1 //===- RewriteStatepointsForGC.cpp - Make GC relocations explicit ---------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Rewrite an existing set of gc.statepoints such that they make potential
11 // relocations performed by the garbage collector explicit in the IR.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Pass.h"
16 #include "llvm/Analysis/CFG.h"
17 #include "llvm/ADT/SetOperations.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/DenseSet.h"
20 #include "llvm/IR/BasicBlock.h"
21 #include "llvm/IR/CallSite.h"
22 #include "llvm/IR/Dominators.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/IRBuilder.h"
25 #include "llvm/IR/InstIterator.h"
26 #include "llvm/IR/Instructions.h"
27 #include "llvm/IR/Intrinsics.h"
28 #include "llvm/IR/IntrinsicInst.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/IR/Statepoint.h"
31 #include "llvm/IR/Value.h"
32 #include "llvm/IR/Verifier.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/CommandLine.h"
35 #include "llvm/Transforms/Scalar.h"
36 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
37 #include "llvm/Transforms/Utils/Cloning.h"
38 #include "llvm/Transforms/Utils/Local.h"
39 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
41 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
45 // Print tracing output
46 static cl::opt<bool> TraceLSP("trace-rewrite-statepoints", cl::Hidden,
49 // Print the liveset found at the insert location
50 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
52 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size",
53 cl::Hidden, cl::init(false));
54 // Print out the base pointers for debugging
55 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers",
56 cl::Hidden, cl::init(false));
59 struct RewriteStatepointsForGC : public FunctionPass {
60 static char ID; // Pass identification, replacement for typeid
62 RewriteStatepointsForGC() : FunctionPass(ID) {
63 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry());
65 bool runOnFunction(Function &F) override;
67 void getAnalysisUsage(AnalysisUsage &AU) const override {
68 // We add and rewrite a bunch of instructions, but don't really do much
69 // else. We could in theory preserve a lot more analyses here.
70 AU.addRequired<DominatorTreeWrapperPass>();
75 char RewriteStatepointsForGC::ID = 0;
77 FunctionPass *llvm::createRewriteStatepointsForGCPass() {
78 return new RewriteStatepointsForGC();
81 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
82 "Make relocations explicit at statepoints", false, false)
83 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
84 INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc",
85 "Make relocations explicit at statepoints", false, false)
88 // The type of the internal cache used inside the findBasePointers family
89 // of functions. From the callers perspective, this is an opaque type and
90 // should not be inspected.
92 // In the actual implementation this caches two relations:
93 // - The base relation itself (i.e. this pointer is based on that one)
94 // - The base defining value relation (i.e. before base_phi insertion)
95 // Generally, after the execution of a full findBasePointer call, only the
96 // base relation will remain. Internally, we add a mixture of the two
97 // types, then update all the second type to the first type
98 typedef DenseMap<Value *, Value *> DefiningValueMapTy;
99 typedef DenseSet<llvm::Value *> StatepointLiveSetTy;
101 struct PartiallyConstructedSafepointRecord {
102 /// The set of values known to be live accross this safepoint
103 StatepointLiveSetTy liveset;
105 /// Mapping from live pointers to a base-defining-value
106 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
108 /// Any new values which were added to the IR during base pointer analysis
109 /// for this safepoint
110 DenseSet<llvm::Value *> NewInsertedDefs;
112 /// The *new* gc.statepoint instruction itself. This produces the token
113 /// that normal path gc.relocates and the gc.result are tied to.
114 Instruction *StatepointToken;
116 /// Instruction to which exceptional gc relocates are attached
117 /// Makes it easier to iterate through them during relocationViaAlloca.
118 Instruction *UnwindToken;
122 // TODO: Once we can get to the GCStrategy, this becomes
123 // Optional<bool> isGCManagedPointer(const Value *V) const override {
125 static bool isGCPointerType(const Type *T) {
126 if (const PointerType *PT = dyn_cast<PointerType>(T))
127 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
128 // GC managed heap. We know that a pointer into this heap needs to be
129 // updated and that no other pointer does.
130 return (1 == PT->getAddressSpace());
134 /// Return true if the Value is a gc reference type which is potentially used
135 /// after the instruction 'loc'. This is only used with the edge reachability
136 /// liveness code. Note: It is assumed the V dominates loc.
137 static bool isLiveGCReferenceAt(Value &V, Instruction *loc, DominatorTree &DT,
139 if (!isGCPointerType(V.getType()))
145 // Given assumption that V dominates loc, this may be live
150 static bool isAggWhichContainsGCPtrType(Type *Ty) {
151 if (VectorType *VT = dyn_cast<VectorType>(Ty))
152 return isGCPointerType(VT->getScalarType());
153 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
154 return isGCPointerType(AT->getElementType()) ||
155 isAggWhichContainsGCPtrType(AT->getElementType());
156 if (StructType *ST = dyn_cast<StructType>(Ty))
157 return std::any_of(ST->subtypes().begin(), ST->subtypes().end(),
159 return isGCPointerType(SubType) ||
160 isAggWhichContainsGCPtrType(SubType);
166 // Conservatively identifies any definitions which might be live at the
167 // given instruction. The analysis is performed immediately before the
168 // given instruction. Values defined by that instruction are not considered
169 // live. Values used by that instruction are considered live.
171 // preconditions: valid IR graph, term is either a terminator instruction or
172 // a call instruction, pred is the basic block of term, DT, LI are valid
174 // side effects: none, does not mutate IR
176 // postconditions: populates liveValues as discussed above
177 static void findLiveGCValuesAtInst(Instruction *term, BasicBlock *pred,
178 DominatorTree &DT, LoopInfo *LI,
179 StatepointLiveSetTy &liveValues) {
182 assert(isa<CallInst>(term) || isa<InvokeInst>(term) || term->isTerminator());
184 Function *F = pred->getParent();
186 auto is_live_gc_reference =
187 [&](Value &V) { return isLiveGCReferenceAt(V, term, DT, LI); };
189 // Are there any gc pointer arguments live over this point? This needs to be
190 // special cased since arguments aren't defined in basic blocks.
191 for (Argument &arg : F->args()) {
192 assert(!isAggWhichContainsGCPtrType(arg.getType()) &&
193 "support for FCA unimplemented");
195 if (is_live_gc_reference(arg)) {
196 liveValues.insert(&arg);
200 // Walk through all dominating blocks - the ones which can contain
201 // definitions used in this block - and check to see if any of the values
202 // they define are used in locations potentially reachable from the
203 // interesting instruction.
204 BasicBlock *BBI = pred;
207 errs() << "[LSP] Looking at dominating block " << pred->getName() << "\n";
209 assert(DT.dominates(BBI, pred));
210 assert(isPotentiallyReachable(BBI, pred, &DT) &&
211 "dominated block must be reachable");
213 // Walk through the instructions in dominating blocks and keep any
214 // that have a use potentially reachable from the block we're
215 // considering putting the safepoint in
216 for (Instruction &inst : *BBI) {
218 errs() << "[LSP] Looking at instruction ";
222 if (pred == BBI && (&inst) == term) {
224 errs() << "[LSP] stopped because we encountered the safepoint "
228 // If we're in the block which defines the interesting instruction,
229 // we don't want to include any values as live which are defined
230 // _after_ the interesting line or as part of the line itself
231 // i.e. "term" is the call instruction for a call safepoint, the
232 // results of the call should not be considered live in that stackmap
236 assert(!isAggWhichContainsGCPtrType(inst.getType()) &&
237 "support for FCA unimplemented");
239 if (is_live_gc_reference(inst)) {
241 errs() << "[LSP] found live value for this safepoint ";
245 liveValues.insert(&inst);
248 if (!DT.getNode(BBI)->getIDom()) {
249 assert(BBI == &F->getEntryBlock() &&
250 "failed to find a dominator for something other than "
254 BBI = DT.getNode(BBI)->getIDom()->getBlock();
258 static bool order_by_name(llvm::Value *a, llvm::Value *b) {
259 if (a->hasName() && b->hasName()) {
260 return -1 == a->getName().compare(b->getName());
261 } else if (a->hasName() && !b->hasName()) {
263 } else if (!a->hasName() && b->hasName()) {
266 // Better than nothing, but not stable
271 /// Find the initial live set. Note that due to base pointer
272 /// insertion, the live set may be incomplete.
274 analyzeParsePointLiveness(DominatorTree &DT, const CallSite &CS,
275 PartiallyConstructedSafepointRecord &result) {
276 Instruction *inst = CS.getInstruction();
278 BasicBlock *BB = inst->getParent();
279 StatepointLiveSetTy liveset;
280 findLiveGCValuesAtInst(inst, BB, DT, nullptr, liveset);
283 // Note: This output is used by several of the test cases
284 // The order of elemtns in a set is not stable, put them in a vec and sort
286 SmallVector<Value *, 64> temp;
287 temp.insert(temp.end(), liveset.begin(), liveset.end());
288 std::sort(temp.begin(), temp.end(), order_by_name);
289 errs() << "Live Variables:\n";
290 for (Value *V : temp) {
291 errs() << " " << V->getName(); // no newline
295 if (PrintLiveSetSize) {
296 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
297 errs() << "Number live values: " << liveset.size() << "\n";
299 result.liveset = liveset;
302 /// Helper function for findBasePointer - Will return a value which either a)
303 /// defines the base pointer for the input or b) blocks the simple search
304 /// (i.e. a PHI or Select of two derived pointers)
305 static Value *findBaseDefiningValue(Value *I) {
306 assert(I->getType()->isPointerTy() &&
307 "Illegal to ask for the base pointer of a non-pointer type");
309 // There are instructions which can never return gc pointer values. Sanity
310 // check that this is actually true.
311 assert(!isa<InsertElementInst>(I) && !isa<ExtractElementInst>(I) &&
312 !isa<ShuffleVectorInst>(I) && "Vector types are not gc pointers");
314 if (isa<Argument>(I))
315 // An incoming argument to the function is a base pointer
316 // We should have never reached here if this argument isn't an gc value
319 if (isa<GlobalVariable>(I))
323 // inlining could possibly introduce phi node that contains
324 // undef if callee has multiple returns
325 if (isa<UndefValue>(I))
326 // utterly meaningless, but useful for dealing with
327 // partially optimized code.
330 // Due to inheritance, this must be _after_ the global variable and undef
332 if (Constant *Con = dyn_cast<Constant>(I)) {
333 assert(!isa<GlobalVariable>(I) && !isa<UndefValue>(I) &&
334 "order of checks wrong!");
335 // Note: Finding a constant base for something marked for relocation
336 // doesn't really make sense. The most likely case is either a) some
337 // screwed up the address space usage or b) your validating against
338 // compiled C++ code w/o the proper separation. The only real exception
339 // is a null pointer. You could have generic code written to index of
340 // off a potentially null value and have proven it null. We also use
341 // null pointers in dead paths of relocation phis (which we might later
342 // want to find a base pointer for).
343 assert(isa<ConstantPointerNull>(Con) &&
344 "null is the only case which makes sense");
348 if (CastInst *CI = dyn_cast<CastInst>(I)) {
349 Value *Def = CI->stripPointerCasts();
350 // If we find a cast instruction here, it means we've found a cast which is
351 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
352 // handle int->ptr conversion.
353 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
354 return findBaseDefiningValue(Def);
357 if (isa<LoadInst>(I))
358 return I; // The value loaded is an gc base itself
360 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
361 // The base of this GEP is the base
362 return findBaseDefiningValue(GEP->getPointerOperand());
364 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
365 switch (II->getIntrinsicID()) {
366 case Intrinsic::experimental_gc_result_ptr:
368 // fall through to general call handling
370 case Intrinsic::experimental_gc_statepoint:
371 case Intrinsic::experimental_gc_result_float:
372 case Intrinsic::experimental_gc_result_int:
373 llvm_unreachable("these don't produce pointers");
374 case Intrinsic::experimental_gc_relocate: {
375 // Rerunning safepoint insertion after safepoints are already
376 // inserted is not supported. It could probably be made to work,
377 // but why are you doing this? There's no good reason.
378 llvm_unreachable("repeat safepoint insertion is not supported");
380 case Intrinsic::gcroot:
381 // Currently, this mechanism hasn't been extended to work with gcroot.
382 // There's no reason it couldn't be, but I haven't thought about the
383 // implications much.
385 "interaction with the gcroot mechanism is not supported");
388 // We assume that functions in the source language only return base
389 // pointers. This should probably be generalized via attributes to support
390 // both source language and internal functions.
391 if (isa<CallInst>(I) || isa<InvokeInst>(I))
394 // I have absolutely no idea how to implement this part yet. It's not
395 // neccessarily hard, I just haven't really looked at it yet.
396 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
398 if (isa<AtomicCmpXchgInst>(I))
399 // A CAS is effectively a atomic store and load combined under a
400 // predicate. From the perspective of base pointers, we just treat it
404 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
405 "binary ops which don't apply to pointers");
407 // The aggregate ops. Aggregates can either be in the heap or on the
408 // stack, but in either case, this is simply a field load. As a result,
409 // this is a defining definition of the base just like a load is.
410 if (isa<ExtractValueInst>(I))
413 // We should never see an insert vector since that would require we be
414 // tracing back a struct value not a pointer value.
415 assert(!isa<InsertValueInst>(I) &&
416 "Base pointer for a struct is meaningless");
418 // The last two cases here don't return a base pointer. Instead, they
419 // return a value which dynamically selects from amoung several base
420 // derived pointers (each with it's own base potentially). It's the job of
421 // the caller to resolve these.
422 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
423 "missing instruction case in findBaseDefiningValing");
427 /// Returns the base defining value for this value.
428 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
429 Value *&Cached = Cache[I];
431 Cached = findBaseDefiningValue(I);
433 assert(Cache[I] != nullptr);
436 dbgs() << "fBDV-cached: " << I->getName() << " -> " << Cached->getName()
442 /// Return a base pointer for this value if known. Otherwise, return it's
443 /// base defining value.
444 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
445 Value *Def = findBaseDefiningValueCached(I, Cache);
446 auto Found = Cache.find(Def);
447 if (Found != Cache.end()) {
448 // Either a base-of relation, or a self reference. Caller must check.
449 return Found->second;
451 // Only a BDV available
455 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
456 /// is it known to be a base pointer? Or do we need to continue searching.
457 static bool isKnownBaseResult(Value *V) {
458 if (!isa<PHINode>(V) && !isa<SelectInst>(V)) {
459 // no recursion possible
462 if (isa<Instruction>(V) &&
463 cast<Instruction>(V)->getMetadata("is_base_value")) {
464 // This is a previously inserted base phi or select. We know
465 // that this is a base value.
469 // We need to keep searching
473 // TODO: find a better name for this
477 enum Status { Unknown, Base, Conflict };
479 PhiState(Status s, Value *b = nullptr) : status(s), base(b) {
480 assert(status != Base || b);
482 PhiState(Value *b) : status(Base), base(b) {}
483 PhiState() : status(Unknown), base(nullptr) {}
485 Status getStatus() const { return status; }
486 Value *getBase() const { return base; }
488 bool isBase() const { return getStatus() == Base; }
489 bool isUnknown() const { return getStatus() == Unknown; }
490 bool isConflict() const { return getStatus() == Conflict; }
492 bool operator==(const PhiState &other) const {
493 return base == other.base && status == other.status;
496 bool operator!=(const PhiState &other) const { return !(*this == other); }
499 errs() << status << " (" << base << " - "
500 << (base ? base->getName() : "nullptr") << "): ";
505 Value *base; // non null only if status == base
508 typedef DenseMap<Value *, PhiState> ConflictStateMapTy;
509 // Values of type PhiState form a lattice, and this is a helper
510 // class that implementes the meet operation. The meat of the meet
511 // operation is implemented in MeetPhiStates::pureMeet
512 class MeetPhiStates {
514 // phiStates is a mapping from PHINodes and SelectInst's to PhiStates.
515 explicit MeetPhiStates(const ConflictStateMapTy &phiStates)
516 : phiStates(phiStates) {}
518 // Destructively meet the current result with the base V. V can
519 // either be a merge instruction (SelectInst / PHINode), in which
520 // case its status is looked up in the phiStates map; or a regular
521 // SSA value, in which case it is assumed to be a base.
522 void meetWith(Value *V) {
523 PhiState otherState = getStateForBDV(V);
524 assert((MeetPhiStates::pureMeet(otherState, currentResult) ==
525 MeetPhiStates::pureMeet(currentResult, otherState)) &&
526 "math is wrong: meet does not commute!");
527 currentResult = MeetPhiStates::pureMeet(otherState, currentResult);
530 PhiState getResult() const { return currentResult; }
533 const ConflictStateMapTy &phiStates;
534 PhiState currentResult;
536 /// Return a phi state for a base defining value. We'll generate a new
537 /// base state for known bases and expect to find a cached state otherwise
538 PhiState getStateForBDV(Value *baseValue) {
539 if (isKnownBaseResult(baseValue)) {
540 return PhiState(baseValue);
542 return lookupFromMap(baseValue);
546 PhiState lookupFromMap(Value *V) {
547 auto I = phiStates.find(V);
548 assert(I != phiStates.end() && "lookup failed!");
552 static PhiState pureMeet(const PhiState &stateA, const PhiState &stateB) {
553 switch (stateA.getStatus()) {
554 case PhiState::Unknown:
558 assert(stateA.getBase() && "can't be null");
559 if (stateB.isUnknown())
562 if (stateB.isBase()) {
563 if (stateA.getBase() == stateB.getBase()) {
564 assert(stateA == stateB && "equality broken!");
567 return PhiState(PhiState::Conflict);
569 assert(stateB.isConflict() && "only three states!");
570 return PhiState(PhiState::Conflict);
572 case PhiState::Conflict:
575 llvm_unreachable("only three states!");
579 /// For a given value or instruction, figure out what base ptr it's derived
580 /// from. For gc objects, this is simply itself. On success, returns a value
581 /// which is the base pointer. (This is reliable and can be used for
582 /// relocation.) On failure, returns nullptr.
583 static Value *findBasePointer(Value *I, DefiningValueMapTy &cache,
584 DenseSet<llvm::Value *> &NewInsertedDefs) {
585 Value *def = findBaseOrBDV(I, cache);
587 if (isKnownBaseResult(def)) {
591 // Here's the rough algorithm:
592 // - For every SSA value, construct a mapping to either an actual base
593 // pointer or a PHI which obscures the base pointer.
594 // - Construct a mapping from PHI to unknown TOP state. Use an
595 // optimistic algorithm to propagate base pointer information. Lattice
600 // When algorithm terminates, all PHIs will either have a single concrete
601 // base or be in a conflict state.
602 // - For every conflict, insert a dummy PHI node without arguments. Add
603 // these to the base[Instruction] = BasePtr mapping. For every
604 // non-conflict, add the actual base.
605 // - For every conflict, add arguments for the base[a] of each input
608 // Note: A simpler form of this would be to add the conflict form of all
609 // PHIs without running the optimistic algorithm. This would be
610 // analougous to pessimistic data flow and would likely lead to an
611 // overall worse solution.
613 ConflictStateMapTy states;
614 states[def] = PhiState();
615 // Recursively fill in all phis & selects reachable from the initial one
616 // for which we don't already know a definite base value for
617 // TODO: This should be rewritten with a worklist
621 // Since we're adding elements to 'states' as we run, we can't keep
622 // iterators into the set.
623 SmallVector<Value*, 16> Keys;
624 Keys.reserve(states.size());
625 for (auto Pair : states) {
626 Value *V = Pair.first;
629 for (Value *v : Keys) {
630 assert(!isKnownBaseResult(v) && "why did it get added?");
631 if (PHINode *phi = dyn_cast<PHINode>(v)) {
632 assert(phi->getNumIncomingValues() > 0 &&
633 "zero input phis are illegal");
634 for (Value *InVal : phi->incoming_values()) {
635 Value *local = findBaseOrBDV(InVal, cache);
636 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
637 states[local] = PhiState();
641 } else if (SelectInst *sel = dyn_cast<SelectInst>(v)) {
642 Value *local = findBaseOrBDV(sel->getTrueValue(), cache);
643 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
644 states[local] = PhiState();
647 local = findBaseOrBDV(sel->getFalseValue(), cache);
648 if (!isKnownBaseResult(local) && states.find(local) == states.end()) {
649 states[local] = PhiState();
657 errs() << "States after initialization:\n";
658 for (auto Pair : states) {
659 Instruction *v = cast<Instruction>(Pair.first);
660 PhiState state = Pair.second;
666 // TODO: come back and revisit the state transitions around inputs which
667 // have reached conflict state. The current version seems too conservative.
669 bool progress = true;
672 size_t oldSize = states.size();
675 // We're only changing keys in this loop, thus safe to keep iterators
676 for (auto Pair : states) {
677 MeetPhiStates calculateMeet(states);
678 Value *v = Pair.first;
679 assert(!isKnownBaseResult(v) && "why did it get added?");
680 if (SelectInst *select = dyn_cast<SelectInst>(v)) {
681 calculateMeet.meetWith(findBaseOrBDV(select->getTrueValue(), cache));
682 calculateMeet.meetWith(findBaseOrBDV(select->getFalseValue(), cache));
684 for (Value *Val : cast<PHINode>(v)->incoming_values())
685 calculateMeet.meetWith(findBaseOrBDV(Val, cache));
687 PhiState oldState = states[v];
688 PhiState newState = calculateMeet.getResult();
689 if (oldState != newState) {
691 states[v] = newState;
695 assert(oldSize <= states.size());
696 assert(oldSize == states.size() || progress);
700 errs() << "States after meet iteration:\n";
701 for (auto Pair : states) {
702 Instruction *v = cast<Instruction>(Pair.first);
703 PhiState state = Pair.second;
709 // Insert Phis for all conflicts
710 // We want to keep naming deterministic in the loop that follows, so
711 // sort the keys before iteration. This is useful in allowing us to
712 // write stable tests. Note that there is no invalidation issue here.
713 SmallVector<Value*, 16> Keys;
714 Keys.reserve(states.size());
715 for (auto Pair : states) {
716 Value *V = Pair.first;
719 std::sort(Keys.begin(), Keys.end(), order_by_name);
720 // TODO: adjust naming patterns to avoid this order of iteration dependency
721 for (Value *V : Keys) {
722 Instruction *v = cast<Instruction>(V);
723 PhiState state = states[V];
724 assert(!isKnownBaseResult(v) && "why did it get added?");
725 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
726 if (!state.isConflict())
729 if (isa<PHINode>(v)) {
731 std::distance(pred_begin(v->getParent()), pred_end(v->getParent()));
732 assert(num_preds > 0 && "how did we reach here");
733 PHINode *phi = PHINode::Create(v->getType(), num_preds, "base_phi", v);
734 NewInsertedDefs.insert(phi);
735 // Add metadata marking this as a base value
736 auto *const_1 = ConstantInt::get(
738 v->getParent()->getParent()->getParent()->getContext()),
740 auto MDConst = ConstantAsMetadata::get(const_1);
741 MDNode *md = MDNode::get(
742 v->getParent()->getParent()->getParent()->getContext(), MDConst);
743 phi->setMetadata("is_base_value", md);
744 states[v] = PhiState(PhiState::Conflict, phi);
746 SelectInst *sel = cast<SelectInst>(v);
747 // The undef will be replaced later
748 UndefValue *undef = UndefValue::get(sel->getType());
749 SelectInst *basesel = SelectInst::Create(sel->getCondition(), undef,
750 undef, "base_select", sel);
751 NewInsertedDefs.insert(basesel);
752 // Add metadata marking this as a base value
753 auto *const_1 = ConstantInt::get(
755 v->getParent()->getParent()->getParent()->getContext()),
757 auto MDConst = ConstantAsMetadata::get(const_1);
758 MDNode *md = MDNode::get(
759 v->getParent()->getParent()->getParent()->getContext(), MDConst);
760 basesel->setMetadata("is_base_value", md);
761 states[v] = PhiState(PhiState::Conflict, basesel);
765 // Fixup all the inputs of the new PHIs
766 for (auto Pair : states) {
767 Instruction *v = cast<Instruction>(Pair.first);
768 PhiState state = Pair.second;
770 assert(!isKnownBaseResult(v) && "why did it get added?");
771 assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
772 if (!state.isConflict())
775 if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
776 PHINode *phi = cast<PHINode>(v);
777 unsigned NumPHIValues = phi->getNumIncomingValues();
778 for (unsigned i = 0; i < NumPHIValues; i++) {
779 Value *InVal = phi->getIncomingValue(i);
780 BasicBlock *InBB = phi->getIncomingBlock(i);
782 // If we've already seen InBB, add the same incoming value
783 // we added for it earlier. The IR verifier requires phi
784 // nodes with multiple entries from the same basic block
785 // to have the same incoming value for each of those
786 // entries. If we don't do this check here and basephi
787 // has a different type than base, we'll end up adding two
788 // bitcasts (and hence two distinct values) as incoming
789 // values for the same basic block.
791 int blockIndex = basephi->getBasicBlockIndex(InBB);
792 if (blockIndex != -1) {
793 Value *oldBase = basephi->getIncomingValue(blockIndex);
794 basephi->addIncoming(oldBase, InBB);
796 Value *base = findBaseOrBDV(InVal, cache);
797 if (!isKnownBaseResult(base)) {
798 // Either conflict or base.
799 assert(states.count(base));
800 base = states[base].getBase();
801 assert(base != nullptr && "unknown PhiState!");
802 assert(NewInsertedDefs.count(base) &&
803 "should have already added this in a prev. iteration!");
806 // In essense this assert states: the only way two
807 // values incoming from the same basic block may be
808 // different is by being different bitcasts of the same
809 // value. A cleanup that remains TODO is changing
810 // findBaseOrBDV to return an llvm::Value of the correct
811 // type (and still remain pure). This will remove the
812 // need to add bitcasts.
813 assert(base->stripPointerCasts() == oldBase->stripPointerCasts() &&
814 "sanity -- findBaseOrBDV should be pure!");
819 // Find either the defining value for the PHI or the normal base for
821 Value *base = findBaseOrBDV(InVal, cache);
822 if (!isKnownBaseResult(base)) {
823 // Either conflict or base.
824 assert(states.count(base));
825 base = states[base].getBase();
826 assert(base != nullptr && "unknown PhiState!");
828 assert(base && "can't be null");
829 // Must use original input BB since base may not be Instruction
830 // The cast is needed since base traversal may strip away bitcasts
831 if (base->getType() != basephi->getType()) {
832 base = new BitCastInst(base, basephi->getType(), "cast",
833 InBB->getTerminator());
834 NewInsertedDefs.insert(base);
836 basephi->addIncoming(base, InBB);
838 assert(basephi->getNumIncomingValues() == NumPHIValues);
840 SelectInst *basesel = cast<SelectInst>(state.getBase());
841 SelectInst *sel = cast<SelectInst>(v);
842 // Operand 1 & 2 are true, false path respectively. TODO: refactor to
843 // something more safe and less hacky.
844 for (int i = 1; i <= 2; i++) {
845 Value *InVal = sel->getOperand(i);
846 // Find either the defining value for the PHI or the normal base for
848 Value *base = findBaseOrBDV(InVal, cache);
849 if (!isKnownBaseResult(base)) {
850 // Either conflict or base.
851 assert(states.count(base));
852 base = states[base].getBase();
853 assert(base != nullptr && "unknown PhiState!");
855 assert(base && "can't be null");
856 // Must use original input BB since base may not be Instruction
857 // The cast is needed since base traversal may strip away bitcasts
858 if (base->getType() != basesel->getType()) {
859 base = new BitCastInst(base, basesel->getType(), "cast", basesel);
860 NewInsertedDefs.insert(base);
862 basesel->setOperand(i, base);
867 // Cache all of our results so we can cheaply reuse them
868 // NOTE: This is actually two caches: one of the base defining value
869 // relation and one of the base pointer relation! FIXME
870 for (auto item : states) {
871 Value *v = item.first;
872 Value *base = item.second.getBase();
874 assert(!isKnownBaseResult(v) && "why did it get added?");
877 std::string fromstr =
878 cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
880 errs() << "Updating base value cache"
881 << " for: " << (v->hasName() ? v->getName() : "")
882 << " from: " << fromstr
883 << " to: " << (base->hasName() ? base->getName() : "") << "\n";
886 assert(isKnownBaseResult(base) &&
887 "must be something we 'know' is a base pointer");
888 if (cache.count(v)) {
889 // Once we transition from the BDV relation being store in the cache to
890 // the base relation being stored, it must be stable
891 assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
892 "base relation should be stable");
896 assert(cache.find(def) != cache.end());
900 // For a set of live pointers (base and/or derived), identify the base
901 // pointer of the object which they are derived from. This routine will
902 // mutate the IR graph as needed to make the 'base' pointer live at the
903 // definition site of 'derived'. This ensures that any use of 'derived' can
904 // also use 'base'. This may involve the insertion of a number of
905 // additional PHI nodes.
907 // preconditions: live is a set of pointer type Values
909 // side effects: may insert PHI nodes into the existing CFG, will preserve
910 // CFG, will not remove or mutate any existing nodes
912 // post condition: PointerToBase contains one (derived, base) pair for every
913 // pointer in live. Note that derived can be equal to base if the original
914 // pointer was a base pointer.
915 static void findBasePointers(const StatepointLiveSetTy &live,
916 DenseMap<llvm::Value *, llvm::Value *> &PointerToBase,
917 DominatorTree *DT, DefiningValueMapTy &DVCache,
918 DenseSet<llvm::Value *> &NewInsertedDefs) {
919 // For the naming of values inserted to be deterministic - which makes for
920 // much cleaner and more stable tests - we need to assign an order to the
921 // live values. DenseSets do not provide a deterministic order across runs.
922 SmallVector<Value*, 64> Temp;
923 Temp.insert(Temp.end(), live.begin(), live.end());
924 std::sort(Temp.begin(), Temp.end(), order_by_name);
925 for (Value *ptr : Temp) {
926 Value *base = findBasePointer(ptr, DVCache, NewInsertedDefs);
927 assert(base && "failed to find base pointer");
928 PointerToBase[ptr] = base;
929 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
930 DT->dominates(cast<Instruction>(base)->getParent(),
931 cast<Instruction>(ptr)->getParent())) &&
932 "The base we found better dominate the derived pointer");
934 // If you see this trip and like to live really dangerously, the code should
935 // be correct, just with idioms the verifier can't handle. You can try
936 // disabling the verifier at your own substaintial risk.
937 assert(!isa<ConstantPointerNull>(base) &&
938 "the relocation code needs adjustment to handle the relocation of "
939 "a null pointer constant without causing false positives in the "
940 "safepoint ir verifier.");
944 /// Find the required based pointers (and adjust the live set) for the given
946 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
948 PartiallyConstructedSafepointRecord &result) {
949 DenseMap<llvm::Value *, llvm::Value *> PointerToBase;
950 DenseSet<llvm::Value *> NewInsertedDefs;
951 findBasePointers(result.liveset, PointerToBase, &DT, DVCache, NewInsertedDefs);
953 if (PrintBasePointers) {
954 // Note: Need to print these in a stable order since this is checked in
956 errs() << "Base Pairs (w/o Relocation):\n";
957 SmallVector<Value*, 64> Temp;
958 Temp.reserve(PointerToBase.size());
959 for (auto Pair : PointerToBase) {
960 Temp.push_back(Pair.first);
962 std::sort(Temp.begin(), Temp.end(), order_by_name);
963 for (Value *Ptr : Temp) {
964 Value *Base = PointerToBase[Ptr];
965 errs() << " derived %" << Ptr->getName() << " base %"
966 << Base->getName() << "\n";
970 result.PointerToBase = PointerToBase;
971 result.NewInsertedDefs = NewInsertedDefs;
974 /// Check for liveness of items in the insert defs and add them to the live
975 /// and base pointer sets
976 static void fixupLiveness(DominatorTree &DT, const CallSite &CS,
977 const DenseSet<Value *> &allInsertedDefs,
978 PartiallyConstructedSafepointRecord &result) {
979 Instruction *inst = CS.getInstruction();
981 auto liveset = result.liveset;
982 auto PointerToBase = result.PointerToBase;
984 auto is_live_gc_reference =
985 [&](Value &V) { return isLiveGCReferenceAt(V, inst, DT, nullptr); };
987 // For each new definition, check to see if a) the definition dominates the
988 // instruction we're interested in, and b) one of the uses of that definition
989 // is edge-reachable from the instruction we're interested in. This is the
990 // same definition of liveness we used in the intial liveness analysis
991 for (Value *newDef : allInsertedDefs) {
992 if (liveset.count(newDef)) {
993 // already live, no action needed
997 // PERF: Use DT to check instruction domination might not be good for
998 // compilation time, and we could change to optimal solution if this
999 // turn to be a issue
1000 if (!DT.dominates(cast<Instruction>(newDef), inst)) {
1001 // can't possibly be live at inst
1005 if (is_live_gc_reference(*newDef)) {
1006 // Add the live new defs into liveset and PointerToBase
1007 liveset.insert(newDef);
1008 PointerToBase[newDef] = newDef;
1012 result.liveset = liveset;
1013 result.PointerToBase = PointerToBase;
1016 static void fixupLiveReferences(
1017 Function &F, DominatorTree &DT, Pass *P,
1018 const DenseSet<llvm::Value *> &allInsertedDefs,
1019 ArrayRef<CallSite> toUpdate,
1020 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1021 for (size_t i = 0; i < records.size(); i++) {
1022 struct PartiallyConstructedSafepointRecord &info = records[i];
1023 const CallSite &CS = toUpdate[i];
1024 fixupLiveness(DT, CS, allInsertedDefs, info);
1028 // Normalize basic block to make it ready to be target of invoke statepoint.
1029 // It means spliting it to have single predecessor. Return newly created BB
1030 // ready to be successor of invoke statepoint.
1031 static BasicBlock *normalizeBBForInvokeSafepoint(BasicBlock *BB,
1032 BasicBlock *InvokeParent,
1034 BasicBlock *ret = BB;
1036 if (!BB->getUniquePredecessor()) {
1037 ret = SplitBlockPredecessors(BB, InvokeParent, "");
1040 // Another requirement for such basic blocks is to not have any phi nodes.
1041 // Since we just ensured that new BB will have single predecessor,
1042 // all phi nodes in it will have one value. Here it would be naturall place
1044 // remove them all. But we can not do this because we are risking to remove
1045 // one of the values stored in liveset of another statepoint. We will do it
1046 // later after placing all safepoints.
1051 static int find_index(ArrayRef<Value *> livevec, Value *val) {
1052 auto itr = std::find(livevec.begin(), livevec.end(), val);
1053 assert(livevec.end() != itr);
1054 size_t index = std::distance(livevec.begin(), itr);
1055 assert(index < livevec.size());
1059 // Create new attribute set containing only attributes which can be transfered
1060 // from original call to the safepoint.
1061 static AttributeSet legalizeCallAttributes(AttributeSet AS) {
1064 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) {
1065 unsigned index = AS.getSlotIndex(Slot);
1067 if (index == AttributeSet::ReturnIndex ||
1068 index == AttributeSet::FunctionIndex) {
1070 for (auto it = AS.begin(Slot), it_end = AS.end(Slot); it != it_end;
1072 Attribute attr = *it;
1074 // Do not allow certain attributes - just skip them
1075 // Safepoint can not be read only or read none.
1076 if (attr.hasAttribute(Attribute::ReadNone) ||
1077 attr.hasAttribute(Attribute::ReadOnly))
1080 ret = ret.addAttributes(
1081 AS.getContext(), index,
1082 AttributeSet::get(AS.getContext(), index, AttrBuilder(attr)));
1086 // Just skip parameter attributes for now
1092 /// Helper function to place all gc relocates necessary for the given
1095 /// liveVariables - list of variables to be relocated.
1096 /// liveStart - index of the first live variable.
1097 /// basePtrs - base pointers.
1098 /// statepointToken - statepoint instruction to which relocates should be
1100 /// Builder - Llvm IR builder to be used to construct new calls.
1101 static void CreateGCRelocates(ArrayRef<llvm::Value *> liveVariables,
1102 const int liveStart,
1103 ArrayRef<llvm::Value *> basePtrs,
1104 Instruction *statepointToken,
1105 IRBuilder<> Builder) {
1106 SmallVector<Instruction *, 64> NewDefs;
1107 NewDefs.reserve(liveVariables.size());
1109 Module *M = statepointToken->getParent()->getParent()->getParent();
1111 for (unsigned i = 0; i < liveVariables.size(); i++) {
1112 // We generate a (potentially) unique declaration for every pointer type
1113 // combination. This results is some blow up the function declarations in
1114 // the IR, but removes the need for argument bitcasts which shrinks the IR
1115 // greatly and makes it much more readable.
1116 SmallVector<Type *, 1> types; // one per 'any' type
1117 types.push_back(liveVariables[i]->getType()); // result type
1118 Value *gc_relocate_decl = Intrinsic::getDeclaration(
1119 M, Intrinsic::experimental_gc_relocate, types);
1121 // Generate the gc.relocate call and save the result
1123 ConstantInt::get(Type::getInt32Ty(M->getContext()),
1124 liveStart + find_index(liveVariables, basePtrs[i]));
1125 Value *liveIdx = ConstantInt::get(
1126 Type::getInt32Ty(M->getContext()),
1127 liveStart + find_index(liveVariables, liveVariables[i]));
1129 // only specify a debug name if we can give a useful one
1130 Value *reloc = Builder.CreateCall3(
1131 gc_relocate_decl, statepointToken, baseIdx, liveIdx,
1132 liveVariables[i]->hasName() ? liveVariables[i]->getName() + ".relocated"
1134 // Trick CodeGen into thinking there are lots of free registers at this
1136 cast<CallInst>(reloc)->setCallingConv(CallingConv::Cold);
1138 NewDefs.push_back(cast<Instruction>(reloc));
1140 assert(NewDefs.size() == liveVariables.size() &&
1141 "missing or extra redefinition at safepoint");
1145 makeStatepointExplicitImpl(const CallSite &CS, /* to replace */
1146 const SmallVectorImpl<llvm::Value *> &basePtrs,
1147 const SmallVectorImpl<llvm::Value *> &liveVariables,
1149 PartiallyConstructedSafepointRecord &result) {
1150 assert(basePtrs.size() == liveVariables.size());
1151 assert(isStatepoint(CS) &&
1152 "This method expects to be rewriting a statepoint");
1154 BasicBlock *BB = CS.getInstruction()->getParent();
1156 Function *F = BB->getParent();
1157 assert(F && "must be set");
1158 Module *M = F->getParent();
1160 assert(M && "must be set");
1162 // We're not changing the function signature of the statepoint since the gc
1163 // arguments go into the var args section.
1164 Function *gc_statepoint_decl = CS.getCalledFunction();
1166 // Then go ahead and use the builder do actually do the inserts. We insert
1167 // immediately before the previous instruction under the assumption that all
1168 // arguments will be available here. We can't insert afterwards since we may
1169 // be replacing a terminator.
1170 Instruction *insertBefore = CS.getInstruction();
1171 IRBuilder<> Builder(insertBefore);
1172 // Copy all of the arguments from the original statepoint - this includes the
1173 // target, call args, and deopt args
1174 SmallVector<llvm::Value *, 64> args;
1175 args.insert(args.end(), CS.arg_begin(), CS.arg_end());
1176 // TODO: Clear the 'needs rewrite' flag
1178 // add all the pointers to be relocated (gc arguments)
1179 // Capture the start of the live variable list for use in the gc_relocates
1180 const int live_start = args.size();
1181 args.insert(args.end(), liveVariables.begin(), liveVariables.end());
1183 // Create the statepoint given all the arguments
1184 Instruction *token = nullptr;
1185 AttributeSet return_attributes;
1187 CallInst *toReplace = cast<CallInst>(CS.getInstruction());
1189 Builder.CreateCall(gc_statepoint_decl, args, "safepoint_token");
1190 call->setTailCall(toReplace->isTailCall());
1191 call->setCallingConv(toReplace->getCallingConv());
1193 // Currently we will fail on parameter attributes and on certain
1194 // function attributes.
1195 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1196 // In case if we can handle this set of sttributes - set up function attrs
1197 // directly on statepoint and return attrs later for gc_result intrinsic.
1198 call->setAttributes(new_attrs.getFnAttributes());
1199 return_attributes = new_attrs.getRetAttributes();
1203 // Put the following gc_result and gc_relocate calls immediately after the
1204 // the old call (which we're about to delete)
1205 BasicBlock::iterator next(toReplace);
1206 assert(BB->end() != next && "not a terminator, must have next");
1208 Instruction *IP = &*(next);
1209 Builder.SetInsertPoint(IP);
1210 Builder.SetCurrentDebugLocation(IP->getDebugLoc());
1213 InvokeInst *toReplace = cast<InvokeInst>(CS.getInstruction());
1215 // Insert the new invoke into the old block. We'll remove the old one in a
1216 // moment at which point this will become the new terminator for the
1218 InvokeInst *invoke = InvokeInst::Create(
1219 gc_statepoint_decl, toReplace->getNormalDest(),
1220 toReplace->getUnwindDest(), args, "", toReplace->getParent());
1221 invoke->setCallingConv(toReplace->getCallingConv());
1223 // Currently we will fail on parameter attributes and on certain
1224 // function attributes.
1225 AttributeSet new_attrs = legalizeCallAttributes(toReplace->getAttributes());
1226 // In case if we can handle this set of sttributes - set up function attrs
1227 // directly on statepoint and return attrs later for gc_result intrinsic.
1228 invoke->setAttributes(new_attrs.getFnAttributes());
1229 return_attributes = new_attrs.getRetAttributes();
1233 // Generate gc relocates in exceptional path
1234 BasicBlock *unwindBlock = normalizeBBForInvokeSafepoint(
1235 toReplace->getUnwindDest(), invoke->getParent(), P);
1237 Instruction *IP = &*(unwindBlock->getFirstInsertionPt());
1238 Builder.SetInsertPoint(IP);
1239 Builder.SetCurrentDebugLocation(toReplace->getDebugLoc());
1241 // Extract second element from landingpad return value. We will attach
1242 // exceptional gc relocates to it.
1243 const unsigned idx = 1;
1244 Instruction *exceptional_token =
1245 cast<Instruction>(Builder.CreateExtractValue(
1246 unwindBlock->getLandingPadInst(), idx, "relocate_token"));
1247 result.UnwindToken = exceptional_token;
1249 // Just throw away return value. We will use the one we got for normal
1251 (void)CreateGCRelocates(liveVariables, live_start, basePtrs,
1252 exceptional_token, Builder);
1254 // Generate gc relocates and returns for normal block
1255 BasicBlock *normalDest = normalizeBBForInvokeSafepoint(
1256 toReplace->getNormalDest(), invoke->getParent(), P);
1258 IP = &*(normalDest->getFirstInsertionPt());
1259 Builder.SetInsertPoint(IP);
1261 // gc relocates will be generated later as if it were regular call
1266 // Take the name of the original value call if it had one.
1267 token->takeName(CS.getInstruction());
1269 // The GCResult is already inserted, we just need to find it
1271 Instruction *toReplace = CS.getInstruction();
1272 assert((toReplace->hasNUses(0) || toReplace->hasNUses(1)) &&
1273 "only valid use before rewrite is gc.result");
1274 assert(!toReplace->hasOneUse() ||
1275 isGCResult(cast<Instruction>(*toReplace->user_begin())));
1278 // Update the gc.result of the original statepoint (if any) to use the newly
1279 // inserted statepoint. This is safe to do here since the token can't be
1280 // considered a live reference.
1281 CS.getInstruction()->replaceAllUsesWith(token);
1283 result.StatepointToken = token;
1285 // Second, create a gc.relocate for every live variable
1286 CreateGCRelocates(liveVariables, live_start, basePtrs, token, Builder);
1291 struct name_ordering {
1294 bool operator()(name_ordering const &a, name_ordering const &b) {
1295 return -1 == a.derived->getName().compare(b.derived->getName());
1299 static void stablize_order(SmallVectorImpl<Value *> &basevec,
1300 SmallVectorImpl<Value *> &livevec) {
1301 assert(basevec.size() == livevec.size());
1303 SmallVector<name_ordering, 64> temp;
1304 for (size_t i = 0; i < basevec.size(); i++) {
1306 v.base = basevec[i];
1307 v.derived = livevec[i];
1310 std::sort(temp.begin(), temp.end(), name_ordering());
1311 for (size_t i = 0; i < basevec.size(); i++) {
1312 basevec[i] = temp[i].base;
1313 livevec[i] = temp[i].derived;
1317 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1318 // which make the relocations happening at this safepoint explicit.
1320 // WARNING: Does not do any fixup to adjust users of the original live
1321 // values. That's the callers responsibility.
1323 makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, Pass *P,
1324 PartiallyConstructedSafepointRecord &result) {
1325 auto liveset = result.liveset;
1326 auto PointerToBase = result.PointerToBase;
1328 // Convert to vector for efficient cross referencing.
1329 SmallVector<Value *, 64> basevec, livevec;
1330 livevec.reserve(liveset.size());
1331 basevec.reserve(liveset.size());
1332 for (Value *L : liveset) {
1333 livevec.push_back(L);
1335 assert(PointerToBase.find(L) != PointerToBase.end());
1336 Value *base = PointerToBase[L];
1337 basevec.push_back(base);
1339 assert(livevec.size() == basevec.size());
1341 // To make the output IR slightly more stable (for use in diffs), ensure a
1342 // fixed order of the values in the safepoint (by sorting the value name).
1343 // The order is otherwise meaningless.
1344 stablize_order(basevec, livevec);
1346 // Do the actual rewriting and delete the old statepoint
1347 makeStatepointExplicitImpl(CS, basevec, livevec, P, result);
1348 CS.getInstruction()->eraseFromParent();
1351 // Helper function for the relocationViaAlloca.
1352 // It receives iterator to the statepoint gc relocates and emits store to the
1354 // location (via allocaMap) for the each one of them.
1355 // Add visited values into the visitedLiveValues set we will later use them
1356 // for sanity check.
1358 insertRelocationStores(iterator_range<Value::user_iterator> gcRelocs,
1359 DenseMap<Value *, Value *> &allocaMap,
1360 DenseSet<Value *> &visitedLiveValues) {
1362 for (User *U : gcRelocs) {
1363 if (!isa<IntrinsicInst>(U))
1366 IntrinsicInst *relocatedValue = cast<IntrinsicInst>(U);
1368 // We only care about relocates
1369 if (relocatedValue->getIntrinsicID() !=
1370 Intrinsic::experimental_gc_relocate) {
1374 GCRelocateOperands relocateOperands(relocatedValue);
1375 Value *originalValue = const_cast<Value *>(relocateOperands.derivedPtr());
1376 assert(allocaMap.count(originalValue));
1377 Value *alloca = allocaMap[originalValue];
1379 // Emit store into the related alloca
1380 StoreInst *store = new StoreInst(relocatedValue, alloca);
1381 store->insertAfter(relocatedValue);
1384 visitedLiveValues.insert(originalValue);
1389 /// do all the relocation update via allocas and mem2reg
1390 static void relocationViaAlloca(
1391 Function &F, DominatorTree &DT, ArrayRef<Value *> live,
1392 ArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1394 // record initial number of (static) allocas; we'll check we have the same
1395 // number when we get done.
1396 int InitialAllocaNum = 0;
1397 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end();
1399 if (isa<AllocaInst>(*I))
1403 // TODO-PERF: change data structures, reserve
1404 DenseMap<Value *, Value *> allocaMap;
1405 SmallVector<AllocaInst *, 200> PromotableAllocas;
1406 PromotableAllocas.reserve(live.size());
1408 // emit alloca for each live gc pointer
1409 for (unsigned i = 0; i < live.size(); i++) {
1410 Value *liveValue = live[i];
1411 AllocaInst *alloca = new AllocaInst(liveValue->getType(), "",
1412 F.getEntryBlock().getFirstNonPHI());
1413 allocaMap[liveValue] = alloca;
1414 PromotableAllocas.push_back(alloca);
1417 // The next two loops are part of the same conceptual operation. We need to
1418 // insert a store to the alloca after the original def and at each
1419 // redefinition. We need to insert a load before each use. These are split
1420 // into distinct loops for performance reasons.
1422 // update gc pointer after each statepoint
1423 // either store a relocated value or null (if no relocated value found for
1424 // this gc pointer and it is not a gc_result)
1425 // this must happen before we update the statepoint with load of alloca
1426 // otherwise we lose the link between statepoint and old def
1427 for (size_t i = 0; i < records.size(); i++) {
1428 const struct PartiallyConstructedSafepointRecord &info = records[i];
1429 Value *Statepoint = info.StatepointToken;
1431 // This will be used for consistency check
1432 DenseSet<Value *> visitedLiveValues;
1434 // Insert stores for normal statepoint gc relocates
1435 insertRelocationStores(Statepoint->users(), allocaMap, visitedLiveValues);
1437 // In case if it was invoke statepoint
1438 // we will insert stores for exceptional path gc relocates.
1439 if (isa<InvokeInst>(Statepoint)) {
1440 insertRelocationStores(info.UnwindToken->users(),
1441 allocaMap, visitedLiveValues);
1445 // As a debuging aid, pretend that an unrelocated pointer becomes null at
1446 // the gc.statepoint. This will turn some subtle GC problems into slightly
1447 // easier to debug SEGVs
1448 SmallVector<AllocaInst *, 64> ToClobber;
1449 for (auto Pair : allocaMap) {
1450 Value *Def = Pair.first;
1451 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1453 // This value was relocated
1454 if (visitedLiveValues.count(Def)) {
1457 ToClobber.push_back(Alloca);
1460 auto InsertClobbersAt = [&](Instruction *IP) {
1461 for (auto *AI : ToClobber) {
1462 auto AIType = cast<PointerType>(AI->getType());
1463 auto PT = cast<PointerType>(AIType->getElementType());
1464 Constant *CPN = ConstantPointerNull::get(PT);
1465 StoreInst *store = new StoreInst(CPN, AI);
1466 store->insertBefore(IP);
1470 // Insert the clobbering stores. These may get intermixed with the
1471 // gc.results and gc.relocates, but that's fine.
1472 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1473 InsertClobbersAt(II->getNormalDest()->getFirstInsertionPt());
1474 InsertClobbersAt(II->getUnwindDest()->getFirstInsertionPt());
1476 BasicBlock::iterator Next(cast<CallInst>(Statepoint));
1478 InsertClobbersAt(Next);
1482 // update use with load allocas and add store for gc_relocated
1483 for (auto Pair : allocaMap) {
1484 Value *def = Pair.first;
1485 Value *alloca = Pair.second;
1487 // we pre-record the uses of allocas so that we dont have to worry about
1489 // that change the user information.
1490 SmallVector<Instruction *, 20> uses;
1491 // PERF: trade a linear scan for repeated reallocation
1492 uses.reserve(std::distance(def->user_begin(), def->user_end()));
1493 for (User *U : def->users()) {
1494 if (!isa<ConstantExpr>(U)) {
1495 // If the def has a ConstantExpr use, then the def is either a
1496 // ConstantExpr use itself or null. In either case
1497 // (recursively in the first, directly in the second), the oop
1498 // it is ultimately dependent on is null and this particular
1499 // use does not need to be fixed up.
1500 uses.push_back(cast<Instruction>(U));
1504 std::sort(uses.begin(), uses.end());
1505 auto last = std::unique(uses.begin(), uses.end());
1506 uses.erase(last, uses.end());
1508 for (Instruction *use : uses) {
1509 if (isa<PHINode>(use)) {
1510 PHINode *phi = cast<PHINode>(use);
1511 for (unsigned i = 0; i < phi->getNumIncomingValues(); i++) {
1512 if (def == phi->getIncomingValue(i)) {
1513 LoadInst *load = new LoadInst(
1514 alloca, "", phi->getIncomingBlock(i)->getTerminator());
1515 phi->setIncomingValue(i, load);
1519 LoadInst *load = new LoadInst(alloca, "", use);
1520 use->replaceUsesOfWith(def, load);
1524 // emit store for the initial gc value
1525 // store must be inserted after load, otherwise store will be in alloca's
1526 // use list and an extra load will be inserted before it
1527 StoreInst *store = new StoreInst(def, alloca);
1528 if (Instruction *inst = dyn_cast<Instruction>(def)) {
1529 if (InvokeInst *invoke = dyn_cast<InvokeInst>(inst)) {
1530 // InvokeInst is a TerminatorInst so the store need to be inserted
1531 // into its normal destination block.
1532 BasicBlock *normalDest = invoke->getNormalDest();
1533 store->insertBefore(normalDest->getFirstNonPHI());
1535 assert(!inst->isTerminator() &&
1536 "The only TerminatorInst that can produce a value is "
1537 "InvokeInst which is handled above.");
1538 store->insertAfter(inst);
1541 assert((isa<Argument>(def) || isa<GlobalVariable>(def) ||
1542 isa<ConstantPointerNull>(def)) &&
1543 "Must be argument or global");
1544 store->insertAfter(cast<Instruction>(alloca));
1548 assert(PromotableAllocas.size() == live.size() &&
1549 "we must have the same allocas with lives");
1550 if (!PromotableAllocas.empty()) {
1551 // apply mem2reg to promote alloca to SSA
1552 PromoteMemToReg(PromotableAllocas, DT);
1556 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end();
1558 if (isa<AllocaInst>(*I))
1560 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1564 /// Implement a unique function which doesn't require we sort the input
1565 /// vector. Doing so has the effect of changing the output of a couple of
1566 /// tests in ways which make them less useful in testing fused safepoints.
1567 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1569 SmallVector<T, 128> TempVec;
1570 TempVec.reserve(Vec.size());
1571 for (auto Element : Vec)
1572 TempVec.push_back(Element);
1574 for (auto V : TempVec) {
1575 if (Seen.insert(V).second) {
1581 static Function *getUseHolder(Module &M) {
1582 FunctionType *ftype =
1583 FunctionType::get(Type::getVoidTy(M.getContext()), true);
1584 Function *Func = cast<Function>(M.getOrInsertFunction("__tmp_use", ftype));
1588 /// Insert holders so that each Value is obviously live through the entire
1589 /// liftetime of the call.
1590 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1591 SmallVectorImpl<CallInst *> &holders) {
1592 Module *M = CS.getInstruction()->getParent()->getParent()->getParent();
1593 Function *Func = getUseHolder(*M);
1595 // For call safepoints insert dummy calls right after safepoint
1596 BasicBlock::iterator next(CS.getInstruction());
1598 CallInst *base_holder = CallInst::Create(Func, Values, "", next);
1599 holders.push_back(base_holder);
1600 } else if (CS.isInvoke()) {
1601 // For invoke safepooints insert dummy calls both in normal and
1602 // exceptional destination blocks
1603 InvokeInst *invoke = cast<InvokeInst>(CS.getInstruction());
1604 CallInst *normal_holder = CallInst::Create(
1605 Func, Values, "", invoke->getNormalDest()->getFirstInsertionPt());
1606 CallInst *unwind_holder = CallInst::Create(
1607 Func, Values, "", invoke->getUnwindDest()->getFirstInsertionPt());
1608 holders.push_back(normal_holder);
1609 holders.push_back(unwind_holder);
1611 llvm_unreachable("unsupported call type");
1614 static void findLiveReferences(
1615 Function &F, DominatorTree &DT, Pass *P, ArrayRef<CallSite> toUpdate,
1616 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1617 for (size_t i = 0; i < records.size(); i++) {
1618 struct PartiallyConstructedSafepointRecord &info = records[i];
1619 const CallSite &CS = toUpdate[i];
1620 analyzeParsePointLiveness(DT, CS, info);
1624 static void addBasesAsLiveValues(StatepointLiveSetTy &liveset,
1625 DenseMap<Value *, Value *> &PointerToBase) {
1626 // Identify any base pointers which are used in this safepoint, but not
1627 // themselves relocated. We need to relocate them so that later inserted
1628 // safepoints can get the properly relocated base register.
1629 DenseSet<Value *> missing;
1630 for (Value *L : liveset) {
1631 assert(PointerToBase.find(L) != PointerToBase.end());
1632 Value *base = PointerToBase[L];
1634 if (liveset.find(base) == liveset.end()) {
1635 assert(PointerToBase.find(base) == PointerToBase.end());
1636 // uniqued by set insert
1637 missing.insert(base);
1641 // Note that we want these at the end of the list, otherwise
1642 // register placement gets screwed up once we lower to STATEPOINT
1643 // instructions. This is an utter hack, but there doesn't seem to be a
1645 for (Value *base : missing) {
1647 liveset.insert(base);
1648 PointerToBase[base] = base;
1650 assert(liveset.size() == PointerToBase.size());
1653 static bool insertParsePoints(Function &F, DominatorTree &DT, Pass *P,
1654 SmallVectorImpl<CallSite> &toUpdate) {
1656 // sanity check the input
1657 std::set<CallSite> uniqued;
1658 uniqued.insert(toUpdate.begin(), toUpdate.end());
1659 assert(uniqued.size() == toUpdate.size() && "no duplicates please!");
1661 for (size_t i = 0; i < toUpdate.size(); i++) {
1662 CallSite &CS = toUpdate[i];
1663 assert(CS.getInstruction()->getParent()->getParent() == &F);
1664 assert(isStatepoint(CS) && "expected to already be a deopt statepoint");
1668 // A list of dummy calls added to the IR to keep various values obviously
1669 // live in the IR. We'll remove all of these when done.
1670 SmallVector<CallInst *, 64> holders;
1672 // Insert a dummy call with all of the arguments to the vm_state we'll need
1673 // for the actual safepoint insertion. This ensures reference arguments in
1674 // the deopt argument list are considered live through the safepoint (and
1675 // thus makes sure they get relocated.)
1676 for (size_t i = 0; i < toUpdate.size(); i++) {
1677 CallSite &CS = toUpdate[i];
1678 Statepoint StatepointCS(CS);
1680 SmallVector<Value *, 64> DeoptValues;
1681 for (Use &U : StatepointCS.vm_state_args()) {
1682 Value *Arg = cast<Value>(&U);
1683 if (isGCPointerType(Arg->getType()))
1684 DeoptValues.push_back(Arg);
1686 insertUseHolderAfter(CS, DeoptValues, holders);
1689 SmallVector<struct PartiallyConstructedSafepointRecord, 64> records;
1690 records.reserve(toUpdate.size());
1691 for (size_t i = 0; i < toUpdate.size(); i++) {
1692 struct PartiallyConstructedSafepointRecord info;
1693 records.push_back(info);
1695 assert(records.size() == toUpdate.size());
1697 // A) Identify all gc pointers which are staticly live at the given call
1699 findLiveReferences(F, DT, P, toUpdate, records);
1701 // B) Find the base pointers for each live pointer
1702 /* scope for caching */ {
1703 // Cache the 'defining value' relation used in the computation and
1704 // insertion of base phis and selects. This ensures that we don't insert
1705 // large numbers of duplicate base_phis.
1706 DefiningValueMapTy DVCache;
1708 for (size_t i = 0; i < records.size(); i++) {
1709 struct PartiallyConstructedSafepointRecord &info = records[i];
1710 CallSite &CS = toUpdate[i];
1711 findBasePointers(DT, DVCache, CS, info);
1713 } // end of cache scope
1715 // The base phi insertion logic (for any safepoint) may have inserted new
1716 // instructions which are now live at some safepoint. The simplest such
1719 // phi a <-- will be a new base_phi here
1720 // safepoint 1 <-- that needs to be live here
1724 DenseSet<llvm::Value *> allInsertedDefs;
1725 for (size_t i = 0; i < records.size(); i++) {
1726 struct PartiallyConstructedSafepointRecord &info = records[i];
1727 allInsertedDefs.insert(info.NewInsertedDefs.begin(),
1728 info.NewInsertedDefs.end());
1731 // We insert some dummy calls after each safepoint to definitely hold live
1732 // the base pointers which were identified for that safepoint. We'll then
1733 // ask liveness for _every_ base inserted to see what is now live. Then we
1734 // remove the dummy calls.
1735 holders.reserve(holders.size() + records.size());
1736 for (size_t i = 0; i < records.size(); i++) {
1737 struct PartiallyConstructedSafepointRecord &info = records[i];
1738 CallSite &CS = toUpdate[i];
1740 SmallVector<Value *, 128> Bases;
1741 for (auto Pair : info.PointerToBase) {
1742 Bases.push_back(Pair.second);
1744 insertUseHolderAfter(CS, Bases, holders);
1747 // Add the bases explicitly to the live vector set. This may result in a few
1748 // extra relocations, but the base has to be available whenever a pointer
1749 // derived from it is used. Thus, we need it to be part of the statepoint's
1750 // gc arguments list. TODO: Introduce an explicit notion (in the following
1751 // code) of the GC argument list as seperate from the live Values at a
1752 // given statepoint.
1753 for (size_t i = 0; i < records.size(); i++) {
1754 struct PartiallyConstructedSafepointRecord &info = records[i];
1755 addBasesAsLiveValues(info.liveset, info.PointerToBase);
1758 // If we inserted any new values, we need to adjust our notion of what is
1759 // live at a particular safepoint.
1760 if (!allInsertedDefs.empty()) {
1761 fixupLiveReferences(F, DT, P, allInsertedDefs, toUpdate, records);
1763 if (PrintBasePointers) {
1764 for (size_t i = 0; i < records.size(); i++) {
1765 struct PartiallyConstructedSafepointRecord &info = records[i];
1766 errs() << "Base Pairs: (w/Relocation)\n";
1767 for (auto Pair : info.PointerToBase) {
1768 errs() << " derived %" << Pair.first->getName() << " base %"
1769 << Pair.second->getName() << "\n";
1773 for (size_t i = 0; i < holders.size(); i++) {
1774 holders[i]->eraseFromParent();
1775 holders[i] = nullptr;
1779 // Now run through and replace the existing statepoints with new ones with
1780 // the live variables listed. We do not yet update uses of the values being
1781 // relocated. We have references to live variables that need to
1782 // survive to the last iteration of this loop. (By construction, the
1783 // previous statepoint can not be a live variable, thus we can and remove
1784 // the old statepoint calls as we go.)
1785 for (size_t i = 0; i < records.size(); i++) {
1786 struct PartiallyConstructedSafepointRecord &info = records[i];
1787 CallSite &CS = toUpdate[i];
1788 makeStatepointExplicit(DT, CS, P, info);
1790 toUpdate.clear(); // prevent accident use of invalid CallSites
1792 // In case if we inserted relocates in a different basic block than the
1793 // original safepoint (this can happen for invokes). We need to be sure that
1794 // original values were not used in any of the phi nodes at the
1795 // beginning of basic block containing them. Because we know that all such
1796 // blocks will have single predecessor we can safely assume that all phi
1797 // nodes have single entry (because of normalizeBBForInvokeSafepoint).
1798 // Just remove them all here.
1799 for (size_t i = 0; i < records.size(); i++) {
1800 Instruction *I = records[i].StatepointToken;
1802 if (InvokeInst *invoke = dyn_cast<InvokeInst>(I)) {
1803 FoldSingleEntryPHINodes(invoke->getNormalDest());
1804 assert(!isa<PHINode>(invoke->getNormalDest()->begin()));
1806 FoldSingleEntryPHINodes(invoke->getUnwindDest());
1807 assert(!isa<PHINode>(invoke->getUnwindDest()->begin()));
1811 // Do all the fixups of the original live variables to their relocated selves
1812 SmallVector<Value *, 128> live;
1813 for (size_t i = 0; i < records.size(); i++) {
1814 struct PartiallyConstructedSafepointRecord &info = records[i];
1815 // We can't simply save the live set from the original insertion. One of
1816 // the live values might be the result of a call which needs a safepoint.
1817 // That Value* no longer exists and we need to use the new gc_result.
1818 // Thankfully, the liveset is embedded in the statepoint (and updated), so
1819 // we just grab that.
1820 Statepoint statepoint(info.StatepointToken);
1821 live.insert(live.end(), statepoint.gc_args_begin(),
1822 statepoint.gc_args_end());
1824 unique_unsorted(live);
1828 for (auto ptr : live) {
1829 assert(isGCPointerType(ptr->getType()) && "must be a gc pointer type");
1833 relocationViaAlloca(F, DT, live, records);
1834 return !records.empty();
1837 /// Returns true if this function should be rewritten by this pass. The main
1838 /// point of this function is as an extension point for custom logic.
1839 static bool shouldRewriteStatepointsIn(Function &F) {
1840 // TODO: This should check the GCStrategy
1842 const std::string StatepointExampleName("statepoint-example");
1843 return StatepointExampleName == F.getGC();
1848 bool RewriteStatepointsForGC::runOnFunction(Function &F) {
1849 // Nothing to do for declarations.
1850 if (F.isDeclaration() || F.empty())
1853 // Policy choice says not to rewrite - the most common reason is that we're
1854 // compiling code without a GCStrategy.
1855 if (!shouldRewriteStatepointsIn(F))
1858 // Gather all the statepoints which need rewritten.
1859 SmallVector<CallSite, 64> ParsePointNeeded;
1860 for (Instruction &I : inst_range(F)) {
1861 // TODO: only the ones with the flag set!
1862 if (isStatepoint(I))
1863 ParsePointNeeded.push_back(CallSite(&I));
1866 // Return early if no work to do.
1867 if (ParsePointNeeded.empty())
1870 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1871 return insertParsePoints(F, DT, this, ParsePointNeeded);