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 call/invoke instructions so as to make potential relocations
11 // performed by the garbage collector explicit in the IR.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Scalar/RewriteStatepointsForGC.h"
17 #include "llvm/ADT/ArrayRef.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/DenseSet.h"
20 #include "llvm/ADT/MapVector.h"
21 #include "llvm/ADT/None.h"
22 #include "llvm/ADT/Optional.h"
23 #include "llvm/ADT/STLExtras.h"
24 #include "llvm/ADT/SetVector.h"
25 #include "llvm/ADT/SmallSet.h"
26 #include "llvm/ADT/SmallVector.h"
27 #include "llvm/ADT/StringRef.h"
28 #include "llvm/ADT/iterator_range.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/Analysis/TargetTransformInfo.h"
31 #include "llvm/IR/Argument.h"
32 #include "llvm/IR/Attributes.h"
33 #include "llvm/IR/BasicBlock.h"
34 #include "llvm/IR/CallSite.h"
35 #include "llvm/IR/CallingConv.h"
36 #include "llvm/IR/Constant.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DataLayout.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/InstIterator.h"
44 #include "llvm/IR/InstrTypes.h"
45 #include "llvm/IR/Instruction.h"
46 #include "llvm/IR/Instructions.h"
47 #include "llvm/IR/IntrinsicInst.h"
48 #include "llvm/IR/Intrinsics.h"
49 #include "llvm/IR/LLVMContext.h"
50 #include "llvm/IR/MDBuilder.h"
51 #include "llvm/IR/Metadata.h"
52 #include "llvm/IR/Module.h"
53 #include "llvm/IR/Statepoint.h"
54 #include "llvm/IR/Type.h"
55 #include "llvm/IR/User.h"
56 #include "llvm/IR/Value.h"
57 #include "llvm/IR/ValueHandle.h"
58 #include "llvm/Pass.h"
59 #include "llvm/Support/Casting.h"
60 #include "llvm/Support/CommandLine.h"
61 #include "llvm/Support/Compiler.h"
62 #include "llvm/Support/Debug.h"
63 #include "llvm/Support/ErrorHandling.h"
64 #include "llvm/Support/raw_ostream.h"
65 #include "llvm/Transforms/Scalar.h"
66 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
67 #include "llvm/Transforms/Utils/Local.h"
68 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
79 #define DEBUG_TYPE "rewrite-statepoints-for-gc"
83 // Print the liveset found at the insert location
84 static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden,
86 static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden,
89 // Print out the base pointers for debugging
90 static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden,
93 // Cost threshold measuring when it is profitable to rematerialize value instead
95 static cl::opt<unsigned>
96 RematerializationThreshold("spp-rematerialization-threshold", cl::Hidden,
99 #ifdef EXPENSIVE_CHECKS
100 static bool ClobberNonLive = true;
102 static bool ClobberNonLive = false;
105 static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live",
106 cl::location(ClobberNonLive),
110 AllowStatepointWithNoDeoptInfo("rs4gc-allow-statepoint-with-no-deopt-info",
111 cl::Hidden, cl::init(true));
113 /// The IR fed into RewriteStatepointsForGC may have had attributes and
114 /// metadata implying dereferenceability that are no longer valid/correct after
115 /// RewriteStatepointsForGC has run. This is because semantically, after
116 /// RewriteStatepointsForGC runs, all calls to gc.statepoint "free" the entire
117 /// heap. stripNonValidData (conservatively) restores
118 /// correctness by erasing all attributes in the module that externally imply
119 /// dereferenceability. Similar reasoning also applies to the noalias
120 /// attributes and metadata. gc.statepoint can touch the entire heap including
122 /// Apart from attributes and metadata, we also remove instructions that imply
123 /// constant physical memory: llvm.invariant.start.
124 static void stripNonValidData(Module &M);
126 static bool shouldRewriteStatepointsIn(Function &F);
128 PreservedAnalyses RewriteStatepointsForGC::run(Module &M,
129 ModuleAnalysisManager &AM) {
130 bool Changed = false;
131 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
132 for (Function &F : M) {
133 // Nothing to do for declarations.
134 if (F.isDeclaration() || F.empty())
137 // Policy choice says not to rewrite - the most common reason is that we're
138 // compiling code without a GCStrategy.
139 if (!shouldRewriteStatepointsIn(F))
142 auto &DT = FAM.getResult<DominatorTreeAnalysis>(F);
143 auto &TTI = FAM.getResult<TargetIRAnalysis>(F);
144 auto &TLI = FAM.getResult<TargetLibraryAnalysis>(F);
145 Changed |= runOnFunction(F, DT, TTI, TLI);
148 return PreservedAnalyses::all();
150 // stripNonValidData asserts that shouldRewriteStatepointsIn
151 // returns true for at least one function in the module. Since at least
152 // one function changed, we know that the precondition is satisfied.
153 stripNonValidData(M);
155 PreservedAnalyses PA;
156 PA.preserve<TargetIRAnalysis>();
157 PA.preserve<TargetLibraryAnalysis>();
163 class RewriteStatepointsForGCLegacyPass : public ModulePass {
164 RewriteStatepointsForGC Impl;
167 static char ID; // Pass identification, replacement for typeid
169 RewriteStatepointsForGCLegacyPass() : ModulePass(ID), Impl() {
170 initializeRewriteStatepointsForGCLegacyPassPass(
171 *PassRegistry::getPassRegistry());
174 bool runOnModule(Module &M) override {
175 bool Changed = false;
176 const TargetLibraryInfo &TLI =
177 getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
178 for (Function &F : M) {
179 // Nothing to do for declarations.
180 if (F.isDeclaration() || F.empty())
183 // Policy choice says not to rewrite - the most common reason is that
184 // we're compiling code without a GCStrategy.
185 if (!shouldRewriteStatepointsIn(F))
188 TargetTransformInfo &TTI =
189 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
190 auto &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
192 Changed |= Impl.runOnFunction(F, DT, TTI, TLI);
198 // stripNonValidData asserts that shouldRewriteStatepointsIn
199 // returns true for at least one function in the module. Since at least
200 // one function changed, we know that the precondition is satisfied.
201 stripNonValidData(M);
205 void getAnalysisUsage(AnalysisUsage &AU) const override {
206 // We add and rewrite a bunch of instructions, but don't really do much
207 // else. We could in theory preserve a lot more analyses here.
208 AU.addRequired<DominatorTreeWrapperPass>();
209 AU.addRequired<TargetTransformInfoWrapperPass>();
210 AU.addRequired<TargetLibraryInfoWrapperPass>();
214 } // end anonymous namespace
216 char RewriteStatepointsForGCLegacyPass::ID = 0;
218 ModulePass *llvm::createRewriteStatepointsForGCLegacyPass() {
219 return new RewriteStatepointsForGCLegacyPass();
222 INITIALIZE_PASS_BEGIN(RewriteStatepointsForGCLegacyPass,
223 "rewrite-statepoints-for-gc",
224 "Make relocations explicit at statepoints", false, false)
225 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
226 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
227 INITIALIZE_PASS_END(RewriteStatepointsForGCLegacyPass,
228 "rewrite-statepoints-for-gc",
229 "Make relocations explicit at statepoints", false, false)
233 struct GCPtrLivenessData {
234 /// Values defined in this block.
235 MapVector<BasicBlock *, SetVector<Value *>> KillSet;
237 /// Values used in this block (and thus live); does not included values
238 /// killed within this block.
239 MapVector<BasicBlock *, SetVector<Value *>> LiveSet;
241 /// Values live into this basic block (i.e. used by any
242 /// instruction in this basic block or ones reachable from here)
243 MapVector<BasicBlock *, SetVector<Value *>> LiveIn;
245 /// Values live out of this basic block (i.e. live into
246 /// any successor block)
247 MapVector<BasicBlock *, SetVector<Value *>> LiveOut;
250 // The type of the internal cache used inside the findBasePointers family
251 // of functions. From the callers perspective, this is an opaque type and
252 // should not be inspected.
254 // In the actual implementation this caches two relations:
255 // - The base relation itself (i.e. this pointer is based on that one)
256 // - The base defining value relation (i.e. before base_phi insertion)
257 // Generally, after the execution of a full findBasePointer call, only the
258 // base relation will remain. Internally, we add a mixture of the two
259 // types, then update all the second type to the first type
260 using DefiningValueMapTy = MapVector<Value *, Value *>;
261 using StatepointLiveSetTy = SetVector<Value *>;
262 using RematerializedValueMapTy =
263 MapVector<AssertingVH<Instruction>, AssertingVH<Value>>;
265 struct PartiallyConstructedSafepointRecord {
266 /// The set of values known to be live across this safepoint
267 StatepointLiveSetTy LiveSet;
269 /// Mapping from live pointers to a base-defining-value
270 MapVector<Value *, Value *> PointerToBase;
272 /// The *new* gc.statepoint instruction itself. This produces the token
273 /// that normal path gc.relocates and the gc.result are tied to.
274 Instruction *StatepointToken;
276 /// Instruction to which exceptional gc relocates are attached
277 /// Makes it easier to iterate through them during relocationViaAlloca.
278 Instruction *UnwindToken;
280 /// Record live values we are rematerialized instead of relocating.
281 /// They are not included into 'LiveSet' field.
282 /// Maps rematerialized copy to it's original value.
283 RematerializedValueMapTy RematerializedValues;
286 } // end anonymous namespace
288 static ArrayRef<Use> GetDeoptBundleOperands(ImmutableCallSite CS) {
289 Optional<OperandBundleUse> DeoptBundle =
290 CS.getOperandBundle(LLVMContext::OB_deopt);
292 if (!DeoptBundle.hasValue()) {
293 assert(AllowStatepointWithNoDeoptInfo &&
294 "Found non-leaf call without deopt info!");
298 return DeoptBundle.getValue().Inputs;
301 /// Compute the live-in set for every basic block in the function
302 static void computeLiveInValues(DominatorTree &DT, Function &F,
303 GCPtrLivenessData &Data);
305 /// Given results from the dataflow liveness computation, find the set of live
306 /// Values at a particular instruction.
307 static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data,
308 StatepointLiveSetTy &out);
310 // TODO: Once we can get to the GCStrategy, this becomes
311 // Optional<bool> isGCManagedPointer(const Type *Ty) const override {
313 static bool isGCPointerType(Type *T) {
314 if (auto *PT = dyn_cast<PointerType>(T))
315 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our
316 // GC managed heap. We know that a pointer into this heap needs to be
317 // updated and that no other pointer does.
318 return PT->getAddressSpace() == 1;
322 // Return true if this type is one which a) is a gc pointer or contains a GC
323 // pointer and b) is of a type this code expects to encounter as a live value.
324 // (The insertion code will assert that a type which matches (a) and not (b)
325 // is not encountered.)
326 static bool isHandledGCPointerType(Type *T) {
327 // We fully support gc pointers
328 if (isGCPointerType(T))
330 // We partially support vectors of gc pointers. The code will assert if it
331 // can't handle something.
332 if (auto VT = dyn_cast<VectorType>(T))
333 if (isGCPointerType(VT->getElementType()))
339 /// Returns true if this type contains a gc pointer whether we know how to
340 /// handle that type or not.
341 static bool containsGCPtrType(Type *Ty) {
342 if (isGCPointerType(Ty))
344 if (VectorType *VT = dyn_cast<VectorType>(Ty))
345 return isGCPointerType(VT->getScalarType());
346 if (ArrayType *AT = dyn_cast<ArrayType>(Ty))
347 return containsGCPtrType(AT->getElementType());
348 if (StructType *ST = dyn_cast<StructType>(Ty))
349 return llvm::any_of(ST->subtypes(), containsGCPtrType);
353 // Returns true if this is a type which a) is a gc pointer or contains a GC
354 // pointer and b) is of a type which the code doesn't expect (i.e. first class
355 // aggregates). Used to trip assertions.
356 static bool isUnhandledGCPointerType(Type *Ty) {
357 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty);
361 // Return the name of the value suffixed with the provided value, or if the
362 // value didn't have a name, the default value specified.
363 static std::string suffixed_name_or(Value *V, StringRef Suffix,
364 StringRef DefaultName) {
365 return V->hasName() ? (V->getName() + Suffix).str() : DefaultName.str();
368 // Conservatively identifies any definitions which might be live at the
369 // given instruction. The analysis is performed immediately before the
370 // given instruction. Values defined by that instruction are not considered
371 // live. Values used by that instruction are considered live.
373 analyzeParsePointLiveness(DominatorTree &DT,
374 GCPtrLivenessData &OriginalLivenessData, CallSite CS,
375 PartiallyConstructedSafepointRecord &Result) {
376 Instruction *Inst = CS.getInstruction();
378 StatepointLiveSetTy LiveSet;
379 findLiveSetAtInst(Inst, OriginalLivenessData, LiveSet);
382 dbgs() << "Live Variables:\n";
383 for (Value *V : LiveSet)
384 dbgs() << " " << V->getName() << " " << *V << "\n";
386 if (PrintLiveSetSize) {
387 dbgs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n";
388 dbgs() << "Number live values: " << LiveSet.size() << "\n";
390 Result.LiveSet = LiveSet;
393 static bool isKnownBaseResult(Value *V);
397 /// A single base defining value - An immediate base defining value for an
398 /// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'.
399 /// For instructions which have multiple pointer [vector] inputs or that
400 /// transition between vector and scalar types, there is no immediate base
401 /// defining value. The 'base defining value' for 'Def' is the transitive
402 /// closure of this relation stopping at the first instruction which has no
403 /// immediate base defining value. The b.d.v. might itself be a base pointer,
404 /// but it can also be an arbitrary derived pointer.
405 struct BaseDefiningValueResult {
406 /// Contains the value which is the base defining value.
409 /// True if the base defining value is also known to be an actual base
411 const bool IsKnownBase;
413 BaseDefiningValueResult(Value *BDV, bool IsKnownBase)
414 : BDV(BDV), IsKnownBase(IsKnownBase) {
416 // Check consistency between new and old means of checking whether a BDV is
418 bool MustBeBase = isKnownBaseResult(BDV);
419 assert(!MustBeBase || MustBeBase == IsKnownBase);
424 } // end anonymous namespace
426 static BaseDefiningValueResult findBaseDefiningValue(Value *I);
428 /// Return a base defining value for the 'Index' element of the given vector
429 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
430 /// 'I'. As an optimization, this method will try to determine when the
431 /// element is known to already be a base pointer. If this can be established,
432 /// the second value in the returned pair will be true. Note that either a
433 /// vector or a pointer typed value can be returned. For the former, the
434 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
435 /// If the later, the return pointer is a BDV (or possibly a base) for the
436 /// particular element in 'I'.
437 static BaseDefiningValueResult
438 findBaseDefiningValueOfVector(Value *I) {
439 // Each case parallels findBaseDefiningValue below, see that code for
440 // detailed motivation.
442 if (isa<Argument>(I))
443 // An incoming argument to the function is a base pointer
444 return BaseDefiningValueResult(I, true);
446 if (isa<Constant>(I))
447 // Base of constant vector consists only of constant null pointers.
448 // For reasoning see similar case inside 'findBaseDefiningValue' function.
449 return BaseDefiningValueResult(ConstantAggregateZero::get(I->getType()),
452 if (isa<LoadInst>(I))
453 return BaseDefiningValueResult(I, true);
455 if (isa<InsertElementInst>(I))
456 // We don't know whether this vector contains entirely base pointers or
457 // not. To be conservatively correct, we treat it as a BDV and will
458 // duplicate code as needed to construct a parallel vector of bases.
459 return BaseDefiningValueResult(I, false);
461 if (isa<ShuffleVectorInst>(I))
462 // We don't know whether this vector contains entirely base pointers or
463 // not. To be conservatively correct, we treat it as a BDV and will
464 // duplicate code as needed to construct a parallel vector of bases.
465 // TODO: There a number of local optimizations which could be applied here
466 // for particular sufflevector patterns.
467 return BaseDefiningValueResult(I, false);
469 // The behavior of getelementptr instructions is the same for vector and
470 // non-vector data types.
471 if (auto *GEP = dyn_cast<GetElementPtrInst>(I))
472 return findBaseDefiningValue(GEP->getPointerOperand());
474 // If the pointer comes through a bitcast of a vector of pointers to
475 // a vector of another type of pointer, then look through the bitcast
476 if (auto *BC = dyn_cast<BitCastInst>(I))
477 return findBaseDefiningValue(BC->getOperand(0));
479 // A PHI or Select is a base defining value. The outer findBasePointer
480 // algorithm is responsible for constructing a base value for this BDV.
481 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
482 "unknown vector instruction - no base found for vector element");
483 return BaseDefiningValueResult(I, false);
486 /// Helper function for findBasePointer - Will return a value which either a)
487 /// defines the base pointer for the input, b) blocks the simple search
488 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
489 /// from pointer to vector type or back.
490 static BaseDefiningValueResult findBaseDefiningValue(Value *I) {
491 assert(I->getType()->isPtrOrPtrVectorTy() &&
492 "Illegal to ask for the base pointer of a non-pointer type");
494 if (I->getType()->isVectorTy())
495 return findBaseDefiningValueOfVector(I);
497 if (isa<Argument>(I))
498 // An incoming argument to the function is a base pointer
499 // We should have never reached here if this argument isn't an gc value
500 return BaseDefiningValueResult(I, true);
502 if (isa<Constant>(I)) {
503 // We assume that objects with a constant base (e.g. a global) can't move
504 // and don't need to be reported to the collector because they are always
505 // live. Besides global references, all kinds of constants (e.g. undef,
506 // constant expressions, null pointers) can be introduced by the inliner or
507 // the optimizer, especially on dynamically dead paths.
508 // Here we treat all of them as having single null base. By doing this we
509 // trying to avoid problems reporting various conflicts in a form of
510 // "phi (const1, const2)" or "phi (const, regular gc ptr)".
511 // See constant.ll file for relevant test cases.
513 return BaseDefiningValueResult(
514 ConstantPointerNull::get(cast<PointerType>(I->getType())), true);
517 if (CastInst *CI = dyn_cast<CastInst>(I)) {
518 Value *Def = CI->stripPointerCasts();
519 // If stripping pointer casts changes the address space there is an
520 // addrspacecast in between.
521 assert(cast<PointerType>(Def->getType())->getAddressSpace() ==
522 cast<PointerType>(CI->getType())->getAddressSpace() &&
523 "unsupported addrspacecast");
524 // If we find a cast instruction here, it means we've found a cast which is
525 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
526 // handle int->ptr conversion.
527 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
528 return findBaseDefiningValue(Def);
531 if (isa<LoadInst>(I))
532 // The value loaded is an gc base itself
533 return BaseDefiningValueResult(I, true);
535 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
536 // The base of this GEP is the base
537 return findBaseDefiningValue(GEP->getPointerOperand());
539 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
540 switch (II->getIntrinsicID()) {
542 // fall through to general call handling
544 case Intrinsic::experimental_gc_statepoint:
545 llvm_unreachable("statepoints don't produce pointers");
546 case Intrinsic::experimental_gc_relocate:
547 // Rerunning safepoint insertion after safepoints are already
548 // inserted is not supported. It could probably be made to work,
549 // but why are you doing this? There's no good reason.
550 llvm_unreachable("repeat safepoint insertion is not supported");
551 case Intrinsic::gcroot:
552 // Currently, this mechanism hasn't been extended to work with gcroot.
553 // There's no reason it couldn't be, but I haven't thought about the
554 // implications much.
556 "interaction with the gcroot mechanism is not supported");
559 // We assume that functions in the source language only return base
560 // pointers. This should probably be generalized via attributes to support
561 // both source language and internal functions.
562 if (isa<CallInst>(I) || isa<InvokeInst>(I))
563 return BaseDefiningValueResult(I, true);
565 // TODO: I have absolutely no idea how to implement this part yet. It's not
566 // necessarily hard, I just haven't really looked at it yet.
567 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
569 if (isa<AtomicCmpXchgInst>(I))
570 // A CAS is effectively a atomic store and load combined under a
571 // predicate. From the perspective of base pointers, we just treat it
573 return BaseDefiningValueResult(I, true);
575 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
576 "binary ops which don't apply to pointers");
578 // The aggregate ops. Aggregates can either be in the heap or on the
579 // stack, but in either case, this is simply a field load. As a result,
580 // this is a defining definition of the base just like a load is.
581 if (isa<ExtractValueInst>(I))
582 return BaseDefiningValueResult(I, true);
584 // We should never see an insert vector since that would require we be
585 // tracing back a struct value not a pointer value.
586 assert(!isa<InsertValueInst>(I) &&
587 "Base pointer for a struct is meaningless");
589 // An extractelement produces a base result exactly when it's input does.
590 // We may need to insert a parallel instruction to extract the appropriate
591 // element out of the base vector corresponding to the input. Given this,
592 // it's analogous to the phi and select case even though it's not a merge.
593 if (isa<ExtractElementInst>(I))
594 // Note: There a lot of obvious peephole cases here. This are deliberately
595 // handled after the main base pointer inference algorithm to make writing
596 // test cases to exercise that code easier.
597 return BaseDefiningValueResult(I, false);
599 // The last two cases here don't return a base pointer. Instead, they
600 // return a value which dynamically selects from among several base
601 // derived pointers (each with it's own base potentially). It's the job of
602 // the caller to resolve these.
603 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
604 "missing instruction case in findBaseDefiningValing");
605 return BaseDefiningValueResult(I, false);
608 /// Returns the base defining value for this value.
609 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
610 Value *&Cached = Cache[I];
612 Cached = findBaseDefiningValue(I).BDV;
613 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
614 << Cached->getName() << "\n");
616 assert(Cache[I] != nullptr);
620 /// Return a base pointer for this value if known. Otherwise, return it's
621 /// base defining value.
622 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
623 Value *Def = findBaseDefiningValueCached(I, Cache);
624 auto Found = Cache.find(Def);
625 if (Found != Cache.end()) {
626 // Either a base-of relation, or a self reference. Caller must check.
627 return Found->second;
629 // Only a BDV available
633 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
634 /// is it known to be a base pointer? Or do we need to continue searching.
635 static bool isKnownBaseResult(Value *V) {
636 if (!isa<PHINode>(V) && !isa<SelectInst>(V) &&
637 !isa<ExtractElementInst>(V) && !isa<InsertElementInst>(V) &&
638 !isa<ShuffleVectorInst>(V)) {
639 // no recursion possible
642 if (isa<Instruction>(V) &&
643 cast<Instruction>(V)->getMetadata("is_base_value")) {
644 // This is a previously inserted base phi or select. We know
645 // that this is a base value.
649 // We need to keep searching
655 /// Models the state of a single base defining value in the findBasePointer
656 /// algorithm for determining where a new instruction is needed to propagate
657 /// the base of this BDV.
660 enum Status { Unknown, Base, Conflict };
662 BDVState() : BaseValue(nullptr) {}
664 explicit BDVState(Status Status, Value *BaseValue = nullptr)
665 : Status(Status), BaseValue(BaseValue) {
666 assert(Status != Base || BaseValue);
669 explicit BDVState(Value *BaseValue) : Status(Base), BaseValue(BaseValue) {}
671 Status getStatus() const { return Status; }
672 Value *getBaseValue() const { return BaseValue; }
674 bool isBase() const { return getStatus() == Base; }
675 bool isUnknown() const { return getStatus() == Unknown; }
676 bool isConflict() const { return getStatus() == Conflict; }
678 bool operator==(const BDVState &Other) const {
679 return BaseValue == Other.BaseValue && Status == Other.Status;
682 bool operator!=(const BDVState &other) const { return !(*this == other); }
690 void print(raw_ostream &OS) const {
691 switch (getStatus()) {
702 OS << " (" << getBaseValue() << " - "
703 << (getBaseValue() ? getBaseValue()->getName() : "nullptr") << "): ";
707 Status Status = Unknown;
708 AssertingVH<Value> BaseValue; // Non-null only if Status == Base.
711 } // end anonymous namespace
714 static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
720 static BDVState meetBDVStateImpl(const BDVState &LHS, const BDVState &RHS) {
721 switch (LHS.getStatus()) {
722 case BDVState::Unknown:
726 assert(LHS.getBaseValue() && "can't be null");
731 if (LHS.getBaseValue() == RHS.getBaseValue()) {
732 assert(LHS == RHS && "equality broken!");
735 return BDVState(BDVState::Conflict);
737 assert(RHS.isConflict() && "only three states!");
738 return BDVState(BDVState::Conflict);
740 case BDVState::Conflict:
743 llvm_unreachable("only three states!");
746 // Values of type BDVState form a lattice, and this function implements the meet
748 static BDVState meetBDVState(const BDVState &LHS, const BDVState &RHS) {
749 BDVState Result = meetBDVStateImpl(LHS, RHS);
750 assert(Result == meetBDVStateImpl(RHS, LHS) &&
751 "Math is wrong: meet does not commute!");
755 /// For a given value or instruction, figure out what base ptr its derived from.
756 /// For gc objects, this is simply itself. On success, returns a value which is
757 /// the base pointer. (This is reliable and can be used for relocation.) On
758 /// failure, returns nullptr.
759 static Value *findBasePointer(Value *I, DefiningValueMapTy &Cache) {
760 Value *Def = findBaseOrBDV(I, Cache);
762 if (isKnownBaseResult(Def))
765 // Here's the rough algorithm:
766 // - For every SSA value, construct a mapping to either an actual base
767 // pointer or a PHI which obscures the base pointer.
768 // - Construct a mapping from PHI to unknown TOP state. Use an
769 // optimistic algorithm to propagate base pointer information. Lattice
774 // When algorithm terminates, all PHIs will either have a single concrete
775 // base or be in a conflict state.
776 // - For every conflict, insert a dummy PHI node without arguments. Add
777 // these to the base[Instruction] = BasePtr mapping. For every
778 // non-conflict, add the actual base.
779 // - For every conflict, add arguments for the base[a] of each input
782 // Note: A simpler form of this would be to add the conflict form of all
783 // PHIs without running the optimistic algorithm. This would be
784 // analogous to pessimistic data flow and would likely lead to an
785 // overall worse solution.
788 auto isExpectedBDVType = [](Value *BDV) {
789 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) ||
790 isa<ExtractElementInst>(BDV) || isa<InsertElementInst>(BDV) ||
791 isa<ShuffleVectorInst>(BDV);
795 // Once populated, will contain a mapping from each potentially non-base BDV
796 // to a lattice value (described above) which corresponds to that BDV.
797 // We use the order of insertion (DFS over the def/use graph) to provide a
798 // stable deterministic ordering for visiting DenseMaps (which are unordered)
799 // below. This is important for deterministic compilation.
800 MapVector<Value *, BDVState> States;
802 // Recursively fill in all base defining values reachable from the initial
803 // one for which we don't already know a definite base value for
805 SmallVector<Value*, 16> Worklist;
806 Worklist.push_back(Def);
807 States.insert({Def, BDVState()});
808 while (!Worklist.empty()) {
809 Value *Current = Worklist.pop_back_val();
810 assert(!isKnownBaseResult(Current) && "why did it get added?");
812 auto visitIncomingValue = [&](Value *InVal) {
813 Value *Base = findBaseOrBDV(InVal, Cache);
814 if (isKnownBaseResult(Base))
815 // Known bases won't need new instructions introduced and can be
818 assert(isExpectedBDVType(Base) && "the only non-base values "
819 "we see should be base defining values");
820 if (States.insert(std::make_pair(Base, BDVState())).second)
821 Worklist.push_back(Base);
823 if (PHINode *PN = dyn_cast<PHINode>(Current)) {
824 for (Value *InVal : PN->incoming_values())
825 visitIncomingValue(InVal);
826 } else if (SelectInst *SI = dyn_cast<SelectInst>(Current)) {
827 visitIncomingValue(SI->getTrueValue());
828 visitIncomingValue(SI->getFalseValue());
829 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
830 visitIncomingValue(EE->getVectorOperand());
831 } else if (auto *IE = dyn_cast<InsertElementInst>(Current)) {
832 visitIncomingValue(IE->getOperand(0)); // vector operand
833 visitIncomingValue(IE->getOperand(1)); // scalar operand
834 } else if (auto *SV = dyn_cast<ShuffleVectorInst>(Current)) {
835 visitIncomingValue(SV->getOperand(0));
836 visitIncomingValue(SV->getOperand(1));
839 llvm_unreachable("Unimplemented instruction case");
845 DEBUG(dbgs() << "States after initialization:\n");
846 for (auto Pair : States) {
847 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
851 // Return a phi state for a base defining value. We'll generate a new
852 // base state for known bases and expect to find a cached state otherwise.
853 auto getStateForBDV = [&](Value *baseValue) {
854 if (isKnownBaseResult(baseValue))
855 return BDVState(baseValue);
856 auto I = States.find(baseValue);
857 assert(I != States.end() && "lookup failed!");
861 bool Progress = true;
864 const size_t OldSize = States.size();
867 // We're only changing values in this loop, thus safe to keep iterators.
868 // Since this is computing a fixed point, the order of visit does not
869 // effect the result. TODO: We could use a worklist here and make this run
871 for (auto Pair : States) {
872 Value *BDV = Pair.first;
873 assert(!isKnownBaseResult(BDV) && "why did it get added?");
875 // Given an input value for the current instruction, return a BDVState
876 // instance which represents the BDV of that value.
877 auto getStateForInput = [&](Value *V) mutable {
878 Value *BDV = findBaseOrBDV(V, Cache);
879 return getStateForBDV(BDV);
883 if (SelectInst *SI = dyn_cast<SelectInst>(BDV)) {
884 NewState = meetBDVState(NewState, getStateForInput(SI->getTrueValue()));
886 meetBDVState(NewState, getStateForInput(SI->getFalseValue()));
887 } else if (PHINode *PN = dyn_cast<PHINode>(BDV)) {
888 for (Value *Val : PN->incoming_values())
889 NewState = meetBDVState(NewState, getStateForInput(Val));
890 } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) {
891 // The 'meet' for an extractelement is slightly trivial, but it's still
892 // useful in that it drives us to conflict if our input is.
894 meetBDVState(NewState, getStateForInput(EE->getVectorOperand()));
895 } else if (auto *IE = dyn_cast<InsertElementInst>(BDV)){
896 // Given there's a inherent type mismatch between the operands, will
897 // *always* produce Conflict.
898 NewState = meetBDVState(NewState, getStateForInput(IE->getOperand(0)));
899 NewState = meetBDVState(NewState, getStateForInput(IE->getOperand(1)));
901 // The only instance this does not return a Conflict is when both the
902 // vector operands are the same vector.
903 auto *SV = cast<ShuffleVectorInst>(BDV);
904 NewState = meetBDVState(NewState, getStateForInput(SV->getOperand(0)));
905 NewState = meetBDVState(NewState, getStateForInput(SV->getOperand(1)));
908 BDVState OldState = States[BDV];
909 if (OldState != NewState) {
911 States[BDV] = NewState;
915 assert(OldSize == States.size() &&
916 "fixed point shouldn't be adding any new nodes to state");
920 DEBUG(dbgs() << "States after meet iteration:\n");
921 for (auto Pair : States) {
922 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
926 // Insert Phis for all conflicts
927 // TODO: adjust naming patterns to avoid this order of iteration dependency
928 for (auto Pair : States) {
929 Instruction *I = cast<Instruction>(Pair.first);
930 BDVState State = Pair.second;
931 assert(!isKnownBaseResult(I) && "why did it get added?");
932 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
934 // extractelement instructions are a bit special in that we may need to
935 // insert an extract even when we know an exact base for the instruction.
936 // The problem is that we need to convert from a vector base to a scalar
937 // base for the particular indice we're interested in.
938 if (State.isBase() && isa<ExtractElementInst>(I) &&
939 isa<VectorType>(State.getBaseValue()->getType())) {
940 auto *EE = cast<ExtractElementInst>(I);
941 // TODO: In many cases, the new instruction is just EE itself. We should
942 // exploit this, but can't do it here since it would break the invariant
943 // about the BDV not being known to be a base.
944 auto *BaseInst = ExtractElementInst::Create(
945 State.getBaseValue(), EE->getIndexOperand(), "base_ee", EE);
946 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
947 States[I] = BDVState(BDVState::Base, BaseInst);
950 // Since we're joining a vector and scalar base, they can never be the
951 // same. As a result, we should always see insert element having reached
952 // the conflict state.
953 assert(!isa<InsertElementInst>(I) || State.isConflict());
955 if (!State.isConflict())
958 /// Create and insert a new instruction which will represent the base of
959 /// the given instruction 'I'.
960 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
961 if (isa<PHINode>(I)) {
962 BasicBlock *BB = I->getParent();
963 int NumPreds = std::distance(pred_begin(BB), pred_end(BB));
964 assert(NumPreds > 0 && "how did we reach here");
965 std::string Name = suffixed_name_or(I, ".base", "base_phi");
966 return PHINode::Create(I->getType(), NumPreds, Name, I);
967 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
968 // The undef will be replaced later
969 UndefValue *Undef = UndefValue::get(SI->getType());
970 std::string Name = suffixed_name_or(I, ".base", "base_select");
971 return SelectInst::Create(SI->getCondition(), Undef, Undef, Name, SI);
972 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
973 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
974 std::string Name = suffixed_name_or(I, ".base", "base_ee");
975 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
977 } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
978 UndefValue *VecUndef = UndefValue::get(IE->getOperand(0)->getType());
979 UndefValue *ScalarUndef = UndefValue::get(IE->getOperand(1)->getType());
980 std::string Name = suffixed_name_or(I, ".base", "base_ie");
981 return InsertElementInst::Create(VecUndef, ScalarUndef,
982 IE->getOperand(2), Name, IE);
984 auto *SV = cast<ShuffleVectorInst>(I);
985 UndefValue *VecUndef = UndefValue::get(SV->getOperand(0)->getType());
986 std::string Name = suffixed_name_or(I, ".base", "base_sv");
987 return new ShuffleVectorInst(VecUndef, VecUndef, SV->getOperand(2),
991 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
992 // Add metadata marking this as a base value
993 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
994 States[I] = BDVState(BDVState::Conflict, BaseInst);
997 // Returns a instruction which produces the base pointer for a given
998 // instruction. The instruction is assumed to be an input to one of the BDVs
999 // seen in the inference algorithm above. As such, we must either already
1000 // know it's base defining value is a base, or have inserted a new
1001 // instruction to propagate the base of it's BDV and have entered that newly
1002 // introduced instruction into the state table. In either case, we are
1003 // assured to be able to determine an instruction which produces it's base
1005 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
1006 Value *BDV = findBaseOrBDV(Input, Cache);
1007 Value *Base = nullptr;
1008 if (isKnownBaseResult(BDV)) {
1011 // Either conflict or base.
1012 assert(States.count(BDV));
1013 Base = States[BDV].getBaseValue();
1015 assert(Base && "Can't be null");
1016 // The cast is needed since base traversal may strip away bitcasts
1017 if (Base->getType() != Input->getType() && InsertPt)
1018 Base = new BitCastInst(Base, Input->getType(), "cast", InsertPt);
1022 // Fixup all the inputs of the new PHIs. Visit order needs to be
1023 // deterministic and predictable because we're naming newly created
1025 for (auto Pair : States) {
1026 Instruction *BDV = cast<Instruction>(Pair.first);
1027 BDVState State = Pair.second;
1029 assert(!isKnownBaseResult(BDV) && "why did it get added?");
1030 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
1031 if (!State.isConflict())
1034 if (PHINode *BasePHI = dyn_cast<PHINode>(State.getBaseValue())) {
1035 PHINode *PN = cast<PHINode>(BDV);
1036 unsigned NumPHIValues = PN->getNumIncomingValues();
1037 for (unsigned i = 0; i < NumPHIValues; i++) {
1038 Value *InVal = PN->getIncomingValue(i);
1039 BasicBlock *InBB = PN->getIncomingBlock(i);
1041 // If we've already seen InBB, add the same incoming value
1042 // we added for it earlier. The IR verifier requires phi
1043 // nodes with multiple entries from the same basic block
1044 // to have the same incoming value for each of those
1045 // entries. If we don't do this check here and basephi
1046 // has a different type than base, we'll end up adding two
1047 // bitcasts (and hence two distinct values) as incoming
1048 // values for the same basic block.
1050 int BlockIndex = BasePHI->getBasicBlockIndex(InBB);
1051 if (BlockIndex != -1) {
1052 Value *OldBase = BasePHI->getIncomingValue(BlockIndex);
1053 BasePHI->addIncoming(OldBase, InBB);
1056 Value *Base = getBaseForInput(InVal, nullptr);
1057 // In essence this assert states: the only way two values
1058 // incoming from the same basic block may be different is by
1059 // being different bitcasts of the same value. A cleanup
1060 // that remains TODO is changing findBaseOrBDV to return an
1061 // llvm::Value of the correct type (and still remain pure).
1062 // This will remove the need to add bitcasts.
1063 assert(Base->stripPointerCasts() == OldBase->stripPointerCasts() &&
1064 "Sanity -- findBaseOrBDV should be pure!");
1069 // Find the instruction which produces the base for each input. We may
1070 // need to insert a bitcast in the incoming block.
1071 // TODO: Need to split critical edges if insertion is needed
1072 Value *Base = getBaseForInput(InVal, InBB->getTerminator());
1073 BasePHI->addIncoming(Base, InBB);
1075 assert(BasePHI->getNumIncomingValues() == NumPHIValues);
1076 } else if (SelectInst *BaseSI =
1077 dyn_cast<SelectInst>(State.getBaseValue())) {
1078 SelectInst *SI = cast<SelectInst>(BDV);
1080 // Find the instruction which produces the base for each input.
1081 // We may need to insert a bitcast.
1082 BaseSI->setTrueValue(getBaseForInput(SI->getTrueValue(), BaseSI));
1083 BaseSI->setFalseValue(getBaseForInput(SI->getFalseValue(), BaseSI));
1084 } else if (auto *BaseEE =
1085 dyn_cast<ExtractElementInst>(State.getBaseValue())) {
1086 Value *InVal = cast<ExtractElementInst>(BDV)->getVectorOperand();
1087 // Find the instruction which produces the base for each input. We may
1088 // need to insert a bitcast.
1089 BaseEE->setOperand(0, getBaseForInput(InVal, BaseEE));
1090 } else if (auto *BaseIE = dyn_cast<InsertElementInst>(State.getBaseValue())){
1091 auto *BdvIE = cast<InsertElementInst>(BDV);
1092 auto UpdateOperand = [&](int OperandIdx) {
1093 Value *InVal = BdvIE->getOperand(OperandIdx);
1094 Value *Base = getBaseForInput(InVal, BaseIE);
1095 BaseIE->setOperand(OperandIdx, Base);
1097 UpdateOperand(0); // vector operand
1098 UpdateOperand(1); // scalar operand
1100 auto *BaseSV = cast<ShuffleVectorInst>(State.getBaseValue());
1101 auto *BdvSV = cast<ShuffleVectorInst>(BDV);
1102 auto UpdateOperand = [&](int OperandIdx) {
1103 Value *InVal = BdvSV->getOperand(OperandIdx);
1104 Value *Base = getBaseForInput(InVal, BaseSV);
1105 BaseSV->setOperand(OperandIdx, Base);
1107 UpdateOperand(0); // vector operand
1108 UpdateOperand(1); // vector operand
1112 // Cache all of our results so we can cheaply reuse them
1113 // NOTE: This is actually two caches: one of the base defining value
1114 // relation and one of the base pointer relation! FIXME
1115 for (auto Pair : States) {
1116 auto *BDV = Pair.first;
1117 Value *Base = Pair.second.getBaseValue();
1118 assert(BDV && Base);
1119 assert(!isKnownBaseResult(BDV) && "why did it get added?");
1121 DEBUG(dbgs() << "Updating base value cache"
1122 << " for: " << BDV->getName() << " from: "
1123 << (Cache.count(BDV) ? Cache[BDV]->getName().str() : "none")
1124 << " to: " << Base->getName() << "\n");
1126 if (Cache.count(BDV)) {
1127 assert(isKnownBaseResult(Base) &&
1128 "must be something we 'know' is a base pointer");
1129 // Once we transition from the BDV relation being store in the Cache to
1130 // the base relation being stored, it must be stable
1131 assert((!isKnownBaseResult(Cache[BDV]) || Cache[BDV] == Base) &&
1132 "base relation should be stable");
1136 assert(Cache.count(Def));
1140 // For a set of live pointers (base and/or derived), identify the base
1141 // pointer of the object which they are derived from. This routine will
1142 // mutate the IR graph as needed to make the 'base' pointer live at the
1143 // definition site of 'derived'. This ensures that any use of 'derived' can
1144 // also use 'base'. This may involve the insertion of a number of
1145 // additional PHI nodes.
1147 // preconditions: live is a set of pointer type Values
1149 // side effects: may insert PHI nodes into the existing CFG, will preserve
1150 // CFG, will not remove or mutate any existing nodes
1152 // post condition: PointerToBase contains one (derived, base) pair for every
1153 // pointer in live. Note that derived can be equal to base if the original
1154 // pointer was a base pointer.
1156 findBasePointers(const StatepointLiveSetTy &live,
1157 MapVector<Value *, Value *> &PointerToBase,
1158 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1159 for (Value *ptr : live) {
1160 Value *base = findBasePointer(ptr, DVCache);
1161 assert(base && "failed to find base pointer");
1162 PointerToBase[ptr] = base;
1163 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1164 DT->dominates(cast<Instruction>(base)->getParent(),
1165 cast<Instruction>(ptr)->getParent())) &&
1166 "The base we found better dominate the derived pointer");
1170 /// Find the required based pointers (and adjust the live set) for the given
1172 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1174 PartiallyConstructedSafepointRecord &result) {
1175 MapVector<Value *, Value *> PointerToBase;
1176 findBasePointers(result.LiveSet, PointerToBase, &DT, DVCache);
1178 if (PrintBasePointers) {
1179 errs() << "Base Pairs (w/o Relocation):\n";
1180 for (auto &Pair : PointerToBase) {
1181 errs() << " derived ";
1182 Pair.first->printAsOperand(errs(), false);
1184 Pair.second->printAsOperand(errs(), false);
1189 result.PointerToBase = PointerToBase;
1192 /// Given an updated version of the dataflow liveness results, update the
1193 /// liveset and base pointer maps for the call site CS.
1194 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1196 PartiallyConstructedSafepointRecord &result);
1198 static void recomputeLiveInValues(
1199 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1200 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1201 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1202 // again. The old values are still live and will help it stabilize quickly.
1203 GCPtrLivenessData RevisedLivenessData;
1204 computeLiveInValues(DT, F, RevisedLivenessData);
1205 for (size_t i = 0; i < records.size(); i++) {
1206 struct PartiallyConstructedSafepointRecord &info = records[i];
1207 recomputeLiveInValues(RevisedLivenessData, toUpdate[i], info);
1211 // When inserting gc.relocate and gc.result calls, we need to ensure there are
1212 // no uses of the original value / return value between the gc.statepoint and
1213 // the gc.relocate / gc.result call. One case which can arise is a phi node
1214 // starting one of the successor blocks. We also need to be able to insert the
1215 // gc.relocates only on the path which goes through the statepoint. We might
1216 // need to split an edge to make this possible.
1218 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1219 DominatorTree &DT) {
1220 BasicBlock *Ret = BB;
1221 if (!BB->getUniquePredecessor())
1222 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1224 // Now that 'Ret' has unique predecessor we can safely remove all phi nodes
1226 FoldSingleEntryPHINodes(Ret);
1227 assert(!isa<PHINode>(Ret->begin()) &&
1228 "All PHI nodes should have been removed!");
1230 // At this point, we can safely insert a gc.relocate or gc.result as the first
1231 // instruction in Ret if needed.
1235 // Create new attribute set containing only attributes which can be transferred
1236 // from original call to the safepoint.
1237 static AttributeList legalizeCallAttributes(AttributeList AL) {
1241 // Remove the readonly, readnone, and statepoint function attributes.
1242 AttrBuilder FnAttrs = AL.getFnAttributes();
1243 FnAttrs.removeAttribute(Attribute::ReadNone);
1244 FnAttrs.removeAttribute(Attribute::ReadOnly);
1245 for (Attribute A : AL.getFnAttributes()) {
1246 if (isStatepointDirectiveAttr(A))
1250 // Just skip parameter and return attributes for now
1251 LLVMContext &Ctx = AL.getContext();
1252 return AttributeList::get(Ctx, AttributeList::FunctionIndex,
1253 AttributeSet::get(Ctx, FnAttrs));
1256 /// Helper function to place all gc relocates necessary for the given
1259 /// liveVariables - list of variables to be relocated.
1260 /// liveStart - index of the first live variable.
1261 /// basePtrs - base pointers.
1262 /// statepointToken - statepoint instruction to which relocates should be
1264 /// Builder - Llvm IR builder to be used to construct new calls.
1265 static void CreateGCRelocates(ArrayRef<Value *> LiveVariables,
1266 const int LiveStart,
1267 ArrayRef<Value *> BasePtrs,
1268 Instruction *StatepointToken,
1269 IRBuilder<> Builder) {
1270 if (LiveVariables.empty())
1273 auto FindIndex = [](ArrayRef<Value *> LiveVec, Value *Val) {
1274 auto ValIt = llvm::find(LiveVec, Val);
1275 assert(ValIt != LiveVec.end() && "Val not found in LiveVec!");
1276 size_t Index = std::distance(LiveVec.begin(), ValIt);
1277 assert(Index < LiveVec.size() && "Bug in std::find?");
1280 Module *M = StatepointToken->getModule();
1282 // All gc_relocate are generated as i8 addrspace(1)* (or a vector type whose
1283 // element type is i8 addrspace(1)*). We originally generated unique
1284 // declarations for each pointer type, but this proved problematic because
1285 // the intrinsic mangling code is incomplete and fragile. Since we're moving
1286 // towards a single unified pointer type anyways, we can just cast everything
1287 // to an i8* of the right address space. A bitcast is added later to convert
1288 // gc_relocate to the actual value's type.
1289 auto getGCRelocateDecl = [&] (Type *Ty) {
1290 assert(isHandledGCPointerType(Ty));
1291 auto AS = Ty->getScalarType()->getPointerAddressSpace();
1292 Type *NewTy = Type::getInt8PtrTy(M->getContext(), AS);
1293 if (auto *VT = dyn_cast<VectorType>(Ty))
1294 NewTy = VectorType::get(NewTy, VT->getNumElements());
1295 return Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate,
1299 // Lazily populated map from input types to the canonicalized form mentioned
1300 // in the comment above. This should probably be cached somewhere more
1302 DenseMap<Type*, Value*> TypeToDeclMap;
1304 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1305 // Generate the gc.relocate call and save the result
1307 Builder.getInt32(LiveStart + FindIndex(LiveVariables, BasePtrs[i]));
1308 Value *LiveIdx = Builder.getInt32(LiveStart + i);
1310 Type *Ty = LiveVariables[i]->getType();
1311 if (!TypeToDeclMap.count(Ty))
1312 TypeToDeclMap[Ty] = getGCRelocateDecl(Ty);
1313 Value *GCRelocateDecl = TypeToDeclMap[Ty];
1315 // only specify a debug name if we can give a useful one
1316 CallInst *Reloc = Builder.CreateCall(
1317 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1318 suffixed_name_or(LiveVariables[i], ".relocated", ""));
1319 // Trick CodeGen into thinking there are lots of free registers at this
1321 Reloc->setCallingConv(CallingConv::Cold);
1327 /// This struct is used to defer RAUWs and `eraseFromParent` s. Using this
1328 /// avoids having to worry about keeping around dangling pointers to Values.
1329 class DeferredReplacement {
1330 AssertingVH<Instruction> Old;
1331 AssertingVH<Instruction> New;
1332 bool IsDeoptimize = false;
1334 DeferredReplacement() = default;
1337 static DeferredReplacement createRAUW(Instruction *Old, Instruction *New) {
1338 assert(Old != New && Old && New &&
1339 "Cannot RAUW equal values or to / from null!");
1341 DeferredReplacement D;
1347 static DeferredReplacement createDelete(Instruction *ToErase) {
1348 DeferredReplacement D;
1353 static DeferredReplacement createDeoptimizeReplacement(Instruction *Old) {
1355 auto *F = cast<CallInst>(Old)->getCalledFunction();
1356 assert(F && F->getIntrinsicID() == Intrinsic::experimental_deoptimize &&
1357 "Only way to construct a deoptimize deferred replacement");
1359 DeferredReplacement D;
1361 D.IsDeoptimize = true;
1365 /// Does the task represented by this instance.
1366 void doReplacement() {
1367 Instruction *OldI = Old;
1368 Instruction *NewI = New;
1370 assert(OldI != NewI && "Disallowed at construction?!");
1371 assert((!IsDeoptimize || !New) &&
1372 "Deoptimize instrinsics are not replaced!");
1378 OldI->replaceAllUsesWith(NewI);
1381 // Note: we've inserted instructions, so the call to llvm.deoptimize may
1382 // not necessarilly be followed by the matching return.
1383 auto *RI = cast<ReturnInst>(OldI->getParent()->getTerminator());
1384 new UnreachableInst(RI->getContext(), RI);
1385 RI->eraseFromParent();
1388 OldI->eraseFromParent();
1392 } // end anonymous namespace
1394 static StringRef getDeoptLowering(CallSite CS) {
1395 const char *DeoptLowering = "deopt-lowering";
1396 if (CS.hasFnAttr(DeoptLowering)) {
1397 // FIXME: CallSite has a *really* confusing interface around attributes
1399 const AttributeList &CSAS = CS.getAttributes();
1400 if (CSAS.hasAttribute(AttributeList::FunctionIndex, DeoptLowering))
1401 return CSAS.getAttribute(AttributeList::FunctionIndex, DeoptLowering)
1402 .getValueAsString();
1403 Function *F = CS.getCalledFunction();
1404 assert(F && F->hasFnAttribute(DeoptLowering));
1405 return F->getFnAttribute(DeoptLowering).getValueAsString();
1407 return "live-through";
1411 makeStatepointExplicitImpl(const CallSite CS, /* to replace */
1412 const SmallVectorImpl<Value *> &BasePtrs,
1413 const SmallVectorImpl<Value *> &LiveVariables,
1414 PartiallyConstructedSafepointRecord &Result,
1415 std::vector<DeferredReplacement> &Replacements) {
1416 assert(BasePtrs.size() == LiveVariables.size());
1418 // Then go ahead and use the builder do actually do the inserts. We insert
1419 // immediately before the previous instruction under the assumption that all
1420 // arguments will be available here. We can't insert afterwards since we may
1421 // be replacing a terminator.
1422 Instruction *InsertBefore = CS.getInstruction();
1423 IRBuilder<> Builder(InsertBefore);
1425 ArrayRef<Value *> GCArgs(LiveVariables);
1426 uint64_t StatepointID = StatepointDirectives::DefaultStatepointID;
1427 uint32_t NumPatchBytes = 0;
1428 uint32_t Flags = uint32_t(StatepointFlags::None);
1430 ArrayRef<Use> CallArgs(CS.arg_begin(), CS.arg_end());
1431 ArrayRef<Use> DeoptArgs = GetDeoptBundleOperands(CS);
1432 ArrayRef<Use> TransitionArgs;
1433 if (auto TransitionBundle =
1434 CS.getOperandBundle(LLVMContext::OB_gc_transition)) {
1435 Flags |= uint32_t(StatepointFlags::GCTransition);
1436 TransitionArgs = TransitionBundle->Inputs;
1439 // Instead of lowering calls to @llvm.experimental.deoptimize as normal calls
1440 // with a return value, we lower then as never returning calls to
1441 // __llvm_deoptimize that are followed by unreachable to get better codegen.
1442 bool IsDeoptimize = false;
1444 StatepointDirectives SD =
1445 parseStatepointDirectivesFromAttrs(CS.getAttributes());
1446 if (SD.NumPatchBytes)
1447 NumPatchBytes = *SD.NumPatchBytes;
1448 if (SD.StatepointID)
1449 StatepointID = *SD.StatepointID;
1451 // Pass through the requested lowering if any. The default is live-through.
1452 StringRef DeoptLowering = getDeoptLowering(CS);
1453 if (DeoptLowering.equals("live-in"))
1454 Flags |= uint32_t(StatepointFlags::DeoptLiveIn);
1456 assert(DeoptLowering.equals("live-through") && "Unsupported value!");
1459 Value *CallTarget = CS.getCalledValue();
1460 if (Function *F = dyn_cast<Function>(CallTarget)) {
1461 if (F->getIntrinsicID() == Intrinsic::experimental_deoptimize) {
1462 // Calls to llvm.experimental.deoptimize are lowered to calls to the
1463 // __llvm_deoptimize symbol. We want to resolve this now, since the
1464 // verifier does not allow taking the address of an intrinsic function.
1466 SmallVector<Type *, 8> DomainTy;
1467 for (Value *Arg : CallArgs)
1468 DomainTy.push_back(Arg->getType());
1469 auto *FTy = FunctionType::get(Type::getVoidTy(F->getContext()), DomainTy,
1470 /* isVarArg = */ false);
1472 // Note: CallTarget can be a bitcast instruction of a symbol if there are
1473 // calls to @llvm.experimental.deoptimize with different argument types in
1474 // the same module. This is fine -- we assume the frontend knew what it
1475 // was doing when generating this kind of IR.
1477 F->getParent()->getOrInsertFunction("__llvm_deoptimize", FTy);
1479 IsDeoptimize = true;
1483 // Create the statepoint given all the arguments
1484 Instruction *Token = nullptr;
1486 CallInst *ToReplace = cast<CallInst>(CS.getInstruction());
1487 CallInst *Call = Builder.CreateGCStatepointCall(
1488 StatepointID, NumPatchBytes, CallTarget, Flags, CallArgs,
1489 TransitionArgs, DeoptArgs, GCArgs, "safepoint_token");
1491 Call->setTailCallKind(ToReplace->getTailCallKind());
1492 Call->setCallingConv(ToReplace->getCallingConv());
1494 // Currently we will fail on parameter attributes and on certain
1495 // function attributes. In case if we can handle this set of attributes -
1496 // set up function attrs directly on statepoint and return attrs later for
1497 // gc_result intrinsic.
1498 Call->setAttributes(legalizeCallAttributes(ToReplace->getAttributes()));
1502 // Put the following gc_result and gc_relocate calls immediately after the
1503 // the old call (which we're about to delete)
1504 assert(ToReplace->getNextNode() && "Not a terminator, must have next!");
1505 Builder.SetInsertPoint(ToReplace->getNextNode());
1506 Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc());
1508 InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction());
1510 // Insert the new invoke into the old block. We'll remove the old one in a
1511 // moment at which point this will become the new terminator for the
1513 InvokeInst *Invoke = Builder.CreateGCStatepointInvoke(
1514 StatepointID, NumPatchBytes, CallTarget, ToReplace->getNormalDest(),
1515 ToReplace->getUnwindDest(), Flags, CallArgs, TransitionArgs, DeoptArgs,
1516 GCArgs, "statepoint_token");
1518 Invoke->setCallingConv(ToReplace->getCallingConv());
1520 // Currently we will fail on parameter attributes and on certain
1521 // function attributes. In case if we can handle this set of attributes -
1522 // set up function attrs directly on statepoint and return attrs later for
1523 // gc_result intrinsic.
1524 Invoke->setAttributes(legalizeCallAttributes(ToReplace->getAttributes()));
1528 // Generate gc relocates in exceptional path
1529 BasicBlock *UnwindBlock = ToReplace->getUnwindDest();
1530 assert(!isa<PHINode>(UnwindBlock->begin()) &&
1531 UnwindBlock->getUniquePredecessor() &&
1532 "can't safely insert in this block!");
1534 Builder.SetInsertPoint(&*UnwindBlock->getFirstInsertionPt());
1535 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc());
1537 // Attach exceptional gc relocates to the landingpad.
1538 Instruction *ExceptionalToken = UnwindBlock->getLandingPadInst();
1539 Result.UnwindToken = ExceptionalToken;
1541 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1542 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, ExceptionalToken,
1545 // Generate gc relocates and returns for normal block
1546 BasicBlock *NormalDest = ToReplace->getNormalDest();
1547 assert(!isa<PHINode>(NormalDest->begin()) &&
1548 NormalDest->getUniquePredecessor() &&
1549 "can't safely insert in this block!");
1551 Builder.SetInsertPoint(&*NormalDest->getFirstInsertionPt());
1553 // gc relocates will be generated later as if it were regular call
1556 assert(Token && "Should be set in one of the above branches!");
1559 // If we're wrapping an @llvm.experimental.deoptimize in a statepoint, we
1560 // transform the tail-call like structure to a call to a void function
1561 // followed by unreachable to get better codegen.
1562 Replacements.push_back(
1563 DeferredReplacement::createDeoptimizeReplacement(CS.getInstruction()));
1565 Token->setName("statepoint_token");
1566 if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) {
1568 CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : "";
1569 CallInst *GCResult = Builder.CreateGCResult(Token, CS.getType(), Name);
1570 GCResult->setAttributes(
1571 AttributeList::get(GCResult->getContext(), AttributeList::ReturnIndex,
1572 CS.getAttributes().getRetAttributes()));
1574 // We cannot RAUW or delete CS.getInstruction() because it could be in the
1575 // live set of some other safepoint, in which case that safepoint's
1576 // PartiallyConstructedSafepointRecord will hold a raw pointer to this
1577 // llvm::Instruction. Instead, we defer the replacement and deletion to
1578 // after the live sets have been made explicit in the IR, and we no longer
1579 // have raw pointers to worry about.
1580 Replacements.emplace_back(
1581 DeferredReplacement::createRAUW(CS.getInstruction(), GCResult));
1583 Replacements.emplace_back(
1584 DeferredReplacement::createDelete(CS.getInstruction()));
1588 Result.StatepointToken = Token;
1590 // Second, create a gc.relocate for every live variable
1591 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx();
1592 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, Token, Builder);
1595 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1596 // which make the relocations happening at this safepoint explicit.
1598 // WARNING: Does not do any fixup to adjust users of the original live
1599 // values. That's the callers responsibility.
1601 makeStatepointExplicit(DominatorTree &DT, CallSite CS,
1602 PartiallyConstructedSafepointRecord &Result,
1603 std::vector<DeferredReplacement> &Replacements) {
1604 const auto &LiveSet = Result.LiveSet;
1605 const auto &PointerToBase = Result.PointerToBase;
1607 // Convert to vector for efficient cross referencing.
1608 SmallVector<Value *, 64> BaseVec, LiveVec;
1609 LiveVec.reserve(LiveSet.size());
1610 BaseVec.reserve(LiveSet.size());
1611 for (Value *L : LiveSet) {
1612 LiveVec.push_back(L);
1613 assert(PointerToBase.count(L));
1614 Value *Base = PointerToBase.find(L)->second;
1615 BaseVec.push_back(Base);
1617 assert(LiveVec.size() == BaseVec.size());
1619 // Do the actual rewriting and delete the old statepoint
1620 makeStatepointExplicitImpl(CS, BaseVec, LiveVec, Result, Replacements);
1623 // Helper function for the relocationViaAlloca.
1625 // It receives iterator to the statepoint gc relocates and emits a store to the
1626 // assigned location (via allocaMap) for the each one of them. It adds the
1627 // visited values into the visitedLiveValues set, which we will later use them
1628 // for sanity checking.
1630 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1631 DenseMap<Value *, Value *> &AllocaMap,
1632 DenseSet<Value *> &VisitedLiveValues) {
1633 for (User *U : GCRelocs) {
1634 GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U);
1638 Value *OriginalValue = Relocate->getDerivedPtr();
1639 assert(AllocaMap.count(OriginalValue));
1640 Value *Alloca = AllocaMap[OriginalValue];
1642 // Emit store into the related alloca
1643 // All gc_relocates are i8 addrspace(1)* typed, and it must be bitcasted to
1644 // the correct type according to alloca.
1645 assert(Relocate->getNextNode() &&
1646 "Should always have one since it's not a terminator");
1647 IRBuilder<> Builder(Relocate->getNextNode());
1648 Value *CastedRelocatedValue =
1649 Builder.CreateBitCast(Relocate,
1650 cast<AllocaInst>(Alloca)->getAllocatedType(),
1651 suffixed_name_or(Relocate, ".casted", ""));
1653 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca);
1654 Store->insertAfter(cast<Instruction>(CastedRelocatedValue));
1657 VisitedLiveValues.insert(OriginalValue);
1662 // Helper function for the "relocationViaAlloca". Similar to the
1663 // "insertRelocationStores" but works for rematerialized values.
1664 static void insertRematerializationStores(
1665 const RematerializedValueMapTy &RematerializedValues,
1666 DenseMap<Value *, Value *> &AllocaMap,
1667 DenseSet<Value *> &VisitedLiveValues) {
1668 for (auto RematerializedValuePair: RematerializedValues) {
1669 Instruction *RematerializedValue = RematerializedValuePair.first;
1670 Value *OriginalValue = RematerializedValuePair.second;
1672 assert(AllocaMap.count(OriginalValue) &&
1673 "Can not find alloca for rematerialized value");
1674 Value *Alloca = AllocaMap[OriginalValue];
1676 StoreInst *Store = new StoreInst(RematerializedValue, Alloca);
1677 Store->insertAfter(RematerializedValue);
1680 VisitedLiveValues.insert(OriginalValue);
1685 /// Do all the relocation update via allocas and mem2reg
1686 static void relocationViaAlloca(
1687 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1688 ArrayRef<PartiallyConstructedSafepointRecord> Records) {
1690 // record initial number of (static) allocas; we'll check we have the same
1691 // number when we get done.
1692 int InitialAllocaNum = 0;
1693 for (Instruction &I : F.getEntryBlock())
1694 if (isa<AllocaInst>(I))
1698 // TODO-PERF: change data structures, reserve
1699 DenseMap<Value *, Value *> AllocaMap;
1700 SmallVector<AllocaInst *, 200> PromotableAllocas;
1701 // Used later to chack that we have enough allocas to store all values
1702 std::size_t NumRematerializedValues = 0;
1703 PromotableAllocas.reserve(Live.size());
1705 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1706 // "PromotableAllocas"
1707 const DataLayout &DL = F.getParent()->getDataLayout();
1708 auto emitAllocaFor = [&](Value *LiveValue) {
1709 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(),
1710 DL.getAllocaAddrSpace(), "",
1711 F.getEntryBlock().getFirstNonPHI());
1712 AllocaMap[LiveValue] = Alloca;
1713 PromotableAllocas.push_back(Alloca);
1716 // Emit alloca for each live gc pointer
1717 for (Value *V : Live)
1720 // Emit allocas for rematerialized values
1721 for (const auto &Info : Records)
1722 for (auto RematerializedValuePair : Info.RematerializedValues) {
1723 Value *OriginalValue = RematerializedValuePair.second;
1724 if (AllocaMap.count(OriginalValue) != 0)
1727 emitAllocaFor(OriginalValue);
1728 ++NumRematerializedValues;
1731 // The next two loops are part of the same conceptual operation. We need to
1732 // insert a store to the alloca after the original def and at each
1733 // redefinition. We need to insert a load before each use. These are split
1734 // into distinct loops for performance reasons.
1736 // Update gc pointer after each statepoint: either store a relocated value or
1737 // null (if no relocated value was found for this gc pointer and it is not a
1738 // gc_result). This must happen before we update the statepoint with load of
1739 // alloca otherwise we lose the link between statepoint and old def.
1740 for (const auto &Info : Records) {
1741 Value *Statepoint = Info.StatepointToken;
1743 // This will be used for consistency check
1744 DenseSet<Value *> VisitedLiveValues;
1746 // Insert stores for normal statepoint gc relocates
1747 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1749 // In case if it was invoke statepoint
1750 // we will insert stores for exceptional path gc relocates.
1751 if (isa<InvokeInst>(Statepoint)) {
1752 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1756 // Do similar thing with rematerialized values
1757 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1760 if (ClobberNonLive) {
1761 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1762 // the gc.statepoint. This will turn some subtle GC problems into
1763 // slightly easier to debug SEGVs. Note that on large IR files with
1764 // lots of gc.statepoints this is extremely costly both memory and time
1766 SmallVector<AllocaInst *, 64> ToClobber;
1767 for (auto Pair : AllocaMap) {
1768 Value *Def = Pair.first;
1769 AllocaInst *Alloca = cast<AllocaInst>(Pair.second);
1771 // This value was relocated
1772 if (VisitedLiveValues.count(Def)) {
1775 ToClobber.push_back(Alloca);
1778 auto InsertClobbersAt = [&](Instruction *IP) {
1779 for (auto *AI : ToClobber) {
1780 auto PT = cast<PointerType>(AI->getAllocatedType());
1781 Constant *CPN = ConstantPointerNull::get(PT);
1782 StoreInst *Store = new StoreInst(CPN, AI);
1783 Store->insertBefore(IP);
1787 // Insert the clobbering stores. These may get intermixed with the
1788 // gc.results and gc.relocates, but that's fine.
1789 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1790 InsertClobbersAt(&*II->getNormalDest()->getFirstInsertionPt());
1791 InsertClobbersAt(&*II->getUnwindDest()->getFirstInsertionPt());
1793 InsertClobbersAt(cast<Instruction>(Statepoint)->getNextNode());
1798 // Update use with load allocas and add store for gc_relocated.
1799 for (auto Pair : AllocaMap) {
1800 Value *Def = Pair.first;
1801 Value *Alloca = Pair.second;
1803 // We pre-record the uses of allocas so that we dont have to worry about
1804 // later update that changes the user information..
1806 SmallVector<Instruction *, 20> Uses;
1807 // PERF: trade a linear scan for repeated reallocation
1808 Uses.reserve(std::distance(Def->user_begin(), Def->user_end()));
1809 for (User *U : Def->users()) {
1810 if (!isa<ConstantExpr>(U)) {
1811 // If the def has a ConstantExpr use, then the def is either a
1812 // ConstantExpr use itself or null. In either case
1813 // (recursively in the first, directly in the second), the oop
1814 // it is ultimately dependent on is null and this particular
1815 // use does not need to be fixed up.
1816 Uses.push_back(cast<Instruction>(U));
1820 std::sort(Uses.begin(), Uses.end());
1821 auto Last = std::unique(Uses.begin(), Uses.end());
1822 Uses.erase(Last, Uses.end());
1824 for (Instruction *Use : Uses) {
1825 if (isa<PHINode>(Use)) {
1826 PHINode *Phi = cast<PHINode>(Use);
1827 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1828 if (Def == Phi->getIncomingValue(i)) {
1829 LoadInst *Load = new LoadInst(
1830 Alloca, "", Phi->getIncomingBlock(i)->getTerminator());
1831 Phi->setIncomingValue(i, Load);
1835 LoadInst *Load = new LoadInst(Alloca, "", Use);
1836 Use->replaceUsesOfWith(Def, Load);
1840 // Emit store for the initial gc value. Store must be inserted after load,
1841 // otherwise store will be in alloca's use list and an extra load will be
1842 // inserted before it.
1843 StoreInst *Store = new StoreInst(Def, Alloca);
1844 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1845 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1846 // InvokeInst is a TerminatorInst so the store need to be inserted
1847 // into its normal destination block.
1848 BasicBlock *NormalDest = Invoke->getNormalDest();
1849 Store->insertBefore(NormalDest->getFirstNonPHI());
1851 assert(!Inst->isTerminator() &&
1852 "The only TerminatorInst that can produce a value is "
1853 "InvokeInst which is handled above.");
1854 Store->insertAfter(Inst);
1857 assert(isa<Argument>(Def));
1858 Store->insertAfter(cast<Instruction>(Alloca));
1862 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1863 "we must have the same allocas with lives");
1864 if (!PromotableAllocas.empty()) {
1865 // Apply mem2reg to promote alloca to SSA
1866 PromoteMemToReg(PromotableAllocas, DT);
1870 for (auto &I : F.getEntryBlock())
1871 if (isa<AllocaInst>(I))
1873 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1877 /// Implement a unique function which doesn't require we sort the input
1878 /// vector. Doing so has the effect of changing the output of a couple of
1879 /// tests in ways which make them less useful in testing fused safepoints.
1880 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1881 SmallSet<T, 8> Seen;
1882 Vec.erase(remove_if(Vec, [&](const T &V) { return !Seen.insert(V).second; }),
1886 /// Insert holders so that each Value is obviously live through the entire
1887 /// lifetime of the call.
1888 static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values,
1889 SmallVectorImpl<CallInst *> &Holders) {
1891 // No values to hold live, might as well not insert the empty holder
1894 Module *M = CS.getInstruction()->getModule();
1895 // Use a dummy vararg function to actually hold the values live
1896 Function *Func = cast<Function>(M->getOrInsertFunction(
1897 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true)));
1899 // For call safepoints insert dummy calls right after safepoint
1900 Holders.push_back(CallInst::Create(Func, Values, "",
1901 &*++CS.getInstruction()->getIterator()));
1904 // For invoke safepooints insert dummy calls both in normal and
1905 // exceptional destination blocks
1906 auto *II = cast<InvokeInst>(CS.getInstruction());
1907 Holders.push_back(CallInst::Create(
1908 Func, Values, "", &*II->getNormalDest()->getFirstInsertionPt()));
1909 Holders.push_back(CallInst::Create(
1910 Func, Values, "", &*II->getUnwindDest()->getFirstInsertionPt()));
1913 static void findLiveReferences(
1914 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate,
1915 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1916 GCPtrLivenessData OriginalLivenessData;
1917 computeLiveInValues(DT, F, OriginalLivenessData);
1918 for (size_t i = 0; i < records.size(); i++) {
1919 struct PartiallyConstructedSafepointRecord &info = records[i];
1920 analyzeParsePointLiveness(DT, OriginalLivenessData, toUpdate[i], info);
1924 // Helper function for the "rematerializeLiveValues". It walks use chain
1925 // starting from the "CurrentValue" until it reaches the root of the chain, i.e.
1926 // the base or a value it cannot process. Only "simple" values are processed
1927 // (currently it is GEP's and casts). The returned root is examined by the
1928 // callers of findRematerializableChainToBasePointer. Fills "ChainToBase" array
1929 // with all visited values.
1930 static Value* findRematerializableChainToBasePointer(
1931 SmallVectorImpl<Instruction*> &ChainToBase,
1932 Value *CurrentValue) {
1933 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
1934 ChainToBase.push_back(GEP);
1935 return findRematerializableChainToBasePointer(ChainToBase,
1936 GEP->getPointerOperand());
1939 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
1940 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
1943 ChainToBase.push_back(CI);
1944 return findRematerializableChainToBasePointer(ChainToBase,
1948 // We have reached the root of the chain, which is either equal to the base or
1949 // is the first unsupported value along the use chain.
1950 return CurrentValue;
1953 // Helper function for the "rematerializeLiveValues". Compute cost of the use
1954 // chain we are going to rematerialize.
1956 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
1957 TargetTransformInfo &TTI) {
1960 for (Instruction *Instr : Chain) {
1961 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
1962 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
1963 "non noop cast is found during rematerialization");
1965 Type *SrcTy = CI->getOperand(0)->getType();
1966 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy, CI);
1968 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
1969 // Cost of the address calculation
1970 Type *ValTy = GEP->getSourceElementType();
1971 Cost += TTI.getAddressComputationCost(ValTy);
1973 // And cost of the GEP itself
1974 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
1975 // allowed for the external usage)
1976 if (!GEP->hasAllConstantIndices())
1980 llvm_unreachable("unsupported instruciton type during rematerialization");
1987 static bool AreEquivalentPhiNodes(PHINode &OrigRootPhi, PHINode &AlternateRootPhi) {
1988 unsigned PhiNum = OrigRootPhi.getNumIncomingValues();
1989 if (PhiNum != AlternateRootPhi.getNumIncomingValues() ||
1990 OrigRootPhi.getParent() != AlternateRootPhi.getParent())
1992 // Map of incoming values and their corresponding basic blocks of
1994 SmallDenseMap<Value *, BasicBlock *, 8> CurrentIncomingValues;
1995 for (unsigned i = 0; i < PhiNum; i++)
1996 CurrentIncomingValues[OrigRootPhi.getIncomingValue(i)] =
1997 OrigRootPhi.getIncomingBlock(i);
1999 // Both current and base PHIs should have same incoming values and
2000 // the same basic blocks corresponding to the incoming values.
2001 for (unsigned i = 0; i < PhiNum; i++) {
2003 CurrentIncomingValues.find(AlternateRootPhi.getIncomingValue(i));
2004 if (CIVI == CurrentIncomingValues.end())
2006 BasicBlock *CurrentIncomingBB = CIVI->second;
2007 if (CurrentIncomingBB != AlternateRootPhi.getIncomingBlock(i))
2013 // From the statepoint live set pick values that are cheaper to recompute then
2014 // to relocate. Remove this values from the live set, rematerialize them after
2015 // statepoint and record them in "Info" structure. Note that similar to
2016 // relocated values we don't do any user adjustments here.
2017 static void rematerializeLiveValues(CallSite CS,
2018 PartiallyConstructedSafepointRecord &Info,
2019 TargetTransformInfo &TTI) {
2020 const unsigned int ChainLengthThreshold = 10;
2022 // Record values we are going to delete from this statepoint live set.
2023 // We can not di this in following loop due to iterator invalidation.
2024 SmallVector<Value *, 32> LiveValuesToBeDeleted;
2026 for (Value *LiveValue: Info.LiveSet) {
2027 // For each live pointer find it's defining chain
2028 SmallVector<Instruction *, 3> ChainToBase;
2029 assert(Info.PointerToBase.count(LiveValue));
2030 Value *RootOfChain =
2031 findRematerializableChainToBasePointer(ChainToBase,
2034 // Nothing to do, or chain is too long
2035 if ( ChainToBase.size() == 0 ||
2036 ChainToBase.size() > ChainLengthThreshold)
2039 // Handle the scenario where the RootOfChain is not equal to the
2040 // Base Value, but they are essentially the same phi values.
2041 if (RootOfChain != Info.PointerToBase[LiveValue]) {
2042 PHINode *OrigRootPhi = dyn_cast<PHINode>(RootOfChain);
2043 PHINode *AlternateRootPhi = dyn_cast<PHINode>(Info.PointerToBase[LiveValue]);
2044 if (!OrigRootPhi || !AlternateRootPhi)
2046 // PHI nodes that have the same incoming values, and belonging to the same
2047 // basic blocks are essentially the same SSA value. When the original phi
2048 // has incoming values with different base pointers, the original phi is
2049 // marked as conflict, and an additional `AlternateRootPhi` with the same
2050 // incoming values get generated by the findBasePointer function. We need
2051 // to identify the newly generated AlternateRootPhi (.base version of phi)
2052 // and RootOfChain (the original phi node itself) are the same, so that we
2053 // can rematerialize the gep and casts. This is a workaround for the
2054 // deficiency in the findBasePointer algorithm.
2055 if (!AreEquivalentPhiNodes(*OrigRootPhi, *AlternateRootPhi))
2057 // Now that the phi nodes are proved to be the same, assert that
2058 // findBasePointer's newly generated AlternateRootPhi is present in the
2059 // liveset of the call.
2060 assert(Info.LiveSet.count(AlternateRootPhi));
2062 // Compute cost of this chain
2063 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
2064 // TODO: We can also account for cases when we will be able to remove some
2065 // of the rematerialized values by later optimization passes. I.e if
2066 // we rematerialized several intersecting chains. Or if original values
2067 // don't have any uses besides this statepoint.
2069 // For invokes we need to rematerialize each chain twice - for normal and
2070 // for unwind basic blocks. Model this by multiplying cost by two.
2071 if (CS.isInvoke()) {
2074 // If it's too expensive - skip it
2075 if (Cost >= RematerializationThreshold)
2078 // Remove value from the live set
2079 LiveValuesToBeDeleted.push_back(LiveValue);
2081 // Clone instructions and record them inside "Info" structure
2083 // Walk backwards to visit top-most instructions first
2084 std::reverse(ChainToBase.begin(), ChainToBase.end());
2086 // Utility function which clones all instructions from "ChainToBase"
2087 // and inserts them before "InsertBefore". Returns rematerialized value
2088 // which should be used after statepoint.
2089 auto rematerializeChain = [&ChainToBase](
2090 Instruction *InsertBefore, Value *RootOfChain, Value *AlternateLiveBase) {
2091 Instruction *LastClonedValue = nullptr;
2092 Instruction *LastValue = nullptr;
2093 for (Instruction *Instr: ChainToBase) {
2094 // Only GEP's and casts are supported as we need to be careful to not
2095 // introduce any new uses of pointers not in the liveset.
2096 // Note that it's fine to introduce new uses of pointers which were
2097 // otherwise not used after this statepoint.
2098 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2100 Instruction *ClonedValue = Instr->clone();
2101 ClonedValue->insertBefore(InsertBefore);
2102 ClonedValue->setName(Instr->getName() + ".remat");
2104 // If it is not first instruction in the chain then it uses previously
2105 // cloned value. We should update it to use cloned value.
2106 if (LastClonedValue) {
2108 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2110 for (auto OpValue : ClonedValue->operand_values()) {
2111 // Assert that cloned instruction does not use any instructions from
2112 // this chain other than LastClonedValue
2113 assert(!is_contained(ChainToBase, OpValue) &&
2114 "incorrect use in rematerialization chain");
2115 // Assert that the cloned instruction does not use the RootOfChain
2116 // or the AlternateLiveBase.
2117 assert(OpValue != RootOfChain && OpValue != AlternateLiveBase);
2121 // For the first instruction, replace the use of unrelocated base i.e.
2122 // RootOfChain/OrigRootPhi, with the corresponding PHI present in the
2123 // live set. They have been proved to be the same PHI nodes. Note
2124 // that the *only* use of the RootOfChain in the ChainToBase list is
2125 // the first Value in the list.
2126 if (RootOfChain != AlternateLiveBase)
2127 ClonedValue->replaceUsesOfWith(RootOfChain, AlternateLiveBase);
2130 LastClonedValue = ClonedValue;
2133 assert(LastClonedValue);
2134 return LastClonedValue;
2137 // Different cases for calls and invokes. For invokes we need to clone
2138 // instructions both on normal and unwind path.
2140 Instruction *InsertBefore = CS.getInstruction()->getNextNode();
2141 assert(InsertBefore);
2142 Instruction *RematerializedValue = rematerializeChain(
2143 InsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2144 Info.RematerializedValues[RematerializedValue] = LiveValue;
2146 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction());
2148 Instruction *NormalInsertBefore =
2149 &*Invoke->getNormalDest()->getFirstInsertionPt();
2150 Instruction *UnwindInsertBefore =
2151 &*Invoke->getUnwindDest()->getFirstInsertionPt();
2153 Instruction *NormalRematerializedValue = rematerializeChain(
2154 NormalInsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2155 Instruction *UnwindRematerializedValue = rematerializeChain(
2156 UnwindInsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2158 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2159 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2163 // Remove rematerializaed values from the live set
2164 for (auto LiveValue: LiveValuesToBeDeleted) {
2165 Info.LiveSet.remove(LiveValue);
2169 static bool insertParsePoints(Function &F, DominatorTree &DT,
2170 TargetTransformInfo &TTI,
2171 SmallVectorImpl<CallSite> &ToUpdate) {
2173 // sanity check the input
2174 std::set<CallSite> Uniqued;
2175 Uniqued.insert(ToUpdate.begin(), ToUpdate.end());
2176 assert(Uniqued.size() == ToUpdate.size() && "no duplicates please!");
2178 for (CallSite CS : ToUpdate)
2179 assert(CS.getInstruction()->getFunction() == &F);
2182 // When inserting gc.relocates for invokes, we need to be able to insert at
2183 // the top of the successor blocks. See the comment on
2184 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2185 // may restructure the CFG.
2186 for (CallSite CS : ToUpdate) {
2189 auto *II = cast<InvokeInst>(CS.getInstruction());
2190 normalizeForInvokeSafepoint(II->getNormalDest(), II->getParent(), DT);
2191 normalizeForInvokeSafepoint(II->getUnwindDest(), II->getParent(), DT);
2194 // A list of dummy calls added to the IR to keep various values obviously
2195 // live in the IR. We'll remove all of these when done.
2196 SmallVector<CallInst *, 64> Holders;
2198 // Insert a dummy call with all of the deopt operands we'll need for the
2199 // actual safepoint insertion as arguments. This ensures reference operands
2200 // in the deopt argument list are considered live through the safepoint (and
2201 // thus makes sure they get relocated.)
2202 for (CallSite CS : ToUpdate) {
2203 SmallVector<Value *, 64> DeoptValues;
2205 for (Value *Arg : GetDeoptBundleOperands(CS)) {
2206 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2207 "support for FCA unimplemented");
2208 if (isHandledGCPointerType(Arg->getType()))
2209 DeoptValues.push_back(Arg);
2212 insertUseHolderAfter(CS, DeoptValues, Holders);
2215 SmallVector<PartiallyConstructedSafepointRecord, 64> Records(ToUpdate.size());
2217 // A) Identify all gc pointers which are statically live at the given call
2219 findLiveReferences(F, DT, ToUpdate, Records);
2221 // B) Find the base pointers for each live pointer
2222 /* scope for caching */ {
2223 // Cache the 'defining value' relation used in the computation and
2224 // insertion of base phis and selects. This ensures that we don't insert
2225 // large numbers of duplicate base_phis.
2226 DefiningValueMapTy DVCache;
2228 for (size_t i = 0; i < Records.size(); i++) {
2229 PartiallyConstructedSafepointRecord &info = Records[i];
2230 findBasePointers(DT, DVCache, ToUpdate[i], info);
2232 } // end of cache scope
2234 // The base phi insertion logic (for any safepoint) may have inserted new
2235 // instructions which are now live at some safepoint. The simplest such
2238 // phi a <-- will be a new base_phi here
2239 // safepoint 1 <-- that needs to be live here
2243 // We insert some dummy calls after each safepoint to definitely hold live
2244 // the base pointers which were identified for that safepoint. We'll then
2245 // ask liveness for _every_ base inserted to see what is now live. Then we
2246 // remove the dummy calls.
2247 Holders.reserve(Holders.size() + Records.size());
2248 for (size_t i = 0; i < Records.size(); i++) {
2249 PartiallyConstructedSafepointRecord &Info = Records[i];
2251 SmallVector<Value *, 128> Bases;
2252 for (auto Pair : Info.PointerToBase)
2253 Bases.push_back(Pair.second);
2255 insertUseHolderAfter(ToUpdate[i], Bases, Holders);
2258 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2259 // need to rerun liveness. We may *also* have inserted new defs, but that's
2260 // not the key issue.
2261 recomputeLiveInValues(F, DT, ToUpdate, Records);
2263 if (PrintBasePointers) {
2264 for (auto &Info : Records) {
2265 errs() << "Base Pairs: (w/Relocation)\n";
2266 for (auto Pair : Info.PointerToBase) {
2267 errs() << " derived ";
2268 Pair.first->printAsOperand(errs(), false);
2270 Pair.second->printAsOperand(errs(), false);
2276 // It is possible that non-constant live variables have a constant base. For
2277 // example, a GEP with a variable offset from a global. In this case we can
2278 // remove it from the liveset. We already don't add constants to the liveset
2279 // because we assume they won't move at runtime and the GC doesn't need to be
2280 // informed about them. The same reasoning applies if the base is constant.
2281 // Note that the relocation placement code relies on this filtering for
2282 // correctness as it expects the base to be in the liveset, which isn't true
2283 // if the base is constant.
2284 for (auto &Info : Records)
2285 for (auto &BasePair : Info.PointerToBase)
2286 if (isa<Constant>(BasePair.second))
2287 Info.LiveSet.remove(BasePair.first);
2289 for (CallInst *CI : Holders)
2290 CI->eraseFromParent();
2294 // In order to reduce live set of statepoint we might choose to rematerialize
2295 // some values instead of relocating them. This is purely an optimization and
2296 // does not influence correctness.
2297 for (size_t i = 0; i < Records.size(); i++)
2298 rematerializeLiveValues(ToUpdate[i], Records[i], TTI);
2300 // We need this to safely RAUW and delete call or invoke return values that
2301 // may themselves be live over a statepoint. For details, please see usage in
2302 // makeStatepointExplicitImpl.
2303 std::vector<DeferredReplacement> Replacements;
2305 // Now run through and replace the existing statepoints with new ones with
2306 // the live variables listed. We do not yet update uses of the values being
2307 // relocated. We have references to live variables that need to
2308 // survive to the last iteration of this loop. (By construction, the
2309 // previous statepoint can not be a live variable, thus we can and remove
2310 // the old statepoint calls as we go.)
2311 for (size_t i = 0; i < Records.size(); i++)
2312 makeStatepointExplicit(DT, ToUpdate[i], Records[i], Replacements);
2314 ToUpdate.clear(); // prevent accident use of invalid CallSites
2316 for (auto &PR : Replacements)
2319 Replacements.clear();
2321 for (auto &Info : Records) {
2322 // These live sets may contain state Value pointers, since we replaced calls
2323 // with operand bundles with calls wrapped in gc.statepoint, and some of
2324 // those calls may have been def'ing live gc pointers. Clear these out to
2325 // avoid accidentally using them.
2327 // TODO: We should create a separate data structure that does not contain
2328 // these live sets, and migrate to using that data structure from this point
2330 Info.LiveSet.clear();
2331 Info.PointerToBase.clear();
2334 // Do all the fixups of the original live variables to their relocated selves
2335 SmallVector<Value *, 128> Live;
2336 for (size_t i = 0; i < Records.size(); i++) {
2337 PartiallyConstructedSafepointRecord &Info = Records[i];
2339 // We can't simply save the live set from the original insertion. One of
2340 // the live values might be the result of a call which needs a safepoint.
2341 // That Value* no longer exists and we need to use the new gc_result.
2342 // Thankfully, the live set is embedded in the statepoint (and updated), so
2343 // we just grab that.
2344 Statepoint Statepoint(Info.StatepointToken);
2345 Live.insert(Live.end(), Statepoint.gc_args_begin(),
2346 Statepoint.gc_args_end());
2348 // Do some basic sanity checks on our liveness results before performing
2349 // relocation. Relocation can and will turn mistakes in liveness results
2350 // into non-sensical code which is must harder to debug.
2351 // TODO: It would be nice to test consistency as well
2352 assert(DT.isReachableFromEntry(Info.StatepointToken->getParent()) &&
2353 "statepoint must be reachable or liveness is meaningless");
2354 for (Value *V : Statepoint.gc_args()) {
2355 if (!isa<Instruction>(V))
2356 // Non-instruction values trivial dominate all possible uses
2358 auto *LiveInst = cast<Instruction>(V);
2359 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2360 "unreachable values should never be live");
2361 assert(DT.dominates(LiveInst, Info.StatepointToken) &&
2362 "basic SSA liveness expectation violated by liveness analysis");
2366 unique_unsorted(Live);
2370 for (auto *Ptr : Live)
2371 assert(isHandledGCPointerType(Ptr->getType()) &&
2372 "must be a gc pointer type");
2375 relocationViaAlloca(F, DT, Live, Records);
2376 return !Records.empty();
2379 // Handles both return values and arguments for Functions and CallSites.
2380 template <typename AttrHolder>
2381 static void RemoveNonValidAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2384 if (AH.getDereferenceableBytes(Index))
2385 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2386 AH.getDereferenceableBytes(Index)));
2387 if (AH.getDereferenceableOrNullBytes(Index))
2388 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2389 AH.getDereferenceableOrNullBytes(Index)));
2390 if (AH.getAttributes().hasAttribute(Index, Attribute::NoAlias))
2391 R.addAttribute(Attribute::NoAlias);
2394 AH.setAttributes(AH.getAttributes().removeAttributes(Ctx, Index, R));
2397 static void stripNonValidAttributesFromPrototype(Function &F) {
2398 LLVMContext &Ctx = F.getContext();
2400 for (Argument &A : F.args())
2401 if (isa<PointerType>(A.getType()))
2402 RemoveNonValidAttrAtIndex(Ctx, F,
2403 A.getArgNo() + AttributeList::FirstArgIndex);
2405 if (isa<PointerType>(F.getReturnType()))
2406 RemoveNonValidAttrAtIndex(Ctx, F, AttributeList::ReturnIndex);
2409 /// Certain metadata on instructions are invalid after running RS4GC.
2410 /// Optimizations that run after RS4GC can incorrectly use this metadata to
2411 /// optimize functions. We drop such metadata on the instruction.
2412 static void stripInvalidMetadataFromInstruction(Instruction &I) {
2413 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
2415 // These are the attributes that are still valid on loads and stores after
2417 // The metadata implying dereferenceability and noalias are (conservatively)
2418 // dropped. This is because semantically, after RewriteStatepointsForGC runs,
2419 // all calls to gc.statepoint "free" the entire heap. Also, gc.statepoint can
2420 // touch the entire heap including noalias objects. Note: The reasoning is
2421 // same as stripping the dereferenceability and noalias attributes that are
2422 // analogous to the metadata counterparts.
2423 // We also drop the invariant.load metadata on the load because that metadata
2424 // implies the address operand to the load points to memory that is never
2425 // changed once it became dereferenceable. This is no longer true after RS4GC.
2426 // Similar reasoning applies to invariant.group metadata, which applies to
2427 // loads within a group.
2428 unsigned ValidMetadataAfterRS4GC[] = {LLVMContext::MD_tbaa,
2429 LLVMContext::MD_range,
2430 LLVMContext::MD_alias_scope,
2431 LLVMContext::MD_nontemporal,
2432 LLVMContext::MD_nonnull,
2433 LLVMContext::MD_align,
2434 LLVMContext::MD_type};
2436 // Drops all metadata on the instruction other than ValidMetadataAfterRS4GC.
2437 I.dropUnknownNonDebugMetadata(ValidMetadataAfterRS4GC);
2440 static void stripNonValidDataFromBody(Function &F) {
2444 LLVMContext &Ctx = F.getContext();
2445 MDBuilder Builder(Ctx);
2447 // Set of invariantstart instructions that we need to remove.
2448 // Use this to avoid invalidating the instruction iterator.
2449 SmallVector<IntrinsicInst*, 12> InvariantStartInstructions;
2451 for (Instruction &I : instructions(F)) {
2452 // invariant.start on memory location implies that the referenced memory
2453 // location is constant and unchanging. This is no longer true after
2454 // RewriteStatepointsForGC runs because there can be calls to gc.statepoint
2455 // which frees the entire heap and the presence of invariant.start allows
2456 // the optimizer to sink the load of a memory location past a statepoint,
2457 // which is incorrect.
2458 if (auto *II = dyn_cast<IntrinsicInst>(&I))
2459 if (II->getIntrinsicID() == Intrinsic::invariant_start) {
2460 InvariantStartInstructions.push_back(II);
2464 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) {
2465 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!");
2466 bool IsImmutableTBAA =
2467 MD->getNumOperands() == 4 &&
2468 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1;
2470 if (!IsImmutableTBAA)
2471 continue; // no work to do, MD_tbaa is already marked mutable
2473 MDNode *Base = cast<MDNode>(MD->getOperand(0));
2474 MDNode *Access = cast<MDNode>(MD->getOperand(1));
2476 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue();
2478 MDNode *MutableTBAA =
2479 Builder.createTBAAStructTagNode(Base, Access, Offset);
2480 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2483 stripInvalidMetadataFromInstruction(I);
2485 if (CallSite CS = CallSite(&I)) {
2486 for (int i = 0, e = CS.arg_size(); i != e; i++)
2487 if (isa<PointerType>(CS.getArgument(i)->getType()))
2488 RemoveNonValidAttrAtIndex(Ctx, CS, i + AttributeList::FirstArgIndex);
2489 if (isa<PointerType>(CS.getType()))
2490 RemoveNonValidAttrAtIndex(Ctx, CS, AttributeList::ReturnIndex);
2494 // Delete the invariant.start instructions and RAUW undef.
2495 for (auto *II : InvariantStartInstructions) {
2496 II->replaceAllUsesWith(UndefValue::get(II->getType()));
2497 II->eraseFromParent();
2501 /// Returns true if this function should be rewritten by this pass. The main
2502 /// point of this function is as an extension point for custom logic.
2503 static bool shouldRewriteStatepointsIn(Function &F) {
2504 // TODO: This should check the GCStrategy
2506 const auto &FunctionGCName = F.getGC();
2507 const StringRef StatepointExampleName("statepoint-example");
2508 const StringRef CoreCLRName("coreclr");
2509 return (StatepointExampleName == FunctionGCName) ||
2510 (CoreCLRName == FunctionGCName);
2515 static void stripNonValidData(Module &M) {
2517 assert(llvm::any_of(M, shouldRewriteStatepointsIn) && "precondition!");
2520 for (Function &F : M)
2521 stripNonValidAttributesFromPrototype(F);
2523 for (Function &F : M)
2524 stripNonValidDataFromBody(F);
2527 bool RewriteStatepointsForGC::runOnFunction(Function &F, DominatorTree &DT,
2528 TargetTransformInfo &TTI,
2529 const TargetLibraryInfo &TLI) {
2530 assert(!F.isDeclaration() && !F.empty() &&
2531 "need function body to rewrite statepoints in");
2532 assert(shouldRewriteStatepointsIn(F) && "mismatch in rewrite decision");
2534 auto NeedsRewrite = [&TLI](Instruction &I) {
2535 if (ImmutableCallSite CS = ImmutableCallSite(&I))
2536 return !callsGCLeafFunction(CS, TLI) && !isStatepoint(CS);
2540 // Gather all the statepoints which need rewritten. Be careful to only
2541 // consider those in reachable code since we need to ask dominance queries
2542 // when rewriting. We'll delete the unreachable ones in a moment.
2543 SmallVector<CallSite, 64> ParsePointNeeded;
2544 bool HasUnreachableStatepoint = false;
2545 for (Instruction &I : instructions(F)) {
2546 // TODO: only the ones with the flag set!
2547 if (NeedsRewrite(I)) {
2548 if (DT.isReachableFromEntry(I.getParent()))
2549 ParsePointNeeded.push_back(CallSite(&I));
2551 HasUnreachableStatepoint = true;
2555 bool MadeChange = false;
2557 // Delete any unreachable statepoints so that we don't have unrewritten
2558 // statepoints surviving this pass. This makes testing easier and the
2559 // resulting IR less confusing to human readers. Rather than be fancy, we
2560 // just reuse a utility function which removes the unreachable blocks.
2561 if (HasUnreachableStatepoint)
2562 MadeChange |= removeUnreachableBlocks(F);
2564 // Return early if no work to do.
2565 if (ParsePointNeeded.empty())
2568 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2569 // These are created by LCSSA. They have the effect of increasing the size
2570 // of liveness sets for no good reason. It may be harder to do this post
2571 // insertion since relocations and base phis can confuse things.
2572 for (BasicBlock &BB : F)
2573 if (BB.getUniquePredecessor()) {
2575 FoldSingleEntryPHINodes(&BB);
2578 // Before we start introducing relocations, we want to tweak the IR a bit to
2579 // avoid unfortunate code generation effects. The main example is that we
2580 // want to try to make sure the comparison feeding a branch is after any
2581 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2582 // values feeding a branch after relocation. This is semantically correct,
2583 // but results in extra register pressure since both the pre-relocation and
2584 // post-relocation copies must be available in registers. For code without
2585 // relocations this is handled elsewhere, but teaching the scheduler to
2586 // reverse the transform we're about to do would be slightly complex.
2587 // Note: This may extend the live range of the inputs to the icmp and thus
2588 // increase the liveset of any statepoint we move over. This is profitable
2589 // as long as all statepoints are in rare blocks. If we had in-register
2590 // lowering for live values this would be a much safer transform.
2591 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* {
2592 if (auto *BI = dyn_cast<BranchInst>(TI))
2593 if (BI->isConditional())
2594 return dyn_cast<Instruction>(BI->getCondition());
2595 // TODO: Extend this to handle switches
2598 for (BasicBlock &BB : F) {
2599 TerminatorInst *TI = BB.getTerminator();
2600 if (auto *Cond = getConditionInst(TI))
2601 // TODO: Handle more than just ICmps here. We should be able to move
2602 // most instructions without side effects or memory access.
2603 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2605 Cond->moveBefore(TI);
2609 MadeChange |= insertParsePoints(F, DT, TTI, ParsePointNeeded);
2613 // liveness computation via standard dataflow
2614 // -------------------------------------------------------------------
2616 // TODO: Consider using bitvectors for liveness, the set of potentially
2617 // interesting values should be small and easy to pre-compute.
2619 /// Compute the live-in set for the location rbegin starting from
2620 /// the live-out set of the basic block
2621 static void computeLiveInValues(BasicBlock::reverse_iterator Begin,
2622 BasicBlock::reverse_iterator End,
2623 SetVector<Value *> &LiveTmp) {
2624 for (auto &I : make_range(Begin, End)) {
2625 // KILL/Def - Remove this definition from LiveIn
2628 // Don't consider *uses* in PHI nodes, we handle their contribution to
2629 // predecessor blocks when we seed the LiveOut sets
2630 if (isa<PHINode>(I))
2633 // USE - Add to the LiveIn set for this instruction
2634 for (Value *V : I.operands()) {
2635 assert(!isUnhandledGCPointerType(V->getType()) &&
2636 "support for FCA unimplemented");
2637 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2638 // The choice to exclude all things constant here is slightly subtle.
2639 // There are two independent reasons:
2640 // - We assume that things which are constant (from LLVM's definition)
2641 // do not move at runtime. For example, the address of a global
2642 // variable is fixed, even though it's contents may not be.
2643 // - Second, we can't disallow arbitrary inttoptr constants even
2644 // if the language frontend does. Optimization passes are free to
2645 // locally exploit facts without respect to global reachability. This
2646 // can create sections of code which are dynamically unreachable and
2647 // contain just about anything. (see constants.ll in tests)
2654 static void computeLiveOutSeed(BasicBlock *BB, SetVector<Value *> &LiveTmp) {
2655 for (BasicBlock *Succ : successors(BB)) {
2656 for (auto &I : *Succ) {
2657 PHINode *PN = dyn_cast<PHINode>(&I);
2661 Value *V = PN->getIncomingValueForBlock(BB);
2662 assert(!isUnhandledGCPointerType(V->getType()) &&
2663 "support for FCA unimplemented");
2664 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V))
2670 static SetVector<Value *> computeKillSet(BasicBlock *BB) {
2671 SetVector<Value *> KillSet;
2672 for (Instruction &I : *BB)
2673 if (isHandledGCPointerType(I.getType()))
2679 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2680 /// sanity check for the liveness computation.
2681 static void checkBasicSSA(DominatorTree &DT, SetVector<Value *> &Live,
2682 TerminatorInst *TI, bool TermOkay = false) {
2683 for (Value *V : Live) {
2684 if (auto *I = dyn_cast<Instruction>(V)) {
2685 // The terminator can be a member of the LiveOut set. LLVM's definition
2686 // of instruction dominance states that V does not dominate itself. As
2687 // such, we need to special case this to allow it.
2688 if (TermOkay && TI == I)
2690 assert(DT.dominates(I, TI) &&
2691 "basic SSA liveness expectation violated by liveness analysis");
2696 /// Check that all the liveness sets used during the computation of liveness
2697 /// obey basic SSA properties. This is useful for finding cases where we miss
2699 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2701 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2702 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2703 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2707 static void computeLiveInValues(DominatorTree &DT, Function &F,
2708 GCPtrLivenessData &Data) {
2709 SmallSetVector<BasicBlock *, 32> Worklist;
2711 // Seed the liveness for each individual block
2712 for (BasicBlock &BB : F) {
2713 Data.KillSet[&BB] = computeKillSet(&BB);
2714 Data.LiveSet[&BB].clear();
2715 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2718 for (Value *Kill : Data.KillSet[&BB])
2719 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2722 Data.LiveOut[&BB] = SetVector<Value *>();
2723 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2724 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2725 Data.LiveIn[&BB].set_union(Data.LiveOut[&BB]);
2726 Data.LiveIn[&BB].set_subtract(Data.KillSet[&BB]);
2727 if (!Data.LiveIn[&BB].empty())
2728 Worklist.insert(pred_begin(&BB), pred_end(&BB));
2731 // Propagate that liveness until stable
2732 while (!Worklist.empty()) {
2733 BasicBlock *BB = Worklist.pop_back_val();
2735 // Compute our new liveout set, then exit early if it hasn't changed despite
2736 // the contribution of our successor.
2737 SetVector<Value *> LiveOut = Data.LiveOut[BB];
2738 const auto OldLiveOutSize = LiveOut.size();
2739 for (BasicBlock *Succ : successors(BB)) {
2740 assert(Data.LiveIn.count(Succ));
2741 LiveOut.set_union(Data.LiveIn[Succ]);
2743 // assert OutLiveOut is a subset of LiveOut
2744 if (OldLiveOutSize == LiveOut.size()) {
2745 // If the sets are the same size, then we didn't actually add anything
2746 // when unioning our successors LiveIn. Thus, the LiveIn of this block
2750 Data.LiveOut[BB] = LiveOut;
2752 // Apply the effects of this basic block
2753 SetVector<Value *> LiveTmp = LiveOut;
2754 LiveTmp.set_union(Data.LiveSet[BB]);
2755 LiveTmp.set_subtract(Data.KillSet[BB]);
2757 assert(Data.LiveIn.count(BB));
2758 const SetVector<Value *> &OldLiveIn = Data.LiveIn[BB];
2759 // assert: OldLiveIn is a subset of LiveTmp
2760 if (OldLiveIn.size() != LiveTmp.size()) {
2761 Data.LiveIn[BB] = LiveTmp;
2762 Worklist.insert(pred_begin(BB), pred_end(BB));
2764 } // while (!Worklist.empty())
2767 // Sanity check our output against SSA properties. This helps catch any
2768 // missing kills during the above iteration.
2769 for (BasicBlock &BB : F)
2770 checkBasicSSA(DT, Data, BB);
2774 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2775 StatepointLiveSetTy &Out) {
2776 BasicBlock *BB = Inst->getParent();
2778 // Note: The copy is intentional and required
2779 assert(Data.LiveOut.count(BB));
2780 SetVector<Value *> LiveOut = Data.LiveOut[BB];
2782 // We want to handle the statepoint itself oddly. It's
2783 // call result is not live (normal), nor are it's arguments
2784 // (unless they're used again later). This adjustment is
2785 // specifically what we need to relocate
2786 computeLiveInValues(BB->rbegin(), ++Inst->getIterator().getReverse(),
2788 LiveOut.remove(Inst);
2789 Out.insert(LiveOut.begin(), LiveOut.end());
2792 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2794 PartiallyConstructedSafepointRecord &Info) {
2795 Instruction *Inst = CS.getInstruction();
2796 StatepointLiveSetTy Updated;
2797 findLiveSetAtInst(Inst, RevisedLivenessData, Updated);
2799 // We may have base pointers which are now live that weren't before. We need
2800 // to update the PointerToBase structure to reflect this.
2801 for (auto V : Updated)
2802 if (Info.PointerToBase.insert({V, V}).second) {
2803 assert(isKnownBaseResult(V) &&
2804 "Can't find base for unexpected live value!");
2809 for (auto V : Updated)
2810 assert(Info.PointerToBase.count(V) &&
2811 "Must be able to find base for live value!");
2814 // Remove any stale base mappings - this can happen since our liveness is
2815 // more precise then the one inherent in the base pointer analysis.
2816 DenseSet<Value *> ToErase;
2817 for (auto KVPair : Info.PointerToBase)
2818 if (!Updated.count(KVPair.first))
2819 ToErase.insert(KVPair.first);
2821 for (auto *V : ToErase)
2822 Info.PointerToBase.erase(V);
2825 for (auto KVPair : Info.PointerToBase)
2826 assert(Updated.count(KVPair.first) && "record for non-live value");
2829 Info.LiveSet = Updated;