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/TargetTransformInfo.h"
18 #include "llvm/ADT/SetOperations.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/ADT/DenseSet.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/StringRef.h"
23 #include "llvm/IR/BasicBlock.h"
24 #include "llvm/IR/CallSite.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/IR/IRBuilder.h"
28 #include "llvm/IR/InstIterator.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/Module.h"
33 #include "llvm/IR/Statepoint.h"
34 #include "llvm/IR/Value.h"
35 #include "llvm/IR/Verifier.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Transforms/Scalar.h"
39 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
40 #include "llvm/Transforms/Utils/Cloning.h"
41 #include "llvm/Transforms/Utils/Local.h"
42 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
44 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
48 // Print tracing output
49 static cl::opt<bool> TraceLSP("trace-rewrite-statepoints", cl::Hidden,
52 // Print the liveset found at the insert location
53 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
55 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden,
57 // Print out the base pointers for debugging
58 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden,
61 // Cost threshold measuring when it is profitable to rematerialize value instead
63 static cl::opt<unsigned>
64 RematerializationThreshold("spp-rematerialization-threshold", cl::Hidden,
68 static bool ClobberNonLive = true;
70 static bool ClobberNonLive = false;
72 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
73 cl::location(ClobberNonLive),
77 struct RewriteStatepointsForGC : public ModulePass {
78 static char ID; // Pass identification, replacement for typeid
80 RewriteStatepointsForGC() : ModulePass(ID) {
81 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
83 bool runOnFunction(Function &F);
84 bool runOnModule(Module &M) override {
87 Changed |= runOnFunction(F);
91 void getAnalysisUsage(AnalysisUsage &AU) const override {
92 // We add and rewrite a bunch of instructions, but don't really do much
93 // else. We could in theory preserve a lot more analyses here.
94 AU.addRequired<DominatorTreeWrapperPass>();
95 AU.addRequired<TargetTransformInfoWrapperPass>();
100 char RewriteStatepointsForGC::ID = 0;
102 ModulePass *llvm::createRewriteStatepointsForGCPass() {
103 return new RewriteStatepointsForGC();
106 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
107 "Make relocations explicit at statepoints", false, false)
108 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
109 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
110 "Make relocations explicit at statepoints", false, false)
113 struct GCPtrLivenessData {
114 /// Values defined in this block.
115 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet;
116 /// Values used in this block (and thus live); does not included values
117 /// killed within this block.
118 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet;
120 /// Values live into this basic block (i.e. used by any
121 /// instruction in this basic block or ones reachable from here)
122 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn;
124 /// Values live out of this basic block (i.e. live into
125 /// any successor block)
126 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut;
129 // The type of the internal cache used inside the findBasePointers family
130 // of functions. From the callers perspective, this is an opaque type and
131 // should not be inspected.
133 // In the actual implementation this caches two relations:
134 // - The base relation itself (i.e. this pointer is based on that one)
135 // - The base defining value relation (i.e. before base_phi insertion)
136 // Generally, after the execution of a full findBasePointer call, only the
137 // base relation will remain. Internally, we add a mixture of the two
138 // types, then update all the second type to the first type
139 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
140 typedef DenseSet<llvm::Value *> StatepointLiveSetTy;
141 typedef DenseMap<Instruction *, Value *> RematerializedValueMapTy;
143 struct PartiallyConstructedSafepointRecord {
144 /// The set of values known to be live accross this safepoint
145 StatepointLiveSetTy liveset;
147 /// Mapping from live pointers to a base-defining-value
148 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
150 /// The *new* gc.statepoint instruction itself. This produces the token
151 /// that normal path gc.relocates and the gc.result are tied to.
152 Instruction *StatepointToken;
154 /// Instruction to which exceptional gc relocates are attached
155 /// Makes it easier to iterate through them during relocationViaAlloca.
156 Instruction *UnwindToken;
158 /// Record live values we are rematerialized instead of relocating.
159 /// They are not included into 'liveset' field.
160 /// Maps rematerialized copy to it's original value.
161 RematerializedValueMapTy RematerializedValues;
165 /// Compute the live-in set for every basic block in the function
166 static void computeLiveInValues(DominatorTree &DT, Function &F,
167 GCPtrLivenessData &Data);
169 /// Given results from the dataflow liveness computation, find the set of live
170 /// Values at a particular instruction.
171 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
172 StatepointLiveSetTy &out);
174 // TODO: Once we can get to the GCStrategy, this becomes
175 // Optional<bool> isGCManagedPointer(const Value *V) const override {
177 static bool isGCPointerType(const Type *T) {
178 if (const PointerType *PT = dyn_cast<PointerType>(T))
179 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
180 // GC managed heap. We know that a pointer into this heap needs to be
181 // updated and that no other pointer does.
182 return (1 == PT->getAddressSpace());
186 // Return true if this type is one which a) is a gc pointer or contains a GC
187 // pointer and b) is of a type this code expects to encounter as a live value.
188 // (The insertion code will assert that a type which matches (a) and not (b)
189 // is not encountered.)
190 static bool isHandledGCPointerType(Type *T) {
191 // We fully support gc pointers
192 if (isGCPointerType(T))
194 // We partially support vectors of gc pointers. The code will assert if it
195 // can't handle something.
196 if (auto VT = dyn_cast<VectorType>(T))
197 if (isGCPointerType(VT->getElementType()))
203 /// Returns true if this type contains a gc pointer whether we know how to
204 /// handle that type or not.
205 static bool containsGCPtrType(Type *Ty) {
206 if (isGCPointerType(Ty))
208 if (VectorType *VT = dyn_cast<VectorType>(Ty))
209 return isGCPointerType(VT->getScalarType());
210 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
211 return containsGCPtrType(AT->getElementType());
212 if (StructType *ST = dyn_cast<StructType>(Ty))
214 ST->subtypes().begin(), ST->subtypes().end(),
215 [](Type *SubType) { return containsGCPtrType(SubType); });
219 // Returns true if this is a type which a) is a gc pointer or contains a GC
220 // pointer and b) is of a type which the code doesn't expect (i.e. first class
221 // aggregates). Used to trip assertions.
222 static bool isUnhandledGCPointerType(Type *Ty) {
223 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
227 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
228 if (a->hasName() && b->hasName()) {
229 return -1 == a->getName().compare(b->getName());
230 } else if (a->hasName() && !b->hasName()) {
232 } else if (!a->hasName() && b->hasName()) {
235 // Better than nothing, but not stable
240 // Conservatively identifies any definitions which might be live at the
241 // given instruction. The analysis is performed immediately before the
242 // given instruction. Values defined by that instruction are not considered
243 // live. Values used by that instruction are considered live.
244 static void analyzeParsePointLiveness(
245 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData,
246 const CallSite &CS, PartiallyConstructedSafepointRecord &result) {
247 Instruction *inst = CS.getInstruction();
249 StatepointLiveSetTy liveset;
250 findLiveSetAtInst(inst, OriginalLivenessData, liveset);
253 // Note: This output is used by several of the test cases
254 // The order of elemtns in a set is not stable, put them in a vec and sort
256 SmallVector<Value *, 64> temp;
257 temp.insert(temp.end(), liveset.begin(), liveset.end());
258 std::sort(temp.begin(), temp.end(), order_by_name);
259 errs() << "Live Variables:\n";
260 for (Value *V : temp) {
261 errs() << " " << V->getName(); // no newline
265 if (PrintLiveSetSize) {
266 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
267 errs() << "Number live values: " << liveset.size() << "\n";
269 result.liveset = liveset;
272 static Value *findBaseDefiningValue(Value *I);
274 /// If we can trivially determine that the index specified in the given vector
275 /// is a base pointer, return it. In cases where the entire vector is known to
276 /// consist of base pointers, the entire vector will be returned. This
277 /// indicates that the relevant extractelement is a valid base pointer and
278 /// should be used directly.
279 static Value *findBaseOfVector(Value *I, Value *Index) {
280 assert(I->getType()->isVectorTy() &&
281 cast<VectorType>(I->getType())->getElementType()->isPointerTy() &&
282 "Illegal to ask for the base pointer of a non-pointer type");
284 // Each case parallels findBaseDefiningValue below, see that code for
285 // detailed motivation.
287 if (isa<Argument>(I))
288 // An incoming argument to the function is a base pointer
291 // We shouldn't see the address of a global as a vector value?
292 assert(!isa<GlobalVariable>(I) &&
293 "unexpected global variable found in base of vector");
295 // inlining could possibly introduce phi node that contains
296 // undef if callee has multiple returns
297 if (isa<UndefValue>(I))
298 // utterly meaningless, but useful for dealing with partially optimized
302 // Due to inheritance, this must be _after_ the global variable and undef
304 if (Constant *Con = dyn_cast<Constant>(I)) {
305 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
306 "order of checks wrong!");
307 assert(Con->isNullValue() && "null is the only case which makes sense");
311 if (isa<LoadInst>(I))
314 // For an insert element, we might be able to look through it if we know
315 // something about the indexes, but if the indices are arbitrary values, we
316 // can't without much more extensive scalarization.
317 if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(I)) {
318 Value *InsertIndex = IEI->getOperand(2);
319 // This index is inserting the value, look for it's base
320 if (InsertIndex == Index)
321 return findBaseDefiningValue(IEI->getOperand(1));
322 // Both constant, and can't be equal per above. This insert is definitely
323 // not relevant, look back at the rest of the vector and keep trying.
324 if (isa<ConstantInt>(Index) && isa<ConstantInt>(InsertIndex))
325 return findBaseOfVector(IEI->getOperand(0), Index);
328 // Note: This code is currently rather incomplete. We are essentially only
329 // handling cases where the vector element is trivially a base pointer. We
330 // need to update the entire base pointer construction algorithm to know how
331 // to track vector elements and potentially scalarize, but the case which
332 // would motivate the work hasn't shown up in real workloads yet.
333 llvm_unreachable("no base found for vector element");
336 /// Helper function for findBasePointer - Will return a value which either a)
337 /// defines the base pointer for the input or b) blocks the simple search
338 /// (i.e. a PHI or Select of two derived pointers)
339 static Value *findBaseDefiningValue(Value *I) {
340 assert(I->getType()->isPointerTy() &&
341 "Illegal to ask for the base pointer of a non-pointer type");
343 // This case is a bit of a hack - it only handles extracts from vectors which
344 // trivially contain only base pointers or cases where we can directly match
345 // the index of the original extract element to an insertion into the vector.
346 // See note inside the function for how to improve this.
347 if (auto *EEI = dyn_cast<ExtractElementInst>(I)) {
348 Value *VectorOperand = EEI->getVectorOperand();
349 Value *Index = EEI->getIndexOperand();
350 Value *VectorBase = findBaseOfVector(VectorOperand, Index);
351 // If the result returned is a vector, we know the entire vector must
352 // contain base pointers. In that case, the extractelement is a valid base
354 if (VectorBase->getType()->isVectorTy())
356 // Otherwise, we needed to look through the vector to find the base for
357 // this particular element.
358 assert(VectorBase->getType()->isPointerTy());
362 if (isa<Argument>(I))
363 // An incoming argument to the function is a base pointer
364 // We should have never reached here if this argument isn't an gc value
367 if (isa<GlobalVariable>(I))
371 // inlining could possibly introduce phi node that contains
372 // undef if callee has multiple returns
373 if (isa<UndefValue>(I))
374 // utterly meaningless, but useful for dealing with
375 // partially optimized code.
378 // Due to inheritance, this must be _after_ the global variable and undef
380 if (Constant *Con = dyn_cast<Constant>(I)) {
381 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
382 "order of checks wrong!");
383 // Note: Finding a constant base for something marked for relocation
384 // doesn't really make sense. The most likely case is either a) some
385 // screwed up the address space usage or b) your validating against
386 // compiled C++ code w/o the proper separation. The only real exception
387 // is a null pointer. You could have generic code written to index of
388 // off a potentially null value and have proven it null. We also use
389 // null pointers in dead paths of relocation phis (which we might later
390 // want to find a base pointer for).
391 assert(isa<ConstantPointerNull>(Con) &&
392 "null is the only case which makes sense");
396 if (CastInst *CI = dyn_cast<CastInst>(I)) {
397 Value *Def = CI->stripPointerCasts();
398 // If we find a cast instruction here, it means we've found a cast which is
399 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
400 // handle int->ptr conversion.
401 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
402 return findBaseDefiningValue(Def);
405 if (isa<LoadInst>(I))
406 return I; // The value loaded is an gc base itself
408 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
409 // The base of this GEP is the base
410 return findBaseDefiningValue(GEP->getPointerOperand());
412 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
413 switch (II->getIntrinsicID()) {
414 case Intrinsic::experimental_gc_result_ptr:
416 // fall through to general call handling
418 case Intrinsic::experimental_gc_statepoint:
419 case Intrinsic::experimental_gc_result_float:
420 case Intrinsic::experimental_gc_result_int:
421 llvm_unreachable("these don't produce pointers");
422 case Intrinsic::experimental_gc_relocate: {
423 // Rerunning safepoint insertion after safepoints are already
424 // inserted is not supported. It could probably be made to work,
425 // but why are you doing this? There's no good reason.
426 llvm_unreachable("repeat safepoint insertion is not supported");
428 case Intrinsic::gcroot:
429 // Currently, this mechanism hasn't been extended to work with gcroot.
430 // There's no reason it couldn't be, but I haven't thought about the
431 // implications much.
433 "interaction with the gcroot mechanism is not supported");
436 // We assume that functions in the source language only return base
437 // pointers. This should probably be generalized via attributes to support
438 // both source language and internal functions.
439 if (isa<CallInst>(I) || isa<InvokeInst>(I))
442 // I have absolutely no idea how to implement this part yet. It's not
443 // neccessarily hard, I just haven't really looked at it yet.
444 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
446 if (isa<AtomicCmpXchgInst>(I))
447 // A CAS is effectively a atomic store and load combined under a
448 // predicate. From the perspective of base pointers, we just treat it
452 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
453 "binary ops which don't apply to pointers");
455 // The aggregate ops. Aggregates can either be in the heap or on the
456 // stack, but in either case, this is simply a field load. As a result,
457 // this is a defining definition of the base just like a load is.
458 if (isa<ExtractValueInst>(I))
461 // We should never see an insert vector since that would require we be
462 // tracing back a struct value not a pointer value.
463 assert(!isa<InsertValueInst>(I) &&
464 "Base pointer for a struct is meaningless");
466 // The last two cases here don't return a base pointer. Instead, they
467 // return a value which dynamically selects from amoung several base
468 // derived pointers (each with it's own base potentially). It's the job of
469 // the caller to resolve these.
470 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
471 "missing instruction case in findBaseDefiningValing");
475 /// Returns the base defining value for this value.
476 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
477 Value *&Cached = Cache[I];
479 Cached = findBaseDefiningValue(I);
481 assert(Cache[I] != nullptr);
484 dbgs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
490 /// Return a base pointer for this value if known. Otherwise, return it's
491 /// base defining value.
492 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
493 Value *Def = findBaseDefiningValueCached(I, Cache);
494 auto Found = Cache.find(Def);
495 if (Found != Cache.end()) {
496 // Either a base-of relation, or a self reference. Caller must check.
497 return Found->second;
499 // Only a BDV available
503 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
504 /// is it known to be a base pointer? Or do we need to continue searching.
505 static bool isKnownBaseResult(Value *V) {
506 if (!isa<PHINode>(V) && !isa<SelectInst>(V)) {
507 // no recursion possible
510 if (isa<Instruction>(V) &&
511 cast<Instruction>(V)->getMetadata("is_base_value")) {
512 // This is a previously inserted base phi or select. We know
513 // that this is a base value.
517 // We need to keep searching
521 // TODO: find a better name for this
525 enum Status { Unknown, Base, Conflict };
527 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
528 assert(status != Base || b);
530 PhiState(Value *b) : status(Base), base(b) {}
531 PhiState() : status(Unknown), base(nullptr) {}
533 Status getStatus() const { return status; }
534 Value *getBase() const { return base; }
536 bool isBase() const { return getStatus() == Base; }
537 bool isUnknown() const { return getStatus() == Unknown; }
538 bool isConflict() const { return getStatus() == Conflict; }
540 bool operator==(const PhiState &other) const {
541 return base == other.base && status == other.status;
544 bool operator!=(const PhiState &other) const { return !(*this == other); }
547 errs() << status << " (" << base << " - "
548 << (base ? base->getName() : "nullptr") << "): ";
553 Value *base; // non null only if status == base
556 typedef DenseMap<Value *, PhiState> ConflictStateMapTy;
557 // Values of type PhiState form a lattice, and this is a helper
558 // class that implementes the meet operation. The meat of the meet
559 // operation is implemented in MeetPhiStates::pureMeet
560 class MeetPhiStates {
562 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
563 explicit MeetPhiStates(const ConflictStateMapTy &phiStates)
564 : phiStates(phiStates) {}
566 // Destructively meet the current result with the base V. V can
567 // either be a merge instruction (SelectInst / PHINode), in which
568 // case its status is looked up in the phiStates map; or a regular
569 // SSA value, in which case it is assumed to be a base.
570 void meetWith(Value *V) {
571 PhiState otherState = getStateForBDV(V);
572 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
573 MeetPhiStates::pureMeet(currentResult, otherState)) &&
574 "math is wrong: meet does not commute!");
575 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
578 PhiState getResult() const { return currentResult; }
581 const ConflictStateMapTy &phiStates;
582 PhiState currentResult;
584 /// Return a phi state for a base defining value. We'll generate a new
585 /// base state for known bases and expect to find a cached state otherwise
586 PhiState getStateForBDV(Value *baseValue) {
587 if (isKnownBaseResult(baseValue)) {
588 return PhiState(baseValue);
590 return lookupFromMap(baseValue);
594 PhiState lookupFromMap(Value *V) {
595 auto I = phiStates.find(V);
596 assert(I != phiStates.end() && "lookup failed!");
600 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
601 switch (stateA.getStatus()) {
602 case PhiState::Unknown:
606 assert(stateA.getBase() && "can't be null");
607 if (stateB.isUnknown())
610 if (stateB.isBase()) {
611 if (stateA.getBase() == stateB.getBase()) {
612 assert(stateA == stateB && "equality broken!");
615 return PhiState(PhiState::Conflict);
617 assert(stateB.isConflict() && "only three states!");
618 return PhiState(PhiState::Conflict);
620 case PhiState::Conflict:
623 llvm_unreachable("only three states!");
627 /// For a given value or instruction, figure out what base ptr it's derived
628 /// from. For gc objects, this is simply itself. On success, returns a value
629 /// which is the base pointer. (This is reliable and can be used for
630 /// relocation.) On failure, returns nullptr.
631 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
632 Value *def = findBaseOrBDV(I, cache);
634 if (isKnownBaseResult(def)) {
638 // Here's the rough algorithm:
639 // - For every SSA value, construct a mapping to either an actual base
640 // pointer or a PHI which obscures the base pointer.
641 // - Construct a mapping from PHI to unknown TOP state. Use an
642 // optimistic algorithm to propagate base pointer information. Lattice
647 // When algorithm terminates, all PHIs will either have a single concrete
648 // base or be in a conflict state.
649 // - For every conflict, insert a dummy PHI node without arguments. Add
650 // these to the base[Instruction] = BasePtr mapping. For every
651 // non-conflict, add the actual base.
652 // - For every conflict, add arguments for the base[a] of each input
655 // Note: A simpler form of this would be to add the conflict form of all
656 // PHIs without running the optimistic algorithm. This would be
657 // analougous to pessimistic data flow and would likely lead to an
658 // overall worse solution.
660 ConflictStateMapTy states;
661 states[def] = PhiState();
662 // Recursively fill in all phis & selects reachable from the initial one
663 // for which we don't already know a definite base value for
664 // TODO: This should be rewritten with a worklist
668 // Since we're adding elements to 'states' as we run, we can't keep
669 // iterators into the set.
670 SmallVector<Value *, 16> Keys;
671 Keys.reserve(states.size());
672 for (auto Pair : states) {
673 Value *V = Pair.first;
676 for (Value *v : Keys) {
677 assert(!isKnownBaseResult(v) && "why did it get added?");
678 if (PHINode *phi = dyn_cast<PHINode>(v)) {
679 assert(phi->getNumIncomingValues() > 0 &&
680 "zero input phis are illegal");
681 for (Value *InVal : phi->incoming_values()) {
682 Value *local = findBaseOrBDV(InVal, cache);
683 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
684 states[local] = PhiState();
688 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
689 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
690 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
691 states[local] = PhiState();
694 local = findBaseOrBDV(sel->getFalseValue(), cache);
695 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
696 states[local] = PhiState();
704 errs() << "States after initialization:\n";
705 for (auto Pair : states) {
706 Instruction *v = cast<Instruction>(Pair.first);
707 PhiState state = Pair.second;
713 // TODO: come back and revisit the state transitions around inputs which
714 // have reached conflict state. The current version seems too conservative.
716 bool progress = true;
719 size_t oldSize = states.size();
722 // We're only changing keys in this loop, thus safe to keep iterators
723 for (auto Pair : states) {
724 MeetPhiStates calculateMeet(states);
725 Value *v = Pair.first;
726 assert(!isKnownBaseResult(v) && "why did it get added?");
727 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
728 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
729 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
731 for (Value *Val : cast<PHINode>(v)->incoming_values())
732 calculateMeet.meetWith(findBaseOrBDV(Val, cache));
734 PhiState oldState = states[v];
735 PhiState newState = calculateMeet.getResult();
736 if (oldState != newState) {
738 states[v] = newState;
742 assert(oldSize <= states.size());
743 assert(oldSize == states.size() || progress);
747 errs() << "States after meet iteration:\n";
748 for (auto Pair : states) {
749 Instruction *v = cast<Instruction>(Pair.first);
750 PhiState state = Pair.second;
756 // Insert Phis for all conflicts
757 // We want to keep naming deterministic in the loop that follows, so
758 // sort the keys before iteration. This is useful in allowing us to
759 // write stable tests. Note that there is no invalidation issue here.
760 SmallVector<Value *, 16> Keys;
761 Keys.reserve(states.size());
762 for (auto Pair : states) {
763 Value *V = Pair.first;
766 std::sort(Keys.begin(), Keys.end(), order_by_name);
767 // TODO: adjust naming patterns to avoid this order of iteration dependency
768 for (Value *V : Keys) {
769 Instruction *v = cast<Instruction>(V);
770 PhiState state = states[V];
771 assert(!isKnownBaseResult(v) && "why did it get added?");
772 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
773 if (!state.isConflict())
776 if (isa<PHINode>(v)) {
778 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
779 assert(num_preds > 0 && "how did we reach here");
780 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
781 // Add metadata marking this as a base value
782 auto *const_1 = ConstantInt::get(
784 v->getParent()->getParent()->getParent()->getContext()),
786 auto MDConst = ConstantAsMetadata::get(const_1);
787 MDNode *md = MDNode::get(
788 v->getParent()->getParent()->getParent()->getContext(), MDConst);
789 phi->setMetadata("is_base_value", md);
790 states[v] = PhiState(PhiState::Conflict, phi);
792 SelectInst *sel = cast<SelectInst>(v);
793 // The undef will be replaced later
794 UndefValue *undef = UndefValue::get(sel->getType());
795 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
796 undef, "base_select", sel);
797 // Add metadata marking this as a base value
798 auto *const_1 = ConstantInt::get(
800 v->getParent()->getParent()->getParent()->getContext()),
802 auto MDConst = ConstantAsMetadata::get(const_1);
803 MDNode *md = MDNode::get(
804 v->getParent()->getParent()->getParent()->getContext(), MDConst);
805 basesel->setMetadata("is_base_value", md);
806 states[v] = PhiState(PhiState::Conflict, basesel);
810 // Fixup all the inputs of the new PHIs
811 for (auto Pair : states) {
812 Instruction *v = cast<Instruction>(Pair.first);
813 PhiState state = Pair.second;
815 assert(!isKnownBaseResult(v) && "why did it get added?");
816 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
817 if (!state.isConflict())
820 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
821 PHINode *phi = cast<PHINode>(v);
822 unsigned NumPHIValues = phi->getNumIncomingValues();
823 for (unsigned i = 0; i < NumPHIValues; i++) {
824 Value *InVal = phi->getIncomingValue(i);
825 BasicBlock *InBB = phi->getIncomingBlock(i);
827 // If we've already seen InBB, add the same incoming value
828 // we added for it earlier. The IR verifier requires phi
829 // nodes with multiple entries from the same basic block
830 // to have the same incoming value for each of those
831 // entries. If we don't do this check here and basephi
832 // has a different type than base, we'll end up adding two
833 // bitcasts (and hence two distinct values) as incoming
834 // values for the same basic block.
836 int blockIndex = basephi->getBasicBlockIndex(InBB);
837 if (blockIndex != -1) {
838 Value *oldBase = basephi->getIncomingValue(blockIndex);
839 basephi->addIncoming(oldBase, InBB);
841 Value *base = findBaseOrBDV(InVal, cache);
842 if (!isKnownBaseResult(base)) {
843 // Either conflict or base.
844 assert(states.count(base));
845 base = states[base].getBase();
846 assert(base != nullptr && "unknown PhiState!");
849 // In essense this assert states: the only way two
850 // values incoming from the same basic block may be
851 // different is by being different bitcasts of the same
852 // value. A cleanup that remains TODO is changing
853 // findBaseOrBDV to return an llvm::Value of the correct
854 // type (and still remain pure). This will remove the
855 // need to add bitcasts.
856 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
857 "sanity -- findBaseOrBDV should be pure!");
862 // Find either the defining value for the PHI or the normal base for
864 Value *base = findBaseOrBDV(InVal, cache);
865 if (!isKnownBaseResult(base)) {
866 // Either conflict or base.
867 assert(states.count(base));
868 base = states[base].getBase();
869 assert(base != nullptr && "unknown PhiState!");
871 assert(base && "can't be null");
872 // Must use original input BB since base may not be Instruction
873 // The cast is needed since base traversal may strip away bitcasts
874 if (base->getType() != basephi->getType()) {
875 base = new BitCastInst(base, basephi->getType(), "cast",
876 InBB->getTerminator());
878 basephi->addIncoming(base, InBB);
880 assert(basephi->getNumIncomingValues() == NumPHIValues);
882 SelectInst *basesel = cast<SelectInst>(state.getBase());
883 SelectInst *sel = cast<SelectInst>(v);
884 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
885 // something more safe and less hacky.
886 for (int i = 1; i <= 2; i++) {
887 Value *InVal = sel->getOperand(i);
888 // Find either the defining value for the PHI or the normal base for
890 Value *base = findBaseOrBDV(InVal, cache);
891 if (!isKnownBaseResult(base)) {
892 // Either conflict or base.
893 assert(states.count(base));
894 base = states[base].getBase();
895 assert(base != nullptr && "unknown PhiState!");
897 assert(base && "can't be null");
898 // Must use original input BB since base may not be Instruction
899 // The cast is needed since base traversal may strip away bitcasts
900 if (base->getType() != basesel->getType()) {
901 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
903 basesel->setOperand(i, base);
908 // Cache all of our results so we can cheaply reuse them
909 // NOTE: This is actually two caches: one of the base defining value
910 // relation and one of the base pointer relation! FIXME
911 for (auto item : states) {
912 Value *v = item.first;
913 Value *base = item.second.getBase();
915 assert(!isKnownBaseResult(v) && "why did it get added?");
918 std::string fromstr =
919 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
921 errs() << "Updating base value cache"
922 << " for: " << (v->hasName() ? v->getName() : "")
923 << " from: " << fromstr
924 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
927 assert(isKnownBaseResult(base) &&
928 "must be something we 'know' is a base pointer");
929 if (cache.count(v)) {
930 // Once we transition from the BDV relation being store in the cache to
931 // the base relation being stored, it must be stable
932 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
933 "base relation should be stable");
937 assert(cache.find(def) != cache.end());
941 // For a set of live pointers (base and/or derived), identify the base
942 // pointer of the object which they are derived from. This routine will
943 // mutate the IR graph as needed to make the 'base' pointer live at the
944 // definition site of 'derived'. This ensures that any use of 'derived' can
945 // also use 'base'. This may involve the insertion of a number of
946 // additional PHI nodes.
948 // preconditions: live is a set of pointer type Values
950 // side effects: may insert PHI nodes into the existing CFG, will preserve
951 // CFG, will not remove or mutate any existing nodes
953 // post condition: PointerToBase contains one (derived, base) pair for every
954 // pointer in live. Note that derived can be equal to base if the original
955 // pointer was a base pointer.
957 findBasePointers(const StatepointLiveSetTy &live,
958 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
959 DominatorTree *DT, DefiningValueMapTy &DVCache) {
960 // For the naming of values inserted to be deterministic - which makes for
961 // much cleaner and more stable tests - we need to assign an order to the
962 // live values. DenseSets do not provide a deterministic order across runs.
963 SmallVector<Value *, 64> Temp;
964 Temp.insert(Temp.end(), live.begin(), live.end());
965 std::sort(Temp.begin(), Temp.end(), order_by_name);
966 for (Value *ptr : Temp) {
967 Value *base = findBasePointer(ptr, DVCache);
968 assert(base && "failed to find base pointer");
969 PointerToBase[ptr] = base;
970 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
971 DT->dominates(cast<Instruction>(base)->getParent(),
972 cast<Instruction>(ptr)->getParent())) &&
973 "The base we found better dominate the derived pointer");
975 // If you see this trip and like to live really dangerously, the code should
976 // be correct, just with idioms the verifier can't handle. You can try
977 // disabling the verifier at your own substaintial risk.
978 assert(!isa<ConstantPointerNull>(base) &&
979 "the relocation code needs adjustment to handle the relocation of "
980 "a null pointer constant without causing false positives in the "
981 "safepoint ir verifier.");
985 /// Find the required based pointers (and adjust the live set) for the given
987 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
989 PartiallyConstructedSafepointRecord &result) {
990 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
991 findBasePointers(result.liveset, PointerToBase, &DT, DVCache);
993 if (PrintBasePointers) {
994 // Note: Need to print these in a stable order since this is checked in
996 errs() << "Base Pairs (w/o Relocation):\n";
997 SmallVector<Value *, 64> Temp;
998 Temp.reserve(PointerToBase.size());
999 for (auto Pair : PointerToBase) {
1000 Temp.push_back(Pair.first);
1002 std::sort(Temp.begin(), Temp.end(), order_by_name);
1003 for (Value *Ptr : Temp) {
1004 Value *Base = PointerToBase[Ptr];
1005 errs() << " derived %" << Ptr->getName() << " base %" << Base->getName()
1010 result.PointerToBase = PointerToBase;
1013 /// Given an updated version of the dataflow liveness results, update the
1014 /// liveset and base pointer maps for the call site CS.
1015 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1017 PartiallyConstructedSafepointRecord &result);
1019 static void recomputeLiveInValues(
1020 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1021 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1022 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1023 // again. The old values are still live and will help it stablize quickly.
1024 GCPtrLivenessData RevisedLivenessData;
1025 computeLiveInValues(DT, F, RevisedLivenessData);
1026 for (size_t i = 0; i < records.size(); i++) {
1027 struct PartiallyConstructedSafepointRecord &info = records[i];
1028 const CallSite &CS = toUpdate[i];
1029 recomputeLiveInValues(RevisedLivenessData, CS, info);
1033 // When inserting gc.relocate calls, we need to ensure there are no uses
1034 // of the original value between the gc.statepoint and the gc.relocate call.
1035 // One case which can arise is a phi node starting one of the successor blocks.
1036 // We also need to be able to insert the gc.relocates only on the path which
1037 // goes through the statepoint. We might need to split an edge to make this
1040 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1041 DominatorTree &DT) {
1042 BasicBlock *Ret = BB;
1043 if (!BB->getUniquePredecessor()) {
1044 Ret = SplitBlockPredecessors(BB, InvokeParent, "", nullptr, &DT);
1047 // Now that 'ret' has unique predecessor we can safely remove all phi nodes
1049 FoldSingleEntryPHINodes(Ret);
1050 assert(!isa<PHINode>(Ret->begin()));
1052 // At this point, we can safely insert a gc.relocate as the first instruction
1053 // in Ret if needed.
1057 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1058 auto itr = std::find(livevec.begin(), livevec.end(), val);
1059 assert(livevec.end() != itr);
1060 size_t index = std::distance(livevec.begin(), itr);
1061 assert(index < livevec.size());
1065 // Create new attribute set containing only attributes which can be transfered
1066 // from original call to the safepoint.
1067 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1070 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1071 unsigned index = AS.getSlotIndex(Slot);
1073 if (index == AttributeSet::ReturnIndex ||
1074 index == AttributeSet::FunctionIndex) {
1076 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1078 Attribute attr = *it;
1080 // Do not allow certain attributes - just skip them
1081 // Safepoint can not be read only or read none.
1082 if (attr.hasAttribute(Attribute::ReadNone) ||
1083 attr.hasAttribute(Attribute::ReadOnly))
1086 ret = ret.addAttributes(
1087 AS.getContext(), index,
1088 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1092 // Just skip parameter attributes for now
1098 /// Helper function to place all gc relocates necessary for the given
1101 /// liveVariables - list of variables to be relocated.
1102 /// liveStart - index of the first live variable.
1103 /// basePtrs - base pointers.
1104 /// statepointToken - statepoint instruction to which relocates should be
1106 /// Builder - Llvm IR builder to be used to construct new calls.
1107 static void CreateGCRelocates(ArrayRef<llvm::Value *> LiveVariables,
1108 const int LiveStart,
1109 ArrayRef<llvm::Value *> BasePtrs,
1110 Instruction *StatepointToken,
1111 IRBuilder<> Builder) {
1112 SmallVector<Instruction *, 64> NewDefs;
1113 NewDefs.reserve(LiveVariables.size());
1115 Module *M = StatepointToken->getParent()->getParent()->getParent();
1117 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1118 // We generate a (potentially) unique declaration for every pointer type
1119 // combination. This results is some blow up the function declarations in
1120 // the IR, but removes the need for argument bitcasts which shrinks the IR
1121 // greatly and makes it much more readable.
1122 SmallVector<Type *, 1> Types; // one per 'any' type
1123 // All gc_relocate are set to i8 addrspace(1)* type. This could help avoid
1124 // cases where the actual value's type mangling is not supported by llvm. A
1125 // bitcast is added later to convert gc_relocate to the actual value's type.
1126 Types.push_back(Type::getInt8PtrTy(M->getContext(), 1));
1127 Value *GCRelocateDecl = Intrinsic::getDeclaration(
1128 M, Intrinsic::experimental_gc_relocate, Types);
1130 // Generate the gc.relocate call and save the result
1132 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1133 LiveStart + find_index(LiveVariables, BasePtrs[i]));
1134 Value *LiveIdx = ConstantInt::get(
1135 Type::getInt32Ty(M->getContext()),
1136 LiveStart + find_index(LiveVariables, LiveVariables[i]));
1138 // only specify a debug name if we can give a useful one
1139 Value *Reloc = Builder.CreateCall(
1140 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1141 LiveVariables[i]->hasName() ? LiveVariables[i]->getName() + ".relocated"
1143 // Trick CodeGen into thinking there are lots of free registers at this
1145 cast<CallInst>(Reloc)->setCallingConv(CallingConv::Cold);
1147 NewDefs.push_back(cast<Instruction>(Reloc));
1149 assert(NewDefs.size() == LiveVariables.size() &&
1150 "missing or extra redefinition at safepoint");
1154 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1155 const SmallVectorImpl<llvm::Value *> &basePtrs,
1156 const SmallVectorImpl<llvm::Value *> &liveVariables,
1158 PartiallyConstructedSafepointRecord &result) {
1159 assert(basePtrs.size() == liveVariables.size());
1160 assert(isStatepoint(CS) &&
1161 "This method expects to be rewriting a statepoint");
1163 BasicBlock *BB = CS.getInstruction()->getParent();
1165 Function *F = BB->getParent();
1166 assert(F && "must be set");
1167 Module *M = F->getParent();
1169 assert(M && "must be set");
1171 // We're not changing the function signature of the statepoint since the gc
1172 // arguments go into the var args section.
1173 Function *gc_statepoint_decl = CS.getCalledFunction();
1175 // Then go ahead and use the builder do actually do the inserts. We insert
1176 // immediately before the previous instruction under the assumption that all
1177 // arguments will be available here. We can't insert afterwards since we may
1178 // be replacing a terminator.
1179 Instruction *insertBefore = CS.getInstruction();
1180 IRBuilder<> Builder(insertBefore);
1181 // Copy all of the arguments from the original statepoint - this includes the
1182 // target, call args, and deopt args
1183 SmallVector<llvm::Value *, 64> args;
1184 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1185 // TODO: Clear the 'needs rewrite' flag
1187 // add all the pointers to be relocated (gc arguments)
1188 // Capture the start of the live variable list for use in the gc_relocates
1189 const int live_start = args.size();
1190 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1192 // Create the statepoint given all the arguments
1193 Instruction *token = nullptr;
1194 AttributeSet return_attributes;
1196 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1198 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1199 call->setTailCall(toReplace->isTailCall());
1200 call->setCallingConv(toReplace->getCallingConv());
1202 // Currently we will fail on parameter attributes and on certain
1203 // function attributes.
1204 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1205 // In case if we can handle this set of sttributes - set up function attrs
1206 // directly on statepoint and return attrs later for gc_result intrinsic.
1207 call->setAttributes(new_attrs.getFnAttributes());
1208 return_attributes = new_attrs.getRetAttributes();
1212 // Put the following gc_result and gc_relocate calls immediately after the
1213 // the old call (which we're about to delete)
1214 BasicBlock::iterator next(toReplace);
1215 assert(BB->end() != next && "not a terminator, must have next");
1217 Instruction *IP = &*(next);
1218 Builder.SetInsertPoint(IP);
1219 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1222 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1224 // Insert the new invoke into the old block. We'll remove the old one in a
1225 // moment at which point this will become the new terminator for the
1227 InvokeInst *invoke = InvokeInst::Create(
1228 gc_statepoint_decl, toReplace->getNormalDest(),
1229 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1230 invoke->setCallingConv(toReplace->getCallingConv());
1232 // Currently we will fail on parameter attributes and on certain
1233 // function attributes.
1234 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1235 // In case if we can handle this set of sttributes - set up function attrs
1236 // directly on statepoint and return attrs later for gc_result intrinsic.
1237 invoke->setAttributes(new_attrs.getFnAttributes());
1238 return_attributes = new_attrs.getRetAttributes();
1242 // Generate gc relocates in exceptional path
1243 BasicBlock *unwindBlock = toReplace->getUnwindDest();
1244 assert(!isa<PHINode>(unwindBlock->begin()) &&
1245 unwindBlock->getUniquePredecessor() &&
1246 "can't safely insert in this block!");
1248 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1249 Builder.SetInsertPoint(IP);
1250 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1252 // Extract second element from landingpad return value. We will attach
1253 // exceptional gc relocates to it.
1254 const unsigned idx = 1;
1255 Instruction *exceptional_token =
1256 cast<Instruction>(Builder.CreateExtractValue(
1257 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1258 result.UnwindToken = exceptional_token;
1260 // Just throw away return value. We will use the one we got for normal
1262 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1263 exceptional_token, Builder);
1265 // Generate gc relocates and returns for normal block
1266 BasicBlock *normalDest = toReplace->getNormalDest();
1267 assert(!isa<PHINode>(normalDest->begin()) &&
1268 normalDest->getUniquePredecessor() &&
1269 "can't safely insert in this block!");
1271 IP = &*(normalDest->getFirstInsertionPt());
1272 Builder.SetInsertPoint(IP);
1274 // gc relocates will be generated later as if it were regular call
1279 // Take the name of the original value call if it had one.
1280 token->takeName(CS.getInstruction());
1282 // The GCResult is already inserted, we just need to find it
1284 Instruction *toReplace = CS.getInstruction();
1285 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1286 "only valid use before rewrite is gc.result");
1287 assert(!toReplace->hasOneUse() ||
1288 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1291 // Update the gc.result of the original statepoint (if any) to use the newly
1292 // inserted statepoint. This is safe to do here since the token can't be
1293 // considered a live reference.
1294 CS.getInstruction()->replaceAllUsesWith(token);
1296 result.StatepointToken = token;
1298 // Second, create a gc.relocate for every live variable
1299 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1303 struct name_ordering {
1306 bool operator()(name_ordering const &a, name_ordering const &b) {
1307 return -1 == a.derived->getName().compare(b.derived->getName());
1311 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1312 SmallVectorImpl<Value *> &livevec) {
1313 assert(basevec.size() == livevec.size());
1315 SmallVector<name_ordering, 64> temp;
1316 for (size_t i = 0; i < basevec.size(); i++) {
1318 v.base = basevec[i];
1319 v.derived = livevec[i];
1322 std::sort(temp.begin(), temp.end(), name_ordering());
1323 for (size_t i = 0; i < basevec.size(); i++) {
1324 basevec[i] = temp[i].base;
1325 livevec[i] = temp[i].derived;
1329 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1330 // which make the relocations happening at this safepoint explicit.
1332 // WARNING: Does not do any fixup to adjust users of the original live
1333 // values. That's the callers responsibility.
1335 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1336 PartiallyConstructedSafepointRecord &result) {
1337 auto liveset = result.liveset;
1338 auto PointerToBase = result.PointerToBase;
1340 // Convert to vector for efficient cross referencing.
1341 SmallVector<Value *, 64> basevec, livevec;
1342 livevec.reserve(liveset.size());
1343 basevec.reserve(liveset.size());
1344 for (Value *L : liveset) {
1345 livevec.push_back(L);
1347 assert(PointerToBase.find(L) != PointerToBase.end());
1348 Value *base = PointerToBase[L];
1349 basevec.push_back(base);
1351 assert(livevec.size() == basevec.size());
1353 // To make the output IR slightly more stable (for use in diffs), ensure a
1354 // fixed order of the values in the safepoint (by sorting the value name).
1355 // The order is otherwise meaningless.
1356 stablize_order(basevec, livevec);
1358 // Do the actual rewriting and delete the old statepoint
1359 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1360 CS.getInstruction()->eraseFromParent();
1363 // Helper function for the relocationViaAlloca.
1364 // It receives iterator to the statepoint gc relocates and emits store to the
1366 // location (via allocaMap) for the each one of them.
1367 // Add visited values into the visitedLiveValues set we will later use them
1368 // for sanity check.
1370 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1371 DenseMap<Value *, Value *> &AllocaMap,
1372 DenseSet<Value *> &VisitedLiveValues) {
1374 for (User *U : GCRelocs) {
1375 if (!isa<IntrinsicInst>(U))
1378 IntrinsicInst *RelocatedValue = cast<IntrinsicInst>(U);
1380 // We only care about relocates
1381 if (RelocatedValue->getIntrinsicID() !=
1382 Intrinsic::experimental_gc_relocate) {
1386 GCRelocateOperands RelocateOperands(RelocatedValue);
1387 Value *OriginalValue =
1388 const_cast<Value *>(RelocateOperands.getDerivedPtr());
1389 assert(AllocaMap.count(OriginalValue));
1390 Value *Alloca = AllocaMap[OriginalValue];
1392 // Emit store into the related alloca
1393 // All gc_relocate are i8 addrspace(1)* typed, and it must be bitcasted to
1394 // the correct type according to alloca.
1395 assert(RelocatedValue->getNextNode() && "Should always have one since it's not a terminator");
1396 IRBuilder<> Builder(RelocatedValue->getNextNode());
1397 Value *CastedRelocatedValue =
1398 Builder.CreateBitCast(RelocatedValue, cast<AllocaInst>(Alloca)->getAllocatedType(),
1399 RelocatedValue->hasName() ? RelocatedValue->getName() + ".casted" : "");
1401 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1402 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1405 VisitedLiveValues.insert(OriginalValue);
1410 // Helper function for the "relocationViaAlloca". Similar to the
1411 // "insertRelocationStores" but works for rematerialized values.
1413 insertRematerializationStores(
1414 RematerializedValueMapTy RematerializedValues,
1415 DenseMap<Value *, Value *> &AllocaMap,
1416 DenseSet<Value *> &VisitedLiveValues) {
1418 for (auto RematerializedValuePair: RematerializedValues) {
1419 Instruction *RematerializedValue = RematerializedValuePair.first;
1420 Value *OriginalValue = RematerializedValuePair.second;
1422 assert(AllocaMap.count(OriginalValue) &&
1423 "Can not find alloca for rematerialized value");
1424 Value *Alloca = AllocaMap[OriginalValue];
1426 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1427 Store->insertAfter(RematerializedValue);
1430 VisitedLiveValues.insert(OriginalValue);
1435 /// do all the relocation update via allocas and mem2reg
1436 static void relocationViaAlloca(
1437 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1438 ArrayRef<struct PartiallyConstructedSafepointRecord> Records) {
1440 // record initial number of (static) allocas; we'll check we have the same
1441 // number when we get done.
1442 int InitialAllocaNum = 0;
1443 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1445 if (isa<AllocaInst>(*I))
1449 // TODO-PERF: change data structures, reserve
1450 DenseMap<Value *, Value *> AllocaMap;
1451 SmallVector<AllocaInst *, 200> PromotableAllocas;
1452 // Used later to chack that we have enough allocas to store all values
1453 std::size_t NumRematerializedValues = 0;
1454 PromotableAllocas.reserve(Live.size());
1456 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1457 // "PromotableAllocas"
1458 auto emitAllocaFor = [&](Value *LiveValue) {
1459 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "",
1460 F.getEntryBlock().getFirstNonPHI());
1461 AllocaMap[LiveValue] = Alloca;
1462 PromotableAllocas.push_back(Alloca);
1465 // emit alloca for each live gc pointer
1466 for (unsigned i = 0; i < Live.size(); i++) {
1467 emitAllocaFor(Live[i]);
1470 // emit allocas for rematerialized values
1471 for (size_t i = 0; i < Records.size(); i++) {
1472 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1474 for (auto RematerializedValuePair : Info.RematerializedValues) {
1475 Value *OriginalValue = RematerializedValuePair.second;
1476 if (AllocaMap.count(OriginalValue) != 0)
1479 emitAllocaFor(OriginalValue);
1480 ++NumRematerializedValues;
1484 // The next two loops are part of the same conceptual operation. We need to
1485 // insert a store to the alloca after the original def and at each
1486 // redefinition. We need to insert a load before each use. These are split
1487 // into distinct loops for performance reasons.
1489 // update gc pointer after each statepoint
1490 // either store a relocated value or null (if no relocated value found for
1491 // this gc pointer and it is not a gc_result)
1492 // this must happen before we update the statepoint with load of alloca
1493 // otherwise we lose the link between statepoint and old def
1494 for (size_t i = 0; i < Records.size(); i++) {
1495 const struct PartiallyConstructedSafepointRecord &Info = Records[i];
1496 Value *Statepoint = Info.StatepointToken;
1498 // This will be used for consistency check
1499 DenseSet<Value *> VisitedLiveValues;
1501 // Insert stores for normal statepoint gc relocates
1502 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1504 // In case if it was invoke statepoint
1505 // we will insert stores for exceptional path gc relocates.
1506 if (isa<InvokeInst>(Statepoint)) {
1507 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1511 // Do similar thing with rematerialized values
1512 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1515 if (ClobberNonLive) {
1516 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1517 // the gc.statepoint. This will turn some subtle GC problems into
1518 // slightly easier to debug SEGVs. Note that on large IR files with
1519 // lots of gc.statepoints this is extremely costly both memory and time
1521 SmallVector<AllocaInst *, 64> ToClobber;
1522 for (auto Pair : AllocaMap) {
1523 Value *Def = Pair.first;
1524 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1526 // This value was relocated
1527 if (VisitedLiveValues.count(Def)) {
1530 ToClobber.push_back(Alloca);
1533 auto InsertClobbersAt = [&](Instruction *IP) {
1534 for (auto *AI : ToClobber) {
1535 auto AIType = cast<PointerType>(AI->getType());
1536 auto PT = cast<PointerType>(AIType->getElementType());
1537 Constant *CPN = ConstantPointerNull::get(PT);
1538 StoreInst *Store = new StoreInst(CPN, AI);
1539 Store->insertBefore(IP);
1543 // Insert the clobbering stores. These may get intermixed with the
1544 // gc.results and gc.relocates, but that's fine.
1545 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1546 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1547 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1549 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1551 InsertClobbersAt(Next);
1555 // update use with load allocas and add store for gc_relocated
1556 for (auto Pair : AllocaMap) {
1557 Value *Def = Pair.first;
1558 Value *Alloca = Pair.second;
1560 // we pre-record the uses of allocas so that we dont have to worry about
1562 // that change the user information.
1563 SmallVector<Instruction *, 20> Uses;
1564 // PERF: trade a linear scan for repeated reallocation
1565 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1566 for (User *U : Def->users()) {
1567 if (!isa<ConstantExpr>(U)) {
1568 // If the def has a ConstantExpr use, then the def is either a
1569 // ConstantExpr use itself or null. In either case
1570 // (recursively in the first, directly in the second), the oop
1571 // it is ultimately dependent on is null and this particular
1572 // use does not need to be fixed up.
1573 Uses.push_back(cast<Instruction>(U));
1577 std::sort(Uses.begin(), Uses.end());
1578 auto Last = std::unique(Uses.begin(), Uses.end());
1579 Uses.erase(Last, Uses.end());
1581 for (Instruction *Use : Uses) {
1582 if (isa<PHINode>(Use)) {
1583 PHINode *Phi = cast<PHINode>(Use);
1584 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1585 if (Def == Phi->getIncomingValue(i)) {
1586 LoadInst *Load = new LoadInst(
1587 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1588 Phi->setIncomingValue(i, Load);
1592 LoadInst *Load = new LoadInst(Alloca, "", Use);
1593 Use->replaceUsesOfWith(Def, Load);
1597 // emit store for the initial gc value
1598 // store must be inserted after load, otherwise store will be in alloca's
1599 // use list and an extra load will be inserted before it
1600 StoreInst *Store = new StoreInst(Def, Alloca);
1601 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1602 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1603 // InvokeInst is a TerminatorInst so the store need to be inserted
1604 // into its normal destination block.
1605 BasicBlock *NormalDest = Invoke->getNormalDest();
1606 Store->insertBefore(NormalDest->getFirstNonPHI());
1608 assert(!Inst->isTerminator() &&
1609 "The only TerminatorInst that can produce a value is "
1610 "InvokeInst which is handled above.");
1611 Store->insertAfter(Inst);
1614 assert(isa<Argument>(Def));
1615 Store->insertAfter(cast<Instruction>(Alloca));
1619 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1620 "we must have the same allocas with lives");
1621 if (!PromotableAllocas.empty()) {
1622 // apply mem2reg to promote alloca to SSA
1623 PromoteMemToReg(PromotableAllocas, DT);
1627 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E;
1629 if (isa<AllocaInst>(*I))
1631 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1635 /// Implement a unique function which doesn't require we sort the input
1636 /// vector. Doing so has the effect of changing the output of a couple of
1637 /// tests in ways which make them less useful in testing fused safepoints.
1638 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1640 SmallVector<T, 128> TempVec;
1641 TempVec.reserve(Vec.size());
1642 for (auto Element : Vec)
1643 TempVec.push_back(Element);
1645 for (auto V : TempVec) {
1646 if (Seen.insert(V).second) {
1652 /// Insert holders so that each Value is obviously live through the entire
1653 /// lifetime of the call.
1654 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1655 SmallVectorImpl<CallInst *> &Holders) {
1657 // No values to hold live, might as well not insert the empty holder
1660 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1661 // Use a dummy vararg function to actually hold the values live
1662 Function *Func = cast<Function>(M->getOrInsertFunction(
1663 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1665 // For call safepoints insert dummy calls right after safepoint
1666 BasicBlock::iterator Next(CS.getInstruction());
1668 Holders.push_back(CallInst::Create(Func, Values, "", Next));
1671 // For invoke safepooints insert dummy calls both in normal and
1672 // exceptional destination blocks
1673 auto *II = cast<InvokeInst>(CS.getInstruction());
1674 Holders.push_back(CallInst::Create(
1675 Func, Values, "", II->getNormalDest()->getFirstInsertionPt()));
1676 Holders.push_back(CallInst::Create(
1677 Func, Values, "", II->getUnwindDest()->getFirstInsertionPt()));
1680 static void findLiveReferences(
1681 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1682 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1683 GCPtrLivenessData OriginalLivenessData;
1684 computeLiveInValues(DT, F, OriginalLivenessData);
1685 for (size_t i = 0; i < records.size(); i++) {
1686 struct PartiallyConstructedSafepointRecord &info = records[i];
1687 const CallSite &CS = toUpdate[i];
1688 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info);
1692 /// Remove any vector of pointers from the liveset by scalarizing them over the
1693 /// statepoint instruction. Adds the scalarized pieces to the liveset. It
1694 /// would be preferrable to include the vector in the statepoint itself, but
1695 /// the lowering code currently does not handle that. Extending it would be
1696 /// slightly non-trivial since it requires a format change. Given how rare
1697 /// such cases are (for the moment?) scalarizing is an acceptable comprimise.
1698 static void splitVectorValues(Instruction *StatepointInst,
1699 StatepointLiveSetTy &LiveSet, DominatorTree &DT) {
1700 SmallVector<Value *, 16> ToSplit;
1701 for (Value *V : LiveSet)
1702 if (isa<VectorType>(V->getType()))
1703 ToSplit.push_back(V);
1705 if (ToSplit.empty())
1708 Function &F = *(StatepointInst->getParent()->getParent());
1710 DenseMap<Value *, AllocaInst *> AllocaMap;
1711 // First is normal return, second is exceptional return (invoke only)
1712 DenseMap<Value *, std::pair<Value *, Value *>> Replacements;
1713 for (Value *V : ToSplit) {
1716 AllocaInst *Alloca =
1717 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI());
1718 AllocaMap[V] = Alloca;
1720 VectorType *VT = cast<VectorType>(V->getType());
1721 IRBuilder<> Builder(StatepointInst);
1722 SmallVector<Value *, 16> Elements;
1723 for (unsigned i = 0; i < VT->getNumElements(); i++)
1724 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i)));
1725 LiveSet.insert(Elements.begin(), Elements.end());
1727 auto InsertVectorReform = [&](Instruction *IP) {
1728 Builder.SetInsertPoint(IP);
1729 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1730 Value *ResultVec = UndefValue::get(VT);
1731 for (unsigned i = 0; i < VT->getNumElements(); i++)
1732 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i],
1733 Builder.getInt32(i));
1737 if (isa<CallInst>(StatepointInst)) {
1738 BasicBlock::iterator Next(StatepointInst);
1740 Instruction *IP = &*(Next);
1741 Replacements[V].first = InsertVectorReform(IP);
1742 Replacements[V].second = nullptr;
1744 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst);
1745 // We've already normalized - check that we don't have shared destination
1747 BasicBlock *NormalDest = Invoke->getNormalDest();
1748 assert(!isa<PHINode>(NormalDest->begin()));
1749 BasicBlock *UnwindDest = Invoke->getUnwindDest();
1750 assert(!isa<PHINode>(UnwindDest->begin()));
1751 // Insert insert element sequences in both successors
1752 Instruction *IP = &*(NormalDest->getFirstInsertionPt());
1753 Replacements[V].first = InsertVectorReform(IP);
1754 IP = &*(UnwindDest->getFirstInsertionPt());
1755 Replacements[V].second = InsertVectorReform(IP);
1758 for (Value *V : ToSplit) {
1759 AllocaInst *Alloca = AllocaMap[V];
1761 // Capture all users before we start mutating use lists
1762 SmallVector<Instruction *, 16> Users;
1763 for (User *U : V->users())
1764 Users.push_back(cast<Instruction>(U));
1766 for (Instruction *I : Users) {
1767 if (auto Phi = dyn_cast<PHINode>(I)) {
1768 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++)
1769 if (V == Phi->getIncomingValue(i)) {
1770 LoadInst *Load = new LoadInst(
1771 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1772 Phi->setIncomingValue(i, Load);
1775 LoadInst *Load = new LoadInst(Alloca, "", I);
1776 I->replaceUsesOfWith(V, Load);
1780 // Store the original value and the replacement value into the alloca
1781 StoreInst *Store = new StoreInst(V, Alloca);
1782 if (auto I = dyn_cast<Instruction>(V))
1783 Store->insertAfter(I);
1785 Store->insertAfter(Alloca);
1787 // Normal return for invoke, or call return
1788 Instruction *Replacement = cast<Instruction>(Replacements[V].first);
1789 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1790 // Unwind return for invoke only
1791 Replacement = cast_or_null<Instruction>(Replacements[V].second);
1793 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement);
1796 // apply mem2reg to promote alloca to SSA
1797 SmallVector<AllocaInst *, 16> Allocas;
1798 for (Value *V : ToSplit)
1799 Allocas.push_back(AllocaMap[V]);
1800 PromoteMemToReg(Allocas, DT);
1803 // Helper function for the "rematerializeLiveValues". It walks use chain
1804 // starting from the "CurrentValue" until it meets "BaseValue". Only "simple"
1805 // values are visited (currently it is GEP's and casts). Returns true if it
1806 // sucessfully reached "BaseValue" and false otherwise.
1807 // Fills "ChainToBase" array with all visited values. "BaseValue" is not
1809 static bool findRematerializableChainToBasePointer(
1810 SmallVectorImpl<Instruction*> &ChainToBase,
1811 Value *CurrentValue, Value *BaseValue) {
1813 // We have found a base value
1814 if (CurrentValue == BaseValue) {
1818 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
1819 ChainToBase.push_back(GEP);
1820 return findRematerializableChainToBasePointer(ChainToBase,
1821 GEP->getPointerOperand(),
1825 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
1826 Value *Def = CI->stripPointerCasts();
1828 // This two checks are basically similar. First one is here for the
1829 // consistency with findBasePointers logic.
1830 assert(!isa<CastInst>(Def) && "not a pointer cast found");
1831 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
1834 ChainToBase.push_back(CI);
1835 return findRematerializableChainToBasePointer(ChainToBase, Def, BaseValue);
1838 // Not supported instruction in the chain
1842 // Helper function for the "rematerializeLiveValues". Compute cost of the use
1843 // chain we are going to rematerialize.
1845 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
1846 TargetTransformInfo &TTI) {
1849 for (Instruction *Instr : Chain) {
1850 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
1851 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
1852 "non noop cast is found during rematerialization");
1854 Type *SrcTy = CI->getOperand(0)->getType();
1855 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy);
1857 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
1858 // Cost of the address calculation
1859 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType();
1860 Cost += TTI.getAddressComputationCost(ValTy);
1862 // And cost of the GEP itself
1863 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
1864 // allowed for the external usage)
1865 if (!GEP->hasAllConstantIndices())
1869 llvm_unreachable("unsupported instruciton type during rematerialization");
1876 // From the statepoint liveset pick values that are cheaper to recompute then to
1877 // relocate. Remove this values from the liveset, rematerialize them after
1878 // statepoint and record them in "Info" structure. Note that similar to
1879 // relocated values we don't do any user adjustments here.
1880 static void rematerializeLiveValues(CallSite CS,
1881 PartiallyConstructedSafepointRecord &Info,
1882 TargetTransformInfo &TTI) {
1883 const unsigned int ChainLengthThreshold = 10;
1885 // Record values we are going to delete from this statepoint live set.
1886 // We can not di this in following loop due to iterator invalidation.
1887 SmallVector<Value *, 32> LiveValuesToBeDeleted;
1889 for (Value *LiveValue: Info.liveset) {
1890 // For each live pointer find it's defining chain
1891 SmallVector<Instruction *, 3> ChainToBase;
1892 assert(Info.PointerToBase.find(LiveValue) != Info.PointerToBase.end());
1894 findRematerializableChainToBasePointer(ChainToBase,
1896 Info.PointerToBase[LiveValue]);
1897 // Nothing to do, or chain is too long
1899 ChainToBase.size() == 0 ||
1900 ChainToBase.size() > ChainLengthThreshold)
1903 // Compute cost of this chain
1904 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
1905 // TODO: We can also account for cases when we will be able to remove some
1906 // of the rematerialized values by later optimization passes. I.e if
1907 // we rematerialized several intersecting chains. Or if original values
1908 // don't have any uses besides this statepoint.
1910 // For invokes we need to rematerialize each chain twice - for normal and
1911 // for unwind basic blocks. Model this by multiplying cost by two.
1912 if (CS.isInvoke()) {
1915 // If it's too expensive - skip it
1916 if (Cost >= RematerializationThreshold)
1919 // Remove value from the live set
1920 LiveValuesToBeDeleted.push_back(LiveValue);
1922 // Clone instructions and record them inside "Info" structure
1924 // Walk backwards to visit top-most instructions first
1925 std::reverse(ChainToBase.begin(), ChainToBase.end());
1927 // Utility function which clones all instructions from "ChainToBase"
1928 // and inserts them before "InsertBefore". Returns rematerialized value
1929 // which should be used after statepoint.
1930 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) {
1931 Instruction *LastClonedValue = nullptr;
1932 Instruction *LastValue = nullptr;
1933 for (Instruction *Instr: ChainToBase) {
1934 // Only GEP's and casts are suported as we need to be careful to not
1935 // introduce any new uses of pointers not in the liveset.
1936 // Note that it's fine to introduce new uses of pointers which were
1937 // otherwise not used after this statepoint.
1938 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
1940 Instruction *ClonedValue = Instr->clone();
1941 ClonedValue->insertBefore(InsertBefore);
1942 ClonedValue->setName(Instr->getName() + ".remat");
1944 // If it is not first instruction in the chain then it uses previously
1945 // cloned value. We should update it to use cloned value.
1946 if (LastClonedValue) {
1948 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
1950 // Assert that cloned instruction does not use any instructions from
1951 // this chain other than LastClonedValue
1952 for (auto OpValue : ClonedValue->operand_values()) {
1953 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) ==
1954 ChainToBase.end() &&
1955 "incorrect use in rematerialization chain");
1960 LastClonedValue = ClonedValue;
1963 assert(LastClonedValue);
1964 return LastClonedValue;
1967 // Different cases for calls and invokes. For invokes we need to clone
1968 // instructions both on normal and unwind path.
1970 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
1971 assert(InsertBefore);
1972 Instruction *RematerializedValue = rematerializeChain(InsertBefore);
1973 Info.RematerializedValues[RematerializedValue] = LiveValue;
1975 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
1977 Instruction *NormalInsertBefore =
1978 Invoke->getNormalDest()->getFirstInsertionPt();
1979 Instruction *UnwindInsertBefore =
1980 Invoke->getUnwindDest()->getFirstInsertionPt();
1982 Instruction *NormalRematerializedValue =
1983 rematerializeChain(NormalInsertBefore);
1984 Instruction *UnwindRematerializedValue =
1985 rematerializeChain(UnwindInsertBefore);
1987 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
1988 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
1992 // Remove rematerializaed values from the live set
1993 for (auto LiveValue: LiveValuesToBeDeleted) {
1994 Info.liveset.erase(LiveValue);
1998 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1999 SmallVectorImpl<CallSite> &toUpdate) {
2001 // sanity check the input
2002 std::set<CallSite> uniqued;
2003 uniqued.insert(toUpdate.begin(), toUpdate.end());
2004 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
2006 for (size_t i = 0; i < toUpdate.size(); i++) {
2007 CallSite &CS = toUpdate[i];
2008 assert(CS.getInstruction()->getParent()->getParent() == &F);
2009 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
2013 // When inserting gc.relocates for invokes, we need to be able to insert at
2014 // the top of the successor blocks. See the comment on
2015 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2016 // may restructure the CFG.
2017 for (CallSite CS : toUpdate) {
2020 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
2021 normalizeForInvokeSafepoint(invoke->getNormalDest(), invoke->getParent(),
2023 normalizeForInvokeSafepoint(invoke->getUnwindDest(), invoke->getParent(),
2027 // A list of dummy calls added to the IR to keep various values obviously
2028 // live in the IR. We'll remove all of these when done.
2029 SmallVector<CallInst *, 64> holders;
2031 // Insert a dummy call with all of the arguments to the vm_state we'll need
2032 // for the actual safepoint insertion. This ensures reference arguments in
2033 // the deopt argument list are considered live through the safepoint (and
2034 // thus makes sure they get relocated.)
2035 for (size_t i = 0; i < toUpdate.size(); i++) {
2036 CallSite &CS = toUpdate[i];
2037 Statepoint StatepointCS(CS);
2039 SmallVector<Value *, 64> DeoptValues;
2040 for (Use &U : StatepointCS.vm_state_args()) {
2041 Value *Arg = cast<Value>(&U);
2042 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2043 "support for FCA unimplemented");
2044 if (isHandledGCPointerType(Arg->getType()))
2045 DeoptValues.push_back(Arg);
2047 insertUseHolderAfter(CS, DeoptValues, holders);
2050 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
2051 records.reserve(toUpdate.size());
2052 for (size_t i = 0; i < toUpdate.size(); i++) {
2053 struct PartiallyConstructedSafepointRecord info;
2054 records.push_back(info);
2056 assert(records.size() == toUpdate.size());
2058 // A) Identify all gc pointers which are staticly live at the given call
2060 findLiveReferences(F, DT, P, toUpdate, records);
2062 // Do a limited scalarization of any live at safepoint vector values which
2063 // contain pointers. This enables this pass to run after vectorization at
2064 // the cost of some possible performance loss. TODO: it would be nice to
2065 // natively support vectors all the way through the backend so we don't need
2066 // to scalarize here.
2067 for (size_t i = 0; i < records.size(); i++) {
2068 struct PartiallyConstructedSafepointRecord &info = records[i];
2069 Instruction *statepoint = toUpdate[i].getInstruction();
2070 splitVectorValues(cast<Instruction>(statepoint), info.liveset, DT);
2073 // B) Find the base pointers for each live pointer
2074 /* scope for caching */ {
2075 // Cache the 'defining value' relation used in the computation and
2076 // insertion of base phis and selects. This ensures that we don't insert
2077 // large numbers of duplicate base_phis.
2078 DefiningValueMapTy DVCache;
2080 for (size_t i = 0; i < records.size(); i++) {
2081 struct PartiallyConstructedSafepointRecord &info = records[i];
2082 CallSite &CS = toUpdate[i];
2083 findBasePointers(DT, DVCache, CS, info);
2085 } // end of cache scope
2087 // The base phi insertion logic (for any safepoint) may have inserted new
2088 // instructions which are now live at some safepoint. The simplest such
2091 // phi a <-- will be a new base_phi here
2092 // safepoint 1 <-- that needs to be live here
2096 // We insert some dummy calls after each safepoint to definitely hold live
2097 // the base pointers which were identified for that safepoint. We'll then
2098 // ask liveness for _every_ base inserted to see what is now live. Then we
2099 // remove the dummy calls.
2100 holders.reserve(holders.size() + records.size());
2101 for (size_t i = 0; i < records.size(); i++) {
2102 struct PartiallyConstructedSafepointRecord &info = records[i];
2103 CallSite &CS = toUpdate[i];
2105 SmallVector<Value *, 128> Bases;
2106 for (auto Pair : info.PointerToBase) {
2107 Bases.push_back(Pair.second);
2109 insertUseHolderAfter(CS, Bases, holders);
2112 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2113 // need to rerun liveness. We may *also* have inserted new defs, but that's
2114 // not the key issue.
2115 recomputeLiveInValues(F, DT, P, toUpdate, records);
2117 if (PrintBasePointers) {
2118 for (size_t i = 0; i < records.size(); i++) {
2119 struct PartiallyConstructedSafepointRecord &info = records[i];
2120 errs() << "Base Pairs: (w/Relocation)\n";
2121 for (auto Pair : info.PointerToBase) {
2122 errs() << " derived %" << Pair.first->getName() << " base %"
2123 << Pair.second->getName() << "\n";
2127 for (size_t i = 0; i < holders.size(); i++) {
2128 holders[i]->eraseFromParent();
2129 holders[i] = nullptr;
2133 // In order to reduce live set of statepoint we might choose to rematerialize
2134 // some values instead of relocating them. This is purelly an optimization and
2135 // does not influence correctness.
2136 TargetTransformInfo &TTI =
2137 P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2139 for (size_t i = 0; i < records.size(); i++) {
2140 struct PartiallyConstructedSafepointRecord &info = records[i];
2141 CallSite &CS = toUpdate[i];
2143 rematerializeLiveValues(CS, info, TTI);
2146 // Now run through and replace the existing statepoints with new ones with
2147 // the live variables listed. We do not yet update uses of the values being
2148 // relocated. We have references to live variables that need to
2149 // survive to the last iteration of this loop. (By construction, the
2150 // previous statepoint can not be a live variable, thus we can and remove
2151 // the old statepoint calls as we go.)
2152 for (size_t i = 0; i < records.size(); i++) {
2153 struct PartiallyConstructedSafepointRecord &info = records[i];
2154 CallSite &CS = toUpdate[i];
2155 makeStatepointExplicit(DT, CS, P, info);
2157 toUpdate.clear(); // prevent accident use of invalid CallSites
2159 // Do all the fixups of the original live variables to their relocated selves
2160 SmallVector<Value *, 128> live;
2161 for (size_t i = 0; i < records.size(); i++) {
2162 struct PartiallyConstructedSafepointRecord &info = records[i];
2163 // We can't simply save the live set from the original insertion. One of
2164 // the live values might be the result of a call which needs a safepoint.
2165 // That Value* no longer exists and we need to use the new gc_result.
2166 // Thankfully, the liveset is embedded in the statepoint (and updated), so
2167 // we just grab that.
2168 Statepoint statepoint(info.StatepointToken);
2169 live.insert(live.end(), statepoint.gc_args_begin(),
2170 statepoint.gc_args_end());
2172 // Do some basic sanity checks on our liveness results before performing
2173 // relocation. Relocation can and will turn mistakes in liveness results
2174 // into non-sensical code which is must harder to debug.
2175 // TODO: It would be nice to test consistency as well
2176 assert(DT.isReachableFromEntry(info.StatepointToken->getParent()) &&
2177 "statepoint must be reachable or liveness is meaningless");
2178 for (Value *V : statepoint.gc_args()) {
2179 if (!isa<Instruction>(V))
2180 // Non-instruction values trivial dominate all possible uses
2182 auto LiveInst = cast<Instruction>(V);
2183 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2184 "unreachable values should never be live");
2185 assert(DT.dominates(LiveInst, info.StatepointToken) &&
2186 "basic SSA liveness expectation violated by liveness analysis");
2190 unique_unsorted(live);
2194 for (auto ptr : live) {
2195 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
2199 relocationViaAlloca(F, DT, live, records);
2200 return !records.empty();
2203 /// Returns true if this function should be rewritten by this pass. The main
2204 /// point of this function is as an extension point for custom logic.
2205 static bool shouldRewriteStatepointsIn(Function &F) {
2206 // TODO: This should check the GCStrategy
2208 const char *FunctionGCName = F.getGC();
2209 const StringRef StatepointExampleName("statepoint-example");
2210 const StringRef CoreCLRName("coreclr");
2211 return (StatepointExampleName == FunctionGCName) ||
2212 (CoreCLRName == FunctionGCName);
2217 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
2218 // Nothing to do for declarations.
2219 if (F.isDeclaration() || F.empty())
2222 // Policy choice says not to rewrite - the most common reason is that we're
2223 // compiling code without a GCStrategy.
2224 if (!shouldRewriteStatepointsIn(F))
2227 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2229 // Gather all the statepoints which need rewritten. Be careful to only
2230 // consider those in reachable code since we need to ask dominance queries
2231 // when rewriting. We'll delete the unreachable ones in a moment.
2232 SmallVector<CallSite, 64> ParsePointNeeded;
2233 bool HasUnreachableStatepoint = false;
2234 for (Instruction &I : inst_range(F)) {
2235 // TODO: only the ones with the flag set!
2236 if (isStatepoint(I)) {
2237 if (DT.isReachableFromEntry(I.getParent()))
2238 ParsePointNeeded.push_back(CallSite(&I));
2240 HasUnreachableStatepoint = true;
2244 bool MadeChange = false;
2246 // Delete any unreachable statepoints so that we don't have unrewritten
2247 // statepoints surviving this pass. This makes testing easier and the
2248 // resulting IR less confusing to human readers. Rather than be fancy, we
2249 // just reuse a utility function which removes the unreachable blocks.
2250 if (HasUnreachableStatepoint)
2251 MadeChange |= removeUnreachableBlocks(F);
2253 // Return early if no work to do.
2254 if (ParsePointNeeded.empty())
2257 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2258 // These are created by LCSSA. They have the effect of increasing the size
2259 // of liveness sets for no good reason. It may be harder to do this post
2260 // insertion since relocations and base phis can confuse things.
2261 for (BasicBlock &BB : F)
2262 if (BB.getUniquePredecessor()) {
2264 FoldSingleEntryPHINodes(&BB);
2267 MadeChange |= insertParsePoints(F, DT, this, ParsePointNeeded);
2271 // liveness computation via standard dataflow
2272 // -------------------------------------------------------------------
2274 // TODO: Consider using bitvectors for liveness, the set of potentially
2275 // interesting values should be small and easy to pre-compute.
2277 /// Compute the live-in set for the location rbegin starting from
2278 /// the live-out set of the basic block
2279 static void computeLiveInValues(BasicBlock::reverse_iterator rbegin,
2280 BasicBlock::reverse_iterator rend,
2281 DenseSet<Value *> &LiveTmp) {
2283 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) {
2284 Instruction *I = &*ritr;
2286 // KILL/Def - Remove this definition from LiveIn
2289 // Don't consider *uses* in PHI nodes, we handle their contribution to
2290 // predecessor blocks when we seed the LiveOut sets
2291 if (isa<PHINode>(I))
2294 // USE - Add to the LiveIn set for this instruction
2295 for (Value *V : I->operands()) {
2296 assert(!isUnhandledGCPointerType(V->getType()) &&
2297 "support for FCA unimplemented");
2298 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2299 // The choice to exclude all things constant here is slightly subtle.
2300 // There are two idependent reasons:
2301 // - We assume that things which are constant (from LLVM's definition)
2302 // do not move at runtime. For example, the address of a global
2303 // variable is fixed, even though it's contents may not be.
2304 // - Second, we can't disallow arbitrary inttoptr constants even
2305 // if the language frontend does. Optimization passes are free to
2306 // locally exploit facts without respect to global reachability. This
2307 // can create sections of code which are dynamically unreachable and
2308 // contain just about anything. (see constants.ll in tests)
2315 static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) {
2317 for (BasicBlock *Succ : successors(BB)) {
2318 const BasicBlock::iterator E(Succ->getFirstNonPHI());
2319 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) {
2320 PHINode *Phi = cast<PHINode>(&*I);
2321 Value *V = Phi->getIncomingValueForBlock(BB);
2322 assert(!isUnhandledGCPointerType(V->getType()) &&
2323 "support for FCA unimplemented");
2324 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2331 static DenseSet<Value *> computeKillSet(BasicBlock *BB) {
2332 DenseSet<Value *> KillSet;
2333 for (Instruction &I : *BB)
2334 if (isHandledGCPointerType(I.getType()))
2340 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2341 /// sanity check for the liveness computation.
2342 static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live,
2343 TerminatorInst *TI, bool TermOkay = false) {
2344 for (Value *V : Live) {
2345 if (auto *I = dyn_cast<Instruction>(V)) {
2346 // The terminator can be a member of the LiveOut set. LLVM's definition
2347 // of instruction dominance states that V does not dominate itself. As
2348 // such, we need to special case this to allow it.
2349 if (TermOkay && TI == I)
2351 assert(DT.dominates(I, TI) &&
2352 "basic SSA liveness expectation violated by liveness analysis");
2357 /// Check that all the liveness sets used during the computation of liveness
2358 /// obey basic SSA properties. This is useful for finding cases where we miss
2360 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2362 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2363 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2364 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2368 static void computeLiveInValues(DominatorTree &DT, Function &F,
2369 GCPtrLivenessData &Data) {
2371 SmallSetVector<BasicBlock *, 200> Worklist;
2372 auto AddPredsToWorklist = [&](BasicBlock *BB) {
2373 // We use a SetVector so that we don't have duplicates in the worklist.
2374 Worklist.insert(pred_begin(BB), pred_end(BB));
2376 auto NextItem = [&]() {
2377 BasicBlock *BB = Worklist.back();
2378 Worklist.pop_back();
2382 // Seed the liveness for each individual block
2383 for (BasicBlock &BB : F) {
2384 Data.KillSet[&BB] = computeKillSet(&BB);
2385 Data.LiveSet[&BB].clear();
2386 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2389 for (Value *Kill : Data.KillSet[&BB])
2390 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2393 Data.LiveOut[&BB] = DenseSet<Value *>();
2394 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2395 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2396 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]);
2397 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]);
2398 if (!Data.LiveIn[&BB].empty())
2399 AddPredsToWorklist(&BB);
2402 // Propagate that liveness until stable
2403 while (!Worklist.empty()) {
2404 BasicBlock *BB = NextItem();
2406 // Compute our new liveout set, then exit early if it hasn't changed
2407 // despite the contribution of our successor.
2408 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2409 const auto OldLiveOutSize = LiveOut.size();
2410 for (BasicBlock *Succ : successors(BB)) {
2411 assert(Data.LiveIn.count(Succ));
2412 set_union(LiveOut, Data.LiveIn[Succ]);
2414 // assert OutLiveOut is a subset of LiveOut
2415 if (OldLiveOutSize == LiveOut.size()) {
2416 // If the sets are the same size, then we didn't actually add anything
2417 // when unioning our successors LiveIn Thus, the LiveIn of this block
2421 Data.LiveOut[BB] = LiveOut;
2423 // Apply the effects of this basic block
2424 DenseSet<Value *> LiveTmp = LiveOut;
2425 set_union(LiveTmp, Data.LiveSet[BB]);
2426 set_subtract(LiveTmp, Data.KillSet[BB]);
2428 assert(Data.LiveIn.count(BB));
2429 const DenseSet<Value *> &OldLiveIn = Data.LiveIn[BB];
2430 // assert: OldLiveIn is a subset of LiveTmp
2431 if (OldLiveIn.size() != LiveTmp.size()) {
2432 Data.LiveIn[BB] = LiveTmp;
2433 AddPredsToWorklist(BB);
2435 } // while( !worklist.empty() )
2438 // Sanity check our ouput against SSA properties. This helps catch any
2439 // missing kills during the above iteration.
2440 for (BasicBlock &BB : F) {
2441 checkBasicSSA(DT, Data, BB);
2446 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2447 StatepointLiveSetTy &Out) {
2449 BasicBlock *BB = Inst->getParent();
2451 // Note: The copy is intentional and required
2452 assert(Data.LiveOut.count(BB));
2453 DenseSet<Value *> LiveOut = Data.LiveOut[BB];
2455 // We want to handle the statepoint itself oddly. It's
2456 // call result is not live (normal), nor are it's arguments
2457 // (unless they're used again later). This adjustment is
2458 // specifically what we need to relocate
2459 BasicBlock::reverse_iterator rend(Inst);
2460 computeLiveInValues(BB->rbegin(), rend, LiveOut);
2461 LiveOut.erase(Inst);
2462 Out.insert(LiveOut.begin(), LiveOut.end());
2465 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2467 PartiallyConstructedSafepointRecord &Info) {
2468 Instruction *Inst = CS.getInstruction();
2469 StatepointLiveSetTy Updated;
2470 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2473 DenseSet<Value *> Bases;
2474 for (auto KVPair : Info.PointerToBase) {
2475 Bases.insert(KVPair.second);
2478 // We may have base pointers which are now live that weren't before. We need
2479 // to update the PointerToBase structure to reflect this.
2480 for (auto V : Updated)
2481 if (!Info.PointerToBase.count(V)) {
2482 assert(Bases.count(V) && "can't find base for unexpected live value");
2483 Info.PointerToBase[V] = V;
2488 for (auto V : Updated) {
2489 assert(Info.PointerToBase.count(V) &&
2490 "must be able to find base for live value");
2494 // Remove any stale base mappings - this can happen since our liveness is
2495 // more precise then the one inherent in the base pointer analysis
2496 DenseSet<Value *> ToErase;
2497 for (auto KVPair : Info.PointerToBase)
2498 if (!Updated.count(KVPair.first))
2499 ToErase.insert(KVPair.first);
2500 for (auto V : ToErase)
2501 Info.PointerToBase.erase(V);
2504 for (auto KVPair : Info.PointerToBase)
2505 assert(Updated.count(KVPair.first) && "record for non-live value");
2508 Info.liveset = Updated;