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);
392 // Has to be a pointer to an gc object, or possibly an array of such?
393 assert(Op->getType()->isPointerTy());
394 return LI; // The value loaded is an gc base itself
397 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
398 Value *Op = GEP->getOperand(0);
399 if (Op->getType()->isPointerTy()) {
400 return findBaseDefiningValue(Op); // The base of this GEP is the base
404 if (AllocaInst *alloc = dyn_cast<AllocaInst>(I)) {
405 // An alloca represents a conceptual stack slot. It's the slot itself
406 // that the GC needs to know about, not the value in the slot.
407 assert(alloc->getType()->isPointerTy() &&
408 "Base for pointer must be another pointer");
412 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
413 switch (II->getIntrinsicID()) {
415 // fall through to general call handling
417 case Intrinsic::experimental_gc_statepoint:
418 case Intrinsic::experimental_gc_result_float:
419 case Intrinsic::experimental_gc_result_int:
420 llvm_unreachable("these don't produce pointers");
421 case Intrinsic::experimental_gc_result_ptr:
422 // This is just a special case of the CallInst check below to handle a
423 // statepoint with deopt args which hasn't been rewritten for GC yet.
424 // TODO: Assert that the statepoint isn't rewritten yet.
426 case Intrinsic::experimental_gc_relocate: {
427 // Rerunning safepoint insertion after safepoints are already
428 // inserted is not supported. It could probably be made to work,
429 // but why are you doing this? There's no good reason.
430 llvm_unreachable("repeat safepoint insertion is not supported");
432 case Intrinsic::gcroot:
433 // Currently, this mechanism hasn't been extended to work with gcroot.
434 // There's no reason it couldn't be, but I haven't thought about the
435 // implications much.
437 "interaction with the gcroot mechanism is not supported");
440 // We assume that functions in the source language only return base
441 // pointers. This should probably be generalized via attributes to support
442 // both source language and internal functions.
443 if (CallInst *call = dyn_cast<CallInst>(I)) {
444 assert(call->getType()->isPointerTy() &&
445 "Base for pointer must be another pointer");
448 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
449 assert(invoke->getType()->isPointerTy() &&
450 "Base for pointer must be another pointer");
454 // I have absolutely no idea how to implement this part yet. It's not
455 // neccessarily hard, I just haven't really looked at it yet.
456 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
458 if (AtomicCmpXchgInst *cas = dyn_cast<AtomicCmpXchgInst>(I)) {
459 // A CAS is effectively a atomic store and load combined under a
460 // predicate. From the perspective of base pointers, we just treat it
461 // like a load. We loaded a pointer from a address in memory, that value
462 // had better be a valid base pointer.
463 return cas->getPointerOperand();
465 if (AtomicRMWInst *atomic = dyn_cast<AtomicRMWInst>(I)) {
466 assert(AtomicRMWInst::Xchg == atomic->getOperation() &&
467 "All others are binary ops which don't apply to base pointers");
468 // semantically, a load, store pair. Treat it the same as a standard load
469 return atomic->getPointerOperand();
472 // The aggregate ops. Aggregates can either be in the heap or on the
473 // stack, but in either case, this is simply a field load. As a result,
474 // this is a defining definition of the base just like a load is.
475 if (ExtractValueInst *ev = dyn_cast<ExtractValueInst>(I)) {
479 // We should never see an insert vector since that would require we be
480 // tracing back a struct value not a pointer value.
481 assert(!isa<InsertValueInst>(I) &&
482 "Base pointer for a struct is meaningless");
484 // The last two cases here don't return a base pointer. Instead, they
485 // return a value which dynamically selects from amoung several base
486 // derived pointers (each with it's own base potentially). It's the job of
487 // the caller to resolve these.
488 if (SelectInst *select = dyn_cast<SelectInst>(I)) {
491 if (PHINode *phi = dyn_cast<PHINode>(I)) {
495 errs() << "unknown type: " << *I << "\n";
496 llvm_unreachable("unknown type");
500 /// Returns the base defining value for this value.
501 Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &cache) {
502 if (cache.find(I) == cache.end()) {
503 cache[I] = findBaseDefiningValue(I);
505 assert(cache.find(I) != cache.end());
508 errs() << "fBDV-cached: " << I->getName() << " -> " << cache[I]->getName()
514 /// Return a base pointer for this value if known. Otherwise, return it's
515 /// base defining value.
516 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &cache) {
517 Value *def = findBaseDefiningValueCached(I, cache);
518 if (cache.count(def)) {
519 // Either a base-of relation, or a self reference. Caller must check.
522 // Only a BDV available
526 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
527 /// is it known to be a base pointer? Or do we need to continue searching.
528 static bool isKnownBaseResult(Value *v) {
529 if (!isa<PHINode>(v) && !isa<SelectInst>(v)) {
530 // no recursion possible
533 if (cast<Instruction>(v)->getMetadata("is_base_value")) {
534 // This is a previously inserted base phi or select. We know
535 // that this is a base value.
539 // We need to keep searching
543 // TODO: find a better name for this
547 enum Status { Unknown, Base, Conflict };
549 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
550 assert(status != Base || b);
552 PhiState(Value *b) : status(Base), base(b) {}
553 PhiState() : status(Unknown), base(nullptr) {}
554 PhiState(const PhiState &other) : status(other.status), base(other.base) {
555 assert(status != Base || base);
558 Status getStatus() const { return status; }
559 Value *getBase() const { return base; }
561 bool isBase() const { return getStatus() == Base; }
562 bool isUnknown() const { return getStatus() == Unknown; }
563 bool isConflict() const { return getStatus() == Conflict; }
565 bool operator==(const PhiState &other) const {
566 return base == other.base && status == other.status;
569 bool operator!=(const PhiState &other) const { return !(*this == other); }
572 errs() << status << " (" << base << " - "
573 << (base ? base->getName() : "nullptr") << "): ";
578 Value *base; // non null only if status == base
581 // Values of type PhiState form a lattice, and this is a helper
582 // class that implementes the meet operation. The meat of the meet
583 // operation is implemented in MeetPhiStates::pureMeet
584 class MeetPhiStates {
586 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
587 explicit MeetPhiStates(const std::map<Value *, PhiState> &phiStates)
588 : phiStates(phiStates) {}
590 // Destructively meet the current result with the base V. V can
591 // either be a merge instruction (SelectInst / PHINode), in which
592 // case its status is looked up in the phiStates map; or a regular
593 // SSA value, in which case it is assumed to be a base.
594 void meetWith(Value *V) {
595 PhiState otherState = getStateForBDV(V);
596 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
597 MeetPhiStates::pureMeet(currentResult, otherState)) &&
598 "math is wrong: meet does not commute!");
599 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
602 PhiState getResult() const { return currentResult; }
605 const std::map<Value *, PhiState> &phiStates;
606 PhiState currentResult;
608 /// Return a phi state for a base defining value. We'll generate a new
609 /// base state for known bases and expect to find a cached state otherwise
610 PhiState getStateForBDV(Value *baseValue) {
611 if (isKnownBaseResult(baseValue)) {
612 return PhiState(baseValue);
614 return lookupFromMap(baseValue);
618 PhiState lookupFromMap(Value *V) {
619 auto I = phiStates.find(V);
620 assert(I != phiStates.end() && "lookup failed!");
624 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
625 switch (stateA.getStatus()) {
626 case PhiState::Unknown:
630 assert(stateA.getBase() && "can't be null");
631 if (stateB.isUnknown()) {
633 } else if (stateB.isBase()) {
634 if (stateA.getBase() == stateB.getBase()) {
635 assert(stateA == stateB && "equality broken!");
638 return PhiState(PhiState::Conflict);
640 assert(stateB.isConflict() && "only three states!");
641 return PhiState(PhiState::Conflict);
644 case PhiState::Conflict:
647 assert(false && "only three states!");
651 /// For a given value or instruction, figure out what base ptr it's derived
652 /// from. For gc objects, this is simply itself. On success, returns a value
653 /// which is the base pointer. (This is reliable and can be used for
654 /// relocation.) On failure, returns nullptr.
655 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache,
656 std::set<llvm::Value *> &newInsertedDefs) {
657 Value *def = findBaseOrBDV(I, cache);
659 if (isKnownBaseResult(def)) {
663 // Here's the rough algorithm:
664 // - For every SSA value, construct a mapping to either an actual base
665 // pointer or a PHI which obscures the base pointer.
666 // - Construct a mapping from PHI to unknown TOP state. Use an
667 // optimistic algorithm to propagate base pointer information. Lattice
672 // When algorithm terminates, all PHIs will either have a single concrete
673 // base or be in a conflict state.
674 // - For every conflict, insert a dummy PHI node without arguments. Add
675 // these to the base[Instruction] = BasePtr mapping. For every
676 // non-conflict, add the actual base.
677 // - For every conflict, add arguments for the base[a] of each input
680 // Note: A simpler form of this would be to add the conflict form of all
681 // PHIs without running the optimistic algorithm. This would be
682 // analougous to pessimistic data flow and would likely lead to an
683 // overall worse solution.
685 std::map<Value *, PhiState> states;
686 states[def] = PhiState();
687 // Recursively fill in all phis & selects reachable from the initial one
688 // for which we don't already know a definite base value for
689 // PERF: Yes, this is as horribly inefficient as it looks.
693 for (auto Pair : states) {
694 Value *v = Pair.first;
695 assert(!isKnownBaseResult(v) && "why did it get added?");
696 if (PHINode *phi = dyn_cast<PHINode>(v)) {
697 unsigned NumPHIValues = phi->getNumIncomingValues();
698 assert(NumPHIValues > 0 && "zero input phis are illegal");
699 for (unsigned i = 0; i != NumPHIValues; ++i) {
700 Value *InVal = phi->getIncomingValue(i);
701 Value *local = findBaseOrBDV(InVal, cache);
702 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
703 states[local] = PhiState();
707 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
708 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
709 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
710 states[local] = PhiState();
713 local = findBaseOrBDV(sel->getFalseValue(), cache);
714 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
715 states[local] = PhiState();
723 errs() << "States after initialization:\n";
724 for (auto Pair : states) {
725 Instruction *v = cast<Instruction>(Pair.first);
726 PhiState state = Pair.second;
732 // TODO: come back and revisit the state transitions around inputs which
733 // have reached conflict state. The current version seems too conservative.
735 bool progress = true;
738 oldSize = states.size();
740 for (auto Pair : states) {
741 MeetPhiStates calculateMeet(states);
742 Value *v = Pair.first;
743 assert(!isKnownBaseResult(v) && "why did it get added?");
744 assert(isa<SelectInst>(v) || isa<PHINode>(v));
745 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
746 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
747 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
748 } else if (PHINode *phi = dyn_cast<PHINode>(v)) {
749 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
750 calculateMeet.meetWith(
751 findBaseOrBDV(phi->getIncomingValue(i), cache));
754 llvm_unreachable("no such state expected");
757 PhiState oldState = states[v];
758 PhiState newState = calculateMeet.getResult();
759 if (oldState != newState) {
761 states[v] = newState;
765 assert(oldSize <= states.size());
766 assert(oldSize == states.size() || progress);
770 errs() << "States after meet iteration:\n";
771 for (auto Pair : states) {
772 Instruction *v = cast<Instruction>(Pair.first);
773 PhiState state = Pair.second;
779 // Insert Phis for all conflicts
780 for (auto Pair : states) {
781 Instruction *v = cast<Instruction>(Pair.first);
782 PhiState state = Pair.second;
783 assert(!isKnownBaseResult(v) && "why did it get added?");
784 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
785 if (state.isConflict()) {
786 if (isa<PHINode>(v)) {
788 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
789 assert(num_preds > 0 && "how did we reach here");
790 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
791 newInsertedDefs.insert(phi);
792 // Add metadata marking this as a base value
793 auto *const_1 = ConstantInt::get(
795 v->getParent()->getParent()->getParent()->getContext()),
797 auto MDConst = ConstantAsMetadata::get(const_1);
798 MDNode *md = MDNode::get(
799 v->getParent()->getParent()->getParent()->getContext(), MDConst);
800 phi->setMetadata("is_base_value", md);
801 states[v] = PhiState(PhiState::Conflict, phi);
802 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
803 // The undef will be replaced later
804 UndefValue *undef = UndefValue::get(sel->getType());
805 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
806 undef, "base_select", sel);
807 newInsertedDefs.insert(basesel);
808 // Add metadata marking this as a base value
809 auto *const_1 = ConstantInt::get(
811 v->getParent()->getParent()->getParent()->getContext()),
813 auto MDConst = ConstantAsMetadata::get(const_1);
814 MDNode *md = MDNode::get(
815 v->getParent()->getParent()->getParent()->getContext(), MDConst);
816 basesel->setMetadata("is_base_value", md);
817 states[v] = PhiState(PhiState::Conflict, basesel);
824 // Fixup all the inputs of the new PHIs
825 for (auto Pair : states) {
826 Instruction *v = cast<Instruction>(Pair.first);
827 PhiState state = Pair.second;
829 assert(!isKnownBaseResult(v) && "why did it get added?");
830 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
831 if (state.isConflict()) {
832 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
833 PHINode *phi = cast<PHINode>(v);
834 unsigned NumPHIValues = phi->getNumIncomingValues();
835 for (unsigned i = 0; i < NumPHIValues; i++) {
836 Value *InVal = phi->getIncomingValue(i);
837 BasicBlock *InBB = phi->getIncomingBlock(i);
839 // If we've already seen InBB, add the same incoming value
840 // we added for it earlier. The IR verifier requires phi
841 // nodes with multiple entries from the same basic block
842 // to have the same incoming value for each of those
843 // entries. If we don't do this check here and basephi
844 // has a different type than base, we'll end up adding two
845 // bitcasts (and hence two distinct values) as incoming
846 // values for the same basic block.
848 int blockIndex = basephi->getBasicBlockIndex(InBB);
849 if (blockIndex != -1) {
850 Value *oldBase = basephi->getIncomingValue(blockIndex);
851 basephi->addIncoming(oldBase, InBB);
853 Value *base = findBaseOrBDV(InVal, cache);
854 if (!isKnownBaseResult(base)) {
855 // Either conflict or base.
856 assert(states.count(base));
857 base = states[base].getBase();
858 assert(base != nullptr && "unknown PhiState!");
859 assert(newInsertedDefs.count(base) &&
860 "should have already added this in a prev. iteration!");
863 // In essense this assert states: the only way two
864 // values incoming from the same basic block may be
865 // different is by being different bitcasts of the same
866 // value. A cleanup that remains TODO is changing
867 // findBaseOrBDV to return an llvm::Value of the correct
868 // type (and still remain pure). This will remove the
869 // need to add bitcasts.
870 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
871 "sanity -- findBaseOrBDV should be pure!");
876 // Find either the defining value for the PHI or the normal base for
878 Value *base = findBaseOrBDV(InVal, cache);
879 if (!isKnownBaseResult(base)) {
880 // Either conflict or base.
881 assert(states.count(base));
882 base = states[base].getBase();
883 assert(base != nullptr && "unknown PhiState!");
885 assert(base && "can't be null");
886 // Must use original input BB since base may not be Instruction
887 // The cast is needed since base traversal may strip away bitcasts
888 if (base->getType() != basephi->getType()) {
889 base = new BitCastInst(base, basephi->getType(), "cast",
890 InBB->getTerminator());
891 newInsertedDefs.insert(base);
893 basephi->addIncoming(base, InBB);
895 assert(basephi->getNumIncomingValues() == NumPHIValues);
896 } else if (SelectInst *basesel = dyn_cast<SelectInst>(state.getBase())) {
897 SelectInst *sel = cast<SelectInst>(v);
898 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
899 // something more safe and less hacky.
900 for (int i = 1; i <= 2; i++) {
901 Value *InVal = sel->getOperand(i);
902 // Find either the defining value for the PHI or the normal base for
904 Value *base = findBaseOrBDV(InVal, cache);
905 if (!isKnownBaseResult(base)) {
906 // Either conflict or base.
907 assert(states.count(base));
908 base = states[base].getBase();
909 assert(base != nullptr && "unknown PhiState!");
911 assert(base && "can't be null");
912 // Must use original input BB since base may not be Instruction
913 // The cast is needed since base traversal may strip away bitcasts
914 if (base->getType() != basesel->getType()) {
915 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
916 newInsertedDefs.insert(base);
918 basesel->setOperand(i, base);
921 assert(false && "unexpected type");
926 // Cache all of our results so we can cheaply reuse them
927 // NOTE: This is actually two caches: one of the base defining value
928 // relation and one of the base pointer relation! FIXME
929 for (auto item : states) {
930 Value *v = item.first;
931 Value *base = item.second.getBase();
933 assert(!isKnownBaseResult(v) && "why did it get added?");
936 std::string fromstr =
937 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
939 errs() << "Updating base value cache"
940 << " for: " << (v->hasName() ? v->getName() : "")
941 << " from: " << fromstr
942 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
945 assert(isKnownBaseResult(base) &&
946 "must be something we 'know' is a base pointer");
947 if (cache.count(v)) {
948 // Once we transition from the BDV relation being store in the cache to
949 // the base relation being stored, it must be stable
950 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
951 "base relation should be stable");
955 assert(cache.find(def) != cache.end());
959 // For a set of live pointers (base and/or derived), identify the base
960 // pointer of the object which they are derived from. This routine will
961 // mutate the IR graph as needed to make the 'base' pointer live at the
962 // definition site of 'derived'. This ensures that any use of 'derived' can
963 // also use 'base'. This may involve the insertion of a number of
964 // additional PHI nodes.
966 // preconditions: live is a set of pointer type Values
968 // side effects: may insert PHI nodes into the existing CFG, will preserve
969 // CFG, will not remove or mutate any existing nodes
971 // post condition: base_pairs contains one (derived, base) pair for every
972 // pointer in live. Note that derived can be equal to base if the original
973 // pointer was a base pointer.
974 static void findBasePointers(const std::set<llvm::Value *> &live,
975 std::map<llvm::Value *, llvm::Value *> &base_pairs,
976 DominatorTree *DT, DefiningValueMapTy &DVCache,
977 std::set<llvm::Value *> &newInsertedDefs) {
978 for (Value *ptr : live) {
979 Value *base = findBasePointer(ptr, DVCache, newInsertedDefs);
980 assert(base && "failed to find base pointer");
981 base_pairs[ptr] = base;
982 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
983 DT->dominates(cast<Instruction>(base)->getParent(),
984 cast<Instruction>(ptr)->getParent())) &&
985 "The base we found better dominate the derived pointer");
987 if (isNullConstant(base))
988 // If you see this trip and like to live really dangerously, the code
989 // should be correct, just with idioms the verifier can't handle. You
990 // can try disabling the verifier at your own substaintial risk.
991 llvm_unreachable("the relocation code needs adjustment to handle the"
992 "relocation of a null pointer constant without causing"
993 "false positives in the safepoint ir verifier.");
997 /// Find the required based pointers (and adjust the live set) for the given
999 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1001 PartiallyConstructedSafepointRecord &result) {
1002 std::map<llvm::Value *, llvm::Value *> base_pairs;
1003 std::set<llvm::Value *> newInsertedDefs;
1004 findBasePointers(result.liveset, base_pairs, &DT, DVCache, newInsertedDefs);
1006 if (PrintBasePointers) {
1007 errs() << "Base Pairs (w/o Relocation):\n";
1008 for (auto Pair : base_pairs) {
1009 errs() << " derived %" << Pair.first->getName() << " base %"
1010 << Pair.second->getName() << "\n";
1014 result.base_pairs = base_pairs;
1015 result.newInsertedDefs = newInsertedDefs;
1018 /// Check for liveness of items in the insert defs and add them to the live
1019 /// and base pointer sets
1020 static void fixupLiveness(DominatorTree &DT, const CallSite &CS,
1021 const std::set<Value *> &allInsertedDefs,
1022 PartiallyConstructedSafepointRecord &result) {
1023 Instruction *inst = CS.getInstruction();
1025 std::set<llvm::Value *> liveset = result.liveset;
1026 std::map<llvm::Value *, llvm::Value *> base_pairs = result.base_pairs;
1028 auto is_live_gc_reference =
1029 [&](Value &V) { return isLiveGCReferenceAt(V, inst, DT, nullptr); };
1031 // For each new definition, check to see if a) the definition dominates the
1032 // instruction we're interested in, and b) one of the uses of that definition
1033 // is edge-reachable from the instruction we're interested in. This is the
1034 // same definition of liveness we used in the intial liveness analysis
1035 for (Value *newDef : allInsertedDefs) {
1036 if (liveset.count(newDef)) {
1037 // already live, no action needed
1041 // PERF: Use DT to check instruction domination might not be good for
1042 // compilation time, and we could change to optimal solution if this
1043 // turn to be a issue
1044 if (!DT.dominates(cast<Instruction>(newDef), inst)) {
1045 // can't possibly be live at inst
1049 if (is_live_gc_reference(*newDef)) {
1050 // Add the live new defs into liveset and base_pairs
1051 liveset.insert(newDef);
1052 base_pairs[newDef] = newDef;
1056 result.liveset = liveset;
1057 result.base_pairs = base_pairs;
1060 static void fixupLiveReferences(
1061 Function &F, DominatorTree &DT, Pass *P,
1062 const std::set<llvm::Value *> &allInsertedDefs,
1063 std::vector<CallSite> &toUpdate,
1064 std::vector<struct PartiallyConstructedSafepointRecord> &records) {
1065 for (size_t i = 0; i < records.size(); i++) {
1066 struct PartiallyConstructedSafepointRecord &info = records[i];
1067 CallSite &CS = toUpdate[i];
1068 fixupLiveness(DT, CS, allInsertedDefs, info);
1072 // Normalize basic block to make it ready to be target of invoke statepoint.
1073 // It means spliting it to have single predecessor. Return newly created BB
1074 // ready to be successor of invoke statepoint.
1075 static BasicBlock *normalizeBBForInvokeSafepoint(BasicBlock *BB,
1076 BasicBlock *InvokeParent,
1078 BasicBlock *ret = BB;
1080 if (!BB->getUniquePredecessor()) {
1081 ret = SplitBlockPredecessors(BB, InvokeParent, "");
1084 // Another requirement for such basic blocks is to not have any phi nodes.
1085 // Since we just ensured that new BB will have single predecessor,
1086 // all phi nodes in it will have one value. Here it would be naturall place
1088 // remove them all. But we can not do this because we are risking to remove
1089 // one of the values stored in liveset of another statepoint. We will do it
1090 // later after placing all safepoints.
1096 VerifySafepointBounds(const std::pair<Instruction *, Instruction *> &bounds) {
1097 assert(bounds.first->getParent() && bounds.second->getParent() &&
1098 "both must belong to basic blocks");
1099 if (bounds.first->getParent() == bounds.second->getParent()) {
1100 // This is a call safepoint
1101 // TODO: scan the range to find the statepoint
1102 // TODO: check that the following instruction is not a gc_relocate or
1105 // This is an invoke safepoint
1106 InvokeInst *invoke = dyn_cast<InvokeInst>(bounds.first);
1107 assert(invoke && "only continues over invokes!");
1108 assert(invoke->getNormalDest() == bounds.second->getParent() &&
1109 "safepoint should continue into normal exit block");
1113 static int find_index(const SmallVectorImpl<Value *> &livevec, Value *val) {
1114 auto itr = std::find(livevec.begin(), livevec.end(), val);
1115 assert(livevec.end() != itr);
1116 size_t index = std::distance(livevec.begin(), itr);
1117 assert(index < livevec.size());
1121 // Create new attribute set containing only attributes which can be transfered
1122 // from original call to the safepoint.
1123 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1126 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1127 unsigned index = AS.getSlotIndex(Slot);
1129 if (index == AttributeSet::ReturnIndex ||
1130 index == AttributeSet::FunctionIndex) {
1132 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1134 Attribute attr = *it;
1136 // Do not allow certain attributes - just skip them
1137 // Safepoint can not be read only or read none.
1138 if (attr.hasAttribute(Attribute::ReadNone) ||
1139 attr.hasAttribute(Attribute::ReadOnly))
1142 ret = ret.addAttributes(
1143 AS.getContext(), index,
1144 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1148 // Just skip parameter attributes for now
1154 /// Helper function to place all gc relocates necessary for the given
1157 /// liveVariables - list of variables to be relocated.
1158 /// liveStart - index of the first live variable.
1159 /// basePtrs - base pointers.
1160 /// statepointToken - statepoint instruction to which relocates should be
1162 /// Builder - Llvm IR builder to be used to construct new calls.
1163 /// Returns array with newly created relocates.
1164 static std::vector<llvm::Instruction *>
1165 CreateGCRelocates(const SmallVectorImpl<llvm::Value *> &liveVariables,
1166 const int liveStart,
1167 const SmallVectorImpl<llvm::Value *> &basePtrs,
1168 Instruction *statepointToken, IRBuilder<> Builder) {
1170 std::vector<llvm::Instruction *> newDefs;
1172 Module *M = statepointToken->getParent()->getParent()->getParent();
1174 for (unsigned i = 0; i < liveVariables.size(); i++) {
1175 // We generate a (potentially) unique declaration for every pointer type
1176 // combination. This results is some blow up the function declarations in
1177 // the IR, but removes the need for argument bitcasts which shrinks the IR
1178 // greatly and makes it much more readable.
1179 std::vector<Type *> types; // one per 'any' type
1180 types.push_back(liveVariables[i]->getType()); // result type
1181 Value *gc_relocate_decl = Intrinsic::getDeclaration(
1182 M, Intrinsic::experimental_gc_relocate, types);
1184 // Generate the gc.relocate call and save the result
1186 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1187 liveStart + find_index(liveVariables, basePtrs[i]));
1188 Value *liveIdx = ConstantInt::get(
1189 Type::getInt32Ty(M->getContext()),
1190 liveStart + find_index(liveVariables, liveVariables[i]));
1192 // only specify a debug name if we can give a useful one
1193 Value *reloc = Builder.CreateCall3(
1194 gc_relocate_decl, statepointToken, baseIdx, liveIdx,
1195 liveVariables[i]->hasName() ? liveVariables[i]->getName() + ".relocated"
1197 // Trick CodeGen into thinking there are lots of free registers at this
1199 cast<CallInst>(reloc)->setCallingConv(CallingConv::Cold);
1201 newDefs.push_back(cast<Instruction>(reloc));
1203 assert(newDefs.size() == liveVariables.size() &&
1204 "missing or extra redefinition at safepoint");
1210 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1211 const SmallVectorImpl<llvm::Value *> &basePtrs,
1212 const SmallVectorImpl<llvm::Value *> &liveVariables,
1214 PartiallyConstructedSafepointRecord &result) {
1215 assert(basePtrs.size() == liveVariables.size());
1216 assert(isStatepoint(CS) &&
1217 "This method expects to be rewriting a statepoint");
1219 BasicBlock *BB = CS.getInstruction()->getParent();
1221 Function *F = BB->getParent();
1222 assert(F && "must be set");
1223 Module *M = F->getParent();
1224 assert(M && "must be set");
1226 // We're not changing the function signature of the statepoint since the gc
1227 // arguments go into the var args section.
1228 Function *gc_statepoint_decl = CS.getCalledFunction();
1230 // Then go ahead and use the builder do actually do the inserts. We insert
1231 // immediately before the previous instruction under the assumption that all
1232 // arguments will be available here. We can't insert afterwards since we may
1233 // be replacing a terminator.
1234 Instruction *insertBefore = CS.getInstruction();
1235 IRBuilder<> Builder(insertBefore);
1236 // Copy all of the arguments from the original statepoint - this includes the
1237 // target, call args, and deopt args
1238 std::vector<llvm::Value *> args;
1239 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1240 // TODO: Clear the 'needs rewrite' flag
1242 // add all the pointers to be relocated (gc arguments)
1243 // Capture the start of the live variable list for use in the gc_relocates
1244 const int live_start = args.size();
1245 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1247 // Create the statepoint given all the arguments
1248 Instruction *token = nullptr;
1249 AttributeSet return_attributes;
1251 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1253 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1254 call->setTailCall(toReplace->isTailCall());
1255 call->setCallingConv(toReplace->getCallingConv());
1257 // Currently we will fail on parameter attributes and on certain
1258 // function attributes.
1259 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1260 // In case if we can handle this set of sttributes - set up function attrs
1261 // directly on statepoint and return attrs later for gc_result intrinsic.
1262 call->setAttributes(new_attrs.getFnAttributes());
1263 return_attributes = new_attrs.getRetAttributes();
1267 // Put the following gc_result and gc_relocate calls immediately after the
1268 // the old call (which we're about to delete)
1269 BasicBlock::iterator next(toReplace);
1270 assert(BB->end() != next && "not a terminator, must have next");
1272 Instruction *IP = &*(next);
1273 Builder.SetInsertPoint(IP);
1274 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1276 } else if (CS.isInvoke()) {
1277 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1279 // Insert the new invoke into the old block. We'll remove the old one in a
1280 // moment at which point this will become the new terminator for the
1282 InvokeInst *invoke = InvokeInst::Create(
1283 gc_statepoint_decl, toReplace->getNormalDest(),
1284 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1285 invoke->setCallingConv(toReplace->getCallingConv());
1287 // Currently we will fail on parameter attributes and on certain
1288 // function attributes.
1289 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1290 // In case if we can handle this set of sttributes - set up function attrs
1291 // directly on statepoint and return attrs later for gc_result intrinsic.
1292 invoke->setAttributes(new_attrs.getFnAttributes());
1293 return_attributes = new_attrs.getRetAttributes();
1297 // Generate gc relocates in exceptional path
1298 BasicBlock *unwindBlock = normalizeBBForInvokeSafepoint(
1299 toReplace->getUnwindDest(), invoke->getParent(), P);
1301 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1302 Builder.SetInsertPoint(IP);
1303 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1305 // Extract second element from landingpad return value. We will attach
1306 // exceptional gc relocates to it.
1307 const unsigned idx = 1;
1308 Instruction *exceptional_token =
1309 cast<Instruction>(Builder.CreateExtractValue(
1310 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1311 result.exceptional_relocates_token = exceptional_token;
1313 // Just throw away return value. We will use the one we got for normal
1315 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1316 exceptional_token, Builder);
1318 // Generate gc relocates and returns for normal block
1319 BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
1320 toReplace->getNormalDest(), invoke->getParent(), P);
1322 IP = &*(normalDest->getFirstInsertionPt());
1323 Builder.SetInsertPoint(IP);
1325 // gc relocates will be generated later as if it were regular call
1328 llvm_unreachable("unexpect type of CallSite");
1332 // Take the name of the original value call if it had one.
1333 token->takeName(CS.getInstruction());
1335 // The GCResult is already inserted, we just need to find it
1336 Instruction *gc_result = nullptr;
1338 Instruction *toReplace = CS.getInstruction();
1339 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1340 "only valid use before rewrite is gc.result");
1341 if (toReplace->hasOneUse()) {
1342 Instruction *GCResult = cast<Instruction>(*toReplace->user_begin());
1343 assert(isGCResult(GCResult));
1344 gc_result = GCResult;
1348 // Update the gc.result of the original statepoint (if any) to use the newly
1349 // inserted statepoint. This is safe to do here since the token can't be
1350 // considered a live reference.
1351 CS.getInstruction()->replaceAllUsesWith(token);
1353 // Second, create a gc.relocate for every live variable
1354 std::vector<llvm::Instruction *> newDefs =
1355 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1357 // Need to pass through the last part of the safepoint block so that we
1358 // don't accidentally update uses in a following gc.relocate which is
1359 // still conceptually part of the same safepoint. Gah.
1360 Instruction *last = nullptr;
1361 if (!newDefs.empty()) {
1362 last = newDefs.back();
1363 } else if (gc_result) {
1368 assert(last && "can't be null");
1369 const auto bounds = std::make_pair(token, last);
1371 // Sanity check our results - this is slightly non-trivial due to invokes
1372 VerifySafepointBounds(bounds);
1374 result.safepoint = bounds;
1378 struct name_ordering {
1381 bool operator()(name_ordering const &a, name_ordering const &b) {
1382 return -1 == a.derived->getName().compare(b.derived->getName());
1386 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1387 SmallVectorImpl<Value *> &livevec) {
1388 assert(basevec.size() == livevec.size());
1390 std::vector<name_ordering> temp;
1391 for (size_t i = 0; i < basevec.size(); i++) {
1393 v.base = basevec[i];
1394 v.derived = livevec[i];
1397 std::sort(temp.begin(), temp.end(), name_ordering());
1398 for (size_t i = 0; i < basevec.size(); i++) {
1399 basevec[i] = temp[i].base;
1400 livevec[i] = temp[i].derived;
1404 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1405 // which make the relocations happening at this safepoint explicit.
1407 // WARNING: Does not do any fixup to adjust users of the original live
1408 // values. That's the callers responsibility.
1410 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1411 PartiallyConstructedSafepointRecord &result) {
1412 std::set<llvm::Value *> liveset = result.liveset;
1413 std::map<llvm::Value *, llvm::Value *> base_pairs = result.base_pairs;
1415 // Convert to vector for efficient cross referencing.
1416 SmallVector<Value *, 64> basevec, livevec;
1417 livevec.reserve(liveset.size());
1418 basevec.reserve(liveset.size());
1419 for (Value *L : liveset) {
1420 livevec.push_back(L);
1422 assert(base_pairs.find(L) != base_pairs.end());
1423 Value *base = base_pairs[L];
1424 basevec.push_back(base);
1426 assert(livevec.size() == basevec.size());
1428 // To make the output IR slightly more stable (for use in diffs), ensure a
1429 // fixed order of the values in the safepoint (by sorting the value name).
1430 // The order is otherwise meaningless.
1431 stablize_order(basevec, livevec);
1433 // Do the actual rewriting and delete the old statepoint
1434 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1435 CS.getInstruction()->eraseFromParent();
1438 // Helper function for the relocationViaAlloca.
1439 // It receives iterator to the statepoint gc relocates and emits store to the
1441 // location (via allocaMap) for the each one of them.
1442 // Add visited values into the visitedLiveValues set we will later use them
1443 // for sanity check.
1445 insertRelocationStores(iterator_range<Value::user_iterator> gcRelocs,
1446 DenseMap<Value *, Value *> &allocaMap,
1447 DenseSet<Value *> &visitedLiveValues) {
1449 for (User *U : gcRelocs) {
1450 if (!isa<IntrinsicInst>(U))
1453 IntrinsicInst *relocatedValue = cast<IntrinsicInst>(U);
1455 // We only care about relocates
1456 if (relocatedValue->getIntrinsicID() !=
1457 Intrinsic::experimental_gc_relocate) {
1461 GCRelocateOperands relocateOperands(relocatedValue);
1462 Value *originalValue = const_cast<Value *>(relocateOperands.derivedPtr());
1463 assert(allocaMap.count(originalValue));
1464 Value *alloca = allocaMap[originalValue];
1466 // Emit store into the related alloca
1467 StoreInst *store = new StoreInst(relocatedValue, alloca);
1468 store->insertAfter(relocatedValue);
1471 visitedLiveValues.insert(originalValue);
1476 /// do all the relocation update via allocas and mem2reg
1477 static void relocationViaAlloca(
1478 Function &F, DominatorTree &DT, const std::vector<Value *> &live,
1479 const std::vector<struct PartiallyConstructedSafepointRecord> &records) {
1481 int initialAllocaNum = 0;
1483 // record initial number of allocas
1484 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1486 if (isa<AllocaInst>(*itr))
1491 // TODO-PERF: change data structures, reserve
1492 DenseMap<Value *, Value *> allocaMap;
1493 SmallVector<AllocaInst *, 200> PromotableAllocas;
1494 PromotableAllocas.reserve(live.size());
1496 // emit alloca for each live gc pointer
1497 for (unsigned i = 0; i < live.size(); i++) {
1498 Value *liveValue = live[i];
1499 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1500 F.getEntryBlock().getFirstNonPHI());
1501 allocaMap[liveValue] = alloca;
1502 PromotableAllocas.push_back(alloca);
1505 // The next two loops are part of the same conceptual operation. We need to
1506 // insert a store to the alloca after the original def and at each
1507 // redefinition. We need to insert a load before each use. These are split
1508 // into distinct loops for performance reasons.
1510 // update gc pointer after each statepoint
1511 // either store a relocated value or null (if no relocated value found for
1512 // this gc pointer and it is not a gc_result)
1513 // this must happen before we update the statepoint with load of alloca
1514 // otherwise we lose the link between statepoint and old def
1515 for (size_t i = 0; i < records.size(); i++) {
1516 const struct PartiallyConstructedSafepointRecord &info = records[i];
1517 Value *statepoint = info.safepoint.first;
1519 // This will be used for consistency check
1520 DenseSet<Value *> visitedLiveValues;
1522 // Insert stores for normal statepoint gc relocates
1523 insertRelocationStores(statepoint->users(), allocaMap, visitedLiveValues);
1525 // In case if it was invoke statepoint
1526 // we will insert stores for exceptional path gc relocates.
1527 if (isa<InvokeInst>(statepoint)) {
1528 insertRelocationStores(info.exceptional_relocates_token->users(),
1529 allocaMap, visitedLiveValues);
1533 // For consistency check store null's into allocas for values that are not
1535 // by this statepoint.
1536 for (auto Pair : allocaMap) {
1537 Value *def = Pair.first;
1538 Value *alloca = Pair.second;
1540 // This value was relocated
1541 if (visitedLiveValues.count(def)) {
1544 // Result should not be relocated
1545 if (def == info.result) {
1550 ConstantPointerNull::get(cast<PointerType>(def->getType()));
1551 StoreInst *store = new StoreInst(CPN, alloca);
1552 store->insertBefore(info.safepoint.second);
1556 // update use with load allocas and add store for gc_relocated
1557 for (auto Pair : allocaMap) {
1558 Value *def = Pair.first;
1559 Value *alloca = Pair.second;
1561 // we pre-record the uses of allocas so that we dont have to worry about
1563 // that change the user information.
1564 SmallVector<Instruction *, 20> uses;
1565 // PERF: trade a linear scan for repeated reallocation
1566 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1567 for (User *U : def->users()) {
1568 if (!isa<ConstantExpr>(U)) {
1569 // If the def has a ConstantExpr use, then the def is either a
1570 // ConstantExpr use itself or null. In either case
1571 // (recursively in the first, directly in the second), the oop
1572 // it is ultimately dependent on is null and this particular
1573 // use does not need to be fixed up.
1574 uses.push_back(cast<Instruction>(U));
1578 std::sort(uses.begin(), uses.end());
1579 auto last = std::unique(uses.begin(), uses.end());
1580 uses.erase(last, uses.end());
1582 for (Instruction *use : uses) {
1583 if (isa<PHINode>(use)) {
1584 PHINode *phi = cast<PHINode>(use);
1585 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1586 if (def == phi->getIncomingValue(i)) {
1587 LoadInst *load = new LoadInst(
1588 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1589 phi->setIncomingValue(i, load);
1593 LoadInst *load = new LoadInst(alloca, "", use);
1594 use->replaceUsesOfWith(def, load);
1598 // emit store for the initial gc value
1599 // store must be inserted after load, otherwise store will be in alloca's
1600 // use list and an extra load will be inserted before it
1601 StoreInst *store = new StoreInst(def, alloca);
1602 if (isa<Instruction>(def)) {
1603 store->insertAfter(cast<Instruction>(def));
1605 assert((isa<Argument>(def) || isa<GlobalVariable>(def) ||
1606 (isa<Constant>(def) && cast<Constant>(def)->isNullValue())) &&
1607 "Must be argument or global");
1608 store->insertAfter(cast<Instruction>(alloca));
1612 assert(PromotableAllocas.size() == live.size() &&
1613 "we must have the same allocas with lives");
1614 if (!PromotableAllocas.empty()) {
1615 // apply mem2reg to promote alloca to SSA
1616 PromoteMemToReg(PromotableAllocas, DT);
1620 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1622 if (isa<AllocaInst>(*itr))
1625 assert(initialAllocaNum == 0 && "We must not introduce any extra allocas");
1629 /// Implement a unique function which doesn't require we sort the input
1630 /// vector. Doing so has the effect of changing the output of a couple of
1631 /// tests in ways which make them less useful in testing fused safepoints.
1632 template <typename T> static void unique_unsorted(std::vector<T> &vec) {
1635 vec.reserve(vec.size());
1636 std::swap(tmp, vec);
1637 for (auto V : tmp) {
1638 if (seen.insert(V).second) {
1644 static Function *getUseHolder(Module &M) {
1645 FunctionType *ftype =
1646 FunctionType::get(Type::getVoidTy(M.getContext()), true);
1647 Function *Func = cast<Function>(M.getOrInsertFunction("__tmp_use", ftype));
1651 /// Insert holders so that each Value is obviously live through the entire
1652 /// liftetime of the call.
1653 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1654 std::vector<CallInst *> &holders) {
1655 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1656 Function *Func = getUseHolder(*M);
1658 // For call safepoints insert dummy calls right after safepoint
1659 BasicBlock::iterator next(CS.getInstruction());
1661 CallInst *base_holder = CallInst::Create(Func, Values, "", next);
1662 holders.push_back(base_holder);
1663 } else if (CS.isInvoke()) {
1664 // For invoke safepooints insert dummy calls both in normal and
1665 // exceptional destination blocks
1666 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1667 CallInst *normal_holder = CallInst::Create(
1668 Func, Values, "", invoke->getNormalDest()->getFirstInsertionPt());
1669 CallInst *unwind_holder = CallInst::Create(
1670 Func, Values, "", invoke->getUnwindDest()->getFirstInsertionPt());
1671 holders.push_back(normal_holder);
1672 holders.push_back(unwind_holder);
1674 assert(false && "Unsupported");
1678 static void findLiveReferences(
1679 Function &F, DominatorTree &DT, Pass *P, std::vector<CallSite> &toUpdate,
1680 std::vector<struct PartiallyConstructedSafepointRecord> &records) {
1681 for (size_t i = 0; i < records.size(); i++) {
1682 struct PartiallyConstructedSafepointRecord &info = records[i];
1683 CallSite &CS = toUpdate[i];
1684 analyzeParsePointLiveness(DT, CS, info);
1688 static void addBasesAsLiveValues(std::set<Value *> &liveset,
1689 std::map<Value *, Value *> &base_pairs) {
1690 // Identify any base pointers which are used in this safepoint, but not
1691 // themselves relocated. We need to relocate them so that later inserted
1692 // safepoints can get the properly relocated base register.
1693 DenseSet<Value *> missing;
1694 for (Value *L : liveset) {
1695 assert(base_pairs.find(L) != base_pairs.end());
1696 Value *base = base_pairs[L];
1698 if (liveset.find(base) == liveset.end()) {
1699 assert(base_pairs.find(base) == base_pairs.end());
1700 // uniqued by set insert
1701 missing.insert(base);
1705 // Note that we want these at the end of the list, otherwise
1706 // register placement gets screwed up once we lower to STATEPOINT
1707 // instructions. This is an utter hack, but there doesn't seem to be a
1709 for (Value *base : missing) {
1711 liveset.insert(base);
1712 base_pairs[base] = base;
1714 assert(liveset.size() == base_pairs.size());
1717 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1718 std::vector<CallSite> &toUpdate) {
1720 // sanity check the input
1721 std::set<CallSite> uniqued;
1722 uniqued.insert(toUpdate.begin(), toUpdate.end());
1723 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1725 for (size_t i = 0; i < toUpdate.size(); i++) {
1726 CallSite &CS = toUpdate[i];
1727 assert(CS.getInstruction()->getParent()->getParent() == &F);
1728 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1732 // A list of dummy calls added to the IR to keep various values obviously
1733 // live in the IR. We'll remove all of these when done.
1734 std::vector<CallInst *> holders;
1736 // Insert a dummy call with all of the arguments to the vm_state we'll need
1737 // for the actual safepoint insertion. This ensures reference arguments in
1738 // the deopt argument list are considered live through the safepoint (and
1739 // thus makes sure they get relocated.)
1740 for (size_t i = 0; i < toUpdate.size(); i++) {
1741 CallSite &CS = toUpdate[i];
1742 Statepoint StatepointCS(CS);
1744 SmallVector<Value *, 64> DeoptValues;
1745 for (Use &U : StatepointCS.vm_state_args()) {
1746 Value *Arg = cast<Value>(&U);
1747 if (isGCPointerType(Arg->getType()))
1748 DeoptValues.push_back(Arg);
1750 insertUseHolderAfter(CS, DeoptValues, holders);
1753 std::vector<struct PartiallyConstructedSafepointRecord> records;
1754 records.reserve(toUpdate.size());
1755 for (size_t i = 0; i < toUpdate.size(); i++) {
1756 struct PartiallyConstructedSafepointRecord info;
1757 records.push_back(info);
1759 assert(records.size() == toUpdate.size());
1761 // A) Identify all gc pointers which are staticly live at the given call
1763 findLiveReferences(F, DT, P, toUpdate, records);
1765 // B) Find the base pointers for each live pointer
1766 /* scope for caching */ {
1767 // Cache the 'defining value' relation used in the computation and
1768 // insertion of base phis and selects. This ensures that we don't insert
1769 // large numbers of duplicate base_phis.
1770 DefiningValueMapTy DVCache;
1772 for (size_t i = 0; i < records.size(); i++) {
1773 struct PartiallyConstructedSafepointRecord &info = records[i];
1774 CallSite &CS = toUpdate[i];
1775 findBasePointers(DT, DVCache, CS, info);
1777 } // end of cache scope
1779 // The base phi insertion logic (for any safepoint) may have inserted new
1780 // instructions which are now live at some safepoint. The simplest such
1783 // phi a <-- will be a new base_phi here
1784 // safepoint 1 <-- that needs to be live here
1788 std::set<llvm::Value *> allInsertedDefs;
1789 for (size_t i = 0; i < records.size(); i++) {
1790 struct PartiallyConstructedSafepointRecord &info = records[i];
1791 allInsertedDefs.insert(info.newInsertedDefs.begin(),
1792 info.newInsertedDefs.end());
1795 // We insert some dummy calls after each safepoint to definitely hold live
1796 // the base pointers which were identified for that safepoint. We'll then
1797 // ask liveness for _every_ base inserted to see what is now live. Then we
1798 // remove the dummy calls.
1799 holders.reserve(holders.size() + records.size());
1800 for (size_t i = 0; i < records.size(); i++) {
1801 struct PartiallyConstructedSafepointRecord &info = records[i];
1802 CallSite &CS = toUpdate[i];
1804 SmallVector<Value *, 128> Bases;
1805 for (auto Pair : info.base_pairs) {
1806 Bases.push_back(Pair.second);
1808 insertUseHolderAfter(CS, Bases, holders);
1811 // Add the bases explicitly to the live vector set. This may result in a few
1812 // extra relocations, but the base has to be available whenever a pointer
1813 // derived from it is used. Thus, we need it to be part of the statepoint's
1814 // gc arguments list. TODO: Introduce an explicit notion (in the following
1815 // code) of the GC argument list as seperate from the live Values at a
1816 // given statepoint.
1817 for (size_t i = 0; i < records.size(); i++) {
1818 struct PartiallyConstructedSafepointRecord &info = records[i];
1819 addBasesAsLiveValues(info.liveset, info.base_pairs);
1822 // If we inserted any new values, we need to adjust our notion of what is
1823 // live at a particular safepoint.
1824 if (!allInsertedDefs.empty()) {
1825 fixupLiveReferences(F, DT, P, allInsertedDefs, toUpdate, records);
1827 if (PrintBasePointers) {
1828 for (size_t i = 0; i < records.size(); i++) {
1829 struct PartiallyConstructedSafepointRecord &info = records[i];
1830 errs() << "Base Pairs: (w/Relocation)\n";
1831 for (auto Pair : info.base_pairs) {
1832 errs() << " derived %" << Pair.first->getName() << " base %"
1833 << Pair.second->getName() << "\n";
1837 for (size_t i = 0; i < holders.size(); i++) {
1838 holders[i]->eraseFromParent();
1839 holders[i] = nullptr;
1843 // Now run through and replace the existing statepoints with new ones with
1844 // the live variables listed. We do not yet update uses of the values being
1845 // relocated. We have references to live variables that need to
1846 // survive to the last iteration of this loop. (By construction, the
1847 // previous statepoint can not be a live variable, thus we can and remove
1848 // the old statepoint calls as we go.)
1849 for (size_t i = 0; i < records.size(); i++) {
1850 struct PartiallyConstructedSafepointRecord &info = records[i];
1851 CallSite &CS = toUpdate[i];
1852 makeStatepointExplicit(DT, CS, P, info);
1854 toUpdate.clear(); // prevent accident use of invalid CallSites
1856 // In case if we inserted relocates in a different basic block than the
1857 // original safepoint (this can happen for invokes). We need to be sure that
1858 // original values were not used in any of the phi nodes at the
1859 // beginning of basic block containing them. Because we know that all such
1860 // blocks will have single predecessor we can safely assume that all phi
1861 // nodes have single entry (because of normalizeBBForInvokeSafepoint).
1862 // Just remove them all here.
1863 for (size_t i = 0; i < records.size(); i++) {
1864 Instruction *I = records[i].safepoint.first;
1866 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
1867 FoldSingleEntryPHINodes(invoke->getNormalDest());
1868 assert(!isa<PHINode>(invoke->getNormalDest()->begin()));
1870 FoldSingleEntryPHINodes(invoke->getUnwindDest());
1871 assert(!isa<PHINode>(invoke->getUnwindDest()->begin()));
1875 // Do all the fixups of the original live variables to their relocated selves
1876 std::vector<Value *> live;
1877 for (size_t i = 0; i < records.size(); i++) {
1878 struct PartiallyConstructedSafepointRecord &info = records[i];
1879 // We can't simply save the live set from the original insertion. One of
1880 // the live values might be the result of a call which needs a safepoint.
1881 // That Value* no longer exists and we need to use the new gc_result.
1882 // Thankfully, the liveset is embedded in the statepoint (and updated), so
1883 // we just grab that.
1884 Statepoint statepoint(info.safepoint.first);
1885 live.insert(live.end(), statepoint.gc_args_begin(),
1886 statepoint.gc_args_end());
1888 unique_unsorted(live);
1891 for (auto ptr : live) {
1892 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
1895 relocationViaAlloca(F, DT, live, records);
1896 return !records.empty();
1899 /// Returns true if this function should be rewritten by this pass. The main
1900 /// point of this function is as an extension point for custom logic.
1901 static bool shouldRewriteStatepointsIn(Function &F) {
1902 // TODO: This should check the GCStrategy
1906 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
1907 // Nothing to do for declarations.
1908 if (F.isDeclaration() || F.empty())
1911 // Policy choice says not to rewrite - the most common reason is that we're
1912 // compiling code without a GCStrategy.
1913 if (!shouldRewriteStatepointsIn(F))
1916 // Gather all the statepoints which need rewritten.
1917 std::vector<CallSite> ParsePointNeeded;
1918 for (inst_iterator itr = inst_begin(F), end = inst_end(F); itr != end;
1920 // TODO: only the ones with the flag set!
1921 if (isStatepoint(*itr))
1922 ParsePointNeeded.push_back(CallSite(&*itr));
1925 // Return early if no work to do.
1926 if (ParsePointNeeded.empty())
1929 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1930 return insertParsePoints(F, DT, this, ParsePointNeeded);