 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1946 // apply mem2reg to promote alloca to SSA
1947 SmallVector<AllocaInst *, 16> Allocas;
1948 for (Value *V : ToSplit)
1949 Allocas.push_back(AllocaMap[V]);
1950 PromoteMemToReg(Allocas, DT);
1952 // Update our tracking of live pointers and base mappings to account for the
1953 // changes we just made.
1954 for (Value *V : ToSplit) {
1955 auto &Elements = ElementMapping[V];
1958 LiveSet.insert(Elements.begin(), Elements.end());
1959 // We need to update the base mapping as well.
1960 assert(PointerToBase.count(V));
1961 Value *OldBase = PointerToBase[V];
1962 auto &BaseElements = ElementMapping[OldBase];
1963 PointerToBase.erase(V);
1964 assert(Elements.size() == BaseElements.size());
1965 for (unsigned i = 0; i < Elements.size(); i++) {
1966 Value *Elem = Elements[i];
1967 PointerToBase[Elem] = BaseElements[i];
1972 // Helper function for the "rematerializeLiveValues". It walks use chain
1973 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
1974 // values are visited (currently it is GEP's and casts). Returns true if it
1975 // successfully reached "BaseValue" and false otherwise.
1976 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
1978 static bool findRematerializableChainToBasePointer(
1979 SmallVectorImpl<Instruction*> &ChainToBase,
1980 Value *CurrentValue, Value *BaseValue) {
1982 // We have found a base value
1983 if (CurrentValue == BaseValue) {
1987 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
1988 ChainToBase.push_back(GEP);
1989 return findRematerializableChainToBasePointer(ChainToBase,
1990 GEP->getPointerOperand(),
1994 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
1995 Value *Def = CI->stripPointerCasts();
1997 // This two checks are basically similar. First one is here for the
1998 // consistency with findBasePointers logic.
1999 assert(!isa<CastInst>(Def) && "not a pointer cast found");
2000 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
2003 ChainToBase.push_back(CI);
2004 return findRematerializableChainToBasePointer(ChainToBase, Def, BaseValue);
2007 // Not supported instruction in the chain
2011 // Helper function for the "rematerializeLiveValues". Compute cost of the use
2012 // chain we are going to rematerialize.
2014 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
2015 TargetTransformInfo &TTI) {
2018 for (Instruction *Instr : Chain) {
2019 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
2020 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
2021 "non noop cast is found during rematerialization");
2023 Type *SrcTy = CI->getOperand(0)->getType();
2024 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
2026 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
2027 // Cost of the address calculation
2028 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
2029 Cost += TTI.getAddressComputationCost(ValTy);
2031 // And cost of the GEP itself
2032 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
2033 // allowed for the external usage)
2034 if (!GEP->hasAllConstantIndices())
2038 llvm_unreachable("unsupported instruciton type during rematerialization");
2045 // From the statepoint liveset pick values that are cheaper to recompute then to
2046 // relocate. Remove this values from the liveset, rematerialize them after
2047 // statepoint and record them in "Info" structure. Note that similar to
2048 // relocated values we don't do any user adjustments here.
2049 static void rematerializeLiveValues(CallSite CS,
2050 PartiallyConstructedSafepointRecord &Info,
2051 TargetTransformInfo &TTI) {
2052 const unsigned int ChainLengthThreshold = 10;
2054 // Record values we are going to delete from this statepoint live set.
2055 // We can not di this in following loop due to iterator invalidation.
2056 SmallVector<Value *, 32> LiveValuesToBeDeleted;
2058 for (Value *LiveValue: Info.liveset) {
2059 // For each live pointer find it's defining chain
2060 SmallVector<Instruction *, 3> ChainToBase;
2061 assert(Info.PointerToBase.count(LiveValue));
2063 findRematerializableChainToBasePointer(ChainToBase,
2065 Info.PointerToBase[LiveValue]);
2066 // Nothing to do, or chain is too long
2068 ChainToBase.size() == 0 ||
2069 ChainToBase.size() > ChainLengthThreshold)
2072 // Compute cost of this chain
2073 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
2074 // TODO: We can also account for cases when we will be able to remove some
2075 // of the rematerialized values by later optimization passes. I.e if
2076 // we rematerialized several intersecting chains. Or if original values
2077 // don't have any uses besides this statepoint.
2079 // For invokes we need to rematerialize each chain twice - for normal and
2080 // for unwind basic blocks. Model this by multiplying cost by two.
2081 if (CS.isInvoke()) {
2084 // If it's too expensive - skip it
2085 if (Cost >= RematerializationThreshold)
2088 // Remove value from the live set
2089 LiveValuesToBeDeleted.push_back(LiveValue);
2091 // Clone instructions and record them inside "Info" structure
2093 // Walk backwards to visit top-most instructions first
2094 std::reverse(ChainToBase.begin(), ChainToBase.end());
2096 // Utility function which clones all instructions from "ChainToBase"
2097 // and inserts them before "InsertBefore". Returns rematerialized value
2098 // which should be used after statepoint.
2099 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
2100 Instruction *LastClonedValue = nullptr;
2101 Instruction *LastValue = nullptr;
2102 for (Instruction *Instr: ChainToBase) {
2103 // Only GEP's and casts are suported as we need to be careful to not
2104 // introduce any new uses of pointers not in the liveset.
2105 // Note that it's fine to introduce new uses of pointers which were
2106 // otherwise not used after this statepoint.
2107 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2109 Instruction *ClonedValue = Instr->clone();
2110 ClonedValue->insertBefore(InsertBefore);
2111 ClonedValue->setName(Instr->getName() + ".remat");
2113 // If it is not first instruction in the chain then it uses previously
2114 // cloned value. We should update it to use cloned value.
2115 if (LastClonedValue) {
2117 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2119 // Assert that cloned instruction does not use any instructions from
2120 // this chain other than LastClonedValue
2121 for (auto OpValue : ClonedValue->operand_values()) {
2122 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2123 ChainToBase.end() &&
2124 "incorrect use in rematerialization chain");
2129 LastClonedValue = ClonedValue;
2132 assert(LastClonedValue);
2133 return LastClonedValue;
2136 // Different cases for calls and invokes. For invokes we need to clone
2137 // instructions both on normal and unwind path.
2139 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2140 assert(InsertBefore);
2141 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2142 Info.RematerializedValues[RematerializedValue] = LiveValue;
2144 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2146 Instruction *NormalInsertBefore =
2147 Invoke->getNormalDest()->getFirstInsertionPt();
2148 Instruction *UnwindInsertBefore =
2149 Invoke->getUnwindDest()->getFirstInsertionPt();
2151 Instruction *NormalRematerializedValue =
2152 rematerializeChain(NormalInsertBefore);
2153 Instruction *UnwindRematerializedValue =
2154 rematerializeChain(UnwindInsertBefore);
2156 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2157 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2161 // Remove rematerializaed values from the live set
2162 for (auto LiveValue: LiveValuesToBeDeleted) {
2163 Info.liveset.erase(LiveValue);
2167 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
2168 SmallVectorImpl<CallSite> &toUpdate) {
2170 // sanity check the input
2171 std::set<CallSite> uniqued;
2172 uniqued.insert(toUpdate.begin(), toUpdate.end());
2173 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
2175 for (size_t i = 0; i < toUpdate.size(); i++) {
2176 CallSite &CS = toUpdate[i];
2177 assert(CS.getInstruction()->getParent()->getParent() == &F);
2178 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
2182 // When inserting gc.relocates for invokes, we need to be able to insert at
2183 // the top of the successor blocks. See the comment on
2184 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2185 // may restructure the CFG.
2186 for (CallSite CS : toUpdate) {
2189 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
2190 normalizeForInvokeSafepoint(invoke->getNormalDest(), invoke->getParent(),
2192 normalizeForInvokeSafepoint(invoke->getUnwindDest(), invoke->getParent(),
2196 // A list of dummy calls added to the IR to keep various values obviously
2197 // live in the IR. We'll remove all of these when done.
2198 SmallVector<CallInst *, 64> holders;
2200 // Insert a dummy call with all of the arguments to the vm_state we'll need
2201 // for the actual safepoint insertion. This ensures reference arguments in
2202 // the deopt argument list are considered live through the safepoint (and
2203 // thus makes sure they get relocated.)
2204 for (size_t i = 0; i < toUpdate.size(); i++) {
2205 CallSite &CS = toUpdate[i];
2206 Statepoint StatepointCS(CS);
2208 SmallVector<Value *, 64> DeoptValues;
2209 for (Use &U : StatepointCS.vm_state_args()) {
2210 Value *Arg = cast<Value>(&U);
2211 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2212 "support for FCA unimplemented");
2213 if (isHandledGCPointerType(Arg->getType()))
2214 DeoptValues.push_back(Arg);
2216 insertUseHolderAfter(CS, DeoptValues, holders);
2219 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
2220 records.reserve(toUpdate.size());
2221 for (size_t i = 0; i < toUpdate.size(); i++) {
2222 struct PartiallyConstructedSafepointRecord info;
2223 records.push_back(info);
2225 assert(records.size() == toUpdate.size());
2227 // A) Identify all gc pointers which are statically live at the given call
2229 findLiveReferences(F, DT, P, toUpdate, records);
2231 // B) Find the base pointers for each live pointer
2232 /* scope for caching */ {
2233 // Cache the 'defining value' relation used in the computation and
2234 // insertion of base phis and selects. This ensures that we don't insert
2235 // large numbers of duplicate base_phis.
2236 DefiningValueMapTy DVCache;
2238 for (size_t i = 0; i < records.size(); i++) {
2239 struct PartiallyConstructedSafepointRecord &info = records[i];
2240 CallSite &CS = toUpdate[i];
2241 findBasePointers(DT, DVCache, CS, info);
2243 } // end of cache scope
2245 // The base phi insertion logic (for any safepoint) may have inserted new
2246 // instructions which are now live at some safepoint. The simplest such
2249 // phi a <-- will be a new base_phi here
2250 // safepoint 1 <-- that needs to be live here
2254 // We insert some dummy calls after each safepoint to definitely hold live
2255 // the base pointers which were identified for that safepoint. We'll then
2256 // ask liveness for _every_ base inserted to see what is now live. Then we
2257 // remove the dummy calls.
2258 holders.reserve(holders.size() + records.size());
2259 for (size_t i = 0; i < records.size(); i++) {
2260 struct PartiallyConstructedSafepointRecord &info = records[i];
2261 CallSite &CS = toUpdate[i];
2263 SmallVector<Value *, 128> Bases;
2264 for (auto Pair : info.PointerToBase) {
2265 Bases.push_back(Pair.second);
2267 insertUseHolderAfter(CS, Bases, holders);
2270 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2271 // need to rerun liveness. We may *also* have inserted new defs, but that's
2272 // not the key issue.
2273 recomputeLiveInValues(F, DT, P, toUpdate, records);
2275 if (PrintBasePointers) {
2276 for (size_t i = 0; i < records.size(); i++) {
2277 struct PartiallyConstructedSafepointRecord &info = records[i];
2278 errs() << "Base Pairs: (w/Relocation)\n";
2279 for (auto Pair : info.PointerToBase) {
2280 errs() << " derived %" << Pair.first->getName() << " base %"
2281 << Pair.second->getName() << "\n";
2285 for (size_t i = 0; i < holders.size(); i++) {
2286 holders[i]->eraseFromParent();
2287 holders[i] = nullptr;
2291 // Do a limited scalarization of any live at safepoint vector values which
2292 // contain pointers. This enables this pass to run after vectorization at
2293 // the cost of some possible performance loss. TODO: it would be nice to
2294 // natively support vectors all the way through the backend so we don't need
2295 // to scalarize here.
2296 for (size_t i = 0; i < records.size(); i++) {
2297 struct PartiallyConstructedSafepointRecord &info = records[i];
2298 Instruction *statepoint = toUpdate[i].getInstruction();
2299 splitVectorValues(cast<Instruction>(statepoint), info.liveset,
2300 info.PointerToBase, DT);
2303 // In order to reduce live set of statepoint we might choose to rematerialize
2304 // some values instead of relocating them. This is purely an optimization and
2305 // does not influence correctness.
2306 TargetTransformInfo &TTI =
2307 P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2309 for (size_t i = 0; i < records.size(); i++) {
2310 struct PartiallyConstructedSafepointRecord &info = records[i];
2311 CallSite &CS = toUpdate[i];
2313 rematerializeLiveValues(CS, info, TTI);
2316 // Now run through and replace the existing statepoints with new ones with
2317 // the live variables listed. We do not yet update uses of the values being
2318 // relocated. We have references to live variables that need to
2319 // survive to the last iteration of this loop. (By construction, the
2320 // previous statepoint can not be a live variable, thus we can and remove
2321 // the old statepoint calls as we go.)
2322 for (size_t i = 0; i < records.size(); i++) {
2323 struct PartiallyConstructedSafepointRecord &info = records[i];
2324 CallSite &CS = toUpdate[i];
2325 makeStatepointExplicit(DT, CS, P, info);
2327 toUpdate.clear(); // prevent accident use of invalid CallSites
2329 // Do all the fixups of the original live variables to their relocated selves
2330 SmallVector<Value *, 128> live;
2331 for (size_t i = 0; i < records.size(); i++) {
2332 struct PartiallyConstructedSafepointRecord &info = records[i];
2333 // We can't simply save the live set from the original insertion. One of
2334 // the live values might be the result of a call which needs a safepoint.
2335 // That Value* no longer exists and we need to use the new gc_result.
2336 // Thankfully, the liveset is embedded in the statepoint (and updated), so
2337 // we just grab that.
2338 Statepoint statepoint(info.StatepointToken);
2339 live.insert(live.end(), statepoint.gc_args_begin(),
2340 statepoint.gc_args_end());
2342 // Do some basic sanity checks on our liveness results before performing
2343 // relocation. Relocation can and will turn mistakes in liveness results
2344 // into non-sensical code which is must harder to debug.
2345 // TODO: It would be nice to test consistency as well
2346 assert(DT.isReachableFromEntry(info.StatepointToken->getParent()) &&
2347 "statepoint must be reachable or liveness is meaningless");
2348 for (Value *V : statepoint.gc_args()) {
2349 if (!isa<Instruction>(V))
2350 // Non-instruction values trivial dominate all possible uses
2352 auto LiveInst = cast<Instruction>(V);
2353 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2354 "unreachable values should never be live");
2355 assert(DT.dominates(LiveInst, info.StatepointToken) &&
2356 "basic SSA liveness expectation violated by liveness analysis");
2360 unique_unsorted(live);
2364 for (auto ptr : live) {
2365 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
2369 relocationViaAlloca(F, DT, live, records);
2370 return !records.empty();
2373 // Handles both return values and arguments for Functions and CallSites.
2374 template <typename AttrHolder>
2375 static void RemoveDerefAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2378 if (AH.getDereferenceableBytes(Index))
2379 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2380 AH.getDereferenceableBytes(Index)));
2381 if (AH.getDereferenceableOrNullBytes(Index))
2382 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2383 AH.getDereferenceableOrNullBytes(Index)));
2386 AH.setAttributes(AH.getAttributes().removeAttributes(
2387 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2391 RewriteStatepointsForGC::stripDereferenceabilityInfoFromPrototype(Function &F) {
2392 LLVMContext &Ctx = F.getContext();
2394 for (Argument &A : F.args())
2395 if (isa<PointerType>(A.getType()))
2396 RemoveDerefAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2398 if (isa<PointerType>(F.getReturnType()))
2399 RemoveDerefAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2402 void RewriteStatepointsForGC::stripDereferenceabilityInfoFromBody(Function &F) {
2406 LLVMContext &Ctx = F.getContext();
2407 MDBuilder Builder(Ctx);
2409 for (Instruction &I : instructions(F)) {
2410 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2411 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2412 bool IsImmutableTBAA =
2413 MD->getNumOperands() == 4 &&
2414 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2416 if (!IsImmutableTBAA)
2417 continue; // no work to do, MD_tbaa is already marked mutable
2419 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2420 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2422 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2424 MDNode *MutableTBAA =
2425 Builder.createTBAAStructTagNode(Base, Access, Offset);
2426 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2429 if (CallSite CS = CallSite(&I)) {
2430 for (int i = 0, e = CS.arg_size(); i != e; i++)
2431 if (isa<PointerType>(CS.getArgument(i)->getType()))
2432 RemoveDerefAttrAtIndex(Ctx, CS, i + 1);
2433 if (isa<PointerType>(CS.getType()))
2434 RemoveDerefAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2439 /// Returns true if this function should be rewritten by this pass. The main
2440 /// point of this function is as an extension point for custom logic.
2441 static bool shouldRewriteStatepointsIn(Function &F) {
2442 // TODO: This should check the GCStrategy
2444 const char *FunctionGCName = F.getGC();
2445 const StringRef StatepointExampleName("statepoint-example");
2446 const StringRef CoreCLRName("coreclr");
2447 return (StatepointExampleName == FunctionGCName) ||
2448 (CoreCLRName == FunctionGCName);
2453 void RewriteStatepointsForGC::stripDereferenceabilityInfo(Module &M) {
2455 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2459 for (Function &F : M)
2460 stripDereferenceabilityInfoFromPrototype(F);
2462 for (Function &F : M)
2463 stripDereferenceabilityInfoFromBody(F);
2466 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2467 // Nothing to do for declarations.
2468 if (F.isDeclaration() || F.empty())
2471 // Policy choice says not to rewrite - the most common reason is that we're
2472 // compiling code without a GCStrategy.
2473 if (!shouldRewriteStatepointsIn(F))
2476 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2478 // Gather all the statepoints which need rewritten. Be careful to only
2479 // consider those in reachable code since we need to ask dominance queries
2480 // when rewriting. We'll delete the unreachable ones in a moment.
2481 SmallVector<CallSite, 64> ParsePointNeeded;
2482 bool HasUnreachableStatepoint = false;
2483 for (Instruction &I : instructions(F)) {
2484 // TODO: only the ones with the flag set!
2485 if (isStatepoint(I)) {
2486 if (DT.isReachableFromEntry(I.getParent()))
2487 ParsePointNeeded.push_back(CallSite(&I));
2489 HasUnreachableStatepoint = true;
2493 bool MadeChange = false;
2495 // Delete any unreachable statepoints so that we don't have unrewritten
2496 // statepoints surviving this pass. This makes testing easier and the
2497 // resulting IR less confusing to human readers. Rather than be fancy, we
2498 // just reuse a utility function which removes the unreachable blocks.
2499 if (HasUnreachableStatepoint)
2500 MadeChange |= removeUnreachableBlocks(F);
2502 // Return early if no work to do.
2503 if (ParsePointNeeded.empty())
2506 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2507 // These are created by LCSSA. They have the effect of increasing the size
2508 // of liveness sets for no good reason. It may be harder to do this post
2509 // insertion since relocations and base phis can confuse things.
2510 for (BasicBlock &BB : F)
2511 if (BB.getUniquePredecessor()) {
2513 FoldSingleEntryPHINodes(&BB);
2516 // Before we start introducing relocations, we want to tweak the IR a bit to
2517 // avoid unfortunate code generation effects. The main example is that we
2518 // want to try to make sure the comparison feeding a branch is after any
2519 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2520 // values feeding a branch after relocation. This is semantically correct,
2521 // but results in extra register pressure since both the pre-relocation and
2522 // post-relocation copies must be available in registers. For code without
2523 // relocations this is handled elsewhere, but teaching the scheduler to
2524 // reverse the transform we're about to do would be slightly complex.
2525 // Note: This may extend the live range of the inputs to the icmp and thus
2526 // increase the liveset of any statepoint we move over. This is profitable
2527 // as long as all statepoints are in rare blocks. If we had in-register
2528 // lowering for live values this would be a much safer transform.
2529 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2530 if (auto *BI = dyn_cast<BranchInst>(TI))
2531 if (BI->isConditional())
2532 return dyn_cast<Instruction>(BI->getCondition());
2533 // TODO: Extend this to handle switches
2536 for (BasicBlock &BB : F) {
2537 TerminatorInst *TI = BB.getTerminator();
2538 if (auto *Cond = getConditionInst(TI))
2539 // TODO: Handle more than just ICmps here. We should be able to move
2540 // most instructions without side effects or memory access.
2541 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2543 Cond->moveBefore(TI);
2547 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
2551 // liveness computation via standard dataflow
2552 // -------------------------------------------------------------------
2554 // TODO: Consider using bitvectors for liveness, the set of potentially
2555 // interesting values should be small and easy to pre-compute.
2557 /// Compute the live-in set for the location rbegin starting from
2558 /// the live-out set of the basic block
2559 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2560 BasicBlock::reverse_iterator rend,
2561 DenseSet<Value *> &LiveTmp) {
2563 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2564 Instruction *I = &*ritr;
2566 // KILL/Def - Remove this definition from LiveIn
2569 // Don't consider *uses* in PHI nodes, we handle their contribution to
2570 // predecessor blocks when we seed the LiveOut sets
2571 if (isa<PHINode>(I))
2574 // USE - Add to the LiveIn set for this instruction
2575 for (Value *V : I->operands()) {
2576 assert(!isUnhandledGCPointerType(V->getType()) &&
2577 "support for FCA unimplemented");
2578 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2579 // The choice to exclude all things constant here is slightly subtle.
2580 // There are two independent reasons:
2581 // - We assume that things which are constant (from LLVM's definition)
2582 // do not move at runtime. For example, the address of a global
2583 // variable is fixed, even though it's contents may not be.
2584 // - Second, we can't disallow arbitrary inttoptr constants even
2585 // if the language frontend does. Optimization passes are free to
2586 // locally exploit facts without respect to global reachability. This
2587 // can create sections of code which are dynamically unreachable and
2588 // contain just about anything. (see constants.ll in tests)
2595 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2597 for (BasicBlock *Succ : successors(BB)) {
2598 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2599 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2600 PHINode *Phi = cast<PHINode>(&*I);
2601 Value *V = Phi->getIncomingValueForBlock(BB);
2602 assert(!isUnhandledGCPointerType(V->getType()) &&
2603 "support for FCA unimplemented");
2604 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2611 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2612 DenseSet<Value *> KillSet;
2613 for (Instruction &I : *BB)
2614 if (isHandledGCPointerType(I.getType()))
2620 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2621 /// sanity check for the liveness computation.
2622 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2623 TerminatorInst *TI, bool TermOkay = false) {
2624 for (Value *V : Live) {
2625 if (auto *I = dyn_cast<Instruction>(V)) {
2626 // The terminator can be a member of the LiveOut set. LLVM's definition
2627 // of instruction dominance states that V does not dominate itself. As
2628 // such, we need to special case this to allow it.
2629 if (TermOkay && TI == I)
2631 assert(DT.dominates(I, TI) &&
2632 "basic SSA liveness expectation violated by liveness analysis");
2637 /// Check that all the liveness sets used during the computation of liveness
2638 /// obey basic SSA properties. This is useful for finding cases where we miss
2640 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2642 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2643 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2644 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2648 static void computeLiveInValues(DominatorTree &DT, Function &F,
2649 GCPtrLivenessData &Data) {
2651 SmallSetVector<BasicBlock *, 200> Worklist;
2652 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2653 // We use a SetVector so that we don't have duplicates in the worklist.
2654 Worklist.insert(pred_begin(BB), pred_end(BB));
2656 auto NextItem = [&]() {
2657 BasicBlock *BB = Worklist.back();
2658 Worklist.pop_back();
2662 // Seed the liveness for each individual block
2663 for (BasicBlock &BB : F) {
2664 Data.KillSet[&BB] = computeKillSet(&BB);
2665 Data.LiveSet[&BB].clear();
2666 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2669 for (Value *Kill : Data.KillSet[&BB])
2670 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2673 Data.LiveOut[&BB] = DenseSet<Value *>();
2674 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2675 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2676 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2677 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2678 if (!Data.LiveIn[&BB].empty())
2679 AddPredsToWorklist(&BB);
2682 // Propagate that liveness until stable
2683 while (!Worklist.empty()) {
2684 BasicBlock *BB = NextItem();
2686 // Compute our new liveout set, then exit early if it hasn't changed
2687 // despite the contribution of our successor.
2688 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2689 const auto OldLiveOutSize = LiveOut.size();
2690 for (BasicBlock *Succ : successors(BB)) {
2691 assert(Data.LiveIn.count(Succ));
2692 set_union(LiveOut, Data.LiveIn[Succ]);
2694 // assert OutLiveOut is a subset of LiveOut
2695 if (OldLiveOutSize == LiveOut.size()) {
2696 // If the sets are the same size, then we didn't actually add anything
2697 // when unioning our successors LiveIn Thus, the LiveIn of this block
2701 Data.LiveOut[BB] = LiveOut;
2703 // Apply the effects of this basic block
2704 DenseSet<Value *> LiveTmp = LiveOut;
2705 set_union(LiveTmp, Data.LiveSet[BB]);
2706 set_subtract(LiveTmp, Data.KillSet[BB]);
2708 assert(Data.LiveIn.count(BB));
2709 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2710 // assert: OldLiveIn is a subset of LiveTmp
2711 if (OldLiveIn.size() != LiveTmp.size()) {
2712 Data.LiveIn[BB] = LiveTmp;
2713 AddPredsToWorklist(BB);
2715 } // while( !worklist.empty() )
2718 // Sanity check our output against SSA properties. This helps catch any
2719 // missing kills during the above iteration.
2720 for (BasicBlock &BB : F) {
2721 checkBasicSSA(DT, Data, BB);
2726 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2727 StatepointLiveSetTy &Out) {
2729 BasicBlock *BB = Inst->getParent();
2731 // Note: The copy is intentional and required
2732 assert(Data.LiveOut.count(BB));
2733 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2735 // We want to handle the statepoint itself oddly. It's
2736 // call result is not live (normal), nor are it's arguments
2737 // (unless they're used again later). This adjustment is
2738 // specifically what we need to relocate
2739 BasicBlock::reverse_iterator rend(Inst);
2740 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2741 LiveOut.erase(Inst);
2742 Out.insert(LiveOut.begin(), LiveOut.end());
2745 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2747 PartiallyConstructedSafepointRecord &Info) {
2748 Instruction *Inst = CS.getInstruction();
2749 StatepointLiveSetTy Updated;
2750 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2753 DenseSet<Value *> Bases;
2754 for (auto KVPair : Info.PointerToBase) {
2755 Bases.insert(KVPair.second);
2758 // We may have base pointers which are now live that weren't before. We need
2759 // to update the PointerToBase structure to reflect this.
2760 for (auto V : Updated)
2761 if (!Info.PointerToBase.count(V)) {
2762 assert(Bases.count(V) && "can't find base for unexpected live value");
2763 Info.PointerToBase[V] = V;
2768 for (auto V : Updated) {
2769 assert(Info.PointerToBase.count(V) &&
2770 "must be able to find base for live value");
2774 // Remove any stale base mappings - this can happen since our liveness is
2775 // more precise then the one inherent in the base pointer analysis
2776 DenseSet<Value *> ToErase;
2777 for (auto KVPair : Info.PointerToBase)
2778 if (!Updated.count(KVPair.first))
2779 ToErase.insert(KVPair.first);
2780 for (auto V : ToErase)
2781 Info.PointerToBase.erase(V);
2784 for (auto KVPair : Info.PointerToBase)
2785 assert(Updated.count(KVPair.first) && "record for non-live value");
2788 Info.liveset = Updated;