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/ADT/SetOperations.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/DenseSet.h"
20 #include "llvm/IR/BasicBlock.h"
21 #include "llvm/IR/CallSite.h"
22 #include "llvm/IR/Dominators.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/IRBuilder.h"
25 #include "llvm/IR/InstIterator.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/Intrinsics.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/IR/Statepoint.h"
31 #include "llvm/IR/Value.h"
32 #include "llvm/IR/Verifier.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Transforms/Scalar.h"
36 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
37 #include "llvm/Transforms/Utils/Cloning.h"
38 #include "llvm/Transforms/Utils/Local.h"
39 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
41 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
45 // Print tracing output
46 static cl::opt<bool> TraceLSP("trace-rewrite-statepoints", cl::Hidden,
49 // Print the liveset found at the insert location
50 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
52 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size",
53 cl::Hidden, cl::init(false));
54 // Print out the base pointers for debugging
55 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers",
56 cl::Hidden, cl::init(false));
59 struct RewriteStatepointsForGC : public FunctionPass {
60 static char ID; // Pass identification, replacement for typeid
62 RewriteStatepointsForGC() : FunctionPass(ID) {
63 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
65 bool runOnFunction(Function &F) override;
67 void getAnalysisUsage(AnalysisUsage &AU) const override {
68 // We add and rewrite a bunch of instructions, but don't really do much
69 // else. We could in theory preserve a lot more analyses here.
70 AU.addRequired<DominatorTreeWrapperPass>();
75 char RewriteStatepointsForGC::ID = 0;
77 FunctionPass *llvm::createRewriteStatepointsForGCPass() {
78 return new RewriteStatepointsForGC();
81 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
82 "Make relocations explicit at statepoints", false, false)
83 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
84 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
85 "Make relocations explicit at statepoints", false, false)
88 // The type of the internal cache used inside the findBasePointers family
89 // of functions. From the callers perspective, this is an opaque type and
90 // should not be inspected.
92 // In the actual implementation this caches two relations:
93 // - The base relation itself (i.e. this pointer is based on that one)
94 // - The base defining value relation (i.e. before base_phi insertion)
95 // Generally, after the execution of a full findBasePointer call, only the
96 // base relation will remain. Internally, we add a mixture of the two
97 // types, then update all the second type to the first type
98 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
99 typedef DenseSet<llvm::Value *> StatepointLiveSetTy;
101 struct PartiallyConstructedSafepointRecord {
102 /// The set of values known to be live accross this safepoint
103 StatepointLiveSetTy liveset;
105 /// Mapping from live pointers to a base-defining-value
106 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
108 /// Any new values which were added to the IR during base pointer analysis
109 /// for this safepoint
110 DenseSet<llvm::Value *> NewInsertedDefs;
112 /// The *new* gc.statepoint instruction itself. This produces the token
113 /// that normal path gc.relocates and the gc.result are tied to.
114 Instruction *StatepointToken;
116 /// Instruction to which exceptional gc relocates are attached
117 /// Makes it easier to iterate through them during relocationViaAlloca.
118 Instruction *UnwindToken;
122 // TODO: Once we can get to the GCStrategy, this becomes
123 // Optional<bool> isGCManagedPointer(const Value *V) const override {
125 static bool isGCPointerType(const Type *T) {
126 if (const PointerType *PT = dyn_cast<PointerType>(T))
127 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
128 // GC managed heap. We know that a pointer into this heap needs to be
129 // updated and that no other pointer does.
130 return (1 == PT->getAddressSpace());
134 /// Return true if the Value is a gc reference type which is potentially used
135 /// after the instruction 'loc'. This is only used with the edge reachability
136 /// liveness code. Note: It is assumed the V dominates loc.
137 static bool isLiveGCReferenceAt(Value &V, Instruction *loc, DominatorTree &DT,
139 if (!isGCPointerType(V.getType()))
145 // Given assumption that V dominates loc, this may be live
150 static bool isAggWhichContainsGCPtrType(Type *Ty) {
151 if (VectorType *VT = dyn_cast<VectorType>(Ty))
152 return isGCPointerType(VT->getScalarType());
153 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
154 return isGCPointerType(AT->getElementType()) ||
155 isAggWhichContainsGCPtrType(AT->getElementType());
156 if (StructType *ST = dyn_cast<StructType>(Ty))
157 return std::any_of(ST->subtypes().begin(), ST->subtypes().end(),
159 return isGCPointerType(SubType) ||
160 isAggWhichContainsGCPtrType(SubType);
166 // Conservatively identifies any definitions which might be live at the
167 // given instruction. The analysis is performed immediately before the
168 // given instruction. Values defined by that instruction are not considered
169 // live. Values used by that instruction are considered live.
171 // preconditions: valid IR graph, term is either a terminator instruction or
172 // a call instruction, pred is the basic block of term, DT, LI are valid
174 // side effects: none, does not mutate IR
176 // postconditions: populates liveValues as discussed above
177 static void findLiveGCValuesAtInst(Instruction *term, BasicBlock *pred,
178 DominatorTree &DT, LoopInfo *LI,
179 StatepointLiveSetTy &liveValues) {
182 assert(isa<CallInst>(term) || isa<InvokeInst>(term) || term->isTerminator());
184 Function *F = pred->getParent();
186 auto is_live_gc_reference =
187 [&](Value &V) { return isLiveGCReferenceAt(V, term, DT, LI); };
189 // Are there any gc pointer arguments live over this point? This needs to be
190 // special cased since arguments aren't defined in basic blocks.
191 for (Argument &arg : F->args()) {
192 assert(!isAggWhichContainsGCPtrType(arg.getType()) &&
193 "support for FCA unimplemented");
195 if (is_live_gc_reference(arg)) {
196 liveValues.insert(&arg);
200 // Walk through all dominating blocks - the ones which can contain
201 // definitions used in this block - and check to see if any of the values
202 // they define are used in locations potentially reachable from the
203 // interesting instruction.
204 BasicBlock *BBI = pred;
207 errs() << "[LSP] Looking at dominating block " << pred->getName() << "\n";
209 assert(DT.dominates(BBI, pred));
210 assert(isPotentiallyReachable(BBI, pred, &DT) &&
211 "dominated block must be reachable");
213 // Walk through the instructions in dominating blocks and keep any
214 // that have a use potentially reachable from the block we're
215 // considering putting the safepoint in
216 for (Instruction &inst : *BBI) {
218 errs() << "[LSP] Looking at instruction ";
222 if (pred == BBI && (&inst) == term) {
224 errs() << "[LSP] stopped because we encountered the safepoint "
228 // If we're in the block which defines the interesting instruction,
229 // we don't want to include any values as live which are defined
230 // _after_ the interesting line or as part of the line itself
231 // i.e. "term" is the call instruction for a call safepoint, the
232 // results of the call should not be considered live in that stackmap
236 assert(!isAggWhichContainsGCPtrType(inst.getType()) &&
237 "support for FCA unimplemented");
239 if (is_live_gc_reference(inst)) {
241 errs() << "[LSP] found live value for this safepoint ";
245 liveValues.insert(&inst);
248 if (!DT.getNode(BBI)->getIDom()) {
249 assert(BBI == &F->getEntryBlock() &&
250 "failed to find a dominator for something other than "
254 BBI = DT.getNode(BBI)->getIDom()->getBlock();
258 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
259 if (a->hasName() && b->hasName()) {
260 return -1 == a->getName().compare(b->getName());
261 } else if (a->hasName() && !b->hasName()) {
263 } else if (!a->hasName() && b->hasName()) {
266 // Better than nothing, but not stable
271 /// Find the initial live set. Note that due to base pointer
272 /// insertion, the live set may be incomplete.
274 analyzeParsePointLiveness(DominatorTree &DT, const CallSite &CS,
275 PartiallyConstructedSafepointRecord &result) {
276 Instruction *inst = CS.getInstruction();
278 BasicBlock *BB = inst->getParent();
279 StatepointLiveSetTy liveset;
280 findLiveGCValuesAtInst(inst, BB, DT, nullptr, liveset);
283 // Note: This output is used by several of the test cases
284 // The order of elemtns in a set is not stable, put them in a vec and sort
286 SmallVector<Value *, 64> temp;
287 temp.insert(temp.end(), liveset.begin(), liveset.end());
288 std::sort(temp.begin(), temp.end(), order_by_name);
289 errs() << "Live Variables:\n";
290 for (Value *V : temp) {
291 errs() << " " << V->getName(); // no newline
295 if (PrintLiveSetSize) {
296 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
297 errs() << "Number live values: " << liveset.size() << "\n";
299 result.liveset = liveset;
302 /// True iff this value is the null pointer constant (of any pointer type)
303 static bool LLVM_ATTRIBUTE_UNUSED isNullConstant(Value *V) {
304 return isa<Constant>(V) && isa<PointerType>(V->getType()) &&
305 cast<Constant>(V)->isNullValue();
308 /// Helper function for findBasePointer - Will return a value which either a)
309 /// defines the base pointer for the input or b) blocks the simple search
310 /// (i.e. a PHI or Select of two derived pointers)
311 static Value *findBaseDefiningValue(Value *I) {
312 assert(I->getType()->isPointerTy() &&
313 "Illegal to ask for the base pointer of a non-pointer type");
315 // There are instructions which can never return gc pointer values. Sanity
317 // that this is actually true.
318 assert(!isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
319 !isa<ShuffleVectorInst>(I) && "Vector types are not gc pointers");
320 assert((!isa<Instruction>(I) || isa<InvokeInst>(I) ||
321 !cast<Instruction>(I)->isTerminator()) &&
322 "With the exception of invoke terminators don't define values");
323 assert(!isa<StoreInst>(I) && !isa<FenceInst>(I) &&
324 "Can't be definitions to start with");
325 assert(!isa<ICmpInst>(I) && !isa<FCmpInst>(I) &&
326 "Comparisons don't give ops");
327 // There's a bunch of instructions which just don't make sense to apply to
328 // a pointer. The only valid reason for this would be pointer bit
329 // twiddling which we're just not going to support.
330 assert((!isa<Instruction>(I) || !cast<Instruction>(I)->isBinaryOp()) &&
331 "Binary ops on pointer values are meaningless. Unless your "
332 "bit-twiddling which we don't support");
334 if (Argument *Arg = dyn_cast<Argument>(I)) {
335 // An incoming argument to the function is a base pointer
336 // We should have never reached here if this argument isn't an gc value
337 assert(Arg->getType()->isPointerTy() &&
338 "Base for pointer must be another pointer");
342 if (GlobalVariable *global = dyn_cast<GlobalVariable>(I)) {
344 assert(global->getType()->isPointerTy() &&
345 "Base for pointer must be another pointer");
349 // inlining could possibly introduce phi node that contains
350 // undef if callee has multiple returns
351 if (UndefValue *undef = dyn_cast<UndefValue>(I)) {
352 assert(undef->getType()->isPointerTy() &&
353 "Base for pointer must be another pointer");
354 return undef; // utterly meaningless, but useful for dealing with
355 // partially optimized code.
358 // Due to inheritance, this must be _after_ the global variable and undef
360 if (Constant *con = dyn_cast<Constant>(I)) {
361 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
362 "order of checks wrong!");
363 // Note: Finding a constant base for something marked for relocation
364 // doesn't really make sense. The most likely case is either a) some
365 // screwed up the address space usage or b) your validating against
366 // compiled C++ code w/o the proper separation. The only real exception
367 // is a null pointer. You could have generic code written to index of
368 // off a potentially null value and have proven it null. We also use
369 // null pointers in dead paths of relocation phis (which we might later
370 // want to find a base pointer for).
371 assert(con->getType()->isPointerTy() &&
372 "Base for pointer must be another pointer");
373 assert(con->isNullValue() && "null is the only case which makes sense");
377 if (CastInst *CI = dyn_cast<CastInst>(I)) {
378 Value *def = CI->stripPointerCasts();
379 assert(def->getType()->isPointerTy() &&
380 "Base for pointer must be another pointer");
381 // If we find a cast instruction here, it means we've found a cast which is
382 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
383 // handle int->ptr conversion.
384 assert(!isa<CastInst>(def) && "shouldn't find another cast here");
385 return findBaseDefiningValue(def);
388 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
389 if (LI->getType()->isPointerTy()) {
390 Value *Op = LI->getOperand(0);
392 // Has to be a pointer to an gc object, or possibly an array of such?
393 assert(Op->getType()->isPointerTy());
394 return LI; // The value loaded is an gc base itself
397 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
398 Value *Op = GEP->getOperand(0);
399 if (Op->getType()->isPointerTy()) {
400 return findBaseDefiningValue(Op); // The base of this GEP is the base
404 if (AllocaInst *alloc = dyn_cast<AllocaInst>(I)) {
405 // An alloca represents a conceptual stack slot. It's the slot itself
406 // that the GC needs to know about, not the value in the slot.
407 assert(alloc->getType()->isPointerTy() &&
408 "Base for pointer must be another pointer");
412 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
413 switch (II->getIntrinsicID()) {
415 // fall through to general call handling
417 case Intrinsic::experimental_gc_statepoint:
418 case Intrinsic::experimental_gc_result_float:
419 case Intrinsic::experimental_gc_result_int:
420 llvm_unreachable("these don't produce pointers");
421 case Intrinsic::experimental_gc_result_ptr:
422 // This is just a special case of the CallInst check below to handle a
423 // statepoint with deopt args which hasn't been rewritten for GC yet.
424 // TODO: Assert that the statepoint isn't rewritten yet.
426 case Intrinsic::experimental_gc_relocate: {
427 // Rerunning safepoint insertion after safepoints are already
428 // inserted is not supported. It could probably be made to work,
429 // but why are you doing this? There's no good reason.
430 llvm_unreachable("repeat safepoint insertion is not supported");
432 case Intrinsic::gcroot:
433 // Currently, this mechanism hasn't been extended to work with gcroot.
434 // There's no reason it couldn't be, but I haven't thought about the
435 // implications much.
437 "interaction with the gcroot mechanism is not supported");
440 // We assume that functions in the source language only return base
441 // pointers. This should probably be generalized via attributes to support
442 // both source language and internal functions.
443 if (CallInst *call = dyn_cast<CallInst>(I)) {
444 assert(call->getType()->isPointerTy() &&
445 "Base for pointer must be another pointer");
448 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
449 assert(invoke->getType()->isPointerTy() &&
450 "Base for pointer must be another pointer");
454 // I have absolutely no idea how to implement this part yet. It's not
455 // neccessarily hard, I just haven't really looked at it yet.
456 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
458 if (AtomicCmpXchgInst *cas = dyn_cast<AtomicCmpXchgInst>(I)) {
459 // A CAS is effectively a atomic store and load combined under a
460 // predicate. From the perspective of base pointers, we just treat it
461 // like a load. We loaded a pointer from a address in memory, that value
462 // had better be a valid base pointer.
463 return cas->getPointerOperand();
465 if (AtomicRMWInst *atomic = dyn_cast<AtomicRMWInst>(I)) {
466 assert(AtomicRMWInst::Xchg == atomic->getOperation() &&
467 "All others are binary ops which don't apply to base pointers");
468 // semantically, a load, store pair. Treat it the same as a standard load
469 return atomic->getPointerOperand();
472 // The aggregate ops. Aggregates can either be in the heap or on the
473 // stack, but in either case, this is simply a field load. As a result,
474 // this is a defining definition of the base just like a load is.
475 if (ExtractValueInst *ev = dyn_cast<ExtractValueInst>(I)) {
479 // We should never see an insert vector since that would require we be
480 // tracing back a struct value not a pointer value.
481 assert(!isa<InsertValueInst>(I) &&
482 "Base pointer for a struct is meaningless");
484 // The last two cases here don't return a base pointer. Instead, they
485 // return a value which dynamically selects from amoung several base
486 // derived pointers (each with it's own base potentially). It's the job of
487 // the caller to resolve these.
488 if (SelectInst *select = dyn_cast<SelectInst>(I)) {
492 return cast<PHINode>(I);
495 /// Returns the base defining value for this value.
496 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &cache) {
497 Value *&Cached = cache[I];
499 Cached = findBaseDefiningValue(I);
501 assert(cache[I] != nullptr);
504 errs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
510 /// Return a base pointer for this value if known. Otherwise, return it's
511 /// base defining value.
512 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &cache) {
513 Value *def = findBaseDefiningValueCached(I, cache);
514 auto Found = cache.find(def);
515 if (Found != cache.end()) {
516 // Either a base-of relation, or a self reference. Caller must check.
517 return Found->second;
519 // Only a BDV available
523 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
524 /// is it known to be a base pointer? Or do we need to continue searching.
525 static bool isKnownBaseResult(Value *v) {
526 if (!isa<PHINode>(v) && !isa<SelectInst>(v)) {
527 // no recursion possible
530 if (cast<Instruction>(v)->getMetadata("is_base_value")) {
531 // This is a previously inserted base phi or select. We know
532 // that this is a base value.
536 // We need to keep searching
540 // TODO: find a better name for this
544 enum Status { Unknown, Base, Conflict };
546 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
547 assert(status != Base || b);
549 PhiState(Value *b) : status(Base), base(b) {}
550 PhiState() : status(Unknown), base(nullptr) {}
551 PhiState(const PhiState &other) : status(other.status), base(other.base) {
552 assert(status != Base || base);
555 Status getStatus() const { return status; }
556 Value *getBase() const { return base; }
558 bool isBase() const { return getStatus() == Base; }
559 bool isUnknown() const { return getStatus() == Unknown; }
560 bool isConflict() const { return getStatus() == Conflict; }
562 bool operator==(const PhiState &other) const {
563 return base == other.base && status == other.status;
566 bool operator!=(const PhiState &other) const { return !(*this == other); }
569 errs() << status << " (" << base << " - "
570 << (base ? base->getName() : "nullptr") << "): ";
575 Value *base; // non null only if status == base
578 typedef DenseMap<Value *, PhiState> ConflictStateMapTy;
579 // Values of type PhiState form a lattice, and this is a helper
580 // class that implementes the meet operation. The meat of the meet
581 // operation is implemented in MeetPhiStates::pureMeet
582 class MeetPhiStates {
584 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
585 explicit MeetPhiStates(const ConflictStateMapTy &phiStates)
586 : phiStates(phiStates) {}
588 // Destructively meet the current result with the base V. V can
589 // either be a merge instruction (SelectInst / PHINode), in which
590 // case its status is looked up in the phiStates map; or a regular
591 // SSA value, in which case it is assumed to be a base.
592 void meetWith(Value *V) {
593 PhiState otherState = getStateForBDV(V);
594 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
595 MeetPhiStates::pureMeet(currentResult, otherState)) &&
596 "math is wrong: meet does not commute!");
597 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
600 PhiState getResult() const { return currentResult; }
603 const ConflictStateMapTy &phiStates;
604 PhiState currentResult;
606 /// Return a phi state for a base defining value. We'll generate a new
607 /// base state for known bases and expect to find a cached state otherwise
608 PhiState getStateForBDV(Value *baseValue) {
609 if (isKnownBaseResult(baseValue)) {
610 return PhiState(baseValue);
612 return lookupFromMap(baseValue);
616 PhiState lookupFromMap(Value *V) {
617 auto I = phiStates.find(V);
618 assert(I != phiStates.end() && "lookup failed!");
622 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
623 switch (stateA.getStatus()) {
624 case PhiState::Unknown:
628 assert(stateA.getBase() && "can't be null");
629 if (stateB.isUnknown())
632 if (stateB.isBase()) {
633 if (stateA.getBase() == stateB.getBase()) {
634 assert(stateA == stateB && "equality broken!");
637 return PhiState(PhiState::Conflict);
639 assert(stateB.isConflict() && "only three states!");
640 return PhiState(PhiState::Conflict);
642 case PhiState::Conflict:
645 llvm_unreachable("only three states!");
649 /// For a given value or instruction, figure out what base ptr it's derived
650 /// from. For gc objects, this is simply itself. On success, returns a value
651 /// which is the base pointer. (This is reliable and can be used for
652 /// relocation.) On failure, returns nullptr.
653 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache,
654 DenseSet<llvm::Value *> &NewInsertedDefs) {
655 Value *def = findBaseOrBDV(I, cache);
657 if (isKnownBaseResult(def)) {
661 // Here's the rough algorithm:
662 // - For every SSA value, construct a mapping to either an actual base
663 // pointer or a PHI which obscures the base pointer.
664 // - Construct a mapping from PHI to unknown TOP state. Use an
665 // optimistic algorithm to propagate base pointer information. Lattice
670 // When algorithm terminates, all PHIs will either have a single concrete
671 // base or be in a conflict state.
672 // - For every conflict, insert a dummy PHI node without arguments. Add
673 // these to the base[Instruction] = BasePtr mapping. For every
674 // non-conflict, add the actual base.
675 // - For every conflict, add arguments for the base[a] of each input
678 // Note: A simpler form of this would be to add the conflict form of all
679 // PHIs without running the optimistic algorithm. This would be
680 // analougous to pessimistic data flow and would likely lead to an
681 // overall worse solution.
683 ConflictStateMapTy states;
684 states[def] = PhiState();
685 // Recursively fill in all phis & selects reachable from the initial one
686 // for which we don't already know a definite base value for
687 // TODO: This should be rewritten with a worklist
691 // Since we're adding elements to 'states' as we run, we can't keep
692 // iterators into the set.
693 SmallVector<Value*, 16> Keys;
694 Keys.reserve(states.size());
695 for (auto Pair : states) {
696 Value *V = Pair.first;
699 for (Value *v : Keys) {
700 assert(!isKnownBaseResult(v) && "why did it get added?");
701 if (PHINode *phi = dyn_cast<PHINode>(v)) {
702 assert(phi->getNumIncomingValues() > 0 &&
703 "zero input phis are illegal");
704 for (Value *InVal : phi->incoming_values()) {
705 Value *local = findBaseOrBDV(InVal, cache);
706 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
707 states[local] = PhiState();
711 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
712 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
713 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
714 states[local] = PhiState();
717 local = findBaseOrBDV(sel->getFalseValue(), cache);
718 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
719 states[local] = PhiState();
727 errs() << "States after initialization:\n";
728 for (auto Pair : states) {
729 Instruction *v = cast<Instruction>(Pair.first);
730 PhiState state = Pair.second;
736 // TODO: come back and revisit the state transitions around inputs which
737 // have reached conflict state. The current version seems too conservative.
739 bool progress = true;
742 size_t oldSize = states.size();
745 // We're only changing keys in this loop, thus safe to keep iterators
746 for (auto Pair : states) {
747 MeetPhiStates calculateMeet(states);
748 Value *v = Pair.first;
749 assert(!isKnownBaseResult(v) && "why did it get added?");
750 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
751 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
752 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
754 for (Value *Val : cast<PHINode>(v)->incoming_values())
755 calculateMeet.meetWith(findBaseOrBDV(Val, cache));
757 PhiState oldState = states[v];
758 PhiState newState = calculateMeet.getResult();
759 if (oldState != newState) {
761 states[v] = newState;
765 assert(oldSize <= states.size());
766 assert(oldSize == states.size() || progress);
770 errs() << "States after meet iteration:\n";
771 for (auto Pair : states) {
772 Instruction *v = cast<Instruction>(Pair.first);
773 PhiState state = Pair.second;
779 // Insert Phis for all conflicts
780 // We want to keep naming deterministic in the loop that follows, so
781 // sort the keys before iteration. This is useful in allowing us to
782 // write stable tests. Note that there is no invalidation issue here.
783 SmallVector<Value*, 16> Keys;
784 Keys.reserve(states.size());
785 for (auto Pair : states) {
786 Value *V = Pair.first;
789 std::sort(Keys.begin(), Keys.end(), order_by_name);
790 // TODO: adjust naming patterns to avoid this order of iteration dependency
791 for (Value *V : Keys) {
792 Instruction *v = cast<Instruction>(V);
793 PhiState state = states[V];
794 assert(!isKnownBaseResult(v) && "why did it get added?");
795 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
796 if (!state.isConflict())
799 if (isa<PHINode>(v)) {
801 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
802 assert(num_preds > 0 && "how did we reach here");
803 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
804 NewInsertedDefs.insert(phi);
805 // Add metadata marking this as a base value
806 auto *const_1 = ConstantInt::get(
808 v->getParent()->getParent()->getParent()->getContext()),
810 auto MDConst = ConstantAsMetadata::get(const_1);
811 MDNode *md = MDNode::get(
812 v->getParent()->getParent()->getParent()->getContext(), MDConst);
813 phi->setMetadata("is_base_value", md);
814 states[v] = PhiState(PhiState::Conflict, phi);
816 SelectInst *sel = cast<SelectInst>(v);
817 // The undef will be replaced later
818 UndefValue *undef = UndefValue::get(sel->getType());
819 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
820 undef, "base_select", sel);
821 NewInsertedDefs.insert(basesel);
822 // Add metadata marking this as a base value
823 auto *const_1 = ConstantInt::get(
825 v->getParent()->getParent()->getParent()->getContext()),
827 auto MDConst = ConstantAsMetadata::get(const_1);
828 MDNode *md = MDNode::get(
829 v->getParent()->getParent()->getParent()->getContext(), MDConst);
830 basesel->setMetadata("is_base_value", md);
831 states[v] = PhiState(PhiState::Conflict, basesel);
835 // Fixup all the inputs of the new PHIs
836 for (auto Pair : states) {
837 Instruction *v = cast<Instruction>(Pair.first);
838 PhiState state = Pair.second;
840 assert(!isKnownBaseResult(v) && "why did it get added?");
841 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
842 if (!state.isConflict())
845 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
846 PHINode *phi = cast<PHINode>(v);
847 unsigned NumPHIValues = phi->getNumIncomingValues();
848 for (unsigned i = 0; i < NumPHIValues; i++) {
849 Value *InVal = phi->getIncomingValue(i);
850 BasicBlock *InBB = phi->getIncomingBlock(i);
852 // If we've already seen InBB, add the same incoming value
853 // we added for it earlier. The IR verifier requires phi
854 // nodes with multiple entries from the same basic block
855 // to have the same incoming value for each of those
856 // entries. If we don't do this check here and basephi
857 // has a different type than base, we'll end up adding two
858 // bitcasts (and hence two distinct values) as incoming
859 // values for the same basic block.
861 int blockIndex = basephi->getBasicBlockIndex(InBB);
862 if (blockIndex != -1) {
863 Value *oldBase = basephi->getIncomingValue(blockIndex);
864 basephi->addIncoming(oldBase, InBB);
866 Value *base = findBaseOrBDV(InVal, cache);
867 if (!isKnownBaseResult(base)) {
868 // Either conflict or base.
869 assert(states.count(base));
870 base = states[base].getBase();
871 assert(base != nullptr && "unknown PhiState!");
872 assert(NewInsertedDefs.count(base) &&
873 "should have already added this in a prev. iteration!");
876 // In essense this assert states: the only way two
877 // values incoming from the same basic block may be
878 // different is by being different bitcasts of the same
879 // value. A cleanup that remains TODO is changing
880 // findBaseOrBDV to return an llvm::Value of the correct
881 // type (and still remain pure). This will remove the
882 // need to add bitcasts.
883 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
884 "sanity -- findBaseOrBDV should be pure!");
889 // Find either the defining value for the PHI or the normal base for
891 Value *base = findBaseOrBDV(InVal, cache);
892 if (!isKnownBaseResult(base)) {
893 // Either conflict or base.
894 assert(states.count(base));
895 base = states[base].getBase();
896 assert(base != nullptr && "unknown PhiState!");
898 assert(base && "can't be null");
899 // Must use original input BB since base may not be Instruction
900 // The cast is needed since base traversal may strip away bitcasts
901 if (base->getType() != basephi->getType()) {
902 base = new BitCastInst(base, basephi->getType(), "cast",
903 InBB->getTerminator());
904 NewInsertedDefs.insert(base);
906 basephi->addIncoming(base, InBB);
908 assert(basephi->getNumIncomingValues() == NumPHIValues);
910 SelectInst *basesel = cast<SelectInst>(state.getBase());
911 SelectInst *sel = cast<SelectInst>(v);
912 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
913 // something more safe and less hacky.
914 for (int i = 1; i <= 2; i++) {
915 Value *InVal = sel->getOperand(i);
916 // Find either the defining value for the PHI or the normal base for
918 Value *base = findBaseOrBDV(InVal, cache);
919 if (!isKnownBaseResult(base)) {
920 // Either conflict or base.
921 assert(states.count(base));
922 base = states[base].getBase();
923 assert(base != nullptr && "unknown PhiState!");
925 assert(base && "can't be null");
926 // Must use original input BB since base may not be Instruction
927 // The cast is needed since base traversal may strip away bitcasts
928 if (base->getType() != basesel->getType()) {
929 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
930 NewInsertedDefs.insert(base);
932 basesel->setOperand(i, base);
937 // Cache all of our results so we can cheaply reuse them
938 // NOTE: This is actually two caches: one of the base defining value
939 // relation and one of the base pointer relation! FIXME
940 for (auto item : states) {
941 Value *v = item.first;
942 Value *base = item.second.getBase();
944 assert(!isKnownBaseResult(v) && "why did it get added?");
947 std::string fromstr =
948 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
950 errs() << "Updating base value cache"
951 << " for: " << (v->hasName() ? v->getName() : "")
952 << " from: " << fromstr
953 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
956 assert(isKnownBaseResult(base) &&
957 "must be something we 'know' is a base pointer");
958 if (cache.count(v)) {
959 // Once we transition from the BDV relation being store in the cache to
960 // the base relation being stored, it must be stable
961 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
962 "base relation should be stable");
966 assert(cache.find(def) != cache.end());
970 // For a set of live pointers (base and/or derived), identify the base
971 // pointer of the object which they are derived from. This routine will
972 // mutate the IR graph as needed to make the 'base' pointer live at the
973 // definition site of 'derived'. This ensures that any use of 'derived' can
974 // also use 'base'. This may involve the insertion of a number of
975 // additional PHI nodes.
977 // preconditions: live is a set of pointer type Values
979 // side effects: may insert PHI nodes into the existing CFG, will preserve
980 // CFG, will not remove or mutate any existing nodes
982 // post condition: PointerToBase contains one (derived, base) pair for every
983 // pointer in live. Note that derived can be equal to base if the original
984 // pointer was a base pointer.
985 static void findBasePointers(const StatepointLiveSetTy &live,
986 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
987 DominatorTree *DT, DefiningValueMapTy &DVCache,
988 DenseSet<llvm::Value *> &NewInsertedDefs) {
989 // For the naming of values inserted to be deterministic - which makes for
990 // much cleaner and more stable tests - we need to assign an order to the
991 // live values. DenseSets do not provide a deterministic order across runs.
992 SmallVector<Value*, 64> Temp;
993 Temp.insert(Temp.end(), live.begin(), live.end());
994 std::sort(Temp.begin(), Temp.end(), order_by_name);
995 for (Value *ptr : Temp) {
996 Value *base = findBasePointer(ptr, DVCache, NewInsertedDefs);
997 assert(base && "failed to find base pointer");
998 PointerToBase[ptr] = base;
999 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1000 DT->dominates(cast<Instruction>(base)->getParent(),
1001 cast<Instruction>(ptr)->getParent())) &&
1002 "The base we found better dominate the derived pointer");
1004 // If you see this trip and like to live really dangerously, the code should
1005 // be correct, just with idioms the verifier can't handle. You can try
1006 // disabling the verifier at your own substaintial risk.
1007 assert(!isNullConstant(base) && "the relocation code needs adjustment to "
1008 "handle the relocation of a null pointer "
1009 "constant without causing false positives "
1010 "in the safepoint ir verifier.");
1014 /// Find the required based pointers (and adjust the live set) for the given
1016 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1018 PartiallyConstructedSafepointRecord &result) {
1019 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
1020 DenseSet<llvm::Value *> NewInsertedDefs;
1021 findBasePointers(result.liveset, PointerToBase, &DT, DVCache, NewInsertedDefs);
1023 if (PrintBasePointers) {
1024 // Note: Need to print these in a stable order since this is checked in
1026 errs() << "Base Pairs (w/o Relocation):\n";
1027 SmallVector<Value*, 64> Temp;
1028 Temp.reserve(PointerToBase.size());
1029 for (auto Pair : PointerToBase) {
1030 Temp.push_back(Pair.first);
1032 std::sort(Temp.begin(), Temp.end(), order_by_name);
1033 for (Value *Ptr : Temp) {
1034 Value *Base = PointerToBase[Ptr];
1035 errs() << " derived %" << Ptr->getName() << " base %"
1036 << Base->getName() << "\n";
1040 result.PointerToBase = PointerToBase;
1041 result.NewInsertedDefs = NewInsertedDefs;
1044 /// Check for liveness of items in the insert defs and add them to the live
1045 /// and base pointer sets
1046 static void fixupLiveness(DominatorTree &DT, const CallSite &CS,
1047 const DenseSet<Value *> &allInsertedDefs,
1048 PartiallyConstructedSafepointRecord &result) {
1049 Instruction *inst = CS.getInstruction();
1051 auto liveset = result.liveset;
1052 auto PointerToBase = result.PointerToBase;
1054 auto is_live_gc_reference =
1055 [&](Value &V) { return isLiveGCReferenceAt(V, inst, DT, nullptr); };
1057 // For each new definition, check to see if a) the definition dominates the
1058 // instruction we're interested in, and b) one of the uses of that definition
1059 // is edge-reachable from the instruction we're interested in. This is the
1060 // same definition of liveness we used in the intial liveness analysis
1061 for (Value *newDef : allInsertedDefs) {
1062 if (liveset.count(newDef)) {
1063 // already live, no action needed
1067 // PERF: Use DT to check instruction domination might not be good for
1068 // compilation time, and we could change to optimal solution if this
1069 // turn to be a issue
1070 if (!DT.dominates(cast<Instruction>(newDef), inst)) {
1071 // can't possibly be live at inst
1075 if (is_live_gc_reference(*newDef)) {
1076 // Add the live new defs into liveset and PointerToBase
1077 liveset.insert(newDef);
1078 PointerToBase[newDef] = newDef;
1082 result.liveset = liveset;
1083 result.PointerToBase = PointerToBase;
1086 static void fixupLiveReferences(
1087 Function &F, DominatorTree &DT, Pass *P,
1088 const DenseSet<llvm::Value *> &allInsertedDefs,
1089 ArrayRef<CallSite> toUpdate,
1090 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1091 for (size_t i = 0; i < records.size(); i++) {
1092 struct PartiallyConstructedSafepointRecord &info = records[i];
1093 const CallSite &CS = toUpdate[i];
1094 fixupLiveness(DT, CS, allInsertedDefs, info);
1098 // Normalize basic block to make it ready to be target of invoke statepoint.
1099 // It means spliting it to have single predecessor. Return newly created BB
1100 // ready to be successor of invoke statepoint.
1101 static BasicBlock *normalizeBBForInvokeSafepoint(BasicBlock *BB,
1102 BasicBlock *InvokeParent,
1104 BasicBlock *ret = BB;
1106 if (!BB->getUniquePredecessor()) {
1107 ret = SplitBlockPredecessors(BB, InvokeParent, "");
1110 // Another requirement for such basic blocks is to not have any phi nodes.
1111 // Since we just ensured that new BB will have single predecessor,
1112 // all phi nodes in it will have one value. Here it would be naturall place
1114 // remove them all. But we can not do this because we are risking to remove
1115 // one of the values stored in liveset of another statepoint. We will do it
1116 // later after placing all safepoints.
1121 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1122 auto itr = std::find(livevec.begin(), livevec.end(), val);
1123 assert(livevec.end() != itr);
1124 size_t index = std::distance(livevec.begin(), itr);
1125 assert(index < livevec.size());
1129 // Create new attribute set containing only attributes which can be transfered
1130 // from original call to the safepoint.
1131 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1134 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1135 unsigned index = AS.getSlotIndex(Slot);
1137 if (index == AttributeSet::ReturnIndex ||
1138 index == AttributeSet::FunctionIndex) {
1140 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1142 Attribute attr = *it;
1144 // Do not allow certain attributes - just skip them
1145 // Safepoint can not be read only or read none.
1146 if (attr.hasAttribute(Attribute::ReadNone) ||
1147 attr.hasAttribute(Attribute::ReadOnly))
1150 ret = ret.addAttributes(
1151 AS.getContext(), index,
1152 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1156 // Just skip parameter attributes for now
1162 /// Helper function to place all gc relocates necessary for the given
1165 /// liveVariables - list of variables to be relocated.
1166 /// liveStart - index of the first live variable.
1167 /// basePtrs - base pointers.
1168 /// statepointToken - statepoint instruction to which relocates should be
1170 /// Builder - Llvm IR builder to be used to construct new calls.
1171 void CreateGCRelocates(ArrayRef<llvm::Value *> liveVariables,
1172 const int liveStart,
1173 ArrayRef<llvm::Value *> basePtrs,
1174 Instruction *statepointToken, IRBuilder<> Builder) {
1176 SmallVector<Instruction *, 64> NewDefs;
1177 NewDefs.reserve(liveVariables.size());
1179 Module *M = statepointToken->getParent()->getParent()->getParent();
1181 for (unsigned i = 0; i < liveVariables.size(); i++) {
1182 // We generate a (potentially) unique declaration for every pointer type
1183 // combination. This results is some blow up the function declarations in
1184 // the IR, but removes the need for argument bitcasts which shrinks the IR
1185 // greatly and makes it much more readable.
1186 SmallVector<Type *, 1> types; // one per 'any' type
1187 types.push_back(liveVariables[i]->getType()); // result type
1188 Value *gc_relocate_decl = Intrinsic::getDeclaration(
1189 M, Intrinsic::experimental_gc_relocate, types);
1191 // Generate the gc.relocate call and save the result
1193 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1194 liveStart + find_index(liveVariables, basePtrs[i]));
1195 Value *liveIdx = ConstantInt::get(
1196 Type::getInt32Ty(M->getContext()),
1197 liveStart + find_index(liveVariables, liveVariables[i]));
1199 // only specify a debug name if we can give a useful one
1200 Value *reloc = Builder.CreateCall3(
1201 gc_relocate_decl, statepointToken, baseIdx, liveIdx,
1202 liveVariables[i]->hasName() ? liveVariables[i]->getName() + ".relocated"
1204 // Trick CodeGen into thinking there are lots of free registers at this
1206 cast<CallInst>(reloc)->setCallingConv(CallingConv::Cold);
1208 NewDefs.push_back(cast<Instruction>(reloc));
1210 assert(NewDefs.size() == liveVariables.size() &&
1211 "missing or extra redefinition at safepoint");
1215 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1216 const SmallVectorImpl<llvm::Value *> &basePtrs,
1217 const SmallVectorImpl<llvm::Value *> &liveVariables,
1219 PartiallyConstructedSafepointRecord &result) {
1220 assert(basePtrs.size() == liveVariables.size());
1221 assert(isStatepoint(CS) &&
1222 "This method expects to be rewriting a statepoint");
1224 BasicBlock *BB = CS.getInstruction()->getParent();
1226 Function *F = BB->getParent();
1227 assert(F && "must be set");
1228 Module *M = F->getParent();
1230 assert(M && "must be set");
1232 // We're not changing the function signature of the statepoint since the gc
1233 // arguments go into the var args section.
1234 Function *gc_statepoint_decl = CS.getCalledFunction();
1236 // Then go ahead and use the builder do actually do the inserts. We insert
1237 // immediately before the previous instruction under the assumption that all
1238 // arguments will be available here. We can't insert afterwards since we may
1239 // be replacing a terminator.
1240 Instruction *insertBefore = CS.getInstruction();
1241 IRBuilder<> Builder(insertBefore);
1242 // Copy all of the arguments from the original statepoint - this includes the
1243 // target, call args, and deopt args
1244 SmallVector<llvm::Value *, 64> args;
1245 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1246 // TODO: Clear the 'needs rewrite' flag
1248 // add all the pointers to be relocated (gc arguments)
1249 // Capture the start of the live variable list for use in the gc_relocates
1250 const int live_start = args.size();
1251 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1253 // Create the statepoint given all the arguments
1254 Instruction *token = nullptr;
1255 AttributeSet return_attributes;
1257 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1259 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1260 call->setTailCall(toReplace->isTailCall());
1261 call->setCallingConv(toReplace->getCallingConv());
1263 // Currently we will fail on parameter attributes and on certain
1264 // function attributes.
1265 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1266 // In case if we can handle this set of sttributes - set up function attrs
1267 // directly on statepoint and return attrs later for gc_result intrinsic.
1268 call->setAttributes(new_attrs.getFnAttributes());
1269 return_attributes = new_attrs.getRetAttributes();
1273 // Put the following gc_result and gc_relocate calls immediately after the
1274 // the old call (which we're about to delete)
1275 BasicBlock::iterator next(toReplace);
1276 assert(BB->end() != next && "not a terminator, must have next");
1278 Instruction *IP = &*(next);
1279 Builder.SetInsertPoint(IP);
1280 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1283 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1285 // Insert the new invoke into the old block. We'll remove the old one in a
1286 // moment at which point this will become the new terminator for the
1288 InvokeInst *invoke = InvokeInst::Create(
1289 gc_statepoint_decl, toReplace->getNormalDest(),
1290 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1291 invoke->setCallingConv(toReplace->getCallingConv());
1293 // Currently we will fail on parameter attributes and on certain
1294 // function attributes.
1295 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1296 // In case if we can handle this set of sttributes - set up function attrs
1297 // directly on statepoint and return attrs later for gc_result intrinsic.
1298 invoke->setAttributes(new_attrs.getFnAttributes());
1299 return_attributes = new_attrs.getRetAttributes();
1303 // Generate gc relocates in exceptional path
1304 BasicBlock *unwindBlock = normalizeBBForInvokeSafepoint(
1305 toReplace->getUnwindDest(), invoke->getParent(), P);
1307 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1308 Builder.SetInsertPoint(IP);
1309 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1311 // Extract second element from landingpad return value. We will attach
1312 // exceptional gc relocates to it.
1313 const unsigned idx = 1;
1314 Instruction *exceptional_token =
1315 cast<Instruction>(Builder.CreateExtractValue(
1316 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1317 result.UnwindToken = exceptional_token;
1319 // Just throw away return value. We will use the one we got for normal
1321 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1322 exceptional_token, Builder);
1324 // Generate gc relocates and returns for normal block
1325 BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
1326 toReplace->getNormalDest(), invoke->getParent(), P);
1328 IP = &*(normalDest->getFirstInsertionPt());
1329 Builder.SetInsertPoint(IP);
1331 // gc relocates will be generated later as if it were regular call
1336 // Take the name of the original value call if it had one.
1337 token->takeName(CS.getInstruction());
1339 // The GCResult is already inserted, we just need to find it
1341 Instruction *toReplace = CS.getInstruction();
1342 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1343 "only valid use before rewrite is gc.result");
1344 assert(!toReplace->hasOneUse() ||
1345 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1348 // Update the gc.result of the original statepoint (if any) to use the newly
1349 // inserted statepoint. This is safe to do here since the token can't be
1350 // considered a live reference.
1351 CS.getInstruction()->replaceAllUsesWith(token);
1353 result.StatepointToken = token;
1355 // Second, create a gc.relocate for every live variable
1356 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1361 struct name_ordering {
1364 bool operator()(name_ordering const &a, name_ordering const &b) {
1365 return -1 == a.derived->getName().compare(b.derived->getName());
1369 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1370 SmallVectorImpl<Value *> &livevec) {
1371 assert(basevec.size() == livevec.size());
1373 SmallVector<name_ordering, 64> temp;
1374 for (size_t i = 0; i < basevec.size(); i++) {
1376 v.base = basevec[i];
1377 v.derived = livevec[i];
1380 std::sort(temp.begin(), temp.end(), name_ordering());
1381 for (size_t i = 0; i < basevec.size(); i++) {
1382 basevec[i] = temp[i].base;
1383 livevec[i] = temp[i].derived;
1387 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1388 // which make the relocations happening at this safepoint explicit.
1390 // WARNING: Does not do any fixup to adjust users of the original live
1391 // values. That's the callers responsibility.
1393 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1394 PartiallyConstructedSafepointRecord &result) {
1395 auto liveset = result.liveset;
1396 auto PointerToBase = result.PointerToBase;
1398 // Convert to vector for efficient cross referencing.
1399 SmallVector<Value *, 64> basevec, livevec;
1400 livevec.reserve(liveset.size());
1401 basevec.reserve(liveset.size());
1402 for (Value *L : liveset) {
1403 livevec.push_back(L);
1405 assert(PointerToBase.find(L) != PointerToBase.end());
1406 Value *base = PointerToBase[L];
1407 basevec.push_back(base);
1409 assert(livevec.size() == basevec.size());
1411 // To make the output IR slightly more stable (for use in diffs), ensure a
1412 // fixed order of the values in the safepoint (by sorting the value name).
1413 // The order is otherwise meaningless.
1414 stablize_order(basevec, livevec);
1416 // Do the actual rewriting and delete the old statepoint
1417 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1418 CS.getInstruction()->eraseFromParent();
1421 // Helper function for the relocationViaAlloca.
1422 // It receives iterator to the statepoint gc relocates and emits store to the
1424 // location (via allocaMap) for the each one of them.
1425 // Add visited values into the visitedLiveValues set we will later use them
1426 // for sanity check.
1428 insertRelocationStores(iterator_range<Value::user_iterator> gcRelocs,
1429 DenseMap<Value *, Value *> &allocaMap,
1430 DenseSet<Value *> &visitedLiveValues) {
1432 for (User *U : gcRelocs) {
1433 if (!isa<IntrinsicInst>(U))
1436 IntrinsicInst *relocatedValue = cast<IntrinsicInst>(U);
1438 // We only care about relocates
1439 if (relocatedValue->getIntrinsicID() !=
1440 Intrinsic::experimental_gc_relocate) {
1444 GCRelocateOperands relocateOperands(relocatedValue);
1445 Value *originalValue = const_cast<Value *>(relocateOperands.derivedPtr());
1446 assert(allocaMap.count(originalValue));
1447 Value *alloca = allocaMap[originalValue];
1449 // Emit store into the related alloca
1450 StoreInst *store = new StoreInst(relocatedValue, alloca);
1451 store->insertAfter(relocatedValue);
1454 visitedLiveValues.insert(originalValue);
1459 /// do all the relocation update via allocas and mem2reg
1460 static void relocationViaAlloca(
1461 Function &F, DominatorTree &DT, ArrayRef<Value *> live,
1462 ArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1464 int initialAllocaNum = 0;
1466 // record initial number of allocas
1467 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1469 if (isa<AllocaInst>(*itr))
1474 // TODO-PERF: change data structures, reserve
1475 DenseMap<Value *, Value *> allocaMap;
1476 SmallVector<AllocaInst *, 200> PromotableAllocas;
1477 PromotableAllocas.reserve(live.size());
1479 // emit alloca for each live gc pointer
1480 for (unsigned i = 0; i < live.size(); i++) {
1481 Value *liveValue = live[i];
1482 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1483 F.getEntryBlock().getFirstNonPHI());
1484 allocaMap[liveValue] = alloca;
1485 PromotableAllocas.push_back(alloca);
1488 // The next two loops are part of the same conceptual operation. We need to
1489 // insert a store to the alloca after the original def and at each
1490 // redefinition. We need to insert a load before each use. These are split
1491 // into distinct loops for performance reasons.
1493 // update gc pointer after each statepoint
1494 // either store a relocated value or null (if no relocated value found for
1495 // this gc pointer and it is not a gc_result)
1496 // this must happen before we update the statepoint with load of alloca
1497 // otherwise we lose the link between statepoint and old def
1498 for (size_t i = 0; i < records.size(); i++) {
1499 const struct PartiallyConstructedSafepointRecord &info = records[i];
1500 Value *Statepoint = info.StatepointToken;
1502 // This will be used for consistency check
1503 DenseSet<Value *> visitedLiveValues;
1505 // Insert stores for normal statepoint gc relocates
1506 insertRelocationStores(Statepoint->users(), allocaMap, visitedLiveValues);
1508 // In case if it was invoke statepoint
1509 // we will insert stores for exceptional path gc relocates.
1510 if (isa<InvokeInst>(Statepoint)) {
1511 insertRelocationStores(info.UnwindToken->users(),
1512 allocaMap, visitedLiveValues);
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 slightly
1518 // easier to debug SEGVs
1519 SmallVector<AllocaInst *, 64> ToClobber;
1520 for (auto Pair : allocaMap) {
1521 Value *Def = Pair.first;
1522 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1524 // This value was relocated
1525 if (visitedLiveValues.count(Def)) {
1528 ToClobber.push_back(Alloca);
1531 auto InsertClobbersAt = [&](Instruction *IP) {
1532 for (auto *AI : ToClobber) {
1533 auto AIType = cast<PointerType>(AI->getType());
1534 auto PT = cast<PointerType>(AIType->getElementType());
1535 Constant *CPN = ConstantPointerNull::get(PT);
1536 StoreInst *store = new StoreInst(CPN, AI);
1537 store->insertBefore(IP);
1541 // Insert the clobbering stores. These may get intermixed with the
1542 // gc.results and gc.relocates, but that's fine.
1543 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1544 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1545 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1547 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1549 InsertClobbersAt(Next);
1553 // update use with load allocas and add store for gc_relocated
1554 for (auto Pair : allocaMap) {
1555 Value *def = Pair.first;
1556 Value *alloca = Pair.second;
1558 // we pre-record the uses of allocas so that we dont have to worry about
1560 // that change the user information.
1561 SmallVector<Instruction *, 20> uses;
1562 // PERF: trade a linear scan for repeated reallocation
1563 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1564 for (User *U : def->users()) {
1565 if (!isa<ConstantExpr>(U)) {
1566 // If the def has a ConstantExpr use, then the def is either a
1567 // ConstantExpr use itself or null. In either case
1568 // (recursively in the first, directly in the second), the oop
1569 // it is ultimately dependent on is null and this particular
1570 // use does not need to be fixed up.
1571 uses.push_back(cast<Instruction>(U));
1575 std::sort(uses.begin(), uses.end());
1576 auto last = std::unique(uses.begin(), uses.end());
1577 uses.erase(last, uses.end());
1579 for (Instruction *use : uses) {
1580 if (isa<PHINode>(use)) {
1581 PHINode *phi = cast<PHINode>(use);
1582 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1583 if (def == phi->getIncomingValue(i)) {
1584 LoadInst *load = new LoadInst(
1585 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1586 phi->setIncomingValue(i, load);
1590 LoadInst *load = new LoadInst(alloca, "", use);
1591 use->replaceUsesOfWith(def, load);
1595 // emit store for the initial gc value
1596 // store must be inserted after load, otherwise store will be in alloca's
1597 // use list and an extra load will be inserted before it
1598 StoreInst *store = new StoreInst(def, alloca);
1599 if (isa<Instruction>(def)) {
1600 store->insertAfter(cast<Instruction>(def));
1602 assert((isa<Argument>(def) || isa<GlobalVariable>(def) ||
1603 (isa<Constant>(def) && cast<Constant>(def)->isNullValue())) &&
1604 "Must be argument or global");
1605 store->insertAfter(cast<Instruction>(alloca));
1609 assert(PromotableAllocas.size() == live.size() &&
1610 "we must have the same allocas with lives");
1611 if (!PromotableAllocas.empty()) {
1612 // apply mem2reg to promote alloca to SSA
1613 PromoteMemToReg(PromotableAllocas, DT);
1617 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1619 if (isa<AllocaInst>(*itr))
1622 assert(initialAllocaNum == 0 && "We must not introduce any extra allocas");
1626 /// Implement a unique function which doesn't require we sort the input
1627 /// vector. Doing so has the effect of changing the output of a couple of
1628 /// tests in ways which make them less useful in testing fused safepoints.
1629 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1631 SmallVector<T, 128> TempVec;
1632 TempVec.reserve(Vec.size());
1633 for (auto Element : Vec)
1634 TempVec.push_back(Element);
1636 for (auto V : TempVec) {
1637 if (Seen.insert(V).second) {
1643 static Function *getUseHolder(Module &M) {
1644 FunctionType *ftype =
1645 FunctionType::get(Type::getVoidTy(M.getContext()), true);
1646 Function *Func = cast<Function>(M.getOrInsertFunction("__tmp_use", ftype));
1650 /// Insert holders so that each Value is obviously live through the entire
1651 /// liftetime of the call.
1652 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1653 SmallVectorImpl<CallInst *> &holders) {
1654 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1655 Function *Func = getUseHolder(*M);
1657 // For call safepoints insert dummy calls right after safepoint
1658 BasicBlock::iterator next(CS.getInstruction());
1660 CallInst *base_holder = CallInst::Create(Func, Values, "", next);
1661 holders.push_back(base_holder);
1662 } else if (CS.isInvoke()) {
1663 // For invoke safepooints insert dummy calls both in normal and
1664 // exceptional destination blocks
1665 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1666 CallInst *normal_holder = CallInst::Create(
1667 Func, Values, "", invoke->getNormalDest()->getFirstInsertionPt());
1668 CallInst *unwind_holder = CallInst::Create(
1669 Func, Values, "", invoke->getUnwindDest()->getFirstInsertionPt());
1670 holders.push_back(normal_holder);
1671 holders.push_back(unwind_holder);
1673 llvm_unreachable("unsupported call type");
1676 static void findLiveReferences(
1677 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1678 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1679 for (size_t i = 0; i < records.size(); i++) {
1680 struct PartiallyConstructedSafepointRecord &info = records[i];
1681 const CallSite &CS = toUpdate[i];
1682 analyzeParsePointLiveness(DT, CS, info);
1686 static void addBasesAsLiveValues(StatepointLiveSetTy &liveset,
1687 DenseMap<Value *, Value *> &PointerToBase) {
1688 // Identify any base pointers which are used in this safepoint, but not
1689 // themselves relocated. We need to relocate them so that later inserted
1690 // safepoints can get the properly relocated base register.
1691 DenseSet<Value *> missing;
1692 for (Value *L : liveset) {
1693 assert(PointerToBase.find(L) != PointerToBase.end());
1694 Value *base = PointerToBase[L];
1696 if (liveset.find(base) == liveset.end()) {
1697 assert(PointerToBase.find(base) == PointerToBase.end());
1698 // uniqued by set insert
1699 missing.insert(base);
1703 // Note that we want these at the end of the list, otherwise
1704 // register placement gets screwed up once we lower to STATEPOINT
1705 // instructions. This is an utter hack, but there doesn't seem to be a
1707 for (Value *base : missing) {
1709 liveset.insert(base);
1710 PointerToBase[base] = base;
1712 assert(liveset.size() == PointerToBase.size());
1715 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1716 SmallVectorImpl<CallSite> &toUpdate) {
1718 // sanity check the input
1719 std::set<CallSite> uniqued;
1720 uniqued.insert(toUpdate.begin(), toUpdate.end());
1721 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1723 for (size_t i = 0; i < toUpdate.size(); i++) {
1724 CallSite &CS = toUpdate[i];
1725 assert(CS.getInstruction()->getParent()->getParent() == &F);
1726 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1730 // A list of dummy calls added to the IR to keep various values obviously
1731 // live in the IR. We'll remove all of these when done.
1732 SmallVector<CallInst *, 64> holders;
1734 // Insert a dummy call with all of the arguments to the vm_state we'll need
1735 // for the actual safepoint insertion. This ensures reference arguments in
1736 // the deopt argument list are considered live through the safepoint (and
1737 // thus makes sure they get relocated.)
1738 for (size_t i = 0; i < toUpdate.size(); i++) {
1739 CallSite &CS = toUpdate[i];
1740 Statepoint StatepointCS(CS);
1742 SmallVector<Value *, 64> DeoptValues;
1743 for (Use &U : StatepointCS.vm_state_args()) {
1744 Value *Arg = cast<Value>(&U);
1745 if (isGCPointerType(Arg->getType()))
1746 DeoptValues.push_back(Arg);
1748 insertUseHolderAfter(CS, DeoptValues, holders);
1751 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
1752 records.reserve(toUpdate.size());
1753 for (size_t i = 0; i < toUpdate.size(); i++) {
1754 struct PartiallyConstructedSafepointRecord info;
1755 records.push_back(info);
1757 assert(records.size() == toUpdate.size());
1759 // A) Identify all gc pointers which are staticly live at the given call
1761 findLiveReferences(F, DT, P, toUpdate, records);
1763 // B) Find the base pointers for each live pointer
1764 /* scope for caching */ {
1765 // Cache the 'defining value' relation used in the computation and
1766 // insertion of base phis and selects. This ensures that we don't insert
1767 // large numbers of duplicate base_phis.
1768 DefiningValueMapTy DVCache;
1770 for (size_t i = 0; i < records.size(); i++) {
1771 struct PartiallyConstructedSafepointRecord &info = records[i];
1772 CallSite &CS = toUpdate[i];
1773 findBasePointers(DT, DVCache, CS, info);
1775 } // end of cache scope
1777 // The base phi insertion logic (for any safepoint) may have inserted new
1778 // instructions which are now live at some safepoint. The simplest such
1781 // phi a <-- will be a new base_phi here
1782 // safepoint 1 <-- that needs to be live here
1786 DenseSet<llvm::Value *> allInsertedDefs;
1787 for (size_t i = 0; i < records.size(); i++) {
1788 struct PartiallyConstructedSafepointRecord &info = records[i];
1789 allInsertedDefs.insert(info.NewInsertedDefs.begin(),
1790 info.NewInsertedDefs.end());
1793 // We insert some dummy calls after each safepoint to definitely hold live
1794 // the base pointers which were identified for that safepoint. We'll then
1795 // ask liveness for _every_ base inserted to see what is now live. Then we
1796 // remove the dummy calls.
1797 holders.reserve(holders.size() + records.size());
1798 for (size_t i = 0; i < records.size(); i++) {
1799 struct PartiallyConstructedSafepointRecord &info = records[i];
1800 CallSite &CS = toUpdate[i];
1802 SmallVector<Value *, 128> Bases;
1803 for (auto Pair : info.PointerToBase) {
1804 Bases.push_back(Pair.second);
1806 insertUseHolderAfter(CS, Bases, holders);
1809 // Add the bases explicitly to the live vector set. This may result in a few
1810 // extra relocations, but the base has to be available whenever a pointer
1811 // derived from it is used. Thus, we need it to be part of the statepoint's
1812 // gc arguments list. TODO: Introduce an explicit notion (in the following
1813 // code) of the GC argument list as seperate from the live Values at a
1814 // given statepoint.
1815 for (size_t i = 0; i < records.size(); i++) {
1816 struct PartiallyConstructedSafepointRecord &info = records[i];
1817 addBasesAsLiveValues(info.liveset, info.PointerToBase);
1820 // If we inserted any new values, we need to adjust our notion of what is
1821 // live at a particular safepoint.
1822 if (!allInsertedDefs.empty()) {
1823 fixupLiveReferences(F, DT, P, allInsertedDefs, toUpdate, records);
1825 if (PrintBasePointers) {
1826 for (size_t i = 0; i < records.size(); i++) {
1827 struct PartiallyConstructedSafepointRecord &info = records[i];
1828 errs() << "Base Pairs: (w/Relocation)\n";
1829 for (auto Pair : info.PointerToBase) {
1830 errs() << " derived %" << Pair.first->getName() << " base %"
1831 << Pair.second->getName() << "\n";
1835 for (size_t i = 0; i < holders.size(); i++) {
1836 holders[i]->eraseFromParent();
1837 holders[i] = nullptr;
1841 // Now run through and replace the existing statepoints with new ones with
1842 // the live variables listed. We do not yet update uses of the values being
1843 // relocated. We have references to live variables that need to
1844 // survive to the last iteration of this loop. (By construction, the
1845 // previous statepoint can not be a live variable, thus we can and remove
1846 // the old statepoint calls as we go.)
1847 for (size_t i = 0; i < records.size(); i++) {
1848 struct PartiallyConstructedSafepointRecord &info = records[i];
1849 CallSite &CS = toUpdate[i];
1850 makeStatepointExplicit(DT, CS, P, info);
1852 toUpdate.clear(); // prevent accident use of invalid CallSites
1854 // In case if we inserted relocates in a different basic block than the
1855 // original safepoint (this can happen for invokes). We need to be sure that
1856 // original values were not used in any of the phi nodes at the
1857 // beginning of basic block containing them. Because we know that all such
1858 // blocks will have single predecessor we can safely assume that all phi
1859 // nodes have single entry (because of normalizeBBForInvokeSafepoint).
1860 // Just remove them all here.
1861 for (size_t i = 0; i < records.size(); i++) {
1862 Instruction *I = records[i].StatepointToken;
1864 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
1865 FoldSingleEntryPHINodes(invoke->getNormalDest());
1866 assert(!isa<PHINode>(invoke->getNormalDest()->begin()));
1868 FoldSingleEntryPHINodes(invoke->getUnwindDest());
1869 assert(!isa<PHINode>(invoke->getUnwindDest()->begin()));
1873 // Do all the fixups of the original live variables to their relocated selves
1874 SmallVector<Value *, 128> live;
1875 for (size_t i = 0; i < records.size(); i++) {
1876 struct PartiallyConstructedSafepointRecord &info = records[i];
1877 // We can't simply save the live set from the original insertion. One of
1878 // the live values might be the result of a call which needs a safepoint.
1879 // That Value* no longer exists and we need to use the new gc_result.
1880 // Thankfully, the liveset is embedded in the statepoint (and updated), so
1881 // we just grab that.
1882 Statepoint statepoint(info.StatepointToken);
1883 live.insert(live.end(), statepoint.gc_args_begin(),
1884 statepoint.gc_args_end());
1886 unique_unsorted(live);
1890 for (auto ptr : live) {
1891 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
1895 relocationViaAlloca(F, DT, live, records);
1896 return !records.empty();
1899 /// Returns true if this function should be rewritten by this pass. The main
1900 /// point of this function is as an extension point for custom logic.
1901 static bool shouldRewriteStatepointsIn(Function &F) {
1902 // TODO: This should check the GCStrategy
1904 const std::string StatepointExampleName("statepoint-example");
1905 return StatepointExampleName == F.getGC();
1910 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
1911 // Nothing to do for declarations.
1912 if (F.isDeclaration() || F.empty())
1915 // Policy choice says not to rewrite - the most common reason is that we're
1916 // compiling code without a GCStrategy.
1917 if (!shouldRewriteStatepointsIn(F))
1920 // Gather all the statepoints which need rewritten.
1921 SmallVector<CallSite, 64> ParsePointNeeded;
1922 for (Instruction &I : inst_range(F)) {
1923 // TODO: only the ones with the flag set!
1924 if (isStatepoint(I))
1925 ParsePointNeeded.push_back(CallSite(&I));
1928 // Return early if no work to do.
1929 if (ParsePointNeeded.empty())
1932 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1933 return insertParsePoints(F, DT, this, ParsePointNeeded);