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));
58 struct RewriteStatepointsForGC : public FunctionPass {
59 static char ID; // Pass identification, replacement for typeid
61 RewriteStatepointsForGC() : FunctionPass(ID) {
62 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
64 bool runOnFunction(Function &F) override;
66 void getAnalysisUsage(AnalysisUsage &AU) const override {
67 // We add and rewrite a bunch of instructions, but don't really do much
68 // else. We could in theory preserve a lot more analyses here.
69 AU.addRequired<DominatorTreeWrapperPass>();
73 char RewriteStatepointsForGC::ID = 0;
75 FunctionPass *llvm::createRewriteStatepointsForGCPass() {
76 return new RewriteStatepointsForGC();
79 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
80 "Make relocations explicit at statepoints", false, false)
81 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
82 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
83 "Make relocations explicit at statepoints", false, false)
86 // The type of the internal cache used inside the findBasePointers family
87 // of functions. From the callers perspective, this is an opaque type and
88 // should not be inspected.
90 // In the actual implementation this caches two relations:
91 // - The base relation itself (i.e. this pointer is based on that one)
92 // - The base defining value relation (i.e. before base_phi insertion)
93 // Generally, after the execution of a full findBasePointer call, only the
94 // base relation will remain. Internally, we add a mixture of the two
95 // types, then update all the second type to the first type
96 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
149 static bool isAggWhichContainsGCPtrType(Type *Ty) {
150 if (VectorType *VT = dyn_cast<VectorType>(Ty))
151 return isGCPointerType(VT->getScalarType());
152 else if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
153 return isGCPointerType(AT->getElementType()) ||
154 isAggWhichContainsGCPtrType(AT->getElementType());
155 } else if (StructType *ST = dyn_cast<StructType>(Ty)) {
156 bool UnsupportedType = false;
157 for (Type *SubType : ST->subtypes())
158 UnsupportedType |= isGCPointerType(SubType) || isAggWhichContainsGCPtrType(SubType);
159 return UnsupportedType;
164 // Conservatively identifies any definitions which might be live at the
165 // given instruction. The analysis is performed immediately before the
166 // given instruction. Values defined by that instruction are not considered
167 // live. Values used by that instruction are considered live.
169 // preconditions: valid IR graph, term is either a terminator instruction or
170 // a call instruction, pred is the basic block of term, DT, LI are valid
172 // side effects: none, does not mutate IR
174 // postconditions: populates liveValues as discussed above
175 static void findLiveGCValuesAtInst(Instruction *term, BasicBlock *pred,
176 DominatorTree &DT, LoopInfo *LI,
177 std::set<llvm::Value *> &liveValues) {
180 assert(isa<CallInst>(term) || isa<InvokeInst>(term) || term->isTerminator());
182 Function *F = pred->getParent();
184 auto is_live_gc_reference =
185 [&](Value &V) { return isLiveGCReferenceAt(V, term, DT, LI); };
187 // Are there any gc pointer arguments live over this point? This needs to be
188 // special cased since arguments aren't defined in basic blocks.
189 for (Argument &arg : F->args()) {
190 assert(!isAggWhichContainsGCPtrType(arg.getType()) &&
191 "support for FCA unimplemented");
193 if (is_live_gc_reference(arg)) {
194 liveValues.insert(&arg);
198 // Walk through all dominating blocks - the ones which can contain
199 // definitions used in this block - and check to see if any of the values
200 // they define are used in locations potentially reachable from the
201 // interesting instruction.
202 BasicBlock *BBI = pred;
205 errs() << "[LSP] Looking at dominating block " << pred->getName() << "\n";
207 assert(DT.dominates(BBI, pred));
208 assert(isPotentiallyReachable(BBI, pred, &DT) &&
209 "dominated block must be reachable");
211 // Walk through the instructions in dominating blocks and keep any
212 // that have a use potentially reachable from the block we're
213 // considering putting the safepoint in
214 for (Instruction &inst : *BBI) {
216 errs() << "[LSP] Looking at instruction ";
220 if (pred == BBI && (&inst) == term) {
222 errs() << "[LSP] stopped because we encountered the safepoint "
226 // If we're in the block which defines the interesting instruction,
227 // we don't want to include any values as live which are defined
228 // _after_ the interesting line or as part of the line itself
229 // i.e. "term" is the call instruction for a call safepoint, the
230 // results of the call should not be considered live in that stackmap
234 assert(!isAggWhichContainsGCPtrType(inst.getType()) &&
235 "support for FCA unimplemented");
237 if (is_live_gc_reference(inst)) {
239 errs() << "[LSP] found live value for this safepoint ";
243 liveValues.insert(&inst);
246 if (!DT.getNode(BBI)->getIDom()) {
247 assert(BBI == &F->getEntryBlock() &&
248 "failed to find a dominator for something other than "
252 BBI = DT.getNode(BBI)->getIDom()->getBlock();
256 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
257 if (a->hasName() && b->hasName()) {
258 return -1 == a->getName().compare(b->getName());
259 } else if (a->hasName() && !b->hasName()) {
261 } else if (!a->hasName() && b->hasName()) {
264 // Better than nothing, but not stable
269 /// Find the initial live set. Note that due to base pointer
270 /// insertion, the live set may be incomplete.
272 analyzeParsePointLiveness(DominatorTree &DT, const CallSite &CS,
273 PartiallyConstructedSafepointRecord &result) {
274 Instruction *inst = CS.getInstruction();
276 BasicBlock *BB = inst->getParent();
277 std::set<Value *> liveset;
278 findLiveGCValuesAtInst(inst, BB, DT, nullptr, liveset);
281 // Note: This output is used by several of the test cases
282 // The order of elemtns in a set is not stable, put them in a vec and sort
284 std::vector<Value *> temp;
285 temp.insert(temp.end(), liveset.begin(), liveset.end());
286 std::sort(temp.begin(), temp.end(), order_by_name);
287 errs() << "Live Variables:\n";
288 for (Value *V : temp) {
289 errs() << " " << V->getName(); // no newline
293 if (PrintLiveSetSize) {
294 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
295 errs() << "Number live values: " << liveset.size() << "\n";
297 result.liveset = liveset;
300 /// True iff this value is the null pointer constant (of any pointer type)
301 static bool isNullConstant(Value *V) {
302 return isa<Constant>(V) && isa<PointerType>(V->getType()) &&
303 cast<Constant>(V)->isNullValue();
306 /// Helper function for findBasePointer - Will return a value which either a)
307 /// defines the base pointer for the input or b) blocks the simple search
308 /// (i.e. a PHI or Select of two derived pointers)
309 static Value *findBaseDefiningValue(Value *I) {
310 assert(I->getType()->isPointerTy() &&
311 "Illegal to ask for the base pointer of a non-pointer type");
313 // There are instructions which can never return gc pointer values. Sanity
315 // that this is actually true.
316 assert(!isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
317 !isa<ShuffleVectorInst>(I) && "Vector types are not gc pointers");
318 assert((!isa<Instruction>(I) || isa<InvokeInst>(I) ||
319 !cast<Instruction>(I)->isTerminator()) &&
320 "With the exception of invoke terminators don't define values");
321 assert(!isa<StoreInst>(I) && !isa<FenceInst>(I) &&
322 "Can't be definitions to start with");
323 assert(!isa<ICmpInst>(I) && !isa<FCmpInst>(I) &&
324 "Comparisons don't give ops");
325 // There's a bunch of instructions which just don't make sense to apply to
326 // a pointer. The only valid reason for this would be pointer bit
327 // twiddling which we're just not going to support.
328 assert((!isa<Instruction>(I) || !cast<Instruction>(I)->isBinaryOp()) &&
329 "Binary ops on pointer values are meaningless. Unless your "
330 "bit-twiddling which we don't support");
332 if (Argument *Arg = dyn_cast<Argument>(I)) {
333 // An incoming argument to the function is a base pointer
334 // We should have never reached here if this argument isn't an gc value
335 assert(Arg->getType()->isPointerTy() &&
336 "Base for pointer must be another pointer");
340 if (GlobalVariable *global = dyn_cast<GlobalVariable>(I)) {
342 assert(global->getType()->isPointerTy() &&
343 "Base for pointer must be another pointer");
347 // inlining could possibly introduce phi node that contains
348 // undef if callee has multiple returns
349 if (UndefValue *undef = dyn_cast<UndefValue>(I)) {
350 assert(undef->getType()->isPointerTy() &&
351 "Base for pointer must be another pointer");
352 return undef; // utterly meaningless, but useful for dealing with
353 // partially optimized code.
356 // Due to inheritance, this must be _after_ the global variable and undef
358 if (Constant *con = dyn_cast<Constant>(I)) {
359 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
360 "order of checks wrong!");
361 // Note: Finding a constant base for something marked for relocation
362 // doesn't really make sense. The most likely case is either a) some
363 // screwed up the address space usage or b) your validating against
364 // compiled C++ code w/o the proper separation. The only real exception
365 // is a null pointer. You could have generic code written to index of
366 // off a potentially null value and have proven it null. We also use
367 // null pointers in dead paths of relocation phis (which we might later
368 // want to find a base pointer for).
369 assert(con->getType()->isPointerTy() &&
370 "Base for pointer must be another pointer");
371 assert(con->isNullValue() && "null is the only case which makes sense");
375 if (CastInst *CI = dyn_cast<CastInst>(I)) {
376 Value *def = CI->stripPointerCasts();
377 assert(def->getType()->isPointerTy() &&
378 "Base for pointer must be another pointer");
379 if (isa<CastInst>(def)) {
380 // If we find a cast instruction here, it means we've found a cast
381 // which is not simply a pointer cast (i.e. an inttoptr). We don't
382 // know how to handle int->ptr conversion.
383 llvm_unreachable("Can not find the base pointers for an inttoptr cast");
385 assert(!isa<CastInst>(def) && "shouldn't find another cast here");
386 return findBaseDefiningValue(def);
389 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
390 if (LI->getType()->isPointerTy()) {
391 Value *Op = LI->getOperand(0);
393 // Has to be a pointer to an gc object, or possibly an array of such?
394 assert(Op->getType()->isPointerTy());
395 return LI; // The value loaded is an gc base itself
398 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
399 Value *Op = GEP->getOperand(0);
400 if (Op->getType()->isPointerTy()) {
401 return findBaseDefiningValue(Op); // The base of this GEP is the base
405 if (AllocaInst *alloc = dyn_cast<AllocaInst>(I)) {
406 // An alloca represents a conceptual stack slot. It's the slot itself
407 // that the GC needs to know about, not the value in the slot.
408 assert(alloc->getType()->isPointerTy() &&
409 "Base for pointer must be another pointer");
413 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
414 switch (II->getIntrinsicID()) {
416 // fall through to general call handling
418 case Intrinsic::experimental_gc_statepoint:
419 case Intrinsic::experimental_gc_result_float:
420 case Intrinsic::experimental_gc_result_int:
421 llvm_unreachable("these don't produce pointers");
422 case Intrinsic::experimental_gc_result_ptr:
423 // This is just a special case of the CallInst check below to handle a
424 // statepoint with deopt args which hasn't been rewritten for GC yet.
425 // TODO: Assert that the statepoint isn't rewritten yet.
427 case Intrinsic::experimental_gc_relocate: {
428 // Rerunning safepoint insertion after safepoints are already
429 // inserted is not supported. It could probably be made to work,
430 // but why are you doing this? There's no good reason.
431 llvm_unreachable("repeat safepoint insertion is not supported");
433 case Intrinsic::gcroot:
434 // Currently, this mechanism hasn't been extended to work with gcroot.
435 // There's no reason it couldn't be, but I haven't thought about the
436 // implications much.
438 "interaction with the gcroot mechanism is not supported");
441 // We assume that functions in the source language only return base
442 // pointers. This should probably be generalized via attributes to support
443 // both source language and internal functions.
444 if (CallInst *call = dyn_cast<CallInst>(I)) {
445 assert(call->getType()->isPointerTy() &&
446 "Base for pointer must be another pointer");
449 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
450 assert(invoke->getType()->isPointerTy() &&
451 "Base for pointer must be another pointer");
455 // I have absolutely no idea how to implement this part yet. It's not
456 // neccessarily hard, I just haven't really looked at it yet.
457 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
459 if (AtomicCmpXchgInst *cas = dyn_cast<AtomicCmpXchgInst>(I)) {
460 // A CAS is effectively a atomic store and load combined under a
461 // predicate. From the perspective of base pointers, we just treat it
462 // like a load. We loaded a pointer from a address in memory, that value
463 // had better be a valid base pointer.
464 return cas->getPointerOperand();
466 if (AtomicRMWInst *atomic = dyn_cast<AtomicRMWInst>(I)) {
467 assert(AtomicRMWInst::Xchg == atomic->getOperation() &&
468 "All others are binary ops which don't apply to base pointers");
469 // semantically, a load, store pair. Treat it the same as a standard load
470 return atomic->getPointerOperand();
473 // The aggregate ops. Aggregates can either be in the heap or on the
474 // stack, but in either case, this is simply a field load. As a result,
475 // this is a defining definition of the base just like a load is.
476 if (ExtractValueInst *ev = dyn_cast<ExtractValueInst>(I)) {
480 // We should never see an insert vector since that would require we be
481 // tracing back a struct value not a pointer value.
482 assert(!isa<InsertValueInst>(I) &&
483 "Base pointer for a struct is meaningless");
485 // The last two cases here don't return a base pointer. Instead, they
486 // return a value which dynamically selects from amoung several base
487 // derived pointers (each with it's own base potentially). It's the job of
488 // the caller to resolve these.
489 if (SelectInst *select = dyn_cast<SelectInst>(I)) {
492 if (PHINode *phi = dyn_cast<PHINode>(I)) {
496 errs() << "unknown type: " << *I << "\n";
497 llvm_unreachable("unknown type");
501 /// Returns the base defining value for this value.
502 Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &cache) {
503 if (cache.find(I) == cache.end()) {
504 cache[I] = findBaseDefiningValue(I);
506 assert(cache.find(I) != cache.end());
509 errs() << "fBDV-cached: " << I->getName() << " -> " << cache[I]->getName()
515 /// Return a base pointer for this value if known. Otherwise, return it's
516 /// base defining value.
517 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &cache) {
518 Value *def = findBaseDefiningValueCached(I, cache);
519 if (cache.count(def)) {
520 // Either a base-of relation, or a self reference. Caller must check.
523 // Only a BDV available
527 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
528 /// is it known to be a base pointer? Or do we need to continue searching.
529 static bool isKnownBaseResult(Value *v) {
530 if (!isa<PHINode>(v) && !isa<SelectInst>(v)) {
531 // no recursion possible
534 if (cast<Instruction>(v)->getMetadata("is_base_value")) {
535 // This is a previously inserted base phi or select. We know
536 // that this is a base value.
540 // We need to keep searching
544 // TODO: find a better name for this
548 enum Status { Unknown, Base, Conflict };
550 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
551 assert(status != Base || b);
553 PhiState(Value *b) : status(Base), base(b) {}
554 PhiState() : status(Unknown), base(nullptr) {}
555 PhiState(const PhiState &other) : status(other.status), base(other.base) {
556 assert(status != Base || base);
559 Status getStatus() const { return status; }
560 Value *getBase() const { return base; }
562 bool isBase() const { return getStatus() == Base; }
563 bool isUnknown() const { return getStatus() == Unknown; }
564 bool isConflict() const { return getStatus() == Conflict; }
566 bool operator==(const PhiState &other) const {
567 return base == other.base && status == other.status;
570 bool operator!=(const PhiState &other) const { return !(*this == other); }
573 errs() << status << " (" << base << " - "
574 << (base ? base->getName() : "nullptr") << "): ";
579 Value *base; // non null only if status == base
582 // Values of type PhiState form a lattice, and this is a helper
583 // class that implementes the meet operation. The meat of the meet
584 // operation is implemented in MeetPhiStates::pureMeet
585 class MeetPhiStates {
587 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
588 explicit MeetPhiStates(const std::map<Value *, PhiState> &phiStates)
589 : phiStates(phiStates) {}
591 // Destructively meet the current result with the base V. V can
592 // either be a merge instruction (SelectInst / PHINode), in which
593 // case its status is looked up in the phiStates map; or a regular
594 // SSA value, in which case it is assumed to be a base.
595 void meetWith(Value *V) {
596 PhiState otherState = getStateForBDV(V);
597 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
598 MeetPhiStates::pureMeet(currentResult, otherState)) &&
599 "math is wrong: meet does not commute!");
600 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
603 PhiState getResult() const { return currentResult; }
606 const std::map<Value *, PhiState> &phiStates;
607 PhiState currentResult;
609 /// Return a phi state for a base defining value. We'll generate a new
610 /// base state for known bases and expect to find a cached state otherwise
611 PhiState getStateForBDV(Value *baseValue) {
612 if (isKnownBaseResult(baseValue)) {
613 return PhiState(baseValue);
615 return lookupFromMap(baseValue);
619 PhiState lookupFromMap(Value *V) {
620 auto I = phiStates.find(V);
621 assert(I != phiStates.end() && "lookup failed!");
625 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
626 switch (stateA.getStatus()) {
627 case PhiState::Unknown:
631 assert(stateA.getBase() && "can't be null");
632 if (stateB.isUnknown()) {
634 } else if (stateB.isBase()) {
635 if (stateA.getBase() == stateB.getBase()) {
636 assert(stateA == stateB && "equality broken!");
639 return PhiState(PhiState::Conflict);
641 assert(stateB.isConflict() && "only three states!");
642 return PhiState(PhiState::Conflict);
645 case PhiState::Conflict:
648 assert(false && "only three states!");
652 /// For a given value or instruction, figure out what base ptr it's derived
653 /// from. For gc objects, this is simply itself. On success, returns a value
654 /// which is the base pointer. (This is reliable and can be used for
655 /// relocation.) On failure, returns nullptr.
656 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache,
657 std::set<llvm::Value *> &newInsertedDefs) {
658 Value *def = findBaseOrBDV(I, cache);
660 if (isKnownBaseResult(def)) {
664 // Here's the rough algorithm:
665 // - For every SSA value, construct a mapping to either an actual base
666 // pointer or a PHI which obscures the base pointer.
667 // - Construct a mapping from PHI to unknown TOP state. Use an
668 // optimistic algorithm to propagate base pointer information. Lattice
673 // When algorithm terminates, all PHIs will either have a single concrete
674 // base or be in a conflict state.
675 // - For every conflict, insert a dummy PHI node without arguments. Add
676 // these to the base[Instruction] = BasePtr mapping. For every
677 // non-conflict, add the actual base.
678 // - For every conflict, add arguments for the base[a] of each input
681 // Note: A simpler form of this would be to add the conflict form of all
682 // PHIs without running the optimistic algorithm. This would be
683 // analougous to pessimistic data flow and would likely lead to an
684 // overall worse solution.
686 std::map<Value *, PhiState> states;
687 states[def] = PhiState();
688 // Recursively fill in all phis & selects reachable from the initial one
689 // for which we don't already know a definite base value for
690 // PERF: Yes, this is as horribly inefficient as it looks.
694 for (auto Pair : states) {
695 Value *v = Pair.first;
696 assert(!isKnownBaseResult(v) && "why did it get added?");
697 if (PHINode *phi = dyn_cast<PHINode>(v)) {
698 unsigned NumPHIValues = phi->getNumIncomingValues();
699 assert(NumPHIValues > 0 && "zero input phis are illegal");
700 for (unsigned i = 0; i != NumPHIValues; ++i) {
701 Value *InVal = phi->getIncomingValue(i);
702 Value *local = findBaseOrBDV(InVal, cache);
703 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
704 states[local] = PhiState();
708 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
709 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
710 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
711 states[local] = PhiState();
714 local = findBaseOrBDV(sel->getFalseValue(), cache);
715 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
716 states[local] = PhiState();
724 errs() << "States after initialization:\n";
725 for (auto Pair : states) {
726 Instruction *v = cast<Instruction>(Pair.first);
727 PhiState state = Pair.second;
733 // TODO: come back and revisit the state transitions around inputs which
734 // have reached conflict state. The current version seems too conservative.
736 bool progress = true;
739 oldSize = states.size();
741 for (auto Pair : states) {
742 MeetPhiStates calculateMeet(states);
743 Value *v = Pair.first;
744 assert(!isKnownBaseResult(v) && "why did it get added?");
745 assert(isa<SelectInst>(v) || isa<PHINode>(v));
746 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
747 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
748 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
749 } else if (PHINode *phi = dyn_cast<PHINode>(v)) {
750 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
751 calculateMeet.meetWith(
752 findBaseOrBDV(phi->getIncomingValue(i), cache));
755 llvm_unreachable("no such state expected");
758 PhiState oldState = states[v];
759 PhiState newState = calculateMeet.getResult();
760 if (oldState != newState) {
762 states[v] = newState;
766 assert(oldSize <= states.size());
767 assert(oldSize == states.size() || progress);
771 errs() << "States after meet iteration:\n";
772 for (auto Pair : states) {
773 Instruction *v = cast<Instruction>(Pair.first);
774 PhiState state = Pair.second;
780 // Insert Phis for all conflicts
781 for (auto Pair : states) {
782 Instruction *v = cast<Instruction>(Pair.first);
783 PhiState state = Pair.second;
784 assert(!isKnownBaseResult(v) && "why did it get added?");
785 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
786 if (state.isConflict()) {
787 if (isa<PHINode>(v)) {
789 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
790 assert(num_preds > 0 && "how did we reach here");
791 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
792 newInsertedDefs.insert(phi);
793 // Add metadata marking this as a base value
794 auto *const_1 = ConstantInt::get(
796 v->getParent()->getParent()->getParent()->getContext()),
798 auto MDConst = ConstantAsMetadata::get(const_1);
799 MDNode *md = MDNode::get(
800 v->getParent()->getParent()->getParent()->getContext(), MDConst);
801 phi->setMetadata("is_base_value", md);
802 states[v] = PhiState(PhiState::Conflict, phi);
803 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
804 // The undef will be replaced later
805 UndefValue *undef = UndefValue::get(sel->getType());
806 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
807 undef, "base_select", sel);
808 newInsertedDefs.insert(basesel);
809 // Add metadata marking this as a base value
810 auto *const_1 = ConstantInt::get(
812 v->getParent()->getParent()->getParent()->getContext()),
814 auto MDConst = ConstantAsMetadata::get(const_1);
815 MDNode *md = MDNode::get(
816 v->getParent()->getParent()->getParent()->getContext(), MDConst);
817 basesel->setMetadata("is_base_value", md);
818 states[v] = PhiState(PhiState::Conflict, basesel);
825 // Fixup all the inputs of the new PHIs
826 for (auto Pair : states) {
827 Instruction *v = cast<Instruction>(Pair.first);
828 PhiState state = Pair.second;
830 assert(!isKnownBaseResult(v) && "why did it get added?");
831 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
832 if (state.isConflict()) {
833 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
834 PHINode *phi = cast<PHINode>(v);
835 unsigned NumPHIValues = phi->getNumIncomingValues();
836 for (unsigned i = 0; i < NumPHIValues; i++) {
837 Value *InVal = phi->getIncomingValue(i);
838 BasicBlock *InBB = phi->getIncomingBlock(i);
840 // If we've already seen InBB, add the same incoming value
841 // we added for it earlier. The IR verifier requires phi
842 // nodes with multiple entries from the same basic block
843 // to have the same incoming value for each of those
844 // entries. If we don't do this check here and basephi
845 // has a different type than base, we'll end up adding two
846 // bitcasts (and hence two distinct values) as incoming
847 // values for the same basic block.
849 int blockIndex = basephi->getBasicBlockIndex(InBB);
850 if (blockIndex != -1) {
851 Value *oldBase = basephi->getIncomingValue(blockIndex);
852 basephi->addIncoming(oldBase, InBB);
854 Value *base = findBaseOrBDV(InVal, cache);
855 if (!isKnownBaseResult(base)) {
856 // Either conflict or base.
857 assert(states.count(base));
858 base = states[base].getBase();
859 assert(base != nullptr && "unknown PhiState!");
860 assert(newInsertedDefs.count(base) &&
861 "should have already added this in a prev. iteration!");
864 // In essense this assert states: the only way two
865 // values incoming from the same basic block may be
866 // different is by being different bitcasts of the same
867 // value. A cleanup that remains TODO is changing
868 // findBaseOrBDV to return an llvm::Value of the correct
869 // type (and still remain pure). This will remove the
870 // need to add bitcasts.
871 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
872 "sanity -- findBaseOrBDV should be pure!");
877 // Find either the defining value for the PHI or the normal base for
879 Value *base = findBaseOrBDV(InVal, cache);
880 if (!isKnownBaseResult(base)) {
881 // Either conflict or base.
882 assert(states.count(base));
883 base = states[base].getBase();
884 assert(base != nullptr && "unknown PhiState!");
886 assert(base && "can't be null");
887 // Must use original input BB since base may not be Instruction
888 // The cast is needed since base traversal may strip away bitcasts
889 if (base->getType() != basephi->getType()) {
890 base = new BitCastInst(base, basephi->getType(), "cast",
891 InBB->getTerminator());
892 newInsertedDefs.insert(base);
894 basephi->addIncoming(base, InBB);
896 assert(basephi->getNumIncomingValues() == NumPHIValues);
897 } else if (SelectInst *basesel = dyn_cast<SelectInst>(state.getBase())) {
898 SelectInst *sel = cast<SelectInst>(v);
899 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
900 // something more safe and less hacky.
901 for (int i = 1; i <= 2; i++) {
902 Value *InVal = sel->getOperand(i);
903 // Find either the defining value for the PHI or the normal base for
905 Value *base = findBaseOrBDV(InVal, cache);
906 if (!isKnownBaseResult(base)) {
907 // Either conflict or base.
908 assert(states.count(base));
909 base = states[base].getBase();
910 assert(base != nullptr && "unknown PhiState!");
912 assert(base && "can't be null");
913 // Must use original input BB since base may not be Instruction
914 // The cast is needed since base traversal may strip away bitcasts
915 if (base->getType() != basesel->getType()) {
916 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
917 newInsertedDefs.insert(base);
919 basesel->setOperand(i, base);
922 assert(false && "unexpected type");
927 // Cache all of our results so we can cheaply reuse them
928 // NOTE: This is actually two caches: one of the base defining value
929 // relation and one of the base pointer relation! FIXME
930 for (auto item : states) {
931 Value *v = item.first;
932 Value *base = item.second.getBase();
934 assert(!isKnownBaseResult(v) && "why did it get added?");
937 std::string fromstr =
938 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
940 errs() << "Updating base value cache"
941 << " for: " << (v->hasName() ? v->getName() : "")
942 << " from: " << fromstr
943 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
946 assert(isKnownBaseResult(base) &&
947 "must be something we 'know' is a base pointer");
948 if (cache.count(v)) {
949 // Once we transition from the BDV relation being store in the cache to
950 // the base relation being stored, it must be stable
951 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
952 "base relation should be stable");
956 assert(cache.find(def) != cache.end());
960 // For a set of live pointers (base and/or derived), identify the base
961 // pointer of the object which they are derived from. This routine will
962 // mutate the IR graph as needed to make the 'base' pointer live at the
963 // definition site of 'derived'. This ensures that any use of 'derived' can
964 // also use 'base'. This may involve the insertion of a number of
965 // additional PHI nodes.
967 // preconditions: live is a set of pointer type Values
969 // side effects: may insert PHI nodes into the existing CFG, will preserve
970 // CFG, will not remove or mutate any existing nodes
972 // post condition: base_pairs contains one (derived, base) pair for every
973 // pointer in live. Note that derived can be equal to base if the original
974 // pointer was a base pointer.
975 static void findBasePointers(const std::set<llvm::Value *> &live,
976 std::map<llvm::Value *, llvm::Value *> &base_pairs,
977 DominatorTree *DT, DefiningValueMapTy &DVCache,
978 std::set<llvm::Value *> &newInsertedDefs) {
979 for (Value *ptr : live) {
980 Value *base = findBasePointer(ptr, DVCache, newInsertedDefs);
981 assert(base && "failed to find base pointer");
982 base_pairs[ptr] = base;
983 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
984 DT->dominates(cast<Instruction>(base)->getParent(),
985 cast<Instruction>(ptr)->getParent())) &&
986 "The base we found better dominate the derived pointer");
988 if (isNullConstant(base))
989 // If you see this trip and like to live really dangerously, the code
990 // should be correct, just with idioms the verifier can't handle. You
991 // can try disabling the verifier at your own substaintial risk.
992 llvm_unreachable("the relocation code needs adjustment to handle the"
993 "relocation of a null pointer constant without causing"
994 "false positives in the safepoint ir verifier.");
998 /// Find the required based pointers (and adjust the live set) for the given
1000 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1002 PartiallyConstructedSafepointRecord &result) {
1003 std::map<llvm::Value *, llvm::Value *> base_pairs;
1004 std::set<llvm::Value *> newInsertedDefs;
1005 findBasePointers(result.liveset, base_pairs, &DT, DVCache, newInsertedDefs);
1007 if (PrintBasePointers) {
1008 errs() << "Base Pairs (w/o Relocation):\n";
1009 for (auto Pair : base_pairs) {
1010 errs() << " derived %" << Pair.first->getName() << " base %"
1011 << Pair.second->getName() << "\n";
1015 result.base_pairs = base_pairs;
1016 result.newInsertedDefs = newInsertedDefs;
1019 /// Check for liveness of items in the insert defs and add them to the live
1020 /// and base pointer sets
1021 static void fixupLiveness(DominatorTree &DT, const CallSite &CS,
1022 const std::set<Value *> &allInsertedDefs,
1023 PartiallyConstructedSafepointRecord &result) {
1024 Instruction *inst = CS.getInstruction();
1026 std::set<llvm::Value *> liveset = result.liveset;
1027 std::map<llvm::Value *, llvm::Value *> base_pairs = result.base_pairs;
1029 auto is_live_gc_reference =
1030 [&](Value &V) { return isLiveGCReferenceAt(V, inst, DT, nullptr); };
1032 // For each new definition, check to see if a) the definition dominates the
1033 // instruction we're interested in, and b) one of the uses of that definition
1034 // is edge-reachable from the instruction we're interested in. This is the
1035 // same definition of liveness we used in the intial liveness analysis
1036 for (Value *newDef : allInsertedDefs) {
1037 if (liveset.count(newDef)) {
1038 // already live, no action needed
1042 // PERF: Use DT to check instruction domination might not be good for
1043 // compilation time, and we could change to optimal solution if this
1044 // turn to be a issue
1045 if (!DT.dominates(cast<Instruction>(newDef), inst)) {
1046 // can't possibly be live at inst
1050 if (is_live_gc_reference(*newDef)) {
1051 // Add the live new defs into liveset and base_pairs
1052 liveset.insert(newDef);
1053 base_pairs[newDef] = newDef;
1057 result.liveset = liveset;
1058 result.base_pairs = base_pairs;
1061 static void fixupLiveReferences(
1062 Function &F, DominatorTree &DT, Pass *P,
1063 const std::set<llvm::Value *> &allInsertedDefs,
1064 std::vector<CallSite> &toUpdate,
1065 std::vector<struct PartiallyConstructedSafepointRecord> &records) {
1066 for (size_t i = 0; i < records.size(); i++) {
1067 struct PartiallyConstructedSafepointRecord &info = records[i];
1068 CallSite &CS = toUpdate[i];
1069 fixupLiveness(DT, CS, allInsertedDefs, info);
1073 // Normalize basic block to make it ready to be target of invoke statepoint.
1074 // It means spliting it to have single predecessor. Return newly created BB
1075 // ready to be successor of invoke statepoint.
1076 static BasicBlock *normalizeBBForInvokeSafepoint(BasicBlock *BB,
1077 BasicBlock *InvokeParent,
1079 BasicBlock *ret = BB;
1081 if (!BB->getUniquePredecessor()) {
1082 ret = SplitBlockPredecessors(BB, InvokeParent, "");
1085 // Another requirement for such basic blocks is to not have any phi nodes.
1086 // Since we just ensured that new BB will have single predecessor,
1087 // all phi nodes in it will have one value. Here it would be naturall place
1089 // remove them all. But we can not do this because we are risking to remove
1090 // one of the values stored in liveset of another statepoint. We will do it
1091 // later after placing all safepoints.
1097 VerifySafepointBounds(const std::pair<Instruction *, Instruction *> &bounds) {
1098 assert(bounds.first->getParent() && bounds.second->getParent() &&
1099 "both must belong to basic blocks");
1100 if (bounds.first->getParent() == bounds.second->getParent()) {
1101 // This is a call safepoint
1102 // TODO: scan the range to find the statepoint
1103 // TODO: check that the following instruction is not a gc_relocate or
1106 // This is an invoke safepoint
1107 InvokeInst *invoke = dyn_cast<InvokeInst>(bounds.first);
1109 assert(invoke && "only continues over invokes!");
1110 assert(invoke->getNormalDest() == bounds.second->getParent() &&
1111 "safepoint should continue into normal exit block");
1115 static int find_index(const SmallVectorImpl<Value *> &livevec, Value *val) {
1116 auto itr = std::find(livevec.begin(), livevec.end(), val);
1117 assert(livevec.end() != itr);
1118 size_t index = std::distance(livevec.begin(), itr);
1119 assert(index < livevec.size());
1123 // Create new attribute set containing only attributes which can be transfered
1124 // from original call to the safepoint.
1125 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1128 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1129 unsigned index = AS.getSlotIndex(Slot);
1131 if (index == AttributeSet::ReturnIndex ||
1132 index == AttributeSet::FunctionIndex) {
1134 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1136 Attribute attr = *it;
1138 // Do not allow certain attributes - just skip them
1139 // Safepoint can not be read only or read none.
1140 if (attr.hasAttribute(Attribute::ReadNone) ||
1141 attr.hasAttribute(Attribute::ReadOnly))
1144 ret = ret.addAttributes(
1145 AS.getContext(), index,
1146 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1150 // Just skip parameter attributes for now
1156 /// Helper function to place all gc relocates necessary for the given
1159 /// liveVariables - list of variables to be relocated.
1160 /// liveStart - index of the first live variable.
1161 /// basePtrs - base pointers.
1162 /// statepointToken - statepoint instruction to which relocates should be
1164 /// Builder - Llvm IR builder to be used to construct new calls.
1165 /// Returns array with newly created relocates.
1166 static std::vector<llvm::Instruction *>
1167 CreateGCRelocates(const SmallVectorImpl<llvm::Value *> &liveVariables,
1168 const int liveStart,
1169 const SmallVectorImpl<llvm::Value *> &basePtrs,
1170 Instruction *statepointToken, IRBuilder<> Builder) {
1172 std::vector<llvm::Instruction *> newDefs;
1174 Module *M = statepointToken->getParent()->getParent()->getParent();
1176 for (unsigned i = 0; i < liveVariables.size(); i++) {
1177 // We generate a (potentially) unique declaration for every pointer type
1178 // combination. This results is some blow up the function declarations in
1179 // the IR, but removes the need for argument bitcasts which shrinks the IR
1180 // greatly and makes it much more readable.
1181 std::vector<Type *> types; // one per 'any' type
1182 types.push_back(liveVariables[i]->getType()); // result type
1183 Value *gc_relocate_decl = Intrinsic::getDeclaration(
1184 M, Intrinsic::experimental_gc_relocate, types);
1186 // Generate the gc.relocate call and save the result
1188 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1189 liveStart + find_index(liveVariables, basePtrs[i]));
1190 Value *liveIdx = ConstantInt::get(
1191 Type::getInt32Ty(M->getContext()),
1192 liveStart + find_index(liveVariables, liveVariables[i]));
1194 // only specify a debug name if we can give a useful one
1195 Value *reloc = Builder.CreateCall3(
1196 gc_relocate_decl, statepointToken, baseIdx, liveIdx,
1197 liveVariables[i]->hasName() ? liveVariables[i]->getName() + ".relocated"
1199 // Trick CodeGen into thinking there are lots of free registers at this
1201 cast<CallInst>(reloc)->setCallingConv(CallingConv::Cold);
1203 newDefs.push_back(cast<Instruction>(reloc));
1205 assert(newDefs.size() == liveVariables.size() &&
1206 "missing or extra redefinition at safepoint");
1212 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1213 const SmallVectorImpl<llvm::Value *> &basePtrs,
1214 const SmallVectorImpl<llvm::Value *> &liveVariables,
1216 PartiallyConstructedSafepointRecord &result) {
1217 assert(basePtrs.size() == liveVariables.size());
1218 assert(isStatepoint(CS) &&
1219 "This method expects to be rewriting a statepoint");
1221 BasicBlock *BB = CS.getInstruction()->getParent();
1223 Function *F = BB->getParent();
1224 assert(F && "must be set");
1225 Module *M = F->getParent();
1227 assert(M && "must be set");
1229 // We're not changing the function signature of the statepoint since the gc
1230 // arguments go into the var args section.
1231 Function *gc_statepoint_decl = CS.getCalledFunction();
1233 // Then go ahead and use the builder do actually do the inserts. We insert
1234 // immediately before the previous instruction under the assumption that all
1235 // arguments will be available here. We can't insert afterwards since we may
1236 // be replacing a terminator.
1237 Instruction *insertBefore = CS.getInstruction();
1238 IRBuilder<> Builder(insertBefore);
1239 // Copy all of the arguments from the original statepoint - this includes the
1240 // target, call args, and deopt args
1241 std::vector<llvm::Value *> args;
1242 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1243 // TODO: Clear the 'needs rewrite' flag
1245 // add all the pointers to be relocated (gc arguments)
1246 // Capture the start of the live variable list for use in the gc_relocates
1247 const int live_start = args.size();
1248 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1250 // Create the statepoint given all the arguments
1251 Instruction *token = nullptr;
1252 AttributeSet return_attributes;
1254 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1256 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1257 call->setTailCall(toReplace->isTailCall());
1258 call->setCallingConv(toReplace->getCallingConv());
1260 // Currently we will fail on parameter attributes and on certain
1261 // function attributes.
1262 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1263 // In case if we can handle this set of sttributes - set up function attrs
1264 // directly on statepoint and return attrs later for gc_result intrinsic.
1265 call->setAttributes(new_attrs.getFnAttributes());
1266 return_attributes = new_attrs.getRetAttributes();
1270 // Put the following gc_result and gc_relocate calls immediately after the
1271 // the old call (which we're about to delete)
1272 BasicBlock::iterator next(toReplace);
1273 assert(BB->end() != next && "not a terminator, must have next");
1275 Instruction *IP = &*(next);
1276 Builder.SetInsertPoint(IP);
1277 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1279 } else if (CS.isInvoke()) {
1280 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1282 // Insert the new invoke into the old block. We'll remove the old one in a
1283 // moment at which point this will become the new terminator for the
1285 InvokeInst *invoke = InvokeInst::Create(
1286 gc_statepoint_decl, toReplace->getNormalDest(),
1287 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1288 invoke->setCallingConv(toReplace->getCallingConv());
1290 // Currently we will fail on parameter attributes and on certain
1291 // function attributes.
1292 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1293 // In case if we can handle this set of sttributes - set up function attrs
1294 // directly on statepoint and return attrs later for gc_result intrinsic.
1295 invoke->setAttributes(new_attrs.getFnAttributes());
1296 return_attributes = new_attrs.getRetAttributes();
1300 // Generate gc relocates in exceptional path
1301 BasicBlock *unwindBlock = normalizeBBForInvokeSafepoint(
1302 toReplace->getUnwindDest(), invoke->getParent(), P);
1304 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1305 Builder.SetInsertPoint(IP);
1306 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1308 // Extract second element from landingpad return value. We will attach
1309 // exceptional gc relocates to it.
1310 const unsigned idx = 1;
1311 Instruction *exceptional_token =
1312 cast<Instruction>(Builder.CreateExtractValue(
1313 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1314 result.exceptional_relocates_token = exceptional_token;
1316 // Just throw away return value. We will use the one we got for normal
1318 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1319 exceptional_token, Builder);
1321 // Generate gc relocates and returns for normal block
1322 BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
1323 toReplace->getNormalDest(), invoke->getParent(), P);
1325 IP = &*(normalDest->getFirstInsertionPt());
1326 Builder.SetInsertPoint(IP);
1328 // gc relocates will be generated later as if it were regular call
1331 llvm_unreachable("unexpect type of CallSite");
1335 // Take the name of the original value call if it had one.
1336 token->takeName(CS.getInstruction());
1338 // The GCResult is already inserted, we just need to find it
1339 Instruction *gc_result = nullptr;
1341 Instruction *toReplace = CS.getInstruction();
1342 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1343 "only valid use before rewrite is gc.result");
1344 if (toReplace->hasOneUse()) {
1345 Instruction *GCResult = cast<Instruction>(*toReplace->user_begin());
1346 assert(isGCResult(GCResult));
1347 gc_result = GCResult;
1351 // Update the gc.result of the original statepoint (if any) to use the newly
1352 // inserted statepoint. This is safe to do here since the token can't be
1353 // considered a live reference.
1354 CS.getInstruction()->replaceAllUsesWith(token);
1356 // Second, create a gc.relocate for every live variable
1357 std::vector<llvm::Instruction *> newDefs =
1358 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1360 // Need to pass through the last part of the safepoint block so that we
1361 // don't accidentally update uses in a following gc.relocate which is
1362 // still conceptually part of the same safepoint. Gah.
1363 Instruction *last = nullptr;
1364 if (!newDefs.empty()) {
1365 last = newDefs.back();
1366 } else if (gc_result) {
1371 assert(last && "can't be null");
1372 const auto bounds = std::make_pair(token, last);
1374 // Sanity check our results - this is slightly non-trivial due to invokes
1375 VerifySafepointBounds(bounds);
1377 result.safepoint = bounds;
1381 struct name_ordering {
1384 bool operator()(name_ordering const &a, name_ordering const &b) {
1385 return -1 == a.derived->getName().compare(b.derived->getName());
1389 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1390 SmallVectorImpl<Value *> &livevec) {
1391 assert(basevec.size() == livevec.size());
1393 std::vector<name_ordering> temp;
1394 for (size_t i = 0; i < basevec.size(); i++) {
1396 v.base = basevec[i];
1397 v.derived = livevec[i];
1400 std::sort(temp.begin(), temp.end(), name_ordering());
1401 for (size_t i = 0; i < basevec.size(); i++) {
1402 basevec[i] = temp[i].base;
1403 livevec[i] = temp[i].derived;
1407 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1408 // which make the relocations happening at this safepoint explicit.
1410 // WARNING: Does not do any fixup to adjust users of the original live
1411 // values. That's the callers responsibility.
1413 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1414 PartiallyConstructedSafepointRecord &result) {
1415 std::set<llvm::Value *> liveset = result.liveset;
1416 std::map<llvm::Value *, llvm::Value *> base_pairs = result.base_pairs;
1418 // Convert to vector for efficient cross referencing.
1419 SmallVector<Value *, 64> basevec, livevec;
1420 livevec.reserve(liveset.size());
1421 basevec.reserve(liveset.size());
1422 for (Value *L : liveset) {
1423 livevec.push_back(L);
1425 assert(base_pairs.find(L) != base_pairs.end());
1426 Value *base = base_pairs[L];
1427 basevec.push_back(base);
1429 assert(livevec.size() == basevec.size());
1431 // To make the output IR slightly more stable (for use in diffs), ensure a
1432 // fixed order of the values in the safepoint (by sorting the value name).
1433 // The order is otherwise meaningless.
1434 stablize_order(basevec, livevec);
1436 // Do the actual rewriting and delete the old statepoint
1437 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1438 CS.getInstruction()->eraseFromParent();
1441 // Helper function for the relocationViaAlloca.
1442 // It receives iterator to the statepoint gc relocates and emits store to the
1444 // location (via allocaMap) for the each one of them.
1445 // Add visited values into the visitedLiveValues set we will later use them
1446 // for sanity check.
1448 insertRelocationStores(iterator_range<Value::user_iterator> gcRelocs,
1449 DenseMap<Value *, Value *> &allocaMap,
1450 DenseSet<Value *> &visitedLiveValues) {
1452 for (User *U : gcRelocs) {
1453 if (!isa<IntrinsicInst>(U))
1456 IntrinsicInst *relocatedValue = cast<IntrinsicInst>(U);
1458 // We only care about relocates
1459 if (relocatedValue->getIntrinsicID() !=
1460 Intrinsic::experimental_gc_relocate) {
1464 GCRelocateOperands relocateOperands(relocatedValue);
1465 Value *originalValue = const_cast<Value *>(relocateOperands.derivedPtr());
1466 assert(allocaMap.count(originalValue));
1467 Value *alloca = allocaMap[originalValue];
1469 // Emit store into the related alloca
1470 StoreInst *store = new StoreInst(relocatedValue, alloca);
1471 store->insertAfter(relocatedValue);
1474 visitedLiveValues.insert(originalValue);
1479 /// do all the relocation update via allocas and mem2reg
1480 static void relocationViaAlloca(
1481 Function &F, DominatorTree &DT, const std::vector<Value *> &live,
1482 const std::vector<struct PartiallyConstructedSafepointRecord> &records) {
1484 int initialAllocaNum = 0;
1486 // record initial number of allocas
1487 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1489 if (isa<AllocaInst>(*itr))
1494 // TODO-PERF: change data structures, reserve
1495 DenseMap<Value *, Value *> allocaMap;
1496 SmallVector<AllocaInst *, 200> PromotableAllocas;
1497 PromotableAllocas.reserve(live.size());
1499 // emit alloca for each live gc pointer
1500 for (unsigned i = 0; i < live.size(); i++) {
1501 Value *liveValue = live[i];
1502 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1503 F.getEntryBlock().getFirstNonPHI());
1504 allocaMap[liveValue] = alloca;
1505 PromotableAllocas.push_back(alloca);
1508 // The next two loops are part of the same conceptual operation. We need to
1509 // insert a store to the alloca after the original def and at each
1510 // redefinition. We need to insert a load before each use. These are split
1511 // into distinct loops for performance reasons.
1513 // update gc pointer after each statepoint
1514 // either store a relocated value or null (if no relocated value found for
1515 // this gc pointer and it is not a gc_result)
1516 // this must happen before we update the statepoint with load of alloca
1517 // otherwise we lose the link between statepoint and old def
1518 for (size_t i = 0; i < records.size(); i++) {
1519 const struct PartiallyConstructedSafepointRecord &info = records[i];
1520 Value *statepoint = info.safepoint.first;
1522 // This will be used for consistency check
1523 DenseSet<Value *> visitedLiveValues;
1525 // Insert stores for normal statepoint gc relocates
1526 insertRelocationStores(statepoint->users(), allocaMap, visitedLiveValues);
1528 // In case if it was invoke statepoint
1529 // we will insert stores for exceptional path gc relocates.
1530 if (isa<InvokeInst>(statepoint)) {
1531 insertRelocationStores(info.exceptional_relocates_token->users(),
1532 allocaMap, visitedLiveValues);
1536 // For consistency check store null's into allocas for values that are not
1538 // by this statepoint.
1539 for (auto Pair : allocaMap) {
1540 Value *def = Pair.first;
1541 Value *alloca = Pair.second;
1543 // This value was relocated
1544 if (visitedLiveValues.count(def)) {
1547 // Result should not be relocated
1548 if (def == info.result) {
1553 ConstantPointerNull::get(cast<PointerType>(def->getType()));
1554 StoreInst *store = new StoreInst(CPN, alloca);
1555 store->insertBefore(info.safepoint.second);
1559 // update use with load allocas and add store for gc_relocated
1560 for (auto Pair : allocaMap) {
1561 Value *def = Pair.first;
1562 Value *alloca = Pair.second;
1564 // we pre-record the uses of allocas so that we dont have to worry about
1566 // that change the user information.
1567 SmallVector<Instruction *, 20> uses;
1568 // PERF: trade a linear scan for repeated reallocation
1569 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1570 for (User *U : def->users()) {
1571 if (!isa<ConstantExpr>(U)) {
1572 // If the def has a ConstantExpr use, then the def is either a
1573 // ConstantExpr use itself or null. In either case
1574 // (recursively in the first, directly in the second), the oop
1575 // it is ultimately dependent on is null and this particular
1576 // use does not need to be fixed up.
1577 uses.push_back(cast<Instruction>(U));
1581 std::sort(uses.begin(), uses.end());
1582 auto last = std::unique(uses.begin(), uses.end());
1583 uses.erase(last, uses.end());
1585 for (Instruction *use : uses) {
1586 if (isa<PHINode>(use)) {
1587 PHINode *phi = cast<PHINode>(use);
1588 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1589 if (def == phi->getIncomingValue(i)) {
1590 LoadInst *load = new LoadInst(
1591 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1592 phi->setIncomingValue(i, load);
1596 LoadInst *load = new LoadInst(alloca, "", use);
1597 use->replaceUsesOfWith(def, load);
1601 // emit store for the initial gc value
1602 // store must be inserted after load, otherwise store will be in alloca's
1603 // use list and an extra load will be inserted before it
1604 StoreInst *store = new StoreInst(def, alloca);
1605 if (isa<Instruction>(def)) {
1606 store->insertAfter(cast<Instruction>(def));
1608 assert((isa<Argument>(def) || isa<GlobalVariable>(def) ||
1609 (isa<Constant>(def) && cast<Constant>(def)->isNullValue())) &&
1610 "Must be argument or global");
1611 store->insertAfter(cast<Instruction>(alloca));
1615 assert(PromotableAllocas.size() == live.size() &&
1616 "we must have the same allocas with lives");
1617 if (!PromotableAllocas.empty()) {
1618 // apply mem2reg to promote alloca to SSA
1619 PromoteMemToReg(PromotableAllocas, DT);
1623 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1625 if (isa<AllocaInst>(*itr))
1628 assert(initialAllocaNum == 0 && "We must not introduce any extra allocas");
1632 /// Implement a unique function which doesn't require we sort the input
1633 /// vector. Doing so has the effect of changing the output of a couple of
1634 /// tests in ways which make them less useful in testing fused safepoints.
1635 template <typename T> static void unique_unsorted(std::vector<T> &vec) {
1638 vec.reserve(vec.size());
1639 std::swap(tmp, vec);
1640 for (auto V : tmp) {
1641 if (seen.insert(V).second) {
1647 static Function *getUseHolder(Module &M) {
1648 FunctionType *ftype =
1649 FunctionType::get(Type::getVoidTy(M.getContext()), true);
1650 Function *Func = cast<Function>(M.getOrInsertFunction("__tmp_use", ftype));
1654 /// Insert holders so that each Value is obviously live through the entire
1655 /// liftetime of the call.
1656 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1657 std::vector<CallInst *> &holders) {
1658 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1659 Function *Func = getUseHolder(*M);
1661 // For call safepoints insert dummy calls right after safepoint
1662 BasicBlock::iterator next(CS.getInstruction());
1664 CallInst *base_holder = CallInst::Create(Func, Values, "", next);
1665 holders.push_back(base_holder);
1666 } else if (CS.isInvoke()) {
1667 // For invoke safepooints insert dummy calls both in normal and
1668 // exceptional destination blocks
1669 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1670 CallInst *normal_holder = CallInst::Create(
1671 Func, Values, "", invoke->getNormalDest()->getFirstInsertionPt());
1672 CallInst *unwind_holder = CallInst::Create(
1673 Func, Values, "", invoke->getUnwindDest()->getFirstInsertionPt());
1674 holders.push_back(normal_holder);
1675 holders.push_back(unwind_holder);
1677 assert(false && "Unsupported");
1681 static void findLiveReferences(
1682 Function &F, DominatorTree &DT, Pass *P, std::vector<CallSite> &toUpdate,
1683 std::vector<struct PartiallyConstructedSafepointRecord> &records) {
1684 for (size_t i = 0; i < records.size(); i++) {
1685 struct PartiallyConstructedSafepointRecord &info = records[i];
1686 CallSite &CS = toUpdate[i];
1687 analyzeParsePointLiveness(DT, CS, info);
1691 static void addBasesAsLiveValues(std::set<Value *> &liveset,
1692 std::map<Value *, Value *> &base_pairs) {
1693 // Identify any base pointers which are used in this safepoint, but not
1694 // themselves relocated. We need to relocate them so that later inserted
1695 // safepoints can get the properly relocated base register.
1696 DenseSet<Value *> missing;
1697 for (Value *L : liveset) {
1698 assert(base_pairs.find(L) != base_pairs.end());
1699 Value *base = base_pairs[L];
1701 if (liveset.find(base) == liveset.end()) {
1702 assert(base_pairs.find(base) == base_pairs.end());
1703 // uniqued by set insert
1704 missing.insert(base);
1708 // Note that we want these at the end of the list, otherwise
1709 // register placement gets screwed up once we lower to STATEPOINT
1710 // instructions. This is an utter hack, but there doesn't seem to be a
1712 for (Value *base : missing) {
1714 liveset.insert(base);
1715 base_pairs[base] = base;
1717 assert(liveset.size() == base_pairs.size());
1720 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1721 std::vector<CallSite> &toUpdate) {
1723 // sanity check the input
1724 std::set<CallSite> uniqued;
1725 uniqued.insert(toUpdate.begin(), toUpdate.end());
1726 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1728 for (size_t i = 0; i < toUpdate.size(); i++) {
1729 CallSite &CS = toUpdate[i];
1730 assert(CS.getInstruction()->getParent()->getParent() == &F);
1731 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1735 // A list of dummy calls added to the IR to keep various values obviously
1736 // live in the IR. We'll remove all of these when done.
1737 std::vector<CallInst *> holders;
1739 // Insert a dummy call with all of the arguments to the vm_state we'll need
1740 // for the actual safepoint insertion. This ensures reference arguments in
1741 // the deopt argument list are considered live through the safepoint (and
1742 // thus makes sure they get relocated.)
1743 for (size_t i = 0; i < toUpdate.size(); i++) {
1744 CallSite &CS = toUpdate[i];
1745 Statepoint StatepointCS(CS);
1747 SmallVector<Value *, 64> DeoptValues;
1748 for (Use &U : StatepointCS.vm_state_args()) {
1749 Value *Arg = cast<Value>(&U);
1750 if (isGCPointerType(Arg->getType()))
1751 DeoptValues.push_back(Arg);
1753 insertUseHolderAfter(CS, DeoptValues, holders);
1756 std::vector<struct PartiallyConstructedSafepointRecord> records;
1757 records.reserve(toUpdate.size());
1758 for (size_t i = 0; i < toUpdate.size(); i++) {
1759 struct PartiallyConstructedSafepointRecord info;
1760 records.push_back(info);
1762 assert(records.size() == toUpdate.size());
1764 // A) Identify all gc pointers which are staticly live at the given call
1766 findLiveReferences(F, DT, P, toUpdate, records);
1768 // B) Find the base pointers for each live pointer
1769 /* scope for caching */ {
1770 // Cache the 'defining value' relation used in the computation and
1771 // insertion of base phis and selects. This ensures that we don't insert
1772 // large numbers of duplicate base_phis.
1773 DefiningValueMapTy DVCache;
1775 for (size_t i = 0; i < records.size(); i++) {
1776 struct PartiallyConstructedSafepointRecord &info = records[i];
1777 CallSite &CS = toUpdate[i];
1778 findBasePointers(DT, DVCache, CS, info);
1780 } // end of cache scope
1782 // The base phi insertion logic (for any safepoint) may have inserted new
1783 // instructions which are now live at some safepoint. The simplest such
1786 // phi a <-- will be a new base_phi here
1787 // safepoint 1 <-- that needs to be live here
1791 std::set<llvm::Value *> allInsertedDefs;
1792 for (size_t i = 0; i < records.size(); i++) {
1793 struct PartiallyConstructedSafepointRecord &info = records[i];
1794 allInsertedDefs.insert(info.newInsertedDefs.begin(),
1795 info.newInsertedDefs.end());
1798 // We insert some dummy calls after each safepoint to definitely hold live
1799 // the base pointers which were identified for that safepoint. We'll then
1800 // ask liveness for _every_ base inserted to see what is now live. Then we
1801 // remove the dummy calls.
1802 holders.reserve(holders.size() + records.size());
1803 for (size_t i = 0; i < records.size(); i++) {
1804 struct PartiallyConstructedSafepointRecord &info = records[i];
1805 CallSite &CS = toUpdate[i];
1807 SmallVector<Value *, 128> Bases;
1808 for (auto Pair : info.base_pairs) {
1809 Bases.push_back(Pair.second);
1811 insertUseHolderAfter(CS, Bases, holders);
1814 // Add the bases explicitly to the live vector set. This may result in a few
1815 // extra relocations, but the base has to be available whenever a pointer
1816 // derived from it is used. Thus, we need it to be part of the statepoint's
1817 // gc arguments list. TODO: Introduce an explicit notion (in the following
1818 // code) of the GC argument list as seperate from the live Values at a
1819 // given statepoint.
1820 for (size_t i = 0; i < records.size(); i++) {
1821 struct PartiallyConstructedSafepointRecord &info = records[i];
1822 addBasesAsLiveValues(info.liveset, info.base_pairs);
1825 // If we inserted any new values, we need to adjust our notion of what is
1826 // live at a particular safepoint.
1827 if (!allInsertedDefs.empty()) {
1828 fixupLiveReferences(F, DT, P, allInsertedDefs, toUpdate, records);
1830 if (PrintBasePointers) {
1831 for (size_t i = 0; i < records.size(); i++) {
1832 struct PartiallyConstructedSafepointRecord &info = records[i];
1833 errs() << "Base Pairs: (w/Relocation)\n";
1834 for (auto Pair : info.base_pairs) {
1835 errs() << " derived %" << Pair.first->getName() << " base %"
1836 << Pair.second->getName() << "\n";
1840 for (size_t i = 0; i < holders.size(); i++) {
1841 holders[i]->eraseFromParent();
1842 holders[i] = nullptr;
1846 // Now run through and replace the existing statepoints with new ones with
1847 // the live variables listed. We do not yet update uses of the values being
1848 // relocated. We have references to live variables that need to
1849 // survive to the last iteration of this loop. (By construction, the
1850 // previous statepoint can not be a live variable, thus we can and remove
1851 // the old statepoint calls as we go.)
1852 for (size_t i = 0; i < records.size(); i++) {
1853 struct PartiallyConstructedSafepointRecord &info = records[i];
1854 CallSite &CS = toUpdate[i];
1855 makeStatepointExplicit(DT, CS, P, info);
1857 toUpdate.clear(); // prevent accident use of invalid CallSites
1859 // In case if we inserted relocates in a different basic block than the
1860 // original safepoint (this can happen for invokes). We need to be sure that
1861 // original values were not used in any of the phi nodes at the
1862 // beginning of basic block containing them. Because we know that all such
1863 // blocks will have single predecessor we can safely assume that all phi
1864 // nodes have single entry (because of normalizeBBForInvokeSafepoint).
1865 // Just remove them all here.
1866 for (size_t i = 0; i < records.size(); i++) {
1867 Instruction *I = records[i].safepoint.first;
1869 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
1870 FoldSingleEntryPHINodes(invoke->getNormalDest());
1871 assert(!isa<PHINode>(invoke->getNormalDest()->begin()));
1873 FoldSingleEntryPHINodes(invoke->getUnwindDest());
1874 assert(!isa<PHINode>(invoke->getUnwindDest()->begin()));
1878 // Do all the fixups of the original live variables to their relocated selves
1879 std::vector<Value *> live;
1880 for (size_t i = 0; i < records.size(); i++) {
1881 struct PartiallyConstructedSafepointRecord &info = records[i];
1882 // We can't simply save the live set from the original insertion. One of
1883 // the live values might be the result of a call which needs a safepoint.
1884 // That Value* no longer exists and we need to use the new gc_result.
1885 // Thankfully, the liveset is embedded in the statepoint (and updated), so
1886 // we just grab that.
1887 Statepoint statepoint(info.safepoint.first);
1888 live.insert(live.end(), statepoint.gc_args_begin(),
1889 statepoint.gc_args_end());
1891 unique_unsorted(live);
1895 for (auto ptr : live) {
1896 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
1900 relocationViaAlloca(F, DT, live, records);
1901 return !records.empty();
1904 /// Returns true if this function should be rewritten by this pass. The main
1905 /// point of this function is as an extension point for custom logic.
1906 static bool shouldRewriteStatepointsIn(Function &F) {
1907 // TODO: This should check the GCStrategy
1908 const std::string StatepointExampleName("statepoint-example");
1909 return StatepointExampleName == F.getGC();
1912 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
1913 // Nothing to do for declarations.
1914 if (F.isDeclaration() || F.empty())
1917 // Policy choice says not to rewrite - the most common reason is that we're
1918 // compiling code without a GCStrategy.
1919 if (!shouldRewriteStatepointsIn(F))
1922 // Gather all the statepoints which need rewritten.
1923 std::vector<CallSite> ParsePointNeeded;
1924 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1926 // TODO: only the ones with the flag set!
1927 if (isStatepoint(*itr))
1928 ParsePointNeeded.push_back(CallSite(&*itr));
1931 // Return early if no work to do.
1932 if (ParsePointNeeded.empty())
1935 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1936 return insertParsePoints(F, DT, this, ParsePointNeeded);