1 | //===-- SafepointIRVerifier.cpp - Verify gc.statepoint invariants ---------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // Run a basic correctness check on the IR to ensure that Safepoints - if |
10 | // they've been inserted - were inserted correctly. In particular, look for use |
11 | // of non-relocated values after a safepoint. It's primary use is to check the |
12 | // correctness of safepoint insertion immediately after insertion, but it can |
13 | // also be used to verify that later transforms have not found a way to break |
14 | // safepoint semenatics. |
15 | // |
16 | // In its current form, this verify checks a property which is sufficient, but |
17 | // not neccessary for correctness. There are some cases where an unrelocated |
18 | // pointer can be used after the safepoint. Consider this example: |
19 | // |
20 | // a = ... |
21 | // b = ... |
22 | // (a',b') = safepoint(a,b) |
23 | // c = cmp eq a b |
24 | // br c, ..., .... |
25 | // |
26 | // Because it is valid to reorder 'c' above the safepoint, this is legal. In |
27 | // practice, this is a somewhat uncommon transform, but CodeGenPrep does create |
28 | // idioms like this. The verifier knows about these cases and avoids reporting |
29 | // false positives. |
30 | // |
31 | //===----------------------------------------------------------------------===// |
32 | |
33 | #include "llvm/IR/SafepointIRVerifier.h" |
34 | #include "llvm/ADT/DenseSet.h" |
35 | #include "llvm/ADT/PostOrderIterator.h" |
36 | #include "llvm/ADT/SetOperations.h" |
37 | #include "llvm/ADT/SetVector.h" |
38 | #include "llvm/IR/BasicBlock.h" |
39 | #include "llvm/IR/Dominators.h" |
40 | #include "llvm/IR/Function.h" |
41 | #include "llvm/IR/InstrTypes.h" |
42 | #include "llvm/IR/Instructions.h" |
43 | #include "llvm/IR/Statepoint.h" |
44 | #include "llvm/IR/Value.h" |
45 | #include "llvm/InitializePasses.h" |
46 | #include "llvm/Support/Allocator.h" |
47 | #include "llvm/Support/CommandLine.h" |
48 | #include "llvm/Support/Debug.h" |
49 | #include "llvm/Support/raw_ostream.h" |
50 | |
51 | #define DEBUG_TYPE "safepoint-ir-verifier" |
52 | |
53 | using namespace llvm; |
54 | |
55 | /// This option is used for writing test cases. Instead of crashing the program |
56 | /// when verification fails, report a message to the console (for FileCheck |
57 | /// usage) and continue execution as if nothing happened. |
58 | static cl::opt<bool> PrintOnly("safepoint-ir-verifier-print-only" , |
59 | cl::init(Val: false)); |
60 | |
61 | namespace { |
62 | |
63 | /// This CFG Deadness finds dead blocks and edges. Algorithm starts with a set |
64 | /// of blocks unreachable from entry then propagates deadness using foldable |
65 | /// conditional branches without modifying CFG. So GVN does but it changes CFG |
66 | /// by splitting critical edges. In most cases passes rely on SimplifyCFG to |
67 | /// clean up dead blocks, but in some cases, like verification or loop passes |
68 | /// it's not possible. |
69 | class CFGDeadness { |
70 | const DominatorTree *DT = nullptr; |
71 | SetVector<const BasicBlock *> DeadBlocks; |
72 | SetVector<const Use *> DeadEdges; // Contains all dead edges from live blocks. |
73 | |
74 | public: |
75 | /// Return the edge that coresponds to the predecessor. |
76 | static const Use& getEdge(const_pred_iterator &PredIt) { |
77 | auto &PU = PredIt.getUse(); |
78 | return PU.getUser()->getOperandUse(i: PU.getOperandNo()); |
79 | } |
80 | |
81 | /// Return true if there is at least one live edge that corresponds to the |
82 | /// basic block InBB listed in the phi node. |
83 | bool hasLiveIncomingEdge(const PHINode *PN, const BasicBlock *InBB) const { |
84 | assert(!isDeadBlock(InBB) && "block must be live" ); |
85 | const BasicBlock* BB = PN->getParent(); |
86 | bool Listed = false; |
87 | for (const_pred_iterator PredIt(BB), End(BB, true); PredIt != End; ++PredIt) { |
88 | if (InBB == *PredIt) { |
89 | if (!isDeadEdge(U: &getEdge(PredIt))) |
90 | return true; |
91 | Listed = true; |
92 | } |
93 | } |
94 | (void)Listed; |
95 | assert(Listed && "basic block is not found among incoming blocks" ); |
96 | return false; |
97 | } |
98 | |
99 | |
100 | bool isDeadBlock(const BasicBlock *BB) const { |
101 | return DeadBlocks.count(key: BB); |
102 | } |
103 | |
104 | bool isDeadEdge(const Use *U) const { |
105 | assert(cast<Instruction>(U->getUser())->isTerminator() && |
106 | "edge must be operand of terminator" ); |
107 | assert(cast_or_null<BasicBlock>(U->get()) && |
108 | "edge must refer to basic block" ); |
109 | assert(!isDeadBlock(cast<Instruction>(U->getUser())->getParent()) && |
110 | "isDeadEdge() must be applied to edge from live block" ); |
111 | return DeadEdges.count(key: U); |
112 | } |
113 | |
114 | bool hasLiveIncomingEdges(const BasicBlock *BB) const { |
115 | // Check if all incoming edges are dead. |
116 | for (const_pred_iterator PredIt(BB), End(BB, true); PredIt != End; ++PredIt) { |
117 | auto &PU = PredIt.getUse(); |
118 | const Use &U = PU.getUser()->getOperandUse(i: PU.getOperandNo()); |
119 | if (!isDeadBlock(BB: *PredIt) && !isDeadEdge(U: &U)) |
120 | return true; // Found a live edge. |
121 | } |
122 | return false; |
123 | } |
124 | |
125 | void processFunction(const Function &F, const DominatorTree &DT) { |
126 | this->DT = &DT; |
127 | |
128 | // Start with all blocks unreachable from entry. |
129 | for (const BasicBlock &BB : F) |
130 | if (!DT.isReachableFromEntry(A: &BB)) |
131 | DeadBlocks.insert(X: &BB); |
132 | |
133 | // Top-down walk of the dominator tree |
134 | ReversePostOrderTraversal<const Function *> RPOT(&F); |
135 | for (const BasicBlock *BB : RPOT) { |
136 | const Instruction *TI = BB->getTerminator(); |
137 | assert(TI && "blocks must be well formed" ); |
138 | |
139 | // For conditional branches, we can perform simple conditional propagation on |
140 | // the condition value itself. |
141 | const BranchInst *BI = dyn_cast<BranchInst>(Val: TI); |
142 | if (!BI || !BI->isConditional() || !isa<Constant>(Val: BI->getCondition())) |
143 | continue; |
144 | |
145 | // If a branch has two identical successors, we cannot declare either dead. |
146 | if (BI->getSuccessor(i: 0) == BI->getSuccessor(i: 1)) |
147 | continue; |
148 | |
149 | ConstantInt *Cond = dyn_cast<ConstantInt>(Val: BI->getCondition()); |
150 | if (!Cond) |
151 | continue; |
152 | |
153 | addDeadEdge(DeadEdge: BI->getOperandUse(i: Cond->getZExtValue() ? 1 : 2)); |
154 | } |
155 | } |
156 | |
157 | protected: |
158 | void addDeadBlock(const BasicBlock *BB) { |
159 | SmallVector<const BasicBlock *, 4> NewDead; |
160 | |
161 | NewDead.push_back(Elt: BB); |
162 | while (!NewDead.empty()) { |
163 | const BasicBlock *D = NewDead.pop_back_val(); |
164 | if (isDeadBlock(BB: D)) |
165 | continue; |
166 | |
167 | // All blocks dominated by D are dead. |
168 | SmallVector<BasicBlock *, 8> Dom; |
169 | DT->getDescendants(R: const_cast<BasicBlock*>(D), Result&: Dom); |
170 | // Do not need to mark all in and out edges dead |
171 | // because BB is marked dead and this is enough |
172 | // to run further. |
173 | DeadBlocks.insert_range(R&: Dom); |
174 | |
175 | // Figure out the dominance-frontier(D). |
176 | for (BasicBlock *B : Dom) |
177 | for (BasicBlock *S : successors(BB: B)) |
178 | if (!isDeadBlock(BB: S) && !hasLiveIncomingEdges(BB: S)) |
179 | NewDead.push_back(Elt: S); |
180 | } |
181 | } |
182 | |
183 | void addDeadEdge(const Use &DeadEdge) { |
184 | if (!DeadEdges.insert(X: &DeadEdge)) |
185 | return; |
186 | |
187 | BasicBlock *BB = cast_or_null<BasicBlock>(Val: DeadEdge.get()); |
188 | if (hasLiveIncomingEdges(BB)) |
189 | return; |
190 | |
191 | addDeadBlock(BB); |
192 | } |
193 | }; |
194 | } // namespace |
195 | |
196 | static void Verify(const Function &F, const DominatorTree &DT, |
197 | const CFGDeadness &CD); |
198 | |
199 | namespace llvm { |
200 | PreservedAnalyses SafepointIRVerifierPass::run(Function &F, |
201 | FunctionAnalysisManager &AM) { |
202 | const auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
203 | CFGDeadness CD; |
204 | CD.processFunction(F, DT); |
205 | Verify(F, DT, CD); |
206 | return PreservedAnalyses::all(); |
207 | } |
208 | } // namespace llvm |
209 | |
210 | namespace { |
211 | |
212 | struct SafepointIRVerifier : public FunctionPass { |
213 | static char ID; // Pass identification, replacement for typeid |
214 | SafepointIRVerifier() : FunctionPass(ID) { |
215 | initializeSafepointIRVerifierPass(*PassRegistry::getPassRegistry()); |
216 | } |
217 | |
218 | bool runOnFunction(Function &F) override { |
219 | auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); |
220 | CFGDeadness CD; |
221 | CD.processFunction(F, DT); |
222 | Verify(F, DT, CD); |
223 | return false; // no modifications |
224 | } |
225 | |
226 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
227 | AU.addRequiredID(ID&: DominatorTreeWrapperPass::ID); |
228 | AU.setPreservesAll(); |
229 | } |
230 | |
231 | StringRef getPassName() const override { return "safepoint verifier" ; } |
232 | }; |
233 | } // namespace |
234 | |
235 | void llvm::verifySafepointIR(Function &F) { |
236 | SafepointIRVerifier pass; |
237 | pass.runOnFunction(F); |
238 | } |
239 | |
240 | char SafepointIRVerifier::ID = 0; |
241 | |
242 | FunctionPass *llvm::createSafepointIRVerifierPass() { |
243 | return new SafepointIRVerifier(); |
244 | } |
245 | |
246 | INITIALIZE_PASS_BEGIN(SafepointIRVerifier, "verify-safepoint-ir" , |
247 | "Safepoint IR Verifier" , false, false) |
248 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) |
249 | INITIALIZE_PASS_END(SafepointIRVerifier, "verify-safepoint-ir" , |
250 | "Safepoint IR Verifier" , false, false) |
251 | |
252 | static bool isGCPointerType(Type *T) { |
253 | if (auto *PT = dyn_cast<PointerType>(Val: T)) |
254 | // For the sake of this example GC, we arbitrarily pick addrspace(1) as our |
255 | // GC managed heap. We know that a pointer into this heap needs to be |
256 | // updated and that no other pointer does. |
257 | return (1 == PT->getAddressSpace()); |
258 | return false; |
259 | } |
260 | |
261 | static bool containsGCPtrType(Type *Ty) { |
262 | if (isGCPointerType(T: Ty)) |
263 | return true; |
264 | if (VectorType *VT = dyn_cast<VectorType>(Val: Ty)) |
265 | return isGCPointerType(T: VT->getScalarType()); |
266 | if (ArrayType *AT = dyn_cast<ArrayType>(Val: Ty)) |
267 | return containsGCPtrType(Ty: AT->getElementType()); |
268 | if (StructType *ST = dyn_cast<StructType>(Val: Ty)) |
269 | return llvm::any_of(Range: ST->elements(), P: containsGCPtrType); |
270 | return false; |
271 | } |
272 | |
273 | // Debugging aid -- prints a [Begin, End) range of values. |
274 | template<typename IteratorTy> |
275 | static void PrintValueSet(raw_ostream &OS, IteratorTy Begin, IteratorTy End) { |
276 | OS << "[ " ; |
277 | while (Begin != End) { |
278 | OS << **Begin << " " ; |
279 | ++Begin; |
280 | } |
281 | OS << "]" ; |
282 | } |
283 | |
284 | /// The verifier algorithm is phrased in terms of availability. The set of |
285 | /// values "available" at a given point in the control flow graph is the set of |
286 | /// correctly relocated value at that point, and is a subset of the set of |
287 | /// definitions dominating that point. |
288 | |
289 | using AvailableValueSet = DenseSet<const Value *>; |
290 | |
291 | namespace { |
292 | /// State we compute and track per basic block. |
293 | struct BasicBlockState { |
294 | // Set of values available coming in, before the phi nodes |
295 | AvailableValueSet AvailableIn; |
296 | |
297 | // Set of values available going out |
298 | AvailableValueSet AvailableOut; |
299 | |
300 | // AvailableOut minus AvailableIn. |
301 | // All elements are Instructions |
302 | AvailableValueSet Contribution; |
303 | |
304 | // True if this block contains a safepoint and thus AvailableIn does not |
305 | // contribute to AvailableOut. |
306 | bool Cleared = false; |
307 | }; |
308 | } // namespace |
309 | |
310 | /// A given derived pointer can have multiple base pointers through phi/selects. |
311 | /// This type indicates when the base pointer is exclusively constant |
312 | /// (ExclusivelySomeConstant), and if that constant is proven to be exclusively |
313 | /// null, we record that as ExclusivelyNull. In all other cases, the BaseType is |
314 | /// NonConstant. |
315 | enum BaseType { |
316 | NonConstant = 1, // Base pointers is not exclusively constant. |
317 | ExclusivelyNull, |
318 | ExclusivelySomeConstant // Base pointers for a given derived pointer is from a |
319 | // set of constants, but they are not exclusively |
320 | // null. |
321 | }; |
322 | |
323 | /// Return the baseType for Val which states whether Val is exclusively |
324 | /// derived from constant/null, or not exclusively derived from constant. |
325 | /// Val is exclusively derived off a constant base when all operands of phi and |
326 | /// selects are derived off a constant base. |
327 | static enum BaseType getBaseType(const Value *Val) { |
328 | |
329 | SmallVector<const Value *, 32> Worklist; |
330 | DenseSet<const Value *> Visited; |
331 | bool isExclusivelyDerivedFromNull = true; |
332 | Worklist.push_back(Elt: Val); |
333 | // Strip through all the bitcasts and geps to get base pointer. Also check for |
334 | // the exclusive value when there can be multiple base pointers (through phis |
335 | // or selects). |
336 | while(!Worklist.empty()) { |
337 | const Value *V = Worklist.pop_back_val(); |
338 | if (!Visited.insert(V).second) |
339 | continue; |
340 | |
341 | if (const auto *CI = dyn_cast<CastInst>(Val: V)) { |
342 | Worklist.push_back(Elt: CI->stripPointerCasts()); |
343 | continue; |
344 | } |
345 | if (const auto *GEP = dyn_cast<GetElementPtrInst>(Val: V)) { |
346 | Worklist.push_back(Elt: GEP->getPointerOperand()); |
347 | continue; |
348 | } |
349 | // Push all the incoming values of phi node into the worklist for |
350 | // processing. |
351 | if (const auto *PN = dyn_cast<PHINode>(Val: V)) { |
352 | append_range(C&: Worklist, R: PN->incoming_values()); |
353 | continue; |
354 | } |
355 | if (const auto *SI = dyn_cast<SelectInst>(Val: V)) { |
356 | // Push in the true and false values |
357 | Worklist.push_back(Elt: SI->getTrueValue()); |
358 | Worklist.push_back(Elt: SI->getFalseValue()); |
359 | continue; |
360 | } |
361 | if (const auto *GCRelocate = dyn_cast<GCRelocateInst>(Val: V)) { |
362 | // GCRelocates do not change null-ness or constant-ness of the value. |
363 | // So we can continue with derived pointer this instruction relocates. |
364 | Worklist.push_back(Elt: GCRelocate->getDerivedPtr()); |
365 | continue; |
366 | } |
367 | if (const auto *FI = dyn_cast<FreezeInst>(Val: V)) { |
368 | // Freeze does not change null-ness or constant-ness of the value. |
369 | Worklist.push_back(Elt: FI->getOperand(i_nocapture: 0)); |
370 | continue; |
371 | } |
372 | if (isa<Constant>(Val: V)) { |
373 | // We found at least one base pointer which is non-null, so this derived |
374 | // pointer is not exclusively derived from null. |
375 | if (V != Constant::getNullValue(Ty: V->getType())) |
376 | isExclusivelyDerivedFromNull = false; |
377 | // Continue processing the remaining values to make sure it's exclusively |
378 | // constant. |
379 | continue; |
380 | } |
381 | // At this point, we know that the base pointer is not exclusively |
382 | // constant. |
383 | return BaseType::NonConstant; |
384 | } |
385 | // Now, we know that the base pointer is exclusively constant, but we need to |
386 | // differentiate between exclusive null constant and non-null constant. |
387 | return isExclusivelyDerivedFromNull ? BaseType::ExclusivelyNull |
388 | : BaseType::ExclusivelySomeConstant; |
389 | } |
390 | |
391 | static bool isNotExclusivelyConstantDerived(const Value *V) { |
392 | return getBaseType(Val: V) == BaseType::NonConstant; |
393 | } |
394 | |
395 | namespace { |
396 | class InstructionVerifier; |
397 | |
398 | /// Builds BasicBlockState for each BB of the function. |
399 | /// It can traverse function for verification and provides all required |
400 | /// information. |
401 | /// |
402 | /// GC pointer may be in one of three states: relocated, unrelocated and |
403 | /// poisoned. |
404 | /// Relocated pointer may be used without any restrictions. |
405 | /// Unrelocated pointer cannot be dereferenced, passed as argument to any call |
406 | /// or returned. Unrelocated pointer may be safely compared against another |
407 | /// unrelocated pointer or against a pointer exclusively derived from null. |
408 | /// Poisoned pointers are produced when we somehow derive pointer from relocated |
409 | /// and unrelocated pointers (e.g. phi, select). This pointers may be safely |
410 | /// used in a very limited number of situations. Currently the only way to use |
411 | /// it is comparison against constant exclusively derived from null. All |
412 | /// limitations arise due to their undefined state: this pointers should be |
413 | /// treated as relocated and unrelocated simultaneously. |
414 | /// Rules of deriving: |
415 | /// R + U = P - that's where the poisoned pointers come from |
416 | /// P + X = P |
417 | /// U + U = U |
418 | /// R + R = R |
419 | /// X + C = X |
420 | /// Where "+" - any operation that somehow derive pointer, U - unrelocated, |
421 | /// R - relocated and P - poisoned, C - constant, X - U or R or P or C or |
422 | /// nothing (in case when "+" is unary operation). |
423 | /// Deriving of pointers by itself is always safe. |
424 | /// NOTE: when we are making decision on the status of instruction's result: |
425 | /// a) for phi we need to check status of each input *at the end of |
426 | /// corresponding predecessor BB*. |
427 | /// b) for other instructions we need to check status of each input *at the |
428 | /// current point*. |
429 | /// |
430 | /// FIXME: This works fairly well except one case |
431 | /// bb1: |
432 | /// p = *some GC-ptr def* |
433 | /// p1 = gep p, offset |
434 | /// / | |
435 | /// / | |
436 | /// bb2: | |
437 | /// safepoint | |
438 | /// \ | |
439 | /// \ | |
440 | /// bb3: |
441 | /// p2 = phi [p, bb2] [p1, bb1] |
442 | /// p3 = phi [p, bb2] [p, bb1] |
443 | /// here p and p1 is unrelocated |
444 | /// p2 and p3 is poisoned (though they shouldn't be) |
445 | /// |
446 | /// This leads to some weird results: |
447 | /// cmp eq p, p2 - illegal instruction (false-positive) |
448 | /// cmp eq p1, p2 - illegal instruction (false-positive) |
449 | /// cmp eq p, p3 - illegal instruction (false-positive) |
450 | /// cmp eq p, p1 - ok |
451 | /// To fix this we need to introduce conception of generations and be able to |
452 | /// check if two values belong to one generation or not. This way p2 will be |
453 | /// considered to be unrelocated and no false alarm will happen. |
454 | class GCPtrTracker { |
455 | const Function &F; |
456 | const CFGDeadness &CD; |
457 | SpecificBumpPtrAllocator<BasicBlockState> BSAllocator; |
458 | DenseMap<const BasicBlock *, BasicBlockState *> BlockMap; |
459 | // This set contains defs of unrelocated pointers that are proved to be legal |
460 | // and don't need verification. |
461 | DenseSet<const Instruction *> ValidUnrelocatedDefs; |
462 | // This set contains poisoned defs. They can be safely ignored during |
463 | // verification too. |
464 | DenseSet<const Value *> PoisonedDefs; |
465 | |
466 | public: |
467 | GCPtrTracker(const Function &F, const DominatorTree &DT, |
468 | const CFGDeadness &CD); |
469 | |
470 | bool hasLiveIncomingEdge(const PHINode *PN, const BasicBlock *InBB) const { |
471 | return CD.hasLiveIncomingEdge(PN, InBB); |
472 | } |
473 | |
474 | BasicBlockState *getBasicBlockState(const BasicBlock *BB); |
475 | const BasicBlockState *getBasicBlockState(const BasicBlock *BB) const; |
476 | |
477 | bool isValuePoisoned(const Value *V) const { return PoisonedDefs.count(V); } |
478 | |
479 | /// Traverse each BB of the function and call |
480 | /// InstructionVerifier::verifyInstruction for each possibly invalid |
481 | /// instruction. |
482 | /// It destructively modifies GCPtrTracker so it's passed via rvalue reference |
483 | /// in order to prohibit further usages of GCPtrTracker as it'll be in |
484 | /// inconsistent state. |
485 | static void verifyFunction(GCPtrTracker &&Tracker, |
486 | InstructionVerifier &Verifier); |
487 | |
488 | /// Returns true for reachable and live blocks. |
489 | bool isMapped(const BasicBlock *BB) const { return BlockMap.contains(Val: BB); } |
490 | |
491 | private: |
492 | /// Returns true if the instruction may be safely skipped during verification. |
493 | bool instructionMayBeSkipped(const Instruction *I) const; |
494 | |
495 | /// Iterates over all BBs from BlockMap and recalculates AvailableIn/Out for |
496 | /// each of them until it converges. |
497 | void recalculateBBsStates(); |
498 | |
499 | /// Remove from Contribution all defs that legally produce unrelocated |
500 | /// pointers and saves them to ValidUnrelocatedDefs. |
501 | /// Though Contribution should belong to BBS it is passed separately with |
502 | /// different const-modifier in order to emphasize (and guarantee) that only |
503 | /// Contribution will be changed. |
504 | /// Returns true if Contribution was changed otherwise false. |
505 | bool removeValidUnrelocatedDefs(const BasicBlock *BB, |
506 | const BasicBlockState *BBS, |
507 | AvailableValueSet &Contribution); |
508 | |
509 | /// Gather all the definitions dominating the start of BB into Result. This is |
510 | /// simply the defs introduced by every dominating basic block and the |
511 | /// function arguments. |
512 | void gatherDominatingDefs(const BasicBlock *BB, AvailableValueSet &Result, |
513 | const DominatorTree &DT); |
514 | |
515 | /// Compute the AvailableOut set for BB, based on the BasicBlockState BBS, |
516 | /// which is the BasicBlockState for BB. |
517 | /// ContributionChanged is set when the verifier runs for the first time |
518 | /// (in this case Contribution was changed from 'empty' to its initial state) |
519 | /// or when Contribution of this BB was changed since last computation. |
520 | static void transferBlock(const BasicBlock *BB, BasicBlockState &BBS, |
521 | bool ContributionChanged); |
522 | |
523 | /// Model the effect of an instruction on the set of available values. |
524 | static void transferInstruction(const Instruction &I, bool &Cleared, |
525 | AvailableValueSet &Available); |
526 | }; |
527 | |
528 | /// It is a visitor for GCPtrTracker::verifyFunction. It decides if the |
529 | /// instruction (which uses heap reference) is legal or not, given our safepoint |
530 | /// semantics. |
531 | class InstructionVerifier { |
532 | bool AnyInvalidUses = false; |
533 | |
534 | public: |
535 | void verifyInstruction(const GCPtrTracker *Tracker, const Instruction &I, |
536 | const AvailableValueSet &AvailableSet); |
537 | |
538 | bool hasAnyInvalidUses() const { return AnyInvalidUses; } |
539 | |
540 | private: |
541 | void reportInvalidUse(const Value &V, const Instruction &I); |
542 | }; |
543 | } // end anonymous namespace |
544 | |
545 | GCPtrTracker::GCPtrTracker(const Function &F, const DominatorTree &DT, |
546 | const CFGDeadness &CD) : F(F), CD(CD) { |
547 | // Calculate Contribution of each live BB. |
548 | // Allocate BB states for live blocks. |
549 | for (const BasicBlock &BB : F) |
550 | if (!CD.isDeadBlock(BB: &BB)) { |
551 | BasicBlockState *BBS = new (BSAllocator.Allocate()) BasicBlockState; |
552 | for (const auto &I : BB) |
553 | transferInstruction(I, Cleared&: BBS->Cleared, Available&: BBS->Contribution); |
554 | BlockMap[&BB] = BBS; |
555 | } |
556 | |
557 | // Initialize AvailableIn/Out sets of each BB using only information about |
558 | // dominating BBs. |
559 | for (auto &BBI : BlockMap) { |
560 | gatherDominatingDefs(BB: BBI.first, Result&: BBI.second->AvailableIn, DT); |
561 | transferBlock(BB: BBI.first, BBS&: *BBI.second, ContributionChanged: true); |
562 | } |
563 | |
564 | // Simulate the flow of defs through the CFG and recalculate AvailableIn/Out |
565 | // sets of each BB until it converges. If any def is proved to be an |
566 | // unrelocated pointer, it will be removed from all BBSs. |
567 | recalculateBBsStates(); |
568 | } |
569 | |
570 | BasicBlockState *GCPtrTracker::getBasicBlockState(const BasicBlock *BB) { |
571 | return BlockMap.lookup(Val: BB); |
572 | } |
573 | |
574 | const BasicBlockState *GCPtrTracker::getBasicBlockState( |
575 | const BasicBlock *BB) const { |
576 | return const_cast<GCPtrTracker *>(this)->getBasicBlockState(BB); |
577 | } |
578 | |
579 | bool GCPtrTracker::instructionMayBeSkipped(const Instruction *I) const { |
580 | // Poisoned defs are skipped since they are always safe by itself by |
581 | // definition (for details see comment to this class). |
582 | return ValidUnrelocatedDefs.count(V: I) || PoisonedDefs.count(V: I); |
583 | } |
584 | |
585 | void GCPtrTracker::verifyFunction(GCPtrTracker &&Tracker, |
586 | InstructionVerifier &Verifier) { |
587 | // We need RPO here to a) report always the first error b) report errors in |
588 | // same order from run to run. |
589 | ReversePostOrderTraversal<const Function *> RPOT(&Tracker.F); |
590 | for (const BasicBlock *BB : RPOT) { |
591 | BasicBlockState *BBS = Tracker.getBasicBlockState(BB); |
592 | if (!BBS) |
593 | continue; |
594 | |
595 | // We destructively modify AvailableIn as we traverse the block instruction |
596 | // by instruction. |
597 | AvailableValueSet &AvailableSet = BBS->AvailableIn; |
598 | for (const Instruction &I : *BB) { |
599 | if (Tracker.instructionMayBeSkipped(I: &I)) |
600 | continue; // This instruction shouldn't be added to AvailableSet. |
601 | |
602 | Verifier.verifyInstruction(Tracker: &Tracker, I, AvailableSet); |
603 | |
604 | // Model the effect of current instruction on AvailableSet to keep the set |
605 | // relevant at each point of BB. |
606 | bool Cleared = false; |
607 | transferInstruction(I, Cleared, Available&: AvailableSet); |
608 | (void)Cleared; |
609 | } |
610 | } |
611 | } |
612 | |
613 | void GCPtrTracker::recalculateBBsStates() { |
614 | // TODO: This order is suboptimal, it's better to replace it with priority |
615 | // queue where priority is RPO number of BB. |
616 | SetVector<const BasicBlock *> Worklist(llvm::from_range, |
617 | llvm::make_first_range(c&: BlockMap)); |
618 | |
619 | // This loop iterates the AvailableIn/Out sets until it converges. |
620 | // The AvailableIn and AvailableOut sets decrease as we iterate. |
621 | while (!Worklist.empty()) { |
622 | const BasicBlock *BB = Worklist.pop_back_val(); |
623 | BasicBlockState *BBS = getBasicBlockState(BB); |
624 | if (!BBS) |
625 | continue; // Ignore dead successors. |
626 | |
627 | size_t OldInCount = BBS->AvailableIn.size(); |
628 | for (const_pred_iterator PredIt(BB), End(BB, true); PredIt != End; ++PredIt) { |
629 | const BasicBlock *PBB = *PredIt; |
630 | BasicBlockState *PBBS = getBasicBlockState(BB: PBB); |
631 | if (PBBS && !CD.isDeadEdge(U: &CFGDeadness::getEdge(PredIt))) |
632 | set_intersect(S1&: BBS->AvailableIn, S2: PBBS->AvailableOut); |
633 | } |
634 | |
635 | assert(OldInCount >= BBS->AvailableIn.size() && "invariant!" ); |
636 | |
637 | bool InputsChanged = OldInCount != BBS->AvailableIn.size(); |
638 | bool ContributionChanged = |
639 | removeValidUnrelocatedDefs(BB, BBS, Contribution&: BBS->Contribution); |
640 | if (!InputsChanged && !ContributionChanged) |
641 | continue; |
642 | |
643 | size_t OldOutCount = BBS->AvailableOut.size(); |
644 | transferBlock(BB, BBS&: *BBS, ContributionChanged); |
645 | if (OldOutCount != BBS->AvailableOut.size()) { |
646 | assert(OldOutCount > BBS->AvailableOut.size() && "invariant!" ); |
647 | Worklist.insert_range(R: successors(BB)); |
648 | } |
649 | } |
650 | } |
651 | |
652 | bool GCPtrTracker::removeValidUnrelocatedDefs(const BasicBlock *BB, |
653 | const BasicBlockState *BBS, |
654 | AvailableValueSet &Contribution) { |
655 | assert(&BBS->Contribution == &Contribution && |
656 | "Passed Contribution should be from the passed BasicBlockState!" ); |
657 | AvailableValueSet AvailableSet = BBS->AvailableIn; |
658 | bool ContributionChanged = false; |
659 | // For explanation why instructions are processed this way see |
660 | // "Rules of deriving" in the comment to this class. |
661 | for (const Instruction &I : *BB) { |
662 | bool ValidUnrelocatedPointerDef = false; |
663 | bool PoisonedPointerDef = false; |
664 | // TODO: `select` instructions should be handled here too. |
665 | if (const PHINode *PN = dyn_cast<PHINode>(Val: &I)) { |
666 | if (containsGCPtrType(Ty: PN->getType())) { |
667 | // If both is true, output is poisoned. |
668 | bool HasRelocatedInputs = false; |
669 | bool HasUnrelocatedInputs = false; |
670 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { |
671 | const BasicBlock *InBB = PN->getIncomingBlock(i); |
672 | if (!isMapped(BB: InBB) || |
673 | !CD.hasLiveIncomingEdge(PN, InBB)) |
674 | continue; // Skip dead block or dead edge. |
675 | |
676 | const Value *InValue = PN->getIncomingValue(i); |
677 | |
678 | if (isNotExclusivelyConstantDerived(V: InValue)) { |
679 | if (isValuePoisoned(V: InValue)) { |
680 | // If any of inputs is poisoned, output is always poisoned too. |
681 | HasRelocatedInputs = true; |
682 | HasUnrelocatedInputs = true; |
683 | break; |
684 | } |
685 | if (BlockMap[InBB]->AvailableOut.count(V: InValue)) |
686 | HasRelocatedInputs = true; |
687 | else |
688 | HasUnrelocatedInputs = true; |
689 | } |
690 | } |
691 | if (HasUnrelocatedInputs) { |
692 | if (HasRelocatedInputs) |
693 | PoisonedPointerDef = true; |
694 | else |
695 | ValidUnrelocatedPointerDef = true; |
696 | } |
697 | } |
698 | } else if ((isa<GetElementPtrInst>(Val: I) || isa<BitCastInst>(Val: I)) && |
699 | containsGCPtrType(Ty: I.getType())) { |
700 | // GEP/bitcast of unrelocated pointer is legal by itself but this def |
701 | // shouldn't appear in any AvailableSet. |
702 | for (const Value *V : I.operands()) |
703 | if (containsGCPtrType(Ty: V->getType()) && |
704 | isNotExclusivelyConstantDerived(V) && !AvailableSet.count(V)) { |
705 | if (isValuePoisoned(V)) |
706 | PoisonedPointerDef = true; |
707 | else |
708 | ValidUnrelocatedPointerDef = true; |
709 | break; |
710 | } |
711 | } |
712 | assert(!(ValidUnrelocatedPointerDef && PoisonedPointerDef) && |
713 | "Value cannot be both unrelocated and poisoned!" ); |
714 | if (ValidUnrelocatedPointerDef) { |
715 | // Remove def of unrelocated pointer from Contribution of this BB and |
716 | // trigger update of all its successors. |
717 | Contribution.erase(V: &I); |
718 | PoisonedDefs.erase(V: &I); |
719 | ValidUnrelocatedDefs.insert(V: &I); |
720 | LLVM_DEBUG(dbgs() << "Removing urelocated " << I |
721 | << " from Contribution of " << BB->getName() << "\n" ); |
722 | ContributionChanged = true; |
723 | } else if (PoisonedPointerDef) { |
724 | // Mark pointer as poisoned, remove its def from Contribution and trigger |
725 | // update of all successors. |
726 | Contribution.erase(V: &I); |
727 | PoisonedDefs.insert(V: &I); |
728 | LLVM_DEBUG(dbgs() << "Removing poisoned " << I << " from Contribution of " |
729 | << BB->getName() << "\n" ); |
730 | ContributionChanged = true; |
731 | } else { |
732 | bool Cleared = false; |
733 | transferInstruction(I, Cleared, Available&: AvailableSet); |
734 | (void)Cleared; |
735 | } |
736 | } |
737 | return ContributionChanged; |
738 | } |
739 | |
740 | void GCPtrTracker::gatherDominatingDefs(const BasicBlock *BB, |
741 | AvailableValueSet &Result, |
742 | const DominatorTree &DT) { |
743 | DomTreeNode *DTN = DT[const_cast<BasicBlock *>(BB)]; |
744 | |
745 | assert(DTN && "Unreachable blocks are ignored" ); |
746 | while (DTN->getIDom()) { |
747 | DTN = DTN->getIDom(); |
748 | auto BBS = getBasicBlockState(BB: DTN->getBlock()); |
749 | assert(BBS && "immediate dominator cannot be dead for a live block" ); |
750 | const auto &Defs = BBS->Contribution; |
751 | Result.insert_range(R: Defs); |
752 | // If this block is 'Cleared', then nothing LiveIn to this block can be |
753 | // available after this block completes. Note: This turns out to be |
754 | // really important for reducing memory consuption of the initial available |
755 | // sets and thus peak memory usage by this verifier. |
756 | if (BBS->Cleared) |
757 | return; |
758 | } |
759 | |
760 | for (const Argument &A : BB->getParent()->args()) |
761 | if (containsGCPtrType(Ty: A.getType())) |
762 | Result.insert(V: &A); |
763 | } |
764 | |
765 | void GCPtrTracker::transferBlock(const BasicBlock *BB, BasicBlockState &BBS, |
766 | bool ContributionChanged) { |
767 | const AvailableValueSet &AvailableIn = BBS.AvailableIn; |
768 | AvailableValueSet &AvailableOut = BBS.AvailableOut; |
769 | |
770 | if (BBS.Cleared) { |
771 | // AvailableOut will change only when Contribution changed. |
772 | if (ContributionChanged) |
773 | AvailableOut = BBS.Contribution; |
774 | } else { |
775 | // Otherwise, we need to reduce the AvailableOut set by things which are no |
776 | // longer in our AvailableIn |
777 | AvailableValueSet Temp = BBS.Contribution; |
778 | set_union(S1&: Temp, S2: AvailableIn); |
779 | AvailableOut = std::move(Temp); |
780 | } |
781 | |
782 | LLVM_DEBUG(dbgs() << "Transfered block " << BB->getName() << " from " ; |
783 | PrintValueSet(dbgs(), AvailableIn.begin(), AvailableIn.end()); |
784 | dbgs() << " to " ; |
785 | PrintValueSet(dbgs(), AvailableOut.begin(), AvailableOut.end()); |
786 | dbgs() << "\n" ;); |
787 | } |
788 | |
789 | void GCPtrTracker::transferInstruction(const Instruction &I, bool &Cleared, |
790 | AvailableValueSet &Available) { |
791 | if (isa<GCStatepointInst>(Val: I)) { |
792 | Cleared = true; |
793 | Available.clear(); |
794 | } else if (containsGCPtrType(Ty: I.getType())) |
795 | Available.insert(V: &I); |
796 | } |
797 | |
798 | void InstructionVerifier::verifyInstruction( |
799 | const GCPtrTracker *Tracker, const Instruction &I, |
800 | const AvailableValueSet &AvailableSet) { |
801 | if (const PHINode *PN = dyn_cast<PHINode>(Val: &I)) { |
802 | if (containsGCPtrType(Ty: PN->getType())) |
803 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { |
804 | const BasicBlock *InBB = PN->getIncomingBlock(i); |
805 | const BasicBlockState *InBBS = Tracker->getBasicBlockState(BB: InBB); |
806 | if (!InBBS || |
807 | !Tracker->hasLiveIncomingEdge(PN, InBB)) |
808 | continue; // Skip dead block or dead edge. |
809 | |
810 | const Value *InValue = PN->getIncomingValue(i); |
811 | |
812 | if (isNotExclusivelyConstantDerived(V: InValue) && |
813 | !InBBS->AvailableOut.count(V: InValue)) |
814 | reportInvalidUse(V: *InValue, I: *PN); |
815 | } |
816 | } else if (isa<CmpInst>(Val: I) && |
817 | containsGCPtrType(Ty: I.getOperand(i: 0)->getType())) { |
818 | Value *LHS = I.getOperand(i: 0), *RHS = I.getOperand(i: 1); |
819 | enum BaseType baseTyLHS = getBaseType(Val: LHS), |
820 | baseTyRHS = getBaseType(Val: RHS); |
821 | |
822 | // Returns true if LHS and RHS are unrelocated pointers and they are |
823 | // valid unrelocated uses. |
824 | auto hasValidUnrelocatedUse = [&AvailableSet, Tracker, baseTyLHS, baseTyRHS, |
825 | &LHS, &RHS] () { |
826 | // A cmp instruction has valid unrelocated pointer operands only if |
827 | // both operands are unrelocated pointers. |
828 | // In the comparison between two pointers, if one is an unrelocated |
829 | // use, the other *should be* an unrelocated use, for this |
830 | // instruction to contain valid unrelocated uses. This unrelocated |
831 | // use can be a null constant as well, or another unrelocated |
832 | // pointer. |
833 | if (AvailableSet.count(V: LHS) || AvailableSet.count(V: RHS)) |
834 | return false; |
835 | // Constant pointers (that are not exclusively null) may have |
836 | // meaning in different VMs, so we cannot reorder the compare |
837 | // against constant pointers before the safepoint. In other words, |
838 | // comparison of an unrelocated use against a non-null constant |
839 | // maybe invalid. |
840 | if ((baseTyLHS == BaseType::ExclusivelySomeConstant && |
841 | baseTyRHS == BaseType::NonConstant) || |
842 | (baseTyLHS == BaseType::NonConstant && |
843 | baseTyRHS == BaseType::ExclusivelySomeConstant)) |
844 | return false; |
845 | |
846 | // If one of pointers is poisoned and other is not exclusively derived |
847 | // from null it is an invalid expression: it produces poisoned result |
848 | // and unless we want to track all defs (not only gc pointers) the only |
849 | // option is to prohibit such instructions. |
850 | if ((Tracker->isValuePoisoned(V: LHS) && baseTyRHS != ExclusivelyNull) || |
851 | (Tracker->isValuePoisoned(V: RHS) && baseTyLHS != ExclusivelyNull)) |
852 | return false; |
853 | |
854 | // All other cases are valid cases enumerated below: |
855 | // 1. Comparison between an exclusively derived null pointer and a |
856 | // constant base pointer. |
857 | // 2. Comparison between an exclusively derived null pointer and a |
858 | // non-constant unrelocated base pointer. |
859 | // 3. Comparison between 2 unrelocated pointers. |
860 | // 4. Comparison between a pointer exclusively derived from null and a |
861 | // non-constant poisoned pointer. |
862 | return true; |
863 | }; |
864 | if (!hasValidUnrelocatedUse()) { |
865 | // Print out all non-constant derived pointers that are unrelocated |
866 | // uses, which are invalid. |
867 | if (baseTyLHS == BaseType::NonConstant && !AvailableSet.count(V: LHS)) |
868 | reportInvalidUse(V: *LHS, I); |
869 | if (baseTyRHS == BaseType::NonConstant && !AvailableSet.count(V: RHS)) |
870 | reportInvalidUse(V: *RHS, I); |
871 | } |
872 | } else { |
873 | for (const Value *V : I.operands()) |
874 | if (containsGCPtrType(Ty: V->getType()) && |
875 | isNotExclusivelyConstantDerived(V) && !AvailableSet.count(V)) |
876 | reportInvalidUse(V: *V, I); |
877 | } |
878 | } |
879 | |
880 | void InstructionVerifier::reportInvalidUse(const Value &V, |
881 | const Instruction &I) { |
882 | errs() << "Illegal use of unrelocated value found!\n" ; |
883 | errs() << "Def: " << V << "\n" ; |
884 | errs() << "Use: " << I << "\n" ; |
885 | if (!PrintOnly) |
886 | abort(); |
887 | AnyInvalidUses = true; |
888 | } |
889 | |
890 | static void Verify(const Function &F, const DominatorTree &DT, |
891 | const CFGDeadness &CD) { |
892 | LLVM_DEBUG(dbgs() << "Verifying gc pointers in function: " << F.getName() |
893 | << "\n" ); |
894 | if (PrintOnly) |
895 | dbgs() << "Verifying gc pointers in function: " << F.getName() << "\n" ; |
896 | |
897 | GCPtrTracker Tracker(F, DT, CD); |
898 | |
899 | // We now have all the information we need to decide if the use of a heap |
900 | // reference is legal or not, given our safepoint semantics. |
901 | |
902 | InstructionVerifier Verifier; |
903 | GCPtrTracker::verifyFunction(Tracker: std::move(Tracker), Verifier); |
904 | |
905 | if (PrintOnly && !Verifier.hasAnyInvalidUses()) { |
906 | dbgs() << "No illegal uses found by SafepointIRVerifier in: " << F.getName() |
907 | << "\n" ; |
908 | } |
909 | } |
910 | |