1 //===- RewriteStatepointsForGC.cpp - Make GC relocations explicit ---------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // Rewrite call/invoke instructions so as to make potential relocations
10 // performed by the garbage collector explicit in the IR.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Scalar/RewriteStatepointsForGC.h"
16 #include "llvm/ADT/ArrayRef.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/DenseSet.h"
19 #include "llvm/ADT/MapVector.h"
20 #include "llvm/ADT/None.h"
21 #include "llvm/ADT/Optional.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/SetVector.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/StringRef.h"
27 #include "llvm/ADT/iterator_range.h"
28 #include "llvm/Analysis/DomTreeUpdater.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/CallingConv.h"
35 #include "llvm/IR/Constant.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/DerivedTypes.h"
39 #include "llvm/IR/Dominators.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/InstIterator.h"
43 #include "llvm/IR/InstrTypes.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Intrinsics.h"
48 #include "llvm/IR/LLVMContext.h"
49 #include "llvm/IR/MDBuilder.h"
50 #include "llvm/IR/Metadata.h"
51 #include "llvm/IR/Module.h"
52 #include "llvm/IR/Statepoint.h"
53 #include "llvm/IR/Type.h"
54 #include "llvm/IR/User.h"
55 #include "llvm/IR/Value.h"
56 #include "llvm/IR/ValueHandle.h"
57 #include "llvm/InitializePasses.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 for (Function &F : M) {
177 // Nothing to do for declarations.
178 if (F.isDeclaration() || F.empty())
181 // Policy choice says not to rewrite - the most common reason is that
182 // we're compiling code without a GCStrategy.
183 if (!shouldRewriteStatepointsIn(F))
186 TargetTransformInfo &TTI =
187 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
188 const TargetLibraryInfo &TLI =
189 getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(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 GCStatepointInst *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(const CallBase *Call) {
289 Optional<OperandBundleUse> DeoptBundle =
290 Call->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->elements(), 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.
372 static void analyzeParsePointLiveness(
373 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData, CallBase *Call,
374 PartiallyConstructedSafepointRecord &Result) {
375 StatepointLiveSetTy LiveSet;
376 findLiveSetAtInst(Call, OriginalLivenessData, LiveSet);
379 dbgs() << "Live Variables:\n";
380 for (Value *V : LiveSet)
381 dbgs() << " " << V->getName() << " " << *V << "\n";
383 if (PrintLiveSetSize) {
384 dbgs() << "Safepoint For: " << Call->getCalledOperand()->getName() << "\n";
385 dbgs() << "Number live values: " << LiveSet.size() << "\n";
387 Result.LiveSet = LiveSet;
390 // Returns true is V is a knownBaseResult.
391 static bool isKnownBaseResult(Value *V);
393 // Returns true if V is a BaseResult that already exists in the IR, i.e. it is
394 // not created by the findBasePointers algorithm.
395 static bool isOriginalBaseResult(Value *V);
399 /// A single base defining value - An immediate base defining value for an
400 /// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'.
401 /// For instructions which have multiple pointer [vector] inputs or that
402 /// transition between vector and scalar types, there is no immediate base
403 /// defining value. The 'base defining value' for 'Def' is the transitive
404 /// closure of this relation stopping at the first instruction which has no
405 /// immediate base defining value. The b.d.v. might itself be a base pointer,
406 /// but it can also be an arbitrary derived pointer.
407 struct BaseDefiningValueResult {
408 /// Contains the value which is the base defining value.
411 /// True if the base defining value is also known to be an actual base
413 const bool IsKnownBase;
415 BaseDefiningValueResult(Value *BDV, bool IsKnownBase)
416 : BDV(BDV), IsKnownBase(IsKnownBase) {
418 // Check consistency between new and old means of checking whether a BDV is
420 bool MustBeBase = isKnownBaseResult(BDV);
421 assert(!MustBeBase || MustBeBase == IsKnownBase);
426 } // end anonymous namespace
428 static BaseDefiningValueResult findBaseDefiningValue(Value *I);
430 /// Return a base defining value for the 'Index' element of the given vector
431 /// instruction 'I'. If Index is null, returns a BDV for the entire vector
432 /// 'I'. As an optimization, this method will try to determine when the
433 /// element is known to already be a base pointer. If this can be established,
434 /// the second value in the returned pair will be true. Note that either a
435 /// vector or a pointer typed value can be returned. For the former, the
436 /// vector returned is a BDV (and possibly a base) of the entire vector 'I'.
437 /// If the later, the return pointer is a BDV (or possibly a base) for the
438 /// particular element in 'I'.
439 static BaseDefiningValueResult
440 findBaseDefiningValueOfVector(Value *I) {
441 // Each case parallels findBaseDefiningValue below, see that code for
442 // detailed motivation.
444 if (isa<Argument>(I))
445 // An incoming argument to the function is a base pointer
446 return BaseDefiningValueResult(I, true);
448 if (isa<Constant>(I))
449 // Base of constant vector consists only of constant null pointers.
450 // For reasoning see similar case inside 'findBaseDefiningValue' function.
451 return BaseDefiningValueResult(ConstantAggregateZero::get(I->getType()),
454 if (isa<LoadInst>(I))
455 return BaseDefiningValueResult(I, true);
457 if (isa<InsertElementInst>(I))
458 // We don't know whether this vector contains entirely base pointers or
459 // not. To be conservatively correct, we treat it as a BDV and will
460 // duplicate code as needed to construct a parallel vector of bases.
461 return BaseDefiningValueResult(I, false);
463 if (isa<ShuffleVectorInst>(I))
464 // We don't know whether this vector contains entirely base pointers or
465 // not. To be conservatively correct, we treat it as a BDV and will
466 // duplicate code as needed to construct a parallel vector of bases.
467 // TODO: There a number of local optimizations which could be applied here
468 // for particular sufflevector patterns.
469 return BaseDefiningValueResult(I, false);
471 // The behavior of getelementptr instructions is the same for vector and
472 // non-vector data types.
473 if (auto *GEP = dyn_cast<GetElementPtrInst>(I))
474 return findBaseDefiningValue(GEP->getPointerOperand());
476 // If the pointer comes through a bitcast of a vector of pointers to
477 // a vector of another type of pointer, then look through the bitcast
478 if (auto *BC = dyn_cast<BitCastInst>(I))
479 return findBaseDefiningValue(BC->getOperand(0));
481 // We assume that functions in the source language only return base
482 // pointers. This should probably be generalized via attributes to support
483 // both source language and internal functions.
484 if (isa<CallInst>(I) || isa<InvokeInst>(I))
485 return BaseDefiningValueResult(I, true);
487 // A PHI or Select is a base defining value. The outer findBasePointer
488 // algorithm is responsible for constructing a base value for this BDV.
489 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
490 "unknown vector instruction - no base found for vector element");
491 return BaseDefiningValueResult(I, false);
494 /// Helper function for findBasePointer - Will return a value which either a)
495 /// defines the base pointer for the input, b) blocks the simple search
496 /// (i.e. a PHI or Select of two derived pointers), or c) involves a change
497 /// from pointer to vector type or back.
498 static BaseDefiningValueResult findBaseDefiningValue(Value *I) {
499 assert(I->getType()->isPtrOrPtrVectorTy() &&
500 "Illegal to ask for the base pointer of a non-pointer type");
502 if (I->getType()->isVectorTy())
503 return findBaseDefiningValueOfVector(I);
505 if (isa<Argument>(I))
506 // An incoming argument to the function is a base pointer
507 // We should have never reached here if this argument isn't an gc value
508 return BaseDefiningValueResult(I, true);
510 if (isa<Constant>(I)) {
511 // We assume that objects with a constant base (e.g. a global) can't move
512 // and don't need to be reported to the collector because they are always
513 // live. Besides global references, all kinds of constants (e.g. undef,
514 // constant expressions, null pointers) can be introduced by the inliner or
515 // the optimizer, especially on dynamically dead paths.
516 // Here we treat all of them as having single null base. By doing this we
517 // trying to avoid problems reporting various conflicts in a form of
518 // "phi (const1, const2)" or "phi (const, regular gc ptr)".
519 // See constant.ll file for relevant test cases.
521 return BaseDefiningValueResult(
522 ConstantPointerNull::get(cast<PointerType>(I->getType())), true);
525 if (CastInst *CI = dyn_cast<CastInst>(I)) {
526 Value *Def = CI->stripPointerCasts();
527 // If stripping pointer casts changes the address space there is an
528 // addrspacecast in between.
529 assert(cast<PointerType>(Def->getType())->getAddressSpace() ==
530 cast<PointerType>(CI->getType())->getAddressSpace() &&
531 "unsupported addrspacecast");
532 // If we find a cast instruction here, it means we've found a cast which is
533 // not simply a pointer cast (i.e. an inttoptr). We don't know how to
534 // handle int->ptr conversion.
535 assert(!isa<CastInst>(Def) && "shouldn't find another cast here");
536 return findBaseDefiningValue(Def);
539 if (isa<LoadInst>(I))
540 // The value loaded is an gc base itself
541 return BaseDefiningValueResult(I, true);
543 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
544 // The base of this GEP is the base
545 return findBaseDefiningValue(GEP->getPointerOperand());
547 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
548 switch (II->getIntrinsicID()) {
550 // fall through to general call handling
552 case Intrinsic::experimental_gc_statepoint:
553 llvm_unreachable("statepoints don't produce pointers");
554 case Intrinsic::experimental_gc_relocate:
555 // Rerunning safepoint insertion after safepoints are already
556 // inserted is not supported. It could probably be made to work,
557 // but why are you doing this? There's no good reason.
558 llvm_unreachable("repeat safepoint insertion is not supported");
559 case Intrinsic::gcroot:
560 // Currently, this mechanism hasn't been extended to work with gcroot.
561 // There's no reason it couldn't be, but I haven't thought about the
562 // implications much.
564 "interaction with the gcroot mechanism is not supported");
567 // We assume that functions in the source language only return base
568 // pointers. This should probably be generalized via attributes to support
569 // both source language and internal functions.
570 if (isa<CallInst>(I) || isa<InvokeInst>(I))
571 return BaseDefiningValueResult(I, true);
573 // TODO: I have absolutely no idea how to implement this part yet. It's not
574 // necessarily hard, I just haven't really looked at it yet.
575 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented");
577 if (isa<AtomicCmpXchgInst>(I))
578 // A CAS is effectively a atomic store and load combined under a
579 // predicate. From the perspective of base pointers, we just treat it
581 return BaseDefiningValueResult(I, true);
583 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are "
584 "binary ops which don't apply to pointers");
586 // The aggregate ops. Aggregates can either be in the heap or on the
587 // stack, but in either case, this is simply a field load. As a result,
588 // this is a defining definition of the base just like a load is.
589 if (isa<ExtractValueInst>(I))
590 return BaseDefiningValueResult(I, true);
592 // We should never see an insert vector since that would require we be
593 // tracing back a struct value not a pointer value.
594 assert(!isa<InsertValueInst>(I) &&
595 "Base pointer for a struct is meaningless");
597 // An extractelement produces a base result exactly when it's input does.
598 // We may need to insert a parallel instruction to extract the appropriate
599 // element out of the base vector corresponding to the input. Given this,
600 // it's analogous to the phi and select case even though it's not a merge.
601 if (isa<ExtractElementInst>(I))
602 // Note: There a lot of obvious peephole cases here. This are deliberately
603 // handled after the main base pointer inference algorithm to make writing
604 // test cases to exercise that code easier.
605 return BaseDefiningValueResult(I, false);
607 // The last two cases here don't return a base pointer. Instead, they
608 // return a value which dynamically selects from among several base
609 // derived pointers (each with it's own base potentially). It's the job of
610 // the caller to resolve these.
611 assert((isa<SelectInst>(I) || isa<PHINode>(I)) &&
612 "missing instruction case in findBaseDefiningValing");
613 return BaseDefiningValueResult(I, false);
616 /// Returns the base defining value for this value.
617 static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) {
618 Value *&Cached = Cache[I];
620 Cached = findBaseDefiningValue(I).BDV;
621 LLVM_DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> "
622 << Cached->getName() << "\n");
624 assert(Cache[I] != nullptr);
628 /// Return a base pointer for this value if known. Otherwise, return it's
629 /// base defining value.
630 static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) {
631 Value *Def = findBaseDefiningValueCached(I, Cache);
632 auto Found = Cache.find(Def);
633 if (Found != Cache.end()) {
634 // Either a base-of relation, or a self reference. Caller must check.
635 return Found->second;
637 // Only a BDV available
641 /// This value is a base pointer that is not generated by RS4GC, i.e. it already
642 /// exists in the code.
643 static bool isOriginalBaseResult(Value *V) {
644 // no recursion possible
645 return !isa<PHINode>(V) && !isa<SelectInst>(V) &&
646 !isa<ExtractElementInst>(V) && !isa<InsertElementInst>(V) &&
647 !isa<ShuffleVectorInst>(V);
650 /// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV,
651 /// is it known to be a base pointer? Or do we need to continue searching.
652 static bool isKnownBaseResult(Value *V) {
653 if (isOriginalBaseResult(V))
655 if (isa<Instruction>(V) &&
656 cast<Instruction>(V)->getMetadata("is_base_value")) {
657 // This is a previously inserted base phi or select. We know
658 // that this is a base value.
662 // We need to keep searching
666 // Returns true if First and Second values are both scalar or both vector.
667 static bool areBothVectorOrScalar(Value *First, Value *Second) {
668 return isa<VectorType>(First->getType()) ==
669 isa<VectorType>(Second->getType());
674 /// Models the state of a single base defining value in the findBasePointer
675 /// algorithm for determining where a new instruction is needed to propagate
676 /// the base of this BDV.
679 enum Status { Unknown, Base, Conflict };
681 BDVState() : BaseValue(nullptr) {}
683 explicit BDVState(Status Status, Value *BaseValue = nullptr)
684 : Status(Status), BaseValue(BaseValue) {
685 assert(Status != Base || BaseValue);
688 explicit BDVState(Value *BaseValue) : Status(Base), BaseValue(BaseValue) {}
690 Status getStatus() const { return Status; }
691 Value *getBaseValue() const { return BaseValue; }
693 bool isBase() const { return getStatus() == Base; }
694 bool isUnknown() const { return getStatus() == Unknown; }
695 bool isConflict() const { return getStatus() == Conflict; }
697 bool operator==(const BDVState &Other) const {
698 return BaseValue == Other.BaseValue && Status == Other.Status;
701 bool operator!=(const BDVState &other) const { return !(*this == other); }
709 void print(raw_ostream &OS) const {
710 switch (getStatus()) {
721 OS << " (" << getBaseValue() << " - "
722 << (getBaseValue() ? getBaseValue()->getName() : "nullptr") << "): ";
726 Status Status = Unknown;
727 AssertingVH<Value> BaseValue; // Non-null only if Status == Base.
730 } // end anonymous namespace
733 static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) {
739 static BDVState meetBDVStateImpl(const BDVState &LHS, const BDVState &RHS) {
740 switch (LHS.getStatus()) {
741 case BDVState::Unknown:
745 assert(LHS.getBaseValue() && "can't be null");
750 if (LHS.getBaseValue() == RHS.getBaseValue()) {
751 assert(LHS == RHS && "equality broken!");
754 return BDVState(BDVState::Conflict);
756 assert(RHS.isConflict() && "only three states!");
757 return BDVState(BDVState::Conflict);
759 case BDVState::Conflict:
762 llvm_unreachable("only three states!");
765 // Values of type BDVState form a lattice, and this function implements the meet
767 static BDVState meetBDVState(const BDVState &LHS, const BDVState &RHS) {
768 BDVState Result = meetBDVStateImpl(LHS, RHS);
769 assert(Result == meetBDVStateImpl(RHS, LHS) &&
770 "Math is wrong: meet does not commute!");
774 /// For a given value or instruction, figure out what base ptr its derived from.
775 /// For gc objects, this is simply itself. On success, returns a value which is
776 /// the base pointer. (This is reliable and can be used for relocation.) On
777 /// failure, returns nullptr.
778 static Value *findBasePointer(Value *I, DefiningValueMapTy &Cache) {
779 Value *Def = findBaseOrBDV(I, Cache);
781 if (isKnownBaseResult(Def) && areBothVectorOrScalar(Def, I))
784 // Here's the rough algorithm:
785 // - For every SSA value, construct a mapping to either an actual base
786 // pointer or a PHI which obscures the base pointer.
787 // - Construct a mapping from PHI to unknown TOP state. Use an
788 // optimistic algorithm to propagate base pointer information. Lattice
793 // When algorithm terminates, all PHIs will either have a single concrete
794 // base or be in a conflict state.
795 // - For every conflict, insert a dummy PHI node without arguments. Add
796 // these to the base[Instruction] = BasePtr mapping. For every
797 // non-conflict, add the actual base.
798 // - For every conflict, add arguments for the base[a] of each input
801 // Note: A simpler form of this would be to add the conflict form of all
802 // PHIs without running the optimistic algorithm. This would be
803 // analogous to pessimistic data flow and would likely lead to an
804 // overall worse solution.
807 auto isExpectedBDVType = [](Value *BDV) {
808 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) ||
809 isa<ExtractElementInst>(BDV) || isa<InsertElementInst>(BDV) ||
810 isa<ShuffleVectorInst>(BDV);
814 // Once populated, will contain a mapping from each potentially non-base BDV
815 // to a lattice value (described above) which corresponds to that BDV.
816 // We use the order of insertion (DFS over the def/use graph) to provide a
817 // stable deterministic ordering for visiting DenseMaps (which are unordered)
818 // below. This is important for deterministic compilation.
819 MapVector<Value *, BDVState> States;
821 // Recursively fill in all base defining values reachable from the initial
822 // one for which we don't already know a definite base value for
824 SmallVector<Value*, 16> Worklist;
825 Worklist.push_back(Def);
826 States.insert({Def, BDVState()});
827 while (!Worklist.empty()) {
828 Value *Current = Worklist.pop_back_val();
829 assert(!isOriginalBaseResult(Current) && "why did it get added?");
831 auto visitIncomingValue = [&](Value *InVal) {
832 Value *Base = findBaseOrBDV(InVal, Cache);
833 if (isKnownBaseResult(Base) && areBothVectorOrScalar(Base, InVal))
834 // Known bases won't need new instructions introduced and can be
835 // ignored safely. However, this can only be done when InVal and Base
836 // are both scalar or both vector. Otherwise, we need to find a
837 // correct BDV for InVal, by creating an entry in the lattice
840 assert(isExpectedBDVType(Base) && "the only non-base values "
841 "we see should be base defining values");
842 if (States.insert(std::make_pair(Base, BDVState())).second)
843 Worklist.push_back(Base);
845 if (PHINode *PN = dyn_cast<PHINode>(Current)) {
846 for (Value *InVal : PN->incoming_values())
847 visitIncomingValue(InVal);
848 } else if (SelectInst *SI = dyn_cast<SelectInst>(Current)) {
849 visitIncomingValue(SI->getTrueValue());
850 visitIncomingValue(SI->getFalseValue());
851 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
852 visitIncomingValue(EE->getVectorOperand());
853 } else if (auto *IE = dyn_cast<InsertElementInst>(Current)) {
854 visitIncomingValue(IE->getOperand(0)); // vector operand
855 visitIncomingValue(IE->getOperand(1)); // scalar operand
856 } else if (auto *SV = dyn_cast<ShuffleVectorInst>(Current)) {
857 visitIncomingValue(SV->getOperand(0));
858 visitIncomingValue(SV->getOperand(1));
861 llvm_unreachable("Unimplemented instruction case");
867 LLVM_DEBUG(dbgs() << "States after initialization:\n");
868 for (auto Pair : States) {
869 LLVM_DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
873 // Return a phi state for a base defining value. We'll generate a new
874 // base state for known bases and expect to find a cached state otherwise.
875 auto GetStateForBDV = [&](Value *BaseValue, Value *Input) {
876 if (isKnownBaseResult(BaseValue) && areBothVectorOrScalar(BaseValue, Input))
877 return BDVState(BaseValue);
878 auto I = States.find(BaseValue);
879 assert(I != States.end() && "lookup failed!");
883 bool Progress = true;
886 const size_t OldSize = States.size();
889 // We're only changing values in this loop, thus safe to keep iterators.
890 // Since this is computing a fixed point, the order of visit does not
891 // effect the result. TODO: We could use a worklist here and make this run
893 for (auto Pair : States) {
894 Value *BDV = Pair.first;
895 // Only values that do not have known bases or those that have differing
896 // type (scalar versus vector) from a possible known base should be in the
898 assert((!isKnownBaseResult(BDV) ||
899 !areBothVectorOrScalar(BDV, Pair.second.getBaseValue())) &&
900 "why did it get added?");
902 // Given an input value for the current instruction, return a BDVState
903 // instance which represents the BDV of that value.
904 auto getStateForInput = [&](Value *V) mutable {
905 Value *BDV = findBaseOrBDV(V, Cache);
906 return GetStateForBDV(BDV, V);
910 if (SelectInst *SI = dyn_cast<SelectInst>(BDV)) {
911 NewState = meetBDVState(NewState, getStateForInput(SI->getTrueValue()));
913 meetBDVState(NewState, getStateForInput(SI->getFalseValue()));
914 } else if (PHINode *PN = dyn_cast<PHINode>(BDV)) {
915 for (Value *Val : PN->incoming_values())
916 NewState = meetBDVState(NewState, getStateForInput(Val));
917 } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) {
918 // The 'meet' for an extractelement is slightly trivial, but it's still
919 // useful in that it drives us to conflict if our input is.
921 meetBDVState(NewState, getStateForInput(EE->getVectorOperand()));
922 } else if (auto *IE = dyn_cast<InsertElementInst>(BDV)){
923 // Given there's a inherent type mismatch between the operands, will
924 // *always* produce Conflict.
925 NewState = meetBDVState(NewState, getStateForInput(IE->getOperand(0)));
926 NewState = meetBDVState(NewState, getStateForInput(IE->getOperand(1)));
928 // The only instance this does not return a Conflict is when both the
929 // vector operands are the same vector.
930 auto *SV = cast<ShuffleVectorInst>(BDV);
931 NewState = meetBDVState(NewState, getStateForInput(SV->getOperand(0)));
932 NewState = meetBDVState(NewState, getStateForInput(SV->getOperand(1)));
935 BDVState OldState = States[BDV];
936 if (OldState != NewState) {
938 States[BDV] = NewState;
942 assert(OldSize == States.size() &&
943 "fixed point shouldn't be adding any new nodes to state");
947 LLVM_DEBUG(dbgs() << "States after meet iteration:\n");
948 for (auto Pair : States) {
949 LLVM_DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
953 // Handle all instructions that have a vector BDV, but the instruction itself
954 // is of scalar type.
955 for (auto Pair : States) {
956 Instruction *I = cast<Instruction>(Pair.first);
957 BDVState State = Pair.second;
958 auto *BaseValue = State.getBaseValue();
959 // Only values that do not have known bases or those that have differing
960 // type (scalar versus vector) from a possible known base should be in the
962 assert((!isKnownBaseResult(I) || !areBothVectorOrScalar(I, BaseValue)) &&
963 "why did it get added?");
964 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
966 if (!State.isBase() || !isa<VectorType>(BaseValue->getType()))
968 // extractelement instructions are a bit special in that we may need to
969 // insert an extract even when we know an exact base for the instruction.
970 // The problem is that we need to convert from a vector base to a scalar
971 // base for the particular indice we're interested in.
972 if (isa<ExtractElementInst>(I)) {
973 auto *EE = cast<ExtractElementInst>(I);
974 // TODO: In many cases, the new instruction is just EE itself. We should
975 // exploit this, but can't do it here since it would break the invariant
976 // about the BDV not being known to be a base.
977 auto *BaseInst = ExtractElementInst::Create(
978 State.getBaseValue(), EE->getIndexOperand(), "base_ee", EE);
979 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
980 States[I] = BDVState(BDVState::Base, BaseInst);
981 } else if (!isa<VectorType>(I->getType())) {
982 // We need to handle cases that have a vector base but the instruction is
983 // a scalar type (these could be phis or selects or any instruction that
984 // are of scalar type, but the base can be a vector type). We
985 // conservatively set this as conflict. Setting the base value for these
986 // conflicts is handled in the next loop which traverses States.
987 States[I] = BDVState(BDVState::Conflict);
991 // Insert Phis for all conflicts
992 // TODO: adjust naming patterns to avoid this order of iteration dependency
993 for (auto Pair : States) {
994 Instruction *I = cast<Instruction>(Pair.first);
995 BDVState State = Pair.second;
996 // Only values that do not have known bases or those that have differing
997 // type (scalar versus vector) from a possible known base should be in the
999 assert((!isKnownBaseResult(I) || !areBothVectorOrScalar(I, State.getBaseValue())) &&
1000 "why did it get added?");
1001 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
1003 // Since we're joining a vector and scalar base, they can never be the
1004 // same. As a result, we should always see insert element having reached
1005 // the conflict state.
1006 assert(!isa<InsertElementInst>(I) || State.isConflict());
1008 if (!State.isConflict())
1011 /// Create and insert a new instruction which will represent the base of
1012 /// the given instruction 'I'.
1013 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* {
1014 if (isa<PHINode>(I)) {
1015 BasicBlock *BB = I->getParent();
1016 int NumPreds = pred_size(BB);
1017 assert(NumPreds > 0 && "how did we reach here");
1018 std::string Name = suffixed_name_or(I, ".base", "base_phi");
1019 return PHINode::Create(I->getType(), NumPreds, Name, I);
1020 } else if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
1021 // The undef will be replaced later
1022 UndefValue *Undef = UndefValue::get(SI->getType());
1023 std::string Name = suffixed_name_or(I, ".base", "base_select");
1024 return SelectInst::Create(SI->getCondition(), Undef, Undef, Name, SI);
1025 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) {
1026 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType());
1027 std::string Name = suffixed_name_or(I, ".base", "base_ee");
1028 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name,
1030 } else if (auto *IE = dyn_cast<InsertElementInst>(I)) {
1031 UndefValue *VecUndef = UndefValue::get(IE->getOperand(0)->getType());
1032 UndefValue *ScalarUndef = UndefValue::get(IE->getOperand(1)->getType());
1033 std::string Name = suffixed_name_or(I, ".base", "base_ie");
1034 return InsertElementInst::Create(VecUndef, ScalarUndef,
1035 IE->getOperand(2), Name, IE);
1037 auto *SV = cast<ShuffleVectorInst>(I);
1038 UndefValue *VecUndef = UndefValue::get(SV->getOperand(0)->getType());
1039 std::string Name = suffixed_name_or(I, ".base", "base_sv");
1040 return new ShuffleVectorInst(VecUndef, VecUndef, SV->getShuffleMask(),
1044 Instruction *BaseInst = MakeBaseInstPlaceholder(I);
1045 // Add metadata marking this as a base value
1046 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {}));
1047 States[I] = BDVState(BDVState::Conflict, BaseInst);
1050 // Returns a instruction which produces the base pointer for a given
1051 // instruction. The instruction is assumed to be an input to one of the BDVs
1052 // seen in the inference algorithm above. As such, we must either already
1053 // know it's base defining value is a base, or have inserted a new
1054 // instruction to propagate the base of it's BDV and have entered that newly
1055 // introduced instruction into the state table. In either case, we are
1056 // assured to be able to determine an instruction which produces it's base
1058 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
1059 Value *BDV = findBaseOrBDV(Input, Cache);
1060 Value *Base = nullptr;
1061 if (isKnownBaseResult(BDV) && areBothVectorOrScalar(BDV, Input)) {
1064 // Either conflict or base.
1065 assert(States.count(BDV));
1066 Base = States[BDV].getBaseValue();
1068 assert(Base && "Can't be null");
1069 // The cast is needed since base traversal may strip away bitcasts
1070 if (Base->getType() != Input->getType() && InsertPt)
1071 Base = new BitCastInst(Base, Input->getType(), "cast", InsertPt);
1075 // Fixup all the inputs of the new PHIs. Visit order needs to be
1076 // deterministic and predictable because we're naming newly created
1078 for (auto Pair : States) {
1079 Instruction *BDV = cast<Instruction>(Pair.first);
1080 BDVState State = Pair.second;
1082 // Only values that do not have known bases or those that have differing
1083 // type (scalar versus vector) from a possible known base should be in the
1085 assert((!isKnownBaseResult(BDV) ||
1086 !areBothVectorOrScalar(BDV, State.getBaseValue())) &&
1087 "why did it get added?");
1088 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
1089 if (!State.isConflict())
1092 if (PHINode *BasePHI = dyn_cast<PHINode>(State.getBaseValue())) {
1093 PHINode *PN = cast<PHINode>(BDV);
1094 unsigned NumPHIValues = PN->getNumIncomingValues();
1095 for (unsigned i = 0; i < NumPHIValues; i++) {
1096 Value *InVal = PN->getIncomingValue(i);
1097 BasicBlock *InBB = PN->getIncomingBlock(i);
1099 // If we've already seen InBB, add the same incoming value
1100 // we added for it earlier. The IR verifier requires phi
1101 // nodes with multiple entries from the same basic block
1102 // to have the same incoming value for each of those
1103 // entries. If we don't do this check here and basephi
1104 // has a different type than base, we'll end up adding two
1105 // bitcasts (and hence two distinct values) as incoming
1106 // values for the same basic block.
1108 int BlockIndex = BasePHI->getBasicBlockIndex(InBB);
1109 if (BlockIndex != -1) {
1110 Value *OldBase = BasePHI->getIncomingValue(BlockIndex);
1111 BasePHI->addIncoming(OldBase, InBB);
1114 Value *Base = getBaseForInput(InVal, nullptr);
1115 // In essence this assert states: the only way two values
1116 // incoming from the same basic block may be different is by
1117 // being different bitcasts of the same value. A cleanup
1118 // that remains TODO is changing findBaseOrBDV to return an
1119 // llvm::Value of the correct type (and still remain pure).
1120 // This will remove the need to add bitcasts.
1121 assert(Base->stripPointerCasts() == OldBase->stripPointerCasts() &&
1122 "Sanity -- findBaseOrBDV should be pure!");
1127 // Find the instruction which produces the base for each input. We may
1128 // need to insert a bitcast in the incoming block.
1129 // TODO: Need to split critical edges if insertion is needed
1130 Value *Base = getBaseForInput(InVal, InBB->getTerminator());
1131 BasePHI->addIncoming(Base, InBB);
1133 assert(BasePHI->getNumIncomingValues() == NumPHIValues);
1134 } else if (SelectInst *BaseSI =
1135 dyn_cast<SelectInst>(State.getBaseValue())) {
1136 SelectInst *SI = cast<SelectInst>(BDV);
1138 // Find the instruction which produces the base for each input.
1139 // We may need to insert a bitcast.
1140 BaseSI->setTrueValue(getBaseForInput(SI->getTrueValue(), BaseSI));
1141 BaseSI->setFalseValue(getBaseForInput(SI->getFalseValue(), BaseSI));
1142 } else if (auto *BaseEE =
1143 dyn_cast<ExtractElementInst>(State.getBaseValue())) {
1144 Value *InVal = cast<ExtractElementInst>(BDV)->getVectorOperand();
1145 // Find the instruction which produces the base for each input. We may
1146 // need to insert a bitcast.
1147 BaseEE->setOperand(0, getBaseForInput(InVal, BaseEE));
1148 } else if (auto *BaseIE = dyn_cast<InsertElementInst>(State.getBaseValue())){
1149 auto *BdvIE = cast<InsertElementInst>(BDV);
1150 auto UpdateOperand = [&](int OperandIdx) {
1151 Value *InVal = BdvIE->getOperand(OperandIdx);
1152 Value *Base = getBaseForInput(InVal, BaseIE);
1153 BaseIE->setOperand(OperandIdx, Base);
1155 UpdateOperand(0); // vector operand
1156 UpdateOperand(1); // scalar operand
1158 auto *BaseSV = cast<ShuffleVectorInst>(State.getBaseValue());
1159 auto *BdvSV = cast<ShuffleVectorInst>(BDV);
1160 auto UpdateOperand = [&](int OperandIdx) {
1161 Value *InVal = BdvSV->getOperand(OperandIdx);
1162 Value *Base = getBaseForInput(InVal, BaseSV);
1163 BaseSV->setOperand(OperandIdx, Base);
1165 UpdateOperand(0); // vector operand
1166 UpdateOperand(1); // vector operand
1170 // Cache all of our results so we can cheaply reuse them
1171 // NOTE: This is actually two caches: one of the base defining value
1172 // relation and one of the base pointer relation! FIXME
1173 for (auto Pair : States) {
1174 auto *BDV = Pair.first;
1175 Value *Base = Pair.second.getBaseValue();
1176 assert(BDV && Base);
1177 // Only values that do not have known bases or those that have differing
1178 // type (scalar versus vector) from a possible known base should be in the
1180 assert((!isKnownBaseResult(BDV) || !areBothVectorOrScalar(BDV, Base)) &&
1181 "why did it get added?");
1184 dbgs() << "Updating base value cache"
1185 << " for: " << BDV->getName() << " from: "
1186 << (Cache.count(BDV) ? Cache[BDV]->getName().str() : "none")
1187 << " to: " << Base->getName() << "\n");
1189 if (Cache.count(BDV)) {
1190 assert(isKnownBaseResult(Base) &&
1191 "must be something we 'know' is a base pointer");
1192 // Once we transition from the BDV relation being store in the Cache to
1193 // the base relation being stored, it must be stable
1194 assert((!isKnownBaseResult(Cache[BDV]) || Cache[BDV] == Base) &&
1195 "base relation should be stable");
1199 assert(Cache.count(Def));
1203 // For a set of live pointers (base and/or derived), identify the base
1204 // pointer of the object which they are derived from. This routine will
1205 // mutate the IR graph as needed to make the 'base' pointer live at the
1206 // definition site of 'derived'. This ensures that any use of 'derived' can
1207 // also use 'base'. This may involve the insertion of a number of
1208 // additional PHI nodes.
1210 // preconditions: live is a set of pointer type Values
1212 // side effects: may insert PHI nodes into the existing CFG, will preserve
1213 // CFG, will not remove or mutate any existing nodes
1215 // post condition: PointerToBase contains one (derived, base) pair for every
1216 // pointer in live. Note that derived can be equal to base if the original
1217 // pointer was a base pointer.
1219 findBasePointers(const StatepointLiveSetTy &live,
1220 MapVector<Value *, Value *> &PointerToBase,
1221 DominatorTree *DT, DefiningValueMapTy &DVCache) {
1222 for (Value *ptr : live) {
1223 Value *base = findBasePointer(ptr, DVCache);
1224 assert(base && "failed to find base pointer");
1225 PointerToBase[ptr] = base;
1226 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) ||
1227 DT->dominates(cast<Instruction>(base)->getParent(),
1228 cast<Instruction>(ptr)->getParent())) &&
1229 "The base we found better dominate the derived pointer");
1233 /// Find the required based pointers (and adjust the live set) for the given
1235 static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache,
1237 PartiallyConstructedSafepointRecord &result) {
1238 MapVector<Value *, Value *> PointerToBase;
1239 findBasePointers(result.LiveSet, PointerToBase, &DT, DVCache);
1241 if (PrintBasePointers) {
1242 errs() << "Base Pairs (w/o Relocation):\n";
1243 for (auto &Pair : PointerToBase) {
1244 errs() << " derived ";
1245 Pair.first->printAsOperand(errs(), false);
1247 Pair.second->printAsOperand(errs(), false);
1252 result.PointerToBase = PointerToBase;
1255 /// Given an updated version of the dataflow liveness results, update the
1256 /// liveset and base pointer maps for the call site CS.
1257 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
1259 PartiallyConstructedSafepointRecord &result);
1261 static void recomputeLiveInValues(
1262 Function &F, DominatorTree &DT, ArrayRef<CallBase *> toUpdate,
1263 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1264 // TODO-PERF: reuse the original liveness, then simply run the dataflow
1265 // again. The old values are still live and will help it stabilize quickly.
1266 GCPtrLivenessData RevisedLivenessData;
1267 computeLiveInValues(DT, F, RevisedLivenessData);
1268 for (size_t i = 0; i < records.size(); i++) {
1269 struct PartiallyConstructedSafepointRecord &info = records[i];
1270 recomputeLiveInValues(RevisedLivenessData, toUpdate[i], info);
1274 // When inserting gc.relocate and gc.result calls, we need to ensure there are
1275 // no uses of the original value / return value between the gc.statepoint and
1276 // the gc.relocate / gc.result call. One case which can arise is a phi node
1277 // starting one of the successor blocks. We also need to be able to insert the
1278 // gc.relocates only on the path which goes through the statepoint. We might
1279 // need to split an edge to make this possible.
1281 normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent,
1282 DominatorTree &DT) {
1283 BasicBlock *Ret = BB;
1284 if (!BB->getUniquePredecessor())
1285 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT);
1287 // Now that 'Ret' has unique predecessor we can safely remove all phi nodes
1289 FoldSingleEntryPHINodes(Ret);
1290 assert(!isa<PHINode>(Ret->begin()) &&
1291 "All PHI nodes should have been removed!");
1293 // At this point, we can safely insert a gc.relocate or gc.result as the first
1294 // instruction in Ret if needed.
1298 // Create new attribute set containing only attributes which can be transferred
1299 // from original call to the safepoint.
1300 static AttributeList legalizeCallAttributes(LLVMContext &Ctx,
1305 // Remove the readonly, readnone, and statepoint function attributes.
1306 AttrBuilder FnAttrs = AL.getFnAttributes();
1307 FnAttrs.removeAttribute(Attribute::ReadNone);
1308 FnAttrs.removeAttribute(Attribute::ReadOnly);
1309 for (Attribute A : AL.getFnAttributes()) {
1310 if (isStatepointDirectiveAttr(A))
1314 // Just skip parameter and return attributes for now
1315 return AttributeList::get(Ctx, AttributeList::FunctionIndex,
1316 AttributeSet::get(Ctx, FnAttrs));
1319 /// Helper function to place all gc relocates necessary for the given
1322 /// liveVariables - list of variables to be relocated.
1323 /// basePtrs - base pointers.
1324 /// statepointToken - statepoint instruction to which relocates should be
1326 /// Builder - Llvm IR builder to be used to construct new calls.
1327 static void CreateGCRelocates(ArrayRef<Value *> LiveVariables,
1328 ArrayRef<Value *> BasePtrs,
1329 Instruction *StatepointToken,
1330 IRBuilder<> &Builder) {
1331 if (LiveVariables.empty())
1334 auto FindIndex = [](ArrayRef<Value *> LiveVec, Value *Val) {
1335 auto ValIt = llvm::find(LiveVec, Val);
1336 assert(ValIt != LiveVec.end() && "Val not found in LiveVec!");
1337 size_t Index = std::distance(LiveVec.begin(), ValIt);
1338 assert(Index < LiveVec.size() && "Bug in std::find?");
1341 Module *M = StatepointToken->getModule();
1343 // All gc_relocate are generated as i8 addrspace(1)* (or a vector type whose
1344 // element type is i8 addrspace(1)*). We originally generated unique
1345 // declarations for each pointer type, but this proved problematic because
1346 // the intrinsic mangling code is incomplete and fragile. Since we're moving
1347 // towards a single unified pointer type anyways, we can just cast everything
1348 // to an i8* of the right address space. A bitcast is added later to convert
1349 // gc_relocate to the actual value's type.
1350 auto getGCRelocateDecl = [&] (Type *Ty) {
1351 assert(isHandledGCPointerType(Ty));
1352 auto AS = Ty->getScalarType()->getPointerAddressSpace();
1353 Type *NewTy = Type::getInt8PtrTy(M->getContext(), AS);
1354 if (auto *VT = dyn_cast<VectorType>(Ty))
1355 NewTy = FixedVectorType::get(NewTy,
1356 cast<FixedVectorType>(VT)->getNumElements());
1357 return Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate,
1361 // Lazily populated map from input types to the canonicalized form mentioned
1362 // in the comment above. This should probably be cached somewhere more
1364 DenseMap<Type *, Function *> TypeToDeclMap;
1366 for (unsigned i = 0; i < LiveVariables.size(); i++) {
1367 // Generate the gc.relocate call and save the result
1368 Value *BaseIdx = Builder.getInt32(FindIndex(LiveVariables, BasePtrs[i]));
1369 Value *LiveIdx = Builder.getInt32(i);
1371 Type *Ty = LiveVariables[i]->getType();
1372 if (!TypeToDeclMap.count(Ty))
1373 TypeToDeclMap[Ty] = getGCRelocateDecl(Ty);
1374 Function *GCRelocateDecl = TypeToDeclMap[Ty];
1376 // only specify a debug name if we can give a useful one
1377 CallInst *Reloc = Builder.CreateCall(
1378 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx},
1379 suffixed_name_or(LiveVariables[i], ".relocated", ""));
1380 // Trick CodeGen into thinking there are lots of free registers at this
1382 Reloc->setCallingConv(CallingConv::Cold);
1388 /// This struct is used to defer RAUWs and `eraseFromParent` s. Using this
1389 /// avoids having to worry about keeping around dangling pointers to Values.
1390 class DeferredReplacement {
1391 AssertingVH<Instruction> Old;
1392 AssertingVH<Instruction> New;
1393 bool IsDeoptimize = false;
1395 DeferredReplacement() = default;
1398 static DeferredReplacement createRAUW(Instruction *Old, Instruction *New) {
1399 assert(Old != New && Old && New &&
1400 "Cannot RAUW equal values or to / from null!");
1402 DeferredReplacement D;
1408 static DeferredReplacement createDelete(Instruction *ToErase) {
1409 DeferredReplacement D;
1414 static DeferredReplacement createDeoptimizeReplacement(Instruction *Old) {
1416 auto *F = cast<CallInst>(Old)->getCalledFunction();
1417 assert(F && F->getIntrinsicID() == Intrinsic::experimental_deoptimize &&
1418 "Only way to construct a deoptimize deferred replacement");
1420 DeferredReplacement D;
1422 D.IsDeoptimize = true;
1426 /// Does the task represented by this instance.
1427 void doReplacement() {
1428 Instruction *OldI = Old;
1429 Instruction *NewI = New;
1431 assert(OldI != NewI && "Disallowed at construction?!");
1432 assert((!IsDeoptimize || !New) &&
1433 "Deoptimize intrinsics are not replaced!");
1439 OldI->replaceAllUsesWith(NewI);
1442 // Note: we've inserted instructions, so the call to llvm.deoptimize may
1443 // not necessarily be followed by the matching return.
1444 auto *RI = cast<ReturnInst>(OldI->getParent()->getTerminator());
1445 new UnreachableInst(RI->getContext(), RI);
1446 RI->eraseFromParent();
1449 OldI->eraseFromParent();
1453 } // end anonymous namespace
1455 static StringRef getDeoptLowering(CallBase *Call) {
1456 const char *DeoptLowering = "deopt-lowering";
1457 if (Call->hasFnAttr(DeoptLowering)) {
1458 // FIXME: Calls have a *really* confusing interface around attributes
1460 const AttributeList &CSAS = Call->getAttributes();
1461 if (CSAS.hasAttribute(AttributeList::FunctionIndex, DeoptLowering))
1462 return CSAS.getAttribute(AttributeList::FunctionIndex, DeoptLowering)
1463 .getValueAsString();
1464 Function *F = Call->getCalledFunction();
1465 assert(F && F->hasFnAttribute(DeoptLowering));
1466 return F->getFnAttribute(DeoptLowering).getValueAsString();
1468 return "live-through";
1472 makeStatepointExplicitImpl(CallBase *Call, /* to replace */
1473 const SmallVectorImpl<Value *> &BasePtrs,
1474 const SmallVectorImpl<Value *> &LiveVariables,
1475 PartiallyConstructedSafepointRecord &Result,
1476 std::vector<DeferredReplacement> &Replacements) {
1477 assert(BasePtrs.size() == LiveVariables.size());
1479 // Then go ahead and use the builder do actually do the inserts. We insert
1480 // immediately before the previous instruction under the assumption that all
1481 // arguments will be available here. We can't insert afterwards since we may
1482 // be replacing a terminator.
1483 IRBuilder<> Builder(Call);
1485 ArrayRef<Value *> GCArgs(LiveVariables);
1486 uint64_t StatepointID = StatepointDirectives::DefaultStatepointID;
1487 uint32_t NumPatchBytes = 0;
1488 uint32_t Flags = uint32_t(StatepointFlags::None);
1490 ArrayRef<Use> CallArgs(Call->arg_begin(), Call->arg_end());
1491 Optional<ArrayRef<Use>> DeoptArgs;
1492 if (auto Bundle = Call->getOperandBundle(LLVMContext::OB_deopt))
1493 DeoptArgs = Bundle->Inputs;
1494 Optional<ArrayRef<Use>> TransitionArgs;
1495 if (auto Bundle = Call->getOperandBundle(LLVMContext::OB_gc_transition)) {
1496 TransitionArgs = Bundle->Inputs;
1497 // TODO: This flag no longer serves a purpose and can be removed later
1498 Flags |= uint32_t(StatepointFlags::GCTransition);
1501 // Instead of lowering calls to @llvm.experimental.deoptimize as normal calls
1502 // with a return value, we lower then as never returning calls to
1503 // __llvm_deoptimize that are followed by unreachable to get better codegen.
1504 bool IsDeoptimize = false;
1506 StatepointDirectives SD =
1507 parseStatepointDirectivesFromAttrs(Call->getAttributes());
1508 if (SD.NumPatchBytes)
1509 NumPatchBytes = *SD.NumPatchBytes;
1510 if (SD.StatepointID)
1511 StatepointID = *SD.StatepointID;
1513 // Pass through the requested lowering if any. The default is live-through.
1514 StringRef DeoptLowering = getDeoptLowering(Call);
1515 if (DeoptLowering.equals("live-in"))
1516 Flags |= uint32_t(StatepointFlags::DeoptLiveIn);
1518 assert(DeoptLowering.equals("live-through") && "Unsupported value!");
1521 Value *CallTarget = Call->getCalledOperand();
1522 if (Function *F = dyn_cast<Function>(CallTarget)) {
1523 if (F->getIntrinsicID() == Intrinsic::experimental_deoptimize) {
1524 // Calls to llvm.experimental.deoptimize are lowered to calls to the
1525 // __llvm_deoptimize symbol. We want to resolve this now, since the
1526 // verifier does not allow taking the address of an intrinsic function.
1528 SmallVector<Type *, 8> DomainTy;
1529 for (Value *Arg : CallArgs)
1530 DomainTy.push_back(Arg->getType());
1531 auto *FTy = FunctionType::get(Type::getVoidTy(F->getContext()), DomainTy,
1532 /* isVarArg = */ false);
1534 // Note: CallTarget can be a bitcast instruction of a symbol if there are
1535 // calls to @llvm.experimental.deoptimize with different argument types in
1536 // the same module. This is fine -- we assume the frontend knew what it
1537 // was doing when generating this kind of IR.
1538 CallTarget = F->getParent()
1539 ->getOrInsertFunction("__llvm_deoptimize", FTy)
1542 IsDeoptimize = true;
1546 // Create the statepoint given all the arguments
1547 GCStatepointInst *Token = nullptr;
1548 if (auto *CI = dyn_cast<CallInst>(Call)) {
1549 CallInst *SPCall = Builder.CreateGCStatepointCall(
1550 StatepointID, NumPatchBytes, CallTarget, Flags, CallArgs,
1551 TransitionArgs, DeoptArgs, GCArgs, "safepoint_token");
1553 SPCall->setTailCallKind(CI->getTailCallKind());
1554 SPCall->setCallingConv(CI->getCallingConv());
1556 // Currently we will fail on parameter attributes and on certain
1557 // function attributes. In case if we can handle this set of attributes -
1558 // set up function attrs directly on statepoint and return attrs later for
1559 // gc_result intrinsic.
1560 SPCall->setAttributes(
1561 legalizeCallAttributes(CI->getContext(), CI->getAttributes()));
1563 Token = cast<GCStatepointInst>(SPCall);
1565 // Put the following gc_result and gc_relocate calls immediately after the
1566 // the old call (which we're about to delete)
1567 assert(CI->getNextNode() && "Not a terminator, must have next!");
1568 Builder.SetInsertPoint(CI->getNextNode());
1569 Builder.SetCurrentDebugLocation(CI->getNextNode()->getDebugLoc());
1571 auto *II = cast<InvokeInst>(Call);
1573 // Insert the new invoke into the old block. We'll remove the old one in a
1574 // moment at which point this will become the new terminator for the
1576 InvokeInst *SPInvoke = Builder.CreateGCStatepointInvoke(
1577 StatepointID, NumPatchBytes, CallTarget, II->getNormalDest(),
1578 II->getUnwindDest(), Flags, CallArgs, TransitionArgs, DeoptArgs, GCArgs,
1579 "statepoint_token");
1581 SPInvoke->setCallingConv(II->getCallingConv());
1583 // Currently we will fail on parameter attributes and on certain
1584 // function attributes. In case if we can handle this set of attributes -
1585 // set up function attrs directly on statepoint and return attrs later for
1586 // gc_result intrinsic.
1587 SPInvoke->setAttributes(
1588 legalizeCallAttributes(II->getContext(), II->getAttributes()));
1590 Token = cast<GCStatepointInst>(SPInvoke);
1592 // Generate gc relocates in exceptional path
1593 BasicBlock *UnwindBlock = II->getUnwindDest();
1594 assert(!isa<PHINode>(UnwindBlock->begin()) &&
1595 UnwindBlock->getUniquePredecessor() &&
1596 "can't safely insert in this block!");
1598 Builder.SetInsertPoint(&*UnwindBlock->getFirstInsertionPt());
1599 Builder.SetCurrentDebugLocation(II->getDebugLoc());
1601 // Attach exceptional gc relocates to the landingpad.
1602 Instruction *ExceptionalToken = UnwindBlock->getLandingPadInst();
1603 Result.UnwindToken = ExceptionalToken;
1605 CreateGCRelocates(LiveVariables, BasePtrs, ExceptionalToken, Builder);
1607 // Generate gc relocates and returns for normal block
1608 BasicBlock *NormalDest = II->getNormalDest();
1609 assert(!isa<PHINode>(NormalDest->begin()) &&
1610 NormalDest->getUniquePredecessor() &&
1611 "can't safely insert in this block!");
1613 Builder.SetInsertPoint(&*NormalDest->getFirstInsertionPt());
1615 // gc relocates will be generated later as if it were regular call
1618 assert(Token && "Should be set in one of the above branches!");
1621 // If we're wrapping an @llvm.experimental.deoptimize in a statepoint, we
1622 // transform the tail-call like structure to a call to a void function
1623 // followed by unreachable to get better codegen.
1624 Replacements.push_back(
1625 DeferredReplacement::createDeoptimizeReplacement(Call));
1627 Token->setName("statepoint_token");
1628 if (!Call->getType()->isVoidTy() && !Call->use_empty()) {
1629 StringRef Name = Call->hasName() ? Call->getName() : "";
1630 CallInst *GCResult = Builder.CreateGCResult(Token, Call->getType(), Name);
1631 GCResult->setAttributes(
1632 AttributeList::get(GCResult->getContext(), AttributeList::ReturnIndex,
1633 Call->getAttributes().getRetAttributes()));
1635 // We cannot RAUW or delete CS.getInstruction() because it could be in the
1636 // live set of some other safepoint, in which case that safepoint's
1637 // PartiallyConstructedSafepointRecord will hold a raw pointer to this
1638 // llvm::Instruction. Instead, we defer the replacement and deletion to
1639 // after the live sets have been made explicit in the IR, and we no longer
1640 // have raw pointers to worry about.
1641 Replacements.emplace_back(
1642 DeferredReplacement::createRAUW(Call, GCResult));
1644 Replacements.emplace_back(DeferredReplacement::createDelete(Call));
1648 Result.StatepointToken = Token;
1650 // Second, create a gc.relocate for every live variable
1651 CreateGCRelocates(LiveVariables, BasePtrs, Token, Builder);
1654 // Replace an existing gc.statepoint with a new one and a set of gc.relocates
1655 // which make the relocations happening at this safepoint explicit.
1657 // WARNING: Does not do any fixup to adjust users of the original live
1658 // values. That's the callers responsibility.
1660 makeStatepointExplicit(DominatorTree &DT, CallBase *Call,
1661 PartiallyConstructedSafepointRecord &Result,
1662 std::vector<DeferredReplacement> &Replacements) {
1663 const auto &LiveSet = Result.LiveSet;
1664 const auto &PointerToBase = Result.PointerToBase;
1666 // Convert to vector for efficient cross referencing.
1667 SmallVector<Value *, 64> BaseVec, LiveVec;
1668 LiveVec.reserve(LiveSet.size());
1669 BaseVec.reserve(LiveSet.size());
1670 for (Value *L : LiveSet) {
1671 LiveVec.push_back(L);
1672 assert(PointerToBase.count(L));
1673 Value *Base = PointerToBase.find(L)->second;
1674 BaseVec.push_back(Base);
1676 assert(LiveVec.size() == BaseVec.size());
1678 // Do the actual rewriting and delete the old statepoint
1679 makeStatepointExplicitImpl(Call, BaseVec, LiveVec, Result, Replacements);
1682 // Helper function for the relocationViaAlloca.
1684 // It receives iterator to the statepoint gc relocates and emits a store to the
1685 // assigned location (via allocaMap) for the each one of them. It adds the
1686 // visited values into the visitedLiveValues set, which we will later use them
1687 // for sanity checking.
1689 insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs,
1690 DenseMap<Value *, AllocaInst *> &AllocaMap,
1691 DenseSet<Value *> &VisitedLiveValues) {
1692 for (User *U : GCRelocs) {
1693 GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U);
1697 Value *OriginalValue = Relocate->getDerivedPtr();
1698 assert(AllocaMap.count(OriginalValue));
1699 Value *Alloca = AllocaMap[OriginalValue];
1701 // Emit store into the related alloca
1702 // All gc_relocates are i8 addrspace(1)* typed, and it must be bitcasted to
1703 // the correct type according to alloca.
1704 assert(Relocate->getNextNode() &&
1705 "Should always have one since it's not a terminator");
1706 IRBuilder<> Builder(Relocate->getNextNode());
1707 Value *CastedRelocatedValue =
1708 Builder.CreateBitCast(Relocate,
1709 cast<AllocaInst>(Alloca)->getAllocatedType(),
1710 suffixed_name_or(Relocate, ".casted", ""));
1712 new StoreInst(CastedRelocatedValue, Alloca,
1713 cast<Instruction>(CastedRelocatedValue)->getNextNode());
1716 VisitedLiveValues.insert(OriginalValue);
1721 // Helper function for the "relocationViaAlloca". Similar to the
1722 // "insertRelocationStores" but works for rematerialized values.
1723 static void insertRematerializationStores(
1724 const RematerializedValueMapTy &RematerializedValues,
1725 DenseMap<Value *, AllocaInst *> &AllocaMap,
1726 DenseSet<Value *> &VisitedLiveValues) {
1727 for (auto RematerializedValuePair: RematerializedValues) {
1728 Instruction *RematerializedValue = RematerializedValuePair.first;
1729 Value *OriginalValue = RematerializedValuePair.second;
1731 assert(AllocaMap.count(OriginalValue) &&
1732 "Can not find alloca for rematerialized value");
1733 Value *Alloca = AllocaMap[OriginalValue];
1735 new StoreInst(RematerializedValue, Alloca,
1736 RematerializedValue->getNextNode());
1739 VisitedLiveValues.insert(OriginalValue);
1744 /// Do all the relocation update via allocas and mem2reg
1745 static void relocationViaAlloca(
1746 Function &F, DominatorTree &DT, ArrayRef<Value *> Live,
1747 ArrayRef<PartiallyConstructedSafepointRecord> Records) {
1749 // record initial number of (static) allocas; we'll check we have the same
1750 // number when we get done.
1751 int InitialAllocaNum = 0;
1752 for (Instruction &I : F.getEntryBlock())
1753 if (isa<AllocaInst>(I))
1757 // TODO-PERF: change data structures, reserve
1758 DenseMap<Value *, AllocaInst *> AllocaMap;
1759 SmallVector<AllocaInst *, 200> PromotableAllocas;
1760 // Used later to chack that we have enough allocas to store all values
1761 std::size_t NumRematerializedValues = 0;
1762 PromotableAllocas.reserve(Live.size());
1764 // Emit alloca for "LiveValue" and record it in "allocaMap" and
1765 // "PromotableAllocas"
1766 const DataLayout &DL = F.getParent()->getDataLayout();
1767 auto emitAllocaFor = [&](Value *LiveValue) {
1768 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(),
1769 DL.getAllocaAddrSpace(), "",
1770 F.getEntryBlock().getFirstNonPHI());
1771 AllocaMap[LiveValue] = Alloca;
1772 PromotableAllocas.push_back(Alloca);
1775 // Emit alloca for each live gc pointer
1776 for (Value *V : Live)
1779 // Emit allocas for rematerialized values
1780 for (const auto &Info : Records)
1781 for (auto RematerializedValuePair : Info.RematerializedValues) {
1782 Value *OriginalValue = RematerializedValuePair.second;
1783 if (AllocaMap.count(OriginalValue) != 0)
1786 emitAllocaFor(OriginalValue);
1787 ++NumRematerializedValues;
1790 // The next two loops are part of the same conceptual operation. We need to
1791 // insert a store to the alloca after the original def and at each
1792 // redefinition. We need to insert a load before each use. These are split
1793 // into distinct loops for performance reasons.
1795 // Update gc pointer after each statepoint: either store a relocated value or
1796 // null (if no relocated value was found for this gc pointer and it is not a
1797 // gc_result). This must happen before we update the statepoint with load of
1798 // alloca otherwise we lose the link between statepoint and old def.
1799 for (const auto &Info : Records) {
1800 Value *Statepoint = Info.StatepointToken;
1802 // This will be used for consistency check
1803 DenseSet<Value *> VisitedLiveValues;
1805 // Insert stores for normal statepoint gc relocates
1806 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues);
1808 // In case if it was invoke statepoint
1809 // we will insert stores for exceptional path gc relocates.
1810 if (isa<InvokeInst>(Statepoint)) {
1811 insertRelocationStores(Info.UnwindToken->users(), AllocaMap,
1815 // Do similar thing with rematerialized values
1816 insertRematerializationStores(Info.RematerializedValues, AllocaMap,
1819 if (ClobberNonLive) {
1820 // As a debugging aid, pretend that an unrelocated pointer becomes null at
1821 // the gc.statepoint. This will turn some subtle GC problems into
1822 // slightly easier to debug SEGVs. Note that on large IR files with
1823 // lots of gc.statepoints this is extremely costly both memory and time
1825 SmallVector<AllocaInst *, 64> ToClobber;
1826 for (auto Pair : AllocaMap) {
1827 Value *Def = Pair.first;
1828 AllocaInst *Alloca = Pair.second;
1830 // This value was relocated
1831 if (VisitedLiveValues.count(Def)) {
1834 ToClobber.push_back(Alloca);
1837 auto InsertClobbersAt = [&](Instruction *IP) {
1838 for (auto *AI : ToClobber) {
1839 auto PT = cast<PointerType>(AI->getAllocatedType());
1840 Constant *CPN = ConstantPointerNull::get(PT);
1841 new StoreInst(CPN, AI, IP);
1845 // Insert the clobbering stores. These may get intermixed with the
1846 // gc.results and gc.relocates, but that's fine.
1847 if (auto II = dyn_cast<InvokeInst>(Statepoint)) {
1848 InsertClobbersAt(&*II->getNormalDest()->getFirstInsertionPt());
1849 InsertClobbersAt(&*II->getUnwindDest()->getFirstInsertionPt());
1851 InsertClobbersAt(cast<Instruction>(Statepoint)->getNextNode());
1856 // Update use with load allocas and add store for gc_relocated.
1857 for (auto Pair : AllocaMap) {
1858 Value *Def = Pair.first;
1859 AllocaInst *Alloca = Pair.second;
1861 // We pre-record the uses of allocas so that we dont have to worry about
1862 // later update that changes the user information..
1864 SmallVector<Instruction *, 20> Uses;
1865 // PERF: trade a linear scan for repeated reallocation
1866 Uses.reserve(Def->getNumUses());
1867 for (User *U : Def->users()) {
1868 if (!isa<ConstantExpr>(U)) {
1869 // If the def has a ConstantExpr use, then the def is either a
1870 // ConstantExpr use itself or null. In either case
1871 // (recursively in the first, directly in the second), the oop
1872 // it is ultimately dependent on is null and this particular
1873 // use does not need to be fixed up.
1874 Uses.push_back(cast<Instruction>(U));
1879 auto Last = std::unique(Uses.begin(), Uses.end());
1880 Uses.erase(Last, Uses.end());
1882 for (Instruction *Use : Uses) {
1883 if (isa<PHINode>(Use)) {
1884 PHINode *Phi = cast<PHINode>(Use);
1885 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) {
1886 if (Def == Phi->getIncomingValue(i)) {
1888 new LoadInst(Alloca->getAllocatedType(), Alloca, "",
1889 Phi->getIncomingBlock(i)->getTerminator());
1890 Phi->setIncomingValue(i, Load);
1895 new LoadInst(Alloca->getAllocatedType(), Alloca, "", Use);
1896 Use->replaceUsesOfWith(Def, Load);
1900 // Emit store for the initial gc value. Store must be inserted after load,
1901 // otherwise store will be in alloca's use list and an extra load will be
1902 // inserted before it.
1903 StoreInst *Store = new StoreInst(Def, Alloca, /*volatile*/ false,
1904 DL.getABITypeAlign(Def->getType()));
1905 if (Instruction *Inst = dyn_cast<Instruction>(Def)) {
1906 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) {
1907 // InvokeInst is a terminator so the store need to be inserted into its
1908 // normal destination block.
1909 BasicBlock *NormalDest = Invoke->getNormalDest();
1910 Store->insertBefore(NormalDest->getFirstNonPHI());
1912 assert(!Inst->isTerminator() &&
1913 "The only terminator that can produce a value is "
1914 "InvokeInst which is handled above.");
1915 Store->insertAfter(Inst);
1918 assert(isa<Argument>(Def));
1919 Store->insertAfter(cast<Instruction>(Alloca));
1923 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues &&
1924 "we must have the same allocas with lives");
1925 if (!PromotableAllocas.empty()) {
1926 // Apply mem2reg to promote alloca to SSA
1927 PromoteMemToReg(PromotableAllocas, DT);
1931 for (auto &I : F.getEntryBlock())
1932 if (isa<AllocaInst>(I))
1934 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas");
1938 /// Implement a unique function which doesn't require we sort the input
1939 /// vector. Doing so has the effect of changing the output of a couple of
1940 /// tests in ways which make them less useful in testing fused safepoints.
1941 template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) {
1942 SmallSet<T, 8> Seen;
1943 Vec.erase(remove_if(Vec, [&](const T &V) { return !Seen.insert(V).second; }),
1947 /// Insert holders so that each Value is obviously live through the entire
1948 /// lifetime of the call.
1949 static void insertUseHolderAfter(CallBase *Call, const ArrayRef<Value *> Values,
1950 SmallVectorImpl<CallInst *> &Holders) {
1952 // No values to hold live, might as well not insert the empty holder
1955 Module *M = Call->getModule();
1956 // Use a dummy vararg function to actually hold the values live
1957 FunctionCallee Func = M->getOrInsertFunction(
1958 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true));
1959 if (isa<CallInst>(Call)) {
1960 // For call safepoints insert dummy calls right after safepoint
1962 CallInst::Create(Func, Values, "", &*++Call->getIterator()));
1965 // For invoke safepooints insert dummy calls both in normal and
1966 // exceptional destination blocks
1967 auto *II = cast<InvokeInst>(Call);
1968 Holders.push_back(CallInst::Create(
1969 Func, Values, "", &*II->getNormalDest()->getFirstInsertionPt()));
1970 Holders.push_back(CallInst::Create(
1971 Func, Values, "", &*II->getUnwindDest()->getFirstInsertionPt()));
1974 static void findLiveReferences(
1975 Function &F, DominatorTree &DT, ArrayRef<CallBase *> toUpdate,
1976 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) {
1977 GCPtrLivenessData OriginalLivenessData;
1978 computeLiveInValues(DT, F, OriginalLivenessData);
1979 for (size_t i = 0; i < records.size(); i++) {
1980 struct PartiallyConstructedSafepointRecord &info = records[i];
1981 analyzeParsePointLiveness(DT, OriginalLivenessData, toUpdate[i], info);
1985 // Helper function for the "rematerializeLiveValues". It walks use chain
1986 // starting from the "CurrentValue" until it reaches the root of the chain, i.e.
1987 // the base or a value it cannot process. Only "simple" values are processed
1988 // (currently it is GEP's and casts). The returned root is examined by the
1989 // callers of findRematerializableChainToBasePointer. Fills "ChainToBase" array
1990 // with all visited values.
1991 static Value* findRematerializableChainToBasePointer(
1992 SmallVectorImpl<Instruction*> &ChainToBase,
1993 Value *CurrentValue) {
1994 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) {
1995 ChainToBase.push_back(GEP);
1996 return findRematerializableChainToBasePointer(ChainToBase,
1997 GEP->getPointerOperand());
2000 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) {
2001 if (!CI->isNoopCast(CI->getModule()->getDataLayout()))
2004 ChainToBase.push_back(CI);
2005 return findRematerializableChainToBasePointer(ChainToBase,
2009 // We have reached the root of the chain, which is either equal to the base or
2010 // is the first unsupported value along the use chain.
2011 return CurrentValue;
2014 // Helper function for the "rematerializeLiveValues". Compute cost of the use
2015 // chain we are going to rematerialize.
2017 chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain,
2018 TargetTransformInfo &TTI) {
2021 for (Instruction *Instr : Chain) {
2022 if (CastInst *CI = dyn_cast<CastInst>(Instr)) {
2023 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) &&
2024 "non noop cast is found during rematerialization");
2026 Type *SrcTy = CI->getOperand(0)->getType();
2027 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy,
2028 TargetTransformInfo::TCK_SizeAndLatency,
2031 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) {
2032 // Cost of the address calculation
2033 Type *ValTy = GEP->getSourceElementType();
2034 Cost += TTI.getAddressComputationCost(ValTy);
2036 // And cost of the GEP itself
2037 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not
2038 // allowed for the external usage)
2039 if (!GEP->hasAllConstantIndices())
2043 llvm_unreachable("unsupported instruction type during rematerialization");
2050 static bool AreEquivalentPhiNodes(PHINode &OrigRootPhi, PHINode &AlternateRootPhi) {
2051 unsigned PhiNum = OrigRootPhi.getNumIncomingValues();
2052 if (PhiNum != AlternateRootPhi.getNumIncomingValues() ||
2053 OrigRootPhi.getParent() != AlternateRootPhi.getParent())
2055 // Map of incoming values and their corresponding basic blocks of
2057 SmallDenseMap<Value *, BasicBlock *, 8> CurrentIncomingValues;
2058 for (unsigned i = 0; i < PhiNum; i++)
2059 CurrentIncomingValues[OrigRootPhi.getIncomingValue(i)] =
2060 OrigRootPhi.getIncomingBlock(i);
2062 // Both current and base PHIs should have same incoming values and
2063 // the same basic blocks corresponding to the incoming values.
2064 for (unsigned i = 0; i < PhiNum; i++) {
2066 CurrentIncomingValues.find(AlternateRootPhi.getIncomingValue(i));
2067 if (CIVI == CurrentIncomingValues.end())
2069 BasicBlock *CurrentIncomingBB = CIVI->second;
2070 if (CurrentIncomingBB != AlternateRootPhi.getIncomingBlock(i))
2076 // From the statepoint live set pick values that are cheaper to recompute then
2077 // to relocate. Remove this values from the live set, rematerialize them after
2078 // statepoint and record them in "Info" structure. Note that similar to
2079 // relocated values we don't do any user adjustments here.
2080 static void rematerializeLiveValues(CallBase *Call,
2081 PartiallyConstructedSafepointRecord &Info,
2082 TargetTransformInfo &TTI) {
2083 const unsigned int ChainLengthThreshold = 10;
2085 // Record values we are going to delete from this statepoint live set.
2086 // We can not di this in following loop due to iterator invalidation.
2087 SmallVector<Value *, 32> LiveValuesToBeDeleted;
2089 for (Value *LiveValue: Info.LiveSet) {
2090 // For each live pointer find its defining chain
2091 SmallVector<Instruction *, 3> ChainToBase;
2092 assert(Info.PointerToBase.count(LiveValue));
2093 Value *RootOfChain =
2094 findRematerializableChainToBasePointer(ChainToBase,
2097 // Nothing to do, or chain is too long
2098 if ( ChainToBase.size() == 0 ||
2099 ChainToBase.size() > ChainLengthThreshold)
2102 // Handle the scenario where the RootOfChain is not equal to the
2103 // Base Value, but they are essentially the same phi values.
2104 if (RootOfChain != Info.PointerToBase[LiveValue]) {
2105 PHINode *OrigRootPhi = dyn_cast<PHINode>(RootOfChain);
2106 PHINode *AlternateRootPhi = dyn_cast<PHINode>(Info.PointerToBase[LiveValue]);
2107 if (!OrigRootPhi || !AlternateRootPhi)
2109 // PHI nodes that have the same incoming values, and belonging to the same
2110 // basic blocks are essentially the same SSA value. When the original phi
2111 // has incoming values with different base pointers, the original phi is
2112 // marked as conflict, and an additional `AlternateRootPhi` with the same
2113 // incoming values get generated by the findBasePointer function. We need
2114 // to identify the newly generated AlternateRootPhi (.base version of phi)
2115 // and RootOfChain (the original phi node itself) are the same, so that we
2116 // can rematerialize the gep and casts. This is a workaround for the
2117 // deficiency in the findBasePointer algorithm.
2118 if (!AreEquivalentPhiNodes(*OrigRootPhi, *AlternateRootPhi))
2120 // Now that the phi nodes are proved to be the same, assert that
2121 // findBasePointer's newly generated AlternateRootPhi is present in the
2122 // liveset of the call.
2123 assert(Info.LiveSet.count(AlternateRootPhi));
2125 // Compute cost of this chain
2126 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI);
2127 // TODO: We can also account for cases when we will be able to remove some
2128 // of the rematerialized values by later optimization passes. I.e if
2129 // we rematerialized several intersecting chains. Or if original values
2130 // don't have any uses besides this statepoint.
2132 // For invokes we need to rematerialize each chain twice - for normal and
2133 // for unwind basic blocks. Model this by multiplying cost by two.
2134 if (isa<InvokeInst>(Call)) {
2137 // If it's too expensive - skip it
2138 if (Cost >= RematerializationThreshold)
2141 // Remove value from the live set
2142 LiveValuesToBeDeleted.push_back(LiveValue);
2144 // Clone instructions and record them inside "Info" structure
2146 // Walk backwards to visit top-most instructions first
2147 std::reverse(ChainToBase.begin(), ChainToBase.end());
2149 // Utility function which clones all instructions from "ChainToBase"
2150 // and inserts them before "InsertBefore". Returns rematerialized value
2151 // which should be used after statepoint.
2152 auto rematerializeChain = [&ChainToBase](
2153 Instruction *InsertBefore, Value *RootOfChain, Value *AlternateLiveBase) {
2154 Instruction *LastClonedValue = nullptr;
2155 Instruction *LastValue = nullptr;
2156 for (Instruction *Instr: ChainToBase) {
2157 // Only GEP's and casts are supported as we need to be careful to not
2158 // introduce any new uses of pointers not in the liveset.
2159 // Note that it's fine to introduce new uses of pointers which were
2160 // otherwise not used after this statepoint.
2161 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr));
2163 Instruction *ClonedValue = Instr->clone();
2164 ClonedValue->insertBefore(InsertBefore);
2165 ClonedValue->setName(Instr->getName() + ".remat");
2167 // If it is not first instruction in the chain then it uses previously
2168 // cloned value. We should update it to use cloned value.
2169 if (LastClonedValue) {
2171 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue);
2173 for (auto OpValue : ClonedValue->operand_values()) {
2174 // Assert that cloned instruction does not use any instructions from
2175 // this chain other than LastClonedValue
2176 assert(!is_contained(ChainToBase, OpValue) &&
2177 "incorrect use in rematerialization chain");
2178 // Assert that the cloned instruction does not use the RootOfChain
2179 // or the AlternateLiveBase.
2180 assert(OpValue != RootOfChain && OpValue != AlternateLiveBase);
2184 // For the first instruction, replace the use of unrelocated base i.e.
2185 // RootOfChain/OrigRootPhi, with the corresponding PHI present in the
2186 // live set. They have been proved to be the same PHI nodes. Note
2187 // that the *only* use of the RootOfChain in the ChainToBase list is
2188 // the first Value in the list.
2189 if (RootOfChain != AlternateLiveBase)
2190 ClonedValue->replaceUsesOfWith(RootOfChain, AlternateLiveBase);
2193 LastClonedValue = ClonedValue;
2196 assert(LastClonedValue);
2197 return LastClonedValue;
2200 // Different cases for calls and invokes. For invokes we need to clone
2201 // instructions both on normal and unwind path.
2202 if (isa<CallInst>(Call)) {
2203 Instruction *InsertBefore = Call->getNextNode();
2204 assert(InsertBefore);
2205 Instruction *RematerializedValue = rematerializeChain(
2206 InsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2207 Info.RematerializedValues[RematerializedValue] = LiveValue;
2209 auto *Invoke = cast<InvokeInst>(Call);
2211 Instruction *NormalInsertBefore =
2212 &*Invoke->getNormalDest()->getFirstInsertionPt();
2213 Instruction *UnwindInsertBefore =
2214 &*Invoke->getUnwindDest()->getFirstInsertionPt();
2216 Instruction *NormalRematerializedValue = rematerializeChain(
2217 NormalInsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2218 Instruction *UnwindRematerializedValue = rematerializeChain(
2219 UnwindInsertBefore, RootOfChain, Info.PointerToBase[LiveValue]);
2221 Info.RematerializedValues[NormalRematerializedValue] = LiveValue;
2222 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue;
2226 // Remove rematerializaed values from the live set
2227 for (auto LiveValue: LiveValuesToBeDeleted) {
2228 Info.LiveSet.remove(LiveValue);
2232 static bool insertParsePoints(Function &F, DominatorTree &DT,
2233 TargetTransformInfo &TTI,
2234 SmallVectorImpl<CallBase *> &ToUpdate) {
2236 // sanity check the input
2237 std::set<CallBase *> Uniqued;
2238 Uniqued.insert(ToUpdate.begin(), ToUpdate.end());
2239 assert(Uniqued.size() == ToUpdate.size() && "no duplicates please!");
2241 for (CallBase *Call : ToUpdate)
2242 assert(Call->getFunction() == &F);
2245 // When inserting gc.relocates for invokes, we need to be able to insert at
2246 // the top of the successor blocks. See the comment on
2247 // normalForInvokeSafepoint on exactly what is needed. Note that this step
2248 // may restructure the CFG.
2249 for (CallBase *Call : ToUpdate) {
2250 auto *II = dyn_cast<InvokeInst>(Call);
2253 normalizeForInvokeSafepoint(II->getNormalDest(), II->getParent(), DT);
2254 normalizeForInvokeSafepoint(II->getUnwindDest(), II->getParent(), DT);
2257 // A list of dummy calls added to the IR to keep various values obviously
2258 // live in the IR. We'll remove all of these when done.
2259 SmallVector<CallInst *, 64> Holders;
2261 // Insert a dummy call with all of the deopt operands we'll need for the
2262 // actual safepoint insertion as arguments. This ensures reference operands
2263 // in the deopt argument list are considered live through the safepoint (and
2264 // thus makes sure they get relocated.)
2265 for (CallBase *Call : ToUpdate) {
2266 SmallVector<Value *, 64> DeoptValues;
2268 for (Value *Arg : GetDeoptBundleOperands(Call)) {
2269 assert(!isUnhandledGCPointerType(Arg->getType()) &&
2270 "support for FCA unimplemented");
2271 if (isHandledGCPointerType(Arg->getType()))
2272 DeoptValues.push_back(Arg);
2275 insertUseHolderAfter(Call, DeoptValues, Holders);
2278 SmallVector<PartiallyConstructedSafepointRecord, 64> Records(ToUpdate.size());
2280 // A) Identify all gc pointers which are statically live at the given call
2282 findLiveReferences(F, DT, ToUpdate, Records);
2284 // B) Find the base pointers for each live pointer
2285 /* scope for caching */ {
2286 // Cache the 'defining value' relation used in the computation and
2287 // insertion of base phis and selects. This ensures that we don't insert
2288 // large numbers of duplicate base_phis.
2289 DefiningValueMapTy DVCache;
2291 for (size_t i = 0; i < Records.size(); i++) {
2292 PartiallyConstructedSafepointRecord &info = Records[i];
2293 findBasePointers(DT, DVCache, ToUpdate[i], info);
2295 } // end of cache scope
2297 // The base phi insertion logic (for any safepoint) may have inserted new
2298 // instructions which are now live at some safepoint. The simplest such
2301 // phi a <-- will be a new base_phi here
2302 // safepoint 1 <-- that needs to be live here
2306 // We insert some dummy calls after each safepoint to definitely hold live
2307 // the base pointers which were identified for that safepoint. We'll then
2308 // ask liveness for _every_ base inserted to see what is now live. Then we
2309 // remove the dummy calls.
2310 Holders.reserve(Holders.size() + Records.size());
2311 for (size_t i = 0; i < Records.size(); i++) {
2312 PartiallyConstructedSafepointRecord &Info = Records[i];
2314 SmallVector<Value *, 128> Bases;
2315 for (auto Pair : Info.PointerToBase)
2316 Bases.push_back(Pair.second);
2318 insertUseHolderAfter(ToUpdate[i], Bases, Holders);
2321 // By selecting base pointers, we've effectively inserted new uses. Thus, we
2322 // need to rerun liveness. We may *also* have inserted new defs, but that's
2323 // not the key issue.
2324 recomputeLiveInValues(F, DT, ToUpdate, Records);
2326 if (PrintBasePointers) {
2327 for (auto &Info : Records) {
2328 errs() << "Base Pairs: (w/Relocation)\n";
2329 for (auto Pair : Info.PointerToBase) {
2330 errs() << " derived ";
2331 Pair.first->printAsOperand(errs(), false);
2333 Pair.second->printAsOperand(errs(), false);
2339 // It is possible that non-constant live variables have a constant base. For
2340 // example, a GEP with a variable offset from a global. In this case we can
2341 // remove it from the liveset. We already don't add constants to the liveset
2342 // because we assume they won't move at runtime and the GC doesn't need to be
2343 // informed about them. The same reasoning applies if the base is constant.
2344 // Note that the relocation placement code relies on this filtering for
2345 // correctness as it expects the base to be in the liveset, which isn't true
2346 // if the base is constant.
2347 for (auto &Info : Records)
2348 for (auto &BasePair : Info.PointerToBase)
2349 if (isa<Constant>(BasePair.second))
2350 Info.LiveSet.remove(BasePair.first);
2352 for (CallInst *CI : Holders)
2353 CI->eraseFromParent();
2357 // In order to reduce live set of statepoint we might choose to rematerialize
2358 // some values instead of relocating them. This is purely an optimization and
2359 // does not influence correctness.
2360 for (size_t i = 0; i < Records.size(); i++)
2361 rematerializeLiveValues(ToUpdate[i], Records[i], TTI);
2363 // We need this to safely RAUW and delete call or invoke return values that
2364 // may themselves be live over a statepoint. For details, please see usage in
2365 // makeStatepointExplicitImpl.
2366 std::vector<DeferredReplacement> Replacements;
2368 // Now run through and replace the existing statepoints with new ones with
2369 // the live variables listed. We do not yet update uses of the values being
2370 // relocated. We have references to live variables that need to
2371 // survive to the last iteration of this loop. (By construction, the
2372 // previous statepoint can not be a live variable, thus we can and remove
2373 // the old statepoint calls as we go.)
2374 for (size_t i = 0; i < Records.size(); i++)
2375 makeStatepointExplicit(DT, ToUpdate[i], Records[i], Replacements);
2377 ToUpdate.clear(); // prevent accident use of invalid calls.
2379 for (auto &PR : Replacements)
2382 Replacements.clear();
2384 for (auto &Info : Records) {
2385 // These live sets may contain state Value pointers, since we replaced calls
2386 // with operand bundles with calls wrapped in gc.statepoint, and some of
2387 // those calls may have been def'ing live gc pointers. Clear these out to
2388 // avoid accidentally using them.
2390 // TODO: We should create a separate data structure that does not contain
2391 // these live sets, and migrate to using that data structure from this point
2393 Info.LiveSet.clear();
2394 Info.PointerToBase.clear();
2397 // Do all the fixups of the original live variables to their relocated selves
2398 SmallVector<Value *, 128> Live;
2399 for (size_t i = 0; i < Records.size(); i++) {
2400 PartiallyConstructedSafepointRecord &Info = Records[i];
2402 // We can't simply save the live set from the original insertion. One of
2403 // the live values might be the result of a call which needs a safepoint.
2404 // That Value* no longer exists and we need to use the new gc_result.
2405 // Thankfully, the live set is embedded in the statepoint (and updated), so
2406 // we just grab that.
2407 Live.insert(Live.end(), Info.StatepointToken->gc_args_begin(),
2408 Info.StatepointToken->gc_args_end());
2410 // Do some basic sanity checks on our liveness results before performing
2411 // relocation. Relocation can and will turn mistakes in liveness results
2412 // into non-sensical code which is must harder to debug.
2413 // TODO: It would be nice to test consistency as well
2414 assert(DT.isReachableFromEntry(Info.StatepointToken->getParent()) &&
2415 "statepoint must be reachable or liveness is meaningless");
2416 for (Value *V : Info.StatepointToken->gc_args()) {
2417 if (!isa<Instruction>(V))
2418 // Non-instruction values trivial dominate all possible uses
2420 auto *LiveInst = cast<Instruction>(V);
2421 assert(DT.isReachableFromEntry(LiveInst->getParent()) &&
2422 "unreachable values should never be live");
2423 assert(DT.dominates(LiveInst, Info.StatepointToken) &&
2424 "basic SSA liveness expectation violated by liveness analysis");
2428 unique_unsorted(Live);
2432 for (auto *Ptr : Live)
2433 assert(isHandledGCPointerType(Ptr->getType()) &&
2434 "must be a gc pointer type");
2437 relocationViaAlloca(F, DT, Live, Records);
2438 return !Records.empty();
2441 // Handles both return values and arguments for Functions and calls.
2442 template <typename AttrHolder>
2443 static void RemoveNonValidAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH,
2446 if (AH.getDereferenceableBytes(Index))
2447 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable,
2448 AH.getDereferenceableBytes(Index)));
2449 if (AH.getDereferenceableOrNullBytes(Index))
2450 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull,
2451 AH.getDereferenceableOrNullBytes(Index)));
2452 if (AH.getAttributes().hasAttribute(Index, Attribute::NoAlias))
2453 R.addAttribute(Attribute::NoAlias);
2456 AH.setAttributes(AH.getAttributes().removeAttributes(Ctx, Index, R));
2459 static void stripNonValidAttributesFromPrototype(Function &F) {
2460 LLVMContext &Ctx = F.getContext();
2462 for (Argument &A : F.args())
2463 if (isa<PointerType>(A.getType()))
2464 RemoveNonValidAttrAtIndex(Ctx, F,
2465 A.getArgNo() + AttributeList::FirstArgIndex);
2467 if (isa<PointerType>(F.getReturnType()))
2468 RemoveNonValidAttrAtIndex(Ctx, F, AttributeList::ReturnIndex);
2471 /// Certain metadata on instructions are invalid after running RS4GC.
2472 /// Optimizations that run after RS4GC can incorrectly use this metadata to
2473 /// optimize functions. We drop such metadata on the instruction.
2474 static void stripInvalidMetadataFromInstruction(Instruction &I) {
2475 if (!isa<LoadInst>(I) && !isa<StoreInst>(I))
2477 // These are the attributes that are still valid on loads and stores after
2479 // The metadata implying dereferenceability and noalias are (conservatively)
2480 // dropped. This is because semantically, after RewriteStatepointsForGC runs,
2481 // all calls to gc.statepoint "free" the entire heap. Also, gc.statepoint can
2482 // touch the entire heap including noalias objects. Note: The reasoning is
2483 // same as stripping the dereferenceability and noalias attributes that are
2484 // analogous to the metadata counterparts.
2485 // We also drop the invariant.load metadata on the load because that metadata
2486 // implies the address operand to the load points to memory that is never
2487 // changed once it became dereferenceable. This is no longer true after RS4GC.
2488 // Similar reasoning applies to invariant.group metadata, which applies to
2489 // loads within a group.
2490 unsigned ValidMetadataAfterRS4GC[] = {LLVMContext::MD_tbaa,
2491 LLVMContext::MD_range,
2492 LLVMContext::MD_alias_scope,
2493 LLVMContext::MD_nontemporal,
2494 LLVMContext::MD_nonnull,
2495 LLVMContext::MD_align,
2496 LLVMContext::MD_type};
2498 // Drops all metadata on the instruction other than ValidMetadataAfterRS4GC.
2499 I.dropUnknownNonDebugMetadata(ValidMetadataAfterRS4GC);
2502 static void stripNonValidDataFromBody(Function &F) {
2506 LLVMContext &Ctx = F.getContext();
2507 MDBuilder Builder(Ctx);
2509 // Set of invariantstart instructions that we need to remove.
2510 // Use this to avoid invalidating the instruction iterator.
2511 SmallVector<IntrinsicInst*, 12> InvariantStartInstructions;
2513 for (Instruction &I : instructions(F)) {
2514 // invariant.start on memory location implies that the referenced memory
2515 // location is constant and unchanging. This is no longer true after
2516 // RewriteStatepointsForGC runs because there can be calls to gc.statepoint
2517 // which frees the entire heap and the presence of invariant.start allows
2518 // the optimizer to sink the load of a memory location past a statepoint,
2519 // which is incorrect.
2520 if (auto *II = dyn_cast<IntrinsicInst>(&I))
2521 if (II->getIntrinsicID() == Intrinsic::invariant_start) {
2522 InvariantStartInstructions.push_back(II);
2526 if (MDNode *Tag = I.getMetadata(LLVMContext::MD_tbaa)) {
2527 MDNode *MutableTBAA = Builder.createMutableTBAAAccessTag(Tag);
2528 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA);
2531 stripInvalidMetadataFromInstruction(I);
2533 if (auto *Call = dyn_cast<CallBase>(&I)) {
2534 for (int i = 0, e = Call->arg_size(); i != e; i++)
2535 if (isa<PointerType>(Call->getArgOperand(i)->getType()))
2536 RemoveNonValidAttrAtIndex(Ctx, *Call,
2537 i + AttributeList::FirstArgIndex);
2538 if (isa<PointerType>(Call->getType()))
2539 RemoveNonValidAttrAtIndex(Ctx, *Call, AttributeList::ReturnIndex);
2543 // Delete the invariant.start instructions and RAUW undef.
2544 for (auto *II : InvariantStartInstructions) {
2545 II->replaceAllUsesWith(UndefValue::get(II->getType()));
2546 II->eraseFromParent();
2550 /// Returns true if this function should be rewritten by this pass. The main
2551 /// point of this function is as an extension point for custom logic.
2552 static bool shouldRewriteStatepointsIn(Function &F) {
2553 // TODO: This should check the GCStrategy
2555 const auto &FunctionGCName = F.getGC();
2556 const StringRef StatepointExampleName("statepoint-example");
2557 const StringRef CoreCLRName("coreclr");
2558 return (StatepointExampleName == FunctionGCName) ||
2559 (CoreCLRName == FunctionGCName);
2564 static void stripNonValidData(Module &M) {
2566 assert(llvm::any_of(M, shouldRewriteStatepointsIn) && "precondition!");
2569 for (Function &F : M)
2570 stripNonValidAttributesFromPrototype(F);
2572 for (Function &F : M)
2573 stripNonValidDataFromBody(F);
2576 bool RewriteStatepointsForGC::runOnFunction(Function &F, DominatorTree &DT,
2577 TargetTransformInfo &TTI,
2578 const TargetLibraryInfo &TLI) {
2579 assert(!F.isDeclaration() && !F.empty() &&
2580 "need function body to rewrite statepoints in");
2581 assert(shouldRewriteStatepointsIn(F) && "mismatch in rewrite decision");
2583 auto NeedsRewrite = [&TLI](Instruction &I) {
2584 if (const auto *Call = dyn_cast<CallBase>(&I))
2585 return !callsGCLeafFunction(Call, TLI) && !isa<GCStatepointInst>(Call);
2589 // Delete any unreachable statepoints so that we don't have unrewritten
2590 // statepoints surviving this pass. This makes testing easier and the
2591 // resulting IR less confusing to human readers.
2592 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
2593 bool MadeChange = removeUnreachableBlocks(F, &DTU);
2594 // Flush the Dominator Tree.
2597 // Gather all the statepoints which need rewritten. Be careful to only
2598 // consider those in reachable code since we need to ask dominance queries
2599 // when rewriting. We'll delete the unreachable ones in a moment.
2600 SmallVector<CallBase *, 64> ParsePointNeeded;
2601 for (Instruction &I : instructions(F)) {
2602 // TODO: only the ones with the flag set!
2603 if (NeedsRewrite(I)) {
2604 // NOTE removeUnreachableBlocks() is stronger than
2605 // DominatorTree::isReachableFromEntry(). In other words
2606 // removeUnreachableBlocks can remove some blocks for which
2607 // isReachableFromEntry() returns true.
2608 assert(DT.isReachableFromEntry(I.getParent()) &&
2609 "no unreachable blocks expected");
2610 ParsePointNeeded.push_back(cast<CallBase>(&I));
2614 // Return early if no work to do.
2615 if (ParsePointNeeded.empty())
2618 // As a prepass, go ahead and aggressively destroy single entry phi nodes.
2619 // These are created by LCSSA. They have the effect of increasing the size
2620 // of liveness sets for no good reason. It may be harder to do this post
2621 // insertion since relocations and base phis can confuse things.
2622 for (BasicBlock &BB : F)
2623 if (BB.getUniquePredecessor()) {
2625 FoldSingleEntryPHINodes(&BB);
2628 // Before we start introducing relocations, we want to tweak the IR a bit to
2629 // avoid unfortunate code generation effects. The main example is that we
2630 // want to try to make sure the comparison feeding a branch is after any
2631 // safepoints. Otherwise, we end up with a comparison of pre-relocation
2632 // values feeding a branch after relocation. This is semantically correct,
2633 // but results in extra register pressure since both the pre-relocation and
2634 // post-relocation copies must be available in registers. For code without
2635 // relocations this is handled elsewhere, but teaching the scheduler to
2636 // reverse the transform we're about to do would be slightly complex.
2637 // Note: This may extend the live range of the inputs to the icmp and thus
2638 // increase the liveset of any statepoint we move over. This is profitable
2639 // as long as all statepoints are in rare blocks. If we had in-register
2640 // lowering for live values this would be a much safer transform.
2641 auto getConditionInst = [](Instruction *TI) -> Instruction * {
2642 if (auto *BI = dyn_cast<BranchInst>(TI))
2643 if (BI->isConditional())
2644 return dyn_cast<Instruction>(BI->getCondition());
2645 // TODO: Extend this to handle switches
2648 for (BasicBlock &BB : F) {
2649 Instruction *TI = BB.getTerminator();
2650 if (auto *Cond = getConditionInst(TI))
2651 // TODO: Handle more than just ICmps here. We should be able to move
2652 // most instructions without side effects or memory access.
2653 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) {
2655 Cond->moveBefore(TI);
2659 // Nasty workaround - The base computation code in the main algorithm doesn't
2660 // consider the fact that a GEP can be used to convert a scalar to a vector.
2661 // The right fix for this is to integrate GEPs into the base rewriting
2662 // algorithm properly, this is just a short term workaround to prevent
2663 // crashes by canonicalizing such GEPs into fully vector GEPs.
2664 for (Instruction &I : instructions(F)) {
2665 if (!isa<GetElementPtrInst>(I))
2669 for (unsigned i = 0; i < I.getNumOperands(); i++)
2670 if (auto *OpndVTy = dyn_cast<VectorType>(I.getOperand(i)->getType())) {
2672 VF == cast<FixedVectorType>(OpndVTy)->getNumElements());
2673 VF = cast<FixedVectorType>(OpndVTy)->getNumElements();
2676 // It's the vector to scalar traversal through the pointer operand which
2677 // confuses base pointer rewriting, so limit ourselves to that case.
2678 if (!I.getOperand(0)->getType()->isVectorTy() && VF != 0) {
2680 auto *Splat = B.CreateVectorSplat(VF, I.getOperand(0));
2681 I.setOperand(0, Splat);
2686 MadeChange |= insertParsePoints(F, DT, TTI, ParsePointNeeded);
2690 // liveness computation via standard dataflow
2691 // -------------------------------------------------------------------
2693 // TODO: Consider using bitvectors for liveness, the set of potentially
2694 // interesting values should be small and easy to pre-compute.
2696 /// Compute the live-in set for the location rbegin starting from
2697 /// the live-out set of the basic block
2698 static void computeLiveInValues(BasicBlock::reverse_iterator Begin,
2699 BasicBlock::reverse_iterator End,
2700 SetVector<Value *> &LiveTmp) {
2701 for (auto &I : make_range(Begin, End)) {
2702 // KILL/Def - Remove this definition from LiveIn
2705 // Don't consider *uses* in PHI nodes, we handle their contribution to
2706 // predecessor blocks when we seed the LiveOut sets
2707 if (isa<PHINode>(I))
2710 // USE - Add to the LiveIn set for this instruction
2711 for (Value *V : I.operands()) {
2712 assert(!isUnhandledGCPointerType(V->getType()) &&
2713 "support for FCA unimplemented");
2714 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) {
2715 // The choice to exclude all things constant here is slightly subtle.
2716 // There are two independent reasons:
2717 // - We assume that things which are constant (from LLVM's definition)
2718 // do not move at runtime. For example, the address of a global
2719 // variable is fixed, even though it's contents may not be.
2720 // - Second, we can't disallow arbitrary inttoptr constants even
2721 // if the language frontend does. Optimization passes are free to
2722 // locally exploit facts without respect to global reachability. This
2723 // can create sections of code which are dynamically unreachable and
2724 // contain just about anything. (see constants.ll in tests)
2731 static void computeLiveOutSeed(BasicBlock *BB, SetVector<Value *> &LiveTmp) {
2732 for (BasicBlock *Succ : successors(BB)) {
2733 for (auto &I : *Succ) {
2734 PHINode *PN = dyn_cast<PHINode>(&I);
2738 Value *V = PN->getIncomingValueForBlock(BB);
2739 assert(!isUnhandledGCPointerType(V->getType()) &&
2740 "support for FCA unimplemented");
2741 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V))
2747 static SetVector<Value *> computeKillSet(BasicBlock *BB) {
2748 SetVector<Value *> KillSet;
2749 for (Instruction &I : *BB)
2750 if (isHandledGCPointerType(I.getType()))
2756 /// Check that the items in 'Live' dominate 'TI'. This is used as a basic
2757 /// sanity check for the liveness computation.
2758 static void checkBasicSSA(DominatorTree &DT, SetVector<Value *> &Live,
2759 Instruction *TI, bool TermOkay = false) {
2760 for (Value *V : Live) {
2761 if (auto *I = dyn_cast<Instruction>(V)) {
2762 // The terminator can be a member of the LiveOut set. LLVM's definition
2763 // of instruction dominance states that V does not dominate itself. As
2764 // such, we need to special case this to allow it.
2765 if (TermOkay && TI == I)
2767 assert(DT.dominates(I, TI) &&
2768 "basic SSA liveness expectation violated by liveness analysis");
2773 /// Check that all the liveness sets used during the computation of liveness
2774 /// obey basic SSA properties. This is useful for finding cases where we miss
2776 static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data,
2778 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator());
2779 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true);
2780 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator());
2784 static void computeLiveInValues(DominatorTree &DT, Function &F,
2785 GCPtrLivenessData &Data) {
2786 SmallSetVector<BasicBlock *, 32> Worklist;
2788 // Seed the liveness for each individual block
2789 for (BasicBlock &BB : F) {
2790 Data.KillSet[&BB] = computeKillSet(&BB);
2791 Data.LiveSet[&BB].clear();
2792 computeLiveInValues(BB.rbegin(), BB.rend(), Data.LiveSet[&BB]);
2795 for (Value *Kill : Data.KillSet[&BB])
2796 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill");
2799 Data.LiveOut[&BB] = SetVector<Value *>();
2800 computeLiveOutSeed(&BB, Data.LiveOut[&BB]);
2801 Data.LiveIn[&BB] = Data.LiveSet[&BB];
2802 Data.LiveIn[&BB].set_union(Data.LiveOut[&BB]);
2803 Data.LiveIn[&BB].set_subtract(Data.KillSet[&BB]);
2804 if (!Data.LiveIn[&BB].empty())
2805 Worklist.insert(pred_begin(&BB), pred_end(&BB));
2808 // Propagate that liveness until stable
2809 while (!Worklist.empty()) {
2810 BasicBlock *BB = Worklist.pop_back_val();
2812 // Compute our new liveout set, then exit early if it hasn't changed despite
2813 // the contribution of our successor.
2814 SetVector<Value *> LiveOut = Data.LiveOut[BB];
2815 const auto OldLiveOutSize = LiveOut.size();
2816 for (BasicBlock *Succ : successors(BB)) {
2817 assert(Data.LiveIn.count(Succ));
2818 LiveOut.set_union(Data.LiveIn[Succ]);
2820 // assert OutLiveOut is a subset of LiveOut
2821 if (OldLiveOutSize == LiveOut.size()) {
2822 // If the sets are the same size, then we didn't actually add anything
2823 // when unioning our successors LiveIn. Thus, the LiveIn of this block
2827 Data.LiveOut[BB] = LiveOut;
2829 // Apply the effects of this basic block
2830 SetVector<Value *> LiveTmp = LiveOut;
2831 LiveTmp.set_union(Data.LiveSet[BB]);
2832 LiveTmp.set_subtract(Data.KillSet[BB]);
2834 assert(Data.LiveIn.count(BB));
2835 const SetVector<Value *> &OldLiveIn = Data.LiveIn[BB];
2836 // assert: OldLiveIn is a subset of LiveTmp
2837 if (OldLiveIn.size() != LiveTmp.size()) {
2838 Data.LiveIn[BB] = LiveTmp;
2839 Worklist.insert(pred_begin(BB), pred_end(BB));
2841 } // while (!Worklist.empty())
2844 // Sanity check our output against SSA properties. This helps catch any
2845 // missing kills during the above iteration.
2846 for (BasicBlock &BB : F)
2847 checkBasicSSA(DT, Data, BB);
2851 static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data,
2852 StatepointLiveSetTy &Out) {
2853 BasicBlock *BB = Inst->getParent();
2855 // Note: The copy is intentional and required
2856 assert(Data.LiveOut.count(BB));
2857 SetVector<Value *> LiveOut = Data.LiveOut[BB];
2859 // We want to handle the statepoint itself oddly. It's
2860 // call result is not live (normal), nor are it's arguments
2861 // (unless they're used again later). This adjustment is
2862 // specifically what we need to relocate
2863 computeLiveInValues(BB->rbegin(), ++Inst->getIterator().getReverse(),
2865 LiveOut.remove(Inst);
2866 Out.insert(LiveOut.begin(), LiveOut.end());
2869 static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData,
2871 PartiallyConstructedSafepointRecord &Info) {
2872 StatepointLiveSetTy Updated;
2873 findLiveSetAtInst(Call, RevisedLivenessData, Updated);
2875 // We may have base pointers which are now live that weren't before. We need
2876 // to update the PointerToBase structure to reflect this.
2877 for (auto V : Updated)
2878 if (Info.PointerToBase.insert({V, V}).second) {
2879 assert(isKnownBaseResult(V) &&
2880 "Can't find base for unexpected live value!");
2885 for (auto V : Updated)
2886 assert(Info.PointerToBase.count(V) &&
2887 "Must be able to find base for live value!");
2890 // Remove any stale base mappings - this can happen since our liveness is
2891 // more precise then the one inherent in the base pointer analysis.
2892 DenseSet<Value *> ToErase;
2893 for (auto KVPair : Info.PointerToBase)
2894 if (!Updated.count(KVPair.first))
2895 ToErase.insert(KVPair.first);
2897 for (auto *V : ToErase)
2898 Info.PointerToBase.erase(V);
2901 for (auto KVPair : Info.PointerToBase)
2902 assert(Updated.count(KVPair.first) && "record for non-live value");
2905 Info.LiveSet = Updated;