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 std::map<Value *, Value *> DefiningValueMapTy;
100 struct PartiallyConstructedSafepointRecord {
101 /// The set of values known to be live accross this safepoint
102 std::set<llvm::Value *> liveset;
104 /// Mapping from live pointers to a base-defining-value
105 std::map<llvm::Value *, llvm::Value *> base_pairs;
107 /// Any new values which were added to the IR during base pointer analysis
108 /// for this safepoint
109 std::set<llvm::Value *> newInsertedDefs;
111 /// The bounds of the inserted code for the safepoint
112 std::pair<Instruction *, Instruction *> safepoint;
114 // Instruction to which exceptional gc relocates are attached
115 // Makes it easier to iterate through them during relocationViaAlloca.
116 Instruction *exceptional_relocates_token;
118 /// The result of the safepointing call (or nullptr)
123 // TODO: Once we can get to the GCStrategy, this becomes
124 // Optional<bool> isGCManagedPointer(const Value *V) const override {
126 static bool isGCPointerType(const Type *T) {
127 if (const PointerType *PT = dyn_cast<PointerType>(T))
128 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
129 // GC managed heap. We know that a pointer into this heap needs to be
130 // updated and that no other pointer does.
131 return (1 == PT->getAddressSpace());
135 /// Return true if the Value is a gc reference type which is potentially used
136 /// after the instruction 'loc'. This is only used with the edge reachability
137 /// liveness code. Note: It is assumed the V dominates loc.
138 static bool isLiveGCReferenceAt(Value &V, Instruction *loc, DominatorTree &DT,
140 if (!isGCPointerType(V.getType()))
146 // Given assumption that V dominates loc, this may be live
151 static bool isAggWhichContainsGCPtrType(Type *Ty) {
152 if (VectorType *VT = dyn_cast<VectorType>(Ty))
153 return isGCPointerType(VT->getScalarType());
154 else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
155 return isGCPointerType(AT->getElementType()) ||
156 isAggWhichContainsGCPtrType(AT->getElementType());
157 } else if (StructType *ST = dyn_cast<StructType>(Ty)) {
158 bool UnsupportedType = false;
159 for (Type *SubType : ST->subtypes())
161 isGCPointerType(SubType) || isAggWhichContainsGCPtrType(SubType);
162 return UnsupportedType;
168 // Conservatively identifies any definitions which might be live at the
169 // given instruction. The analysis is performed immediately before the
170 // given instruction. Values defined by that instruction are not considered
171 // live. Values used by that instruction are considered live.
173 // preconditions: valid IR graph, term is either a terminator instruction or
174 // a call instruction, pred is the basic block of term, DT, LI are valid
176 // side effects: none, does not mutate IR
178 // postconditions: populates liveValues as discussed above
179 static void findLiveGCValuesAtInst(Instruction *term, BasicBlock *pred,
180 DominatorTree &DT, LoopInfo *LI,
181 std::set<llvm::Value *> &liveValues) {
184 assert(isa<CallInst>(term) || isa<InvokeInst>(term) || term->isTerminator());
186 Function *F = pred->getParent();
188 auto is_live_gc_reference =
189 [&](Value &V) { return isLiveGCReferenceAt(V, term, DT, LI); };
191 // Are there any gc pointer arguments live over this point? This needs to be
192 // special cased since arguments aren't defined in basic blocks.
193 for (Argument &arg : F->args()) {
194 assert(!isAggWhichContainsGCPtrType(arg.getType()) &&
195 "support for FCA unimplemented");
197 if (is_live_gc_reference(arg)) {
198 liveValues.insert(&arg);
202 // Walk through all dominating blocks - the ones which can contain
203 // definitions used in this block - and check to see if any of the values
204 // they define are used in locations potentially reachable from the
205 // interesting instruction.
206 BasicBlock *BBI = pred;
209 errs() << "[LSP] Looking at dominating block " << pred->getName() << "\n";
211 assert(DT.dominates(BBI, pred));
212 assert(isPotentiallyReachable(BBI, pred, &DT) &&
213 "dominated block must be reachable");
215 // Walk through the instructions in dominating blocks and keep any
216 // that have a use potentially reachable from the block we're
217 // considering putting the safepoint in
218 for (Instruction &inst : *BBI) {
220 errs() << "[LSP] Looking at instruction ";
224 if (pred == BBI && (&inst) == term) {
226 errs() << "[LSP] stopped because we encountered the safepoint "
230 // If we're in the block which defines the interesting instruction,
231 // we don't want to include any values as live which are defined
232 // _after_ the interesting line or as part of the line itself
233 // i.e. "term" is the call instruction for a call safepoint, the
234 // results of the call should not be considered live in that stackmap
238 assert(!isAggWhichContainsGCPtrType(inst.getType()) &&
239 "support for FCA unimplemented");
241 if (is_live_gc_reference(inst)) {
243 errs() << "[LSP] found live value for this safepoint ";
247 liveValues.insert(&inst);
250 if (!DT.getNode(BBI)->getIDom()) {
251 assert(BBI == &F->getEntryBlock() &&
252 "failed to find a dominator for something other than "
256 BBI = DT.getNode(BBI)->getIDom()->getBlock();
260 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
261 if (a->hasName() && b->hasName()) {
262 return -1 == a->getName().compare(b->getName());
263 } else if (a->hasName() && !b->hasName()) {
265 } else if (!a->hasName() && b->hasName()) {
268 // Better than nothing, but not stable
273 /// Find the initial live set. Note that due to base pointer
274 /// insertion, the live set may be incomplete.
276 analyzeParsePointLiveness(DominatorTree &DT, const CallSite &CS,
277 PartiallyConstructedSafepointRecord &result) {
278 Instruction *inst = CS.getInstruction();
280 BasicBlock *BB = inst->getParent();
281 std::set<Value *> liveset;
282 findLiveGCValuesAtInst(inst, BB, DT, nullptr, liveset);
285 // Note: This output is used by several of the test cases
286 // The order of elemtns in a set is not stable, put them in a vec and sort
288 std::vector<Value *> temp;
289 temp.insert(temp.end(), liveset.begin(), liveset.end());
290 std::sort(temp.begin(), temp.end(), order_by_name);
291 errs() << "Live Variables:\n";
292 for (Value *V : temp) {
293 errs() << " " << V->getName(); // no newline
297 if (PrintLiveSetSize) {
298 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
299 errs() << "Number live values: " << liveset.size() << "\n";
301 result.liveset = liveset;
304 /// True iff this value is the null pointer constant (of any pointer type)
305 static bool isNullConstant(Value *V) {
306 return isa<Constant>(V) && isa<PointerType>(V->getType()) &&
307 cast<Constant>(V)->isNullValue();
310 /// Helper function for findBasePointer - Will return a value which either a)
311 /// defines the base pointer for the input or b) blocks the simple search
312 /// (i.e. a PHI or Select of two derived pointers)
313 static Value *findBaseDefiningValue(Value *I) {
314 assert(I->getType()->isPointerTy() &&
315 "Illegal to ask for the base pointer of a non-pointer type");
317 // There are instructions which can never return gc pointer values. Sanity
319 // that this is actually true.
320 assert(!isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
321 !isa<ShuffleVectorInst>(I) && "Vector types are not gc pointers");
322 assert((!isa<Instruction>(I) || isa<InvokeInst>(I) ||
323 !cast<Instruction>(I)->isTerminator()) &&
324 "With the exception of invoke terminators don't define values");
325 assert(!isa<StoreInst>(I) && !isa<FenceInst>(I) &&
326 "Can't be definitions to start with");
327 assert(!isa<ICmpInst>(I) && !isa<FCmpInst>(I) &&
328 "Comparisons don't give ops");
329 // There's a bunch of instructions which just don't make sense to apply to
330 // a pointer. The only valid reason for this would be pointer bit
331 // twiddling which we're just not going to support.
332 assert((!isa<Instruction>(I) || !cast<Instruction>(I)->isBinaryOp()) &&
333 "Binary ops on pointer values are meaningless. Unless your "
334 "bit-twiddling which we don't support");
336 if (Argument *Arg = dyn_cast<Argument>(I)) {
337 // An incoming argument to the function is a base pointer
338 // We should have never reached here if this argument isn't an gc value
339 assert(Arg->getType()->isPointerTy() &&
340 "Base for pointer must be another pointer");
344 if (GlobalVariable *global = dyn_cast<GlobalVariable>(I)) {
346 assert(global->getType()->isPointerTy() &&
347 "Base for pointer must be another pointer");
351 // inlining could possibly introduce phi node that contains
352 // undef if callee has multiple returns
353 if (UndefValue *undef = dyn_cast<UndefValue>(I)) {
354 assert(undef->getType()->isPointerTy() &&
355 "Base for pointer must be another pointer");
356 return undef; // utterly meaningless, but useful for dealing with
357 // partially optimized code.
360 // Due to inheritance, this must be _after_ the global variable and undef
362 if (Constant *con = dyn_cast<Constant>(I)) {
363 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
364 "order of checks wrong!");
365 // Note: Finding a constant base for something marked for relocation
366 // doesn't really make sense. The most likely case is either a) some
367 // screwed up the address space usage or b) your validating against
368 // compiled C++ code w/o the proper separation. The only real exception
369 // is a null pointer. You could have generic code written to index of
370 // off a potentially null value and have proven it null. We also use
371 // null pointers in dead paths of relocation phis (which we might later
372 // want to find a base pointer for).
373 assert(con->getType()->isPointerTy() &&
374 "Base for pointer must be another pointer");
375 assert(con->isNullValue() && "null is the only case which makes sense");
379 if (CastInst *CI = dyn_cast<CastInst>(I)) {
380 Value *def = CI->stripPointerCasts();
381 assert(def->getType()->isPointerTy() &&
382 "Base for pointer must be another pointer");
383 if (isa<CastInst>(def)) {
384 // If we find a cast instruction here, it means we've found a cast
385 // which is not simply a pointer cast (i.e. an inttoptr). We don't
386 // know how to handle int->ptr conversion.
387 llvm_unreachable("Can not find the base pointers for an inttoptr cast");
389 assert(!isa<CastInst>(def) && "shouldn't find another cast here");
390 return findBaseDefiningValue(def);
393 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
394 if (LI->getType()->isPointerTy()) {
395 Value *Op = LI->getOperand(0);
397 // Has to be a pointer to an gc object, or possibly an array of such?
398 assert(Op->getType()->isPointerTy());
399 return LI; // The value loaded is an gc base itself
402 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
403 Value *Op = GEP->getOperand(0);
404 if (Op->getType()->isPointerTy()) {
405 return findBaseDefiningValue(Op); // The base of this GEP is the base
409 if (AllocaInst *alloc = dyn_cast<AllocaInst>(I)) {
410 // An alloca represents a conceptual stack slot. It's the slot itself
411 // that the GC needs to know about, not the value in the slot.
412 assert(alloc->getType()->isPointerTy() &&
413 "Base for pointer must be another pointer");
417 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
418 switch (II->getIntrinsicID()) {
420 // fall through to general call handling
422 case Intrinsic::experimental_gc_statepoint:
423 case Intrinsic::experimental_gc_result_float:
424 case Intrinsic::experimental_gc_result_int:
425 llvm_unreachable("these don't produce pointers");
426 case Intrinsic::experimental_gc_result_ptr:
427 // This is just a special case of the CallInst check below to handle a
428 // statepoint with deopt args which hasn't been rewritten for GC yet.
429 // TODO: Assert that the statepoint isn't rewritten yet.
431 case Intrinsic::experimental_gc_relocate: {
432 // Rerunning safepoint insertion after safepoints are already
433 // inserted is not supported. It could probably be made to work,
434 // but why are you doing this? There's no good reason.
435 llvm_unreachable("repeat safepoint insertion is not supported");
437 case Intrinsic::gcroot:
438 // Currently, this mechanism hasn't been extended to work with gcroot.
439 // There's no reason it couldn't be, but I haven't thought about the
440 // implications much.
442 "interaction with the gcroot mechanism is not supported");
445 // We assume that functions in the source language only return base
446 // pointers. This should probably be generalized via attributes to support
447 // both source language and internal functions.
448 if (CallInst *call = dyn_cast<CallInst>(I)) {
449 assert(call->getType()->isPointerTy() &&
450 "Base for pointer must be another pointer");
453 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
454 assert(invoke->getType()->isPointerTy() &&
455 "Base for pointer must be another pointer");
459 // I have absolutely no idea how to implement this part yet. It's not
460 // neccessarily hard, I just haven't really looked at it yet.
461 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
463 if (AtomicCmpXchgInst *cas = dyn_cast<AtomicCmpXchgInst>(I)) {
464 // A CAS is effectively a atomic store and load combined under a
465 // predicate. From the perspective of base pointers, we just treat it
466 // like a load. We loaded a pointer from a address in memory, that value
467 // had better be a valid base pointer.
468 return cas->getPointerOperand();
470 if (AtomicRMWInst *atomic = dyn_cast<AtomicRMWInst>(I)) {
471 assert(AtomicRMWInst::Xchg == atomic->getOperation() &&
472 "All others are binary ops which don't apply to base pointers");
473 // semantically, a load, store pair. Treat it the same as a standard load
474 return atomic->getPointerOperand();
477 // The aggregate ops. Aggregates can either be in the heap or on the
478 // stack, but in either case, this is simply a field load. As a result,
479 // this is a defining definition of the base just like a load is.
480 if (ExtractValueInst *ev = dyn_cast<ExtractValueInst>(I)) {
484 // We should never see an insert vector since that would require we be
485 // tracing back a struct value not a pointer value.
486 assert(!isa<InsertValueInst>(I) &&
487 "Base pointer for a struct is meaningless");
489 // The last two cases here don't return a base pointer. Instead, they
490 // return a value which dynamically selects from amoung several base
491 // derived pointers (each with it's own base potentially). It's the job of
492 // the caller to resolve these.
493 if (SelectInst *select = dyn_cast<SelectInst>(I)) {
496 if (PHINode *phi = dyn_cast<PHINode>(I)) {
500 errs() << "unknown type: " << *I << "\n";
501 llvm_unreachable("unknown type");
505 /// Returns the base defining value for this value.
506 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &cache) {
507 Value *&Cached = cache[I];
509 Cached = findBaseDefiningValue(I);
511 assert(cache[I] != nullptr);
514 errs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
520 /// Return a base pointer for this value if known. Otherwise, return it's
521 /// base defining value.
522 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &cache) {
523 Value *def = findBaseDefiningValueCached(I, cache);
524 auto Found = cache.find(def);
525 if (Found != cache.end()) {
526 // Either a base-of relation, or a self reference. Caller must check.
527 return Found->second;
529 // Only a BDV available
533 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
534 /// is it known to be a base pointer? Or do we need to continue searching.
535 static bool isKnownBaseResult(Value *v) {
536 if (!isa<PHINode>(v) && !isa<SelectInst>(v)) {
537 // no recursion possible
540 if (cast<Instruction>(v)->getMetadata("is_base_value")) {
541 // This is a previously inserted base phi or select. We know
542 // that this is a base value.
546 // We need to keep searching
550 // TODO: find a better name for this
554 enum Status { Unknown, Base, Conflict };
556 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
557 assert(status != Base || b);
559 PhiState(Value *b) : status(Base), base(b) {}
560 PhiState() : status(Unknown), base(nullptr) {}
561 PhiState(const PhiState &other) : status(other.status), base(other.base) {
562 assert(status != Base || base);
565 Status getStatus() const { return status; }
566 Value *getBase() const { return base; }
568 bool isBase() const { return getStatus() == Base; }
569 bool isUnknown() const { return getStatus() == Unknown; }
570 bool isConflict() const { return getStatus() == Conflict; }
572 bool operator==(const PhiState &other) const {
573 return base == other.base && status == other.status;
576 bool operator!=(const PhiState &other) const { return !(*this == other); }
579 errs() << status << " (" << base << " - "
580 << (base ? base->getName() : "nullptr") << "): ";
585 Value *base; // non null only if status == base
588 // Values of type PhiState form a lattice, and this is a helper
589 // class that implementes the meet operation. The meat of the meet
590 // operation is implemented in MeetPhiStates::pureMeet
591 class MeetPhiStates {
593 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
594 explicit MeetPhiStates(const std::map<Value *, PhiState> &phiStates)
595 : phiStates(phiStates) {}
597 // Destructively meet the current result with the base V. V can
598 // either be a merge instruction (SelectInst / PHINode), in which
599 // case its status is looked up in the phiStates map; or a regular
600 // SSA value, in which case it is assumed to be a base.
601 void meetWith(Value *V) {
602 PhiState otherState = getStateForBDV(V);
603 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
604 MeetPhiStates::pureMeet(currentResult, otherState)) &&
605 "math is wrong: meet does not commute!");
606 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
609 PhiState getResult() const { return currentResult; }
612 const std::map<Value *, PhiState> &phiStates;
613 PhiState currentResult;
615 /// Return a phi state for a base defining value. We'll generate a new
616 /// base state for known bases and expect to find a cached state otherwise
617 PhiState getStateForBDV(Value *baseValue) {
618 if (isKnownBaseResult(baseValue)) {
619 return PhiState(baseValue);
621 return lookupFromMap(baseValue);
625 PhiState lookupFromMap(Value *V) {
626 auto I = phiStates.find(V);
627 assert(I != phiStates.end() && "lookup failed!");
631 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
632 switch (stateA.getStatus()) {
633 case PhiState::Unknown:
637 assert(stateA.getBase() && "can't be null");
638 if (stateB.isUnknown()) {
640 } else if (stateB.isBase()) {
641 if (stateA.getBase() == stateB.getBase()) {
642 assert(stateA == stateB && "equality broken!");
645 return PhiState(PhiState::Conflict);
647 assert(stateB.isConflict() && "only three states!");
648 return PhiState(PhiState::Conflict);
651 case PhiState::Conflict:
654 assert(false && "only three states!");
658 /// For a given value or instruction, figure out what base ptr it's derived
659 /// from. For gc objects, this is simply itself. On success, returns a value
660 /// which is the base pointer. (This is reliable and can be used for
661 /// relocation.) On failure, returns nullptr.
662 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache,
663 std::set<llvm::Value *> &newInsertedDefs) {
664 Value *def = findBaseOrBDV(I, cache);
666 if (isKnownBaseResult(def)) {
670 // Here's the rough algorithm:
671 // - For every SSA value, construct a mapping to either an actual base
672 // pointer or a PHI which obscures the base pointer.
673 // - Construct a mapping from PHI to unknown TOP state. Use an
674 // optimistic algorithm to propagate base pointer information. Lattice
679 // When algorithm terminates, all PHIs will either have a single concrete
680 // base or be in a conflict state.
681 // - For every conflict, insert a dummy PHI node without arguments. Add
682 // these to the base[Instruction] = BasePtr mapping. For every
683 // non-conflict, add the actual base.
684 // - For every conflict, add arguments for the base[a] of each input
687 // Note: A simpler form of this would be to add the conflict form of all
688 // PHIs without running the optimistic algorithm. This would be
689 // analougous to pessimistic data flow and would likely lead to an
690 // overall worse solution.
692 std::map<Value *, PhiState> states;
693 states[def] = PhiState();
694 // Recursively fill in all phis & selects reachable from the initial one
695 // for which we don't already know a definite base value for
696 // PERF: Yes, this is as horribly inefficient as it looks.
700 for (auto Pair : states) {
701 Value *v = Pair.first;
702 assert(!isKnownBaseResult(v) && "why did it get added?");
703 if (PHINode *phi = dyn_cast<PHINode>(v)) {
704 unsigned NumPHIValues = phi->getNumIncomingValues();
705 assert(NumPHIValues > 0 && "zero input phis are illegal");
706 for (unsigned i = 0; i != NumPHIValues; ++i) {
707 Value *InVal = phi->getIncomingValue(i);
708 Value *local = findBaseOrBDV(InVal, cache);
709 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
710 states[local] = PhiState();
714 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
715 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
716 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
717 states[local] = PhiState();
720 local = findBaseOrBDV(sel->getFalseValue(), cache);
721 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
722 states[local] = PhiState();
730 errs() << "States after initialization:\n";
731 for (auto Pair : states) {
732 Instruction *v = cast<Instruction>(Pair.first);
733 PhiState state = Pair.second;
739 // TODO: come back and revisit the state transitions around inputs which
740 // have reached conflict state. The current version seems too conservative.
742 bool progress = true;
745 oldSize = states.size();
747 for (auto Pair : states) {
748 MeetPhiStates calculateMeet(states);
749 Value *v = Pair.first;
750 assert(!isKnownBaseResult(v) && "why did it get added?");
751 assert(isa<SelectInst>(v) || isa<PHINode>(v));
752 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
753 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
754 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
755 } else if (PHINode *phi = dyn_cast<PHINode>(v)) {
756 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
757 calculateMeet.meetWith(
758 findBaseOrBDV(phi->getIncomingValue(i), cache));
761 llvm_unreachable("no such state expected");
764 PhiState oldState = states[v];
765 PhiState newState = calculateMeet.getResult();
766 if (oldState != newState) {
768 states[v] = newState;
772 assert(oldSize <= states.size());
773 assert(oldSize == states.size() || progress);
777 errs() << "States after meet iteration:\n";
778 for (auto Pair : states) {
779 Instruction *v = cast<Instruction>(Pair.first);
780 PhiState state = Pair.second;
786 // Insert Phis for all conflicts
787 for (auto Pair : states) {
788 Instruction *v = cast<Instruction>(Pair.first);
789 PhiState state = Pair.second;
790 assert(!isKnownBaseResult(v) && "why did it get added?");
791 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
792 if (state.isConflict()) {
793 if (isa<PHINode>(v)) {
795 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
796 assert(num_preds > 0 && "how did we reach here");
797 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
798 newInsertedDefs.insert(phi);
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 phi->setMetadata("is_base_value", md);
808 states[v] = PhiState(PhiState::Conflict, phi);
809 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
810 // The undef will be replaced later
811 UndefValue *undef = UndefValue::get(sel->getType());
812 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
813 undef, "base_select", sel);
814 newInsertedDefs.insert(basesel);
815 // Add metadata marking this as a base value
816 auto *const_1 = ConstantInt::get(
818 v->getParent()->getParent()->getParent()->getContext()),
820 auto MDConst = ConstantAsMetadata::get(const_1);
821 MDNode *md = MDNode::get(
822 v->getParent()->getParent()->getParent()->getContext(), MDConst);
823 basesel->setMetadata("is_base_value", md);
824 states[v] = PhiState(PhiState::Conflict, basesel);
831 // Fixup all the inputs of the new PHIs
832 for (auto Pair : states) {
833 Instruction *v = cast<Instruction>(Pair.first);
834 PhiState state = Pair.second;
836 assert(!isKnownBaseResult(v) && "why did it get added?");
837 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
838 if (state.isConflict()) {
839 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
840 PHINode *phi = cast<PHINode>(v);
841 unsigned NumPHIValues = phi->getNumIncomingValues();
842 for (unsigned i = 0; i < NumPHIValues; i++) {
843 Value *InVal = phi->getIncomingValue(i);
844 BasicBlock *InBB = phi->getIncomingBlock(i);
846 // If we've already seen InBB, add the same incoming value
847 // we added for it earlier. The IR verifier requires phi
848 // nodes with multiple entries from the same basic block
849 // to have the same incoming value for each of those
850 // entries. If we don't do this check here and basephi
851 // has a different type than base, we'll end up adding two
852 // bitcasts (and hence two distinct values) as incoming
853 // values for the same basic block.
855 int blockIndex = basephi->getBasicBlockIndex(InBB);
856 if (blockIndex != -1) {
857 Value *oldBase = basephi->getIncomingValue(blockIndex);
858 basephi->addIncoming(oldBase, InBB);
860 Value *base = findBaseOrBDV(InVal, cache);
861 if (!isKnownBaseResult(base)) {
862 // Either conflict or base.
863 assert(states.count(base));
864 base = states[base].getBase();
865 assert(base != nullptr && "unknown PhiState!");
866 assert(newInsertedDefs.count(base) &&
867 "should have already added this in a prev. iteration!");
870 // In essense this assert states: the only way two
871 // values incoming from the same basic block may be
872 // different is by being different bitcasts of the same
873 // value. A cleanup that remains TODO is changing
874 // findBaseOrBDV to return an llvm::Value of the correct
875 // type (and still remain pure). This will remove the
876 // need to add bitcasts.
877 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
878 "sanity -- findBaseOrBDV should be pure!");
883 // Find either the defining value for the PHI or the normal base for
885 Value *base = findBaseOrBDV(InVal, cache);
886 if (!isKnownBaseResult(base)) {
887 // Either conflict or base.
888 assert(states.count(base));
889 base = states[base].getBase();
890 assert(base != nullptr && "unknown PhiState!");
892 assert(base && "can't be null");
893 // Must use original input BB since base may not be Instruction
894 // The cast is needed since base traversal may strip away bitcasts
895 if (base->getType() != basephi->getType()) {
896 base = new BitCastInst(base, basephi->getType(), "cast",
897 InBB->getTerminator());
898 newInsertedDefs.insert(base);
900 basephi->addIncoming(base, InBB);
902 assert(basephi->getNumIncomingValues() == NumPHIValues);
903 } else if (SelectInst *basesel = dyn_cast<SelectInst>(state.getBase())) {
904 SelectInst *sel = cast<SelectInst>(v);
905 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
906 // something more safe and less hacky.
907 for (int i = 1; i <= 2; i++) {
908 Value *InVal = sel->getOperand(i);
909 // Find either the defining value for the PHI or the normal base for
911 Value *base = findBaseOrBDV(InVal, cache);
912 if (!isKnownBaseResult(base)) {
913 // Either conflict or base.
914 assert(states.count(base));
915 base = states[base].getBase();
916 assert(base != nullptr && "unknown PhiState!");
918 assert(base && "can't be null");
919 // Must use original input BB since base may not be Instruction
920 // The cast is needed since base traversal may strip away bitcasts
921 if (base->getType() != basesel->getType()) {
922 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
923 newInsertedDefs.insert(base);
925 basesel->setOperand(i, base);
928 assert(false && "unexpected type");
933 // Cache all of our results so we can cheaply reuse them
934 // NOTE: This is actually two caches: one of the base defining value
935 // relation and one of the base pointer relation! FIXME
936 for (auto item : states) {
937 Value *v = item.first;
938 Value *base = item.second.getBase();
940 assert(!isKnownBaseResult(v) && "why did it get added?");
943 std::string fromstr =
944 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
946 errs() << "Updating base value cache"
947 << " for: " << (v->hasName() ? v->getName() : "")
948 << " from: " << fromstr
949 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
952 assert(isKnownBaseResult(base) &&
953 "must be something we 'know' is a base pointer");
954 if (cache.count(v)) {
955 // Once we transition from the BDV relation being store in the cache to
956 // the base relation being stored, it must be stable
957 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
958 "base relation should be stable");
962 assert(cache.find(def) != cache.end());
966 // For a set of live pointers (base and/or derived), identify the base
967 // pointer of the object which they are derived from. This routine will
968 // mutate the IR graph as needed to make the 'base' pointer live at the
969 // definition site of 'derived'. This ensures that any use of 'derived' can
970 // also use 'base'. This may involve the insertion of a number of
971 // additional PHI nodes.
973 // preconditions: live is a set of pointer type Values
975 // side effects: may insert PHI nodes into the existing CFG, will preserve
976 // CFG, will not remove or mutate any existing nodes
978 // post condition: base_pairs contains one (derived, base) pair for every
979 // pointer in live. Note that derived can be equal to base if the original
980 // pointer was a base pointer.
981 static void findBasePointers(const std::set<llvm::Value *> &live,
982 std::map<llvm::Value *, llvm::Value *> &base_pairs,
983 DominatorTree *DT, DefiningValueMapTy &DVCache,
984 std::set<llvm::Value *> &newInsertedDefs) {
985 for (Value *ptr : live) {
986 Value *base = findBasePointer(ptr, DVCache, newInsertedDefs);
987 assert(base && "failed to find base pointer");
988 base_pairs[ptr] = base;
989 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
990 DT->dominates(cast<Instruction>(base)->getParent(),
991 cast<Instruction>(ptr)->getParent())) &&
992 "The base we found better dominate the derived pointer");
994 if (isNullConstant(base))
995 // If you see this trip and like to live really dangerously, the code
996 // should be correct, just with idioms the verifier can't handle. You
997 // can try disabling the verifier at your own substaintial risk.
998 llvm_unreachable("the relocation code needs adjustment to handle the"
999 "relocation of a null pointer constant without causing"
1000 "false positives in the safepoint ir verifier.");
1004 /// Find the required based pointers (and adjust the live set) for the given
1006 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1008 PartiallyConstructedSafepointRecord &result) {
1009 std::map<llvm::Value *, llvm::Value *> base_pairs;
1010 std::set<llvm::Value *> newInsertedDefs;
1011 findBasePointers(result.liveset, base_pairs, &DT, DVCache, newInsertedDefs);
1013 if (PrintBasePointers) {
1014 errs() << "Base Pairs (w/o Relocation):\n";
1015 for (auto Pair : base_pairs) {
1016 errs() << " derived %" << Pair.first->getName() << " base %"
1017 << Pair.second->getName() << "\n";
1021 result.base_pairs = base_pairs;
1022 result.newInsertedDefs = newInsertedDefs;
1025 /// Check for liveness of items in the insert defs and add them to the live
1026 /// and base pointer sets
1027 static void fixupLiveness(DominatorTree &DT, const CallSite &CS,
1028 const std::set<Value *> &allInsertedDefs,
1029 PartiallyConstructedSafepointRecord &result) {
1030 Instruction *inst = CS.getInstruction();
1032 std::set<llvm::Value *> liveset = result.liveset;
1033 std::map<llvm::Value *, llvm::Value *> base_pairs = result.base_pairs;
1035 auto is_live_gc_reference =
1036 [&](Value &V) { return isLiveGCReferenceAt(V, inst, DT, nullptr); };
1038 // For each new definition, check to see if a) the definition dominates the
1039 // instruction we're interested in, and b) one of the uses of that definition
1040 // is edge-reachable from the instruction we're interested in. This is the
1041 // same definition of liveness we used in the intial liveness analysis
1042 for (Value *newDef : allInsertedDefs) {
1043 if (liveset.count(newDef)) {
1044 // already live, no action needed
1048 // PERF: Use DT to check instruction domination might not be good for
1049 // compilation time, and we could change to optimal solution if this
1050 // turn to be a issue
1051 if (!DT.dominates(cast<Instruction>(newDef), inst)) {
1052 // can't possibly be live at inst
1056 if (is_live_gc_reference(*newDef)) {
1057 // Add the live new defs into liveset and base_pairs
1058 liveset.insert(newDef);
1059 base_pairs[newDef] = newDef;
1063 result.liveset = liveset;
1064 result.base_pairs = base_pairs;
1067 static void fixupLiveReferences(
1068 Function &F, DominatorTree &DT, Pass *P,
1069 const std::set<llvm::Value *> &allInsertedDefs,
1070 std::vector<CallSite> &toUpdate,
1071 std::vector<struct PartiallyConstructedSafepointRecord> &records) {
1072 for (size_t i = 0; i < records.size(); i++) {
1073 struct PartiallyConstructedSafepointRecord &info = records[i];
1074 CallSite &CS = toUpdate[i];
1075 fixupLiveness(DT, CS, allInsertedDefs, info);
1079 // Normalize basic block to make it ready to be target of invoke statepoint.
1080 // It means spliting it to have single predecessor. Return newly created BB
1081 // ready to be successor of invoke statepoint.
1082 static BasicBlock *normalizeBBForInvokeSafepoint(BasicBlock *BB,
1083 BasicBlock *InvokeParent,
1085 BasicBlock *ret = BB;
1087 if (!BB->getUniquePredecessor()) {
1088 ret = SplitBlockPredecessors(BB, InvokeParent, "");
1091 // Another requirement for such basic blocks is to not have any phi nodes.
1092 // Since we just ensured that new BB will have single predecessor,
1093 // all phi nodes in it will have one value. Here it would be naturall place
1095 // remove them all. But we can not do this because we are risking to remove
1096 // one of the values stored in liveset of another statepoint. We will do it
1097 // later after placing all safepoints.
1103 VerifySafepointBounds(const std::pair<Instruction *, Instruction *> &bounds) {
1104 assert(bounds.first->getParent() && bounds.second->getParent() &&
1105 "both must belong to basic blocks");
1106 if (bounds.first->getParent() == bounds.second->getParent()) {
1107 // This is a call safepoint
1108 // TODO: scan the range to find the statepoint
1109 // TODO: check that the following instruction is not a gc_relocate or
1112 // This is an invoke safepoint
1113 InvokeInst *invoke = dyn_cast<InvokeInst>(bounds.first);
1115 assert(invoke && "only continues over invokes!");
1116 assert(invoke->getNormalDest() == bounds.second->getParent() &&
1117 "safepoint should continue into normal exit block");
1121 static int find_index(const SmallVectorImpl<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 /// Returns array with newly created relocates.
1172 static std::vector<llvm::Instruction *>
1173 CreateGCRelocates(const SmallVectorImpl<llvm::Value *> &liveVariables,
1174 const int liveStart,
1175 const SmallVectorImpl<llvm::Value *> &basePtrs,
1176 Instruction *statepointToken, IRBuilder<> Builder) {
1178 std::vector<llvm::Instruction *> newDefs;
1180 Module *M = statepointToken->getParent()->getParent()->getParent();
1182 for (unsigned i = 0; i < liveVariables.size(); i++) {
1183 // We generate a (potentially) unique declaration for every pointer type
1184 // combination. This results is some blow up the function declarations in
1185 // the IR, but removes the need for argument bitcasts which shrinks the IR
1186 // greatly and makes it much more readable.
1187 std::vector<Type *> types; // one per 'any' type
1188 types.push_back(liveVariables[i]->getType()); // result type
1189 Value *gc_relocate_decl = Intrinsic::getDeclaration(
1190 M, Intrinsic::experimental_gc_relocate, types);
1192 // Generate the gc.relocate call and save the result
1194 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1195 liveStart + find_index(liveVariables, basePtrs[i]));
1196 Value *liveIdx = ConstantInt::get(
1197 Type::getInt32Ty(M->getContext()),
1198 liveStart + find_index(liveVariables, liveVariables[i]));
1200 // only specify a debug name if we can give a useful one
1201 Value *reloc = Builder.CreateCall3(
1202 gc_relocate_decl, statepointToken, baseIdx, liveIdx,
1203 liveVariables[i]->hasName() ? liveVariables[i]->getName() + ".relocated"
1205 // Trick CodeGen into thinking there are lots of free registers at this
1207 cast<CallInst>(reloc)->setCallingConv(CallingConv::Cold);
1209 newDefs.push_back(cast<Instruction>(reloc));
1211 assert(newDefs.size() == liveVariables.size() &&
1212 "missing or extra redefinition at safepoint");
1218 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1219 const SmallVectorImpl<llvm::Value *> &basePtrs,
1220 const SmallVectorImpl<llvm::Value *> &liveVariables,
1222 PartiallyConstructedSafepointRecord &result) {
1223 assert(basePtrs.size() == liveVariables.size());
1224 assert(isStatepoint(CS) &&
1225 "This method expects to be rewriting a statepoint");
1227 BasicBlock *BB = CS.getInstruction()->getParent();
1229 Function *F = BB->getParent();
1230 assert(F && "must be set");
1231 Module *M = F->getParent();
1233 assert(M && "must be set");
1235 // We're not changing the function signature of the statepoint since the gc
1236 // arguments go into the var args section.
1237 Function *gc_statepoint_decl = CS.getCalledFunction();
1239 // Then go ahead and use the builder do actually do the inserts. We insert
1240 // immediately before the previous instruction under the assumption that all
1241 // arguments will be available here. We can't insert afterwards since we may
1242 // be replacing a terminator.
1243 Instruction *insertBefore = CS.getInstruction();
1244 IRBuilder<> Builder(insertBefore);
1245 // Copy all of the arguments from the original statepoint - this includes the
1246 // target, call args, and deopt args
1247 std::vector<llvm::Value *> args;
1248 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1249 // TODO: Clear the 'needs rewrite' flag
1251 // add all the pointers to be relocated (gc arguments)
1252 // Capture the start of the live variable list for use in the gc_relocates
1253 const int live_start = args.size();
1254 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1256 // Create the statepoint given all the arguments
1257 Instruction *token = nullptr;
1258 AttributeSet return_attributes;
1260 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1262 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1263 call->setTailCall(toReplace->isTailCall());
1264 call->setCallingConv(toReplace->getCallingConv());
1266 // Currently we will fail on parameter attributes and on certain
1267 // function attributes.
1268 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1269 // In case if we can handle this set of sttributes - set up function attrs
1270 // directly on statepoint and return attrs later for gc_result intrinsic.
1271 call->setAttributes(new_attrs.getFnAttributes());
1272 return_attributes = new_attrs.getRetAttributes();
1276 // Put the following gc_result and gc_relocate calls immediately after the
1277 // the old call (which we're about to delete)
1278 BasicBlock::iterator next(toReplace);
1279 assert(BB->end() != next && "not a terminator, must have next");
1281 Instruction *IP = &*(next);
1282 Builder.SetInsertPoint(IP);
1283 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1285 } else if (CS.isInvoke()) {
1286 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1288 // Insert the new invoke into the old block. We'll remove the old one in a
1289 // moment at which point this will become the new terminator for the
1291 InvokeInst *invoke = InvokeInst::Create(
1292 gc_statepoint_decl, toReplace->getNormalDest(),
1293 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1294 invoke->setCallingConv(toReplace->getCallingConv());
1296 // Currently we will fail on parameter attributes and on certain
1297 // function attributes.
1298 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1299 // In case if we can handle this set of sttributes - set up function attrs
1300 // directly on statepoint and return attrs later for gc_result intrinsic.
1301 invoke->setAttributes(new_attrs.getFnAttributes());
1302 return_attributes = new_attrs.getRetAttributes();
1306 // Generate gc relocates in exceptional path
1307 BasicBlock *unwindBlock = normalizeBBForInvokeSafepoint(
1308 toReplace->getUnwindDest(), invoke->getParent(), P);
1310 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1311 Builder.SetInsertPoint(IP);
1312 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1314 // Extract second element from landingpad return value. We will attach
1315 // exceptional gc relocates to it.
1316 const unsigned idx = 1;
1317 Instruction *exceptional_token =
1318 cast<Instruction>(Builder.CreateExtractValue(
1319 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1320 result.exceptional_relocates_token = exceptional_token;
1322 // Just throw away return value. We will use the one we got for normal
1324 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1325 exceptional_token, Builder);
1327 // Generate gc relocates and returns for normal block
1328 BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
1329 toReplace->getNormalDest(), invoke->getParent(), P);
1331 IP = &*(normalDest->getFirstInsertionPt());
1332 Builder.SetInsertPoint(IP);
1334 // gc relocates will be generated later as if it were regular call
1337 llvm_unreachable("unexpect type of CallSite");
1341 // Take the name of the original value call if it had one.
1342 token->takeName(CS.getInstruction());
1344 // The GCResult is already inserted, we just need to find it
1345 Instruction *gc_result = nullptr;
1347 Instruction *toReplace = CS.getInstruction();
1348 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1349 "only valid use before rewrite is gc.result");
1350 if (toReplace->hasOneUse()) {
1351 Instruction *GCResult = cast<Instruction>(*toReplace->user_begin());
1352 assert(isGCResult(GCResult));
1353 gc_result = GCResult;
1357 // Update the gc.result of the original statepoint (if any) to use the newly
1358 // inserted statepoint. This is safe to do here since the token can't be
1359 // considered a live reference.
1360 CS.getInstruction()->replaceAllUsesWith(token);
1362 // Second, create a gc.relocate for every live variable
1363 std::vector<llvm::Instruction *> newDefs =
1364 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1366 // Need to pass through the last part of the safepoint block so that we
1367 // don't accidentally update uses in a following gc.relocate which is
1368 // still conceptually part of the same safepoint. Gah.
1369 Instruction *last = nullptr;
1370 if (!newDefs.empty()) {
1371 last = newDefs.back();
1372 } else if (gc_result) {
1377 assert(last && "can't be null");
1378 const auto bounds = std::make_pair(token, last);
1380 // Sanity check our results - this is slightly non-trivial due to invokes
1381 VerifySafepointBounds(bounds);
1383 result.safepoint = bounds;
1387 struct name_ordering {
1390 bool operator()(name_ordering const &a, name_ordering const &b) {
1391 return -1 == a.derived->getName().compare(b.derived->getName());
1395 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1396 SmallVectorImpl<Value *> &livevec) {
1397 assert(basevec.size() == livevec.size());
1399 std::vector<name_ordering> temp;
1400 for (size_t i = 0; i < basevec.size(); i++) {
1402 v.base = basevec[i];
1403 v.derived = livevec[i];
1406 std::sort(temp.begin(), temp.end(), name_ordering());
1407 for (size_t i = 0; i < basevec.size(); i++) {
1408 basevec[i] = temp[i].base;
1409 livevec[i] = temp[i].derived;
1413 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1414 // which make the relocations happening at this safepoint explicit.
1416 // WARNING: Does not do any fixup to adjust users of the original live
1417 // values. That's the callers responsibility.
1419 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1420 PartiallyConstructedSafepointRecord &result) {
1421 std::set<llvm::Value *> liveset = result.liveset;
1422 std::map<llvm::Value *, llvm::Value *> base_pairs = result.base_pairs;
1424 // Convert to vector for efficient cross referencing.
1425 SmallVector<Value *, 64> basevec, livevec;
1426 livevec.reserve(liveset.size());
1427 basevec.reserve(liveset.size());
1428 for (Value *L : liveset) {
1429 livevec.push_back(L);
1431 assert(base_pairs.find(L) != base_pairs.end());
1432 Value *base = base_pairs[L];
1433 basevec.push_back(base);
1435 assert(livevec.size() == basevec.size());
1437 // To make the output IR slightly more stable (for use in diffs), ensure a
1438 // fixed order of the values in the safepoint (by sorting the value name).
1439 // The order is otherwise meaningless.
1440 stablize_order(basevec, livevec);
1442 // Do the actual rewriting and delete the old statepoint
1443 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1444 CS.getInstruction()->eraseFromParent();
1447 // Helper function for the relocationViaAlloca.
1448 // It receives iterator to the statepoint gc relocates and emits store to the
1450 // location (via allocaMap) for the each one of them.
1451 // Add visited values into the visitedLiveValues set we will later use them
1452 // for sanity check.
1454 insertRelocationStores(iterator_range<Value::user_iterator> gcRelocs,
1455 DenseMap<Value *, Value *> &allocaMap,
1456 DenseSet<Value *> &visitedLiveValues) {
1458 for (User *U : gcRelocs) {
1459 if (!isa<IntrinsicInst>(U))
1462 IntrinsicInst *relocatedValue = cast<IntrinsicInst>(U);
1464 // We only care about relocates
1465 if (relocatedValue->getIntrinsicID() !=
1466 Intrinsic::experimental_gc_relocate) {
1470 GCRelocateOperands relocateOperands(relocatedValue);
1471 Value *originalValue = const_cast<Value *>(relocateOperands.derivedPtr());
1472 assert(allocaMap.count(originalValue));
1473 Value *alloca = allocaMap[originalValue];
1475 // Emit store into the related alloca
1476 StoreInst *store = new StoreInst(relocatedValue, alloca);
1477 store->insertAfter(relocatedValue);
1480 visitedLiveValues.insert(originalValue);
1485 /// do all the relocation update via allocas and mem2reg
1486 static void relocationViaAlloca(
1487 Function &F, DominatorTree &DT, const std::vector<Value *> &live,
1488 const std::vector<struct PartiallyConstructedSafepointRecord> &records) {
1490 int initialAllocaNum = 0;
1492 // record initial number of allocas
1493 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1495 if (isa<AllocaInst>(*itr))
1500 // TODO-PERF: change data structures, reserve
1501 DenseMap<Value *, Value *> allocaMap;
1502 SmallVector<AllocaInst *, 200> PromotableAllocas;
1503 PromotableAllocas.reserve(live.size());
1505 // emit alloca for each live gc pointer
1506 for (unsigned i = 0; i < live.size(); i++) {
1507 Value *liveValue = live[i];
1508 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1509 F.getEntryBlock().getFirstNonPHI());
1510 allocaMap[liveValue] = alloca;
1511 PromotableAllocas.push_back(alloca);
1514 // The next two loops are part of the same conceptual operation. We need to
1515 // insert a store to the alloca after the original def and at each
1516 // redefinition. We need to insert a load before each use. These are split
1517 // into distinct loops for performance reasons.
1519 // update gc pointer after each statepoint
1520 // either store a relocated value or null (if no relocated value found for
1521 // this gc pointer and it is not a gc_result)
1522 // this must happen before we update the statepoint with load of alloca
1523 // otherwise we lose the link between statepoint and old def
1524 for (size_t i = 0; i < records.size(); i++) {
1525 const struct PartiallyConstructedSafepointRecord &info = records[i];
1526 Value *statepoint = info.safepoint.first;
1528 // This will be used for consistency check
1529 DenseSet<Value *> visitedLiveValues;
1531 // Insert stores for normal statepoint gc relocates
1532 insertRelocationStores(statepoint->users(), allocaMap, visitedLiveValues);
1534 // In case if it was invoke statepoint
1535 // we will insert stores for exceptional path gc relocates.
1536 if (isa<InvokeInst>(statepoint)) {
1537 insertRelocationStores(info.exceptional_relocates_token->users(),
1538 allocaMap, visitedLiveValues);
1542 // For consistency check store null's into allocas for values that are not
1544 // by this statepoint.
1545 for (auto Pair : allocaMap) {
1546 Value *def = Pair.first;
1547 Value *alloca = Pair.second;
1549 // This value was relocated
1550 if (visitedLiveValues.count(def)) {
1553 // Result should not be relocated
1554 if (def == info.result) {
1559 ConstantPointerNull::get(cast<PointerType>(def->getType()));
1560 StoreInst *store = new StoreInst(CPN, alloca);
1561 store->insertBefore(info.safepoint.second);
1565 // update use with load allocas and add store for gc_relocated
1566 for (auto Pair : allocaMap) {
1567 Value *def = Pair.first;
1568 Value *alloca = Pair.second;
1570 // we pre-record the uses of allocas so that we dont have to worry about
1572 // that change the user information.
1573 SmallVector<Instruction *, 20> uses;
1574 // PERF: trade a linear scan for repeated reallocation
1575 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1576 for (User *U : def->users()) {
1577 if (!isa<ConstantExpr>(U)) {
1578 // If the def has a ConstantExpr use, then the def is either a
1579 // ConstantExpr use itself or null. In either case
1580 // (recursively in the first, directly in the second), the oop
1581 // it is ultimately dependent on is null and this particular
1582 // use does not need to be fixed up.
1583 uses.push_back(cast<Instruction>(U));
1587 std::sort(uses.begin(), uses.end());
1588 auto last = std::unique(uses.begin(), uses.end());
1589 uses.erase(last, uses.end());
1591 for (Instruction *use : uses) {
1592 if (isa<PHINode>(use)) {
1593 PHINode *phi = cast<PHINode>(use);
1594 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1595 if (def == phi->getIncomingValue(i)) {
1596 LoadInst *load = new LoadInst(
1597 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1598 phi->setIncomingValue(i, load);
1602 LoadInst *load = new LoadInst(alloca, "", use);
1603 use->replaceUsesOfWith(def, load);
1607 // emit store for the initial gc value
1608 // store must be inserted after load, otherwise store will be in alloca's
1609 // use list and an extra load will be inserted before it
1610 StoreInst *store = new StoreInst(def, alloca);
1611 if (isa<Instruction>(def)) {
1612 store->insertAfter(cast<Instruction>(def));
1614 assert((isa<Argument>(def) || isa<GlobalVariable>(def) ||
1615 (isa<Constant>(def) && cast<Constant>(def)->isNullValue())) &&
1616 "Must be argument or global");
1617 store->insertAfter(cast<Instruction>(alloca));
1621 assert(PromotableAllocas.size() == live.size() &&
1622 "we must have the same allocas with lives");
1623 if (!PromotableAllocas.empty()) {
1624 // apply mem2reg to promote alloca to SSA
1625 PromoteMemToReg(PromotableAllocas, DT);
1629 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1631 if (isa<AllocaInst>(*itr))
1634 assert(initialAllocaNum == 0 && "We must not introduce any extra allocas");
1638 /// Implement a unique function which doesn't require we sort the input
1639 /// vector. Doing so has the effect of changing the output of a couple of
1640 /// tests in ways which make them less useful in testing fused safepoints.
1641 template <typename T> static void unique_unsorted(std::vector<T> &vec) {
1644 vec.reserve(vec.size());
1645 std::swap(tmp, vec);
1646 for (auto V : tmp) {
1647 if (seen.insert(V).second) {
1653 static Function *getUseHolder(Module &M) {
1654 FunctionType *ftype =
1655 FunctionType::get(Type::getVoidTy(M.getContext()), true);
1656 Function *Func = cast<Function>(M.getOrInsertFunction("__tmp_use", ftype));
1660 /// Insert holders so that each Value is obviously live through the entire
1661 /// liftetime of the call.
1662 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1663 std::vector<CallInst *> &holders) {
1664 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1665 Function *Func = getUseHolder(*M);
1667 // For call safepoints insert dummy calls right after safepoint
1668 BasicBlock::iterator next(CS.getInstruction());
1670 CallInst *base_holder = CallInst::Create(Func, Values, "", next);
1671 holders.push_back(base_holder);
1672 } else if (CS.isInvoke()) {
1673 // For invoke safepooints insert dummy calls both in normal and
1674 // exceptional destination blocks
1675 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1676 CallInst *normal_holder = CallInst::Create(
1677 Func, Values, "", invoke->getNormalDest()->getFirstInsertionPt());
1678 CallInst *unwind_holder = CallInst::Create(
1679 Func, Values, "", invoke->getUnwindDest()->getFirstInsertionPt());
1680 holders.push_back(normal_holder);
1681 holders.push_back(unwind_holder);
1683 assert(false && "Unsupported");
1687 static void findLiveReferences(
1688 Function &F, DominatorTree &DT, Pass *P, std::vector<CallSite> &toUpdate,
1689 std::vector<struct PartiallyConstructedSafepointRecord> &records) {
1690 for (size_t i = 0; i < records.size(); i++) {
1691 struct PartiallyConstructedSafepointRecord &info = records[i];
1692 CallSite &CS = toUpdate[i];
1693 analyzeParsePointLiveness(DT, CS, info);
1697 static void addBasesAsLiveValues(std::set<Value *> &liveset,
1698 std::map<Value *, Value *> &base_pairs) {
1699 // Identify any base pointers which are used in this safepoint, but not
1700 // themselves relocated. We need to relocate them so that later inserted
1701 // safepoints can get the properly relocated base register.
1702 DenseSet<Value *> missing;
1703 for (Value *L : liveset) {
1704 assert(base_pairs.find(L) != base_pairs.end());
1705 Value *base = base_pairs[L];
1707 if (liveset.find(base) == liveset.end()) {
1708 assert(base_pairs.find(base) == base_pairs.end());
1709 // uniqued by set insert
1710 missing.insert(base);
1714 // Note that we want these at the end of the list, otherwise
1715 // register placement gets screwed up once we lower to STATEPOINT
1716 // instructions. This is an utter hack, but there doesn't seem to be a
1718 for (Value *base : missing) {
1720 liveset.insert(base);
1721 base_pairs[base] = base;
1723 assert(liveset.size() == base_pairs.size());
1726 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1727 std::vector<CallSite> &toUpdate) {
1729 // sanity check the input
1730 std::set<CallSite> uniqued;
1731 uniqued.insert(toUpdate.begin(), toUpdate.end());
1732 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1734 for (size_t i = 0; i < toUpdate.size(); i++) {
1735 CallSite &CS = toUpdate[i];
1736 assert(CS.getInstruction()->getParent()->getParent() == &F);
1737 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1741 // A list of dummy calls added to the IR to keep various values obviously
1742 // live in the IR. We'll remove all of these when done.
1743 std::vector<CallInst *> holders;
1745 // Insert a dummy call with all of the arguments to the vm_state we'll need
1746 // for the actual safepoint insertion. This ensures reference arguments in
1747 // the deopt argument list are considered live through the safepoint (and
1748 // thus makes sure they get relocated.)
1749 for (size_t i = 0; i < toUpdate.size(); i++) {
1750 CallSite &CS = toUpdate[i];
1751 Statepoint StatepointCS(CS);
1753 SmallVector<Value *, 64> DeoptValues;
1754 for (Use &U : StatepointCS.vm_state_args()) {
1755 Value *Arg = cast<Value>(&U);
1756 if (isGCPointerType(Arg->getType()))
1757 DeoptValues.push_back(Arg);
1759 insertUseHolderAfter(CS, DeoptValues, holders);
1762 std::vector<struct PartiallyConstructedSafepointRecord> records;
1763 records.reserve(toUpdate.size());
1764 for (size_t i = 0; i < toUpdate.size(); i++) {
1765 struct PartiallyConstructedSafepointRecord info;
1766 records.push_back(info);
1768 assert(records.size() == toUpdate.size());
1770 // A) Identify all gc pointers which are staticly live at the given call
1772 findLiveReferences(F, DT, P, toUpdate, records);
1774 // B) Find the base pointers for each live pointer
1775 /* scope for caching */ {
1776 // Cache the 'defining value' relation used in the computation and
1777 // insertion of base phis and selects. This ensures that we don't insert
1778 // large numbers of duplicate base_phis.
1779 DefiningValueMapTy DVCache;
1781 for (size_t i = 0; i < records.size(); i++) {
1782 struct PartiallyConstructedSafepointRecord &info = records[i];
1783 CallSite &CS = toUpdate[i];
1784 findBasePointers(DT, DVCache, CS, info);
1786 } // end of cache scope
1788 // The base phi insertion logic (for any safepoint) may have inserted new
1789 // instructions which are now live at some safepoint. The simplest such
1792 // phi a <-- will be a new base_phi here
1793 // safepoint 1 <-- that needs to be live here
1797 std::set<llvm::Value *> allInsertedDefs;
1798 for (size_t i = 0; i < records.size(); i++) {
1799 struct PartiallyConstructedSafepointRecord &info = records[i];
1800 allInsertedDefs.insert(info.newInsertedDefs.begin(),
1801 info.newInsertedDefs.end());
1804 // We insert some dummy calls after each safepoint to definitely hold live
1805 // the base pointers which were identified for that safepoint. We'll then
1806 // ask liveness for _every_ base inserted to see what is now live. Then we
1807 // remove the dummy calls.
1808 holders.reserve(holders.size() + records.size());
1809 for (size_t i = 0; i < records.size(); i++) {
1810 struct PartiallyConstructedSafepointRecord &info = records[i];
1811 CallSite &CS = toUpdate[i];
1813 SmallVector<Value *, 128> Bases;
1814 for (auto Pair : info.base_pairs) {
1815 Bases.push_back(Pair.second);
1817 insertUseHolderAfter(CS, Bases, holders);
1820 // Add the bases explicitly to the live vector set. This may result in a few
1821 // extra relocations, but the base has to be available whenever a pointer
1822 // derived from it is used. Thus, we need it to be part of the statepoint's
1823 // gc arguments list. TODO: Introduce an explicit notion (in the following
1824 // code) of the GC argument list as seperate from the live Values at a
1825 // given statepoint.
1826 for (size_t i = 0; i < records.size(); i++) {
1827 struct PartiallyConstructedSafepointRecord &info = records[i];
1828 addBasesAsLiveValues(info.liveset, info.base_pairs);
1831 // If we inserted any new values, we need to adjust our notion of what is
1832 // live at a particular safepoint.
1833 if (!allInsertedDefs.empty()) {
1834 fixupLiveReferences(F, DT, P, allInsertedDefs, toUpdate, records);
1836 if (PrintBasePointers) {
1837 for (size_t i = 0; i < records.size(); i++) {
1838 struct PartiallyConstructedSafepointRecord &info = records[i];
1839 errs() << "Base Pairs: (w/Relocation)\n";
1840 for (auto Pair : info.base_pairs) {
1841 errs() << " derived %" << Pair.first->getName() << " base %"
1842 << Pair.second->getName() << "\n";
1846 for (size_t i = 0; i < holders.size(); i++) {
1847 holders[i]->eraseFromParent();
1848 holders[i] = nullptr;
1852 // Now run through and replace the existing statepoints with new ones with
1853 // the live variables listed. We do not yet update uses of the values being
1854 // relocated. We have references to live variables that need to
1855 // survive to the last iteration of this loop. (By construction, the
1856 // previous statepoint can not be a live variable, thus we can and remove
1857 // the old statepoint calls as we go.)
1858 for (size_t i = 0; i < records.size(); i++) {
1859 struct PartiallyConstructedSafepointRecord &info = records[i];
1860 CallSite &CS = toUpdate[i];
1861 makeStatepointExplicit(DT, CS, P, info);
1863 toUpdate.clear(); // prevent accident use of invalid CallSites
1865 // In case if we inserted relocates in a different basic block than the
1866 // original safepoint (this can happen for invokes). We need to be sure that
1867 // original values were not used in any of the phi nodes at the
1868 // beginning of basic block containing them. Because we know that all such
1869 // blocks will have single predecessor we can safely assume that all phi
1870 // nodes have single entry (because of normalizeBBForInvokeSafepoint).
1871 // Just remove them all here.
1872 for (size_t i = 0; i < records.size(); i++) {
1873 Instruction *I = records[i].safepoint.first;
1875 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
1876 FoldSingleEntryPHINodes(invoke->getNormalDest());
1877 assert(!isa<PHINode>(invoke->getNormalDest()->begin()));
1879 FoldSingleEntryPHINodes(invoke->getUnwindDest());
1880 assert(!isa<PHINode>(invoke->getUnwindDest()->begin()));
1884 // Do all the fixups of the original live variables to their relocated selves
1885 std::vector<Value *> live;
1886 for (size_t i = 0; i < records.size(); i++) {
1887 struct PartiallyConstructedSafepointRecord &info = records[i];
1888 // We can't simply save the live set from the original insertion. One of
1889 // the live values might be the result of a call which needs a safepoint.
1890 // That Value* no longer exists and we need to use the new gc_result.
1891 // Thankfully, the liveset is embedded in the statepoint (and updated), so
1892 // we just grab that.
1893 Statepoint statepoint(info.safepoint.first);
1894 live.insert(live.end(), statepoint.gc_args_begin(),
1895 statepoint.gc_args_end());
1897 unique_unsorted(live);
1901 for (auto ptr : live) {
1902 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
1906 relocationViaAlloca(F, DT, live, records);
1907 return !records.empty();
1910 /// Returns true if this function should be rewritten by this pass. The main
1911 /// point of this function is as an extension point for custom logic.
1912 static bool shouldRewriteStatepointsIn(Function &F) {
1913 // TODO: This should check the GCStrategy
1914 const std::string StatepointExampleName("statepoint-example");
1915 return StatepointExampleName == F.getGC();
1918 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
1919 // Nothing to do for declarations.
1920 if (F.isDeclaration() || F.empty())
1923 // Policy choice says not to rewrite - the most common reason is that we're
1924 // compiling code without a GCStrategy.
1925 if (!shouldRewriteStatepointsIn(F))
1928 // Gather all the statepoints which need rewritten.
1929 std::vector<CallSite> ParsePointNeeded;
1930 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1932 // TODO: only the ones with the flag set!
1933 if (isStatepoint(*itr))
1934 ParsePointNeeded.push_back(CallSite(&*itr));
1937 // Return early if no work to do.
1938 if (ParsePointNeeded.empty())
1941 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1942 return insertParsePoints(F, DT, this, ParsePointNeeded);