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 std::set<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 else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
154 return isGCPointerType(AT->getElementType()) ||
155 isAggWhichContainsGCPtrType(AT->getElementType());
156 } else if (StructType *ST = dyn_cast<StructType>(Ty)) {
157 bool UnsupportedType = false;
158 for (Type *SubType : ST->subtypes())
160 isGCPointerType(SubType) || isAggWhichContainsGCPtrType(SubType);
161 return UnsupportedType;
167 // Conservatively identifies any definitions which might be live at the
168 // given instruction. The analysis is performed immediately before the
169 // given instruction. Values defined by that instruction are not considered
170 // live. Values used by that instruction are considered live.
172 // preconditions: valid IR graph, term is either a terminator instruction or
173 // a call instruction, pred is the basic block of term, DT, LI are valid
175 // side effects: none, does not mutate IR
177 // postconditions: populates liveValues as discussed above
178 static void findLiveGCValuesAtInst(Instruction *term, BasicBlock *pred,
179 DominatorTree &DT, LoopInfo *LI,
180 std::set<llvm::Value *> &liveValues) {
183 assert(isa<CallInst>(term) || isa<InvokeInst>(term) || term->isTerminator());
185 Function *F = pred->getParent();
187 auto is_live_gc_reference =
188 [&](Value &V) { return isLiveGCReferenceAt(V, term, DT, LI); };
190 // Are there any gc pointer arguments live over this point? This needs to be
191 // special cased since arguments aren't defined in basic blocks.
192 for (Argument &arg : F->args()) {
193 assert(!isAggWhichContainsGCPtrType(arg.getType()) &&
194 "support for FCA unimplemented");
196 if (is_live_gc_reference(arg)) {
197 liveValues.insert(&arg);
201 // Walk through all dominating blocks - the ones which can contain
202 // definitions used in this block - and check to see if any of the values
203 // they define are used in locations potentially reachable from the
204 // interesting instruction.
205 BasicBlock *BBI = pred;
208 errs() << "[LSP] Looking at dominating block " << pred->getName() << "\n";
210 assert(DT.dominates(BBI, pred));
211 assert(isPotentiallyReachable(BBI, pred, &DT) &&
212 "dominated block must be reachable");
214 // Walk through the instructions in dominating blocks and keep any
215 // that have a use potentially reachable from the block we're
216 // considering putting the safepoint in
217 for (Instruction &inst : *BBI) {
219 errs() << "[LSP] Looking at instruction ";
223 if (pred == BBI && (&inst) == term) {
225 errs() << "[LSP] stopped because we encountered the safepoint "
229 // If we're in the block which defines the interesting instruction,
230 // we don't want to include any values as live which are defined
231 // _after_ the interesting line or as part of the line itself
232 // i.e. "term" is the call instruction for a call safepoint, the
233 // results of the call should not be considered live in that stackmap
237 assert(!isAggWhichContainsGCPtrType(inst.getType()) &&
238 "support for FCA unimplemented");
240 if (is_live_gc_reference(inst)) {
242 errs() << "[LSP] found live value for this safepoint ";
246 liveValues.insert(&inst);
249 if (!DT.getNode(BBI)->getIDom()) {
250 assert(BBI == &F->getEntryBlock() &&
251 "failed to find a dominator for something other than "
255 BBI = DT.getNode(BBI)->getIDom()->getBlock();
259 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
260 if (a->hasName() && b->hasName()) {
261 return -1 == a->getName().compare(b->getName());
262 } else if (a->hasName() && !b->hasName()) {
264 } else if (!a->hasName() && b->hasName()) {
267 // Better than nothing, but not stable
272 /// Find the initial live set. Note that due to base pointer
273 /// insertion, the live set may be incomplete.
275 analyzeParsePointLiveness(DominatorTree &DT, const CallSite &CS,
276 PartiallyConstructedSafepointRecord &result) {
277 Instruction *inst = CS.getInstruction();
279 BasicBlock *BB = inst->getParent();
280 std::set<Value *> liveset;
281 findLiveGCValuesAtInst(inst, BB, DT, nullptr, liveset);
284 // Note: This output is used by several of the test cases
285 // The order of elemtns in a set is not stable, put them in a vec and sort
287 SmallVector<Value *, 64> temp;
288 temp.insert(temp.end(), liveset.begin(), liveset.end());
289 std::sort(temp.begin(), temp.end(), order_by_name);
290 errs() << "Live Variables:\n";
291 for (Value *V : temp) {
292 errs() << " " << V->getName(); // no newline
296 if (PrintLiveSetSize) {
297 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
298 errs() << "Number live values: " << liveset.size() << "\n";
300 result.liveset = liveset;
303 /// True iff this value is the null pointer constant (of any pointer type)
304 static bool isNullConstant(Value *V) {
305 return isa<Constant>(V) && isa<PointerType>(V->getType()) &&
306 cast<Constant>(V)->isNullValue();
309 /// Helper function for findBasePointer - Will return a value which either a)
310 /// defines the base pointer for the input or b) blocks the simple search
311 /// (i.e. a PHI or Select of two derived pointers)
312 static Value *findBaseDefiningValue(Value *I) {
313 assert(I->getType()->isPointerTy() &&
314 "Illegal to ask for the base pointer of a non-pointer type");
316 // There are instructions which can never return gc pointer values. Sanity
318 // that this is actually true.
319 assert(!isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
320 !isa<ShuffleVectorInst>(I) && "Vector types are not gc pointers");
321 assert((!isa<Instruction>(I) || isa<InvokeInst>(I) ||
322 !cast<Instruction>(I)->isTerminator()) &&
323 "With the exception of invoke terminators don't define values");
324 assert(!isa<StoreInst>(I) && !isa<FenceInst>(I) &&
325 "Can't be definitions to start with");
326 assert(!isa<ICmpInst>(I) && !isa<FCmpInst>(I) &&
327 "Comparisons don't give ops");
328 // There's a bunch of instructions which just don't make sense to apply to
329 // a pointer. The only valid reason for this would be pointer bit
330 // twiddling which we're just not going to support.
331 assert((!isa<Instruction>(I) || !cast<Instruction>(I)->isBinaryOp()) &&
332 "Binary ops on pointer values are meaningless. Unless your "
333 "bit-twiddling which we don't support");
335 if (Argument *Arg = dyn_cast<Argument>(I)) {
336 // An incoming argument to the function is a base pointer
337 // We should have never reached here if this argument isn't an gc value
338 assert(Arg->getType()->isPointerTy() &&
339 "Base for pointer must be another pointer");
343 if (GlobalVariable *global = dyn_cast<GlobalVariable>(I)) {
345 assert(global->getType()->isPointerTy() &&
346 "Base for pointer must be another pointer");
350 // inlining could possibly introduce phi node that contains
351 // undef if callee has multiple returns
352 if (UndefValue *undef = dyn_cast<UndefValue>(I)) {
353 assert(undef->getType()->isPointerTy() &&
354 "Base for pointer must be another pointer");
355 return undef; // utterly meaningless, but useful for dealing with
356 // partially optimized code.
359 // Due to inheritance, this must be _after_ the global variable and undef
361 if (Constant *con = dyn_cast<Constant>(I)) {
362 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
363 "order of checks wrong!");
364 // Note: Finding a constant base for something marked for relocation
365 // doesn't really make sense. The most likely case is either a) some
366 // screwed up the address space usage or b) your validating against
367 // compiled C++ code w/o the proper separation. The only real exception
368 // is a null pointer. You could have generic code written to index of
369 // off a potentially null value and have proven it null. We also use
370 // null pointers in dead paths of relocation phis (which we might later
371 // want to find a base pointer for).
372 assert(con->getType()->isPointerTy() &&
373 "Base for pointer must be another pointer");
374 assert(con->isNullValue() && "null is the only case which makes sense");
378 if (CastInst *CI = dyn_cast<CastInst>(I)) {
379 Value *def = CI->stripPointerCasts();
380 assert(def->getType()->isPointerTy() &&
381 "Base for pointer must be another pointer");
382 if (isa<CastInst>(def)) {
383 // If we find a cast instruction here, it means we've found a cast
384 // which is not simply a pointer cast (i.e. an inttoptr). We don't
385 // know how to handle int->ptr conversion.
386 llvm_unreachable("Can not find the base pointers for an inttoptr cast");
388 assert(!isa<CastInst>(def) && "shouldn't find another cast here");
389 return findBaseDefiningValue(def);
392 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
393 if (LI->getType()->isPointerTy()) {
394 Value *Op = LI->getOperand(0);
396 // Has to be a pointer to an gc object, or possibly an array of such?
397 assert(Op->getType()->isPointerTy());
398 return LI; // The value loaded is an gc base itself
401 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
402 Value *Op = GEP->getOperand(0);
403 if (Op->getType()->isPointerTy()) {
404 return findBaseDefiningValue(Op); // The base of this GEP is the base
408 if (AllocaInst *alloc = dyn_cast<AllocaInst>(I)) {
409 // An alloca represents a conceptual stack slot. It's the slot itself
410 // that the GC needs to know about, not the value in the slot.
411 assert(alloc->getType()->isPointerTy() &&
412 "Base for pointer must be another pointer");
416 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
417 switch (II->getIntrinsicID()) {
419 // fall through to general call handling
421 case Intrinsic::experimental_gc_statepoint:
422 case Intrinsic::experimental_gc_result_float:
423 case Intrinsic::experimental_gc_result_int:
424 llvm_unreachable("these don't produce pointers");
425 case Intrinsic::experimental_gc_result_ptr:
426 // This is just a special case of the CallInst check below to handle a
427 // statepoint with deopt args which hasn't been rewritten for GC yet.
428 // TODO: Assert that the statepoint isn't rewritten yet.
430 case Intrinsic::experimental_gc_relocate: {
431 // Rerunning safepoint insertion after safepoints are already
432 // inserted is not supported. It could probably be made to work,
433 // but why are you doing this? There's no good reason.
434 llvm_unreachable("repeat safepoint insertion is not supported");
436 case Intrinsic::gcroot:
437 // Currently, this mechanism hasn't been extended to work with gcroot.
438 // There's no reason it couldn't be, but I haven't thought about the
439 // implications much.
441 "interaction with the gcroot mechanism is not supported");
444 // We assume that functions in the source language only return base
445 // pointers. This should probably be generalized via attributes to support
446 // both source language and internal functions.
447 if (CallInst *call = dyn_cast<CallInst>(I)) {
448 assert(call->getType()->isPointerTy() &&
449 "Base for pointer must be another pointer");
452 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
453 assert(invoke->getType()->isPointerTy() &&
454 "Base for pointer must be another pointer");
458 // I have absolutely no idea how to implement this part yet. It's not
459 // neccessarily hard, I just haven't really looked at it yet.
460 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
462 if (AtomicCmpXchgInst *cas = dyn_cast<AtomicCmpXchgInst>(I)) {
463 // A CAS is effectively a atomic store and load combined under a
464 // predicate. From the perspective of base pointers, we just treat it
465 // like a load. We loaded a pointer from a address in memory, that value
466 // had better be a valid base pointer.
467 return cas->getPointerOperand();
469 if (AtomicRMWInst *atomic = dyn_cast<AtomicRMWInst>(I)) {
470 assert(AtomicRMWInst::Xchg == atomic->getOperation() &&
471 "All others are binary ops which don't apply to base pointers");
472 // semantically, a load, store pair. Treat it the same as a standard load
473 return atomic->getPointerOperand();
476 // The aggregate ops. Aggregates can either be in the heap or on the
477 // stack, but in either case, this is simply a field load. As a result,
478 // this is a defining definition of the base just like a load is.
479 if (ExtractValueInst *ev = dyn_cast<ExtractValueInst>(I)) {
483 // We should never see an insert vector since that would require we be
484 // tracing back a struct value not a pointer value.
485 assert(!isa<InsertValueInst>(I) &&
486 "Base pointer for a struct is meaningless");
488 // The last two cases here don't return a base pointer. Instead, they
489 // return a value which dynamically selects from amoung several base
490 // derived pointers (each with it's own base potentially). It's the job of
491 // the caller to resolve these.
492 if (SelectInst *select = dyn_cast<SelectInst>(I)) {
495 if (PHINode *phi = dyn_cast<PHINode>(I)) {
499 errs() << "unknown type: " << *I << "\n";
500 llvm_unreachable("unknown type");
504 /// Returns the base defining value for this value.
505 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &cache) {
506 Value *&Cached = cache[I];
508 Cached = findBaseDefiningValue(I);
510 assert(cache[I] != nullptr);
513 errs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
519 /// Return a base pointer for this value if known. Otherwise, return it's
520 /// base defining value.
521 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &cache) {
522 Value *def = findBaseDefiningValueCached(I, cache);
523 auto Found = cache.find(def);
524 if (Found != cache.end()) {
525 // Either a base-of relation, or a self reference. Caller must check.
526 return Found->second;
528 // Only a BDV available
532 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
533 /// is it known to be a base pointer? Or do we need to continue searching.
534 static bool isKnownBaseResult(Value *v) {
535 if (!isa<PHINode>(v) && !isa<SelectInst>(v)) {
536 // no recursion possible
539 if (cast<Instruction>(v)->getMetadata("is_base_value")) {
540 // This is a previously inserted base phi or select. We know
541 // that this is a base value.
545 // We need to keep searching
549 // TODO: find a better name for this
553 enum Status { Unknown, Base, Conflict };
555 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
556 assert(status != Base || b);
558 PhiState(Value *b) : status(Base), base(b) {}
559 PhiState() : status(Unknown), base(nullptr) {}
560 PhiState(const PhiState &other) : status(other.status), base(other.base) {
561 assert(status != Base || base);
564 Status getStatus() const { return status; }
565 Value *getBase() const { return base; }
567 bool isBase() const { return getStatus() == Base; }
568 bool isUnknown() const { return getStatus() == Unknown; }
569 bool isConflict() const { return getStatus() == Conflict; }
571 bool operator==(const PhiState &other) const {
572 return base == other.base && status == other.status;
575 bool operator!=(const PhiState &other) const { return !(*this == other); }
578 errs() << status << " (" << base << " - "
579 << (base ? base->getName() : "nullptr") << "): ";
584 Value *base; // non null only if status == base
587 typedef DenseMap<Value *, PhiState> ConflictStateMapTy;
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 ConflictStateMapTy &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 ConflictStateMapTy &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 llvm_unreachable("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 DenseSet<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 ConflictStateMapTy 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);
826 llvm_unreachable("unknown conflict type");
830 // Fixup all the inputs of the new PHIs
831 for (auto Pair : states) {
832 Instruction *v = cast<Instruction>(Pair.first);
833 PhiState state = Pair.second;
835 assert(!isKnownBaseResult(v) && "why did it get added?");
836 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
837 if (state.isConflict()) {
838 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
839 PHINode *phi = cast<PHINode>(v);
840 unsigned NumPHIValues = phi->getNumIncomingValues();
841 for (unsigned i = 0; i < NumPHIValues; i++) {
842 Value *InVal = phi->getIncomingValue(i);
843 BasicBlock *InBB = phi->getIncomingBlock(i);
845 // If we've already seen InBB, add the same incoming value
846 // we added for it earlier. The IR verifier requires phi
847 // nodes with multiple entries from the same basic block
848 // to have the same incoming value for each of those
849 // entries. If we don't do this check here and basephi
850 // has a different type than base, we'll end up adding two
851 // bitcasts (and hence two distinct values) as incoming
852 // values for the same basic block.
854 int blockIndex = basephi->getBasicBlockIndex(InBB);
855 if (blockIndex != -1) {
856 Value *oldBase = basephi->getIncomingValue(blockIndex);
857 basephi->addIncoming(oldBase, InBB);
859 Value *base = findBaseOrBDV(InVal, cache);
860 if (!isKnownBaseResult(base)) {
861 // Either conflict or base.
862 assert(states.count(base));
863 base = states[base].getBase();
864 assert(base != nullptr && "unknown PhiState!");
865 assert(NewInsertedDefs.count(base) &&
866 "should have already added this in a prev. iteration!");
869 // In essense this assert states: the only way two
870 // values incoming from the same basic block may be
871 // different is by being different bitcasts of the same
872 // value. A cleanup that remains TODO is changing
873 // findBaseOrBDV to return an llvm::Value of the correct
874 // type (and still remain pure). This will remove the
875 // need to add bitcasts.
876 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
877 "sanity -- findBaseOrBDV should be pure!");
882 // Find either the defining value for the PHI or the normal base for
884 Value *base = findBaseOrBDV(InVal, cache);
885 if (!isKnownBaseResult(base)) {
886 // Either conflict or base.
887 assert(states.count(base));
888 base = states[base].getBase();
889 assert(base != nullptr && "unknown PhiState!");
891 assert(base && "can't be null");
892 // Must use original input BB since base may not be Instruction
893 // The cast is needed since base traversal may strip away bitcasts
894 if (base->getType() != basephi->getType()) {
895 base = new BitCastInst(base, basephi->getType(), "cast",
896 InBB->getTerminator());
897 NewInsertedDefs.insert(base);
899 basephi->addIncoming(base, InBB);
901 assert(basephi->getNumIncomingValues() == NumPHIValues);
902 } else if (SelectInst *basesel = dyn_cast<SelectInst>(state.getBase())) {
903 SelectInst *sel = cast<SelectInst>(v);
904 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
905 // something more safe and less hacky.
906 for (int i = 1; i <= 2; i++) {
907 Value *InVal = sel->getOperand(i);
908 // Find either the defining value for the PHI or the normal base for
910 Value *base = findBaseOrBDV(InVal, cache);
911 if (!isKnownBaseResult(base)) {
912 // Either conflict or base.
913 assert(states.count(base));
914 base = states[base].getBase();
915 assert(base != nullptr && "unknown PhiState!");
917 assert(base && "can't be null");
918 // Must use original input BB since base may not be Instruction
919 // The cast is needed since base traversal may strip away bitcasts
920 if (base->getType() != basesel->getType()) {
921 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
922 NewInsertedDefs.insert(base);
924 basesel->setOperand(i, base);
927 llvm_unreachable("unexpected conflict type");
931 // Cache all of our results so we can cheaply reuse them
932 // NOTE: This is actually two caches: one of the base defining value
933 // relation and one of the base pointer relation! FIXME
934 for (auto item : states) {
935 Value *v = item.first;
936 Value *base = item.second.getBase();
938 assert(!isKnownBaseResult(v) && "why did it get added?");
941 std::string fromstr =
942 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
944 errs() << "Updating base value cache"
945 << " for: " << (v->hasName() ? v->getName() : "")
946 << " from: " << fromstr
947 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
950 assert(isKnownBaseResult(base) &&
951 "must be something we 'know' is a base pointer");
952 if (cache.count(v)) {
953 // Once we transition from the BDV relation being store in the cache to
954 // the base relation being stored, it must be stable
955 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
956 "base relation should be stable");
960 assert(cache.find(def) != cache.end());
964 // For a set of live pointers (base and/or derived), identify the base
965 // pointer of the object which they are derived from. This routine will
966 // mutate the IR graph as needed to make the 'base' pointer live at the
967 // definition site of 'derived'. This ensures that any use of 'derived' can
968 // also use 'base'. This may involve the insertion of a number of
969 // additional PHI nodes.
971 // preconditions: live is a set of pointer type Values
973 // side effects: may insert PHI nodes into the existing CFG, will preserve
974 // CFG, will not remove or mutate any existing nodes
976 // post condition: PointerToBase contains one (derived, base) pair for every
977 // pointer in live. Note that derived can be equal to base if the original
978 // pointer was a base pointer.
979 static void findBasePointers(const std::set<llvm::Value *> &live,
980 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
981 DominatorTree *DT, DefiningValueMapTy &DVCache,
982 DenseSet<llvm::Value *> &NewInsertedDefs) {
983 for (Value *ptr : live) {
984 Value *base = findBasePointer(ptr, DVCache, NewInsertedDefs);
985 assert(base && "failed to find base pointer");
986 PointerToBase[ptr] = base;
987 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
988 DT->dominates(cast<Instruction>(base)->getParent(),
989 cast<Instruction>(ptr)->getParent())) &&
990 "The base we found better dominate the derived pointer");
992 if (isNullConstant(base))
993 // If you see this trip and like to live really dangerously, the code
994 // should be correct, just with idioms the verifier can't handle. You
995 // can try disabling the verifier at your own substaintial risk.
996 llvm_unreachable("the relocation code needs adjustment to handle the"
997 "relocation of a null pointer constant without causing"
998 "false positives in the safepoint ir verifier.");
1002 /// Find the required based pointers (and adjust the live set) for the given
1004 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1006 PartiallyConstructedSafepointRecord &result) {
1007 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
1008 DenseSet<llvm::Value *> NewInsertedDefs;
1009 findBasePointers(result.liveset, PointerToBase, &DT, DVCache, NewInsertedDefs);
1011 if (PrintBasePointers) {
1012 errs() << "Base Pairs (w/o Relocation):\n";
1013 for (auto Pair : PointerToBase) {
1014 errs() << " derived %" << Pair.first->getName() << " base %"
1015 << Pair.second->getName() << "\n";
1019 result.PointerToBase = PointerToBase;
1020 result.NewInsertedDefs = NewInsertedDefs;
1023 /// Check for liveness of items in the insert defs and add them to the live
1024 /// and base pointer sets
1025 static void fixupLiveness(DominatorTree &DT, const CallSite &CS,
1026 const std::set<Value *> &allInsertedDefs,
1027 PartiallyConstructedSafepointRecord &result) {
1028 Instruction *inst = CS.getInstruction();
1030 auto liveset = result.liveset;
1031 auto PointerToBase = result.PointerToBase;
1033 auto is_live_gc_reference =
1034 [&](Value &V) { return isLiveGCReferenceAt(V, inst, DT, nullptr); };
1036 // For each new definition, check to see if a) the definition dominates the
1037 // instruction we're interested in, and b) one of the uses of that definition
1038 // is edge-reachable from the instruction we're interested in. This is the
1039 // same definition of liveness we used in the intial liveness analysis
1040 for (Value *newDef : allInsertedDefs) {
1041 if (liveset.count(newDef)) {
1042 // already live, no action needed
1046 // PERF: Use DT to check instruction domination might not be good for
1047 // compilation time, and we could change to optimal solution if this
1048 // turn to be a issue
1049 if (!DT.dominates(cast<Instruction>(newDef), inst)) {
1050 // can't possibly be live at inst
1054 if (is_live_gc_reference(*newDef)) {
1055 // Add the live new defs into liveset and PointerToBase
1056 liveset.insert(newDef);
1057 PointerToBase[newDef] = newDef;
1061 result.liveset = liveset;
1062 result.PointerToBase = PointerToBase;
1065 static void fixupLiveReferences(
1066 Function &F, DominatorTree &DT, Pass *P,
1067 const std::set<llvm::Value *> &allInsertedDefs,
1068 ArrayRef<CallSite> toUpdate,
1069 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1070 for (size_t i = 0; i < records.size(); i++) {
1071 struct PartiallyConstructedSafepointRecord &info = records[i];
1072 const CallSite &CS = toUpdate[i];
1073 fixupLiveness(DT, CS, allInsertedDefs, info);
1077 // Normalize basic block to make it ready to be target of invoke statepoint.
1078 // It means spliting it to have single predecessor. Return newly created BB
1079 // ready to be successor of invoke statepoint.
1080 static BasicBlock *normalizeBBForInvokeSafepoint(BasicBlock *BB,
1081 BasicBlock *InvokeParent,
1083 BasicBlock *ret = BB;
1085 if (!BB->getUniquePredecessor()) {
1086 ret = SplitBlockPredecessors(BB, InvokeParent, "");
1089 // Another requirement for such basic blocks is to not have any phi nodes.
1090 // Since we just ensured that new BB will have single predecessor,
1091 // all phi nodes in it will have one value. Here it would be naturall place
1093 // remove them all. But we can not do this because we are risking to remove
1094 // one of the values stored in liveset of another statepoint. We will do it
1095 // later after placing all safepoints.
1100 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1101 auto itr = std::find(livevec.begin(), livevec.end(), val);
1102 assert(livevec.end() != itr);
1103 size_t index = std::distance(livevec.begin(), itr);
1104 assert(index < livevec.size());
1108 // Create new attribute set containing only attributes which can be transfered
1109 // from original call to the safepoint.
1110 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1113 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1114 unsigned index = AS.getSlotIndex(Slot);
1116 if (index == AttributeSet::ReturnIndex ||
1117 index == AttributeSet::FunctionIndex) {
1119 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1121 Attribute attr = *it;
1123 // Do not allow certain attributes - just skip them
1124 // Safepoint can not be read only or read none.
1125 if (attr.hasAttribute(Attribute::ReadNone) ||
1126 attr.hasAttribute(Attribute::ReadOnly))
1129 ret = ret.addAttributes(
1130 AS.getContext(), index,
1131 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1135 // Just skip parameter attributes for now
1141 /// Helper function to place all gc relocates necessary for the given
1144 /// liveVariables - list of variables to be relocated.
1145 /// liveStart - index of the first live variable.
1146 /// basePtrs - base pointers.
1147 /// statepointToken - statepoint instruction to which relocates should be
1149 /// Builder - Llvm IR builder to be used to construct new calls.
1150 void CreateGCRelocates(ArrayRef<llvm::Value *> liveVariables,
1151 const int liveStart,
1152 ArrayRef<llvm::Value *> basePtrs,
1153 Instruction *statepointToken, IRBuilder<> Builder) {
1155 SmallVector<Instruction *, 64> NewDefs;
1156 NewDefs.reserve(liveVariables.size());
1158 Module *M = statepointToken->getParent()->getParent()->getParent();
1160 for (unsigned i = 0; i < liveVariables.size(); i++) {
1161 // We generate a (potentially) unique declaration for every pointer type
1162 // combination. This results is some blow up the function declarations in
1163 // the IR, but removes the need for argument bitcasts which shrinks the IR
1164 // greatly and makes it much more readable.
1165 SmallVector<Type *, 1> types; // one per 'any' type
1166 types.push_back(liveVariables[i]->getType()); // result type
1167 Value *gc_relocate_decl = Intrinsic::getDeclaration(
1168 M, Intrinsic::experimental_gc_relocate, types);
1170 // Generate the gc.relocate call and save the result
1172 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1173 liveStart + find_index(liveVariables, basePtrs[i]));
1174 Value *liveIdx = ConstantInt::get(
1175 Type::getInt32Ty(M->getContext()),
1176 liveStart + find_index(liveVariables, liveVariables[i]));
1178 // only specify a debug name if we can give a useful one
1179 Value *reloc = Builder.CreateCall3(
1180 gc_relocate_decl, statepointToken, baseIdx, liveIdx,
1181 liveVariables[i]->hasName() ? liveVariables[i]->getName() + ".relocated"
1183 // Trick CodeGen into thinking there are lots of free registers at this
1185 cast<CallInst>(reloc)->setCallingConv(CallingConv::Cold);
1187 NewDefs.push_back(cast<Instruction>(reloc));
1189 assert(NewDefs.size() == liveVariables.size() &&
1190 "missing or extra redefinition at safepoint");
1194 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1195 const SmallVectorImpl<llvm::Value *> &basePtrs,
1196 const SmallVectorImpl<llvm::Value *> &liveVariables,
1198 PartiallyConstructedSafepointRecord &result) {
1199 assert(basePtrs.size() == liveVariables.size());
1200 assert(isStatepoint(CS) &&
1201 "This method expects to be rewriting a statepoint");
1203 BasicBlock *BB = CS.getInstruction()->getParent();
1205 Function *F = BB->getParent();
1206 assert(F && "must be set");
1207 Module *M = F->getParent();
1209 assert(M && "must be set");
1211 // We're not changing the function signature of the statepoint since the gc
1212 // arguments go into the var args section.
1213 Function *gc_statepoint_decl = CS.getCalledFunction();
1215 // Then go ahead and use the builder do actually do the inserts. We insert
1216 // immediately before the previous instruction under the assumption that all
1217 // arguments will be available here. We can't insert afterwards since we may
1218 // be replacing a terminator.
1219 Instruction *insertBefore = CS.getInstruction();
1220 IRBuilder<> Builder(insertBefore);
1221 // Copy all of the arguments from the original statepoint - this includes the
1222 // target, call args, and deopt args
1223 SmallVector<llvm::Value *, 64> args;
1224 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1225 // TODO: Clear the 'needs rewrite' flag
1227 // add all the pointers to be relocated (gc arguments)
1228 // Capture the start of the live variable list for use in the gc_relocates
1229 const int live_start = args.size();
1230 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1232 // Create the statepoint given all the arguments
1233 Instruction *token = nullptr;
1234 AttributeSet return_attributes;
1236 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1238 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1239 call->setTailCall(toReplace->isTailCall());
1240 call->setCallingConv(toReplace->getCallingConv());
1242 // Currently we will fail on parameter attributes and on certain
1243 // function attributes.
1244 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1245 // In case if we can handle this set of sttributes - set up function attrs
1246 // directly on statepoint and return attrs later for gc_result intrinsic.
1247 call->setAttributes(new_attrs.getFnAttributes());
1248 return_attributes = new_attrs.getRetAttributes();
1252 // Put the following gc_result and gc_relocate calls immediately after the
1253 // the old call (which we're about to delete)
1254 BasicBlock::iterator next(toReplace);
1255 assert(BB->end() != next && "not a terminator, must have next");
1257 Instruction *IP = &*(next);
1258 Builder.SetInsertPoint(IP);
1259 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1261 } else if (CS.isInvoke()) {
1262 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1264 // Insert the new invoke into the old block. We'll remove the old one in a
1265 // moment at which point this will become the new terminator for the
1267 InvokeInst *invoke = InvokeInst::Create(
1268 gc_statepoint_decl, toReplace->getNormalDest(),
1269 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1270 invoke->setCallingConv(toReplace->getCallingConv());
1272 // Currently we will fail on parameter attributes and on certain
1273 // function attributes.
1274 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1275 // In case if we can handle this set of sttributes - set up function attrs
1276 // directly on statepoint and return attrs later for gc_result intrinsic.
1277 invoke->setAttributes(new_attrs.getFnAttributes());
1278 return_attributes = new_attrs.getRetAttributes();
1282 // Generate gc relocates in exceptional path
1283 BasicBlock *unwindBlock = normalizeBBForInvokeSafepoint(
1284 toReplace->getUnwindDest(), invoke->getParent(), P);
1286 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1287 Builder.SetInsertPoint(IP);
1288 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1290 // Extract second element from landingpad return value. We will attach
1291 // exceptional gc relocates to it.
1292 const unsigned idx = 1;
1293 Instruction *exceptional_token =
1294 cast<Instruction>(Builder.CreateExtractValue(
1295 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1296 result.UnwindToken = exceptional_token;
1298 // Just throw away return value. We will use the one we got for normal
1300 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1301 exceptional_token, Builder);
1303 // Generate gc relocates and returns for normal block
1304 BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
1305 toReplace->getNormalDest(), invoke->getParent(), P);
1307 IP = &*(normalDest->getFirstInsertionPt());
1308 Builder.SetInsertPoint(IP);
1310 // gc relocates will be generated later as if it were regular call
1313 llvm_unreachable("unexpect type of CallSite");
1317 // Take the name of the original value call if it had one.
1318 token->takeName(CS.getInstruction());
1320 // The GCResult is already inserted, we just need to find it
1322 Instruction *toReplace = CS.getInstruction();
1323 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1324 "only valid use before rewrite is gc.result");
1325 if (toReplace->hasOneUse()) {
1326 Instruction *GCResult = cast<Instruction>(*toReplace->user_begin());
1327 assert(isGCResult(GCResult));
1331 // Update the gc.result of the original statepoint (if any) to use the newly
1332 // inserted statepoint. This is safe to do here since the token can't be
1333 // considered a live reference.
1334 CS.getInstruction()->replaceAllUsesWith(token);
1336 result.StatepointToken = token;
1338 // Second, create a gc.relocate for every live variable
1339 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1344 struct name_ordering {
1347 bool operator()(name_ordering const &a, name_ordering const &b) {
1348 return -1 == a.derived->getName().compare(b.derived->getName());
1352 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1353 SmallVectorImpl<Value *> &livevec) {
1354 assert(basevec.size() == livevec.size());
1356 SmallVector<name_ordering, 64> temp;
1357 for (size_t i = 0; i < basevec.size(); i++) {
1359 v.base = basevec[i];
1360 v.derived = livevec[i];
1363 std::sort(temp.begin(), temp.end(), name_ordering());
1364 for (size_t i = 0; i < basevec.size(); i++) {
1365 basevec[i] = temp[i].base;
1366 livevec[i] = temp[i].derived;
1370 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1371 // which make the relocations happening at this safepoint explicit.
1373 // WARNING: Does not do any fixup to adjust users of the original live
1374 // values. That's the callers responsibility.
1376 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1377 PartiallyConstructedSafepointRecord &result) {
1378 auto liveset = result.liveset;
1379 auto PointerToBase = result.PointerToBase;
1381 // Convert to vector for efficient cross referencing.
1382 SmallVector<Value *, 64> basevec, livevec;
1383 livevec.reserve(liveset.size());
1384 basevec.reserve(liveset.size());
1385 for (Value *L : liveset) {
1386 livevec.push_back(L);
1388 assert(PointerToBase.find(L) != PointerToBase.end());
1389 Value *base = PointerToBase[L];
1390 basevec.push_back(base);
1392 assert(livevec.size() == basevec.size());
1394 // To make the output IR slightly more stable (for use in diffs), ensure a
1395 // fixed order of the values in the safepoint (by sorting the value name).
1396 // The order is otherwise meaningless.
1397 stablize_order(basevec, livevec);
1399 // Do the actual rewriting and delete the old statepoint
1400 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1401 CS.getInstruction()->eraseFromParent();
1404 // Helper function for the relocationViaAlloca.
1405 // It receives iterator to the statepoint gc relocates and emits store to the
1407 // location (via allocaMap) for the each one of them.
1408 // Add visited values into the visitedLiveValues set we will later use them
1409 // for sanity check.
1411 insertRelocationStores(iterator_range<Value::user_iterator> gcRelocs,
1412 DenseMap<Value *, Value *> &allocaMap,
1413 DenseSet<Value *> &visitedLiveValues) {
1415 for (User *U : gcRelocs) {
1416 if (!isa<IntrinsicInst>(U))
1419 IntrinsicInst *relocatedValue = cast<IntrinsicInst>(U);
1421 // We only care about relocates
1422 if (relocatedValue->getIntrinsicID() !=
1423 Intrinsic::experimental_gc_relocate) {
1427 GCRelocateOperands relocateOperands(relocatedValue);
1428 Value *originalValue = const_cast<Value *>(relocateOperands.derivedPtr());
1429 assert(allocaMap.count(originalValue));
1430 Value *alloca = allocaMap[originalValue];
1432 // Emit store into the related alloca
1433 StoreInst *store = new StoreInst(relocatedValue, alloca);
1434 store->insertAfter(relocatedValue);
1437 visitedLiveValues.insert(originalValue);
1442 /// do all the relocation update via allocas and mem2reg
1443 static void relocationViaAlloca(
1444 Function &F, DominatorTree &DT, ArrayRef<Value *> live,
1445 ArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1447 int initialAllocaNum = 0;
1449 // record initial number of allocas
1450 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1452 if (isa<AllocaInst>(*itr))
1457 // TODO-PERF: change data structures, reserve
1458 DenseMap<Value *, Value *> allocaMap;
1459 SmallVector<AllocaInst *, 200> PromotableAllocas;
1460 PromotableAllocas.reserve(live.size());
1462 // emit alloca for each live gc pointer
1463 for (unsigned i = 0; i < live.size(); i++) {
1464 Value *liveValue = live[i];
1465 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1466 F.getEntryBlock().getFirstNonPHI());
1467 allocaMap[liveValue] = alloca;
1468 PromotableAllocas.push_back(alloca);
1471 // The next two loops are part of the same conceptual operation. We need to
1472 // insert a store to the alloca after the original def and at each
1473 // redefinition. We need to insert a load before each use. These are split
1474 // into distinct loops for performance reasons.
1476 // update gc pointer after each statepoint
1477 // either store a relocated value or null (if no relocated value found for
1478 // this gc pointer and it is not a gc_result)
1479 // this must happen before we update the statepoint with load of alloca
1480 // otherwise we lose the link between statepoint and old def
1481 for (size_t i = 0; i < records.size(); i++) {
1482 const struct PartiallyConstructedSafepointRecord &info = records[i];
1483 Value *Statepoint = info.StatepointToken;
1485 // This will be used for consistency check
1486 DenseSet<Value *> visitedLiveValues;
1488 // Insert stores for normal statepoint gc relocates
1489 insertRelocationStores(Statepoint->users(), allocaMap, visitedLiveValues);
1491 // In case if it was invoke statepoint
1492 // we will insert stores for exceptional path gc relocates.
1493 if (isa<InvokeInst>(Statepoint)) {
1494 insertRelocationStores(info.UnwindToken->users(),
1495 allocaMap, visitedLiveValues);
1499 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1500 // the gc.statepoint. This will turn some subtle GC problems into slightly
1501 // easier to debug SEGVs
1502 SmallVector<AllocaInst *, 64> ToClobber;
1503 for (auto Pair : allocaMap) {
1504 Value *Def = Pair.first;
1505 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1507 // This value was relocated
1508 if (visitedLiveValues.count(Def)) {
1511 ToClobber.push_back(Alloca);
1514 auto InsertClobbersAt = [&](Instruction *IP) {
1515 for (auto *AI : ToClobber) {
1516 auto AIType = cast<PointerType>(AI->getType());
1517 auto PT = cast<PointerType>(AIType->getElementType());
1518 Constant *CPN = ConstantPointerNull::get(PT);
1519 StoreInst *store = new StoreInst(CPN, AI);
1520 store->insertBefore(IP);
1524 // Insert the clobbering stores. These may get intermixed with the
1525 // gc.results and gc.relocates, but that's fine.
1526 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1527 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1528 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1529 } else if (auto CI = dyn_cast<CallInst>(Statepoint)) {
1530 BasicBlock::iterator Next(CI);
1532 InsertClobbersAt(Next);
1534 llvm_unreachable("illegal statepoint instruction type?");
1537 // update use with load allocas and add store for gc_relocated
1538 for (auto Pair : allocaMap) {
1539 Value *def = Pair.first;
1540 Value *alloca = Pair.second;
1542 // we pre-record the uses of allocas so that we dont have to worry about
1544 // that change the user information.
1545 SmallVector<Instruction *, 20> uses;
1546 // PERF: trade a linear scan for repeated reallocation
1547 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1548 for (User *U : def->users()) {
1549 if (!isa<ConstantExpr>(U)) {
1550 // If the def has a ConstantExpr use, then the def is either a
1551 // ConstantExpr use itself or null. In either case
1552 // (recursively in the first, directly in the second), the oop
1553 // it is ultimately dependent on is null and this particular
1554 // use does not need to be fixed up.
1555 uses.push_back(cast<Instruction>(U));
1559 std::sort(uses.begin(), uses.end());
1560 auto last = std::unique(uses.begin(), uses.end());
1561 uses.erase(last, uses.end());
1563 for (Instruction *use : uses) {
1564 if (isa<PHINode>(use)) {
1565 PHINode *phi = cast<PHINode>(use);
1566 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1567 if (def == phi->getIncomingValue(i)) {
1568 LoadInst *load = new LoadInst(
1569 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1570 phi->setIncomingValue(i, load);
1574 LoadInst *load = new LoadInst(alloca, "", use);
1575 use->replaceUsesOfWith(def, load);
1579 // emit store for the initial gc value
1580 // store must be inserted after load, otherwise store will be in alloca's
1581 // use list and an extra load will be inserted before it
1582 StoreInst *store = new StoreInst(def, alloca);
1583 if (isa<Instruction>(def)) {
1584 store->insertAfter(cast<Instruction>(def));
1586 assert((isa<Argument>(def) || isa<GlobalVariable>(def) ||
1587 (isa<Constant>(def) && cast<Constant>(def)->isNullValue())) &&
1588 "Must be argument or global");
1589 store->insertAfter(cast<Instruction>(alloca));
1593 assert(PromotableAllocas.size() == live.size() &&
1594 "we must have the same allocas with lives");
1595 if (!PromotableAllocas.empty()) {
1596 // apply mem2reg to promote alloca to SSA
1597 PromoteMemToReg(PromotableAllocas, DT);
1601 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1603 if (isa<AllocaInst>(*itr))
1606 assert(initialAllocaNum == 0 && "We must not introduce any extra allocas");
1610 /// Implement a unique function which doesn't require we sort the input
1611 /// vector. Doing so has the effect of changing the output of a couple of
1612 /// tests in ways which make them less useful in testing fused safepoints.
1613 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1615 SmallVector<T, 128> TempVec;
1616 TempVec.reserve(Vec.size());
1617 for (auto Element : Vec)
1618 TempVec.push_back(Element);
1620 for (auto V : TempVec) {
1621 if (Seen.insert(V).second) {
1627 static Function *getUseHolder(Module &M) {
1628 FunctionType *ftype =
1629 FunctionType::get(Type::getVoidTy(M.getContext()), true);
1630 Function *Func = cast<Function>(M.getOrInsertFunction("__tmp_use", ftype));
1634 /// Insert holders so that each Value is obviously live through the entire
1635 /// liftetime of the call.
1636 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1637 SmallVectorImpl<CallInst *> &holders) {
1638 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1639 Function *Func = getUseHolder(*M);
1641 // For call safepoints insert dummy calls right after safepoint
1642 BasicBlock::iterator next(CS.getInstruction());
1644 CallInst *base_holder = CallInst::Create(Func, Values, "", next);
1645 holders.push_back(base_holder);
1646 } else if (CS.isInvoke()) {
1647 // For invoke safepooints insert dummy calls both in normal and
1648 // exceptional destination blocks
1649 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1650 CallInst *normal_holder = CallInst::Create(
1651 Func, Values, "", invoke->getNormalDest()->getFirstInsertionPt());
1652 CallInst *unwind_holder = CallInst::Create(
1653 Func, Values, "", invoke->getUnwindDest()->getFirstInsertionPt());
1654 holders.push_back(normal_holder);
1655 holders.push_back(unwind_holder);
1657 llvm_unreachable("unsupported call type");
1660 static void findLiveReferences(
1661 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1662 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1663 for (size_t i = 0; i < records.size(); i++) {
1664 struct PartiallyConstructedSafepointRecord &info = records[i];
1665 const CallSite &CS = toUpdate[i];
1666 analyzeParsePointLiveness(DT, CS, info);
1670 static void addBasesAsLiveValues(std::set<Value *> &liveset,
1671 DenseMap<Value *, Value *> &PointerToBase) {
1672 // Identify any base pointers which are used in this safepoint, but not
1673 // themselves relocated. We need to relocate them so that later inserted
1674 // safepoints can get the properly relocated base register.
1675 DenseSet<Value *> missing;
1676 for (Value *L : liveset) {
1677 assert(PointerToBase.find(L) != PointerToBase.end());
1678 Value *base = PointerToBase[L];
1680 if (liveset.find(base) == liveset.end()) {
1681 assert(PointerToBase.find(base) == PointerToBase.end());
1682 // uniqued by set insert
1683 missing.insert(base);
1687 // Note that we want these at the end of the list, otherwise
1688 // register placement gets screwed up once we lower to STATEPOINT
1689 // instructions. This is an utter hack, but there doesn't seem to be a
1691 for (Value *base : missing) {
1693 liveset.insert(base);
1694 PointerToBase[base] = base;
1696 assert(liveset.size() == PointerToBase.size());
1699 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1700 SmallVectorImpl<CallSite> &toUpdate) {
1702 // sanity check the input
1703 std::set<CallSite> uniqued;
1704 uniqued.insert(toUpdate.begin(), toUpdate.end());
1705 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1707 for (size_t i = 0; i < toUpdate.size(); i++) {
1708 CallSite &CS = toUpdate[i];
1709 assert(CS.getInstruction()->getParent()->getParent() == &F);
1710 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1714 // A list of dummy calls added to the IR to keep various values obviously
1715 // live in the IR. We'll remove all of these when done.
1716 SmallVector<CallInst *, 64> holders;
1718 // Insert a dummy call with all of the arguments to the vm_state we'll need
1719 // for the actual safepoint insertion. This ensures reference arguments in
1720 // the deopt argument list are considered live through the safepoint (and
1721 // thus makes sure they get relocated.)
1722 for (size_t i = 0; i < toUpdate.size(); i++) {
1723 CallSite &CS = toUpdate[i];
1724 Statepoint StatepointCS(CS);
1726 SmallVector<Value *, 64> DeoptValues;
1727 for (Use &U : StatepointCS.vm_state_args()) {
1728 Value *Arg = cast<Value>(&U);
1729 if (isGCPointerType(Arg->getType()))
1730 DeoptValues.push_back(Arg);
1732 insertUseHolderAfter(CS, DeoptValues, holders);
1735 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
1736 records.reserve(toUpdate.size());
1737 for (size_t i = 0; i < toUpdate.size(); i++) {
1738 struct PartiallyConstructedSafepointRecord info;
1739 records.push_back(info);
1741 assert(records.size() == toUpdate.size());
1743 // A) Identify all gc pointers which are staticly live at the given call
1745 findLiveReferences(F, DT, P, toUpdate, records);
1747 // B) Find the base pointers for each live pointer
1748 /* scope for caching */ {
1749 // Cache the 'defining value' relation used in the computation and
1750 // insertion of base phis and selects. This ensures that we don't insert
1751 // large numbers of duplicate base_phis.
1752 DefiningValueMapTy DVCache;
1754 for (size_t i = 0; i < records.size(); i++) {
1755 struct PartiallyConstructedSafepointRecord &info = records[i];
1756 CallSite &CS = toUpdate[i];
1757 findBasePointers(DT, DVCache, CS, info);
1759 } // end of cache scope
1761 // The base phi insertion logic (for any safepoint) may have inserted new
1762 // instructions which are now live at some safepoint. The simplest such
1765 // phi a <-- will be a new base_phi here
1766 // safepoint 1 <-- that needs to be live here
1770 std::set<llvm::Value *> allInsertedDefs;
1771 for (size_t i = 0; i < records.size(); i++) {
1772 struct PartiallyConstructedSafepointRecord &info = records[i];
1773 allInsertedDefs.insert(info.NewInsertedDefs.begin(),
1774 info.NewInsertedDefs.end());
1777 // We insert some dummy calls after each safepoint to definitely hold live
1778 // the base pointers which were identified for that safepoint. We'll then
1779 // ask liveness for _every_ base inserted to see what is now live. Then we
1780 // remove the dummy calls.
1781 holders.reserve(holders.size() + records.size());
1782 for (size_t i = 0; i < records.size(); i++) {
1783 struct PartiallyConstructedSafepointRecord &info = records[i];
1784 CallSite &CS = toUpdate[i];
1786 SmallVector<Value *, 128> Bases;
1787 for (auto Pair : info.PointerToBase) {
1788 Bases.push_back(Pair.second);
1790 insertUseHolderAfter(CS, Bases, holders);
1793 // Add the bases explicitly to the live vector set. This may result in a few
1794 // extra relocations, but the base has to be available whenever a pointer
1795 // derived from it is used. Thus, we need it to be part of the statepoint's
1796 // gc arguments list. TODO: Introduce an explicit notion (in the following
1797 // code) of the GC argument list as seperate from the live Values at a
1798 // given statepoint.
1799 for (size_t i = 0; i < records.size(); i++) {
1800 struct PartiallyConstructedSafepointRecord &info = records[i];
1801 addBasesAsLiveValues(info.liveset, info.PointerToBase);
1804 // If we inserted any new values, we need to adjust our notion of what is
1805 // live at a particular safepoint.
1806 if (!allInsertedDefs.empty()) {
1807 fixupLiveReferences(F, DT, P, allInsertedDefs, toUpdate, records);
1809 if (PrintBasePointers) {
1810 for (size_t i = 0; i < records.size(); i++) {
1811 struct PartiallyConstructedSafepointRecord &info = records[i];
1812 errs() << "Base Pairs: (w/Relocation)\n";
1813 for (auto Pair : info.PointerToBase) {
1814 errs() << " derived %" << Pair.first->getName() << " base %"
1815 << Pair.second->getName() << "\n";
1819 for (size_t i = 0; i < holders.size(); i++) {
1820 holders[i]->eraseFromParent();
1821 holders[i] = nullptr;
1825 // Now run through and replace the existing statepoints with new ones with
1826 // the live variables listed. We do not yet update uses of the values being
1827 // relocated. We have references to live variables that need to
1828 // survive to the last iteration of this loop. (By construction, the
1829 // previous statepoint can not be a live variable, thus we can and remove
1830 // the old statepoint calls as we go.)
1831 for (size_t i = 0; i < records.size(); i++) {
1832 struct PartiallyConstructedSafepointRecord &info = records[i];
1833 CallSite &CS = toUpdate[i];
1834 makeStatepointExplicit(DT, CS, P, info);
1836 toUpdate.clear(); // prevent accident use of invalid CallSites
1838 // In case if we inserted relocates in a different basic block than the
1839 // original safepoint (this can happen for invokes). We need to be sure that
1840 // original values were not used in any of the phi nodes at the
1841 // beginning of basic block containing them. Because we know that all such
1842 // blocks will have single predecessor we can safely assume that all phi
1843 // nodes have single entry (because of normalizeBBForInvokeSafepoint).
1844 // Just remove them all here.
1845 for (size_t i = 0; i < records.size(); i++) {
1846 Instruction *I = records[i].StatepointToken;
1848 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
1849 FoldSingleEntryPHINodes(invoke->getNormalDest());
1850 assert(!isa<PHINode>(invoke->getNormalDest()->begin()));
1852 FoldSingleEntryPHINodes(invoke->getUnwindDest());
1853 assert(!isa<PHINode>(invoke->getUnwindDest()->begin()));
1857 // Do all the fixups of the original live variables to their relocated selves
1858 SmallVector<Value *, 128> live;
1859 for (size_t i = 0; i < records.size(); i++) {
1860 struct PartiallyConstructedSafepointRecord &info = records[i];
1861 // We can't simply save the live set from the original insertion. One of
1862 // the live values might be the result of a call which needs a safepoint.
1863 // That Value* no longer exists and we need to use the new gc_result.
1864 // Thankfully, the liveset is embedded in the statepoint (and updated), so
1865 // we just grab that.
1866 Statepoint statepoint(info.StatepointToken);
1867 live.insert(live.end(), statepoint.gc_args_begin(),
1868 statepoint.gc_args_end());
1870 unique_unsorted(live);
1874 for (auto ptr : live) {
1875 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
1879 relocationViaAlloca(F, DT, live, records);
1880 return !records.empty();
1883 /// Returns true if this function should be rewritten by this pass. The main
1884 /// point of this function is as an extension point for custom logic.
1885 static bool shouldRewriteStatepointsIn(Function &F) {
1886 // TODO: This should check the GCStrategy
1888 const std::string StatepointExampleName("statepoint-example");
1889 return StatepointExampleName == F.getGC();
1894 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
1895 // Nothing to do for declarations.
1896 if (F.isDeclaration() || F.empty())
1899 // Policy choice says not to rewrite - the most common reason is that we're
1900 // compiling code without a GCStrategy.
1901 if (!shouldRewriteStatepointsIn(F))
1904 // Gather all the statepoints which need rewritten.
1905 SmallVector<CallSite, 64> ParsePointNeeded;
1906 for (Instruction &I : inst_range(F)) {
1907 // TODO: only the ones with the flag set!
1908 if (isStatepoint(I))
1909 ParsePointNeeded.push_back(CallSite(&I));
1912 // Return early if no work to do.
1913 if (ParsePointNeeded.empty())
1916 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1917 return insertParsePoints(F, DT, this, ParsePointNeeded);