1 | //===- MustExecute.cpp - Printer for isGuaranteedToExecute ----------------===// |
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 | #include "llvm/Analysis/MustExecute.h" |
10 | #include "llvm/ADT/PostOrderIterator.h" |
11 | #include "llvm/ADT/StringExtras.h" |
12 | #include "llvm/Analysis/CFG.h" |
13 | #include "llvm/Analysis/InstructionSimplify.h" |
14 | #include "llvm/Analysis/LoopInfo.h" |
15 | #include "llvm/Analysis/PostDominators.h" |
16 | #include "llvm/Analysis/ValueTracking.h" |
17 | #include "llvm/IR/AssemblyAnnotationWriter.h" |
18 | #include "llvm/IR/Dominators.h" |
19 | #include "llvm/IR/InstIterator.h" |
20 | #include "llvm/IR/Module.h" |
21 | #include "llvm/IR/PassManager.h" |
22 | #include "llvm/Support/FormattedStream.h" |
23 | #include "llvm/Support/raw_ostream.h" |
24 | |
25 | using namespace llvm; |
26 | |
27 | #define DEBUG_TYPE "must-execute" |
28 | |
29 | const DenseMap<BasicBlock *, ColorVector> & |
30 | LoopSafetyInfo::getBlockColors() const { |
31 | return BlockColors; |
32 | } |
33 | |
34 | void LoopSafetyInfo::copyColors(BasicBlock *New, BasicBlock *Old) { |
35 | ColorVector &ColorsForNewBlock = BlockColors[New]; |
36 | ColorVector &ColorsForOldBlock = BlockColors[Old]; |
37 | ColorsForNewBlock = ColorsForOldBlock; |
38 | } |
39 | |
40 | bool SimpleLoopSafetyInfo::blockMayThrow(const BasicBlock *BB) const { |
41 | (void)BB; |
42 | return anyBlockMayThrow(); |
43 | } |
44 | |
45 | bool SimpleLoopSafetyInfo::anyBlockMayThrow() const { |
46 | return MayThrow; |
47 | } |
48 | |
49 | void SimpleLoopSafetyInfo::computeLoopSafetyInfo(const Loop *CurLoop) { |
50 | assert(CurLoop != nullptr && "CurLoop can't be null" ); |
51 | BasicBlock * = CurLoop->getHeader(); |
52 | // Iterate over header and compute safety info. |
53 | HeaderMayThrow = !isGuaranteedToTransferExecutionToSuccessor(BB: Header); |
54 | MayThrow = HeaderMayThrow; |
55 | // Iterate over loop instructions and compute safety info. |
56 | // Skip header as it has been computed and stored in HeaderMayThrow. |
57 | // The first block in loopinfo.Blocks is guaranteed to be the header. |
58 | assert(Header == *CurLoop->getBlocks().begin() && |
59 | "First block must be header" ); |
60 | for (const BasicBlock *BB : llvm::drop_begin(RangeOrContainer: CurLoop->blocks())) { |
61 | MayThrow |= !isGuaranteedToTransferExecutionToSuccessor(BB); |
62 | if (MayThrow) |
63 | break; |
64 | } |
65 | |
66 | computeBlockColors(CurLoop); |
67 | } |
68 | |
69 | bool ICFLoopSafetyInfo::blockMayThrow(const BasicBlock *BB) const { |
70 | return ICF.hasICF(BB); |
71 | } |
72 | |
73 | bool ICFLoopSafetyInfo::anyBlockMayThrow() const { |
74 | return MayThrow; |
75 | } |
76 | |
77 | void ICFLoopSafetyInfo::computeLoopSafetyInfo(const Loop *CurLoop) { |
78 | assert(CurLoop != nullptr && "CurLoop can't be null" ); |
79 | ICF.clear(); |
80 | MW.clear(); |
81 | MayThrow = false; |
82 | // Figure out the fact that at least one block may throw. |
83 | for (const auto &BB : CurLoop->blocks()) |
84 | if (ICF.hasICF(BB: &*BB)) { |
85 | MayThrow = true; |
86 | break; |
87 | } |
88 | computeBlockColors(CurLoop); |
89 | } |
90 | |
91 | void ICFLoopSafetyInfo::insertInstructionTo(const Instruction *Inst, |
92 | const BasicBlock *BB) { |
93 | ICF.insertInstructionTo(Inst, BB); |
94 | MW.insertInstructionTo(Inst, BB); |
95 | } |
96 | |
97 | void ICFLoopSafetyInfo::removeInstruction(const Instruction *Inst) { |
98 | ICF.removeInstruction(Inst); |
99 | MW.removeInstruction(Inst); |
100 | } |
101 | |
102 | void LoopSafetyInfo::computeBlockColors(const Loop *CurLoop) { |
103 | // Compute funclet colors if we might sink/hoist in a function with a funclet |
104 | // personality routine. |
105 | Function *Fn = CurLoop->getHeader()->getParent(); |
106 | if (Fn->hasPersonalityFn()) |
107 | if (Constant *PersonalityFn = Fn->getPersonalityFn()) |
108 | if (isScopedEHPersonality(Pers: classifyEHPersonality(Pers: PersonalityFn))) |
109 | BlockColors = colorEHFunclets(F&: *Fn); |
110 | } |
111 | |
112 | /// Return true if we can prove that the given ExitBlock is not reached on the |
113 | /// first iteration of the given loop. That is, the backedge of the loop must |
114 | /// be executed before the ExitBlock is executed in any dynamic execution trace. |
115 | static bool CanProveNotTakenFirstIteration(const BasicBlock *ExitBlock, |
116 | const DominatorTree *DT, |
117 | const Loop *CurLoop) { |
118 | auto *CondExitBlock = ExitBlock->getSinglePredecessor(); |
119 | if (!CondExitBlock) |
120 | // expect unique exits |
121 | return false; |
122 | assert(CurLoop->contains(CondExitBlock) && "meaning of exit block" ); |
123 | auto *BI = dyn_cast<BranchInst>(Val: CondExitBlock->getTerminator()); |
124 | if (!BI || !BI->isConditional()) |
125 | return false; |
126 | // If condition is constant and false leads to ExitBlock then we always |
127 | // execute the true branch. |
128 | if (auto *Cond = dyn_cast<ConstantInt>(Val: BI->getCondition())) |
129 | return BI->getSuccessor(i: Cond->getZExtValue() ? 1 : 0) == ExitBlock; |
130 | auto *Cond = dyn_cast<CmpInst>(Val: BI->getCondition()); |
131 | if (!Cond) |
132 | return false; |
133 | // todo: this would be a lot more powerful if we used scev, but all the |
134 | // plumbing is currently missing to pass a pointer in from the pass |
135 | // Check for cmp (phi [x, preheader] ...), y where (pred x, y is known |
136 | ICmpInst::Predicate Pred = Cond->getPredicate(); |
137 | auto *LHS = dyn_cast<PHINode>(Val: Cond->getOperand(i_nocapture: 0)); |
138 | auto *RHS = Cond->getOperand(i_nocapture: 1); |
139 | if (!LHS || LHS->getParent() != CurLoop->getHeader()) { |
140 | Pred = Cond->getSwappedPredicate(); |
141 | LHS = dyn_cast<PHINode>(Val: Cond->getOperand(i_nocapture: 1)); |
142 | RHS = Cond->getOperand(i_nocapture: 0); |
143 | if (!LHS || LHS->getParent() != CurLoop->getHeader()) |
144 | return false; |
145 | } |
146 | |
147 | auto DL = ExitBlock->getModule()->getDataLayout(); |
148 | auto *IVStart = LHS->getIncomingValueForBlock(BB: CurLoop->getLoopPreheader()); |
149 | auto *SimpleValOrNull = simplifyCmpInst( |
150 | Predicate: Pred, LHS: IVStart, RHS, Q: {DL, /*TLI*/ nullptr, DT, /*AC*/ nullptr, BI}); |
151 | auto *SimpleCst = dyn_cast_or_null<Constant>(Val: SimpleValOrNull); |
152 | if (!SimpleCst) |
153 | return false; |
154 | if (ExitBlock == BI->getSuccessor(i: 0)) |
155 | return SimpleCst->isZeroValue(); |
156 | assert(ExitBlock == BI->getSuccessor(1) && "implied by above" ); |
157 | return SimpleCst->isAllOnesValue(); |
158 | } |
159 | |
160 | /// Collect all blocks from \p CurLoop which lie on all possible paths from |
161 | /// the header of \p CurLoop (inclusive) to BB (exclusive) into the set |
162 | /// \p Predecessors. If \p BB is the header, \p Predecessors will be empty. |
163 | /// Note: It's possible that we encounter Irreducible control flow, due to |
164 | /// which, we may find that a few predecessors of \p BB are not a part of the |
165 | /// \p CurLoop. We only return Predecessors that are a part of \p CurLoop. |
166 | static void collectTransitivePredecessors( |
167 | const Loop *CurLoop, const BasicBlock *BB, |
168 | SmallPtrSetImpl<const BasicBlock *> &Predecessors) { |
169 | assert(Predecessors.empty() && "Garbage in predecessors set?" ); |
170 | assert(CurLoop->contains(BB) && "Should only be called for loop blocks!" ); |
171 | if (BB == CurLoop->getHeader()) |
172 | return; |
173 | SmallVector<const BasicBlock *, 4> WorkList; |
174 | for (const auto *Pred : predecessors(BB)) { |
175 | if (!CurLoop->contains(BB: Pred)) |
176 | continue; |
177 | Predecessors.insert(Ptr: Pred); |
178 | WorkList.push_back(Elt: Pred); |
179 | } |
180 | while (!WorkList.empty()) { |
181 | auto *Pred = WorkList.pop_back_val(); |
182 | assert(CurLoop->contains(Pred) && "Should only reach loop blocks!" ); |
183 | // We are not interested in backedges and we don't want to leave loop. |
184 | if (Pred == CurLoop->getHeader()) |
185 | continue; |
186 | // TODO: If BB lies in an inner loop of CurLoop, this will traverse over all |
187 | // blocks of this inner loop, even those that are always executed AFTER the |
188 | // BB. It may make our analysis more conservative than it could be, see test |
189 | // @nested and @nested_no_throw in test/Analysis/MustExecute/loop-header.ll. |
190 | // We can ignore backedge of all loops containing BB to get a sligtly more |
191 | // optimistic result. |
192 | for (const auto *PredPred : predecessors(BB: Pred)) |
193 | if (CurLoop->contains(BB: PredPred) && Predecessors.insert(Ptr: PredPred).second) |
194 | WorkList.push_back(Elt: PredPred); |
195 | } |
196 | } |
197 | |
198 | bool LoopSafetyInfo::allLoopPathsLeadToBlock(const Loop *CurLoop, |
199 | const BasicBlock *BB, |
200 | const DominatorTree *DT) const { |
201 | assert(CurLoop->contains(BB) && "Should only be called for loop blocks!" ); |
202 | |
203 | // Fast path: header is always reached once the loop is entered. |
204 | if (BB == CurLoop->getHeader()) |
205 | return true; |
206 | |
207 | // Collect all transitive predecessors of BB in the same loop. This set will |
208 | // be a subset of the blocks within the loop. |
209 | SmallPtrSet<const BasicBlock *, 4> Predecessors; |
210 | collectTransitivePredecessors(CurLoop, BB, Predecessors); |
211 | |
212 | // Bail out if a latch block is part of the predecessor set. In this case |
213 | // we may take the backedge to the header and not execute other latch |
214 | // successors. |
215 | for (const BasicBlock *Pred : predecessors(BB: CurLoop->getHeader())) |
216 | // Predecessors only contains loop blocks, so we don't have to worry about |
217 | // preheader predecessors here. |
218 | if (Predecessors.contains(Ptr: Pred)) |
219 | return false; |
220 | |
221 | // Make sure that all successors of, all predecessors of BB which are not |
222 | // dominated by BB, are either: |
223 | // 1) BB, |
224 | // 2) Also predecessors of BB, |
225 | // 3) Exit blocks which are not taken on 1st iteration. |
226 | // Memoize blocks we've already checked. |
227 | SmallPtrSet<const BasicBlock *, 4> CheckedSuccessors; |
228 | for (const auto *Pred : Predecessors) { |
229 | // Predecessor block may throw, so it has a side exit. |
230 | if (blockMayThrow(BB: Pred)) |
231 | return false; |
232 | |
233 | // BB dominates Pred, so if Pred runs, BB must run. |
234 | // This is true when Pred is a loop latch. |
235 | if (DT->dominates(A: BB, B: Pred)) |
236 | continue; |
237 | |
238 | for (const auto *Succ : successors(BB: Pred)) |
239 | if (CheckedSuccessors.insert(Ptr: Succ).second && |
240 | Succ != BB && !Predecessors.count(Ptr: Succ)) |
241 | // By discharging conditions that are not executed on the 1st iteration, |
242 | // we guarantee that *at least* on the first iteration all paths from |
243 | // header that *may* execute will lead us to the block of interest. So |
244 | // that if we had virtually peeled one iteration away, in this peeled |
245 | // iteration the set of predecessors would contain only paths from |
246 | // header to BB without any exiting edges that may execute. |
247 | // |
248 | // TODO: We only do it for exiting edges currently. We could use the |
249 | // same function to skip some of the edges within the loop if we know |
250 | // that they will not be taken on the 1st iteration. |
251 | // |
252 | // TODO: If we somehow know the number of iterations in loop, the same |
253 | // check may be done for any arbitrary N-th iteration as long as N is |
254 | // not greater than minimum number of iterations in this loop. |
255 | if (CurLoop->contains(BB: Succ) || |
256 | !CanProveNotTakenFirstIteration(ExitBlock: Succ, DT, CurLoop)) |
257 | return false; |
258 | } |
259 | |
260 | // All predecessors can only lead us to BB. |
261 | return true; |
262 | } |
263 | |
264 | /// Returns true if the instruction in a loop is guaranteed to execute at least |
265 | /// once. |
266 | bool SimpleLoopSafetyInfo::isGuaranteedToExecute(const Instruction &Inst, |
267 | const DominatorTree *DT, |
268 | const Loop *CurLoop) const { |
269 | // If the instruction is in the header block for the loop (which is very |
270 | // common), it is always guaranteed to dominate the exit blocks. Since this |
271 | // is a common case, and can save some work, check it now. |
272 | if (Inst.getParent() == CurLoop->getHeader()) |
273 | // If there's a throw in the header block, we can't guarantee we'll reach |
274 | // Inst unless we can prove that Inst comes before the potential implicit |
275 | // exit. At the moment, we use a (cheap) hack for the common case where |
276 | // the instruction of interest is the first one in the block. |
277 | return !HeaderMayThrow || |
278 | &*Inst.getParent()->getFirstNonPHIOrDbg() == &Inst; |
279 | |
280 | // If there is a path from header to exit or latch that doesn't lead to our |
281 | // instruction's block, return false. |
282 | return allLoopPathsLeadToBlock(CurLoop, BB: Inst.getParent(), DT); |
283 | } |
284 | |
285 | bool ICFLoopSafetyInfo::isGuaranteedToExecute(const Instruction &Inst, |
286 | const DominatorTree *DT, |
287 | const Loop *CurLoop) const { |
288 | return !ICF.isDominatedByICFIFromSameBlock(Insn: &Inst) && |
289 | allLoopPathsLeadToBlock(CurLoop, BB: Inst.getParent(), DT); |
290 | } |
291 | |
292 | bool ICFLoopSafetyInfo::doesNotWriteMemoryBefore(const BasicBlock *BB, |
293 | const Loop *CurLoop) const { |
294 | assert(CurLoop->contains(BB) && "Should only be called for loop blocks!" ); |
295 | |
296 | // Fast path: there are no instructions before header. |
297 | if (BB == CurLoop->getHeader()) |
298 | return true; |
299 | |
300 | // Collect all transitive predecessors of BB in the same loop. This set will |
301 | // be a subset of the blocks within the loop. |
302 | SmallPtrSet<const BasicBlock *, 4> Predecessors; |
303 | collectTransitivePredecessors(CurLoop, BB, Predecessors); |
304 | // Find if there any instruction in either predecessor that could write |
305 | // to memory. |
306 | for (const auto *Pred : Predecessors) |
307 | if (MW.mayWriteToMemory(BB: Pred)) |
308 | return false; |
309 | return true; |
310 | } |
311 | |
312 | bool ICFLoopSafetyInfo::doesNotWriteMemoryBefore(const Instruction &I, |
313 | const Loop *CurLoop) const { |
314 | auto *BB = I.getParent(); |
315 | assert(CurLoop->contains(BB) && "Should only be called for loop blocks!" ); |
316 | return !MW.isDominatedByMemoryWriteFromSameBlock(Insn: &I) && |
317 | doesNotWriteMemoryBefore(BB, CurLoop); |
318 | } |
319 | |
320 | static bool isMustExecuteIn(const Instruction &I, Loop *L, DominatorTree *DT) { |
321 | // TODO: merge these two routines. For the moment, we display the best |
322 | // result obtained by *either* implementation. This is a bit unfair since no |
323 | // caller actually gets the full power at the moment. |
324 | SimpleLoopSafetyInfo LSI; |
325 | LSI.computeLoopSafetyInfo(CurLoop: L); |
326 | return LSI.isGuaranteedToExecute(Inst: I, DT, CurLoop: L) || |
327 | isGuaranteedToExecuteForEveryIteration(I: &I, L); |
328 | } |
329 | |
330 | namespace { |
331 | /// An assembly annotator class to print must execute information in |
332 | /// comments. |
333 | class MustExecuteAnnotatedWriter : public AssemblyAnnotationWriter { |
334 | DenseMap<const Value*, SmallVector<Loop*, 4> > MustExec; |
335 | |
336 | public: |
337 | MustExecuteAnnotatedWriter(const Function &F, |
338 | DominatorTree &DT, LoopInfo &LI) { |
339 | for (const auto &I: instructions(F)) { |
340 | Loop *L = LI.getLoopFor(BB: I.getParent()); |
341 | while (L) { |
342 | if (isMustExecuteIn(I, L, DT: &DT)) { |
343 | MustExec[&I].push_back(Elt: L); |
344 | } |
345 | L = L->getParentLoop(); |
346 | }; |
347 | } |
348 | } |
349 | MustExecuteAnnotatedWriter(const Module &M, |
350 | DominatorTree &DT, LoopInfo &LI) { |
351 | for (const auto &F : M) |
352 | for (const auto &I: instructions(F)) { |
353 | Loop *L = LI.getLoopFor(BB: I.getParent()); |
354 | while (L) { |
355 | if (isMustExecuteIn(I, L, DT: &DT)) { |
356 | MustExec[&I].push_back(Elt: L); |
357 | } |
358 | L = L->getParentLoop(); |
359 | }; |
360 | } |
361 | } |
362 | |
363 | |
364 | void (const Value &V, formatted_raw_ostream &OS) override { |
365 | if (!MustExec.count(Val: &V)) |
366 | return; |
367 | |
368 | const auto &Loops = MustExec.lookup(Val: &V); |
369 | const auto NumLoops = Loops.size(); |
370 | if (NumLoops > 1) |
371 | OS << " ; (mustexec in " << NumLoops << " loops: " ; |
372 | else |
373 | OS << " ; (mustexec in: " ; |
374 | |
375 | ListSeparator LS; |
376 | for (const Loop *L : Loops) |
377 | OS << LS << L->getHeader()->getName(); |
378 | OS << ")" ; |
379 | } |
380 | }; |
381 | } // namespace |
382 | |
383 | /// Return true if \p L might be an endless loop. |
384 | static bool maybeEndlessLoop(const Loop &L) { |
385 | if (L.getHeader()->getParent()->hasFnAttribute(Kind: Attribute::WillReturn)) |
386 | return false; |
387 | // TODO: Actually try to prove it is not. |
388 | // TODO: If maybeEndlessLoop is going to be expensive, cache it. |
389 | return true; |
390 | } |
391 | |
392 | bool llvm::mayContainIrreducibleControl(const Function &F, const LoopInfo *LI) { |
393 | if (!LI) |
394 | return false; |
395 | using RPOTraversal = ReversePostOrderTraversal<const Function *>; |
396 | RPOTraversal FuncRPOT(&F); |
397 | return containsIrreducibleCFG<const BasicBlock *, const RPOTraversal, |
398 | const LoopInfo>(RPOTraversal: FuncRPOT, LI: *LI); |
399 | } |
400 | |
401 | /// Lookup \p Key in \p Map and return the result, potentially after |
402 | /// initializing the optional through \p Fn(\p args). |
403 | template <typename K, typename V, typename FnTy, typename... ArgsTy> |
404 | static V getOrCreateCachedOptional(K Key, DenseMap<K, std::optional<V>> &Map, |
405 | FnTy &&Fn, ArgsTy &&...args) { |
406 | std::optional<V> &OptVal = Map[Key]; |
407 | if (!OptVal) |
408 | OptVal = Fn(std::forward<ArgsTy>(args)...); |
409 | return *OptVal; |
410 | } |
411 | |
412 | const BasicBlock * |
413 | MustBeExecutedContextExplorer::findForwardJoinPoint(const BasicBlock *InitBB) { |
414 | const LoopInfo *LI = LIGetter(*InitBB->getParent()); |
415 | const PostDominatorTree *PDT = PDTGetter(*InitBB->getParent()); |
416 | |
417 | LLVM_DEBUG(dbgs() << "\tFind forward join point for " << InitBB->getName() |
418 | << (LI ? " [LI]" : "" ) << (PDT ? " [PDT]" : "" )); |
419 | |
420 | const Function &F = *InitBB->getParent(); |
421 | const Loop *L = LI ? LI->getLoopFor(BB: InitBB) : nullptr; |
422 | const BasicBlock * = L ? L->getHeader() : InitBB; |
423 | bool WillReturnAndNoThrow = (F.hasFnAttribute(Kind: Attribute::WillReturn) || |
424 | (L && !maybeEndlessLoop(L: *L))) && |
425 | F.doesNotThrow(); |
426 | LLVM_DEBUG(dbgs() << (L ? " [in loop]" : "" ) |
427 | << (WillReturnAndNoThrow ? " [WillReturn] [NoUnwind]" : "" ) |
428 | << "\n" ); |
429 | |
430 | // Determine the adjacent blocks in the given direction but exclude (self) |
431 | // loops under certain circumstances. |
432 | SmallVector<const BasicBlock *, 8> Worklist; |
433 | for (const BasicBlock *SuccBB : successors(BB: InitBB)) { |
434 | bool IsLatch = SuccBB == HeaderBB; |
435 | // Loop latches are ignored in forward propagation if the loop cannot be |
436 | // endless and may not throw: control has to go somewhere. |
437 | if (!WillReturnAndNoThrow || !IsLatch) |
438 | Worklist.push_back(Elt: SuccBB); |
439 | } |
440 | LLVM_DEBUG(dbgs() << "\t\t#Worklist: " << Worklist.size() << "\n" ); |
441 | |
442 | // If there are no other adjacent blocks, there is no join point. |
443 | if (Worklist.empty()) |
444 | return nullptr; |
445 | |
446 | // If there is one adjacent block, it is the join point. |
447 | if (Worklist.size() == 1) |
448 | return Worklist[0]; |
449 | |
450 | // Try to determine a join block through the help of the post-dominance |
451 | // tree. If no tree was provided, we perform simple pattern matching for one |
452 | // block conditionals and one block loops only. |
453 | const BasicBlock *JoinBB = nullptr; |
454 | if (PDT) |
455 | if (const auto *InitNode = PDT->getNode(BB: InitBB)) |
456 | if (const auto *IDomNode = InitNode->getIDom()) |
457 | JoinBB = IDomNode->getBlock(); |
458 | |
459 | if (!JoinBB && Worklist.size() == 2) { |
460 | const BasicBlock *Succ0 = Worklist[0]; |
461 | const BasicBlock *Succ1 = Worklist[1]; |
462 | const BasicBlock *Succ0UniqueSucc = Succ0->getUniqueSuccessor(); |
463 | const BasicBlock *Succ1UniqueSucc = Succ1->getUniqueSuccessor(); |
464 | if (Succ0UniqueSucc == InitBB) { |
465 | // InitBB -> Succ0 -> InitBB |
466 | // InitBB -> Succ1 = JoinBB |
467 | JoinBB = Succ1; |
468 | } else if (Succ1UniqueSucc == InitBB) { |
469 | // InitBB -> Succ1 -> InitBB |
470 | // InitBB -> Succ0 = JoinBB |
471 | JoinBB = Succ0; |
472 | } else if (Succ0 == Succ1UniqueSucc) { |
473 | // InitBB -> Succ0 = JoinBB |
474 | // InitBB -> Succ1 -> Succ0 = JoinBB |
475 | JoinBB = Succ0; |
476 | } else if (Succ1 == Succ0UniqueSucc) { |
477 | // InitBB -> Succ0 -> Succ1 = JoinBB |
478 | // InitBB -> Succ1 = JoinBB |
479 | JoinBB = Succ1; |
480 | } else if (Succ0UniqueSucc == Succ1UniqueSucc) { |
481 | // InitBB -> Succ0 -> JoinBB |
482 | // InitBB -> Succ1 -> JoinBB |
483 | JoinBB = Succ0UniqueSucc; |
484 | } |
485 | } |
486 | |
487 | if (!JoinBB && L) |
488 | JoinBB = L->getUniqueExitBlock(); |
489 | |
490 | if (!JoinBB) |
491 | return nullptr; |
492 | |
493 | LLVM_DEBUG(dbgs() << "\t\tJoin block candidate: " << JoinBB->getName() << "\n" ); |
494 | |
495 | // In forward direction we check if control will for sure reach JoinBB from |
496 | // InitBB, thus it can not be "stopped" along the way. Ways to "stop" control |
497 | // are: infinite loops and instructions that do not necessarily transfer |
498 | // execution to their successor. To check for them we traverse the CFG from |
499 | // the adjacent blocks to the JoinBB, looking at all intermediate blocks. |
500 | |
501 | // If we know the function is "will-return" and "no-throw" there is no need |
502 | // for futher checks. |
503 | if (!F.hasFnAttribute(Kind: Attribute::WillReturn) || !F.doesNotThrow()) { |
504 | |
505 | auto BlockTransfersExecutionToSuccessor = [](const BasicBlock *BB) { |
506 | return isGuaranteedToTransferExecutionToSuccessor(BB); |
507 | }; |
508 | |
509 | SmallPtrSet<const BasicBlock *, 16> Visited; |
510 | while (!Worklist.empty()) { |
511 | const BasicBlock *ToBB = Worklist.pop_back_val(); |
512 | if (ToBB == JoinBB) |
513 | continue; |
514 | |
515 | // Make sure all loops in-between are finite. |
516 | if (!Visited.insert(Ptr: ToBB).second) { |
517 | if (!F.hasFnAttribute(Kind: Attribute::WillReturn)) { |
518 | if (!LI) |
519 | return nullptr; |
520 | |
521 | bool MayContainIrreducibleControl = getOrCreateCachedOptional( |
522 | Key: &F, Map&: IrreducibleControlMap, Fn&: mayContainIrreducibleControl, args: F, args&: LI); |
523 | if (MayContainIrreducibleControl) |
524 | return nullptr; |
525 | |
526 | const Loop *L = LI->getLoopFor(BB: ToBB); |
527 | if (L && maybeEndlessLoop(L: *L)) |
528 | return nullptr; |
529 | } |
530 | |
531 | continue; |
532 | } |
533 | |
534 | // Make sure the block has no instructions that could stop control |
535 | // transfer. |
536 | bool TransfersExecution = getOrCreateCachedOptional( |
537 | Key: ToBB, Map&: BlockTransferMap, Fn&: BlockTransfersExecutionToSuccessor, args&: ToBB); |
538 | if (!TransfersExecution) |
539 | return nullptr; |
540 | |
541 | append_range(C&: Worklist, R: successors(BB: ToBB)); |
542 | } |
543 | } |
544 | |
545 | LLVM_DEBUG(dbgs() << "\tJoin block: " << JoinBB->getName() << "\n" ); |
546 | return JoinBB; |
547 | } |
548 | const BasicBlock * |
549 | MustBeExecutedContextExplorer::findBackwardJoinPoint(const BasicBlock *InitBB) { |
550 | const LoopInfo *LI = LIGetter(*InitBB->getParent()); |
551 | const DominatorTree *DT = DTGetter(*InitBB->getParent()); |
552 | LLVM_DEBUG(dbgs() << "\tFind backward join point for " << InitBB->getName() |
553 | << (LI ? " [LI]" : "" ) << (DT ? " [DT]" : "" )); |
554 | |
555 | // Try to determine a join block through the help of the dominance tree. If no |
556 | // tree was provided, we perform simple pattern matching for one block |
557 | // conditionals only. |
558 | if (DT) |
559 | if (const auto *InitNode = DT->getNode(BB: InitBB)) |
560 | if (const auto *IDomNode = InitNode->getIDom()) |
561 | return IDomNode->getBlock(); |
562 | |
563 | const Loop *L = LI ? LI->getLoopFor(BB: InitBB) : nullptr; |
564 | const BasicBlock * = L ? L->getHeader() : nullptr; |
565 | |
566 | // Determine the predecessor blocks but ignore backedges. |
567 | SmallVector<const BasicBlock *, 8> Worklist; |
568 | for (const BasicBlock *PredBB : predecessors(BB: InitBB)) { |
569 | bool IsBackedge = |
570 | (PredBB == InitBB) || (HeaderBB == InitBB && L->contains(BB: PredBB)); |
571 | // Loop backedges are ignored in backwards propagation: control has to come |
572 | // from somewhere. |
573 | if (!IsBackedge) |
574 | Worklist.push_back(Elt: PredBB); |
575 | } |
576 | |
577 | // If there are no other predecessor blocks, there is no join point. |
578 | if (Worklist.empty()) |
579 | return nullptr; |
580 | |
581 | // If there is one predecessor block, it is the join point. |
582 | if (Worklist.size() == 1) |
583 | return Worklist[0]; |
584 | |
585 | const BasicBlock *JoinBB = nullptr; |
586 | if (Worklist.size() == 2) { |
587 | const BasicBlock *Pred0 = Worklist[0]; |
588 | const BasicBlock *Pred1 = Worklist[1]; |
589 | const BasicBlock *Pred0UniquePred = Pred0->getUniquePredecessor(); |
590 | const BasicBlock *Pred1UniquePred = Pred1->getUniquePredecessor(); |
591 | if (Pred0 == Pred1UniquePred) { |
592 | // InitBB <- Pred0 = JoinBB |
593 | // InitBB <- Pred1 <- Pred0 = JoinBB |
594 | JoinBB = Pred0; |
595 | } else if (Pred1 == Pred0UniquePred) { |
596 | // InitBB <- Pred0 <- Pred1 = JoinBB |
597 | // InitBB <- Pred1 = JoinBB |
598 | JoinBB = Pred1; |
599 | } else if (Pred0UniquePred == Pred1UniquePred) { |
600 | // InitBB <- Pred0 <- JoinBB |
601 | // InitBB <- Pred1 <- JoinBB |
602 | JoinBB = Pred0UniquePred; |
603 | } |
604 | } |
605 | |
606 | if (!JoinBB && L) |
607 | JoinBB = L->getHeader(); |
608 | |
609 | // In backwards direction there is no need to show termination of previous |
610 | // instructions. If they do not terminate, the code afterward is dead, making |
611 | // any information/transformation correct anyway. |
612 | return JoinBB; |
613 | } |
614 | |
615 | const Instruction * |
616 | MustBeExecutedContextExplorer::getMustBeExecutedNextInstruction( |
617 | MustBeExecutedIterator &It, const Instruction *PP) { |
618 | if (!PP) |
619 | return PP; |
620 | LLVM_DEBUG(dbgs() << "Find next instruction for " << *PP << "\n" ); |
621 | |
622 | // If we explore only inside a given basic block we stop at terminators. |
623 | if (!ExploreInterBlock && PP->isTerminator()) { |
624 | LLVM_DEBUG(dbgs() << "\tReached terminator in intra-block mode, done\n" ); |
625 | return nullptr; |
626 | } |
627 | |
628 | // If we do not traverse the call graph we check if we can make progress in |
629 | // the current function. First, check if the instruction is guaranteed to |
630 | // transfer execution to the successor. |
631 | bool TransfersExecution = isGuaranteedToTransferExecutionToSuccessor(I: PP); |
632 | if (!TransfersExecution) |
633 | return nullptr; |
634 | |
635 | // If this is not a terminator we know that there is a single instruction |
636 | // after this one that is executed next if control is transfered. If not, |
637 | // we can try to go back to a call site we entered earlier. If none exists, we |
638 | // do not know any instruction that has to be executd next. |
639 | if (!PP->isTerminator()) { |
640 | const Instruction *NextPP = PP->getNextNode(); |
641 | LLVM_DEBUG(dbgs() << "\tIntermediate instruction does transfer control\n" ); |
642 | return NextPP; |
643 | } |
644 | |
645 | // Finally, we have to handle terminators, trivial ones first. |
646 | assert(PP->isTerminator() && "Expected a terminator!" ); |
647 | |
648 | // A terminator without a successor is not handled yet. |
649 | if (PP->getNumSuccessors() == 0) { |
650 | LLVM_DEBUG(dbgs() << "\tUnhandled terminator\n" ); |
651 | return nullptr; |
652 | } |
653 | |
654 | // A terminator with a single successor, we will continue at the beginning of |
655 | // that one. |
656 | if (PP->getNumSuccessors() == 1) { |
657 | LLVM_DEBUG( |
658 | dbgs() << "\tUnconditional terminator, continue with successor\n" ); |
659 | return &PP->getSuccessor(Idx: 0)->front(); |
660 | } |
661 | |
662 | // Multiple successors mean we need to find the join point where control flow |
663 | // converges again. We use the findForwardJoinPoint helper function with |
664 | // information about the function and helper analyses, if available. |
665 | if (const BasicBlock *JoinBB = findForwardJoinPoint(InitBB: PP->getParent())) |
666 | return &JoinBB->front(); |
667 | |
668 | LLVM_DEBUG(dbgs() << "\tNo join point found\n" ); |
669 | return nullptr; |
670 | } |
671 | |
672 | const Instruction * |
673 | MustBeExecutedContextExplorer::getMustBeExecutedPrevInstruction( |
674 | MustBeExecutedIterator &It, const Instruction *PP) { |
675 | if (!PP) |
676 | return PP; |
677 | |
678 | bool IsFirst = !(PP->getPrevNode()); |
679 | LLVM_DEBUG(dbgs() << "Find next instruction for " << *PP |
680 | << (IsFirst ? " [IsFirst]" : "" ) << "\n" ); |
681 | |
682 | // If we explore only inside a given basic block we stop at the first |
683 | // instruction. |
684 | if (!ExploreInterBlock && IsFirst) { |
685 | LLVM_DEBUG(dbgs() << "\tReached block front in intra-block mode, done\n" ); |
686 | return nullptr; |
687 | } |
688 | |
689 | // The block and function that contains the current position. |
690 | const BasicBlock *PPBlock = PP->getParent(); |
691 | |
692 | // If we are inside a block we know what instruction was executed before, the |
693 | // previous one. |
694 | if (!IsFirst) { |
695 | const Instruction *PrevPP = PP->getPrevNode(); |
696 | LLVM_DEBUG( |
697 | dbgs() << "\tIntermediate instruction, continue with previous\n" ); |
698 | // We did not enter a callee so we simply return the previous instruction. |
699 | return PrevPP; |
700 | } |
701 | |
702 | // Finally, we have to handle the case where the program point is the first in |
703 | // a block but not in the function. We use the findBackwardJoinPoint helper |
704 | // function with information about the function and helper analyses, if |
705 | // available. |
706 | if (const BasicBlock *JoinBB = findBackwardJoinPoint(InitBB: PPBlock)) |
707 | return &JoinBB->back(); |
708 | |
709 | LLVM_DEBUG(dbgs() << "\tNo join point found\n" ); |
710 | return nullptr; |
711 | } |
712 | |
713 | MustBeExecutedIterator::MustBeExecutedIterator( |
714 | MustBeExecutedContextExplorer &Explorer, const Instruction *I) |
715 | : Explorer(Explorer), CurInst(I) { |
716 | reset(I); |
717 | } |
718 | |
719 | void MustBeExecutedIterator::reset(const Instruction *I) { |
720 | Visited.clear(); |
721 | resetInstruction(I); |
722 | } |
723 | |
724 | void MustBeExecutedIterator::resetInstruction(const Instruction *I) { |
725 | CurInst = I; |
726 | Head = Tail = nullptr; |
727 | Visited.insert(V: {I, ExplorationDirection::FORWARD}); |
728 | Visited.insert(V: {I, ExplorationDirection::BACKWARD}); |
729 | if (Explorer.ExploreCFGForward) |
730 | Head = I; |
731 | if (Explorer.ExploreCFGBackward) |
732 | Tail = I; |
733 | } |
734 | |
735 | const Instruction *MustBeExecutedIterator::advance() { |
736 | assert(CurInst && "Cannot advance an end iterator!" ); |
737 | Head = Explorer.getMustBeExecutedNextInstruction(It&: *this, PP: Head); |
738 | if (Head && Visited.insert(V: {Head, ExplorationDirection ::FORWARD}).second) |
739 | return Head; |
740 | Head = nullptr; |
741 | |
742 | Tail = Explorer.getMustBeExecutedPrevInstruction(It&: *this, PP: Tail); |
743 | if (Tail && Visited.insert(V: {Tail, ExplorationDirection ::BACKWARD}).second) |
744 | return Tail; |
745 | Tail = nullptr; |
746 | return nullptr; |
747 | } |
748 | |
749 | PreservedAnalyses MustExecutePrinterPass::run(Function &F, |
750 | FunctionAnalysisManager &AM) { |
751 | auto &LI = AM.getResult<LoopAnalysis>(IR&: F); |
752 | auto &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
753 | |
754 | MustExecuteAnnotatedWriter Writer(F, DT, LI); |
755 | F.print(OS, AAW: &Writer); |
756 | return PreservedAnalyses::all(); |
757 | } |
758 | |
759 | PreservedAnalyses |
760 | MustBeExecutedContextPrinterPass::run(Module &M, ModuleAnalysisManager &AM) { |
761 | FunctionAnalysisManager &FAM = |
762 | AM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager(); |
763 | GetterTy<const LoopInfo> LIGetter = [&](const Function &F) { |
764 | return &FAM.getResult<LoopAnalysis>(IR&: const_cast<Function &>(F)); |
765 | }; |
766 | GetterTy<const DominatorTree> DTGetter = [&](const Function &F) { |
767 | return &FAM.getResult<DominatorTreeAnalysis>(IR&: const_cast<Function &>(F)); |
768 | }; |
769 | GetterTy<const PostDominatorTree> PDTGetter = [&](const Function &F) { |
770 | return &FAM.getResult<PostDominatorTreeAnalysis>(IR&: const_cast<Function &>(F)); |
771 | }; |
772 | |
773 | MustBeExecutedContextExplorer Explorer( |
774 | /* ExploreInterBlock */ true, |
775 | /* ExploreCFGForward */ true, |
776 | /* ExploreCFGBackward */ true, LIGetter, DTGetter, PDTGetter); |
777 | |
778 | for (Function &F : M) { |
779 | for (Instruction &I : instructions(F)) { |
780 | OS << "-- Explore context of: " << I << "\n" ; |
781 | for (const Instruction *CI : Explorer.range(PP: &I)) |
782 | OS << " [F: " << CI->getFunction()->getName() << "] " << *CI << "\n" ; |
783 | } |
784 | } |
785 | return PreservedAnalyses::all(); |
786 | } |
787 | |