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 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 // PERF: Yes, this is as horribly inefficient as it looks.
691 for (auto Pair : states) {
692 Value *v = Pair.first;
693 assert(!isKnownBaseResult(v) && "why did it get added?");
694 if (PHINode *phi = dyn_cast<PHINode>(v)) {
695 assert(phi->getNumIncomingValues() > 0 &&
696 "zero input phis are illegal");
697 for (Value *InVal : phi->incoming_values()) {
698 Value *local = findBaseOrBDV(InVal, cache);
699 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
700 states[local] = PhiState();
704 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
705 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
706 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
707 states[local] = PhiState();
710 local = findBaseOrBDV(sel->getFalseValue(), cache);
711 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
712 states[local] = PhiState();
720 errs() << "States after initialization:\n";
721 for (auto Pair : states) {
722 Instruction *v = cast<Instruction>(Pair.first);
723 PhiState state = Pair.second;
729 // TODO: come back and revisit the state transitions around inputs which
730 // have reached conflict state. The current version seems too conservative.
732 bool progress = true;
735 oldSize = states.size();
737 for (auto Pair : states) {
738 MeetPhiStates calculateMeet(states);
739 Value *v = Pair.first;
740 assert(!isKnownBaseResult(v) && "why did it get added?");
741 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
742 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
743 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
745 for (Value *Val : cast<PHINode>(v)->incoming_values())
746 calculateMeet.meetWith(findBaseOrBDV(Val, cache));
748 PhiState oldState = states[v];
749 PhiState newState = calculateMeet.getResult();
750 if (oldState != newState) {
752 states[v] = newState;
756 assert(oldSize <= states.size());
757 assert(oldSize == states.size() || progress);
761 errs() << "States after meet iteration:\n";
762 for (auto Pair : states) {
763 Instruction *v = cast<Instruction>(Pair.first);
764 PhiState state = Pair.second;
770 // Insert Phis for all conflicts
771 for (auto Pair : states) {
772 Instruction *v = cast<Instruction>(Pair.first);
773 PhiState state = Pair.second;
774 assert(!isKnownBaseResult(v) && "why did it get added?");
775 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
776 if (state.isConflict()) {
777 if (isa<PHINode>(v)) {
779 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
780 assert(num_preds > 0 && "how did we reach here");
781 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
782 NewInsertedDefs.insert(phi);
783 // Add metadata marking this as a base value
784 auto *const_1 = ConstantInt::get(
786 v->getParent()->getParent()->getParent()->getContext()),
788 auto MDConst = ConstantAsMetadata::get(const_1);
789 MDNode *md = MDNode::get(
790 v->getParent()->getParent()->getParent()->getContext(), MDConst);
791 phi->setMetadata("is_base_value", md);
792 states[v] = PhiState(PhiState::Conflict, phi);
793 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
794 // The undef will be replaced later
795 UndefValue *undef = UndefValue::get(sel->getType());
796 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
797 undef, "base_select", sel);
798 NewInsertedDefs.insert(basesel);
799 // Add metadata marking this as a base value
800 auto *const_1 = ConstantInt::get(
802 v->getParent()->getParent()->getParent()->getContext()),
804 auto MDConst = ConstantAsMetadata::get(const_1);
805 MDNode *md = MDNode::get(
806 v->getParent()->getParent()->getParent()->getContext(), MDConst);
807 basesel->setMetadata("is_base_value", md);
808 states[v] = PhiState(PhiState::Conflict, basesel);
810 llvm_unreachable("unknown conflict type");
814 // Fixup all the inputs of the new PHIs
815 for (auto Pair : states) {
816 Instruction *v = cast<Instruction>(Pair.first);
817 PhiState state = Pair.second;
819 assert(!isKnownBaseResult(v) && "why did it get added?");
820 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
821 if (state.isConflict()) {
822 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
823 PHINode *phi = cast<PHINode>(v);
824 unsigned NumPHIValues = phi->getNumIncomingValues();
825 for (unsigned i = 0; i < NumPHIValues; i++) {
826 Value *InVal = phi->getIncomingValue(i);
827 BasicBlock *InBB = phi->getIncomingBlock(i);
829 // If we've already seen InBB, add the same incoming value
830 // we added for it earlier. The IR verifier requires phi
831 // nodes with multiple entries from the same basic block
832 // to have the same incoming value for each of those
833 // entries. If we don't do this check here and basephi
834 // has a different type than base, we'll end up adding two
835 // bitcasts (and hence two distinct values) as incoming
836 // values for the same basic block.
838 int blockIndex = basephi->getBasicBlockIndex(InBB);
839 if (blockIndex != -1) {
840 Value *oldBase = basephi->getIncomingValue(blockIndex);
841 basephi->addIncoming(oldBase, InBB);
843 Value *base = findBaseOrBDV(InVal, cache);
844 if (!isKnownBaseResult(base)) {
845 // Either conflict or base.
846 assert(states.count(base));
847 base = states[base].getBase();
848 assert(base != nullptr && "unknown PhiState!");
849 assert(NewInsertedDefs.count(base) &&
850 "should have already added this in a prev. iteration!");
853 // In essense this assert states: the only way two
854 // values incoming from the same basic block may be
855 // different is by being different bitcasts of the same
856 // value. A cleanup that remains TODO is changing
857 // findBaseOrBDV to return an llvm::Value of the correct
858 // type (and still remain pure). This will remove the
859 // need to add bitcasts.
860 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
861 "sanity -- findBaseOrBDV should be pure!");
866 // Find either the defining value for the PHI or the normal base for
868 Value *base = findBaseOrBDV(InVal, cache);
869 if (!isKnownBaseResult(base)) {
870 // Either conflict or base.
871 assert(states.count(base));
872 base = states[base].getBase();
873 assert(base != nullptr && "unknown PhiState!");
875 assert(base && "can't be null");
876 // Must use original input BB since base may not be Instruction
877 // The cast is needed since base traversal may strip away bitcasts
878 if (base->getType() != basephi->getType()) {
879 base = new BitCastInst(base, basephi->getType(), "cast",
880 InBB->getTerminator());
881 NewInsertedDefs.insert(base);
883 basephi->addIncoming(base, InBB);
885 assert(basephi->getNumIncomingValues() == NumPHIValues);
886 } else if (SelectInst *basesel = dyn_cast<SelectInst>(state.getBase())) {
887 SelectInst *sel = cast<SelectInst>(v);
888 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
889 // something more safe and less hacky.
890 for (int i = 1; i <= 2; i++) {
891 Value *InVal = sel->getOperand(i);
892 // Find either the defining value for the PHI or the normal base for
894 Value *base = findBaseOrBDV(InVal, cache);
895 if (!isKnownBaseResult(base)) {
896 // Either conflict or base.
897 assert(states.count(base));
898 base = states[base].getBase();
899 assert(base != nullptr && "unknown PhiState!");
901 assert(base && "can't be null");
902 // Must use original input BB since base may not be Instruction
903 // The cast is needed since base traversal may strip away bitcasts
904 if (base->getType() != basesel->getType()) {
905 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
906 NewInsertedDefs.insert(base);
908 basesel->setOperand(i, base);
911 llvm_unreachable("unexpected conflict type");
915 // Cache all of our results so we can cheaply reuse them
916 // NOTE: This is actually two caches: one of the base defining value
917 // relation and one of the base pointer relation! FIXME
918 for (auto item : states) {
919 Value *v = item.first;
920 Value *base = item.second.getBase();
922 assert(!isKnownBaseResult(v) && "why did it get added?");
925 std::string fromstr =
926 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
928 errs() << "Updating base value cache"
929 << " for: " << (v->hasName() ? v->getName() : "")
930 << " from: " << fromstr
931 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
934 assert(isKnownBaseResult(base) &&
935 "must be something we 'know' is a base pointer");
936 if (cache.count(v)) {
937 // Once we transition from the BDV relation being store in the cache to
938 // the base relation being stored, it must be stable
939 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
940 "base relation should be stable");
944 assert(cache.find(def) != cache.end());
948 // For a set of live pointers (base and/or derived), identify the base
949 // pointer of the object which they are derived from. This routine will
950 // mutate the IR graph as needed to make the 'base' pointer live at the
951 // definition site of 'derived'. This ensures that any use of 'derived' can
952 // also use 'base'. This may involve the insertion of a number of
953 // additional PHI nodes.
955 // preconditions: live is a set of pointer type Values
957 // side effects: may insert PHI nodes into the existing CFG, will preserve
958 // CFG, will not remove or mutate any existing nodes
960 // post condition: PointerToBase contains one (derived, base) pair for every
961 // pointer in live. Note that derived can be equal to base if the original
962 // pointer was a base pointer.
963 static void findBasePointers(const StatepointLiveSetTy &live,
964 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
965 DominatorTree *DT, DefiningValueMapTy &DVCache,
966 DenseSet<llvm::Value *> &NewInsertedDefs) {
967 for (Value *ptr : live) {
968 Value *base = findBasePointer(ptr, DVCache, NewInsertedDefs);
969 assert(base && "failed to find base pointer");
970 PointerToBase[ptr] = base;
971 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
972 DT->dominates(cast<Instruction>(base)->getParent(),
973 cast<Instruction>(ptr)->getParent())) &&
974 "The base we found better dominate the derived pointer");
976 // If you see this trip and like to live really dangerously, the code should
977 // be correct, just with idioms the verifier can't handle. You can try
978 // disabling the verifier at your own substaintial risk.
979 assert(!isNullConstant(base) && "the relocation code needs adjustment to "
980 "handle the relocation of a null pointer "
981 "constant without causing false positives "
982 "in the safepoint ir verifier.");
986 /// Find the required based pointers (and adjust the live set) for the given
988 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
990 PartiallyConstructedSafepointRecord &result) {
991 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
992 DenseSet<llvm::Value *> NewInsertedDefs;
993 findBasePointers(result.liveset, PointerToBase, &DT, DVCache, NewInsertedDefs);
995 if (PrintBasePointers) {
996 errs() << "Base Pairs (w/o Relocation):\n";
997 for (auto Pair : PointerToBase) {
998 errs() << " derived %" << Pair.first->getName() << " base %"
999 << Pair.second->getName() << "\n";
1003 result.PointerToBase = PointerToBase;
1004 result.NewInsertedDefs = NewInsertedDefs;
1007 /// Check for liveness of items in the insert defs and add them to the live
1008 /// and base pointer sets
1009 static void fixupLiveness(DominatorTree &DT, const CallSite &CS,
1010 const DenseSet<Value *> &allInsertedDefs,
1011 PartiallyConstructedSafepointRecord &result) {
1012 Instruction *inst = CS.getInstruction();
1014 auto liveset = result.liveset;
1015 auto PointerToBase = result.PointerToBase;
1017 auto is_live_gc_reference =
1018 [&](Value &V) { return isLiveGCReferenceAt(V, inst, DT, nullptr); };
1020 // For each new definition, check to see if a) the definition dominates the
1021 // instruction we're interested in, and b) one of the uses of that definition
1022 // is edge-reachable from the instruction we're interested in. This is the
1023 // same definition of liveness we used in the intial liveness analysis
1024 for (Value *newDef : allInsertedDefs) {
1025 if (liveset.count(newDef)) {
1026 // already live, no action needed
1030 // PERF: Use DT to check instruction domination might not be good for
1031 // compilation time, and we could change to optimal solution if this
1032 // turn to be a issue
1033 if (!DT.dominates(cast<Instruction>(newDef), inst)) {
1034 // can't possibly be live at inst
1038 if (is_live_gc_reference(*newDef)) {
1039 // Add the live new defs into liveset and PointerToBase
1040 liveset.insert(newDef);
1041 PointerToBase[newDef] = newDef;
1045 result.liveset = liveset;
1046 result.PointerToBase = PointerToBase;
1049 static void fixupLiveReferences(
1050 Function &F, DominatorTree &DT, Pass *P,
1051 const DenseSet<llvm::Value *> &allInsertedDefs,
1052 ArrayRef<CallSite> toUpdate,
1053 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1054 for (size_t i = 0; i < records.size(); i++) {
1055 struct PartiallyConstructedSafepointRecord &info = records[i];
1056 const CallSite &CS = toUpdate[i];
1057 fixupLiveness(DT, CS, allInsertedDefs, info);
1061 // Normalize basic block to make it ready to be target of invoke statepoint.
1062 // It means spliting it to have single predecessor. Return newly created BB
1063 // ready to be successor of invoke statepoint.
1064 static BasicBlock *normalizeBBForInvokeSafepoint(BasicBlock *BB,
1065 BasicBlock *InvokeParent,
1067 BasicBlock *ret = BB;
1069 if (!BB->getUniquePredecessor()) {
1070 ret = SplitBlockPredecessors(BB, InvokeParent, "");
1073 // Another requirement for such basic blocks is to not have any phi nodes.
1074 // Since we just ensured that new BB will have single predecessor,
1075 // all phi nodes in it will have one value. Here it would be naturall place
1077 // remove them all. But we can not do this because we are risking to remove
1078 // one of the values stored in liveset of another statepoint. We will do it
1079 // later after placing all safepoints.
1084 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1085 auto itr = std::find(livevec.begin(), livevec.end(), val);
1086 assert(livevec.end() != itr);
1087 size_t index = std::distance(livevec.begin(), itr);
1088 assert(index < livevec.size());
1092 // Create new attribute set containing only attributes which can be transfered
1093 // from original call to the safepoint.
1094 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1097 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1098 unsigned index = AS.getSlotIndex(Slot);
1100 if (index == AttributeSet::ReturnIndex ||
1101 index == AttributeSet::FunctionIndex) {
1103 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1105 Attribute attr = *it;
1107 // Do not allow certain attributes - just skip them
1108 // Safepoint can not be read only or read none.
1109 if (attr.hasAttribute(Attribute::ReadNone) ||
1110 attr.hasAttribute(Attribute::ReadOnly))
1113 ret = ret.addAttributes(
1114 AS.getContext(), index,
1115 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1119 // Just skip parameter attributes for now
1125 /// Helper function to place all gc relocates necessary for the given
1128 /// liveVariables - list of variables to be relocated.
1129 /// liveStart - index of the first live variable.
1130 /// basePtrs - base pointers.
1131 /// statepointToken - statepoint instruction to which relocates should be
1133 /// Builder - Llvm IR builder to be used to construct new calls.
1134 void CreateGCRelocates(ArrayRef<llvm::Value *> liveVariables,
1135 const int liveStart,
1136 ArrayRef<llvm::Value *> basePtrs,
1137 Instruction *statepointToken, IRBuilder<> Builder) {
1139 SmallVector<Instruction *, 64> NewDefs;
1140 NewDefs.reserve(liveVariables.size());
1142 Module *M = statepointToken->getParent()->getParent()->getParent();
1144 for (unsigned i = 0; i < liveVariables.size(); i++) {
1145 // We generate a (potentially) unique declaration for every pointer type
1146 // combination. This results is some blow up the function declarations in
1147 // the IR, but removes the need for argument bitcasts which shrinks the IR
1148 // greatly and makes it much more readable.
1149 SmallVector<Type *, 1> types; // one per 'any' type
1150 types.push_back(liveVariables[i]->getType()); // result type
1151 Value *gc_relocate_decl = Intrinsic::getDeclaration(
1152 M, Intrinsic::experimental_gc_relocate, types);
1154 // Generate the gc.relocate call and save the result
1156 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1157 liveStart + find_index(liveVariables, basePtrs[i]));
1158 Value *liveIdx = ConstantInt::get(
1159 Type::getInt32Ty(M->getContext()),
1160 liveStart + find_index(liveVariables, liveVariables[i]));
1162 // only specify a debug name if we can give a useful one
1163 Value *reloc = Builder.CreateCall3(
1164 gc_relocate_decl, statepointToken, baseIdx, liveIdx,
1165 liveVariables[i]->hasName() ? liveVariables[i]->getName() + ".relocated"
1167 // Trick CodeGen into thinking there are lots of free registers at this
1169 cast<CallInst>(reloc)->setCallingConv(CallingConv::Cold);
1171 NewDefs.push_back(cast<Instruction>(reloc));
1173 assert(NewDefs.size() == liveVariables.size() &&
1174 "missing or extra redefinition at safepoint");
1178 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1179 const SmallVectorImpl<llvm::Value *> &basePtrs,
1180 const SmallVectorImpl<llvm::Value *> &liveVariables,
1182 PartiallyConstructedSafepointRecord &result) {
1183 assert(basePtrs.size() == liveVariables.size());
1184 assert(isStatepoint(CS) &&
1185 "This method expects to be rewriting a statepoint");
1187 BasicBlock *BB = CS.getInstruction()->getParent();
1189 Function *F = BB->getParent();
1190 assert(F && "must be set");
1191 Module *M = F->getParent();
1193 assert(M && "must be set");
1195 // We're not changing the function signature of the statepoint since the gc
1196 // arguments go into the var args section.
1197 Function *gc_statepoint_decl = CS.getCalledFunction();
1199 // Then go ahead and use the builder do actually do the inserts. We insert
1200 // immediately before the previous instruction under the assumption that all
1201 // arguments will be available here. We can't insert afterwards since we may
1202 // be replacing a terminator.
1203 Instruction *insertBefore = CS.getInstruction();
1204 IRBuilder<> Builder(insertBefore);
1205 // Copy all of the arguments from the original statepoint - this includes the
1206 // target, call args, and deopt args
1207 SmallVector<llvm::Value *, 64> args;
1208 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1209 // TODO: Clear the 'needs rewrite' flag
1211 // add all the pointers to be relocated (gc arguments)
1212 // Capture the start of the live variable list for use in the gc_relocates
1213 const int live_start = args.size();
1214 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1216 // Create the statepoint given all the arguments
1217 Instruction *token = nullptr;
1218 AttributeSet return_attributes;
1220 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1222 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1223 call->setTailCall(toReplace->isTailCall());
1224 call->setCallingConv(toReplace->getCallingConv());
1226 // Currently we will fail on parameter attributes and on certain
1227 // function attributes.
1228 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1229 // In case if we can handle this set of sttributes - set up function attrs
1230 // directly on statepoint and return attrs later for gc_result intrinsic.
1231 call->setAttributes(new_attrs.getFnAttributes());
1232 return_attributes = new_attrs.getRetAttributes();
1236 // Put the following gc_result and gc_relocate calls immediately after the
1237 // the old call (which we're about to delete)
1238 BasicBlock::iterator next(toReplace);
1239 assert(BB->end() != next && "not a terminator, must have next");
1241 Instruction *IP = &*(next);
1242 Builder.SetInsertPoint(IP);
1243 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1246 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1248 // Insert the new invoke into the old block. We'll remove the old one in a
1249 // moment at which point this will become the new terminator for the
1251 InvokeInst *invoke = InvokeInst::Create(
1252 gc_statepoint_decl, toReplace->getNormalDest(),
1253 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1254 invoke->setCallingConv(toReplace->getCallingConv());
1256 // Currently we will fail on parameter attributes and on certain
1257 // function attributes.
1258 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1259 // In case if we can handle this set of sttributes - set up function attrs
1260 // directly on statepoint and return attrs later for gc_result intrinsic.
1261 invoke->setAttributes(new_attrs.getFnAttributes());
1262 return_attributes = new_attrs.getRetAttributes();
1266 // Generate gc relocates in exceptional path
1267 BasicBlock *unwindBlock = normalizeBBForInvokeSafepoint(
1268 toReplace->getUnwindDest(), invoke->getParent(), P);
1270 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1271 Builder.SetInsertPoint(IP);
1272 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1274 // Extract second element from landingpad return value. We will attach
1275 // exceptional gc relocates to it.
1276 const unsigned idx = 1;
1277 Instruction *exceptional_token =
1278 cast<Instruction>(Builder.CreateExtractValue(
1279 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1280 result.UnwindToken = exceptional_token;
1282 // Just throw away return value. We will use the one we got for normal
1284 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1285 exceptional_token, Builder);
1287 // Generate gc relocates and returns for normal block
1288 BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
1289 toReplace->getNormalDest(), invoke->getParent(), P);
1291 IP = &*(normalDest->getFirstInsertionPt());
1292 Builder.SetInsertPoint(IP);
1294 // gc relocates will be generated later as if it were regular call
1299 // Take the name of the original value call if it had one.
1300 token->takeName(CS.getInstruction());
1302 // The GCResult is already inserted, we just need to find it
1304 Instruction *toReplace = CS.getInstruction();
1305 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1306 "only valid use before rewrite is gc.result");
1307 if (toReplace->hasOneUse()) {
1308 assert(isGCResult(cast<Instruction>(*toReplace->user_begin())));
1312 // Update the gc.result of the original statepoint (if any) to use the newly
1313 // inserted statepoint. This is safe to do here since the token can't be
1314 // considered a live reference.
1315 CS.getInstruction()->replaceAllUsesWith(token);
1317 result.StatepointToken = token;
1319 // Second, create a gc.relocate for every live variable
1320 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1325 struct name_ordering {
1328 bool operator()(name_ordering const &a, name_ordering const &b) {
1329 return -1 == a.derived->getName().compare(b.derived->getName());
1333 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1334 SmallVectorImpl<Value *> &livevec) {
1335 assert(basevec.size() == livevec.size());
1337 SmallVector<name_ordering, 64> temp;
1338 for (size_t i = 0; i < basevec.size(); i++) {
1340 v.base = basevec[i];
1341 v.derived = livevec[i];
1344 std::sort(temp.begin(), temp.end(), name_ordering());
1345 for (size_t i = 0; i < basevec.size(); i++) {
1346 basevec[i] = temp[i].base;
1347 livevec[i] = temp[i].derived;
1351 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1352 // which make the relocations happening at this safepoint explicit.
1354 // WARNING: Does not do any fixup to adjust users of the original live
1355 // values. That's the callers responsibility.
1357 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1358 PartiallyConstructedSafepointRecord &result) {
1359 auto liveset = result.liveset;
1360 auto PointerToBase = result.PointerToBase;
1362 // Convert to vector for efficient cross referencing.
1363 SmallVector<Value *, 64> basevec, livevec;
1364 livevec.reserve(liveset.size());
1365 basevec.reserve(liveset.size());
1366 for (Value *L : liveset) {
1367 livevec.push_back(L);
1369 assert(PointerToBase.find(L) != PointerToBase.end());
1370 Value *base = PointerToBase[L];
1371 basevec.push_back(base);
1373 assert(livevec.size() == basevec.size());
1375 // To make the output IR slightly more stable (for use in diffs), ensure a
1376 // fixed order of the values in the safepoint (by sorting the value name).
1377 // The order is otherwise meaningless.
1378 stablize_order(basevec, livevec);
1380 // Do the actual rewriting and delete the old statepoint
1381 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1382 CS.getInstruction()->eraseFromParent();
1385 // Helper function for the relocationViaAlloca.
1386 // It receives iterator to the statepoint gc relocates and emits store to the
1388 // location (via allocaMap) for the each one of them.
1389 // Add visited values into the visitedLiveValues set we will later use them
1390 // for sanity check.
1392 insertRelocationStores(iterator_range<Value::user_iterator> gcRelocs,
1393 DenseMap<Value *, Value *> &allocaMap,
1394 DenseSet<Value *> &visitedLiveValues) {
1396 for (User *U : gcRelocs) {
1397 if (!isa<IntrinsicInst>(U))
1400 IntrinsicInst *relocatedValue = cast<IntrinsicInst>(U);
1402 // We only care about relocates
1403 if (relocatedValue->getIntrinsicID() !=
1404 Intrinsic::experimental_gc_relocate) {
1408 GCRelocateOperands relocateOperands(relocatedValue);
1409 Value *originalValue = const_cast<Value *>(relocateOperands.derivedPtr());
1410 assert(allocaMap.count(originalValue));
1411 Value *alloca = allocaMap[originalValue];
1413 // Emit store into the related alloca
1414 StoreInst *store = new StoreInst(relocatedValue, alloca);
1415 store->insertAfter(relocatedValue);
1418 visitedLiveValues.insert(originalValue);
1423 /// do all the relocation update via allocas and mem2reg
1424 static void relocationViaAlloca(
1425 Function &F, DominatorTree &DT, ArrayRef<Value *> live,
1426 ArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1428 int initialAllocaNum = 0;
1430 // record initial number of allocas
1431 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1433 if (isa<AllocaInst>(*itr))
1438 // TODO-PERF: change data structures, reserve
1439 DenseMap<Value *, Value *> allocaMap;
1440 SmallVector<AllocaInst *, 200> PromotableAllocas;
1441 PromotableAllocas.reserve(live.size());
1443 // emit alloca for each live gc pointer
1444 for (unsigned i = 0; i < live.size(); i++) {
1445 Value *liveValue = live[i];
1446 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1447 F.getEntryBlock().getFirstNonPHI());
1448 allocaMap[liveValue] = alloca;
1449 PromotableAllocas.push_back(alloca);
1452 // The next two loops are part of the same conceptual operation. We need to
1453 // insert a store to the alloca after the original def and at each
1454 // redefinition. We need to insert a load before each use. These are split
1455 // into distinct loops for performance reasons.
1457 // update gc pointer after each statepoint
1458 // either store a relocated value or null (if no relocated value found for
1459 // this gc pointer and it is not a gc_result)
1460 // this must happen before we update the statepoint with load of alloca
1461 // otherwise we lose the link between statepoint and old def
1462 for (size_t i = 0; i < records.size(); i++) {
1463 const struct PartiallyConstructedSafepointRecord &info = records[i];
1464 Value *Statepoint = info.StatepointToken;
1466 // This will be used for consistency check
1467 DenseSet<Value *> visitedLiveValues;
1469 // Insert stores for normal statepoint gc relocates
1470 insertRelocationStores(Statepoint->users(), allocaMap, visitedLiveValues);
1472 // In case if it was invoke statepoint
1473 // we will insert stores for exceptional path gc relocates.
1474 if (isa<InvokeInst>(Statepoint)) {
1475 insertRelocationStores(info.UnwindToken->users(),
1476 allocaMap, visitedLiveValues);
1480 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1481 // the gc.statepoint. This will turn some subtle GC problems into slightly
1482 // easier to debug SEGVs
1483 SmallVector<AllocaInst *, 64> ToClobber;
1484 for (auto Pair : allocaMap) {
1485 Value *Def = Pair.first;
1486 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1488 // This value was relocated
1489 if (visitedLiveValues.count(Def)) {
1492 ToClobber.push_back(Alloca);
1495 auto InsertClobbersAt = [&](Instruction *IP) {
1496 for (auto *AI : ToClobber) {
1497 auto AIType = cast<PointerType>(AI->getType());
1498 auto PT = cast<PointerType>(AIType->getElementType());
1499 Constant *CPN = ConstantPointerNull::get(PT);
1500 StoreInst *store = new StoreInst(CPN, AI);
1501 store->insertBefore(IP);
1505 // Insert the clobbering stores. These may get intermixed with the
1506 // gc.results and gc.relocates, but that's fine.
1507 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1508 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1509 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1511 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1513 InsertClobbersAt(Next);
1517 // update use with load allocas and add store for gc_relocated
1518 for (auto Pair : allocaMap) {
1519 Value *def = Pair.first;
1520 Value *alloca = Pair.second;
1522 // we pre-record the uses of allocas so that we dont have to worry about
1524 // that change the user information.
1525 SmallVector<Instruction *, 20> uses;
1526 // PERF: trade a linear scan for repeated reallocation
1527 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1528 for (User *U : def->users()) {
1529 if (!isa<ConstantExpr>(U)) {
1530 // If the def has a ConstantExpr use, then the def is either a
1531 // ConstantExpr use itself or null. In either case
1532 // (recursively in the first, directly in the second), the oop
1533 // it is ultimately dependent on is null and this particular
1534 // use does not need to be fixed up.
1535 uses.push_back(cast<Instruction>(U));
1539 std::sort(uses.begin(), uses.end());
1540 auto last = std::unique(uses.begin(), uses.end());
1541 uses.erase(last, uses.end());
1543 for (Instruction *use : uses) {
1544 if (isa<PHINode>(use)) {
1545 PHINode *phi = cast<PHINode>(use);
1546 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1547 if (def == phi->getIncomingValue(i)) {
1548 LoadInst *load = new LoadInst(
1549 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1550 phi->setIncomingValue(i, load);
1554 LoadInst *load = new LoadInst(alloca, "", use);
1555 use->replaceUsesOfWith(def, load);
1559 // emit store for the initial gc value
1560 // store must be inserted after load, otherwise store will be in alloca's
1561 // use list and an extra load will be inserted before it
1562 StoreInst *store = new StoreInst(def, alloca);
1563 if (isa<Instruction>(def)) {
1564 store->insertAfter(cast<Instruction>(def));
1566 assert((isa<Argument>(def) || isa<GlobalVariable>(def) ||
1567 (isa<Constant>(def) && cast<Constant>(def)->isNullValue())) &&
1568 "Must be argument or global");
1569 store->insertAfter(cast<Instruction>(alloca));
1573 assert(PromotableAllocas.size() == live.size() &&
1574 "we must have the same allocas with lives");
1575 if (!PromotableAllocas.empty()) {
1576 // apply mem2reg to promote alloca to SSA
1577 PromoteMemToReg(PromotableAllocas, DT);
1581 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1583 if (isa<AllocaInst>(*itr))
1586 assert(initialAllocaNum == 0 && "We must not introduce any extra allocas");
1590 /// Implement a unique function which doesn't require we sort the input
1591 /// vector. Doing so has the effect of changing the output of a couple of
1592 /// tests in ways which make them less useful in testing fused safepoints.
1593 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1595 SmallVector<T, 128> TempVec;
1596 TempVec.reserve(Vec.size());
1597 for (auto Element : Vec)
1598 TempVec.push_back(Element);
1600 for (auto V : TempVec) {
1601 if (Seen.insert(V).second) {
1607 static Function *getUseHolder(Module &M) {
1608 FunctionType *ftype =
1609 FunctionType::get(Type::getVoidTy(M.getContext()), true);
1610 Function *Func = cast<Function>(M.getOrInsertFunction("__tmp_use", ftype));
1614 /// Insert holders so that each Value is obviously live through the entire
1615 /// liftetime of the call.
1616 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1617 SmallVectorImpl<CallInst *> &holders) {
1618 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1619 Function *Func = getUseHolder(*M);
1621 // For call safepoints insert dummy calls right after safepoint
1622 BasicBlock::iterator next(CS.getInstruction());
1624 CallInst *base_holder = CallInst::Create(Func, Values, "", next);
1625 holders.push_back(base_holder);
1626 } else if (CS.isInvoke()) {
1627 // For invoke safepooints insert dummy calls both in normal and
1628 // exceptional destination blocks
1629 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1630 CallInst *normal_holder = CallInst::Create(
1631 Func, Values, "", invoke->getNormalDest()->getFirstInsertionPt());
1632 CallInst *unwind_holder = CallInst::Create(
1633 Func, Values, "", invoke->getUnwindDest()->getFirstInsertionPt());
1634 holders.push_back(normal_holder);
1635 holders.push_back(unwind_holder);
1637 llvm_unreachable("unsupported call type");
1640 static void findLiveReferences(
1641 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1642 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1643 for (size_t i = 0; i < records.size(); i++) {
1644 struct PartiallyConstructedSafepointRecord &info = records[i];
1645 const CallSite &CS = toUpdate[i];
1646 analyzeParsePointLiveness(DT, CS, info);
1650 static void addBasesAsLiveValues(StatepointLiveSetTy &liveset,
1651 DenseMap<Value *, Value *> &PointerToBase) {
1652 // Identify any base pointers which are used in this safepoint, but not
1653 // themselves relocated. We need to relocate them so that later inserted
1654 // safepoints can get the properly relocated base register.
1655 DenseSet<Value *> missing;
1656 for (Value *L : liveset) {
1657 assert(PointerToBase.find(L) != PointerToBase.end());
1658 Value *base = PointerToBase[L];
1660 if (liveset.find(base) == liveset.end()) {
1661 assert(PointerToBase.find(base) == PointerToBase.end());
1662 // uniqued by set insert
1663 missing.insert(base);
1667 // Note that we want these at the end of the list, otherwise
1668 // register placement gets screwed up once we lower to STATEPOINT
1669 // instructions. This is an utter hack, but there doesn't seem to be a
1671 for (Value *base : missing) {
1673 liveset.insert(base);
1674 PointerToBase[base] = base;
1676 assert(liveset.size() == PointerToBase.size());
1679 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1680 SmallVectorImpl<CallSite> &toUpdate) {
1682 // sanity check the input
1683 std::set<CallSite> uniqued;
1684 uniqued.insert(toUpdate.begin(), toUpdate.end());
1685 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1687 for (size_t i = 0; i < toUpdate.size(); i++) {
1688 CallSite &CS = toUpdate[i];
1689 assert(CS.getInstruction()->getParent()->getParent() == &F);
1690 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1694 // A list of dummy calls added to the IR to keep various values obviously
1695 // live in the IR. We'll remove all of these when done.
1696 SmallVector<CallInst *, 64> holders;
1698 // Insert a dummy call with all of the arguments to the vm_state we'll need
1699 // for the actual safepoint insertion. This ensures reference arguments in
1700 // the deopt argument list are considered live through the safepoint (and
1701 // thus makes sure they get relocated.)
1702 for (size_t i = 0; i < toUpdate.size(); i++) {
1703 CallSite &CS = toUpdate[i];
1704 Statepoint StatepointCS(CS);
1706 SmallVector<Value *, 64> DeoptValues;
1707 for (Use &U : StatepointCS.vm_state_args()) {
1708 Value *Arg = cast<Value>(&U);
1709 if (isGCPointerType(Arg->getType()))
1710 DeoptValues.push_back(Arg);
1712 insertUseHolderAfter(CS, DeoptValues, holders);
1715 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
1716 records.reserve(toUpdate.size());
1717 for (size_t i = 0; i < toUpdate.size(); i++) {
1718 struct PartiallyConstructedSafepointRecord info;
1719 records.push_back(info);
1721 assert(records.size() == toUpdate.size());
1723 // A) Identify all gc pointers which are staticly live at the given call
1725 findLiveReferences(F, DT, P, toUpdate, records);
1727 // B) Find the base pointers for each live pointer
1728 /* scope for caching */ {
1729 // Cache the 'defining value' relation used in the computation and
1730 // insertion of base phis and selects. This ensures that we don't insert
1731 // large numbers of duplicate base_phis.
1732 DefiningValueMapTy DVCache;
1734 for (size_t i = 0; i < records.size(); i++) {
1735 struct PartiallyConstructedSafepointRecord &info = records[i];
1736 CallSite &CS = toUpdate[i];
1737 findBasePointers(DT, DVCache, CS, info);
1739 } // end of cache scope
1741 // The base phi insertion logic (for any safepoint) may have inserted new
1742 // instructions which are now live at some safepoint. The simplest such
1745 // phi a <-- will be a new base_phi here
1746 // safepoint 1 <-- that needs to be live here
1750 DenseSet<llvm::Value *> allInsertedDefs;
1751 for (size_t i = 0; i < records.size(); i++) {
1752 struct PartiallyConstructedSafepointRecord &info = records[i];
1753 allInsertedDefs.insert(info.NewInsertedDefs.begin(),
1754 info.NewInsertedDefs.end());
1757 // We insert some dummy calls after each safepoint to definitely hold live
1758 // the base pointers which were identified for that safepoint. We'll then
1759 // ask liveness for _every_ base inserted to see what is now live. Then we
1760 // remove the dummy calls.
1761 holders.reserve(holders.size() + records.size());
1762 for (size_t i = 0; i < records.size(); i++) {
1763 struct PartiallyConstructedSafepointRecord &info = records[i];
1764 CallSite &CS = toUpdate[i];
1766 SmallVector<Value *, 128> Bases;
1767 for (auto Pair : info.PointerToBase) {
1768 Bases.push_back(Pair.second);
1770 insertUseHolderAfter(CS, Bases, holders);
1773 // Add the bases explicitly to the live vector set. This may result in a few
1774 // extra relocations, but the base has to be available whenever a pointer
1775 // derived from it is used. Thus, we need it to be part of the statepoint's
1776 // gc arguments list. TODO: Introduce an explicit notion (in the following
1777 // code) of the GC argument list as seperate from the live Values at a
1778 // given statepoint.
1779 for (size_t i = 0; i < records.size(); i++) {
1780 struct PartiallyConstructedSafepointRecord &info = records[i];
1781 addBasesAsLiveValues(info.liveset, info.PointerToBase);
1784 // If we inserted any new values, we need to adjust our notion of what is
1785 // live at a particular safepoint.
1786 if (!allInsertedDefs.empty()) {
1787 fixupLiveReferences(F, DT, P, allInsertedDefs, toUpdate, records);
1789 if (PrintBasePointers) {
1790 for (size_t i = 0; i < records.size(); i++) {
1791 struct PartiallyConstructedSafepointRecord &info = records[i];
1792 errs() << "Base Pairs: (w/Relocation)\n";
1793 for (auto Pair : info.PointerToBase) {
1794 errs() << " derived %" << Pair.first->getName() << " base %"
1795 << Pair.second->getName() << "\n";
1799 for (size_t i = 0; i < holders.size(); i++) {
1800 holders[i]->eraseFromParent();
1801 holders[i] = nullptr;
1805 // Now run through and replace the existing statepoints with new ones with
1806 // the live variables listed. We do not yet update uses of the values being
1807 // relocated. We have references to live variables that need to
1808 // survive to the last iteration of this loop. (By construction, the
1809 // previous statepoint can not be a live variable, thus we can and remove
1810 // the old statepoint calls as we go.)
1811 for (size_t i = 0; i < records.size(); i++) {
1812 struct PartiallyConstructedSafepointRecord &info = records[i];
1813 CallSite &CS = toUpdate[i];
1814 makeStatepointExplicit(DT, CS, P, info);
1816 toUpdate.clear(); // prevent accident use of invalid CallSites
1818 // In case if we inserted relocates in a different basic block than the
1819 // original safepoint (this can happen for invokes). We need to be sure that
1820 // original values were not used in any of the phi nodes at the
1821 // beginning of basic block containing them. Because we know that all such
1822 // blocks will have single predecessor we can safely assume that all phi
1823 // nodes have single entry (because of normalizeBBForInvokeSafepoint).
1824 // Just remove them all here.
1825 for (size_t i = 0; i < records.size(); i++) {
1826 Instruction *I = records[i].StatepointToken;
1828 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
1829 FoldSingleEntryPHINodes(invoke->getNormalDest());
1830 assert(!isa<PHINode>(invoke->getNormalDest()->begin()));
1832 FoldSingleEntryPHINodes(invoke->getUnwindDest());
1833 assert(!isa<PHINode>(invoke->getUnwindDest()->begin()));
1837 // Do all the fixups of the original live variables to their relocated selves
1838 SmallVector<Value *, 128> live;
1839 for (size_t i = 0; i < records.size(); i++) {
1840 struct PartiallyConstructedSafepointRecord &info = records[i];
1841 // We can't simply save the live set from the original insertion. One of
1842 // the live values might be the result of a call which needs a safepoint.
1843 // That Value* no longer exists and we need to use the new gc_result.
1844 // Thankfully, the liveset is embedded in the statepoint (and updated), so
1845 // we just grab that.
1846 Statepoint statepoint(info.StatepointToken);
1847 live.insert(live.end(), statepoint.gc_args_begin(),
1848 statepoint.gc_args_end());
1850 unique_unsorted(live);
1854 for (auto ptr : live) {
1855 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
1859 relocationViaAlloca(F, DT, live, records);
1860 return !records.empty();
1863 /// Returns true if this function should be rewritten by this pass. The main
1864 /// point of this function is as an extension point for custom logic.
1865 static bool shouldRewriteStatepointsIn(Function &F) {
1866 // TODO: This should check the GCStrategy
1868 const std::string StatepointExampleName("statepoint-example");
1869 return StatepointExampleName == F.getGC();
1874 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
1875 // Nothing to do for declarations.
1876 if (F.isDeclaration() || F.empty())
1879 // Policy choice says not to rewrite - the most common reason is that we're
1880 // compiling code without a GCStrategy.
1881 if (!shouldRewriteStatepointsIn(F))
1884 // Gather all the statepoints which need rewritten.
1885 SmallVector<CallSite, 64> ParsePointNeeded;
1886 for (Instruction &I : inst_range(F)) {
1887 // TODO: only the ones with the flag set!
1888 if (isStatepoint(I))
1889 ParsePointNeeded.push_back(CallSite(&I));
1892 // Return early if no work to do.
1893 if (ParsePointNeeded.empty())
1896 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1897 return insertParsePoints(F, DT, this, ParsePointNeeded);