1//===- RewriteStatepointsForGC.cpp - Make GC relocations explicit ---------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// Rewrite an existing set of gc.statepoints such that they make potential 11// relocations performed by the garbage collector explicit in the IR. 12// 13//===----------------------------------------------------------------------===// 14 15#include "llvm/Pass.h" 16#include "llvm/Analysis/CFG.h" 17#include "llvm/Analysis/InstructionSimplify.h" 18#include "llvm/Analysis/TargetTransformInfo.h" 19#include "llvm/ADT/SetOperations.h" 20#include "llvm/ADT/Statistic.h" 21#include "llvm/ADT/DenseSet.h" 22#include "llvm/ADT/SetVector.h" 23#include "llvm/ADT/StringRef.h" 24#include "llvm/ADT/MapVector.h" 25#include "llvm/IR/BasicBlock.h" 26#include "llvm/IR/CallSite.h" 27#include "llvm/IR/Dominators.h" 28#include "llvm/IR/Function.h" 29#include "llvm/IR/IRBuilder.h" 30#include "llvm/IR/InstIterator.h" 31#include "llvm/IR/Instructions.h" 32#include "llvm/IR/Intrinsics.h" 33#include "llvm/IR/IntrinsicInst.h" 34#include "llvm/IR/Module.h" 35#include "llvm/IR/MDBuilder.h" 36#include "llvm/IR/Statepoint.h" 37#include "llvm/IR/Value.h" 38#include "llvm/IR/Verifier.h" 39#include "llvm/Support/Debug.h" 40#include "llvm/Support/CommandLine.h" 41#include "llvm/Transforms/Scalar.h" 42#include "llvm/Transforms/Utils/BasicBlockUtils.h" 43#include "llvm/Transforms/Utils/Cloning.h" 44#include "llvm/Transforms/Utils/Local.h" 45#include "llvm/Transforms/Utils/PromoteMemToReg.h" 46 47#define DEBUG_TYPE "rewrite-statepoints-for-gc" 48 49using namespace llvm; 50 51// Print the liveset found at the insert location 52static cl::opt<bool> PrintLiveSet("spp-print-liveset", cl::Hidden, 53 cl::init(false)); 54static cl::opt<bool> PrintLiveSetSize("spp-print-liveset-size", cl::Hidden, 55 cl::init(false)); 56// Print out the base pointers for debugging 57static cl::opt<bool> PrintBasePointers("spp-print-base-pointers", cl::Hidden, 58 cl::init(false)); 59 60// Cost threshold measuring when it is profitable to rematerialize value instead 61// of relocating it 62static cl::opt<unsigned> 63RematerializationThreshold("spp-rematerialization-threshold", cl::Hidden, 64 cl::init(6)); 65 66#ifdef XDEBUG 67static bool ClobberNonLive = true; 68#else 69static bool ClobberNonLive = false; 70#endif 71static cl::opt<bool, true> ClobberNonLiveOverride("rs4gc-clobber-non-live", 72 cl::location(ClobberNonLive), 73 cl::Hidden); 74 75static cl::opt<bool> UseDeoptBundles("rs4gc-use-deopt-bundles", cl::Hidden, 76 cl::init(false)); 77static cl::opt<bool> 78 AllowStatepointWithNoDeoptInfo("rs4gc-allow-statepoint-with-no-deopt-info", 79 cl::Hidden, cl::init(true)); 80 81/// Should we split vectors of pointers into their individual elements? This 82/// is known to be buggy, but the alternate implementation isn't yet ready. 83/// This is purely to provide a debugging and dianostic hook until the vector 84/// split is replaced with vector relocations. 85static cl::opt<bool> UseVectorSplit("rs4gc-split-vector-values", cl::Hidden, 86 cl::init(true)); 87 88namespace { 89struct RewriteStatepointsForGC : public ModulePass { 90 static char ID; // Pass identification, replacement for typeid 91 92 RewriteStatepointsForGC() : ModulePass(ID) { 93 initializeRewriteStatepointsForGCPass(*PassRegistry::getPassRegistry()); 94 } 95 bool runOnFunction(Function &F); 96 bool runOnModule(Module &M) override { 97 bool Changed = false; 98 for (Function &F : M) 99 Changed |= runOnFunction(F); 100 101 if (Changed) { 102 // stripNonValidAttributes asserts that shouldRewriteStatepointsIn 103 // returns true for at least one function in the module. Since at least 104 // one function changed, we know that the precondition is satisfied. 105 stripNonValidAttributes(M); 106 } 107 108 return Changed; 109 } 110 111 void getAnalysisUsage(AnalysisUsage &AU) const override { 112 // We add and rewrite a bunch of instructions, but don't really do much 113 // else. We could in theory preserve a lot more analyses here. 114 AU.addRequired<DominatorTreeWrapperPass>(); 115 AU.addRequired<TargetTransformInfoWrapperPass>(); 116 } 117 118 /// The IR fed into RewriteStatepointsForGC may have had attributes implying 119 /// dereferenceability that are no longer valid/correct after 120 /// RewriteStatepointsForGC has run. This is because semantically, after 121 /// RewriteStatepointsForGC runs, all calls to gc.statepoint "free" the entire 122 /// heap. stripNonValidAttributes (conservatively) restores correctness 123 /// by erasing all attributes in the module that externally imply 124 /// dereferenceability. 125 /// Similar reasoning also applies to the noalias attributes. gc.statepoint 126 /// can touch the entire heap including noalias objects. 127 void stripNonValidAttributes(Module &M); 128 129 // Helpers for stripNonValidAttributes 130 void stripNonValidAttributesFromBody(Function &F); 131 void stripNonValidAttributesFromPrototype(Function &F); 132}; 133} // namespace 134 135char RewriteStatepointsForGC::ID = 0; 136 137ModulePass *llvm::createRewriteStatepointsForGCPass() { 138 return new RewriteStatepointsForGC(); 139} 140 141INITIALIZE_PASS_BEGIN(RewriteStatepointsForGC, "rewrite-statepoints-for-gc", 142 "Make relocations explicit at statepoints", false, false) 143INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 144INITIALIZE_PASS_END(RewriteStatepointsForGC, "rewrite-statepoints-for-gc", 145 "Make relocations explicit at statepoints", false, false) 146 147namespace { 148struct GCPtrLivenessData { 149 /// Values defined in this block. 150 DenseMap<BasicBlock *, DenseSet<Value *>> KillSet; 151 /// Values used in this block (and thus live); does not included values 152 /// killed within this block. 153 DenseMap<BasicBlock *, DenseSet<Value *>> LiveSet; 154 155 /// Values live into this basic block (i.e. used by any 156 /// instruction in this basic block or ones reachable from here) 157 DenseMap<BasicBlock *, DenseSet<Value *>> LiveIn; 158 159 /// Values live out of this basic block (i.e. live into 160 /// any successor block) 161 DenseMap<BasicBlock *, DenseSet<Value *>> LiveOut; 162}; 163 164// The type of the internal cache used inside the findBasePointers family 165// of functions. From the callers perspective, this is an opaque type and 166// should not be inspected. 167// 168// In the actual implementation this caches two relations: 169// - The base relation itself (i.e. this pointer is based on that one) 170// - The base defining value relation (i.e. before base_phi insertion) 171// Generally, after the execution of a full findBasePointer call, only the 172// base relation will remain. Internally, we add a mixture of the two 173// types, then update all the second type to the first type 174typedef DenseMap<Value *, Value *> DefiningValueMapTy; 175typedef DenseSet<Value *> StatepointLiveSetTy; 176typedef DenseMap<AssertingVH<Instruction>, AssertingVH<Value>> 177 RematerializedValueMapTy; 178 179struct PartiallyConstructedSafepointRecord { 180 /// The set of values known to be live across this safepoint 181 StatepointLiveSetTy LiveSet; 182 183 /// Mapping from live pointers to a base-defining-value 184 DenseMap<Value *, Value *> PointerToBase; 185 186 /// The *new* gc.statepoint instruction itself. This produces the token 187 /// that normal path gc.relocates and the gc.result are tied to. 188 Instruction *StatepointToken; 189 190 /// Instruction to which exceptional gc relocates are attached 191 /// Makes it easier to iterate through them during relocationViaAlloca. 192 Instruction *UnwindToken; 193 194 /// Record live values we are rematerialized instead of relocating. 195 /// They are not included into 'LiveSet' field. 196 /// Maps rematerialized copy to it's original value. 197 RematerializedValueMapTy RematerializedValues; 198}; 199} 200 201static ArrayRef<Use> GetDeoptBundleOperands(ImmutableCallSite CS) { 202 assert(UseDeoptBundles && "Should not be called otherwise!"); 203 204 Optional<OperandBundleUse> DeoptBundle = CS.getOperandBundle("deopt"); 205 206 if (!DeoptBundle.hasValue()) { 207 assert(AllowStatepointWithNoDeoptInfo && 208 "Found non-leaf call without deopt info!"); 209 return None; 210 } 211 212 return DeoptBundle.getValue().Inputs; 213} 214 215/// Compute the live-in set for every basic block in the function 216static void computeLiveInValues(DominatorTree &DT, Function &F, 217 GCPtrLivenessData &Data); 218 219/// Given results from the dataflow liveness computation, find the set of live 220/// Values at a particular instruction. 221static void findLiveSetAtInst(Instruction *inst, GCPtrLivenessData &Data, 222 StatepointLiveSetTy &out); 223 224// TODO: Once we can get to the GCStrategy, this becomes 225// Optional<bool> isGCManagedPointer(const Type *Ty) const override { 226 227static bool isGCPointerType(Type *T) { 228 if (auto *PT = dyn_cast<PointerType>(T)) 229 // For the sake of this example GC, we arbitrarily pick addrspace(1) as our 230 // GC managed heap. We know that a pointer into this heap needs to be 231 // updated and that no other pointer does. 232 return (1 == PT->getAddressSpace()); 233 return false; 234} 235 236// Return true if this type is one which a) is a gc pointer or contains a GC 237// pointer and b) is of a type this code expects to encounter as a live value. 238// (The insertion code will assert that a type which matches (a) and not (b) 239// is not encountered.) 240static bool isHandledGCPointerType(Type *T) { 241 // We fully support gc pointers 242 if (isGCPointerType(T)) 243 return true; 244 // We partially support vectors of gc pointers. The code will assert if it 245 // can't handle something. 246 if (auto VT = dyn_cast<VectorType>(T)) 247 if (isGCPointerType(VT->getElementType())) 248 return true; 249 return false; 250} 251 252#ifndef NDEBUG 253/// Returns true if this type contains a gc pointer whether we know how to 254/// handle that type or not. 255static bool containsGCPtrType(Type *Ty) { 256 if (isGCPointerType(Ty)) 257 return true; 258 if (VectorType *VT = dyn_cast<VectorType>(Ty)) 259 return isGCPointerType(VT->getScalarType()); 260 if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) 261 return containsGCPtrType(AT->getElementType()); 262 if (StructType *ST = dyn_cast<StructType>(Ty)) 263 return std::any_of(ST->subtypes().begin(), ST->subtypes().end(), 264 containsGCPtrType); 265 return false; 266} 267 268// Returns true if this is a type which a) is a gc pointer or contains a GC 269// pointer and b) is of a type which the code doesn't expect (i.e. first class 270// aggregates). Used to trip assertions. 271static bool isUnhandledGCPointerType(Type *Ty) { 272 return containsGCPtrType(Ty) && !isHandledGCPointerType(Ty); 273} 274#endif 275 276static bool order_by_name(Value *a, Value *b) { 277 if (a->hasName() && b->hasName()) { 278 return -1 == a->getName().compare(b->getName()); 279 } else if (a->hasName() && !b->hasName()) { 280 return true; 281 } else if (!a->hasName() && b->hasName()) { 282 return false; 283 } else { 284 // Better than nothing, but not stable 285 return a < b; 286 } 287} 288 289// Return the name of the value suffixed with the provided value, or if the 290// value didn't have a name, the default value specified. 291static std::string suffixed_name_or(Value *V, StringRef Suffix, 292 StringRef DefaultName) { 293 return V->hasName() ? (V->getName() + Suffix).str() : DefaultName.str(); 294} 295 296// Conservatively identifies any definitions which might be live at the 297// given instruction. The analysis is performed immediately before the 298// given instruction. Values defined by that instruction are not considered 299// live. Values used by that instruction are considered live. 300static void analyzeParsePointLiveness( 301 DominatorTree &DT, GCPtrLivenessData &OriginalLivenessData, 302 const CallSite &CS, PartiallyConstructedSafepointRecord &result) { 303 Instruction *inst = CS.getInstruction(); 304 305 StatepointLiveSetTy LiveSet; 306 findLiveSetAtInst(inst, OriginalLivenessData, LiveSet); 307 308 if (PrintLiveSet) { 309 // Note: This output is used by several of the test cases 310 // The order of elements in a set is not stable, put them in a vec and sort 311 // by name 312 SmallVector<Value *, 64> Temp; 313 Temp.insert(Temp.end(), LiveSet.begin(), LiveSet.end()); 314 std::sort(Temp.begin(), Temp.end(), order_by_name); 315 errs() << "Live Variables:\n"; 316 for (Value *V : Temp) 317 dbgs() << " " << V->getName() << " " << *V << "\n"; 318 } 319 if (PrintLiveSetSize) { 320 errs() << "Safepoint For: " << CS.getCalledValue()->getName() << "\n"; 321 errs() << "Number live values: " << LiveSet.size() << "\n"; 322 } 323 result.LiveSet = LiveSet; 324} 325 326static bool isKnownBaseResult(Value *V); 327namespace { 328/// A single base defining value - An immediate base defining value for an 329/// instruction 'Def' is an input to 'Def' whose base is also a base of 'Def'. 330/// For instructions which have multiple pointer [vector] inputs or that 331/// transition between vector and scalar types, there is no immediate base 332/// defining value. The 'base defining value' for 'Def' is the transitive 333/// closure of this relation stopping at the first instruction which has no 334/// immediate base defining value. The b.d.v. might itself be a base pointer, 335/// but it can also be an arbitrary derived pointer. 336struct BaseDefiningValueResult { 337 /// Contains the value which is the base defining value. 338 Value * const BDV; 339 /// True if the base defining value is also known to be an actual base 340 /// pointer. 341 const bool IsKnownBase; 342 BaseDefiningValueResult(Value *BDV, bool IsKnownBase) 343 : BDV(BDV), IsKnownBase(IsKnownBase) { 344#ifndef NDEBUG 345 // Check consistency between new and old means of checking whether a BDV is 346 // a base. 347 bool MustBeBase = isKnownBaseResult(BDV); 348 assert(!MustBeBase || MustBeBase == IsKnownBase); 349#endif 350 } 351}; 352} 353 354static BaseDefiningValueResult findBaseDefiningValue(Value *I); 355 356/// Return a base defining value for the 'Index' element of the given vector 357/// instruction 'I'. If Index is null, returns a BDV for the entire vector 358/// 'I'. As an optimization, this method will try to determine when the 359/// element is known to already be a base pointer. If this can be established, 360/// the second value in the returned pair will be true. Note that either a 361/// vector or a pointer typed value can be returned. For the former, the 362/// vector returned is a BDV (and possibly a base) of the entire vector 'I'. 363/// If the later, the return pointer is a BDV (or possibly a base) for the 364/// particular element in 'I'. 365static BaseDefiningValueResult 366findBaseDefiningValueOfVector(Value *I) { 367 // Each case parallels findBaseDefiningValue below, see that code for 368 // detailed motivation. 369 370 if (isa<Argument>(I)) 371 // An incoming argument to the function is a base pointer 372 return BaseDefiningValueResult(I, true); 373 374 if (isa<Constant>(I)) 375 // Constant vectors consist only of constant pointers. 376 return BaseDefiningValueResult(I, true); 377 378 if (isa<LoadInst>(I)) 379 return BaseDefiningValueResult(I, true); 380 381 if (isa<InsertElementInst>(I)) 382 // We don't know whether this vector contains entirely base pointers or 383 // not. To be conservatively correct, we treat it as a BDV and will 384 // duplicate code as needed to construct a parallel vector of bases. 385 return BaseDefiningValueResult(I, false); 386 387 if (isa<ShuffleVectorInst>(I)) 388 // We don't know whether this vector contains entirely base pointers or 389 // not. To be conservatively correct, we treat it as a BDV and will 390 // duplicate code as needed to construct a parallel vector of bases. 391 // TODO: There a number of local optimizations which could be applied here 392 // for particular sufflevector patterns. 393 return BaseDefiningValueResult(I, false); 394 395 // A PHI or Select is a base defining value. The outer findBasePointer 396 // algorithm is responsible for constructing a base value for this BDV. 397 assert((isa<SelectInst>(I) || isa<PHINode>(I)) && 398 "unknown vector instruction - no base found for vector element"); 399 return BaseDefiningValueResult(I, false); 400} 401 402/// Helper function for findBasePointer - Will return a value which either a) 403/// defines the base pointer for the input, b) blocks the simple search 404/// (i.e. a PHI or Select of two derived pointers), or c) involves a change 405/// from pointer to vector type or back. 406static BaseDefiningValueResult findBaseDefiningValue(Value *I) { 407 assert(I->getType()->isPtrOrPtrVectorTy() && 408 "Illegal to ask for the base pointer of a non-pointer type"); 409 410 if (I->getType()->isVectorTy()) 411 return findBaseDefiningValueOfVector(I); 412 413 if (isa<Argument>(I)) 414 // An incoming argument to the function is a base pointer 415 // We should have never reached here if this argument isn't an gc value 416 return BaseDefiningValueResult(I, true); 417 418 if (isa<Constant>(I)) 419 // We assume that objects with a constant base (e.g. a global) can't move 420 // and don't need to be reported to the collector because they are always 421 // live. All constants have constant bases. Besides global references, all 422 // kinds of constants (e.g. undef, constant expressions, null pointers) can 423 // be introduced by the inliner or the optimizer, especially on dynamically 424 // dead paths. See e.g. test4 in constants.ll. 425 return BaseDefiningValueResult(I, true); 426 427 if (CastInst *CI = dyn_cast<CastInst>(I)) { 428 Value *Def = CI->stripPointerCasts(); 429 // If stripping pointer casts changes the address space there is an 430 // addrspacecast in between. 431 assert(cast<PointerType>(Def->getType())->getAddressSpace() == 432 cast<PointerType>(CI->getType())->getAddressSpace() && 433 "unsupported addrspacecast"); 434 // If we find a cast instruction here, it means we've found a cast which is 435 // not simply a pointer cast (i.e. an inttoptr). We don't know how to 436 // handle int->ptr conversion. 437 assert(!isa<CastInst>(Def) && "shouldn't find another cast here"); 438 return findBaseDefiningValue(Def); 439 } 440 441 if (isa<LoadInst>(I)) 442 // The value loaded is an gc base itself 443 return BaseDefiningValueResult(I, true); 444 445 446 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) 447 // The base of this GEP is the base 448 return findBaseDefiningValue(GEP->getPointerOperand()); 449 450 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 451 switch (II->getIntrinsicID()) { 452 default: 453 // fall through to general call handling 454 break; 455 case Intrinsic::experimental_gc_statepoint: 456 llvm_unreachable("statepoints don't produce pointers"); 457 case Intrinsic::experimental_gc_relocate: { 458 // Rerunning safepoint insertion after safepoints are already 459 // inserted is not supported. It could probably be made to work, 460 // but why are you doing this? There's no good reason. 461 llvm_unreachable("repeat safepoint insertion is not supported"); 462 } 463 case Intrinsic::gcroot: 464 // Currently, this mechanism hasn't been extended to work with gcroot. 465 // There's no reason it couldn't be, but I haven't thought about the 466 // implications much. 467 llvm_unreachable( 468 "interaction with the gcroot mechanism is not supported"); 469 } 470 } 471 // We assume that functions in the source language only return base 472 // pointers. This should probably be generalized via attributes to support 473 // both source language and internal functions. 474 if (isa<CallInst>(I) || isa<InvokeInst>(I)) 475 return BaseDefiningValueResult(I, true); 476 477 // I have absolutely no idea how to implement this part yet. It's not 478 // necessarily hard, I just haven't really looked at it yet. 479 assert(!isa<LandingPadInst>(I) && "Landing Pad is unimplemented"); 480 481 if (isa<AtomicCmpXchgInst>(I)) 482 // A CAS is effectively a atomic store and load combined under a 483 // predicate. From the perspective of base pointers, we just treat it 484 // like a load. 485 return BaseDefiningValueResult(I, true); 486 487 assert(!isa<AtomicRMWInst>(I) && "Xchg handled above, all others are " 488 "binary ops which don't apply to pointers"); 489 490 // The aggregate ops. Aggregates can either be in the heap or on the 491 // stack, but in either case, this is simply a field load. As a result, 492 // this is a defining definition of the base just like a load is. 493 if (isa<ExtractValueInst>(I)) 494 return BaseDefiningValueResult(I, true); 495 496 // We should never see an insert vector since that would require we be 497 // tracing back a struct value not a pointer value. 498 assert(!isa<InsertValueInst>(I) && 499 "Base pointer for a struct is meaningless"); 500 501 // An extractelement produces a base result exactly when it's input does. 502 // We may need to insert a parallel instruction to extract the appropriate 503 // element out of the base vector corresponding to the input. Given this, 504 // it's analogous to the phi and select case even though it's not a merge. 505 if (isa<ExtractElementInst>(I)) 506 // Note: There a lot of obvious peephole cases here. This are deliberately 507 // handled after the main base pointer inference algorithm to make writing 508 // test cases to exercise that code easier. 509 return BaseDefiningValueResult(I, false); 510 511 // The last two cases here don't return a base pointer. Instead, they 512 // return a value which dynamically selects from among several base 513 // derived pointers (each with it's own base potentially). It's the job of 514 // the caller to resolve these. 515 assert((isa<SelectInst>(I) || isa<PHINode>(I)) && 516 "missing instruction case in findBaseDefiningValing"); 517 return BaseDefiningValueResult(I, false); 518} 519 520/// Returns the base defining value for this value. 521static Value *findBaseDefiningValueCached(Value *I, DefiningValueMapTy &Cache) { 522 Value *&Cached = Cache[I]; 523 if (!Cached) { 524 Cached = findBaseDefiningValue(I).BDV; 525 DEBUG(dbgs() << "fBDV-cached: " << I->getName() << " -> " 526 << Cached->getName() << "\n"); 527 } 528 assert(Cache[I] != nullptr); 529 return Cached; 530} 531 532/// Return a base pointer for this value if known. Otherwise, return it's 533/// base defining value. 534static Value *findBaseOrBDV(Value *I, DefiningValueMapTy &Cache) { 535 Value *Def = findBaseDefiningValueCached(I, Cache); 536 auto Found = Cache.find(Def); 537 if (Found != Cache.end()) { 538 // Either a base-of relation, or a self reference. Caller must check. 539 return Found->second; 540 } 541 // Only a BDV available 542 return Def; 543} 544 545/// Given the result of a call to findBaseDefiningValue, or findBaseOrBDV, 546/// is it known to be a base pointer? Or do we need to continue searching. 547static bool isKnownBaseResult(Value *V) { 548 if (!isa<PHINode>(V) && !isa<SelectInst>(V) && 549 !isa<ExtractElementInst>(V) && !isa<InsertElementInst>(V) && 550 !isa<ShuffleVectorInst>(V)) { 551 // no recursion possible 552 return true; 553 } 554 if (isa<Instruction>(V) && 555 cast<Instruction>(V)->getMetadata("is_base_value")) { 556 // This is a previously inserted base phi or select. We know 557 // that this is a base value. 558 return true; 559 } 560 561 // We need to keep searching 562 return false; 563} 564 565namespace { 566/// Models the state of a single base defining value in the findBasePointer 567/// algorithm for determining where a new instruction is needed to propagate 568/// the base of this BDV. 569class BDVState { 570public: 571 enum Status { Unknown, Base, Conflict }; 572 573 BDVState(Status s, Value *b = nullptr) : status(s), base(b) { 574 assert(status != Base || b); 575 } 576 explicit BDVState(Value *b) : status(Base), base(b) {} 577 BDVState() : status(Unknown), base(nullptr) {} 578 579 Status getStatus() const { return status; } 580 Value *getBase() const { return base; } 581 582 bool isBase() const { return getStatus() == Base; } 583 bool isUnknown() const { return getStatus() == Unknown; } 584 bool isConflict() const { return getStatus() == Conflict; } 585 586 bool operator==(const BDVState &other) const { 587 return base == other.base && status == other.status; 588 } 589 590 bool operator!=(const BDVState &other) const { return !(*this == other); } 591 592 LLVM_DUMP_METHOD 593 void dump() const { print(dbgs()); dbgs() << '\n'; } 594 595 void print(raw_ostream &OS) const { 596 switch (status) { 597 case Unknown: 598 OS << "U"; 599 break; 600 case Base: 601 OS << "B"; 602 break; 603 case Conflict: 604 OS << "C"; 605 break; 606 }; 607 OS << " (" << base << " - " 608 << (base ? base->getName() : "nullptr") << "): "; 609 } 610 611private: 612 Status status; 613 AssertingVH<Value> base; // non null only if status == base 614}; 615} 616 617#ifndef NDEBUG 618static raw_ostream &operator<<(raw_ostream &OS, const BDVState &State) { 619 State.print(OS); 620 return OS; 621} 622#endif 623 624namespace { 625// Values of type BDVState form a lattice, and this is a helper 626// class that implementes the meet operation. The meat of the meet 627// operation is implemented in MeetBDVStates::pureMeet 628class MeetBDVStates { 629public: 630 /// Initializes the currentResult to the TOP state so that if can be met with 631 /// any other state to produce that state. 632 MeetBDVStates() {} 633 634 // Destructively meet the current result with the given BDVState 635 void meetWith(BDVState otherState) { 636 currentResult = meet(otherState, currentResult); 637 } 638 639 BDVState getResult() const { return currentResult; } 640 641private: 642 BDVState currentResult; 643 644 /// Perform a meet operation on two elements of the BDVState lattice. 645 static BDVState meet(BDVState LHS, BDVState RHS) { 646 assert((pureMeet(LHS, RHS) == pureMeet(RHS, LHS)) && 647 "math is wrong: meet does not commute!"); 648 BDVState Result = pureMeet(LHS, RHS); 649 DEBUG(dbgs() << "meet of " << LHS << " with " << RHS 650 << " produced " << Result << "\n"); 651 return Result; 652 } 653 654 static BDVState pureMeet(const BDVState &stateA, const BDVState &stateB) { 655 switch (stateA.getStatus()) { 656 case BDVState::Unknown: 657 return stateB; 658 659 case BDVState::Base: 660 assert(stateA.getBase() && "can't be null"); 661 if (stateB.isUnknown()) 662 return stateA; 663 664 if (stateB.isBase()) { 665 if (stateA.getBase() == stateB.getBase()) { 666 assert(stateA == stateB && "equality broken!"); 667 return stateA; 668 } 669 return BDVState(BDVState::Conflict); 670 } 671 assert(stateB.isConflict() && "only three states!"); 672 return BDVState(BDVState::Conflict); 673 674 case BDVState::Conflict: 675 return stateA; 676 } 677 llvm_unreachable("only three states!"); 678 } 679}; 680} 681 682 683/// For a given value or instruction, figure out what base ptr it's derived 684/// from. For gc objects, this is simply itself. On success, returns a value 685/// which is the base pointer. (This is reliable and can be used for 686/// relocation.) On failure, returns nullptr. 687static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) { 688 Value *def = findBaseOrBDV(I, cache); 689 690 if (isKnownBaseResult(def)) { 691 return def; 692 } 693 694 // Here's the rough algorithm: 695 // - For every SSA value, construct a mapping to either an actual base 696 // pointer or a PHI which obscures the base pointer. 697 // - Construct a mapping from PHI to unknown TOP state. Use an 698 // optimistic algorithm to propagate base pointer information. Lattice 699 // looks like: 700 // UNKNOWN 701 // b1 b2 b3 b4 702 // CONFLICT 703 // When algorithm terminates, all PHIs will either have a single concrete 704 // base or be in a conflict state. 705 // - For every conflict, insert a dummy PHI node without arguments. Add 706 // these to the base[Instruction] = BasePtr mapping. For every 707 // non-conflict, add the actual base. 708 // - For every conflict, add arguments for the base[a] of each input 709 // arguments. 710 // 711 // Note: A simpler form of this would be to add the conflict form of all 712 // PHIs without running the optimistic algorithm. This would be 713 // analogous to pessimistic data flow and would likely lead to an 714 // overall worse solution. 715 716#ifndef NDEBUG 717 auto isExpectedBDVType = [](Value *BDV) { 718 return isa<PHINode>(BDV) || isa<SelectInst>(BDV) || 719 isa<ExtractElementInst>(BDV) || isa<InsertElementInst>(BDV); 720 }; 721#endif 722 723 // Once populated, will contain a mapping from each potentially non-base BDV 724 // to a lattice value (described above) which corresponds to that BDV. 725 // We use the order of insertion (DFS over the def/use graph) to provide a 726 // stable deterministic ordering for visiting DenseMaps (which are unordered) 727 // below. This is important for deterministic compilation. 728 MapVector<Value *, BDVState> States; 729 730 // Recursively fill in all base defining values reachable from the initial 731 // one for which we don't already know a definite base value for 732 /* scope */ { 733 SmallVector<Value*, 16> Worklist; 734 Worklist.push_back(def); 735 States.insert(std::make_pair(def, BDVState())); 736 while (!Worklist.empty()) { 737 Value *Current = Worklist.pop_back_val(); 738 assert(!isKnownBaseResult(Current) && "why did it get added?"); 739 740 auto visitIncomingValue = [&](Value *InVal) { 741 Value *Base = findBaseOrBDV(InVal, cache); 742 if (isKnownBaseResult(Base)) 743 // Known bases won't need new instructions introduced and can be 744 // ignored safely 745 return; 746 assert(isExpectedBDVType(Base) && "the only non-base values " 747 "we see should be base defining values"); 748 if (States.insert(std::make_pair(Base, BDVState())).second) 749 Worklist.push_back(Base); 750 }; 751 if (PHINode *Phi = dyn_cast<PHINode>(Current)) { 752 for (Value *InVal : Phi->incoming_values()) 753 visitIncomingValue(InVal); 754 } else if (SelectInst *Sel = dyn_cast<SelectInst>(Current)) { 755 visitIncomingValue(Sel->getTrueValue()); 756 visitIncomingValue(Sel->getFalseValue()); 757 } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) { 758 visitIncomingValue(EE->getVectorOperand()); 759 } else if (auto *IE = dyn_cast<InsertElementInst>(Current)) { 760 visitIncomingValue(IE->getOperand(0)); // vector operand 761 visitIncomingValue(IE->getOperand(1)); // scalar operand 762 } else { 763 // There is one known class of instructions we know we don't handle. 764 assert(isa<ShuffleVectorInst>(Current)); 765 llvm_unreachable("unimplemented instruction case"); 766 } 767 } 768 } 769 770#ifndef NDEBUG 771 DEBUG(dbgs() << "States after initialization:\n"); 772 for (auto Pair : States) { 773 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n"); 774 } 775#endif 776 777 // Return a phi state for a base defining value. We'll generate a new 778 // base state for known bases and expect to find a cached state otherwise. 779 auto getStateForBDV = [&](Value *baseValue) { 780 if (isKnownBaseResult(baseValue)) 781 return BDVState(baseValue); 782 auto I = States.find(baseValue); 783 assert(I != States.end() && "lookup failed!"); 784 return I->second; 785 }; 786 787 bool progress = true; 788 while (progress) { 789#ifndef NDEBUG 790 const size_t oldSize = States.size(); 791#endif 792 progress = false; 793 // We're only changing values in this loop, thus safe to keep iterators. 794 // Since this is computing a fixed point, the order of visit does not 795 // effect the result. TODO: We could use a worklist here and make this run 796 // much faster. 797 for (auto Pair : States) { 798 Value *BDV = Pair.first; 799 assert(!isKnownBaseResult(BDV) && "why did it get added?"); 800 801 // Given an input value for the current instruction, return a BDVState 802 // instance which represents the BDV of that value. 803 auto getStateForInput = [&](Value *V) mutable { 804 Value *BDV = findBaseOrBDV(V, cache); 805 return getStateForBDV(BDV); 806 }; 807 808 MeetBDVStates calculateMeet; 809 if (SelectInst *select = dyn_cast<SelectInst>(BDV)) { 810 calculateMeet.meetWith(getStateForInput(select->getTrueValue())); 811 calculateMeet.meetWith(getStateForInput(select->getFalseValue())); 812 } else if (PHINode *Phi = dyn_cast<PHINode>(BDV)) { 813 for (Value *Val : Phi->incoming_values()) 814 calculateMeet.meetWith(getStateForInput(Val)); 815 } else if (auto *EE = dyn_cast<ExtractElementInst>(BDV)) { 816 // The 'meet' for an extractelement is slightly trivial, but it's still 817 // useful in that it drives us to conflict if our input is. 818 calculateMeet.meetWith(getStateForInput(EE->getVectorOperand())); 819 } else { 820 // Given there's a inherent type mismatch between the operands, will 821 // *always* produce Conflict. 822 auto *IE = cast<InsertElementInst>(BDV); 823 calculateMeet.meetWith(getStateForInput(IE->getOperand(0))); 824 calculateMeet.meetWith(getStateForInput(IE->getOperand(1))); 825 } 826 827 BDVState oldState = States[BDV]; 828 BDVState newState = calculateMeet.getResult(); 829 if (oldState != newState) { 830 progress = true; 831 States[BDV] = newState; 832 } 833 } 834 835 assert(oldSize == States.size() && 836 "fixed point shouldn't be adding any new nodes to state"); 837 } 838 839#ifndef NDEBUG 840 DEBUG(dbgs() << "States after meet iteration:\n"); 841 for (auto Pair : States) { 842 DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n"); 843 } 844#endif 845 846 // Insert Phis for all conflicts 847 // TODO: adjust naming patterns to avoid this order of iteration dependency 848 for (auto Pair : States) { 849 Instruction *I = cast<Instruction>(Pair.first); 850 BDVState State = Pair.second; 851 assert(!isKnownBaseResult(I) && "why did it get added?"); 852 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!"); 853 854 // extractelement instructions are a bit special in that we may need to 855 // insert an extract even when we know an exact base for the instruction. 856 // The problem is that we need to convert from a vector base to a scalar 857 // base for the particular indice we're interested in. 858 if (State.isBase() && isa<ExtractElementInst>(I) && 859 isa<VectorType>(State.getBase()->getType())) { 860 auto *EE = cast<ExtractElementInst>(I); 861 // TODO: In many cases, the new instruction is just EE itself. We should 862 // exploit this, but can't do it here since it would break the invariant 863 // about the BDV not being known to be a base. 864 auto *BaseInst = ExtractElementInst::Create(State.getBase(), 865 EE->getIndexOperand(), 866 "base_ee", EE); 867 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {})); 868 States[I] = BDVState(BDVState::Base, BaseInst); 869 } 870 871 // Since we're joining a vector and scalar base, they can never be the 872 // same. As a result, we should always see insert element having reached 873 // the conflict state. 874 if (isa<InsertElementInst>(I)) { 875 assert(State.isConflict()); 876 } 877 878 if (!State.isConflict()) 879 continue; 880 881 /// Create and insert a new instruction which will represent the base of 882 /// the given instruction 'I'. 883 auto MakeBaseInstPlaceholder = [](Instruction *I) -> Instruction* { 884 if (isa<PHINode>(I)) { 885 BasicBlock *BB = I->getParent(); 886 int NumPreds = std::distance(pred_begin(BB), pred_end(BB)); 887 assert(NumPreds > 0 && "how did we reach here"); 888 std::string Name = suffixed_name_or(I, ".base", "base_phi"); 889 return PHINode::Create(I->getType(), NumPreds, Name, I); 890 } else if (SelectInst *Sel = dyn_cast<SelectInst>(I)) { 891 // The undef will be replaced later 892 UndefValue *Undef = UndefValue::get(Sel->getType()); 893 std::string Name = suffixed_name_or(I, ".base", "base_select"); 894 return SelectInst::Create(Sel->getCondition(), Undef, 895 Undef, Name, Sel); 896 } else if (auto *EE = dyn_cast<ExtractElementInst>(I)) { 897 UndefValue *Undef = UndefValue::get(EE->getVectorOperand()->getType()); 898 std::string Name = suffixed_name_or(I, ".base", "base_ee"); 899 return ExtractElementInst::Create(Undef, EE->getIndexOperand(), Name, 900 EE); 901 } else { 902 auto *IE = cast<InsertElementInst>(I); 903 UndefValue *VecUndef = UndefValue::get(IE->getOperand(0)->getType()); 904 UndefValue *ScalarUndef = UndefValue::get(IE->getOperand(1)->getType()); 905 std::string Name = suffixed_name_or(I, ".base", "base_ie"); 906 return InsertElementInst::Create(VecUndef, ScalarUndef, 907 IE->getOperand(2), Name, IE); 908 } 909 910 }; 911 Instruction *BaseInst = MakeBaseInstPlaceholder(I); 912 // Add metadata marking this as a base value 913 BaseInst->setMetadata("is_base_value", MDNode::get(I->getContext(), {})); 914 States[I] = BDVState(BDVState::Conflict, BaseInst); 915 } 916 917 // Returns a instruction which produces the base pointer for a given 918 // instruction. The instruction is assumed to be an input to one of the BDVs 919 // seen in the inference algorithm above. As such, we must either already 920 // know it's base defining value is a base, or have inserted a new 921 // instruction to propagate the base of it's BDV and have entered that newly 922 // introduced instruction into the state table. In either case, we are 923 // assured to be able to determine an instruction which produces it's base 924 // pointer. 925 auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) { 926 Value *BDV = findBaseOrBDV(Input, cache); 927 Value *Base = nullptr; 928 if (isKnownBaseResult(BDV)) { 929 Base = BDV; 930 } else { 931 // Either conflict or base. 932 assert(States.count(BDV)); 933 Base = States[BDV].getBase(); 934 } 935 assert(Base && "can't be null"); 936 // The cast is needed since base traversal may strip away bitcasts 937 if (Base->getType() != Input->getType() && 938 InsertPt) { 939 Base = new BitCastInst(Base, Input->getType(), "cast", 940 InsertPt); 941 } 942 return Base; 943 }; 944 945 // Fixup all the inputs of the new PHIs. Visit order needs to be 946 // deterministic and predictable because we're naming newly created 947 // instructions. 948 for (auto Pair : States) { 949 Instruction *BDV = cast<Instruction>(Pair.first); 950 BDVState State = Pair.second; 951 952 assert(!isKnownBaseResult(BDV) && "why did it get added?"); 953 assert(!State.isUnknown() && "Optimistic algorithm didn't complete!"); 954 if (!State.isConflict()) 955 continue; 956 957 if (PHINode *basephi = dyn_cast<PHINode>(State.getBase())) { 958 PHINode *phi = cast<PHINode>(BDV); 959 unsigned NumPHIValues = phi->getNumIncomingValues(); 960 for (unsigned i = 0; i < NumPHIValues; i++) { 961 Value *InVal = phi->getIncomingValue(i); 962 BasicBlock *InBB = phi->getIncomingBlock(i); 963 964 // If we've already seen InBB, add the same incoming value 965 // we added for it earlier. The IR verifier requires phi 966 // nodes with multiple entries from the same basic block 967 // to have the same incoming value for each of those 968 // entries. If we don't do this check here and basephi 969 // has a different type than base, we'll end up adding two 970 // bitcasts (and hence two distinct values) as incoming 971 // values for the same basic block. 972 973 int blockIndex = basephi->getBasicBlockIndex(InBB); 974 if (blockIndex != -1) { 975 Value *oldBase = basephi->getIncomingValue(blockIndex); 976 basephi->addIncoming(oldBase, InBB); 977 978#ifndef NDEBUG 979 Value *Base = getBaseForInput(InVal, nullptr); 980 // In essence this assert states: the only way two 981 // values incoming from the same basic block may be 982 // different is by being different bitcasts of the same 983 // value. A cleanup that remains TODO is changing 984 // findBaseOrBDV to return an llvm::Value of the correct 985 // type (and still remain pure). This will remove the 986 // need to add bitcasts. 987 assert(Base->stripPointerCasts() == oldBase->stripPointerCasts() && 988 "sanity -- findBaseOrBDV should be pure!"); 989#endif 990 continue; 991 } 992 993 // Find the instruction which produces the base for each input. We may 994 // need to insert a bitcast in the incoming block. 995 // TODO: Need to split critical edges if insertion is needed 996 Value *Base = getBaseForInput(InVal, InBB->getTerminator()); 997 basephi->addIncoming(Base, InBB); 998 } 999 assert(basephi->getNumIncomingValues() == NumPHIValues); 1000 } else if (SelectInst *BaseSel = dyn_cast<SelectInst>(State.getBase())) { 1001 SelectInst *Sel = cast<SelectInst>(BDV); 1002 // Operand 1 & 2 are true, false path respectively. TODO: refactor to 1003 // something more safe and less hacky. 1004 for (int i = 1; i <= 2; i++) { 1005 Value *InVal = Sel->getOperand(i); 1006 // Find the instruction which produces the base for each input. We may 1007 // need to insert a bitcast. 1008 Value *Base = getBaseForInput(InVal, BaseSel); 1009 BaseSel->setOperand(i, Base); 1010 } 1011 } else if (auto *BaseEE = dyn_cast<ExtractElementInst>(State.getBase())) { 1012 Value *InVal = cast<ExtractElementInst>(BDV)->getVectorOperand(); 1013 // Find the instruction which produces the base for each input. We may 1014 // need to insert a bitcast. 1015 Value *Base = getBaseForInput(InVal, BaseEE); 1016 BaseEE->setOperand(0, Base); 1017 } else { 1018 auto *BaseIE = cast<InsertElementInst>(State.getBase()); 1019 auto *BdvIE = cast<InsertElementInst>(BDV); 1020 auto UpdateOperand = [&](int OperandIdx) { 1021 Value *InVal = BdvIE->getOperand(OperandIdx); 1022 Value *Base = getBaseForInput(InVal, BaseIE); 1023 BaseIE->setOperand(OperandIdx, Base); 1024 }; 1025 UpdateOperand(0); // vector operand 1026 UpdateOperand(1); // scalar operand 1027 } 1028 1029 } 1030 1031 // Now that we're done with the algorithm, see if we can optimize the 1032 // results slightly by reducing the number of new instructions needed. 1033 // Arguably, this should be integrated into the algorithm above, but 1034 // doing as a post process step is easier to reason about for the moment. 1035 DenseMap<Value *, Value *> ReverseMap; 1036 SmallPtrSet<Instruction *, 16> NewInsts; 1037 SmallSetVector<AssertingVH<Instruction>, 16> Worklist; 1038 // Note: We need to visit the states in a deterministic order. We uses the 1039 // Keys we sorted above for this purpose. Note that we are papering over a 1040 // bigger problem with the algorithm above - it's visit order is not 1041 // deterministic. A larger change is needed to fix this. 1042 for (auto Pair : States) { 1043 auto *BDV = Pair.first; 1044 auto State = Pair.second; 1045 Value *Base = State.getBase(); 1046 assert(BDV && Base); 1047 assert(!isKnownBaseResult(BDV) && "why did it get added?"); 1048 assert(isKnownBaseResult(Base) && 1049 "must be something we 'know' is a base pointer"); 1050 if (!State.isConflict()) 1051 continue; 1052 1053 ReverseMap[Base] = BDV; 1054 if (auto *BaseI = dyn_cast<Instruction>(Base)) { 1055 NewInsts.insert(BaseI); 1056 Worklist.insert(BaseI); 1057 } 1058 } 1059 auto ReplaceBaseInstWith = [&](Value *BDV, Instruction *BaseI, 1060 Value *Replacement) { 1061 // Add users which are new instructions (excluding self references) 1062 for (User *U : BaseI->users()) 1063 if (auto *UI = dyn_cast<Instruction>(U)) 1064 if (NewInsts.count(UI) && UI != BaseI) 1065 Worklist.insert(UI); 1066 // Then do the actual replacement 1067 NewInsts.erase(BaseI); 1068 ReverseMap.erase(BaseI); 1069 BaseI->replaceAllUsesWith(Replacement); 1070 assert(States.count(BDV)); 1071 assert(States[BDV].isConflict() && States[BDV].getBase() == BaseI); 1072 States[BDV] = BDVState(BDVState::Conflict, Replacement); 1073 BaseI->eraseFromParent(); 1074 }; 1075 const DataLayout &DL = cast<Instruction>(def)->getModule()->getDataLayout(); 1076 while (!Worklist.empty()) { 1077 Instruction *BaseI = Worklist.pop_back_val(); 1078 assert(NewInsts.count(BaseI)); 1079 Value *Bdv = ReverseMap[BaseI]; 1080 if (auto *BdvI = dyn_cast<Instruction>(Bdv)) 1081 if (BaseI->isIdenticalTo(BdvI)) { 1082 DEBUG(dbgs() << "Identical Base: " << *BaseI << "\n"); 1083 ReplaceBaseInstWith(Bdv, BaseI, Bdv); 1084 continue; 1085 } 1086 if (Value *V = SimplifyInstruction(BaseI, DL)) { 1087 DEBUG(dbgs() << "Base " << *BaseI << " simplified to " << *V << "\n"); 1088 ReplaceBaseInstWith(Bdv, BaseI, V); 1089 continue; 1090 } 1091 } 1092 1093 // Cache all of our results so we can cheaply reuse them 1094 // NOTE: This is actually two caches: one of the base defining value 1095 // relation and one of the base pointer relation! FIXME 1096 for (auto Pair : States) { 1097 auto *BDV = Pair.first; 1098 Value *base = Pair.second.getBase(); 1099 assert(BDV && base); 1100 1101 std::string fromstr = cache.count(BDV) ? cache[BDV]->getName() : "none"; 1102 DEBUG(dbgs() << "Updating base value cache" 1103 << " for: " << BDV->getName() 1104 << " from: " << fromstr 1105 << " to: " << base->getName() << "\n"); 1106 1107 if (cache.count(BDV)) { 1108 // Once we transition from the BDV relation being store in the cache to 1109 // the base relation being stored, it must be stable 1110 assert((!isKnownBaseResult(cache[BDV]) || cache[BDV] == base) && 1111 "base relation should be stable"); 1112 } 1113 cache[BDV] = base; 1114 } 1115 assert(cache.count(def)); 1116 return cache[def]; 1117} 1118 1119// For a set of live pointers (base and/or derived), identify the base 1120// pointer of the object which they are derived from. This routine will 1121// mutate the IR graph as needed to make the 'base' pointer live at the 1122// definition site of 'derived'. This ensures that any use of 'derived' can 1123// also use 'base'. This may involve the insertion of a number of 1124// additional PHI nodes. 1125// 1126// preconditions: live is a set of pointer type Values 1127// 1128// side effects: may insert PHI nodes into the existing CFG, will preserve 1129// CFG, will not remove or mutate any existing nodes 1130// 1131// post condition: PointerToBase contains one (derived, base) pair for every 1132// pointer in live. Note that derived can be equal to base if the original 1133// pointer was a base pointer. 1134static void 1135findBasePointers(const StatepointLiveSetTy &live, 1136 DenseMap<Value *, Value *> &PointerToBase, 1137 DominatorTree *DT, DefiningValueMapTy &DVCache) { 1138 // For the naming of values inserted to be deterministic - which makes for 1139 // much cleaner and more stable tests - we need to assign an order to the 1140 // live values. DenseSets do not provide a deterministic order across runs. 1141 SmallVector<Value *, 64> Temp; 1142 Temp.insert(Temp.end(), live.begin(), live.end()); 1143 std::sort(Temp.begin(), Temp.end(), order_by_name); 1144 for (Value *ptr : Temp) { 1145 Value *base = findBasePointer(ptr, DVCache); 1146 assert(base && "failed to find base pointer"); 1147 PointerToBase[ptr] = base; 1148 assert((!isa<Instruction>(base) || !isa<Instruction>(ptr) || 1149 DT->dominates(cast<Instruction>(base)->getParent(), 1150 cast<Instruction>(ptr)->getParent())) && 1151 "The base we found better dominate the derived pointer"); 1152 1153 // If you see this trip and like to live really dangerously, the code should 1154 // be correct, just with idioms the verifier can't handle. You can try 1155 // disabling the verifier at your own substantial risk. 1156 assert(!isa<ConstantPointerNull>(base) && 1157 "the relocation code needs adjustment to handle the relocation of " 1158 "a null pointer constant without causing false positives in the " 1159 "safepoint ir verifier."); 1160 } 1161} 1162 1163/// Find the required based pointers (and adjust the live set) for the given 1164/// parse point. 1165static void findBasePointers(DominatorTree &DT, DefiningValueMapTy &DVCache, 1166 const CallSite &CS, 1167 PartiallyConstructedSafepointRecord &result) { 1168 DenseMap<Value *, Value *> PointerToBase; 1169 findBasePointers(result.LiveSet, PointerToBase, &DT, DVCache); 1170 1171 if (PrintBasePointers) { 1172 // Note: Need to print these in a stable order since this is checked in 1173 // some tests. 1174 errs() << "Base Pairs (w/o Relocation):\n"; 1175 SmallVector<Value *, 64> Temp; 1176 Temp.reserve(PointerToBase.size()); 1177 for (auto Pair : PointerToBase) { 1178 Temp.push_back(Pair.first); 1179 } 1180 std::sort(Temp.begin(), Temp.end(), order_by_name); 1181 for (Value *Ptr : Temp) { 1182 Value *Base = PointerToBase[Ptr]; 1183 errs() << " derived "; 1184 Ptr->printAsOperand(errs(), false); 1185 errs() << " base "; 1186 Base->printAsOperand(errs(), false); 1187 errs() << "\n";; 1188 } 1189 } 1190 1191 result.PointerToBase = PointerToBase; 1192} 1193 1194/// Given an updated version of the dataflow liveness results, update the 1195/// liveset and base pointer maps for the call site CS. 1196static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData, 1197 const CallSite &CS, 1198 PartiallyConstructedSafepointRecord &result); 1199 1200static void recomputeLiveInValues( 1201 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate, 1202 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) { 1203 // TODO-PERF: reuse the original liveness, then simply run the dataflow 1204 // again. The old values are still live and will help it stabilize quickly. 1205 GCPtrLivenessData RevisedLivenessData; 1206 computeLiveInValues(DT, F, RevisedLivenessData); 1207 for (size_t i = 0; i < records.size(); i++) { 1208 struct PartiallyConstructedSafepointRecord &info = records[i]; 1209 const CallSite &CS = toUpdate[i]; 1210 recomputeLiveInValues(RevisedLivenessData, CS, info); 1211 } 1212} 1213 1214// When inserting gc.relocate and gc.result calls, we need to ensure there are 1215// no uses of the original value / return value between the gc.statepoint and 1216// the gc.relocate / gc.result call. One case which can arise is a phi node 1217// starting one of the successor blocks. We also need to be able to insert the 1218// gc.relocates only on the path which goes through the statepoint. We might 1219// need to split an edge to make this possible. 1220static BasicBlock * 1221normalizeForInvokeSafepoint(BasicBlock *BB, BasicBlock *InvokeParent, 1222 DominatorTree &DT) { 1223 BasicBlock *Ret = BB; 1224 if (!BB->getUniquePredecessor()) 1225 Ret = SplitBlockPredecessors(BB, InvokeParent, "", &DT); 1226 1227 // Now that 'Ret' has unique predecessor we can safely remove all phi nodes 1228 // from it 1229 FoldSingleEntryPHINodes(Ret); 1230 assert(!isa<PHINode>(Ret->begin()) && 1231 "All PHI nodes should have been removed!"); 1232 1233 // At this point, we can safely insert a gc.relocate or gc.result as the first 1234 // instruction in Ret if needed. 1235 return Ret; 1236} 1237 1238// Create new attribute set containing only attributes which can be transferred 1239// from original call to the safepoint. 1240static AttributeSet legalizeCallAttributes(AttributeSet AS) { 1241 AttributeSet Ret; 1242 1243 for (unsigned Slot = 0; Slot < AS.getNumSlots(); Slot++) { 1244 unsigned Index = AS.getSlotIndex(Slot); 1245 1246 if (Index == AttributeSet::ReturnIndex || 1247 Index == AttributeSet::FunctionIndex) { 1248 1249 for (Attribute Attr : make_range(AS.begin(Slot), AS.end(Slot))) { 1250 1251 // Do not allow certain attributes - just skip them 1252 // Safepoint can not be read only or read none. 1253 if (Attr.hasAttribute(Attribute::ReadNone) || 1254 Attr.hasAttribute(Attribute::ReadOnly)) 1255 continue; 1256 1257 // These attributes control the generation of the gc.statepoint call / 1258 // invoke itself; and once the gc.statepoint is in place, they're of no 1259 // use. 1260 if (Attr.hasAttribute("statepoint-num-patch-bytes") || 1261 Attr.hasAttribute("statepoint-id")) 1262 continue; 1263 1264 Ret = Ret.addAttributes( 1265 AS.getContext(), Index, 1266 AttributeSet::get(AS.getContext(), Index, AttrBuilder(Attr))); 1267 } 1268 } 1269 1270 // Just skip parameter attributes for now 1271 } 1272 1273 return Ret; 1274} 1275 1276/// Helper function to place all gc relocates necessary for the given 1277/// statepoint. 1278/// Inputs: 1279/// liveVariables - list of variables to be relocated. 1280/// liveStart - index of the first live variable. 1281/// basePtrs - base pointers. 1282/// statepointToken - statepoint instruction to which relocates should be 1283/// bound. 1284/// Builder - Llvm IR builder to be used to construct new calls. 1285static void CreateGCRelocates(ArrayRef<Value *> LiveVariables, 1286 const int LiveStart, 1287 ArrayRef<Value *> BasePtrs, 1288 Instruction *StatepointToken, 1289 IRBuilder<> Builder) { 1290 if (LiveVariables.empty()) 1291 return; 1292 1293 auto FindIndex = [](ArrayRef<Value *> LiveVec, Value *Val) { 1294 auto ValIt = std::find(LiveVec.begin(), LiveVec.end(), Val); 1295 assert(ValIt != LiveVec.end() && "Val not found in LiveVec!"); 1296 size_t Index = std::distance(LiveVec.begin(), ValIt); 1297 assert(Index < LiveVec.size() && "Bug in std::find?"); 1298 return Index; 1299 }; 1300 Module *M = StatepointToken->getModule(); 1301 1302 // All gc_relocate are generated as i8 addrspace(1)* (or a vector type whose 1303 // element type is i8 addrspace(1)*). We originally generated unique 1304 // declarations for each pointer type, but this proved problematic because 1305 // the intrinsic mangling code is incomplete and fragile. Since we're moving 1306 // towards a single unified pointer type anyways, we can just cast everything 1307 // to an i8* of the right address space. A bitcast is added later to convert 1308 // gc_relocate to the actual value's type. 1309 auto getGCRelocateDecl = [&] (Type *Ty) { 1310 assert(isHandledGCPointerType(Ty)); 1311 auto AS = Ty->getScalarType()->getPointerAddressSpace(); 1312 Type *NewTy = Type::getInt8PtrTy(M->getContext(), AS); 1313 if (auto *VT = dyn_cast<VectorType>(Ty)) 1314 NewTy = VectorType::get(NewTy, VT->getNumElements()); 1315 return Intrinsic::getDeclaration(M, Intrinsic::experimental_gc_relocate, 1316 {NewTy}); 1317 }; 1318 1319 // Lazily populated map from input types to the canonicalized form mentioned 1320 // in the comment above. This should probably be cached somewhere more 1321 // broadly. 1322 DenseMap<Type*, Value*> TypeToDeclMap; 1323 1324 for (unsigned i = 0; i < LiveVariables.size(); i++) { 1325 // Generate the gc.relocate call and save the result 1326 Value *BaseIdx = 1327 Builder.getInt32(LiveStart + FindIndex(LiveVariables, BasePtrs[i])); 1328 Value *LiveIdx = Builder.getInt32(LiveStart + i); 1329 1330 Type *Ty = LiveVariables[i]->getType(); 1331 if (!TypeToDeclMap.count(Ty)) 1332 TypeToDeclMap[Ty] = getGCRelocateDecl(Ty); 1333 Value *GCRelocateDecl = TypeToDeclMap[Ty]; 1334 1335 // only specify a debug name if we can give a useful one 1336 CallInst *Reloc = Builder.CreateCall( 1337 GCRelocateDecl, {StatepointToken, BaseIdx, LiveIdx}, 1338 suffixed_name_or(LiveVariables[i], ".relocated", "")); 1339 // Trick CodeGen into thinking there are lots of free registers at this 1340 // fake call. 1341 Reloc->setCallingConv(CallingConv::Cold); 1342 } 1343} 1344 1345namespace { 1346 1347/// This struct is used to defer RAUWs and `eraseFromParent` s. Using this 1348/// avoids having to worry about keeping around dangling pointers to Values. 1349class DeferredReplacement { 1350 AssertingVH<Instruction> Old; 1351 AssertingVH<Instruction> New; 1352 1353public: 1354 explicit DeferredReplacement(Instruction *Old, Instruction *New) : 1355 Old(Old), New(New) { 1356 assert(Old != New && "Not allowed!"); 1357 } 1358 1359 /// Does the task represented by this instance. 1360 void doReplacement() { 1361 Instruction *OldI = Old; 1362 Instruction *NewI = New; 1363 1364 assert(OldI != NewI && "Disallowed at construction?!"); 1365 1366 Old = nullptr; 1367 New = nullptr; 1368 1369 if (NewI) 1370 OldI->replaceAllUsesWith(NewI); 1371 OldI->eraseFromParent(); 1372 } 1373}; 1374} 1375 1376static void 1377makeStatepointExplicitImpl(const CallSite CS, /* to replace */ 1378 const SmallVectorImpl<Value *> &BasePtrs, 1379 const SmallVectorImpl<Value *> &LiveVariables, 1380 PartiallyConstructedSafepointRecord &Result, 1381 std::vector<DeferredReplacement> &Replacements) { 1382 assert(BasePtrs.size() == LiveVariables.size()); 1383 assert((UseDeoptBundles || isStatepoint(CS)) && 1384 "This method expects to be rewriting a statepoint"); 1385 1386 // Then go ahead and use the builder do actually do the inserts. We insert 1387 // immediately before the previous instruction under the assumption that all 1388 // arguments will be available here. We can't insert afterwards since we may 1389 // be replacing a terminator. 1390 Instruction *InsertBefore = CS.getInstruction(); 1391 IRBuilder<> Builder(InsertBefore); 1392 1393 ArrayRef<Value *> GCArgs(LiveVariables); 1394 uint64_t StatepointID = 0xABCDEF00; 1395 uint32_t NumPatchBytes = 0; 1396 uint32_t Flags = uint32_t(StatepointFlags::None); 1397 1398 ArrayRef<Use> CallArgs; 1399 ArrayRef<Use> DeoptArgs; 1400 ArrayRef<Use> TransitionArgs; 1401 1402 Value *CallTarget = nullptr; 1403 1404 if (UseDeoptBundles) { 1405 CallArgs = {CS.arg_begin(), CS.arg_end()}; 1406 DeoptArgs = GetDeoptBundleOperands(CS); 1407 // TODO: we don't fill in TransitionArgs or Flags in this branch, but we 1408 // could have an operand bundle for that too. 1409 AttributeSet OriginalAttrs = CS.getAttributes(); 1410 1411 Attribute AttrID = OriginalAttrs.getAttribute(AttributeSet::FunctionIndex, 1412 "statepoint-id"); 1413 if (AttrID.isStringAttribute()) 1414 AttrID.getValueAsString().getAsInteger(10, StatepointID); 1415 1416 Attribute AttrNumPatchBytes = OriginalAttrs.getAttribute( 1417 AttributeSet::FunctionIndex, "statepoint-num-patch-bytes"); 1418 if (AttrNumPatchBytes.isStringAttribute()) 1419 AttrNumPatchBytes.getValueAsString().getAsInteger(10, NumPatchBytes); 1420 1421 CallTarget = CS.getCalledValue(); 1422 } else { 1423 // This branch will be gone soon, and we will soon only support the 1424 // UseDeoptBundles == true configuration. 1425 Statepoint OldSP(CS); 1426 StatepointID = OldSP.getID(); 1427 NumPatchBytes = OldSP.getNumPatchBytes(); 1428 Flags = OldSP.getFlags(); 1429 1430 CallArgs = {OldSP.arg_begin(), OldSP.arg_end()}; 1431 DeoptArgs = {OldSP.vm_state_begin(), OldSP.vm_state_end()}; 1432 TransitionArgs = {OldSP.gc_transition_args_begin(), 1433 OldSP.gc_transition_args_end()}; 1434 CallTarget = OldSP.getCalledValue(); 1435 } 1436 1437 // Create the statepoint given all the arguments 1438 Instruction *Token = nullptr; 1439 AttributeSet ReturnAttrs; 1440 if (CS.isCall()) { 1441 CallInst *ToReplace = cast<CallInst>(CS.getInstruction()); 1442 CallInst *Call = Builder.CreateGCStatepointCall( 1443 StatepointID, NumPatchBytes, CallTarget, Flags, CallArgs, 1444 TransitionArgs, DeoptArgs, GCArgs, "safepoint_token"); 1445 1446 Call->setTailCall(ToReplace->isTailCall()); 1447 Call->setCallingConv(ToReplace->getCallingConv()); 1448 1449 // Currently we will fail on parameter attributes and on certain 1450 // function attributes. 1451 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes()); 1452 // In case if we can handle this set of attributes - set up function attrs 1453 // directly on statepoint and return attrs later for gc_result intrinsic. 1454 Call->setAttributes(NewAttrs.getFnAttributes()); 1455 ReturnAttrs = NewAttrs.getRetAttributes(); 1456 1457 Token = Call; 1458 1459 // Put the following gc_result and gc_relocate calls immediately after the 1460 // the old call (which we're about to delete) 1461 assert(ToReplace->getNextNode() && "Not a terminator, must have next!"); 1462 Builder.SetInsertPoint(ToReplace->getNextNode()); 1463 Builder.SetCurrentDebugLocation(ToReplace->getNextNode()->getDebugLoc()); 1464 } else { 1465 InvokeInst *ToReplace = cast<InvokeInst>(CS.getInstruction()); 1466 1467 // Insert the new invoke into the old block. We'll remove the old one in a 1468 // moment at which point this will become the new terminator for the 1469 // original block. 1470 InvokeInst *Invoke = Builder.CreateGCStatepointInvoke( 1471 StatepointID, NumPatchBytes, CallTarget, ToReplace->getNormalDest(), 1472 ToReplace->getUnwindDest(), Flags, CallArgs, TransitionArgs, DeoptArgs, 1473 GCArgs, "statepoint_token"); 1474 1475 Invoke->setCallingConv(ToReplace->getCallingConv()); 1476 1477 // Currently we will fail on parameter attributes and on certain 1478 // function attributes. 1479 AttributeSet NewAttrs = legalizeCallAttributes(ToReplace->getAttributes()); 1480 // In case if we can handle this set of attributes - set up function attrs 1481 // directly on statepoint and return attrs later for gc_result intrinsic. 1482 Invoke->setAttributes(NewAttrs.getFnAttributes()); 1483 ReturnAttrs = NewAttrs.getRetAttributes(); 1484 1485 Token = Invoke; 1486 1487 // Generate gc relocates in exceptional path 1488 BasicBlock *UnwindBlock = ToReplace->getUnwindDest(); 1489 assert(!isa<PHINode>(UnwindBlock->begin()) && 1490 UnwindBlock->getUniquePredecessor() && 1491 "can't safely insert in this block!"); 1492 1493 Builder.SetInsertPoint(&*UnwindBlock->getFirstInsertionPt()); 1494 Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc()); 1495 1496 // Attach exceptional gc relocates to the landingpad. 1497 Instruction *ExceptionalToken = UnwindBlock->getLandingPadInst(); 1498 Result.UnwindToken = ExceptionalToken; 1499 1500 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx(); 1501 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, ExceptionalToken, 1502 Builder); 1503 1504 // Generate gc relocates and returns for normal block 1505 BasicBlock *NormalDest = ToReplace->getNormalDest(); 1506 assert(!isa<PHINode>(NormalDest->begin()) && 1507 NormalDest->getUniquePredecessor() && 1508 "can't safely insert in this block!"); 1509 1510 Builder.SetInsertPoint(&*NormalDest->getFirstInsertionPt()); 1511 1512 // gc relocates will be generated later as if it were regular call 1513 // statepoint 1514 } 1515 assert(Token && "Should be set in one of the above branches!"); 1516 1517 if (UseDeoptBundles) { 1518 Token->setName("statepoint_token"); 1519 if (!CS.getType()->isVoidTy() && !CS.getInstruction()->use_empty()) { 1520 StringRef Name = 1521 CS.getInstruction()->hasName() ? CS.getInstruction()->getName() : ""; 1522 CallInst *GCResult = Builder.CreateGCResult(Token, CS.getType(), Name); 1523 GCResult->setAttributes(CS.getAttributes().getRetAttributes()); 1524 1525 // We cannot RAUW or delete CS.getInstruction() because it could be in the 1526 // live set of some other safepoint, in which case that safepoint's 1527 // PartiallyConstructedSafepointRecord will hold a raw pointer to this 1528 // llvm::Instruction. Instead, we defer the replacement and deletion to 1529 // after the live sets have been made explicit in the IR, and we no longer 1530 // have raw pointers to worry about. 1531 Replacements.emplace_back(CS.getInstruction(), GCResult); 1532 } else { 1533 Replacements.emplace_back(CS.getInstruction(), nullptr); 1534 } 1535 } else { 1536 assert(!CS.getInstruction()->hasNUsesOrMore(2) && 1537 "only valid use before rewrite is gc.result"); 1538 assert(!CS.getInstruction()->hasOneUse() || 1539 isGCResult(cast<Instruction>(*CS.getInstruction()->user_begin()))); 1540 1541 // Take the name of the original statepoint token if there was one. 1542 Token->takeName(CS.getInstruction()); 1543 1544 // Update the gc.result of the original statepoint (if any) to use the newly 1545 // inserted statepoint. This is safe to do here since the token can't be 1546 // considered a live reference. 1547 CS.getInstruction()->replaceAllUsesWith(Token); 1548 CS.getInstruction()->eraseFromParent(); 1549 } 1550 1551 Result.StatepointToken = Token; 1552 1553 // Second, create a gc.relocate for every live variable 1554 const unsigned LiveStartIdx = Statepoint(Token).gcArgsStartIdx(); 1555 CreateGCRelocates(LiveVariables, LiveStartIdx, BasePtrs, Token, Builder); 1556} 1557 1558namespace { 1559struct NameOrdering { 1560 Value *Base; 1561 Value *Derived; 1562 1563 bool operator()(NameOrdering const &a, NameOrdering const &b) { 1564 return -1 == a.Derived->getName().compare(b.Derived->getName()); 1565 } 1566}; 1567} 1568 1569static void StabilizeOrder(SmallVectorImpl<Value *> &BaseVec, 1570 SmallVectorImpl<Value *> &LiveVec) { 1571 assert(BaseVec.size() == LiveVec.size()); 1572 1573 SmallVector<NameOrdering, 64> Temp; 1574 for (size_t i = 0; i < BaseVec.size(); i++) { 1575 NameOrdering v; 1576 v.Base = BaseVec[i]; 1577 v.Derived = LiveVec[i]; 1578 Temp.push_back(v); 1579 } 1580 1581 std::sort(Temp.begin(), Temp.end(), NameOrdering()); 1582 for (size_t i = 0; i < BaseVec.size(); i++) { 1583 BaseVec[i] = Temp[i].Base; 1584 LiveVec[i] = Temp[i].Derived; 1585 } 1586} 1587 1588// Replace an existing gc.statepoint with a new one and a set of gc.relocates 1589// which make the relocations happening at this safepoint explicit. 1590// 1591// WARNING: Does not do any fixup to adjust users of the original live 1592// values. That's the callers responsibility. 1593static void 1594makeStatepointExplicit(DominatorTree &DT, const CallSite &CS, 1595 PartiallyConstructedSafepointRecord &Result, 1596 std::vector<DeferredReplacement> &Replacements) { 1597 const auto &LiveSet = Result.LiveSet; 1598 const auto &PointerToBase = Result.PointerToBase; 1599 1600 // Convert to vector for efficient cross referencing. 1601 SmallVector<Value *, 64> BaseVec, LiveVec; 1602 LiveVec.reserve(LiveSet.size()); 1603 BaseVec.reserve(LiveSet.size()); 1604 for (Value *L : LiveSet) { 1605 LiveVec.push_back(L); 1606 assert(PointerToBase.count(L)); 1607 Value *Base = PointerToBase.find(L)->second; 1608 BaseVec.push_back(Base); 1609 } 1610 assert(LiveVec.size() == BaseVec.size()); 1611 1612 // To make the output IR slightly more stable (for use in diffs), ensure a 1613 // fixed order of the values in the safepoint (by sorting the value name). 1614 // The order is otherwise meaningless. 1615 StabilizeOrder(BaseVec, LiveVec); 1616 1617 // Do the actual rewriting and delete the old statepoint 1618 makeStatepointExplicitImpl(CS, BaseVec, LiveVec, Result, Replacements); 1619} 1620 1621// Helper function for the relocationViaAlloca. 1622// 1623// It receives iterator to the statepoint gc relocates and emits a store to the 1624// assigned location (via allocaMap) for the each one of them. It adds the 1625// visited values into the visitedLiveValues set, which we will later use them 1626// for sanity checking. 1627static void 1628insertRelocationStores(iterator_range<Value::user_iterator> GCRelocs, 1629 DenseMap<Value *, Value *> &AllocaMap, 1630 DenseSet<Value *> &VisitedLiveValues) { 1631 1632 for (User *U : GCRelocs) { 1633 GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U); 1634 if (!Relocate) 1635 continue; 1636 1637 Value *OriginalValue = const_cast<Value *>(Relocate->getDerivedPtr()); 1638 assert(AllocaMap.count(OriginalValue)); 1639 Value *Alloca = AllocaMap[OriginalValue]; 1640 1641 // Emit store into the related alloca 1642 // All gc_relocates are i8 addrspace(1)* typed, and it must be bitcasted to 1643 // the correct type according to alloca. 1644 assert(Relocate->getNextNode() && 1645 "Should always have one since it's not a terminator"); 1646 IRBuilder<> Builder(Relocate->getNextNode()); 1647 Value *CastedRelocatedValue = 1648 Builder.CreateBitCast(Relocate, 1649 cast<AllocaInst>(Alloca)->getAllocatedType(), 1650 suffixed_name_or(Relocate, ".casted", "")); 1651 1652 StoreInst *Store = new StoreInst(CastedRelocatedValue, Alloca); 1653 Store->insertAfter(cast<Instruction>(CastedRelocatedValue)); 1654 1655#ifndef NDEBUG 1656 VisitedLiveValues.insert(OriginalValue); 1657#endif 1658 } 1659} 1660 1661// Helper function for the "relocationViaAlloca". Similar to the 1662// "insertRelocationStores" but works for rematerialized values. 1663static void 1664insertRematerializationStores( 1665 RematerializedValueMapTy RematerializedValues, 1666 DenseMap<Value *, Value *> &AllocaMap, 1667 DenseSet<Value *> &VisitedLiveValues) { 1668 1669 for (auto RematerializedValuePair: RematerializedValues) { 1670 Instruction *RematerializedValue = RematerializedValuePair.first; 1671 Value *OriginalValue = RematerializedValuePair.second; 1672 1673 assert(AllocaMap.count(OriginalValue) && 1674 "Can not find alloca for rematerialized value"); 1675 Value *Alloca = AllocaMap[OriginalValue]; 1676 1677 StoreInst *Store = new StoreInst(RematerializedValue, Alloca); 1678 Store->insertAfter(RematerializedValue); 1679 1680#ifndef NDEBUG 1681 VisitedLiveValues.insert(OriginalValue); 1682#endif 1683 } 1684} 1685 1686/// Do all the relocation update via allocas and mem2reg 1687static void relocationViaAlloca( 1688 Function &F, DominatorTree &DT, ArrayRef<Value *> Live, 1689 ArrayRef<PartiallyConstructedSafepointRecord> Records) { 1690#ifndef NDEBUG 1691 // record initial number of (static) allocas; we'll check we have the same 1692 // number when we get done. 1693 int InitialAllocaNum = 0; 1694 for (auto I = F.getEntryBlock().begin(), E = F.getEntryBlock().end(); I != E; 1695 I++) 1696 if (isa<AllocaInst>(*I)) 1697 InitialAllocaNum++; 1698#endif 1699 1700 // TODO-PERF: change data structures, reserve 1701 DenseMap<Value *, Value *> AllocaMap; 1702 SmallVector<AllocaInst *, 200> PromotableAllocas; 1703 // Used later to chack that we have enough allocas to store all values 1704 std::size_t NumRematerializedValues = 0; 1705 PromotableAllocas.reserve(Live.size()); 1706 1707 // Emit alloca for "LiveValue" and record it in "allocaMap" and 1708 // "PromotableAllocas" 1709 auto emitAllocaFor = [&](Value *LiveValue) { 1710 AllocaInst *Alloca = new AllocaInst(LiveValue->getType(), "", 1711 F.getEntryBlock().getFirstNonPHI()); 1712 AllocaMap[LiveValue] = Alloca; 1713 PromotableAllocas.push_back(Alloca); 1714 }; 1715 1716 // Emit alloca for each live gc pointer 1717 for (Value *V : Live) 1718 emitAllocaFor(V); 1719 1720 // Emit allocas for rematerialized values 1721 for (const auto &Info : Records) 1722 for (auto RematerializedValuePair : Info.RematerializedValues) { 1723 Value *OriginalValue = RematerializedValuePair.second; 1724 if (AllocaMap.count(OriginalValue) != 0) 1725 continue; 1726 1727 emitAllocaFor(OriginalValue); 1728 ++NumRematerializedValues; 1729 } 1730 1731 // The next two loops are part of the same conceptual operation. We need to 1732 // insert a store to the alloca after the original def and at each 1733 // redefinition. We need to insert a load before each use. These are split 1734 // into distinct loops for performance reasons. 1735 1736 // Update gc pointer after each statepoint: either store a relocated value or 1737 // null (if no relocated value was found for this gc pointer and it is not a 1738 // gc_result). This must happen before we update the statepoint with load of 1739 // alloca otherwise we lose the link between statepoint and old def. 1740 for (const auto &Info : Records) { 1741 Value *Statepoint = Info.StatepointToken; 1742 1743 // This will be used for consistency check 1744 DenseSet<Value *> VisitedLiveValues; 1745 1746 // Insert stores for normal statepoint gc relocates 1747 insertRelocationStores(Statepoint->users(), AllocaMap, VisitedLiveValues); 1748 1749 // In case if it was invoke statepoint 1750 // we will insert stores for exceptional path gc relocates. 1751 if (isa<InvokeInst>(Statepoint)) { 1752 insertRelocationStores(Info.UnwindToken->users(), AllocaMap, 1753 VisitedLiveValues); 1754 } 1755 1756 // Do similar thing with rematerialized values 1757 insertRematerializationStores(Info.RematerializedValues, AllocaMap, 1758 VisitedLiveValues); 1759 1760 if (ClobberNonLive) { 1761 // As a debugging aid, pretend that an unrelocated pointer becomes null at 1762 // the gc.statepoint. This will turn some subtle GC problems into 1763 // slightly easier to debug SEGVs. Note that on large IR files with 1764 // lots of gc.statepoints this is extremely costly both memory and time 1765 // wise. 1766 SmallVector<AllocaInst *, 64> ToClobber; 1767 for (auto Pair : AllocaMap) { 1768 Value *Def = Pair.first; 1769 AllocaInst *Alloca = cast<AllocaInst>(Pair.second); 1770 1771 // This value was relocated 1772 if (VisitedLiveValues.count(Def)) { 1773 continue; 1774 } 1775 ToClobber.push_back(Alloca); 1776 } 1777 1778 auto InsertClobbersAt = [&](Instruction *IP) { 1779 for (auto *AI : ToClobber) { 1780 auto AIType = cast<PointerType>(AI->getType()); 1781 auto PT = cast<PointerType>(AIType->getElementType()); 1782 Constant *CPN = ConstantPointerNull::get(PT); 1783 StoreInst *Store = new StoreInst(CPN, AI); 1784 Store->insertBefore(IP); 1785 } 1786 }; 1787 1788 // Insert the clobbering stores. These may get intermixed with the 1789 // gc.results and gc.relocates, but that's fine. 1790 if (auto II = dyn_cast<InvokeInst>(Statepoint)) { 1791 InsertClobbersAt(&*II->getNormalDest()->getFirstInsertionPt()); 1792 InsertClobbersAt(&*II->getUnwindDest()->getFirstInsertionPt()); 1793 } else { 1794 InsertClobbersAt(cast<Instruction>(Statepoint)->getNextNode()); 1795 } 1796 } 1797 } 1798 1799 // Update use with load allocas and add store for gc_relocated. 1800 for (auto Pair : AllocaMap) { 1801 Value *Def = Pair.first; 1802 Value *Alloca = Pair.second; 1803 1804 // We pre-record the uses of allocas so that we dont have to worry about 1805 // later update that changes the user information.. 1806 1807 SmallVector<Instruction *, 20> Uses; 1808 // PERF: trade a linear scan for repeated reallocation 1809 Uses.reserve(std::distance(Def->user_begin(), Def->user_end())); 1810 for (User *U : Def->users()) { 1811 if (!isa<ConstantExpr>(U)) { 1812 // If the def has a ConstantExpr use, then the def is either a 1813 // ConstantExpr use itself or null. In either case 1814 // (recursively in the first, directly in the second), the oop 1815 // it is ultimately dependent on is null and this particular 1816 // use does not need to be fixed up. 1817 Uses.push_back(cast<Instruction>(U)); 1818 } 1819 } 1820 1821 std::sort(Uses.begin(), Uses.end()); 1822 auto Last = std::unique(Uses.begin(), Uses.end()); 1823 Uses.erase(Last, Uses.end()); 1824 1825 for (Instruction *Use : Uses) { 1826 if (isa<PHINode>(Use)) { 1827 PHINode *Phi = cast<PHINode>(Use); 1828 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) { 1829 if (Def == Phi->getIncomingValue(i)) { 1830 LoadInst *Load = new LoadInst( 1831 Alloca, "", Phi->getIncomingBlock(i)->getTerminator()); 1832 Phi->setIncomingValue(i, Load); 1833 } 1834 } 1835 } else { 1836 LoadInst *Load = new LoadInst(Alloca, "", Use); 1837 Use->replaceUsesOfWith(Def, Load); 1838 } 1839 } 1840 1841 // Emit store for the initial gc value. Store must be inserted after load, 1842 // otherwise store will be in alloca's use list and an extra load will be 1843 // inserted before it. 1844 StoreInst *Store = new StoreInst(Def, Alloca); 1845 if (Instruction *Inst = dyn_cast<Instruction>(Def)) { 1846 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(Inst)) { 1847 // InvokeInst is a TerminatorInst so the store need to be inserted 1848 // into its normal destination block. 1849 BasicBlock *NormalDest = Invoke->getNormalDest(); 1850 Store->insertBefore(NormalDest->getFirstNonPHI()); 1851 } else { 1852 assert(!Inst->isTerminator() && 1853 "The only TerminatorInst that can produce a value is " 1854 "InvokeInst which is handled above."); 1855 Store->insertAfter(Inst); 1856 } 1857 } else { 1858 assert(isa<Argument>(Def)); 1859 Store->insertAfter(cast<Instruction>(Alloca)); 1860 } 1861 } 1862 1863 assert(PromotableAllocas.size() == Live.size() + NumRematerializedValues && 1864 "we must have the same allocas with lives"); 1865 if (!PromotableAllocas.empty()) { 1866 // Apply mem2reg to promote alloca to SSA 1867 PromoteMemToReg(PromotableAllocas, DT); 1868 } 1869 1870#ifndef NDEBUG 1871 for (auto &I : F.getEntryBlock()) 1872 if (isa<AllocaInst>(I)) 1873 InitialAllocaNum--; 1874 assert(InitialAllocaNum == 0 && "We must not introduce any extra allocas"); 1875#endif 1876} 1877 1878/// Implement a unique function which doesn't require we sort the input 1879/// vector. Doing so has the effect of changing the output of a couple of 1880/// tests in ways which make them less useful in testing fused safepoints. 1881template <typename T> static void unique_unsorted(SmallVectorImpl<T> &Vec) { 1882 SmallSet<T, 8> Seen; 1883 Vec.erase(std::remove_if(Vec.begin(), Vec.end(), [&](const T &V) { 1884 return !Seen.insert(V).second; 1885 }), Vec.end()); 1886} 1887 1888/// Insert holders so that each Value is obviously live through the entire 1889/// lifetime of the call. 1890static void insertUseHolderAfter(CallSite &CS, const ArrayRef<Value *> Values, 1891 SmallVectorImpl<CallInst *> &Holders) { 1892 if (Values.empty()) 1893 // No values to hold live, might as well not insert the empty holder 1894 return; 1895 1896 Module *M = CS.getInstruction()->getModule(); 1897 // Use a dummy vararg function to actually hold the values live 1898 Function *Func = cast<Function>(M->getOrInsertFunction( 1899 "__tmp_use", FunctionType::get(Type::getVoidTy(M->getContext()), true))); 1900 if (CS.isCall()) { 1901 // For call safepoints insert dummy calls right after safepoint 1902 Holders.push_back(CallInst::Create(Func, Values, "", 1903 &*++CS.getInstruction()->getIterator())); 1904 return; 1905 } 1906 // For invoke safepooints insert dummy calls both in normal and 1907 // exceptional destination blocks 1908 auto *II = cast<InvokeInst>(CS.getInstruction()); 1909 Holders.push_back(CallInst::Create( 1910 Func, Values, "", &*II->getNormalDest()->getFirstInsertionPt())); 1911 Holders.push_back(CallInst::Create( 1912 Func, Values, "", &*II->getUnwindDest()->getFirstInsertionPt())); 1913} 1914 1915static void findLiveReferences( 1916 Function &F, DominatorTree &DT, ArrayRef<CallSite> toUpdate, 1917 MutableArrayRef<struct PartiallyConstructedSafepointRecord> records) { 1918 GCPtrLivenessData OriginalLivenessData; 1919 computeLiveInValues(DT, F, OriginalLivenessData); 1920 for (size_t i = 0; i < records.size(); i++) { 1921 struct PartiallyConstructedSafepointRecord &info = records[i]; 1922 const CallSite &CS = toUpdate[i]; 1923 analyzeParsePointLiveness(DT, OriginalLivenessData, CS, info); 1924 } 1925} 1926 1927/// Remove any vector of pointers from the live set by scalarizing them over the 1928/// statepoint instruction. Adds the scalarized pieces to the live set. It 1929/// would be preferable to include the vector in the statepoint itself, but 1930/// the lowering code currently does not handle that. Extending it would be 1931/// slightly non-trivial since it requires a format change. Given how rare 1932/// such cases are (for the moment?) scalarizing is an acceptable compromise. 1933static void splitVectorValues(Instruction *StatepointInst, 1934 StatepointLiveSetTy &LiveSet, 1935 DenseMap<Value *, Value *>& PointerToBase, 1936 DominatorTree &DT) { 1937 SmallVector<Value *, 16> ToSplit; 1938 for (Value *V : LiveSet) 1939 if (isa<VectorType>(V->getType())) 1940 ToSplit.push_back(V); 1941 1942 if (ToSplit.empty()) 1943 return; 1944 1945 DenseMap<Value *, SmallVector<Value *, 16>> ElementMapping; 1946 1947 Function &F = *(StatepointInst->getParent()->getParent()); 1948 1949 DenseMap<Value *, AllocaInst *> AllocaMap; 1950 // First is normal return, second is exceptional return (invoke only) 1951 DenseMap<Value *, std::pair<Value *, Value *>> Replacements; 1952 for (Value *V : ToSplit) { 1953 AllocaInst *Alloca = 1954 new AllocaInst(V->getType(), "", F.getEntryBlock().getFirstNonPHI()); 1955 AllocaMap[V] = Alloca; 1956 1957 VectorType *VT = cast<VectorType>(V->getType()); 1958 IRBuilder<> Builder(StatepointInst); 1959 SmallVector<Value *, 16> Elements; 1960 for (unsigned i = 0; i < VT->getNumElements(); i++) 1961 Elements.push_back(Builder.CreateExtractElement(V, Builder.getInt32(i))); 1962 ElementMapping[V] = Elements; 1963 1964 auto InsertVectorReform = [&](Instruction *IP) { 1965 Builder.SetInsertPoint(IP); 1966 Builder.SetCurrentDebugLocation(IP->getDebugLoc()); 1967 Value *ResultVec = UndefValue::get(VT); 1968 for (unsigned i = 0; i < VT->getNumElements(); i++) 1969 ResultVec = Builder.CreateInsertElement(ResultVec, Elements[i], 1970 Builder.getInt32(i)); 1971 return ResultVec; 1972 }; 1973 1974 if (isa<CallInst>(StatepointInst)) { 1975 BasicBlock::iterator Next(StatepointInst); 1976 Next++; 1977 Instruction *IP = &*(Next); 1978 Replacements[V].first = InsertVectorReform(IP); 1979 Replacements[V].second = nullptr; 1980 } else { 1981 InvokeInst *Invoke = cast<InvokeInst>(StatepointInst); 1982 // We've already normalized - check that we don't have shared destination 1983 // blocks 1984 BasicBlock *NormalDest = Invoke->getNormalDest(); 1985 assert(!isa<PHINode>(NormalDest->begin())); 1986 BasicBlock *UnwindDest = Invoke->getUnwindDest(); 1987 assert(!isa<PHINode>(UnwindDest->begin())); 1988 // Insert insert element sequences in both successors 1989 Instruction *IP = &*(NormalDest->getFirstInsertionPt()); 1990 Replacements[V].first = InsertVectorReform(IP); 1991 IP = &*(UnwindDest->getFirstInsertionPt()); 1992 Replacements[V].second = InsertVectorReform(IP); 1993 } 1994 } 1995 1996 for (Value *V : ToSplit) { 1997 AllocaInst *Alloca = AllocaMap[V]; 1998 1999 // Capture all users before we start mutating use lists 2000 SmallVector<Instruction *, 16> Users; 2001 for (User *U : V->users()) 2002 Users.push_back(cast<Instruction>(U)); 2003 2004 for (Instruction *I : Users) { 2005 if (auto Phi = dyn_cast<PHINode>(I)) { 2006 for (unsigned i = 0; i < Phi->getNumIncomingValues(); i++) 2007 if (V == Phi->getIncomingValue(i)) { 2008 LoadInst *Load = new LoadInst( 2009 Alloca, "", Phi->getIncomingBlock(i)->getTerminator()); 2010 Phi->setIncomingValue(i, Load); 2011 } 2012 } else { 2013 LoadInst *Load = new LoadInst(Alloca, "", I); 2014 I->replaceUsesOfWith(V, Load); 2015 } 2016 } 2017 2018 // Store the original value and the replacement value into the alloca 2019 StoreInst *Store = new StoreInst(V, Alloca); 2020 if (auto I = dyn_cast<Instruction>(V)) 2021 Store->insertAfter(I); 2022 else 2023 Store->insertAfter(Alloca); 2024 2025 // Normal return for invoke, or call return 2026 Instruction *Replacement = cast<Instruction>(Replacements[V].first); 2027 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement); 2028 // Unwind return for invoke only 2029 Replacement = cast_or_null<Instruction>(Replacements[V].second); 2030 if (Replacement) 2031 (new StoreInst(Replacement, Alloca))->insertAfter(Replacement); 2032 } 2033 2034 // apply mem2reg to promote alloca to SSA 2035 SmallVector<AllocaInst *, 16> Allocas; 2036 for (Value *V : ToSplit) 2037 Allocas.push_back(AllocaMap[V]); 2038 PromoteMemToReg(Allocas, DT); 2039 2040 // Update our tracking of live pointers and base mappings to account for the 2041 // changes we just made. 2042 for (Value *V : ToSplit) { 2043 auto &Elements = ElementMapping[V]; 2044 2045 LiveSet.erase(V); 2046 LiveSet.insert(Elements.begin(), Elements.end()); 2047 // We need to update the base mapping as well. 2048 assert(PointerToBase.count(V)); 2049 Value *OldBase = PointerToBase[V]; 2050 auto &BaseElements = ElementMapping[OldBase]; 2051 PointerToBase.erase(V); 2052 assert(Elements.size() == BaseElements.size()); 2053 for (unsigned i = 0; i < Elements.size(); i++) { 2054 Value *Elem = Elements[i]; 2055 PointerToBase[Elem] = BaseElements[i]; 2056 } 2057 } 2058} 2059 2060// Helper function for the "rematerializeLiveValues". It walks use chain 2061// starting from the "CurrentValue" until it meets "BaseValue". Only "simple" 2062// values are visited (currently it is GEP's and casts). Returns true if it 2063// successfully reached "BaseValue" and false otherwise. 2064// Fills "ChainToBase" array with all visited values. "BaseValue" is not 2065// recorded. 2066static bool findRematerializableChainToBasePointer( 2067 SmallVectorImpl<Instruction*> &ChainToBase, 2068 Value *CurrentValue, Value *BaseValue) { 2069 2070 // We have found a base value 2071 if (CurrentValue == BaseValue) { 2072 return true; 2073 } 2074 2075 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(CurrentValue)) { 2076 ChainToBase.push_back(GEP); 2077 return findRematerializableChainToBasePointer(ChainToBase, 2078 GEP->getPointerOperand(), 2079 BaseValue); 2080 } 2081 2082 if (CastInst *CI = dyn_cast<CastInst>(CurrentValue)) { 2083 if (!CI->isNoopCast(CI->getModule()->getDataLayout())) 2084 return false; 2085 2086 ChainToBase.push_back(CI); 2087 return findRematerializableChainToBasePointer(ChainToBase, 2088 CI->getOperand(0), BaseValue); 2089 } 2090 2091 // Not supported instruction in the chain 2092 return false; 2093} 2094 2095// Helper function for the "rematerializeLiveValues". Compute cost of the use 2096// chain we are going to rematerialize. 2097static unsigned 2098chainToBasePointerCost(SmallVectorImpl<Instruction*> &Chain, 2099 TargetTransformInfo &TTI) { 2100 unsigned Cost = 0; 2101 2102 for (Instruction *Instr : Chain) { 2103 if (CastInst *CI = dyn_cast<CastInst>(Instr)) { 2104 assert(CI->isNoopCast(CI->getModule()->getDataLayout()) && 2105 "non noop cast is found during rematerialization"); 2106 2107 Type *SrcTy = CI->getOperand(0)->getType(); 2108 Cost += TTI.getCastInstrCost(CI->getOpcode(), CI->getType(), SrcTy); 2109 2110 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Instr)) { 2111 // Cost of the address calculation 2112 Type *ValTy = GEP->getPointerOperandType()->getPointerElementType(); 2113 Cost += TTI.getAddressComputationCost(ValTy); 2114 2115 // And cost of the GEP itself 2116 // TODO: Use TTI->getGEPCost here (it exists, but appears to be not 2117 // allowed for the external usage) 2118 if (!GEP->hasAllConstantIndices()) 2119 Cost += 2; 2120 2121 } else { 2122 llvm_unreachable("unsupported instruciton type during rematerialization"); 2123 } 2124 } 2125 2126 return Cost; 2127} 2128 2129// From the statepoint live set pick values that are cheaper to recompute then 2130// to relocate. Remove this values from the live set, rematerialize them after 2131// statepoint and record them in "Info" structure. Note that similar to 2132// relocated values we don't do any user adjustments here. 2133static void rematerializeLiveValues(CallSite CS, 2134 PartiallyConstructedSafepointRecord &Info, 2135 TargetTransformInfo &TTI) { 2136 const unsigned int ChainLengthThreshold = 10; 2137 2138 // Record values we are going to delete from this statepoint live set. 2139 // We can not di this in following loop due to iterator invalidation. 2140 SmallVector<Value *, 32> LiveValuesToBeDeleted; 2141 2142 for (Value *LiveValue: Info.LiveSet) { 2143 // For each live pointer find it's defining chain 2144 SmallVector<Instruction *, 3> ChainToBase; 2145 assert(Info.PointerToBase.count(LiveValue)); 2146 bool FoundChain = 2147 findRematerializableChainToBasePointer(ChainToBase, 2148 LiveValue, 2149 Info.PointerToBase[LiveValue]); 2150 // Nothing to do, or chain is too long 2151 if (!FoundChain || 2152 ChainToBase.size() == 0 || 2153 ChainToBase.size() > ChainLengthThreshold) 2154 continue; 2155 2156 // Compute cost of this chain 2157 unsigned Cost = chainToBasePointerCost(ChainToBase, TTI); 2158 // TODO: We can also account for cases when we will be able to remove some 2159 // of the rematerialized values by later optimization passes. I.e if 2160 // we rematerialized several intersecting chains. Or if original values 2161 // don't have any uses besides this statepoint. 2162 2163 // For invokes we need to rematerialize each chain twice - for normal and 2164 // for unwind basic blocks. Model this by multiplying cost by two. 2165 if (CS.isInvoke()) { 2166 Cost *= 2; 2167 } 2168 // If it's too expensive - skip it 2169 if (Cost >= RematerializationThreshold) 2170 continue; 2171 2172 // Remove value from the live set 2173 LiveValuesToBeDeleted.push_back(LiveValue); 2174 2175 // Clone instructions and record them inside "Info" structure 2176 2177 // Walk backwards to visit top-most instructions first 2178 std::reverse(ChainToBase.begin(), ChainToBase.end()); 2179 2180 // Utility function which clones all instructions from "ChainToBase" 2181 // and inserts them before "InsertBefore". Returns rematerialized value 2182 // which should be used after statepoint. 2183 auto rematerializeChain = [&ChainToBase](Instruction *InsertBefore) { 2184 Instruction *LastClonedValue = nullptr; 2185 Instruction *LastValue = nullptr; 2186 for (Instruction *Instr: ChainToBase) { 2187 // Only GEP's and casts are suported as we need to be careful to not 2188 // introduce any new uses of pointers not in the liveset. 2189 // Note that it's fine to introduce new uses of pointers which were 2190 // otherwise not used after this statepoint. 2191 assert(isa<GetElementPtrInst>(Instr) || isa<CastInst>(Instr)); 2192 2193 Instruction *ClonedValue = Instr->clone(); 2194 ClonedValue->insertBefore(InsertBefore); 2195 ClonedValue->setName(Instr->getName() + ".remat"); 2196 2197 // If it is not first instruction in the chain then it uses previously 2198 // cloned value. We should update it to use cloned value. 2199 if (LastClonedValue) { 2200 assert(LastValue); 2201 ClonedValue->replaceUsesOfWith(LastValue, LastClonedValue); 2202#ifndef NDEBUG 2203 // Assert that cloned instruction does not use any instructions from 2204 // this chain other than LastClonedValue 2205 for (auto OpValue : ClonedValue->operand_values()) { 2206 assert(std::find(ChainToBase.begin(), ChainToBase.end(), OpValue) == 2207 ChainToBase.end() && 2208 "incorrect use in rematerialization chain"); 2209 } 2210#endif 2211 } 2212 2213 LastClonedValue = ClonedValue; 2214 LastValue = Instr; 2215 } 2216 assert(LastClonedValue); 2217 return LastClonedValue; 2218 }; 2219 2220 // Different cases for calls and invokes. For invokes we need to clone 2221 // instructions both on normal and unwind path. 2222 if (CS.isCall()) { 2223 Instruction *InsertBefore = CS.getInstruction()->getNextNode(); 2224 assert(InsertBefore); 2225 Instruction *RematerializedValue = rematerializeChain(InsertBefore); 2226 Info.RematerializedValues[RematerializedValue] = LiveValue; 2227 } else { 2228 InvokeInst *Invoke = cast<InvokeInst>(CS.getInstruction()); 2229 2230 Instruction *NormalInsertBefore = 2231 &*Invoke->getNormalDest()->getFirstInsertionPt(); 2232 Instruction *UnwindInsertBefore = 2233 &*Invoke->getUnwindDest()->getFirstInsertionPt(); 2234 2235 Instruction *NormalRematerializedValue = 2236 rematerializeChain(NormalInsertBefore); 2237 Instruction *UnwindRematerializedValue = 2238 rematerializeChain(UnwindInsertBefore); 2239 2240 Info.RematerializedValues[NormalRematerializedValue] = LiveValue; 2241 Info.RematerializedValues[UnwindRematerializedValue] = LiveValue; 2242 } 2243 } 2244 2245 // Remove rematerializaed values from the live set 2246 for (auto LiveValue: LiveValuesToBeDeleted) { 2247 Info.LiveSet.erase(LiveValue); 2248 } 2249} 2250 2251static bool insertParsePoints(Function &F, DominatorTree &DT, 2252 TargetTransformInfo &TTI, 2253 SmallVectorImpl<CallSite> &ToUpdate) { 2254#ifndef NDEBUG 2255 // sanity check the input 2256 std::set<CallSite> Uniqued; 2257 Uniqued.insert(ToUpdate.begin(), ToUpdate.end()); 2258 assert(Uniqued.size() == ToUpdate.size() && "no duplicates please!"); 2259 2260 for (CallSite CS : ToUpdate) { 2261 assert(CS.getInstruction()->getParent()->getParent() == &F); 2262 assert((UseDeoptBundles || isStatepoint(CS)) && 2263 "expected to already be a deopt statepoint"); 2264 } 2265#endif 2266 2267 // When inserting gc.relocates for invokes, we need to be able to insert at 2268 // the top of the successor blocks. See the comment on 2269 // normalForInvokeSafepoint on exactly what is needed. Note that this step 2270 // may restructure the CFG. 2271 for (CallSite CS : ToUpdate) { 2272 if (!CS.isInvoke()) 2273 continue; 2274 auto *II = cast<InvokeInst>(CS.getInstruction()); 2275 normalizeForInvokeSafepoint(II->getNormalDest(), II->getParent(), DT); 2276 normalizeForInvokeSafepoint(II->getUnwindDest(), II->getParent(), DT); 2277 } 2278 2279 // A list of dummy calls added to the IR to keep various values obviously 2280 // live in the IR. We'll remove all of these when done. 2281 SmallVector<CallInst *, 64> Holders; 2282 2283 // Insert a dummy call with all of the arguments to the vm_state we'll need 2284 // for the actual safepoint insertion. This ensures reference arguments in 2285 // the deopt argument list are considered live through the safepoint (and 2286 // thus makes sure they get relocated.) 2287 for (CallSite CS : ToUpdate) { 2288 SmallVector<Value *, 64> DeoptValues; 2289 2290 iterator_range<const Use *> DeoptStateRange = 2291 UseDeoptBundles 2292 ? iterator_range<const Use *>(GetDeoptBundleOperands(CS)) 2293 : iterator_range<const Use *>(Statepoint(CS).vm_state_args()); 2294 2295 for (Value *Arg : DeoptStateRange) { 2296 assert(!isUnhandledGCPointerType(Arg->getType()) && 2297 "support for FCA unimplemented"); 2298 if (isHandledGCPointerType(Arg->getType())) 2299 DeoptValues.push_back(Arg); 2300 } 2301 2302 insertUseHolderAfter(CS, DeoptValues, Holders); 2303 } 2304 2305 SmallVector<PartiallyConstructedSafepointRecord, 64> Records(ToUpdate.size()); 2306 2307 // A) Identify all gc pointers which are statically live at the given call 2308 // site. 2309 findLiveReferences(F, DT, ToUpdate, Records); 2310 2311 // B) Find the base pointers for each live pointer 2312 /* scope for caching */ { 2313 // Cache the 'defining value' relation used in the computation and 2314 // insertion of base phis and selects. This ensures that we don't insert 2315 // large numbers of duplicate base_phis. 2316 DefiningValueMapTy DVCache; 2317 2318 for (size_t i = 0; i < Records.size(); i++) { 2319 PartiallyConstructedSafepointRecord &info = Records[i]; 2320 findBasePointers(DT, DVCache, ToUpdate[i], info); 2321 } 2322 } // end of cache scope 2323 2324 // The base phi insertion logic (for any safepoint) may have inserted new 2325 // instructions which are now live at some safepoint. The simplest such 2326 // example is: 2327 // loop: 2328 // phi a <-- will be a new base_phi here 2329 // safepoint 1 <-- that needs to be live here 2330 // gep a + 1 2331 // safepoint 2 2332 // br loop 2333 // We insert some dummy calls after each safepoint to definitely hold live 2334 // the base pointers which were identified for that safepoint. We'll then 2335 // ask liveness for _every_ base inserted to see what is now live. Then we 2336 // remove the dummy calls. 2337 Holders.reserve(Holders.size() + Records.size()); 2338 for (size_t i = 0; i < Records.size(); i++) { 2339 PartiallyConstructedSafepointRecord &Info = Records[i]; 2340 2341 SmallVector<Value *, 128> Bases; 2342 for (auto Pair : Info.PointerToBase) 2343 Bases.push_back(Pair.second); 2344 2345 insertUseHolderAfter(ToUpdate[i], Bases, Holders); 2346 } 2347 2348 // By selecting base pointers, we've effectively inserted new uses. Thus, we 2349 // need to rerun liveness. We may *also* have inserted new defs, but that's 2350 // not the key issue. 2351 recomputeLiveInValues(F, DT, ToUpdate, Records); 2352 2353 if (PrintBasePointers) { 2354 for (auto &Info : Records) { 2355 errs() << "Base Pairs: (w/Relocation)\n"; 2356 for (auto Pair : Info.PointerToBase) { 2357 errs() << " derived "; 2358 Pair.first->printAsOperand(errs(), false); 2359 errs() << " base "; 2360 Pair.second->printAsOperand(errs(), false); 2361 errs() << "\n"; 2362 } 2363 } 2364 } 2365 2366 // It is possible that non-constant live variables have a constant base. For 2367 // example, a GEP with a variable offset from a global. In this case we can 2368 // remove it from the liveset. We already don't add constants to the liveset 2369 // because we assume they won't move at runtime and the GC doesn't need to be 2370 // informed about them. The same reasoning applies if the base is constant. 2371 // Note that the relocation placement code relies on this filtering for 2372 // correctness as it expects the base to be in the liveset, which isn't true 2373 // if the base is constant. 2374 for (auto &Info : Records) 2375 for (auto &BasePair : Info.PointerToBase) 2376 if (isa<Constant>(BasePair.second)) 2377 Info.LiveSet.erase(BasePair.first); 2378 2379 for (CallInst *CI : Holders) 2380 CI->eraseFromParent(); 2381 2382 Holders.clear(); 2383 2384 // Do a limited scalarization of any live at safepoint vector values which 2385 // contain pointers. This enables this pass to run after vectorization at 2386 // the cost of some possible performance loss. Note: This is known to not 2387 // handle updating of the side tables correctly which can lead to relocation 2388 // bugs when the same vector is live at multiple statepoints. We're in the 2389 // process of implementing the alternate lowering - relocating the 2390 // vector-of-pointers as first class item and updating the backend to 2391 // understand that - but that's not yet complete. 2392 if (UseVectorSplit) 2393 for (size_t i = 0; i < Records.size(); i++) { 2394 PartiallyConstructedSafepointRecord &Info = Records[i]; 2395 Instruction *Statepoint = ToUpdate[i].getInstruction(); 2396 splitVectorValues(cast<Instruction>(Statepoint), Info.LiveSet, 2397 Info.PointerToBase, DT); 2398 } 2399 2400 // In order to reduce live set of statepoint we might choose to rematerialize 2401 // some values instead of relocating them. This is purely an optimization and 2402 // does not influence correctness. 2403 for (size_t i = 0; i < Records.size(); i++) 2404 rematerializeLiveValues(ToUpdate[i], Records[i], TTI); 2405 2406 // We need this to safely RAUW and delete call or invoke return values that 2407 // may themselves be live over a statepoint. For details, please see usage in 2408 // makeStatepointExplicitImpl. 2409 std::vector<DeferredReplacement> Replacements; 2410 2411 // Now run through and replace the existing statepoints with new ones with 2412 // the live variables listed. We do not yet update uses of the values being 2413 // relocated. We have references to live variables that need to 2414 // survive to the last iteration of this loop. (By construction, the 2415 // previous statepoint can not be a live variable, thus we can and remove 2416 // the old statepoint calls as we go.) 2417 for (size_t i = 0; i < Records.size(); i++) 2418 makeStatepointExplicit(DT, ToUpdate[i], Records[i], Replacements); 2419 2420 ToUpdate.clear(); // prevent accident use of invalid CallSites 2421 2422 for (auto &PR : Replacements) 2423 PR.doReplacement(); 2424 2425 Replacements.clear(); 2426 2427 for (auto &Info : Records) { 2428 // These live sets may contain state Value pointers, since we replaced calls 2429 // with operand bundles with calls wrapped in gc.statepoint, and some of 2430 // those calls may have been def'ing live gc pointers. Clear these out to 2431 // avoid accidentally using them. 2432 // 2433 // TODO: We should create a separate data structure that does not contain 2434 // these live sets, and migrate to using that data structure from this point 2435 // onward. 2436 Info.LiveSet.clear(); 2437 Info.PointerToBase.clear(); 2438 } 2439 2440 // Do all the fixups of the original live variables to their relocated selves 2441 SmallVector<Value *, 128> Live; 2442 for (size_t i = 0; i < Records.size(); i++) { 2443 PartiallyConstructedSafepointRecord &Info = Records[i]; 2444 2445 // We can't simply save the live set from the original insertion. One of 2446 // the live values might be the result of a call which needs a safepoint. 2447 // That Value* no longer exists and we need to use the new gc_result. 2448 // Thankfully, the live set is embedded in the statepoint (and updated), so 2449 // we just grab that. 2450 Statepoint Statepoint(Info.StatepointToken); 2451 Live.insert(Live.end(), Statepoint.gc_args_begin(), 2452 Statepoint.gc_args_end()); 2453#ifndef NDEBUG 2454 // Do some basic sanity checks on our liveness results before performing 2455 // relocation. Relocation can and will turn mistakes in liveness results 2456 // into non-sensical code which is must harder to debug. 2457 // TODO: It would be nice to test consistency as well 2458 assert(DT.isReachableFromEntry(Info.StatepointToken->getParent()) && 2459 "statepoint must be reachable or liveness is meaningless"); 2460 for (Value *V : Statepoint.gc_args()) { 2461 if (!isa<Instruction>(V)) 2462 // Non-instruction values trivial dominate all possible uses 2463 continue; 2464 auto *LiveInst = cast<Instruction>(V); 2465 assert(DT.isReachableFromEntry(LiveInst->getParent()) && 2466 "unreachable values should never be live"); 2467 assert(DT.dominates(LiveInst, Info.StatepointToken) && 2468 "basic SSA liveness expectation violated by liveness analysis"); 2469 } 2470#endif 2471 } 2472 unique_unsorted(Live); 2473 2474#ifndef NDEBUG 2475 // sanity check 2476 for (auto *Ptr : Live) 2477 assert(isHandledGCPointerType(Ptr->getType()) && 2478 "must be a gc pointer type"); 2479#endif 2480 2481 relocationViaAlloca(F, DT, Live, Records); 2482 return !Records.empty(); 2483} 2484 2485// Handles both return values and arguments for Functions and CallSites. 2486template <typename AttrHolder> 2487static void RemoveNonValidAttrAtIndex(LLVMContext &Ctx, AttrHolder &AH, 2488 unsigned Index) { 2489 AttrBuilder R; 2490 if (AH.getDereferenceableBytes(Index)) 2491 R.addAttribute(Attribute::get(Ctx, Attribute::Dereferenceable, 2492 AH.getDereferenceableBytes(Index))); 2493 if (AH.getDereferenceableOrNullBytes(Index)) 2494 R.addAttribute(Attribute::get(Ctx, Attribute::DereferenceableOrNull, 2495 AH.getDereferenceableOrNullBytes(Index))); 2496 if (AH.doesNotAlias(Index)) 2497 R.addAttribute(Attribute::NoAlias); 2498 2499 if (!R.empty()) 2500 AH.setAttributes(AH.getAttributes().removeAttributes( 2501 Ctx, Index, AttributeSet::get(Ctx, Index, R))); 2502} 2503 2504void 2505RewriteStatepointsForGC::stripNonValidAttributesFromPrototype(Function &F) { 2506 LLVMContext &Ctx = F.getContext(); 2507 2508 for (Argument &A : F.args()) 2509 if (isa<PointerType>(A.getType())) 2510 RemoveNonValidAttrAtIndex(Ctx, F, A.getArgNo() + 1); 2511 2512 if (isa<PointerType>(F.getReturnType())) 2513 RemoveNonValidAttrAtIndex(Ctx, F, AttributeSet::ReturnIndex); 2514} 2515 2516void RewriteStatepointsForGC::stripNonValidAttributesFromBody(Function &F) { 2517 if (F.empty()) 2518 return; 2519 2520 LLVMContext &Ctx = F.getContext(); 2521 MDBuilder Builder(Ctx); 2522 2523 for (Instruction &I : instructions(F)) { 2524 if (const MDNode *MD = I.getMetadata(LLVMContext::MD_tbaa)) { 2525 assert(MD->getNumOperands() < 5 && "unrecognized metadata shape!"); 2526 bool IsImmutableTBAA = 2527 MD->getNumOperands() == 4 && 2528 mdconst::extract<ConstantInt>(MD->getOperand(3))->getValue() == 1; 2529 2530 if (!IsImmutableTBAA) 2531 continue; // no work to do, MD_tbaa is already marked mutable 2532 2533 MDNode *Base = cast<MDNode>(MD->getOperand(0)); 2534 MDNode *Access = cast<MDNode>(MD->getOperand(1)); 2535 uint64_t Offset = 2536 mdconst::extract<ConstantInt>(MD->getOperand(2))->getZExtValue(); 2537 2538 MDNode *MutableTBAA = 2539 Builder.createTBAAStructTagNode(Base, Access, Offset); 2540 I.setMetadata(LLVMContext::MD_tbaa, MutableTBAA); 2541 } 2542 2543 if (CallSite CS = CallSite(&I)) { 2544 for (int i = 0, e = CS.arg_size(); i != e; i++) 2545 if (isa<PointerType>(CS.getArgument(i)->getType())) 2546 RemoveNonValidAttrAtIndex(Ctx, CS, i + 1); 2547 if (isa<PointerType>(CS.getType())) 2548 RemoveNonValidAttrAtIndex(Ctx, CS, AttributeSet::ReturnIndex); 2549 } 2550 } 2551} 2552 2553/// Returns true if this function should be rewritten by this pass. The main 2554/// point of this function is as an extension point for custom logic. 2555static bool shouldRewriteStatepointsIn(Function &F) { 2556 // TODO: This should check the GCStrategy 2557 if (F.hasGC()) { 2558 const auto &FunctionGCName = F.getGC(); 2559 const StringRef StatepointExampleName("statepoint-example"); 2560 const StringRef CoreCLRName("coreclr"); 2561 return (StatepointExampleName == FunctionGCName) || 2562 (CoreCLRName == FunctionGCName); 2563 } else 2564 return false; 2565} 2566 2567void RewriteStatepointsForGC::stripNonValidAttributes(Module &M) { 2568#ifndef NDEBUG 2569 assert(std::any_of(M.begin(), M.end(), shouldRewriteStatepointsIn) && 2570 "precondition!"); 2571#endif 2572 2573 for (Function &F : M) 2574 stripNonValidAttributesFromPrototype(F); 2575 2576 for (Function &F : M) 2577 stripNonValidAttributesFromBody(F); 2578} 2579 2580bool RewriteStatepointsForGC::runOnFunction(Function &F) { 2581 // Nothing to do for declarations. 2582 if (F.isDeclaration() || F.empty()) 2583 return false; 2584 2585 // Policy choice says not to rewrite - the most common reason is that we're 2586 // compiling code without a GCStrategy. 2587 if (!shouldRewriteStatepointsIn(F)) 2588 return false; 2589 2590 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree(); 2591 TargetTransformInfo &TTI = 2592 getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 2593 2594 auto NeedsRewrite = [](Instruction &I) { 2595 if (UseDeoptBundles) { 2596 if (ImmutableCallSite CS = ImmutableCallSite(&I)) 2597 return !callsGCLeafFunction(CS); 2598 return false; 2599 } 2600 2601 return isStatepoint(I); 2602 }; 2603 2604 // Gather all the statepoints which need rewritten. Be careful to only 2605 // consider those in reachable code since we need to ask dominance queries 2606 // when rewriting. We'll delete the unreachable ones in a moment. 2607 SmallVector<CallSite, 64> ParsePointNeeded; 2608 bool HasUnreachableStatepoint = false; 2609 for (Instruction &I : instructions(F)) { 2610 // TODO: only the ones with the flag set! 2611 if (NeedsRewrite(I)) { 2612 if (DT.isReachableFromEntry(I.getParent())) 2613 ParsePointNeeded.push_back(CallSite(&I)); 2614 else 2615 HasUnreachableStatepoint = true; 2616 } 2617 } 2618 2619 bool MadeChange = false; 2620 2621 // Delete any unreachable statepoints so that we don't have unrewritten 2622 // statepoints surviving this pass. This makes testing easier and the 2623 // resulting IR less confusing to human readers. Rather than be fancy, we 2624 // just reuse a utility function which removes the unreachable blocks. 2625 if (HasUnreachableStatepoint) 2626 MadeChange |= removeUnreachableBlocks(F); 2627 2628 // Return early if no work to do. 2629 if (ParsePointNeeded.empty()) 2630 return MadeChange; 2631 2632 // As a prepass, go ahead and aggressively destroy single entry phi nodes. 2633 // These are created by LCSSA. They have the effect of increasing the size 2634 // of liveness sets for no good reason. It may be harder to do this post 2635 // insertion since relocations and base phis can confuse things. 2636 for (BasicBlock &BB : F) 2637 if (BB.getUniquePredecessor()) { 2638 MadeChange = true; 2639 FoldSingleEntryPHINodes(&BB); 2640 } 2641 2642 // Before we start introducing relocations, we want to tweak the IR a bit to 2643 // avoid unfortunate code generation effects. The main example is that we 2644 // want to try to make sure the comparison feeding a branch is after any 2645 // safepoints. Otherwise, we end up with a comparison of pre-relocation 2646 // values feeding a branch after relocation. This is semantically correct, 2647 // but results in extra register pressure since both the pre-relocation and 2648 // post-relocation copies must be available in registers. For code without 2649 // relocations this is handled elsewhere, but teaching the scheduler to 2650 // reverse the transform we're about to do would be slightly complex. 2651 // Note: This may extend the live range of the inputs to the icmp and thus 2652 // increase the liveset of any statepoint we move over. This is profitable 2653 // as long as all statepoints are in rare blocks. If we had in-register 2654 // lowering for live values this would be a much safer transform. 2655 auto getConditionInst = [](TerminatorInst *TI) -> Instruction* { 2656 if (auto *BI = dyn_cast<BranchInst>(TI)) 2657 if (BI->isConditional()) 2658 return dyn_cast<Instruction>(BI->getCondition()); 2659 // TODO: Extend this to handle switches 2660 return nullptr; 2661 }; 2662 for (BasicBlock &BB : F) { 2663 TerminatorInst *TI = BB.getTerminator(); 2664 if (auto *Cond = getConditionInst(TI)) 2665 // TODO: Handle more than just ICmps here. We should be able to move 2666 // most instructions without side effects or memory access. 2667 if (isa<ICmpInst>(Cond) && Cond->hasOneUse()) { 2668 MadeChange = true; 2669 Cond->moveBefore(TI); 2670 } 2671 } 2672 2673 MadeChange |= insertParsePoints(F, DT, TTI, ParsePointNeeded); 2674 return MadeChange; 2675} 2676 2677// liveness computation via standard dataflow 2678// ------------------------------------------------------------------- 2679 2680// TODO: Consider using bitvectors for liveness, the set of potentially 2681// interesting values should be small and easy to pre-compute. 2682 2683/// Compute the live-in set for the location rbegin starting from 2684/// the live-out set of the basic block 2685static void computeLiveInValues(BasicBlock::reverse_iterator rbegin, 2686 BasicBlock::reverse_iterator rend, 2687 DenseSet<Value *> &LiveTmp) { 2688 2689 for (BasicBlock::reverse_iterator ritr = rbegin; ritr != rend; ritr++) { 2690 Instruction *I = &*ritr; 2691 2692 // KILL/Def - Remove this definition from LiveIn 2693 LiveTmp.erase(I); 2694 2695 // Don't consider *uses* in PHI nodes, we handle their contribution to 2696 // predecessor blocks when we seed the LiveOut sets 2697 if (isa<PHINode>(I)) 2698 continue; 2699 2700 // USE - Add to the LiveIn set for this instruction 2701 for (Value *V : I->operands()) { 2702 assert(!isUnhandledGCPointerType(V->getType()) && 2703 "support for FCA unimplemented"); 2704 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) { 2705 // The choice to exclude all things constant here is slightly subtle. 2706 // There are two independent reasons: 2707 // - We assume that things which are constant (from LLVM's definition) 2708 // do not move at runtime. For example, the address of a global 2709 // variable is fixed, even though it's contents may not be. 2710 // - Second, we can't disallow arbitrary inttoptr constants even 2711 // if the language frontend does. Optimization passes are free to 2712 // locally exploit facts without respect to global reachability. This 2713 // can create sections of code which are dynamically unreachable and 2714 // contain just about anything. (see constants.ll in tests) 2715 LiveTmp.insert(V); 2716 } 2717 } 2718 } 2719} 2720 2721static void computeLiveOutSeed(BasicBlock *BB, DenseSet<Value *> &LiveTmp) { 2722 2723 for (BasicBlock *Succ : successors(BB)) { 2724 const BasicBlock::iterator E(Succ->getFirstNonPHI()); 2725 for (BasicBlock::iterator I = Succ->begin(); I != E; I++) { 2726 PHINode *Phi = cast<PHINode>(&*I); 2727 Value *V = Phi->getIncomingValueForBlock(BB); 2728 assert(!isUnhandledGCPointerType(V->getType()) && 2729 "support for FCA unimplemented"); 2730 if (isHandledGCPointerType(V->getType()) && !isa<Constant>(V)) { 2731 LiveTmp.insert(V); 2732 } 2733 } 2734 } 2735} 2736 2737static DenseSet<Value *> computeKillSet(BasicBlock *BB) { 2738 DenseSet<Value *> KillSet; 2739 for (Instruction &I : *BB) 2740 if (isHandledGCPointerType(I.getType())) 2741 KillSet.insert(&I); 2742 return KillSet; 2743} 2744 2745#ifndef NDEBUG 2746/// Check that the items in 'Live' dominate 'TI'. This is used as a basic 2747/// sanity check for the liveness computation. 2748static void checkBasicSSA(DominatorTree &DT, DenseSet<Value *> &Live, 2749 TerminatorInst *TI, bool TermOkay = false) { 2750 for (Value *V : Live) { 2751 if (auto *I = dyn_cast<Instruction>(V)) { 2752 // The terminator can be a member of the LiveOut set. LLVM's definition 2753 // of instruction dominance states that V does not dominate itself. As 2754 // such, we need to special case this to allow it. 2755 if (TermOkay && TI == I) 2756 continue; 2757 assert(DT.dominates(I, TI) && 2758 "basic SSA liveness expectation violated by liveness analysis"); 2759 } 2760 } 2761} 2762 2763/// Check that all the liveness sets used during the computation of liveness 2764/// obey basic SSA properties. This is useful for finding cases where we miss 2765/// a def. 2766static void checkBasicSSA(DominatorTree &DT, GCPtrLivenessData &Data, 2767 BasicBlock &BB) { 2768 checkBasicSSA(DT, Data.LiveSet[&BB], BB.getTerminator()); 2769 checkBasicSSA(DT, Data.LiveOut[&BB], BB.getTerminator(), true); 2770 checkBasicSSA(DT, Data.LiveIn[&BB], BB.getTerminator()); 2771} 2772#endif 2773 2774static void computeLiveInValues(DominatorTree &DT, Function &F, 2775 GCPtrLivenessData &Data) { 2776 2777 SmallSetVector<BasicBlock *, 200> Worklist; 2778 auto AddPredsToWorklist = [&](BasicBlock *BB) { 2779 // We use a SetVector so that we don't have duplicates in the worklist. 2780 Worklist.insert(pred_begin(BB), pred_end(BB)); 2781 }; 2782 auto NextItem = [&]() { 2783 BasicBlock *BB = Worklist.back(); 2784 Worklist.pop_back(); 2785 return BB; 2786 }; 2787 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]); 2793 2794#ifndef NDEBUG 2795 for (Value *Kill : Data.KillSet[&BB]) 2796 assert(!Data.LiveSet[&BB].count(Kill) && "live set contains kill"); 2797#endif 2798 2799 Data.LiveOut[&BB] = DenseSet<Value *>(); 2800 computeLiveOutSeed(&BB, Data.LiveOut[&BB]); 2801 Data.LiveIn[&BB] = Data.LiveSet[&BB]; 2802 set_union(Data.LiveIn[&BB], Data.LiveOut[&BB]); 2803 set_subtract(Data.LiveIn[&BB], Data.KillSet[&BB]); 2804 if (!Data.LiveIn[&BB].empty()) 2805 AddPredsToWorklist(&BB); 2806 } 2807 2808 // Propagate that liveness until stable 2809 while (!Worklist.empty()) { 2810 BasicBlock *BB = NextItem(); 2811 2812 // Compute our new liveout set, then exit early if it hasn't changed 2813 // despite the contribution of our successor. 2814 DenseSet<Value *> LiveOut = Data.LiveOut[BB]; 2815 const auto OldLiveOutSize = LiveOut.size(); 2816 for (BasicBlock *Succ : successors(BB)) { 2817 assert(Data.LiveIn.count(Succ)); 2818 set_union(LiveOut, Data.LiveIn[Succ]); 2819 } 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 2824 // hasn't changed. 2825 continue; 2826 } 2827 Data.LiveOut[BB] = LiveOut; 2828 2829 // Apply the effects of this basic block 2830 DenseSet<Value *> LiveTmp = LiveOut; 2831 set_union(LiveTmp, Data.LiveSet[BB]); 2832 set_subtract(LiveTmp, Data.KillSet[BB]); 2833 2834 assert(Data.LiveIn.count(BB)); 2835 const DenseSet<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 AddPredsToWorklist(BB); 2840 } 2841 } // while( !worklist.empty() ) 2842 2843#ifndef NDEBUG 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); 2848 } 2849#endif 2850} 2851 2852static void findLiveSetAtInst(Instruction *Inst, GCPtrLivenessData &Data, 2853 StatepointLiveSetTy &Out) { 2854 2855 BasicBlock *BB = Inst->getParent(); 2856 2857 // Note: The copy is intentional and required 2858 assert(Data.LiveOut.count(BB)); 2859 DenseSet<Value *> LiveOut = Data.LiveOut[BB]; 2860 2861 // We want to handle the statepoint itself oddly. It's 2862 // call result is not live (normal), nor are it's arguments 2863 // (unless they're used again later). This adjustment is 2864 // specifically what we need to relocate 2865 BasicBlock::reverse_iterator rend(Inst->getIterator()); 2866 computeLiveInValues(BB->rbegin(), rend, LiveOut); 2867 LiveOut.erase(Inst); 2868 Out.insert(LiveOut.begin(), LiveOut.end()); 2869} 2870 2871static void recomputeLiveInValues(GCPtrLivenessData &RevisedLivenessData, 2872 const CallSite &CS, 2873 PartiallyConstructedSafepointRecord &Info) { 2874 Instruction *Inst = CS.getInstruction(); 2875 StatepointLiveSetTy Updated; 2876 findLiveSetAtInst(Inst, RevisedLivenessData, Updated); 2877 2878#ifndef NDEBUG 2879 DenseSet<Value *> Bases; 2880 for (auto KVPair : Info.PointerToBase) { 2881 Bases.insert(KVPair.second); 2882 } 2883#endif 2884 // We may have base pointers which are now live that weren't before. We need 2885 // to update the PointerToBase structure to reflect this. 2886 for (auto V : Updated) 2887 if (!Info.PointerToBase.count(V)) { 2888 assert(Bases.count(V) && "can't find base for unexpected live value"); 2889 Info.PointerToBase[V] = V; 2890 continue; 2891 } 2892 2893#ifndef NDEBUG 2894 for (auto V : Updated) { 2895 assert(Info.PointerToBase.count(V) && 2896 "must be able to find base for live value"); 2897 } 2898#endif 2899 2900 // Remove any stale base mappings - this can happen since our liveness is 2901 // more precise then the one inherent in the base pointer analysis 2902 DenseSet<Value *> ToErase; 2903 for (auto KVPair : Info.PointerToBase) 2904 if (!Updated.count(KVPair.first)) 2905 ToErase.insert(KVPair.first); 2906 for (auto V : ToErase) 2907 Info.PointerToBase.erase(V); 2908 2909#ifndef NDEBUG 2910 for (auto KVPair : Info.PointerToBase) 2911 assert(Updated.count(KVPair.first) && "record for non-live value"); 2912#endif 2913 2914 Info.LiveSet = Updated; 2915} 2916