1 //===- RewriteStatepointsForGC.cpp - Make GC relocations explicit ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Rewrite an existing set of gc.statepoints such that they make potential
11 // relocations performed by the garbage collector explicit in the IR.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Pass.h"
16 #include "llvm/Analysis/CFG.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/TargetTransformInfo.h"
19 #include "llvm/ADT/SetOperations.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/ADT/DenseSet.h"
22 #include "llvm/ADT/SetVector.h"
23 #include "llvm/ADT/StringRef.h"
24 #include "llvm/ADT/MapVector.h"
25 #include "llvm/IR/BasicBlock.h"
26 #include "llvm/IR/CallSite.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/IRBuilder.h"
30 #include "llvm/IR/InstIterator.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/IR/MDBuilder.h"
36 #include "llvm/IR/Statepoint.h"
37 #include "llvm/IR/Value.h"
38 #include "llvm/IR/Verifier.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Transforms/Scalar.h"
42 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 #include "llvm/Transforms/Utils/Cloning.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
47 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
51 // Print the liveset found at the insert location
52 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
54 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden,
56 // Print out the base pointers for debugging
57 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden,
60 // Cost threshold measuring when it is profitable to rematerialize value instead
62 static cl::opt<unsigned>
63 RematerializationThreshold("spp-rematerialization-threshold", cl::Hidden,
67 static bool ClobberNonLive = true;
69 static bool ClobberNonLive = false;
71 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
72 cl::location(ClobberNonLive),
75 static cl::opt<bool> UseDeoptBundles("rs4gc-use-deopt-bundles", cl::Hidden,
78 AllowStatepointWithNoDeoptInfo("rs4gc-allow-statepoint-with-no-deopt-info",
79 cl::Hidden, cl::init(true));
81 /// Should we split vectors of pointers into their individual elements? This
82 /// is known to be buggy, but the alternate implementation isn't yet ready.
83 /// This is purely to provide a debugging and dianostic hook until the vector
84 /// split is replaced with vector relocations.
85 static cl::opt<bool> UseVectorSplit("rs4gc-split-vector-values", cl::Hidden,
89 struct RewriteStatepointsForGC : public ModulePass {
90 static char ID; // Pass identification, replacement for typeid
92 RewriteStatepointsForGC() : ModulePass(ID) {
93 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
95 bool runOnFunction(Function &F);
96 bool runOnModule(Module &M) override {
99 Changed |= runOnFunction(F);
102 // stripNonValidAttributes asserts that shouldRewriteStatepointsIn
103 // returns true for at least one function in the module. Since at least
104 // one function changed, we know that the precondition is satisfied.
105 stripNonValidAttributes(M);
111 void getAnalysisUsage(AnalysisUsage &AU) const override {
112 // We add and rewrite a bunch of instructions, but don't really do much
113 // else. We could in theory preserve a lot more analyses here.
114 AU.addRequired<DominatorTreeWrapperPass>();
115 AU.addRequired<TargetTransformInfoWrapperPass>();
118 /// The IR fed into RewriteStatepointsForGC may have had attributes implying
119 /// dereferenceability that are no longer valid/correct after
120 /// RewriteStatepointsForGC has run. This is because semantically, after
121 /// RewriteStatepointsForGC runs, all calls to gc.statepoint "free" the entire
122 /// heap. stripNonValidAttributes (conservatively) restores correctness
123 /// by erasing all attributes in the module that externally imply
124 /// dereferenceability.
125 /// Similar reasoning also applies to the noalias attributes. gc.statepoint
126 /// can touch the entire heap including noalias objects.
127 void stripNonValidAttributes(Module &M);
129 // Helpers for stripNonValidAttributes
130 void stripNonValidAttributesFromBody(Function &F);
131 void stripNonValidAttributesFromPrototype(Function &F);
135 char RewriteStatepointsForGC::ID = 0;
137 ModulePass *llvm::createRewriteStatepointsForGCPass() {
138 return new RewriteStatepointsForGC();
141 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
142 "Make relocations explicit at statepoints", false, false)
143 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
144 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
145 "Make relocations explicit at statepoints", false, false)
148 struct GCPtrLivenessData {
149 /// Values defined in this block.
150 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet;
151 /// Values used in this block (and thus live); does not included values
152 /// killed within this block.
153 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet;
155 /// Values live into this basic block (i.e. used by any
156 /// instruction in this basic block or ones reachable from here)
157 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn;
159 /// Values live out of this basic block (i.e. live into
160 /// any successor block)
161 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut;
164 // The type of the internal cache used inside the findBasePointers family
165 // of functions. From the callers perspective, this is an opaque type and
166 // should not be inspected.
168 // In the actual implementation this caches two relations:
169 // - The base relation itself (i.e. this pointer is based on that one)
170 // - The base defining value relation (i.e. before base_phi insertion)
171 // Generally, after the execution of a full findBasePointer call, only the
172 // base relation will remain. Internally, we add a mixture of the two
173 // types, then update all the second type to the first type
174 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
175 typedef DenseSet<Value *> StatepointLiveSetTy;
176 typedef DenseMap<AssertingVH<Instruction>, AssertingVH<Value>>
177 RematerializedValueMapTy;
179 struct PartiallyConstructedSafepointRecord {
180 /// The set of values known to be live across this safepoint
181 StatepointLiveSetTy LiveSet;
183 /// Mapping from live pointers to a base-defining-value
184 DenseMap<Value *, Value *> PointerToBase;
186 /// The *new* gc.statepoint instruction itself. This produces the token
187 /// that normal path gc.relocates and the gc.result are tied to.
188 Instruction *StatepointToken;
190 /// Instruction to which exceptional gc relocates are attached
191 /// Makes it easier to iterate through them during relocationViaAlloca.
192 Instruction *UnwindToken;
194 /// Record live values we are rematerialized instead of relocating.
195 /// They are not included into 'LiveSet' field.
196 /// Maps rematerialized copy to it's original value.
197 RematerializedValueMapTy RematerializedValues;
201 static ArrayRef<Use> GetDeoptBundleOperands(ImmutableCallSite CS) {
202 assert(UseDeoptBundles && "Should not be called otherwise!");
204 Optional<OperandBundleUse> DeoptBundle = CS.getOperandBundle("deopt");
206 if (!DeoptBundle.hasValue()) {
207 assert(AllowStatepointWithNoDeoptInfo &&
208 "Found non-leaf call without deopt info!");
212 return DeoptBundle.getValue().Inputs;
215 /// Compute the live-in set for every basic block in the function
216 static void computeLiveInValues(DominatorTree &DT, Function &F,
217 GCPtrLivenessData &Data);
219 /// Given results from the dataflow liveness computation, find the set of live
220 /// Values at a particular instruction.
221 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
222 StatepointLiveSetTy &out);
224 // TODO: Once we can get to the GCStrategy, this becomes
225 // Optional<bool> isGCManagedPointer(const Type *Ty) const override {
227 static bool isGCPointerType(Type *T) {
228 if (auto *PT = dyn_cast<PointerType>(T))
229 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
230 // GC managed heap. We know that a pointer into this heap needs to be
231 // updated and that no other pointer does.
232 return (1 == PT->getAddressSpace());
236 // Return true if this type is one which a) is a gc pointer or contains a GC
237 // pointer and b) is of a type this code expects to encounter as a live value.
238 // (The insertion code will assert that a type which matches (a) and not (b)
239 // is not encountered.)
240 static bool isHandledGCPointerType(Type *T) {
241 // We fully support gc pointers
242 if (isGCPointerType(T))
244 // We partially support vectors of gc pointers. The code will assert if it
245 // can't handle something.
246 if (auto VT = dyn_cast<VectorType>(T))
247 if (isGCPointerType(VT->getElementType()))
253 /// Returns true if this type contains a gc pointer whether we know how to
254 /// handle that type or not.
255 static bool containsGCPtrType(Type *Ty) {
256 if (isGCPointerType(Ty))
258 if (VectorType *VT = dyn_cast<VectorType>(Ty))
259 return isGCPointerType(VT->getScalarType());
260 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
261 return containsGCPtrType(AT->getElementType());
262 if (StructType *ST = dyn_cast<StructType>(Ty))
263 return std::any_of(ST->subtypes().begin(), ST->subtypes().end(),
268 // Returns true if this is a type which a) is a gc pointer or contains a GC
269 // pointer and b) is of a type which the code doesn't expect (i.e. first class
270 // aggregates). Used to trip assertions.
271 static bool isUnhandledGCPointerType(Type *Ty) {
272 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
276 static bool order_by_name(Value *a, Value *b) {
277 if (a->hasName() && b->hasName()) {
278 return -1 == a->getName().compare(b->getName());
279 } else if (a->hasName() && !b->hasName()) {
281 } else if (!a->hasName() && b->hasName()) {
284 // Better than nothing, but not stable
289 // Return the name of the value suffixed with the provided value, or if the
290 // value didn't have a name, the default value specified.
291 static std::string suffixed_name_or(Value *V, StringRef Suffix,
292 StringRef DefaultName) {
293 return V->hasName() ? (V->getName() + Suffix).str() : DefaultName.str();
296 // Conservatively identifies any definitions which might be live at the
297 // given instruction. The analysis is performed immediately before the
298 // given instruction. Values defined by that instruction are not considered
299 // live. Values used by that instruction are considered live.
300 static void analyzeParsePointLiveness(
301 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
302 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
303 Instruction *inst = CS.getInstruction();
305 StatepointLiveSetTy LiveSet;
306 findLiveSetAtInst(inst, OriginalLivenessData, LiveSet);
309 // Note: This output is used by several of the test cases
310 // The order of elements in a set is not stable, put them in a vec and sort
312 SmallVector<Value *, 64> Temp;
313 Temp.insert(Temp.end(), LiveSet.begin(), LiveSet.end());
314 std::sort(Temp.begin(), Temp.end(), order_by_name);
315 errs() << "Live Variables:\n";
316 for (Value *V : Temp)
317 dbgs() << " " << V->getName() << " " << *V << "\n";
319 if (PrintLiveSetSize) {
320 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
321 errs() << "Number live values: " << LiveSet.size() << "\n";
323 result.LiveSet = LiveSet;
326 static bool isKnownBaseResult(Value *V);
328 /// A single base defining value - An immediate base defining value for an
329 /// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'.
330 /// For instructions which have multiple pointer [vector] inputs or that
331 /// transition between vector and scalar types, there is no immediate base
332 /// defining value. The 'base defining value' for 'Def' is the transitive
333 /// closure of this relation stopping at the first instruction which has no
334 /// immediate base defining value. The b.d.v. might itself be a base pointer,
335 /// but it can also be an arbitrary derived pointer.
336 struct BaseDefiningValueResult {
337 /// Contains the value which is the base defining value.
339 /// True if the base defining value is also known to be an actual base
341 const bool IsKnownBase;
342 BaseDefiningValueResult(Value *BDV, bool IsKnownBase)
343 : BDV(BDV), IsKnownBase(IsKnownBase) {
345 // Check consistency between new and old means of checking whether a BDV is
347 bool MustBeBase = isKnownBaseResult(BDV);
348 assert(!MustBeBase || MustBeBase == IsKnownBase);
354 static BaseDefiningValueResult findBaseDefiningValue(Value *I);
356 /// Return a base defining value for the 'Index' element of the given vector
357 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
358 /// 'I'. As an optimization, this method will try to determine when the
359 /// element is known to already be a base pointer. If this can be established,
360 /// the second value in the returned pair will be true. Note that either a
361 /// vector or a pointer typed value can be returned. For the former, the
362 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
363 /// If the later, the return pointer is a BDV (or possibly a base) for the
364 /// particular element in 'I'.
365 static BaseDefiningValueResult
366 findBaseDefiningValueOfVector(Value *I) {
367 assert(I->getType()->isVectorTy() &&
368 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
369 "Illegal to ask for the base pointer of a non-pointer type");
371 // Each case parallels findBaseDefiningValue below, see that code for
372 // detailed motivation.
374 if (isa<Argument>(I))
375 // An incoming argument to the function is a base pointer
376 return BaseDefiningValueResult(I, true);
378 // We shouldn't see the address of a global as a vector value?
379 assert(!isa<GlobalVariable>(I) &&
380 "unexpected global variable found in base of vector");
382 // inlining could possibly introduce phi node that contains
383 // undef if callee has multiple returns
384 if (isa<UndefValue>(I))
385 // utterly meaningless, but useful for dealing with partially optimized
387 return BaseDefiningValueResult(I, true);
389 // Due to inheritance, this must be _after_ the global variable and undef
391 if (Constant *Con = dyn_cast<Constant>(I)) {
392 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
393 "order of checks wrong!");
394 assert(Con->isNullValue() && "null is the only case which makes sense");
395 return BaseDefiningValueResult(Con, true);
398 if (isa<LoadInst>(I))
399 return BaseDefiningValueResult(I, true);
401 if (isa<InsertElementInst>(I))
402 // We don't know whether this vector contains entirely base pointers or
403 // not. To be conservatively correct, we treat it as a BDV and will
404 // duplicate code as needed to construct a parallel vector of bases.
405 return BaseDefiningValueResult(I, false);
407 if (isa<ShuffleVectorInst>(I))
408 // We don't know whether this vector contains entirely base pointers or
409 // not. To be conservatively correct, we treat it as a BDV and will
410 // duplicate code as needed to construct a parallel vector of bases.
411 // TODO: There a number of local optimizations which could be applied here
412 // for particular sufflevector patterns.
413 return BaseDefiningValueResult(I, false);
415 // A PHI or Select is a base defining value. The outer findBasePointer
416 // algorithm is responsible for constructing a base value for this BDV.
417 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
418 "unknown vector instruction - no base found for vector element");
419 return BaseDefiningValueResult(I, false);
422 /// Helper function for findBasePointer - Will return a value which either a)
423 /// defines the base pointer for the input, b) blocks the simple search
424 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
425 /// from pointer to vector type or back.
426 static BaseDefiningValueResult findBaseDefiningValue(Value *I) {
427 if (I->getType()->isVectorTy())
428 return findBaseDefiningValueOfVector(I);
430 assert(I->getType()->isPointerTy() &&
431 "Illegal to ask for the base pointer of a non-pointer type");
433 if (isa<Argument>(I))
434 // An incoming argument to the function is a base pointer
435 // We should have never reached here if this argument isn't an gc value
436 return BaseDefiningValueResult(I, true);
438 if (isa<Constant>(I))
439 // We assume that objects with a constant base (e.g. a global) can't move
440 // and don't need to be reported to the collector because they are always
441 // live. All constants have constant bases. Besides global references, all
442 // kinds of constants (e.g. undef, constant expressions, null pointers) can
443 // be introduced by the inliner or the optimizer, especially on dynamically
444 // dead paths. See e.g. test4 in constants.ll.
445 return BaseDefiningValueResult(I, true);
447 if (CastInst *CI = dyn_cast<CastInst>(I)) {
448 Value *Def = CI->stripPointerCasts();
449 // If stripping pointer casts changes the address space there is an
450 // addrspacecast in between.
451 assert(cast<PointerType>(Def->getType())->getAddressSpace() ==
452 cast<PointerType>(CI->getType())->getAddressSpace() &&
453 "unsupported addrspacecast");
454 // If we find a cast instruction here, it means we've found a cast which is
455 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
456 // handle int->ptr conversion.
457 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
458 return findBaseDefiningValue(Def);
461 if (isa<LoadInst>(I))
462 // The value loaded is an gc base itself
463 return BaseDefiningValueResult(I, true);
466 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
467 // The base of this GEP is the base
468 return findBaseDefiningValue(GEP->getPointerOperand());
470 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
471 switch (II->getIntrinsicID()) {
473 // fall through to general call handling
475 case Intrinsic::experimental_gc_statepoint:
476 llvm_unreachable("statepoints don't produce pointers");
477 case Intrinsic::experimental_gc_relocate: {
478 // Rerunning safepoint insertion after safepoints are already
479 // inserted is not supported. It could probably be made to work,
480 // but why are you doing this? There's no good reason.
481 llvm_unreachable("repeat safepoint insertion is not supported");
483 case Intrinsic::gcroot:
484 // Currently, this mechanism hasn't been extended to work with gcroot.
485 // There's no reason it couldn't be, but I haven't thought about the
486 // implications much.
488 "interaction with the gcroot mechanism is not supported");
491 // We assume that functions in the source language only return base
492 // pointers. This should probably be generalized via attributes to support
493 // both source language and internal functions.
494 if (isa<CallInst>(I) || isa<InvokeInst>(I))
495 return BaseDefiningValueResult(I, true);
497 // I have absolutely no idea how to implement this part yet. It's not
498 // necessarily hard, I just haven't really looked at it yet.
499 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
501 if (isa<AtomicCmpXchgInst>(I))
502 // A CAS is effectively a atomic store and load combined under a
503 // predicate. From the perspective of base pointers, we just treat it
505 return BaseDefiningValueResult(I, true);
507 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
508 "binary ops which don't apply to pointers");
510 // The aggregate ops. Aggregates can either be in the heap or on the
511 // stack, but in either case, this is simply a field load. As a result,
512 // this is a defining definition of the base just like a load is.
513 if (isa<ExtractValueInst>(I))
514 return BaseDefiningValueResult(I, true);
516 // We should never see an insert vector since that would require we be
517 // tracing back a struct value not a pointer value.
518 assert(!isa<InsertValueInst>(I) &&
519 "Base pointer for a struct is meaningless");
521 // An extractelement produces a base result exactly when it's input does.
522 // We may need to insert a parallel instruction to extract the appropriate
523 // element out of the base vector corresponding to the input. Given this,
524 // it's analogous to the phi and select case even though it's not a merge.
525 if (isa<ExtractElementInst>(I))
526 // Note: There a lot of obvious peephole cases here. This are deliberately
527 // handled after the main base pointer inference algorithm to make writing
528 // test cases to exercise that code easier.
529 return BaseDefiningValueResult(I, false);
531 // The last two cases here don't return a base pointer. Instead, they
532 // return a value which dynamically selects from among several base
533 // derived pointers (each with it's own base potentially). It's the job of
534 // the caller to resolve these.
535 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
536 "missing instruction case in findBaseDefiningValing");
537 return BaseDefiningValueResult(I, false);
540 /// Returns the base defining value for this value.
541 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
542 Value *&Cached = Cache[I];
544 Cached = findBaseDefiningValue(I).BDV;
545 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
546 << Cached->getName() << "\n");
548 assert(Cache[I] != nullptr);
552 /// Return a base pointer for this value if known. Otherwise, return it's
553 /// base defining value.
554 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
555 Value *Def = findBaseDefiningValueCached(I, Cache);
556 auto Found = Cache.find(Def);
557 if (Found != Cache.end()) {
558 // Either a base-of relation, or a self reference. Caller must check.
559 return Found->second;
561 // Only a BDV available
565 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
566 /// is it known to be a base pointer? Or do we need to continue searching.
567 static bool isKnownBaseResult(Value *V) {
568 if (!isa<PHINode>(V) && !isa<SelectInst>(V) &&
569 !isa<ExtractElementInst>(V) && !isa<InsertElementInst>(V) &&
570 !isa<ShuffleVectorInst>(V)) {
571 // no recursion possible
574 if (isa<Instruction>(V) &&
575 cast<Instruction>(V)->getMetadata("is_base_value")) {
576 // This is a previously inserted base phi or select. We know
577 // that this is a base value.
581 // We need to keep searching
586 /// Models the state of a single base defining value in the findBasePointer
587 /// algorithm for determining where a new instruction is needed to propagate
588 /// the base of this BDV.
591 enum Status { Unknown, Base, Conflict };
593 BDVState(Status s, Value *b = nullptr) : status(s), base(b) {
594 assert(status != Base || b);
596 explicit BDVState(Value *b) : status(Base), base(b) {}
597 BDVState() : status(Unknown), base(nullptr) {}
599 Status getStatus() const { return status; }
600 Value *getBase() const { return base; }
602 bool isBase() const { return getStatus() == Base; }
603 bool isUnknown() const { return getStatus() == Unknown; }
604 bool isConflict() const { return getStatus() == Conflict; }
606 bool operator==(const BDVState &other) const {
607 return base == other.base && status == other.status;
610 bool operator!=(const BDVState &other) const { return !(*this == other); }
613 void dump() const { print(dbgs()); dbgs() << '\n'; }
615 void print(raw_ostream &OS) const {
627 OS << " (" << base << " - "
628 << (base ? base->getName() : "nullptr") << "): ";
633 AssertingVH<Value> base; // non null only if status == base
638 static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
645 // Values of type BDVState form a lattice, and this is a helper
646 // class that implementes the meet operation. The meat of the meet
647 // operation is implemented in MeetBDVStates::pureMeet
648 class MeetBDVStates {
650 /// Initializes the currentResult to the TOP state so that if can be met with
651 /// any other state to produce that state.
654 // Destructively meet the current result with the given BDVState
655 void meetWith(BDVState otherState) {
656 currentResult = meet(otherState, currentResult);
659 BDVState getResult() const { return currentResult; }
662 BDVState currentResult;
664 /// Perform a meet operation on two elements of the BDVState lattice.
665 static BDVState meet(BDVState LHS, BDVState RHS) {
666 assert((pureMeet(LHS, RHS) == pureMeet(RHS, LHS)) &&
667 "math is wrong: meet does not commute!");
668 BDVState Result = pureMeet(LHS, RHS);
669 DEBUG(dbgs() << "meet of " << LHS << " with " << RHS
670 << " produced " << Result << "\n");
674 static BDVState pureMeet(const BDVState &stateA, const BDVState &stateB) {
675 switch (stateA.getStatus()) {
676 case BDVState::Unknown:
680 assert(stateA.getBase() && "can't be null");
681 if (stateB.isUnknown())
684 if (stateB.isBase()) {
685 if (stateA.getBase() == stateB.getBase()) {
686 assert(stateA == stateB && "equality broken!");
689 return BDVState(BDVState::Conflict);
691 assert(stateB.isConflict() && "only three states!");
692 return BDVState(BDVState::Conflict);
694 case BDVState::Conflict:
697 llvm_unreachable("only three states!");
703 /// For a given value or instruction, figure out what base ptr it's derived
704 /// from. For gc objects, this is simply itself. On success, returns a value
705 /// which is the base pointer. (This is reliable and can be used for
706 /// relocation.) On failure, returns nullptr.
707 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
708 Value *def = findBaseOrBDV(I, cache);
710 if (isKnownBaseResult(def)) {
714 // Here's the rough algorithm:
715 // - For every SSA value, construct a mapping to either an actual base
716 // pointer or a PHI which obscures the base pointer.
717 // - Construct a mapping from PHI to unknown TOP state. Use an
718 // optimistic algorithm to propagate base pointer information. Lattice
723 // When algorithm terminates, all PHIs will either have a single concrete
724 // base or be in a conflict state.
725 // - For every conflict, insert a dummy PHI node without arguments. Add
726 // these to the base[Instruction] = BasePtr mapping. For every
727 // non-conflict, add the actual base.
728 // - For every conflict, add arguments for the base[a] of each input
731 // Note: A simpler form of this would be to add the conflict form of all
732 // PHIs without running the optimistic algorithm. This would be
733 // analogous to pessimistic data flow and would likely lead to an
734 // overall worse solution.
737 auto isExpectedBDVType = [](Value *BDV) {
738 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) ||
739 isa<ExtractElementInst>(BDV) || isa<InsertElementInst>(BDV);
743 // Once populated, will contain a mapping from each potentially non-base BDV
744 // to a lattice value (described above) which corresponds to that BDV.
745 // We use the order of insertion (DFS over the def/use graph) to provide a
746 // stable deterministic ordering for visiting DenseMaps (which are unordered)
747 // below. This is important for deterministic compilation.
748 MapVector<Value *, BDVState> States;
750 // Recursively fill in all base defining values reachable from the initial
751 // one for which we don't already know a definite base value for
753 SmallVector<Value*, 16> Worklist;
754 Worklist.push_back(def);
755 States.insert(std::make_pair(def, BDVState()));
756 while (!Worklist.empty()) {
757 Value *Current = Worklist.pop_back_val();
758 assert(!isKnownBaseResult(Current) && "why did it get added?");
760 auto visitIncomingValue = [&](Value *InVal) {
761 Value *Base = findBaseOrBDV(InVal, cache);
762 if (isKnownBaseResult(Base))
763 // Known bases won't need new instructions introduced and can be
766 assert(isExpectedBDVType(Base) && "the only non-base values "
767 "we see should be base defining values");
768 if (States.insert(std::make_pair(Base, BDVState())).second)
769 Worklist.push_back(Base);
771 if (PHINode *Phi = dyn_cast<PHINode>(Current)) {
772 for (Value *InVal : Phi->incoming_values())
773 visitIncomingValue(InVal);
774 } else if (SelectInst *Sel = dyn_cast<SelectInst>(Current)) {
775 visitIncomingValue(Sel->getTrueValue());
776 visitIncomingValue(Sel->getFalseValue());
777 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
778 visitIncomingValue(EE->getVectorOperand());
779 } else if (auto *IE = dyn_cast<InsertElementInst>(Current)) {
780 visitIncomingValue(IE->getOperand(0)); // vector operand
781 visitIncomingValue(IE->getOperand(1)); // scalar operand
783 // There is one known class of instructions we know we don't handle.
784 assert(isa<ShuffleVectorInst>(Current));
785 llvm_unreachable("unimplemented instruction case");
791 DEBUG(dbgs() << "States after initialization:\n");
792 for (auto Pair : States) {
793 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
797 // Return a phi state for a base defining value. We'll generate a new
798 // base state for known bases and expect to find a cached state otherwise.
799 auto getStateForBDV = [&](Value *baseValue) {
800 if (isKnownBaseResult(baseValue))
801 return BDVState(baseValue);
802 auto I = States.find(baseValue);
803 assert(I != States.end() && "lookup failed!");
807 bool progress = true;
810 const size_t oldSize = States.size();
813 // We're only changing values in this loop, thus safe to keep iterators.
814 // Since this is computing a fixed point, the order of visit does not
815 // effect the result. TODO: We could use a worklist here and make this run
817 for (auto Pair : States) {
818 Value *BDV = Pair.first;
819 assert(!isKnownBaseResult(BDV) && "why did it get added?");
821 // Given an input value for the current instruction, return a BDVState
822 // instance which represents the BDV of that value.
823 auto getStateForInput = [&](Value *V) mutable {
824 Value *BDV = findBaseOrBDV(V, cache);
825 return getStateForBDV(BDV);
828 MeetBDVStates calculateMeet;
829 if (SelectInst *select = dyn_cast<SelectInst>(BDV)) {
830 calculateMeet.meetWith(getStateForInput(select->getTrueValue()));
831 calculateMeet.meetWith(getStateForInput(select->getFalseValue()));
832 } else if (PHINode *Phi = dyn_cast<PHINode>(BDV)) {
833 for (Value *Val : Phi->incoming_values())
834 calculateMeet.meetWith(getStateForInput(Val));
835 } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) {
836 // The 'meet' for an extractelement is slightly trivial, but it's still
837 // useful in that it drives us to conflict if our input is.
838 calculateMeet.meetWith(getStateForInput(EE->getVectorOperand()));
840 // Given there's a inherent type mismatch between the operands, will
841 // *always* produce Conflict.
842 auto *IE = cast<InsertElementInst>(BDV);
843 calculateMeet.meetWith(getStateForInput(IE->getOperand(0)));
844 calculateMeet.meetWith(getStateForInput(IE->getOperand(1)));
847 BDVState oldState = States[BDV];
848 BDVState newState = calculateMeet.getResult();
849 if (oldState != newState) {
851 States[BDV] = newState;
855 assert(oldSize == States.size() &&
856 "fixed point shouldn't be adding any new nodes to state");
860 DEBUG(dbgs() << "States after meet iteration:\n");
861 for (auto Pair : States) {
862 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
866 // Insert Phis for all conflicts
867 // TODO: adjust naming patterns to avoid this order of iteration dependency
868 for (auto Pair : States) {
869 Instruction *I = cast<Instruction>(Pair.first);
870 BDVState State = Pair.second;
871 assert(!isKnownBaseResult(I) && "why did it get added?");
872 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
874 // extractelement instructions are a bit special in that we may need to
875 // insert an extract even when we know an exact base for the instruction.
876 // The problem is that we need to convert from a vector base to a scalar
877 // base for the particular indice we're interested in.
878 if (State.isBase() && isa<ExtractElementInst>(I) &&
879 isa<VectorType>(State.getBase()->getType())) {
880 auto *EE = cast<ExtractElementInst>(I);
881 // TODO: In many cases, the new instruction is just EE itself. We should
882 // exploit this, but can't do it here since it would break the invariant
883 // about the BDV not being known to be a base.
884 auto *BaseInst = ExtractElementInst::Create(State.getBase(),
885 EE->getIndexOperand(),
887 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
888 States[I] = BDVState(BDVState::Base, BaseInst);
891 // Since we're joining a vector and scalar base, they can never be the
892 // same. As a result, we should always see insert element having reached
893 // the conflict state.
894 if (isa<InsertElementInst>(I)) {
895 assert(State.isConflict());
898 if (!State.isConflict())
901 /// Create and insert a new instruction which will represent the base of
902 /// the given instruction 'I'.
903 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
904 if (isa<PHINode>(I)) {
905 BasicBlock *BB = I->getParent();
906 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
907 assert(NumPreds > 0 && "how did we reach here");
908 std::string Name = suffixed_name_or(I, ".base", "base_phi");
909 return PHINode::Create(I->getType(), NumPreds, Name, I);
910 } else if (SelectInst *Sel = dyn_cast<SelectInst>(I)) {
911 // The undef will be replaced later
912 UndefValue *Undef = UndefValue::get(Sel->getType());
913 std::string Name = suffixed_name_or(I, ".base", "base_select");
914 return SelectInst::Create(Sel->getCondition(), Undef,
916 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
917 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
918 std::string Name = suffixed_name_or(I, ".base", "base_ee");
919 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
922 auto *IE = cast<InsertElementInst>(I);
923 UndefValue *VecUndef = UndefValue::get(IE->getOperand(0)->getType());
924 UndefValue *ScalarUndef = UndefValue::get(IE->getOperand(1)->getType());
925 std::string Name = suffixed_name_or(I, ".base", "base_ie");
926 return InsertElementInst::Create(VecUndef, ScalarUndef,
927 IE->getOperand(2), Name, IE);
931 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
932 // Add metadata marking this as a base value
933 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
934 States[I] = BDVState(BDVState::Conflict, BaseInst);
937 // Returns a instruction which produces the base pointer for a given
938 // instruction. The instruction is assumed to be an input to one of the BDVs
939 // seen in the inference algorithm above. As such, we must either already
940 // know it's base defining value is a base, or have inserted a new
941 // instruction to propagate the base of it's BDV and have entered that newly
942 // introduced instruction into the state table. In either case, we are
943 // assured to be able to determine an instruction which produces it's base
945 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
946 Value *BDV = findBaseOrBDV(Input, cache);
947 Value *Base = nullptr;
948 if (isKnownBaseResult(BDV)) {
951 // Either conflict or base.
952 assert(States.count(BDV));
953 Base = States[BDV].getBase();
955 assert(Base && "can't be null");
956 // The cast is needed since base traversal may strip away bitcasts
957 if (Base->getType() != Input->getType() &&
959 Base = new BitCastInst(Base, Input->getType(), "cast",
965 // Fixup all the inputs of the new PHIs. Visit order needs to be
966 // deterministic and predictable because we're naming newly created
968 for (auto Pair : States) {
969 Instruction *BDV = cast<Instruction>(Pair.first);
970 BDVState State = Pair.second;
972 assert(!isKnownBaseResult(BDV) && "why did it get added?");
973 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
974 if (!State.isConflict())
977 if (PHINode *basephi = dyn_cast<PHINode>(State.getBase())) {
978 PHINode *phi = cast<PHINode>(BDV);
979 unsigned NumPHIValues = phi->getNumIncomingValues();
980 for (unsigned i = 0; i < NumPHIValues; i++) {
981 Value *InVal = phi->getIncomingValue(i);
982 BasicBlock *InBB = phi->getIncomingBlock(i);
984 // If we've already seen InBB, add the same incoming value
985 // we added for it earlier. The IR verifier requires phi
986 // nodes with multiple entries from the same basic block
987 // to have the same incoming value for each of those
988 // entries. If we don't do this check here and basephi
989 // has a different type than base, we'll end up adding two
990 // bitcasts (and hence two distinct values) as incoming
991 // values for the same basic block.
993 int blockIndex = basephi->getBasicBlockIndex(InBB);
994 if (blockIndex != -1) {
995 Value *oldBase = basephi->getIncomingValue(blockIndex);
996 basephi->addIncoming(oldBase, InBB);
999 Value *Base = getBaseForInput(InVal, nullptr);
1000 // In essence this assert states: the only way two
1001 // values incoming from the same basic block may be
1002 // different is by being different bitcasts of the same
1003 // value. A cleanup that remains TODO is changing
1004 // findBaseOrBDV to return an llvm::Value of the correct
1005 // type (and still remain pure). This will remove the
1006 // need to add bitcasts.
1007 assert(Base->stripPointerCasts() == oldBase->stripPointerCasts() &&
1008 "sanity -- findBaseOrBDV should be pure!");
1013 // Find the instruction which produces the base for each input. We may
1014 // need to insert a bitcast in the incoming block.
1015 // TODO: Need to split critical edges if insertion is needed
1016 Value *Base = getBaseForInput(InVal, InBB->getTerminator());
1017 basephi->addIncoming(Base, InBB);
1019 assert(basephi->getNumIncomingValues() == NumPHIValues);
1020 } else if (SelectInst *BaseSel = dyn_cast<SelectInst>(State.getBase())) {
1021 SelectInst *Sel = cast<SelectInst>(BDV);
1022 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
1023 // something more safe and less hacky.
1024 for (int i = 1; i <= 2; i++) {
1025 Value *InVal = Sel->getOperand(i);
1026 // Find the instruction which produces the base for each input. We may
1027 // need to insert a bitcast.
1028 Value *Base = getBaseForInput(InVal, BaseSel);
1029 BaseSel->setOperand(i, Base);
1031 } else if (auto *BaseEE = dyn_cast<ExtractElementInst>(State.getBase())) {
1032 Value *InVal = cast<ExtractElementInst>(BDV)->getVectorOperand();
1033 // Find the instruction which produces the base for each input. We may
1034 // need to insert a bitcast.
1035 Value *Base = getBaseForInput(InVal, BaseEE);
1036 BaseEE->setOperand(0, Base);
1038 auto *BaseIE = cast<InsertElementInst>(State.getBase());
1039 auto *BdvIE = cast<InsertElementInst>(BDV);
1040 auto UpdateOperand = [&](int OperandIdx) {
1041 Value *InVal = BdvIE->getOperand(OperandIdx);
1042 Value *Base = getBaseForInput(InVal, BaseIE);
1043 BaseIE->setOperand(OperandIdx, Base);
1045 UpdateOperand(0); // vector operand
1046 UpdateOperand(1); // scalar operand
1051 // Now that we're done with the algorithm, see if we can optimize the
1052 // results slightly by reducing the number of new instructions needed.
1053 // Arguably, this should be integrated into the algorithm above, but
1054 // doing as a post process step is easier to reason about for the moment.
1055 DenseMap<Value *, Value *> ReverseMap;
1056 SmallPtrSet<Instruction *, 16> NewInsts;
1057 SmallSetVector<AssertingVH<Instruction>, 16> Worklist;
1058 // Note: We need to visit the states in a deterministic order. We uses the
1059 // Keys we sorted above for this purpose. Note that we are papering over a
1060 // bigger problem with the algorithm above - it's visit order is not
1061 // deterministic. A larger change is needed to fix this.
1062 for (auto Pair : States) {
1063 auto *BDV = Pair.first;
1064 auto State = Pair.second;
1065 Value *Base = State.getBase();
1066 assert(BDV && Base);
1067 assert(!isKnownBaseResult(BDV) && "why did it get added?");
1068 assert(isKnownBaseResult(Base) &&
1069 "must be something we 'know' is a base pointer");
1070 if (!State.isConflict())
1073 ReverseMap[Base] = BDV;
1074 if (auto *BaseI = dyn_cast<Instruction>(Base)) {
1075 NewInsts.insert(BaseI);
1076 Worklist.insert(BaseI);
1079 auto ReplaceBaseInstWith = [&](Value *BDV, Instruction *BaseI,
1080 Value *Replacement) {
1081 // Add users which are new instructions (excluding self references)
1082 for (User *U : BaseI->users())
1083 if (auto *UI = dyn_cast<Instruction>(U))
1084 if (NewInsts.count(UI) && UI != BaseI)
1085 Worklist.insert(UI);
1086 // Then do the actual replacement
1087 NewInsts.erase(BaseI);
1088 ReverseMap.erase(BaseI);
1089 BaseI->replaceAllUsesWith(Replacement);
1090 assert(States.count(BDV));
1091 assert(States[BDV].isConflict() && States[BDV].getBase() == BaseI);
1092 States[BDV] = BDVState(BDVState::Conflict, Replacement);
1093 BaseI->eraseFromParent();
1095 const DataLayout &DL = cast<Instruction>(def)->getModule()->getDataLayout();
1096 while (!Worklist.empty()) {
1097 Instruction *BaseI = Worklist.pop_back_val();
1098 assert(NewInsts.count(BaseI));
1099 Value *Bdv = ReverseMap[BaseI];
1100 if (auto *BdvI = dyn_cast<Instruction>(Bdv))
1101 if (BaseI->isIdenticalTo(BdvI)) {
1102 DEBUG(dbgs() << "Identical Base: " << *BaseI << "\n");
1103 ReplaceBaseInstWith(Bdv, BaseI, Bdv);
1106 if (Value *V = SimplifyInstruction(BaseI, DL)) {
1107 DEBUG(dbgs() << "Base " << *BaseI << " simplified to " << *V << "\n");
1108 ReplaceBaseInstWith(Bdv, BaseI, V);
1113 // Cache all of our results so we can cheaply reuse them
1114 // NOTE: This is actually two caches: one of the base defining value
1115 // relation and one of the base pointer relation! FIXME
1116 for (auto Pair : States) {
1117 auto *BDV = Pair.first;
1118 Value *base = Pair.second.getBase();
1119 assert(BDV && base);
1121 std::string fromstr = cache.count(BDV) ? cache[BDV]->getName() : "none";
1122 DEBUG(dbgs() << "Updating base value cache"
1123 << " for: " << BDV->getName()
1124 << " from: " << fromstr
1125 << " to: " << base->getName() << "\n");
1127 if (cache.count(BDV)) {
1128 // Once we transition from the BDV relation being store in the cache to
1129 // the base relation being stored, it must be stable
1130 assert((!isKnownBaseResult(cache[BDV]) || cache[BDV] == base) &&
1131 "base relation should be stable");
1135 assert(cache.count(def));
1139 // For a set of live pointers (base and/or derived), identify the base
1140 // pointer of the object which they are derived from. This routine will
1141 // mutate the IR graph as needed to make the 'base' pointer live at the
1142 // definition site of 'derived'. This ensures that any use of 'derived' can
1143 // also use 'base'. This may involve the insertion of a number of
1144 // additional PHI nodes.
1146 // preconditions: live is a set of pointer type Values
1148 // side effects: may insert PHI nodes into the existing CFG, will preserve
1149 // CFG, will not remove or mutate any existing nodes
1151 // post condition: PointerToBase contains one (derived, base) pair for every
1152 // pointer in live. Note that derived can be equal to base if the original
1153 // pointer was a base pointer.
1155 findBasePointers(const StatepointLiveSetTy &live,
1156 DenseMap<Value *, Value *> &PointerToBase,
1157 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1158 // For the naming of values inserted to be deterministic - which makes for
1159 // much cleaner and more stable tests - we need to assign an order to the
1160 // live values. DenseSets do not provide a deterministic order across runs.
1161 SmallVector<Value *, 64> Temp;
1162 Temp.insert(Temp.end(), live.begin(), live.end());
1163 std::sort(Temp.begin(), Temp.end(), order_by_name);
1164 for (Value *ptr : Temp) {
1165 Value *base = findBasePointer(ptr, DVCache);
1166 assert(base && "failed to find base pointer");
1167 PointerToBase[ptr] = base;
1168 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1169 DT->dominates(cast<Instruction>(base)->getParent(),
1170 cast<Instruction>(ptr)->getParent())) &&
1171 "The base we found better dominate the derived pointer");
1173 // If you see this trip and like to live really dangerously, the code should
1174 // be correct, just with idioms the verifier can't handle. You can try
1175 // disabling the verifier at your own substantial risk.
1176 assert(!isa<ConstantPointerNull>(base) &&
1177 "the relocation code needs adjustment to handle the relocation of "
1178 "a null pointer constant without causing false positives in the "
1179 "safepoint ir verifier.");
1183 /// Find the required based pointers (and adjust the live set) for the given
1185 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1187 PartiallyConstructedSafepointRecord &result) {
1188 DenseMap<Value *, Value *> PointerToBase;
1189 findBasePointers(result.LiveSet, PointerToBase, &DT, DVCache);
1191 if (PrintBasePointers) {
1192 // Note: Need to print these in a stable order since this is checked in
1194 errs() << "Base Pairs (w/o Relocation):\n";
1195 SmallVector<Value *, 64> Temp;
1196 Temp.reserve(PointerToBase.size());
1197 for (auto Pair : PointerToBase) {
1198 Temp.push_back(Pair.first);
1200 std::sort(Temp.begin(), Temp.end(), order_by_name);
1201 for (Value *Ptr : Temp) {
1202 Value *Base = PointerToBase[Ptr];
1203 errs() << " derived ";
1204 Ptr->printAsOperand(errs(), false);
1206 Base->printAsOperand(errs(), false);
1211 result.PointerToBase = PointerToBase;
1214 /// Given an updated version of the dataflow liveness results, update the
1215 /// liveset and base pointer maps for the call site CS.
1216 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1218 PartiallyConstructedSafepointRecord &result);
1220 static void recomputeLiveInValues(
1221 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1222 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1223 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1224 // again. The old values are still live and will help it stabilize quickly.
1225 GCPtrLivenessData RevisedLivenessData;
1226 computeLiveInValues(DT, F, RevisedLivenessData);
1227 for (size_t i = 0; i < records.size(); i++) {
1228 struct PartiallyConstructedSafepointRecord &info = records[i];
1229 const CallSite &CS = toUpdate[i];
1230 recomputeLiveInValues(RevisedLivenessData, CS, info);
1234 // When inserting gc.relocate and gc.result calls, we need to ensure there are
1235 // no uses of the original value / return value between the gc.statepoint and
1236 // the gc.relocate / gc.result call. One case which can arise is a phi node
1237 // starting one of the successor blocks. We also need to be able to insert the
1238 // gc.relocates only on the path which goes through the statepoint. We might
1239 // need to split an edge to make this possible.
1241 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1242 DominatorTree &DT) {
1243 BasicBlock *Ret = BB;
1244 if (!BB->getUniquePredecessor())
1245 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1247 // Now that 'Ret' has unique predecessor we can safely remove all phi nodes
1249 FoldSingleEntryPHINodes(Ret);
1250 assert(!isa<PHINode>(Ret->begin()) &&
1251 "All PHI nodes should have been removed!");
1253 // At this point, we can safely insert a gc.relocate or gc.result as the first
1254 // instruction in Ret if needed.
1258 // Create new attribute set containing only attributes which can be transferred
1259 // from original call to the safepoint.
1260 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1263 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1264 unsigned Index = AS.getSlotIndex(Slot);
1266 if (Index == AttributeSet::ReturnIndex ||
1267 Index == AttributeSet::FunctionIndex) {
1269 for (Attribute Attr : make_range(AS.begin(Slot), AS.end(Slot))) {
1271 // Do not allow certain attributes - just skip them
1272 // Safepoint can not be read only or read none.
1273 if (Attr.hasAttribute(Attribute::ReadNone) ||
1274 Attr.hasAttribute(Attribute::ReadOnly))
1277 // These attributes control the generation of the gc.statepoint call /
1278 // invoke itself; and once the gc.statepoint is in place, they're of no
1280 if (Attr.hasAttribute("statepoint-num-patch-bytes") ||
1281 Attr.hasAttribute("statepoint-id"))
1284 Ret = Ret.addAttributes(
1285 AS.getContext(), Index,
1286 AttributeSet::get(AS.getContext(), Index, AttrBuilder(Attr)));
1290 // Just skip parameter attributes for now
1296 /// Helper function to place all gc relocates necessary for the given
1299 /// liveVariables - list of variables to be relocated.
1300 /// liveStart - index of the first live variable.
1301 /// basePtrs - base pointers.
1302 /// statepointToken - statepoint instruction to which relocates should be
1304 /// Builder - Llvm IR builder to be used to construct new calls.
1305 static void CreateGCRelocates(ArrayRef<Value *> LiveVariables,
1306 const int LiveStart,
1307 ArrayRef<Value *> BasePtrs,
1308 Instruction *StatepointToken,
1309 IRBuilder<> Builder) {
1310 if (LiveVariables.empty())
1313 auto FindIndex = [](ArrayRef<Value *> LiveVec, Value *Val) {
1314 auto ValIt = std::find(LiveVec.begin(), LiveVec.end(), Val);
1315 assert(ValIt != LiveVec.end() && "Val not found in LiveVec!");
1316 size_t Index = std::distance(LiveVec.begin(), ValIt);
1317 assert(Index < LiveVec.size() && "Bug in std::find?");
1320 Module *M = StatepointToken->getModule();
1322 // All gc_relocate are generated as i8 addrspace(1)* (or a vector type whose
1323 // element type is i8 addrspace(1)*). We originally generated unique
1324 // declarations for each pointer type, but this proved problematic because
1325 // the intrinsic mangling code is incomplete and fragile. Since we're moving
1326 // towards a single unified pointer type anyways, we can just cast everything
1327 // to an i8* of the right address space. A bitcast is added later to convert
1328 // gc_relocate to the actual value's type.
1329 auto getGCRelocateDecl = [&] (Type *Ty) {
1330 assert(isHandledGCPointerType(Ty));
1331 auto AS = Ty->getScalarType()->getPointerAddressSpace();
1332 Type *NewTy = Type::getInt8PtrTy(M->getContext(), AS);
1333 if (auto *VT = dyn_cast<VectorType>(Ty))
1334 NewTy = VectorType::get(NewTy, VT->getNumElements());
1335 return Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate,
1339 // Lazily populated map from input types to the canonicalized form mentioned
1340 // in the comment above. This should probably be cached somewhere more
1342 DenseMap<Type*, Value*> TypeToDeclMap;
1344 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1345 // Generate the gc.relocate call and save the result
1347 Builder.getInt32(LiveStart + FindIndex(LiveVariables, BasePtrs[i]));
1348 Value *LiveIdx = Builder.getInt32(LiveStart + i);
1350 Type *Ty = LiveVariables[i]->getType();
1351 if (!TypeToDeclMap.count(Ty))
1352 TypeToDeclMap[Ty] = getGCRelocateDecl(Ty);
1353 Value *GCRelocateDecl = TypeToDeclMap[Ty];
1355 // only specify a debug name if we can give a useful one
1356 CallInst *Reloc = Builder.CreateCall(
1357 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1358 suffixed_name_or(LiveVariables[i], ".relocated", ""));
1359 // Trick CodeGen into thinking there are lots of free registers at this
1361 Reloc->setCallingConv(CallingConv::Cold);
1367 /// This struct is used to defer RAUWs and `eraseFromParent` s. Using this
1368 /// avoids having to worry about keeping around dangling pointers to Values.
1369 class DeferredReplacement {
1370 AssertingVH<Instruction> Old;
1371 AssertingVH<Instruction> New;
1374 explicit DeferredReplacement(Instruction *Old, Instruction *New) :
1375 Old(Old), New(New) {
1376 assert(Old != New && "Not allowed!");
1379 /// Does the task represented by this instance.
1380 void doReplacement() {
1381 Instruction *OldI = Old;
1382 Instruction *NewI = New;
1384 assert(OldI != NewI && "Disallowed at construction?!");
1390 OldI->replaceAllUsesWith(NewI);
1391 OldI->eraseFromParent();
1397 makeStatepointExplicitImpl(const CallSite CS, /* to replace */
1398 const SmallVectorImpl<Value *> &BasePtrs,
1399 const SmallVectorImpl<Value *> &LiveVariables,
1400 PartiallyConstructedSafepointRecord &Result,
1401 std::vector<DeferredReplacement> &Replacements) {
1402 assert(BasePtrs.size() == LiveVariables.size());
1403 assert((UseDeoptBundles || isStatepoint(CS)) &&
1404 "This method expects to be rewriting a statepoint");
1406 // Then go ahead and use the builder do actually do the inserts. We insert
1407 // immediately before the previous instruction under the assumption that all
1408 // arguments will be available here. We can't insert afterwards since we may
1409 // be replacing a terminator.
1410 Instruction *InsertBefore = CS.getInstruction();
1411 IRBuilder<> Builder(InsertBefore);
1413 ArrayRef<Value *> GCArgs(LiveVariables);
1414 uint64_t StatepointID = 0xABCDEF00;
1415 uint32_t NumPatchBytes = 0;
1416 uint32_t Flags = uint32_t(StatepointFlags::None);
1418 ArrayRef<Use> CallArgs;
1419 ArrayRef<Use> DeoptArgs;
1420 ArrayRef<Use> TransitionArgs;
1422 Value *CallTarget = nullptr;
1424 if (UseDeoptBundles) {
1425 CallArgs = {CS.arg_begin(), CS.arg_end()};
1426 DeoptArgs = GetDeoptBundleOperands(CS);
1427 // TODO: we don't fill in TransitionArgs or Flags in this branch, but we
1428 // could have an operand bundle for that too.
1429 AttributeSet OriginalAttrs = CS.getAttributes();
1431 Attribute AttrID = OriginalAttrs.getAttribute(AttributeSet::FunctionIndex,
1433 if (AttrID.isStringAttribute())
1434 AttrID.getValueAsString().getAsInteger(10, StatepointID);
1436 Attribute AttrNumPatchBytes = OriginalAttrs.getAttribute(
1437 AttributeSet::FunctionIndex, "statepoint-num-patch-bytes");
1438 if (AttrNumPatchBytes.isStringAttribute())
1439 AttrNumPatchBytes.getValueAsString().getAsInteger(10, NumPatchBytes);
1441 CallTarget = CS.getCalledValue();
1443 // This branch will be gone soon, and we will soon only support the
1444 // UseDeoptBundles == true configuration.
1445 Statepoint OldSP(CS);
1446 StatepointID = OldSP.getID();
1447 NumPatchBytes = OldSP.getNumPatchBytes();
1448 Flags = OldSP.getFlags();
1450 CallArgs = {OldSP.arg_begin(), OldSP.arg_end()};
1451 DeoptArgs = {OldSP.vm_state_begin(), OldSP.vm_state_end()};
1452 TransitionArgs = {OldSP.gc_transition_args_begin(),
1453 OldSP.gc_transition_args_end()};
1454 CallTarget = OldSP.getCalledValue();
1457 // Create the statepoint given all the arguments
1458 Instruction *Token = nullptr;
1459 AttributeSet ReturnAttrs;
1461 CallInst *ToReplace = cast<CallInst>(CS.getInstruction());
1462 CallInst *Call = Builder.CreateGCStatepointCall(
1463 StatepointID, NumPatchBytes, CallTarget, Flags, CallArgs,
1464 TransitionArgs, DeoptArgs, GCArgs, "safepoint_token");
1466 Call->setTailCall(ToReplace->isTailCall());
1467 Call->setCallingConv(ToReplace->getCallingConv());
1469 // Currently we will fail on parameter attributes and on certain
1470 // function attributes.
1471 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes());
1472 // In case if we can handle this set of attributes - set up function attrs
1473 // directly on statepoint and return attrs later for gc_result intrinsic.
1474 Call->setAttributes(NewAttrs.getFnAttributes());
1475 ReturnAttrs = NewAttrs.getRetAttributes();
1479 // Put the following gc_result and gc_relocate calls immediately after the
1480 // the old call (which we're about to delete)
1481 assert(ToReplace->getNextNode() && "Not a terminator, must have next!");
1482 Builder.SetInsertPoint(ToReplace->getNextNode());
1483 Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc());
1485 InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction());
1487 // Insert the new invoke into the old block. We'll remove the old one in a
1488 // moment at which point this will become the new terminator for the
1490 InvokeInst *Invoke = Builder.CreateGCStatepointInvoke(
1491 StatepointID, NumPatchBytes, CallTarget, ToReplace->getNormalDest(),
1492 ToReplace->getUnwindDest(), Flags, CallArgs, TransitionArgs, DeoptArgs,
1493 GCArgs, "statepoint_token");
1495 Invoke->setCallingConv(ToReplace->getCallingConv());
1497 // Currently we will fail on parameter attributes and on certain
1498 // function attributes.
1499 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes());
1500 // In case if we can handle this set of attributes - set up function attrs
1501 // directly on statepoint and return attrs later for gc_result intrinsic.
1502 Invoke->setAttributes(NewAttrs.getFnAttributes());
1503 ReturnAttrs = NewAttrs.getRetAttributes();
1507 // Generate gc relocates in exceptional path
1508 BasicBlock *UnwindBlock = ToReplace->getUnwindDest();
1509 assert(!isa<PHINode>(UnwindBlock->begin()) &&
1510 UnwindBlock->getUniquePredecessor() &&
1511 "can't safely insert in this block!");
1513 Builder.SetInsertPoint(&*UnwindBlock->getFirstInsertionPt());
1514 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1516 // Attach exceptional gc relocates to the landingpad.
1517 Instruction *ExceptionalToken = UnwindBlock->getLandingPadInst();
1518 Result.UnwindToken = ExceptionalToken;
1520 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1521 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, ExceptionalToken,
1524 // Generate gc relocates and returns for normal block
1525 BasicBlock *NormalDest = ToReplace->getNormalDest();
1526 assert(!isa<PHINode>(NormalDest->begin()) &&
1527 NormalDest->getUniquePredecessor() &&
1528 "can't safely insert in this block!");
1530 Builder.SetInsertPoint(&*NormalDest->getFirstInsertionPt());
1532 // gc relocates will be generated later as if it were regular call
1535 assert(Token && "Should be set in one of the above branches!");
1537 if (UseDeoptBundles) {
1538 Token->setName("statepoint_token");
1539 if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
1541 CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "";
1542 CallInst *GCResult = Builder.CreateGCResult(Token, CS.getType(), Name);
1543 GCResult->setAttributes(CS.getAttributes().getRetAttributes());
1545 // We cannot RAUW or delete CS.getInstruction() because it could be in the
1546 // live set of some other safepoint, in which case that safepoint's
1547 // PartiallyConstructedSafepointRecord will hold a raw pointer to this
1548 // llvm::Instruction. Instead, we defer the replacement and deletion to
1549 // after the live sets have been made explicit in the IR, and we no longer
1550 // have raw pointers to worry about.
1551 Replacements.emplace_back(CS.getInstruction(), GCResult);
1553 Replacements.emplace_back(CS.getInstruction(), nullptr);
1556 assert(!CS.getInstruction()->hasNUsesOrMore(2) &&
1557 "only valid use before rewrite is gc.result");
1558 assert(!CS.getInstruction()->hasOneUse() ||
1559 isGCResult(cast<Instruction>(*CS.getInstruction()->user_begin())));
1561 // Take the name of the original statepoint token if there was one.
1562 Token->takeName(CS.getInstruction());
1564 // Update the gc.result of the original statepoint (if any) to use the newly
1565 // inserted statepoint. This is safe to do here since the token can't be
1566 // considered a live reference.
1567 CS.getInstruction()->replaceAllUsesWith(Token);
1568 CS.getInstruction()->eraseFromParent();
1571 Result.StatepointToken = Token;
1573 // Second, create a gc.relocate for every live variable
1574 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1575 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, Token, Builder);
1579 struct NameOrdering {
1583 bool operator()(NameOrdering const &a, NameOrdering const &b) {
1584 return -1 == a.Derived->getName().compare(b.Derived->getName());
1589 static void StabilizeOrder(SmallVectorImpl<Value *> &BaseVec,
1590 SmallVectorImpl<Value *> &LiveVec) {
1591 assert(BaseVec.size() == LiveVec.size());
1593 SmallVector<NameOrdering, 64> Temp;
1594 for (size_t i = 0; i < BaseVec.size(); i++) {
1596 v.Base = BaseVec[i];
1597 v.Derived = LiveVec[i];
1601 std::sort(Temp.begin(), Temp.end(), NameOrdering());
1602 for (size_t i = 0; i < BaseVec.size(); i++) {
1603 BaseVec[i] = Temp[i].Base;
1604 LiveVec[i] = Temp[i].Derived;
1608 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1609 // which make the relocations happening at this safepoint explicit.
1611 // WARNING: Does not do any fixup to adjust users of the original live
1612 // values. That's the callers responsibility.
1614 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS,
1615 PartiallyConstructedSafepointRecord &Result,
1616 std::vector<DeferredReplacement> &Replacements) {
1617 const auto &LiveSet = Result.LiveSet;
1618 const auto &PointerToBase = Result.PointerToBase;
1620 // Convert to vector for efficient cross referencing.
1621 SmallVector<Value *, 64> BaseVec, LiveVec;
1622 LiveVec.reserve(LiveSet.size());
1623 BaseVec.reserve(LiveSet.size());
1624 for (Value *L : LiveSet) {
1625 LiveVec.push_back(L);
1626 assert(PointerToBase.count(L));
1627 Value *Base = PointerToBase.find(L)->second;
1628 BaseVec.push_back(Base);
1630 assert(LiveVec.size() == BaseVec.size());
1632 // To make the output IR slightly more stable (for use in diffs), ensure a
1633 // fixed order of the values in the safepoint (by sorting the value name).
1634 // The order is otherwise meaningless.
1635 StabilizeOrder(BaseVec, LiveVec);
1637 // Do the actual rewriting and delete the old statepoint
1638 makeStatepointExplicitImpl(CS, BaseVec, LiveVec, Result, Replacements);
1641 // Helper function for the relocationViaAlloca.
1643 // It receives iterator to the statepoint gc relocates and emits a store to the
1644 // assigned location (via allocaMap) for the each one of them. It adds the
1645 // visited values into the visitedLiveValues set, which we will later use them
1646 // for sanity checking.
1648 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1649 DenseMap<Value *, Value *> &AllocaMap,
1650 DenseSet<Value *> &VisitedLiveValues) {
1652 for (User *U : GCRelocs) {
1653 GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U);
1657 Value *OriginalValue = const_cast<Value *>(Relocate->getDerivedPtr());
1658 assert(AllocaMap.count(OriginalValue));
1659 Value *Alloca = AllocaMap[OriginalValue];
1661 // Emit store into the related alloca
1662 // All gc_relocates are i8 addrspace(1)* typed, and it must be bitcasted to
1663 // the correct type according to alloca.
1664 assert(Relocate->getNextNode() &&
1665 "Should always have one since it's not a terminator");
1666 IRBuilder<> Builder(Relocate->getNextNode());
1667 Value *CastedRelocatedValue =
1668 Builder.CreateBitCast(Relocate,
1669 cast<AllocaInst>(Alloca)->getAllocatedType(),
1670 suffixed_name_or(Relocate, ".casted", ""));
1672 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1673 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1676 VisitedLiveValues.insert(OriginalValue);
1681 // Helper function for the "relocationViaAlloca". Similar to the
1682 // "insertRelocationStores" but works for rematerialized values.
1684 insertRematerializationStores(
1685 RematerializedValueMapTy RematerializedValues,
1686 DenseMap<Value *, Value *> &AllocaMap,
1687 DenseSet<Value *> &VisitedLiveValues) {
1689 for (auto RematerializedValuePair: RematerializedValues) {
1690 Instruction *RematerializedValue = RematerializedValuePair.first;
1691 Value *OriginalValue = RematerializedValuePair.second;
1693 assert(AllocaMap.count(OriginalValue) &&
1694 "Can not find alloca for rematerialized value");
1695 Value *Alloca = AllocaMap[OriginalValue];
1697 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1698 Store->insertAfter(RematerializedValue);
1701 VisitedLiveValues.insert(OriginalValue);
1706 /// Do all the relocation update via allocas and mem2reg
1707 static void relocationViaAlloca(
1708 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1709 ArrayRef<PartiallyConstructedSafepointRecord> Records) {
1711 // record initial number of (static) allocas; we'll check we have the same
1712 // number when we get done.
1713 int InitialAllocaNum = 0;
1714 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1716 if (isa<AllocaInst>(*I))
1720 // TODO-PERF: change data structures, reserve
1721 DenseMap<Value *, Value *> AllocaMap;
1722 SmallVector<AllocaInst *, 200> PromotableAllocas;
1723 // Used later to chack that we have enough allocas to store all values
1724 std::size_t NumRematerializedValues = 0;
1725 PromotableAllocas.reserve(Live.size());
1727 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1728 // "PromotableAllocas"
1729 auto emitAllocaFor = [&](Value *LiveValue) {
1730 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1731 F.getEntryBlock().getFirstNonPHI());
1732 AllocaMap[LiveValue] = Alloca;
1733 PromotableAllocas.push_back(Alloca);
1736 // Emit alloca for each live gc pointer
1737 for (Value *V : Live)
1740 // Emit allocas for rematerialized values
1741 for (const auto &Info : Records)
1742 for (auto RematerializedValuePair : Info.RematerializedValues) {
1743 Value *OriginalValue = RematerializedValuePair.second;
1744 if (AllocaMap.count(OriginalValue) != 0)
1747 emitAllocaFor(OriginalValue);
1748 ++NumRematerializedValues;
1751 // The next two loops are part of the same conceptual operation. We need to
1752 // insert a store to the alloca after the original def and at each
1753 // redefinition. We need to insert a load before each use. These are split
1754 // into distinct loops for performance reasons.
1756 // Update gc pointer after each statepoint: either store a relocated value or
1757 // null (if no relocated value was found for this gc pointer and it is not a
1758 // gc_result). This must happen before we update the statepoint with load of
1759 // alloca otherwise we lose the link between statepoint and old def.
1760 for (const auto &Info : Records) {
1761 Value *Statepoint = Info.StatepointToken;
1763 // This will be used for consistency check
1764 DenseSet<Value *> VisitedLiveValues;
1766 // Insert stores for normal statepoint gc relocates
1767 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1769 // In case if it was invoke statepoint
1770 // we will insert stores for exceptional path gc relocates.
1771 if (isa<InvokeInst>(Statepoint)) {
1772 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1776 // Do similar thing with rematerialized values
1777 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1780 if (ClobberNonLive) {
1781 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1782 // the gc.statepoint. This will turn some subtle GC problems into
1783 // slightly easier to debug SEGVs. Note that on large IR files with
1784 // lots of gc.statepoints this is extremely costly both memory and time
1786 SmallVector<AllocaInst *, 64> ToClobber;
1787 for (auto Pair : AllocaMap) {
1788 Value *Def = Pair.first;
1789 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1791 // This value was relocated
1792 if (VisitedLiveValues.count(Def)) {
1795 ToClobber.push_back(Alloca);
1798 auto InsertClobbersAt = [&](Instruction *IP) {
1799 for (auto *AI : ToClobber) {
1800 auto AIType = cast<PointerType>(AI->getType());
1801 auto PT = cast<PointerType>(AIType->getElementType());
1802 Constant *CPN = ConstantPointerNull::get(PT);
1803 StoreInst *Store = new StoreInst(CPN, AI);
1804 Store->insertBefore(IP);
1808 // Insert the clobbering stores. These may get intermixed with the
1809 // gc.results and gc.relocates, but that's fine.
1810 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1811 InsertClobbersAt(&*II->getNormalDest()->getFirstInsertionPt());
1812 InsertClobbersAt(&*II->getUnwindDest()->getFirstInsertionPt());
1814 InsertClobbersAt(cast<Instruction>(Statepoint)->getNextNode());
1819 // Update use with load allocas and add store for gc_relocated.
1820 for (auto Pair : AllocaMap) {
1821 Value *Def = Pair.first;
1822 Value *Alloca = Pair.second;
1824 // We pre-record the uses of allocas so that we dont have to worry about
1825 // later update that changes the user information..
1827 SmallVector<Instruction *, 20> Uses;
1828 // PERF: trade a linear scan for repeated reallocation
1829 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1830 for (User *U : Def->users()) {
1831 if (!isa<ConstantExpr>(U)) {
1832 // If the def has a ConstantExpr use, then the def is either a
1833 // ConstantExpr use itself or null. In either case
1834 // (recursively in the first, directly in the second), the oop
1835 // it is ultimately dependent on is null and this particular
1836 // use does not need to be fixed up.
1837 Uses.push_back(cast<Instruction>(U));
1841 std::sort(Uses.begin(), Uses.end());
1842 auto Last = std::unique(Uses.begin(), Uses.end());
1843 Uses.erase(Last, Uses.end());
1845 for (Instruction *Use : Uses) {
1846 if (isa<PHINode>(Use)) {
1847 PHINode *Phi = cast<PHINode>(Use);
1848 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1849 if (Def == Phi->getIncomingValue(i)) {
1850 LoadInst *Load = new LoadInst(
1851 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1852 Phi->setIncomingValue(i, Load);
1856 LoadInst *Load = new LoadInst(Alloca, "", Use);
1857 Use->replaceUsesOfWith(Def, Load);
1861 // Emit store for the initial gc value. Store must be inserted after load,
1862 // otherwise store will be in alloca's use list and an extra load will be
1863 // inserted before it.
1864 StoreInst *Store = new StoreInst(Def, Alloca);
1865 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1866 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1867 // InvokeInst is a TerminatorInst so the store need to be inserted
1868 // into its normal destination block.
1869 BasicBlock *NormalDest = Invoke->getNormalDest();
1870 Store->insertBefore(NormalDest->getFirstNonPHI());
1872 assert(!Inst->isTerminator() &&
1873 "The only TerminatorInst that can produce a value is "
1874 "InvokeInst which is handled above.");
1875 Store->insertAfter(Inst);
1878 assert(isa<Argument>(Def));
1879 Store->insertAfter(cast<Instruction>(Alloca));
1883 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1884 "we must have the same allocas with lives");
1885 if (!PromotableAllocas.empty()) {
1886 // Apply mem2reg to promote alloca to SSA
1887 PromoteMemToReg(PromotableAllocas, DT);
1891 for (auto &I : F.getEntryBlock())
1892 if (isa<AllocaInst>(I))
1894 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1898 /// Implement a unique function which doesn't require we sort the input
1899 /// vector. Doing so has the effect of changing the output of a couple of
1900 /// tests in ways which make them less useful in testing fused safepoints.
1901 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1902 SmallSet<T, 8> Seen;
1903 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) {
1904 return !Seen.insert(V).second;
1908 /// Insert holders so that each Value is obviously live through the entire
1909 /// lifetime of the call.
1910 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1911 SmallVectorImpl<CallInst *> &Holders) {
1913 // No values to hold live, might as well not insert the empty holder
1916 Module *M = CS.getInstruction()->getModule();
1917 // Use a dummy vararg function to actually hold the values live
1918 Function *Func = cast<Function>(M->getOrInsertFunction(
1919 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1921 // For call safepoints insert dummy calls right after safepoint
1922 Holders.push_back(CallInst::Create(Func, Values, "",
1923 &*++CS.getInstruction()->getIterator()));
1926 // For invoke safepooints insert dummy calls both in normal and
1927 // exceptional destination blocks
1928 auto *II = cast<InvokeInst>(CS.getInstruction());
1929 Holders.push_back(CallInst::Create(
1930 Func, Values, "", &*II->getNormalDest()->getFirstInsertionPt()));
1931 Holders.push_back(CallInst::Create(
1932 Func, Values, "", &*II->getUnwindDest()->getFirstInsertionPt()));
1935 static void findLiveReferences(
1936 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1937 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1938 GCPtrLivenessData OriginalLivenessData;
1939 computeLiveInValues(DT, F, OriginalLivenessData);
1940 for (size_t i = 0; i < records.size(); i++) {
1941 struct PartiallyConstructedSafepointRecord &info = records[i];
1942 const CallSite &CS = toUpdate[i];
1943 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1947 /// Remove any vector of pointers from the live set by scalarizing them over the
1948 /// statepoint instruction. Adds the scalarized pieces to the live set. It
1949 /// would be preferable to include the vector in the statepoint itself, but
1950 /// the lowering code currently does not handle that. Extending it would be
1951 /// slightly non-trivial since it requires a format change. Given how rare
1952 /// such cases are (for the moment?) scalarizing is an acceptable compromise.
1953 static void splitVectorValues(Instruction *StatepointInst,
1954 StatepointLiveSetTy &LiveSet,
1955 DenseMap<Value *, Value *>& PointerToBase,
1956 DominatorTree &DT) {
1957 SmallVector<Value *, 16> ToSplit;
1958 for (Value *V : LiveSet)
1959 if (isa<VectorType>(V->getType()))
1960 ToSplit.push_back(V);
1962 if (ToSplit.empty())
1965 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping;
1967 Function &F = *(StatepointInst->getParent()->getParent());
1969 DenseMap<Value *, AllocaInst *> AllocaMap;
1970 // First is normal return, second is exceptional return (invoke only)
1971 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1972 for (Value *V : ToSplit) {
1973 AllocaInst *Alloca =
1974 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1975 AllocaMap[V] = Alloca;
1977 VectorType *VT = cast<VectorType>(V->getType());
1978 IRBuilder<> Builder(StatepointInst);
1979 SmallVector<Value *, 16> Elements;
1980 for (unsigned i = 0; i < VT->getNumElements(); i++)
1981 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1982 ElementMapping[V] = Elements;
1984 auto InsertVectorReform = [&](Instruction *IP) {
1985 Builder.SetInsertPoint(IP);
1986 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1987 Value *ResultVec = UndefValue::get(VT);
1988 for (unsigned i = 0; i < VT->getNumElements(); i++)
1989 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1990 Builder.getInt32(i));
1994 if (isa<CallInst>(StatepointInst)) {
1995 BasicBlock::iterator Next(StatepointInst);
1997 Instruction *IP = &*(Next);
1998 Replacements[V].first = InsertVectorReform(IP);
1999 Replacements[V].second = nullptr;
2001 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
2002 // We've already normalized - check that we don't have shared destination
2004 BasicBlock *NormalDest = Invoke->getNormalDest();
2005 assert(!isa<PHINode>(NormalDest->begin()));
2006 BasicBlock *UnwindDest = Invoke->getUnwindDest();
2007 assert(!isa<PHINode>(UnwindDest->begin()));
2008 // Insert insert element sequences in both successors
2009 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
2010 Replacements[V].first = InsertVectorReform(IP);
2011 IP = &*(UnwindDest->getFirstInsertionPt());
2012 Replacements[V].second = InsertVectorReform(IP);
2016 for (Value *V : ToSplit) {
2017 AllocaInst *Alloca = AllocaMap[V];
2019 // Capture all users before we start mutating use lists
2020 SmallVector<Instruction *, 16> Users;
2021 for (User *U : V->users())
2022 Users.push_back(cast<Instruction>(U));
2024 for (Instruction *I : Users) {
2025 if (auto Phi = dyn_cast<PHINode>(I)) {
2026 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
2027 if (V == Phi->getIncomingValue(i)) {
2028 LoadInst *Load = new LoadInst(
2029 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
2030 Phi->setIncomingValue(i, Load);
2033 LoadInst *Load = new LoadInst(Alloca, "", I);
2034 I->replaceUsesOfWith(V, Load);
2038 // Store the original value and the replacement value into the alloca
2039 StoreInst *Store = new StoreInst(V, Alloca);
2040 if (auto I = dyn_cast<Instruction>(V))
2041 Store->insertAfter(I);
2043 Store->insertAfter(Alloca);
2045 // Normal return for invoke, or call return
2046 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
2047 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
2048 // Unwind return for invoke only
2049 Replacement = cast_or_null<Instruction>(Replacements[V].second);
2051 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
2054 // apply mem2reg to promote alloca to SSA
2055 SmallVector<AllocaInst *, 16> Allocas;
2056 for (Value *V : ToSplit)
2057 Allocas.push_back(AllocaMap[V]);
2058 PromoteMemToReg(Allocas, DT);
2060 // Update our tracking of live pointers and base mappings to account for the
2061 // changes we just made.
2062 for (Value *V : ToSplit) {
2063 auto &Elements = ElementMapping[V];
2066 LiveSet.insert(Elements.begin(), Elements.end());
2067 // We need to update the base mapping as well.
2068 assert(PointerToBase.count(V));
2069 Value *OldBase = PointerToBase[V];
2070 auto &BaseElements = ElementMapping[OldBase];
2071 PointerToBase.erase(V);
2072 assert(Elements.size() == BaseElements.size());
2073 for (unsigned i = 0; i < Elements.size(); i++) {
2074 Value *Elem = Elements[i];
2075 PointerToBase[Elem] = BaseElements[i];
2080 // Helper function for the "rematerializeLiveValues". It walks use chain
2081 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
2082 // values are visited (currently it is GEP's and casts). Returns true if it
2083 // successfully reached "BaseValue" and false otherwise.
2084 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
2086 static bool findRematerializableChainToBasePointer(
2087 SmallVectorImpl<Instruction*> &ChainToBase,
2088 Value *CurrentValue, Value *BaseValue) {
2090 // We have found a base value
2091 if (CurrentValue == BaseValue) {
2095 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
2096 ChainToBase.push_back(GEP);
2097 return findRematerializableChainToBasePointer(ChainToBase,
2098 GEP->getPointerOperand(),
2102 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
2103 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
2106 ChainToBase.push_back(CI);
2107 return findRematerializableChainToBasePointer(ChainToBase,
2108 CI->getOperand(0), BaseValue);
2111 // Not supported instruction in the chain
2115 // Helper function for the "rematerializeLiveValues". Compute cost of the use
2116 // chain we are going to rematerialize.
2118 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
2119 TargetTransformInfo &TTI) {
2122 for (Instruction *Instr : Chain) {
2123 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
2124 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
2125 "non noop cast is found during rematerialization");
2127 Type *SrcTy = CI->getOperand(0)->getType();
2128 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
2130 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
2131 // Cost of the address calculation
2132 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
2133 Cost += TTI.getAddressComputationCost(ValTy);
2135 // And cost of the GEP itself
2136 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
2137 // allowed for the external usage)
2138 if (!GEP->hasAllConstantIndices())
2142 llvm_unreachable("unsupported instruciton type during rematerialization");
2149 // From the statepoint live set pick values that are cheaper to recompute then
2150 // to relocate. Remove this values from the live set, rematerialize them after
2151 // statepoint and record them in "Info" structure. Note that similar to
2152 // relocated values we don't do any user adjustments here.
2153 static void rematerializeLiveValues(CallSite CS,
2154 PartiallyConstructedSafepointRecord &Info,
2155 TargetTransformInfo &TTI) {
2156 const unsigned int ChainLengthThreshold = 10;
2158 // Record values we are going to delete from this statepoint live set.
2159 // We can not di this in following loop due to iterator invalidation.
2160 SmallVector<Value *, 32> LiveValuesToBeDeleted;
2162 for (Value *LiveValue: Info.LiveSet) {
2163 // For each live pointer find it's defining chain
2164 SmallVector<Instruction *, 3> ChainToBase;
2165 assert(Info.PointerToBase.count(LiveValue));
2167 findRematerializableChainToBasePointer(ChainToBase,
2169 Info.PointerToBase[LiveValue]);
2170 // Nothing to do, or chain is too long
2172 ChainToBase.size() == 0 ||
2173 ChainToBase.size() > ChainLengthThreshold)
2176 // Compute cost of this chain
2177 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
2178 // TODO: We can also account for cases when we will be able to remove some
2179 // of the rematerialized values by later optimization passes. I.e if
2180 // we rematerialized several intersecting chains. Or if original values
2181 // don't have any uses besides this statepoint.
2183 // For invokes we need to rematerialize each chain twice - for normal and
2184 // for unwind basic blocks. Model this by multiplying cost by two.
2185 if (CS.isInvoke()) {
2188 // If it's too expensive - skip it
2189 if (Cost >= RematerializationThreshold)
2192 // Remove value from the live set
2193 LiveValuesToBeDeleted.push_back(LiveValue);
2195 // Clone instructions and record them inside "Info" structure
2197 // Walk backwards to visit top-most instructions first
2198 std::reverse(ChainToBase.begin(), ChainToBase.end());
2200 // Utility function which clones all instructions from "ChainToBase"
2201 // and inserts them before "InsertBefore". Returns rematerialized value
2202 // which should be used after statepoint.
2203 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
2204 Instruction *LastClonedValue = nullptr;
2205 Instruction *LastValue = nullptr;
2206 for (Instruction *Instr: ChainToBase) {
2207 // Only GEP's and casts are suported as we need to be careful to not
2208 // introduce any new uses of pointers not in the liveset.
2209 // Note that it's fine to introduce new uses of pointers which were
2210 // otherwise not used after this statepoint.
2211 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2213 Instruction *ClonedValue = Instr->clone();
2214 ClonedValue->insertBefore(InsertBefore);
2215 ClonedValue->setName(Instr->getName() + ".remat");
2217 // If it is not first instruction in the chain then it uses previously
2218 // cloned value. We should update it to use cloned value.
2219 if (LastClonedValue) {
2221 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2223 // Assert that cloned instruction does not use any instructions from
2224 // this chain other than LastClonedValue
2225 for (auto OpValue : ClonedValue->operand_values()) {
2226 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
2227 ChainToBase.end() &&
2228 "incorrect use in rematerialization chain");
2233 LastClonedValue = ClonedValue;
2236 assert(LastClonedValue);
2237 return LastClonedValue;
2240 // Different cases for calls and invokes. For invokes we need to clone
2241 // instructions both on normal and unwind path.
2243 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2244 assert(InsertBefore);
2245 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
2246 Info.RematerializedValues[RematerializedValue] = LiveValue;
2248 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2250 Instruction *NormalInsertBefore =
2251 &*Invoke->getNormalDest()->getFirstInsertionPt();
2252 Instruction *UnwindInsertBefore =
2253 &*Invoke->getUnwindDest()->getFirstInsertionPt();
2255 Instruction *NormalRematerializedValue =
2256 rematerializeChain(NormalInsertBefore);
2257 Instruction *UnwindRematerializedValue =
2258 rematerializeChain(UnwindInsertBefore);
2260 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2261 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2265 // Remove rematerializaed values from the live set
2266 for (auto LiveValue: LiveValuesToBeDeleted) {
2267 Info.LiveSet.erase(LiveValue);
2271 static bool insertParsePoints(Function &F, DominatorTree &DT,
2272 TargetTransformInfo &TTI,
2273 SmallVectorImpl<CallSite> &ToUpdate) {
2275 // sanity check the input
2276 std::set<CallSite> Uniqued;
2277 Uniqued.insert(ToUpdate.begin(), ToUpdate.end());
2278 assert(Uniqued.size() == ToUpdate.size() && "no duplicates please!");
2280 for (CallSite CS : ToUpdate) {
2281 assert(CS.getInstruction()->getParent()->getParent() == &F);
2282 assert((UseDeoptBundles || isStatepoint(CS)) &&
2283 "expected to already be a deopt statepoint");
2287 // When inserting gc.relocates for invokes, we need to be able to insert at
2288 // the top of the successor blocks. See the comment on
2289 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2290 // may restructure the CFG.
2291 for (CallSite CS : ToUpdate) {
2294 auto *II = cast<InvokeInst>(CS.getInstruction());
2295 normalizeForInvokeSafepoint(II->getNormalDest(), II->getParent(), DT);
2296 normalizeForInvokeSafepoint(II->getUnwindDest(), II->getParent(), DT);
2299 // A list of dummy calls added to the IR to keep various values obviously
2300 // live in the IR. We'll remove all of these when done.
2301 SmallVector<CallInst *, 64> Holders;
2303 // Insert a dummy call with all of the arguments to the vm_state we'll need
2304 // for the actual safepoint insertion. This ensures reference arguments in
2305 // the deopt argument list are considered live through the safepoint (and
2306 // thus makes sure they get relocated.)
2307 for (CallSite CS : ToUpdate) {
2308 SmallVector<Value *, 64> DeoptValues;
2310 iterator_range<const Use *> DeoptStateRange =
2312 ? iterator_range<const Use *>(GetDeoptBundleOperands(CS))
2313 : iterator_range<const Use *>(Statepoint(CS).vm_state_args());
2315 for (Value *Arg : DeoptStateRange) {
2316 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2317 "support for FCA unimplemented");
2318 if (isHandledGCPointerType(Arg->getType()))
2319 DeoptValues.push_back(Arg);
2322 insertUseHolderAfter(CS, DeoptValues, Holders);
2325 SmallVector<PartiallyConstructedSafepointRecord, 64> Records(ToUpdate.size());
2327 // A) Identify all gc pointers which are statically live at the given call
2329 findLiveReferences(F, DT, ToUpdate, Records);
2331 // B) Find the base pointers for each live pointer
2332 /* scope for caching */ {
2333 // Cache the 'defining value' relation used in the computation and
2334 // insertion of base phis and selects. This ensures that we don't insert
2335 // large numbers of duplicate base_phis.
2336 DefiningValueMapTy DVCache;
2338 for (size_t i = 0; i < Records.size(); i++) {
2339 PartiallyConstructedSafepointRecord &info = Records[i];
2340 findBasePointers(DT, DVCache, ToUpdate[i], info);
2342 } // end of cache scope
2344 // The base phi insertion logic (for any safepoint) may have inserted new
2345 // instructions which are now live at some safepoint. The simplest such
2348 // phi a <-- will be a new base_phi here
2349 // safepoint 1 <-- that needs to be live here
2353 // We insert some dummy calls after each safepoint to definitely hold live
2354 // the base pointers which were identified for that safepoint. We'll then
2355 // ask liveness for _every_ base inserted to see what is now live. Then we
2356 // remove the dummy calls.
2357 Holders.reserve(Holders.size() + Records.size());
2358 for (size_t i = 0; i < Records.size(); i++) {
2359 PartiallyConstructedSafepointRecord &Info = Records[i];
2361 SmallVector<Value *, 128> Bases;
2362 for (auto Pair : Info.PointerToBase)
2363 Bases.push_back(Pair.second);
2365 insertUseHolderAfter(ToUpdate[i], Bases, Holders);
2368 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2369 // need to rerun liveness. We may *also* have inserted new defs, but that's
2370 // not the key issue.
2371 recomputeLiveInValues(F, DT, ToUpdate, Records);
2373 if (PrintBasePointers) {
2374 for (auto &Info : Records) {
2375 errs() << "Base Pairs: (w/Relocation)\n";
2376 for (auto Pair : Info.PointerToBase) {
2377 errs() << " derived ";
2378 Pair.first->printAsOperand(errs(), false);
2380 Pair.second->printAsOperand(errs(), false);
2386 // It is possible that non-constant live variables have a constant base. For
2387 // example, a GEP with a variable offset from a global. In this case we can
2388 // remove it from the liveset. We already don't add constants to the liveset
2389 // because we assume they won't move at runtime and the GC doesn't need to be
2390 // informed about them. The same reasoning applies if the base is constant.
2391 // Note that the relocation placement code relies on this filtering for
2392 // correctness as it expects the base to be in the liveset, which isn't true
2393 // if the base is constant.
2394 for (auto &Info : Records)
2395 for (auto &BasePair : Info.PointerToBase)
2396 if (isa<Constant>(BasePair.second))
2397 Info.LiveSet.erase(BasePair.first);
2399 for (CallInst *CI : Holders)
2400 CI->eraseFromParent();
2404 // Do a limited scalarization of any live at safepoint vector values which
2405 // contain pointers. This enables this pass to run after vectorization at
2406 // the cost of some possible performance loss. Note: This is known to not
2407 // handle updating of the side tables correctly which can lead to relocation
2408 // bugs when the same vector is live at multiple statepoints. We're in the
2409 // process of implementing the alternate lowering - relocating the
2410 // vector-of-pointers as first class item and updating the backend to
2411 // understand that - but that's not yet complete.
2413 for (size_t i = 0; i < Records.size(); i++) {
2414 PartiallyConstructedSafepointRecord &Info = Records[i];
2415 Instruction *Statepoint = ToUpdate[i].getInstruction();
2416 splitVectorValues(cast<Instruction>(Statepoint), Info.LiveSet,
2417 Info.PointerToBase, DT);
2420 // In order to reduce live set of statepoint we might choose to rematerialize
2421 // some values instead of relocating them. This is purely an optimization and
2422 // does not influence correctness.
2423 for (size_t i = 0; i < Records.size(); i++)
2424 rematerializeLiveValues(ToUpdate[i], Records[i], TTI);
2426 // We need this to safely RAUW and delete call or invoke return values that
2427 // may themselves be live over a statepoint. For details, please see usage in
2428 // makeStatepointExplicitImpl.
2429 std::vector<DeferredReplacement> Replacements;
2431 // Now run through and replace the existing statepoints with new ones with
2432 // the live variables listed. We do not yet update uses of the values being
2433 // relocated. We have references to live variables that need to
2434 // survive to the last iteration of this loop. (By construction, the
2435 // previous statepoint can not be a live variable, thus we can and remove
2436 // the old statepoint calls as we go.)
2437 for (size_t i = 0; i < Records.size(); i++)
2438 makeStatepointExplicit(DT, ToUpdate[i], Records[i], Replacements);
2440 ToUpdate.clear(); // prevent accident use of invalid CallSites
2442 for (auto &PR : Replacements)
2445 Replacements.clear();
2447 for (auto &Info : Records) {
2448 // These live sets may contain state Value pointers, since we replaced calls
2449 // with operand bundles with calls wrapped in gc.statepoint, and some of
2450 // those calls may have been def'ing live gc pointers. Clear these out to
2451 // avoid accidentally using them.
2453 // TODO: We should create a separate data structure that does not contain
2454 // these live sets, and migrate to using that data structure from this point
2456 Info.LiveSet.clear();
2457 Info.PointerToBase.clear();
2460 // Do all the fixups of the original live variables to their relocated selves
2461 SmallVector<Value *, 128> Live;
2462 for (size_t i = 0; i < Records.size(); i++) {
2463 PartiallyConstructedSafepointRecord &Info = Records[i];
2465 // We can't simply save the live set from the original insertion. One of
2466 // the live values might be the result of a call which needs a safepoint.
2467 // That Value* no longer exists and we need to use the new gc_result.
2468 // Thankfully, the live set is embedded in the statepoint (and updated), so
2469 // we just grab that.
2470 Statepoint Statepoint(Info.StatepointToken);
2471 Live.insert(Live.end(), Statepoint.gc_args_begin(),
2472 Statepoint.gc_args_end());
2474 // Do some basic sanity checks on our liveness results before performing
2475 // relocation. Relocation can and will turn mistakes in liveness results
2476 // into non-sensical code which is must harder to debug.
2477 // TODO: It would be nice to test consistency as well
2478 assert(DT.isReachableFromEntry(Info.StatepointToken->getParent()) &&
2479 "statepoint must be reachable or liveness is meaningless");
2480 for (Value *V : Statepoint.gc_args()) {
2481 if (!isa<Instruction>(V))
2482 // Non-instruction values trivial dominate all possible uses
2484 auto *LiveInst = cast<Instruction>(V);
2485 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2486 "unreachable values should never be live");
2487 assert(DT.dominates(LiveInst, Info.StatepointToken) &&
2488 "basic SSA liveness expectation violated by liveness analysis");
2492 unique_unsorted(Live);
2496 for (auto *Ptr : Live)
2497 assert(isHandledGCPointerType(Ptr->getType()) &&
2498 "must be a gc pointer type");
2501 relocationViaAlloca(F, DT, Live, Records);
2502 return !Records.empty();
2505 // Handles both return values and arguments for Functions and CallSites.
2506 template <typename AttrHolder>
2507 static void RemoveNonValidAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2510 if (AH.getDereferenceableBytes(Index))
2511 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2512 AH.getDereferenceableBytes(Index)));
2513 if (AH.getDereferenceableOrNullBytes(Index))
2514 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2515 AH.getDereferenceableOrNullBytes(Index)));
2516 if (AH.doesNotAlias(Index))
2517 R.addAttribute(Attribute::NoAlias);
2520 AH.setAttributes(AH.getAttributes().removeAttributes(
2521 Ctx, Index, AttributeSet::get(Ctx, Index, R)));
2525 RewriteStatepointsForGC::stripNonValidAttributesFromPrototype(Function &F) {
2526 LLVMContext &Ctx = F.getContext();
2528 for (Argument &A : F.args())
2529 if (isa<PointerType>(A.getType()))
2530 RemoveNonValidAttrAtIndex(Ctx, F, A.getArgNo() + 1);
2532 if (isa<PointerType>(F.getReturnType()))
2533 RemoveNonValidAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex);
2536 void RewriteStatepointsForGC::stripNonValidAttributesFromBody(Function &F) {
2540 LLVMContext &Ctx = F.getContext();
2541 MDBuilder Builder(Ctx);
2543 for (Instruction &I : instructions(F)) {
2544 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2545 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2546 bool IsImmutableTBAA =
2547 MD->getNumOperands() == 4 &&
2548 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2550 if (!IsImmutableTBAA)
2551 continue; // no work to do, MD_tbaa is already marked mutable
2553 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2554 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2556 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2558 MDNode *MutableTBAA =
2559 Builder.createTBAAStructTagNode(Base, Access, Offset);
2560 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2563 if (CallSite CS = CallSite(&I)) {
2564 for (int i = 0, e = CS.arg_size(); i != e; i++)
2565 if (isa<PointerType>(CS.getArgument(i)->getType()))
2566 RemoveNonValidAttrAtIndex(Ctx, CS, i + 1);
2567 if (isa<PointerType>(CS.getType()))
2568 RemoveNonValidAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex);
2573 /// Returns true if this function should be rewritten by this pass. The main
2574 /// point of this function is as an extension point for custom logic.
2575 static bool shouldRewriteStatepointsIn(Function &F) {
2576 // TODO: This should check the GCStrategy
2578 const auto &FunctionGCName = F.getGC();
2579 const StringRef StatepointExampleName("statepoint-example");
2580 const StringRef CoreCLRName("coreclr");
2581 return (StatepointExampleName == FunctionGCName) ||
2582 (CoreCLRName == FunctionGCName);
2587 void RewriteStatepointsForGC::stripNonValidAttributes(Module &M) {
2589 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) &&
2593 for (Function &F : M)
2594 stripNonValidAttributesFromPrototype(F);
2596 for (Function &F : M)
2597 stripNonValidAttributesFromBody(F);
2600 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2601 // Nothing to do for declarations.
2602 if (F.isDeclaration() || F.empty())
2605 // Policy choice says not to rewrite - the most common reason is that we're
2606 // compiling code without a GCStrategy.
2607 if (!shouldRewriteStatepointsIn(F))
2610 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2611 TargetTransformInfo &TTI =
2612 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2614 auto NeedsRewrite = [](Instruction &I) {
2615 if (UseDeoptBundles) {
2616 if (ImmutableCallSite CS = ImmutableCallSite(&I))
2617 return !callsGCLeafFunction(CS);
2621 return isStatepoint(I);
2624 // Gather all the statepoints which need rewritten. Be careful to only
2625 // consider those in reachable code since we need to ask dominance queries
2626 // when rewriting. We'll delete the unreachable ones in a moment.
2627 SmallVector<CallSite, 64> ParsePointNeeded;
2628 bool HasUnreachableStatepoint = false;
2629 for (Instruction &I : instructions(F)) {
2630 // TODO: only the ones with the flag set!
2631 if (NeedsRewrite(I)) {
2632 if (DT.isReachableFromEntry(I.getParent()))
2633 ParsePointNeeded.push_back(CallSite(&I));
2635 HasUnreachableStatepoint = true;
2639 bool MadeChange = false;
2641 // Delete any unreachable statepoints so that we don't have unrewritten
2642 // statepoints surviving this pass. This makes testing easier and the
2643 // resulting IR less confusing to human readers. Rather than be fancy, we
2644 // just reuse a utility function which removes the unreachable blocks.
2645 if (HasUnreachableStatepoint)
2646 MadeChange |= removeUnreachableBlocks(F);
2648 // Return early if no work to do.
2649 if (ParsePointNeeded.empty())
2652 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2653 // These are created by LCSSA. They have the effect of increasing the size
2654 // of liveness sets for no good reason. It may be harder to do this post
2655 // insertion since relocations and base phis can confuse things.
2656 for (BasicBlock &BB : F)
2657 if (BB.getUniquePredecessor()) {
2659 FoldSingleEntryPHINodes(&BB);
2662 // Before we start introducing relocations, we want to tweak the IR a bit to
2663 // avoid unfortunate code generation effects. The main example is that we
2664 // want to try to make sure the comparison feeding a branch is after any
2665 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2666 // values feeding a branch after relocation. This is semantically correct,
2667 // but results in extra register pressure since both the pre-relocation and
2668 // post-relocation copies must be available in registers. For code without
2669 // relocations this is handled elsewhere, but teaching the scheduler to
2670 // reverse the transform we're about to do would be slightly complex.
2671 // Note: This may extend the live range of the inputs to the icmp and thus
2672 // increase the liveset of any statepoint we move over. This is profitable
2673 // as long as all statepoints are in rare blocks. If we had in-register
2674 // lowering for live values this would be a much safer transform.
2675 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2676 if (auto *BI = dyn_cast<BranchInst>(TI))
2677 if (BI->isConditional())
2678 return dyn_cast<Instruction>(BI->getCondition());
2679 // TODO: Extend this to handle switches
2682 for (BasicBlock &BB : F) {
2683 TerminatorInst *TI = BB.getTerminator();
2684 if (auto *Cond = getConditionInst(TI))
2685 // TODO: Handle more than just ICmps here. We should be able to move
2686 // most instructions without side effects or memory access.
2687 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2689 Cond->moveBefore(TI);
2693 MadeChange |= insertParsePoints(F, DT, TTI, ParsePointNeeded);
2697 // liveness computation via standard dataflow
2698 // -------------------------------------------------------------------
2700 // TODO: Consider using bitvectors for liveness, the set of potentially
2701 // interesting values should be small and easy to pre-compute.
2703 /// Compute the live-in set for the location rbegin starting from
2704 /// the live-out set of the basic block
2705 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2706 BasicBlock::reverse_iterator rend,
2707 DenseSet<Value *> &LiveTmp) {
2709 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2710 Instruction *I = &*ritr;
2712 // KILL/Def - Remove this definition from LiveIn
2715 // Don't consider *uses* in PHI nodes, we handle their contribution to
2716 // predecessor blocks when we seed the LiveOut sets
2717 if (isa<PHINode>(I))
2720 // USE - Add to the LiveIn set for this instruction
2721 for (Value *V : I->operands()) {
2722 assert(!isUnhandledGCPointerType(V->getType()) &&
2723 "support for FCA unimplemented");
2724 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2725 // The choice to exclude all things constant here is slightly subtle.
2726 // There are two independent reasons:
2727 // - We assume that things which are constant (from LLVM's definition)
2728 // do not move at runtime. For example, the address of a global
2729 // variable is fixed, even though it's contents may not be.
2730 // - Second, we can't disallow arbitrary inttoptr constants even
2731 // if the language frontend does. Optimization passes are free to
2732 // locally exploit facts without respect to global reachability. This
2733 // can create sections of code which are dynamically unreachable and
2734 // contain just about anything. (see constants.ll in tests)
2741 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2743 for (BasicBlock *Succ : successors(BB)) {
2744 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2745 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2746 PHINode *Phi = cast<PHINode>(&*I);
2747 Value *V = Phi->getIncomingValueForBlock(BB);
2748 assert(!isUnhandledGCPointerType(V->getType()) &&
2749 "support for FCA unimplemented");
2750 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2757 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2758 DenseSet<Value *> KillSet;
2759 for (Instruction &I : *BB)
2760 if (isHandledGCPointerType(I.getType()))
2766 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2767 /// sanity check for the liveness computation.
2768 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2769 TerminatorInst *TI, bool TermOkay = false) {
2770 for (Value *V : Live) {
2771 if (auto *I = dyn_cast<Instruction>(V)) {
2772 // The terminator can be a member of the LiveOut set. LLVM's definition
2773 // of instruction dominance states that V does not dominate itself. As
2774 // such, we need to special case this to allow it.
2775 if (TermOkay && TI == I)
2777 assert(DT.dominates(I, TI) &&
2778 "basic SSA liveness expectation violated by liveness analysis");
2783 /// Check that all the liveness sets used during the computation of liveness
2784 /// obey basic SSA properties. This is useful for finding cases where we miss
2786 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2788 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2789 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2790 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2794 static void computeLiveInValues(DominatorTree &DT, Function &F,
2795 GCPtrLivenessData &Data) {
2797 SmallSetVector<BasicBlock *, 200> Worklist;
2798 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2799 // We use a SetVector so that we don't have duplicates in the worklist.
2800 Worklist.insert(pred_begin(BB), pred_end(BB));
2802 auto NextItem = [&]() {
2803 BasicBlock *BB = Worklist.back();
2804 Worklist.pop_back();
2808 // Seed the liveness for each individual block
2809 for (BasicBlock &BB : F) {
2810 Data.KillSet[&BB] = computeKillSet(&BB);
2811 Data.LiveSet[&BB].clear();
2812 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2815 for (Value *Kill : Data.KillSet[&BB])
2816 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2819 Data.LiveOut[&BB] = DenseSet<Value *>();
2820 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2821 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2822 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2823 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2824 if (!Data.LiveIn[&BB].empty())
2825 AddPredsToWorklist(&BB);
2828 // Propagate that liveness until stable
2829 while (!Worklist.empty()) {
2830 BasicBlock *BB = NextItem();
2832 // Compute our new liveout set, then exit early if it hasn't changed
2833 // despite the contribution of our successor.
2834 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2835 const auto OldLiveOutSize = LiveOut.size();
2836 for (BasicBlock *Succ : successors(BB)) {
2837 assert(Data.LiveIn.count(Succ));
2838 set_union(LiveOut, Data.LiveIn[Succ]);
2840 // assert OutLiveOut is a subset of LiveOut
2841 if (OldLiveOutSize == LiveOut.size()) {
2842 // If the sets are the same size, then we didn't actually add anything
2843 // when unioning our successors LiveIn Thus, the LiveIn of this block
2847 Data.LiveOut[BB] = LiveOut;
2849 // Apply the effects of this basic block
2850 DenseSet<Value *> LiveTmp = LiveOut;
2851 set_union(LiveTmp, Data.LiveSet[BB]);
2852 set_subtract(LiveTmp, Data.KillSet[BB]);
2854 assert(Data.LiveIn.count(BB));
2855 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2856 // assert: OldLiveIn is a subset of LiveTmp
2857 if (OldLiveIn.size() != LiveTmp.size()) {
2858 Data.LiveIn[BB] = LiveTmp;
2859 AddPredsToWorklist(BB);
2861 } // while( !worklist.empty() )
2864 // Sanity check our output against SSA properties. This helps catch any
2865 // missing kills during the above iteration.
2866 for (BasicBlock &BB : F) {
2867 checkBasicSSA(DT, Data, BB);
2872 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2873 StatepointLiveSetTy &Out) {
2875 BasicBlock *BB = Inst->getParent();
2877 // Note: The copy is intentional and required
2878 assert(Data.LiveOut.count(BB));
2879 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2881 // We want to handle the statepoint itself oddly. It's
2882 // call result is not live (normal), nor are it's arguments
2883 // (unless they're used again later). This adjustment is
2884 // specifically what we need to relocate
2885 BasicBlock::reverse_iterator rend(Inst->getIterator());
2886 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2887 LiveOut.erase(Inst);
2888 Out.insert(LiveOut.begin(), LiveOut.end());
2891 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2893 PartiallyConstructedSafepointRecord &Info) {
2894 Instruction *Inst = CS.getInstruction();
2895 StatepointLiveSetTy Updated;
2896 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2899 DenseSet<Value *> Bases;
2900 for (auto KVPair : Info.PointerToBase) {
2901 Bases.insert(KVPair.second);
2904 // We may have base pointers which are now live that weren't before. We need
2905 // to update the PointerToBase structure to reflect this.
2906 for (auto V : Updated)
2907 if (!Info.PointerToBase.count(V)) {
2908 assert(Bases.count(V) && "can't find base for unexpected live value");
2909 Info.PointerToBase[V] = V;
2914 for (auto V : Updated) {
2915 assert(Info.PointerToBase.count(V) &&
2916 "must be able to find base for live value");
2920 // Remove any stale base mappings - this can happen since our liveness is
2921 // more precise then the one inherent in the base pointer analysis
2922 DenseSet<Value *> ToErase;
2923 for (auto KVPair : Info.PointerToBase)
2924 if (!Updated.count(KVPair.first))
2925 ToErase.insert(KVPair.first);
2926 for (auto V : ToErase)
2927 Info.PointerToBase.erase(V);
2930 for (auto KVPair : Info.PointerToBase)
2931 assert(Updated.count(KVPair.first) && "record for non-live value");
2934 Info.LiveSet = Updated;