1 | //===- InlineFunction.cpp - Code to perform function inlining -------------===// |
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 | // This file implements inlining of a function into a call site, resolving |
10 | // parameters and the return value as appropriate. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "llvm/ADT/DenseMap.h" |
15 | #include "llvm/ADT/STLExtras.h" |
16 | #include "llvm/ADT/SetVector.h" |
17 | #include "llvm/ADT/SmallPtrSet.h" |
18 | #include "llvm/ADT/SmallVector.h" |
19 | #include "llvm/ADT/StringExtras.h" |
20 | #include "llvm/ADT/iterator_range.h" |
21 | #include "llvm/Analysis/AliasAnalysis.h" |
22 | #include "llvm/Analysis/AssumptionCache.h" |
23 | #include "llvm/Analysis/BlockFrequencyInfo.h" |
24 | #include "llvm/Analysis/CallGraph.h" |
25 | #include "llvm/Analysis/CaptureTracking.h" |
26 | #include "llvm/Analysis/IndirectCallVisitor.h" |
27 | #include "llvm/Analysis/InstructionSimplify.h" |
28 | #include "llvm/Analysis/MemoryProfileInfo.h" |
29 | #include "llvm/Analysis/ObjCARCAnalysisUtils.h" |
30 | #include "llvm/Analysis/ObjCARCUtil.h" |
31 | #include "llvm/Analysis/ProfileSummaryInfo.h" |
32 | #include "llvm/Analysis/ValueTracking.h" |
33 | #include "llvm/Analysis/VectorUtils.h" |
34 | #include "llvm/IR/Argument.h" |
35 | #include "llvm/IR/AttributeMask.h" |
36 | #include "llvm/IR/BasicBlock.h" |
37 | #include "llvm/IR/CFG.h" |
38 | #include "llvm/IR/Constant.h" |
39 | #include "llvm/IR/ConstantRange.h" |
40 | #include "llvm/IR/Constants.h" |
41 | #include "llvm/IR/DataLayout.h" |
42 | #include "llvm/IR/DebugInfo.h" |
43 | #include "llvm/IR/DebugInfoMetadata.h" |
44 | #include "llvm/IR/DebugLoc.h" |
45 | #include "llvm/IR/DerivedTypes.h" |
46 | #include "llvm/IR/Dominators.h" |
47 | #include "llvm/IR/EHPersonalities.h" |
48 | #include "llvm/IR/Function.h" |
49 | #include "llvm/IR/IRBuilder.h" |
50 | #include "llvm/IR/InlineAsm.h" |
51 | #include "llvm/IR/InstrTypes.h" |
52 | #include "llvm/IR/Instruction.h" |
53 | #include "llvm/IR/Instructions.h" |
54 | #include "llvm/IR/IntrinsicInst.h" |
55 | #include "llvm/IR/Intrinsics.h" |
56 | #include "llvm/IR/LLVMContext.h" |
57 | #include "llvm/IR/MDBuilder.h" |
58 | #include "llvm/IR/Metadata.h" |
59 | #include "llvm/IR/Module.h" |
60 | #include "llvm/IR/ProfDataUtils.h" |
61 | #include "llvm/IR/Type.h" |
62 | #include "llvm/IR/User.h" |
63 | #include "llvm/IR/Value.h" |
64 | #include "llvm/Support/Casting.h" |
65 | #include "llvm/Support/CommandLine.h" |
66 | #include "llvm/Support/ErrorHandling.h" |
67 | #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" |
68 | #include "llvm/Transforms/Utils/Cloning.h" |
69 | #include "llvm/Transforms/Utils/Local.h" |
70 | #include "llvm/Transforms/Utils/ValueMapper.h" |
71 | #include <algorithm> |
72 | #include <cassert> |
73 | #include <cstdint> |
74 | #include <iterator> |
75 | #include <limits> |
76 | #include <optional> |
77 | #include <string> |
78 | #include <utility> |
79 | #include <vector> |
80 | |
81 | #define DEBUG_TYPE "inline-function" |
82 | |
83 | using namespace llvm; |
84 | using namespace llvm::memprof; |
85 | using ProfileCount = Function::ProfileCount; |
86 | |
87 | static cl::opt<bool> |
88 | EnableNoAliasConversion("enable-noalias-to-md-conversion" , cl::init(Val: true), |
89 | cl::Hidden, |
90 | cl::desc("Convert noalias attributes to metadata during inlining." )); |
91 | |
92 | static cl::opt<bool> |
93 | UseNoAliasIntrinsic("use-noalias-intrinsic-during-inlining" , cl::Hidden, |
94 | cl::init(Val: true), |
95 | cl::desc("Use the llvm.experimental.noalias.scope.decl " |
96 | "intrinsic during inlining." )); |
97 | |
98 | // Disabled by default, because the added alignment assumptions may increase |
99 | // compile-time and block optimizations. This option is not suitable for use |
100 | // with frontends that emit comprehensive parameter alignment annotations. |
101 | static cl::opt<bool> |
102 | PreserveAlignmentAssumptions("preserve-alignment-assumptions-during-inlining" , |
103 | cl::init(Val: false), cl::Hidden, |
104 | cl::desc("Convert align attributes to assumptions during inlining." )); |
105 | |
106 | static cl::opt<unsigned> InlinerAttributeWindow( |
107 | "max-inst-checked-for-throw-during-inlining" , cl::Hidden, |
108 | cl::desc("the maximum number of instructions analyzed for may throw during " |
109 | "attribute inference in inlined body" ), |
110 | cl::init(Val: 4)); |
111 | |
112 | namespace { |
113 | |
114 | /// A class for recording information about inlining a landing pad. |
115 | class LandingPadInliningInfo { |
116 | /// Destination of the invoke's unwind. |
117 | BasicBlock *OuterResumeDest; |
118 | |
119 | /// Destination for the callee's resume. |
120 | BasicBlock *InnerResumeDest = nullptr; |
121 | |
122 | /// LandingPadInst associated with the invoke. |
123 | LandingPadInst *CallerLPad = nullptr; |
124 | |
125 | /// PHI for EH values from landingpad insts. |
126 | PHINode *InnerEHValuesPHI = nullptr; |
127 | |
128 | SmallVector<Value*, 8> UnwindDestPHIValues; |
129 | |
130 | public: |
131 | LandingPadInliningInfo(InvokeInst *II) |
132 | : OuterResumeDest(II->getUnwindDest()) { |
133 | // If there are PHI nodes in the unwind destination block, we need to keep |
134 | // track of which values came into them from the invoke before removing |
135 | // the edge from this block. |
136 | BasicBlock *InvokeBB = II->getParent(); |
137 | BasicBlock::iterator I = OuterResumeDest->begin(); |
138 | for (; isa<PHINode>(Val: I); ++I) { |
139 | // Save the value to use for this edge. |
140 | PHINode *PHI = cast<PHINode>(Val&: I); |
141 | UnwindDestPHIValues.push_back(Elt: PHI->getIncomingValueForBlock(BB: InvokeBB)); |
142 | } |
143 | |
144 | CallerLPad = cast<LandingPadInst>(Val&: I); |
145 | } |
146 | |
147 | /// The outer unwind destination is the target of |
148 | /// unwind edges introduced for calls within the inlined function. |
149 | BasicBlock *getOuterResumeDest() const { |
150 | return OuterResumeDest; |
151 | } |
152 | |
153 | BasicBlock *getInnerResumeDest(); |
154 | |
155 | LandingPadInst *getLandingPadInst() const { return CallerLPad; } |
156 | |
157 | /// Forward the 'resume' instruction to the caller's landing pad block. |
158 | /// When the landing pad block has only one predecessor, this is |
159 | /// a simple branch. When there is more than one predecessor, we need to |
160 | /// split the landing pad block after the landingpad instruction and jump |
161 | /// to there. |
162 | void forwardResume(ResumeInst *RI, |
163 | SmallPtrSetImpl<LandingPadInst*> &InlinedLPads); |
164 | |
165 | /// Add incoming-PHI values to the unwind destination block for the given |
166 | /// basic block, using the values for the original invoke's source block. |
167 | void addIncomingPHIValuesFor(BasicBlock *BB) const { |
168 | addIncomingPHIValuesForInto(src: BB, dest: OuterResumeDest); |
169 | } |
170 | |
171 | void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const { |
172 | BasicBlock::iterator I = dest->begin(); |
173 | for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { |
174 | PHINode *phi = cast<PHINode>(Val&: I); |
175 | phi->addIncoming(V: UnwindDestPHIValues[i], BB: src); |
176 | } |
177 | } |
178 | }; |
179 | |
180 | } // end anonymous namespace |
181 | |
182 | /// Get or create a target for the branch from ResumeInsts. |
183 | BasicBlock *LandingPadInliningInfo::getInnerResumeDest() { |
184 | if (InnerResumeDest) return InnerResumeDest; |
185 | |
186 | // Split the landing pad. |
187 | BasicBlock::iterator SplitPoint = ++CallerLPad->getIterator(); |
188 | InnerResumeDest = |
189 | OuterResumeDest->splitBasicBlock(I: SplitPoint, |
190 | BBName: OuterResumeDest->getName() + ".body" ); |
191 | |
192 | // The number of incoming edges we expect to the inner landing pad. |
193 | const unsigned PHICapacity = 2; |
194 | |
195 | // Create corresponding new PHIs for all the PHIs in the outer landing pad. |
196 | BasicBlock::iterator InsertPoint = InnerResumeDest->begin(); |
197 | BasicBlock::iterator I = OuterResumeDest->begin(); |
198 | for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { |
199 | PHINode *OuterPHI = cast<PHINode>(Val&: I); |
200 | PHINode *InnerPHI = PHINode::Create(Ty: OuterPHI->getType(), NumReservedValues: PHICapacity, |
201 | NameStr: OuterPHI->getName() + ".lpad-body" ); |
202 | InnerPHI->insertBefore(InsertPos: InsertPoint); |
203 | OuterPHI->replaceAllUsesWith(V: InnerPHI); |
204 | InnerPHI->addIncoming(V: OuterPHI, BB: OuterResumeDest); |
205 | } |
206 | |
207 | // Create a PHI for the exception values. |
208 | InnerEHValuesPHI = |
209 | PHINode::Create(Ty: CallerLPad->getType(), NumReservedValues: PHICapacity, NameStr: "eh.lpad-body" ); |
210 | InnerEHValuesPHI->insertBefore(InsertPos: InsertPoint); |
211 | CallerLPad->replaceAllUsesWith(V: InnerEHValuesPHI); |
212 | InnerEHValuesPHI->addIncoming(V: CallerLPad, BB: OuterResumeDest); |
213 | |
214 | // All done. |
215 | return InnerResumeDest; |
216 | } |
217 | |
218 | /// Forward the 'resume' instruction to the caller's landing pad block. |
219 | /// When the landing pad block has only one predecessor, this is a simple |
220 | /// branch. When there is more than one predecessor, we need to split the |
221 | /// landing pad block after the landingpad instruction and jump to there. |
222 | void LandingPadInliningInfo::forwardResume( |
223 | ResumeInst *RI, SmallPtrSetImpl<LandingPadInst *> &InlinedLPads) { |
224 | BasicBlock *Dest = getInnerResumeDest(); |
225 | BasicBlock *Src = RI->getParent(); |
226 | |
227 | BranchInst::Create(IfTrue: Dest, InsertBefore: Src); |
228 | |
229 | // Update the PHIs in the destination. They were inserted in an order which |
230 | // makes this work. |
231 | addIncomingPHIValuesForInto(src: Src, dest: Dest); |
232 | |
233 | InnerEHValuesPHI->addIncoming(V: RI->getOperand(i_nocapture: 0), BB: Src); |
234 | RI->eraseFromParent(); |
235 | } |
236 | |
237 | /// Helper for getUnwindDestToken/getUnwindDestTokenHelper. |
238 | static Value *getParentPad(Value *EHPad) { |
239 | if (auto *FPI = dyn_cast<FuncletPadInst>(Val: EHPad)) |
240 | return FPI->getParentPad(); |
241 | return cast<CatchSwitchInst>(Val: EHPad)->getParentPad(); |
242 | } |
243 | |
244 | using UnwindDestMemoTy = DenseMap<Instruction *, Value *>; |
245 | |
246 | /// Helper for getUnwindDestToken that does the descendant-ward part of |
247 | /// the search. |
248 | static Value *getUnwindDestTokenHelper(Instruction *EHPad, |
249 | UnwindDestMemoTy &MemoMap) { |
250 | SmallVector<Instruction *, 8> Worklist(1, EHPad); |
251 | |
252 | while (!Worklist.empty()) { |
253 | Instruction *CurrentPad = Worklist.pop_back_val(); |
254 | // We only put pads on the worklist that aren't in the MemoMap. When |
255 | // we find an unwind dest for a pad we may update its ancestors, but |
256 | // the queue only ever contains uncles/great-uncles/etc. of CurrentPad, |
257 | // so they should never get updated while queued on the worklist. |
258 | assert(!MemoMap.count(CurrentPad)); |
259 | Value *UnwindDestToken = nullptr; |
260 | if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Val: CurrentPad)) { |
261 | if (CatchSwitch->hasUnwindDest()) { |
262 | UnwindDestToken = CatchSwitch->getUnwindDest()->getFirstNonPHI(); |
263 | } else { |
264 | // Catchswitch doesn't have a 'nounwind' variant, and one might be |
265 | // annotated as "unwinds to caller" when really it's nounwind (see |
266 | // e.g. SimplifyCFGOpt::SimplifyUnreachable), so we can't infer the |
267 | // parent's unwind dest from this. We can check its catchpads' |
268 | // descendants, since they might include a cleanuppad with an |
269 | // "unwinds to caller" cleanupret, which can be trusted. |
270 | for (auto HI = CatchSwitch->handler_begin(), |
271 | HE = CatchSwitch->handler_end(); |
272 | HI != HE && !UnwindDestToken; ++HI) { |
273 | BasicBlock *HandlerBlock = *HI; |
274 | auto *CatchPad = cast<CatchPadInst>(Val: HandlerBlock->getFirstNonPHI()); |
275 | for (User *Child : CatchPad->users()) { |
276 | // Intentionally ignore invokes here -- since the catchswitch is |
277 | // marked "unwind to caller", it would be a verifier error if it |
278 | // contained an invoke which unwinds out of it, so any invoke we'd |
279 | // encounter must unwind to some child of the catch. |
280 | if (!isa<CleanupPadInst>(Val: Child) && !isa<CatchSwitchInst>(Val: Child)) |
281 | continue; |
282 | |
283 | Instruction *ChildPad = cast<Instruction>(Val: Child); |
284 | auto Memo = MemoMap.find(Val: ChildPad); |
285 | if (Memo == MemoMap.end()) { |
286 | // Haven't figured out this child pad yet; queue it. |
287 | Worklist.push_back(Elt: ChildPad); |
288 | continue; |
289 | } |
290 | // We've already checked this child, but might have found that |
291 | // it offers no proof either way. |
292 | Value *ChildUnwindDestToken = Memo->second; |
293 | if (!ChildUnwindDestToken) |
294 | continue; |
295 | // We already know the child's unwind dest, which can either |
296 | // be ConstantTokenNone to indicate unwind to caller, or can |
297 | // be another child of the catchpad. Only the former indicates |
298 | // the unwind dest of the catchswitch. |
299 | if (isa<ConstantTokenNone>(Val: ChildUnwindDestToken)) { |
300 | UnwindDestToken = ChildUnwindDestToken; |
301 | break; |
302 | } |
303 | assert(getParentPad(ChildUnwindDestToken) == CatchPad); |
304 | } |
305 | } |
306 | } |
307 | } else { |
308 | auto *CleanupPad = cast<CleanupPadInst>(Val: CurrentPad); |
309 | for (User *U : CleanupPad->users()) { |
310 | if (auto *CleanupRet = dyn_cast<CleanupReturnInst>(Val: U)) { |
311 | if (BasicBlock *RetUnwindDest = CleanupRet->getUnwindDest()) |
312 | UnwindDestToken = RetUnwindDest->getFirstNonPHI(); |
313 | else |
314 | UnwindDestToken = ConstantTokenNone::get(Context&: CleanupPad->getContext()); |
315 | break; |
316 | } |
317 | Value *ChildUnwindDestToken; |
318 | if (auto *Invoke = dyn_cast<InvokeInst>(Val: U)) { |
319 | ChildUnwindDestToken = Invoke->getUnwindDest()->getFirstNonPHI(); |
320 | } else if (isa<CleanupPadInst>(Val: U) || isa<CatchSwitchInst>(Val: U)) { |
321 | Instruction *ChildPad = cast<Instruction>(Val: U); |
322 | auto Memo = MemoMap.find(Val: ChildPad); |
323 | if (Memo == MemoMap.end()) { |
324 | // Haven't resolved this child yet; queue it and keep searching. |
325 | Worklist.push_back(Elt: ChildPad); |
326 | continue; |
327 | } |
328 | // We've checked this child, but still need to ignore it if it |
329 | // had no proof either way. |
330 | ChildUnwindDestToken = Memo->second; |
331 | if (!ChildUnwindDestToken) |
332 | continue; |
333 | } else { |
334 | // Not a relevant user of the cleanuppad |
335 | continue; |
336 | } |
337 | // In a well-formed program, the child/invoke must either unwind to |
338 | // an(other) child of the cleanup, or exit the cleanup. In the |
339 | // first case, continue searching. |
340 | if (isa<Instruction>(Val: ChildUnwindDestToken) && |
341 | getParentPad(EHPad: ChildUnwindDestToken) == CleanupPad) |
342 | continue; |
343 | UnwindDestToken = ChildUnwindDestToken; |
344 | break; |
345 | } |
346 | } |
347 | // If we haven't found an unwind dest for CurrentPad, we may have queued its |
348 | // children, so move on to the next in the worklist. |
349 | if (!UnwindDestToken) |
350 | continue; |
351 | |
352 | // Now we know that CurrentPad unwinds to UnwindDestToken. It also exits |
353 | // any ancestors of CurrentPad up to but not including UnwindDestToken's |
354 | // parent pad. Record this in the memo map, and check to see if the |
355 | // original EHPad being queried is one of the ones exited. |
356 | Value *UnwindParent; |
357 | if (auto *UnwindPad = dyn_cast<Instruction>(Val: UnwindDestToken)) |
358 | UnwindParent = getParentPad(EHPad: UnwindPad); |
359 | else |
360 | UnwindParent = nullptr; |
361 | bool ExitedOriginalPad = false; |
362 | for (Instruction *ExitedPad = CurrentPad; |
363 | ExitedPad && ExitedPad != UnwindParent; |
364 | ExitedPad = dyn_cast<Instruction>(Val: getParentPad(EHPad: ExitedPad))) { |
365 | // Skip over catchpads since they just follow their catchswitches. |
366 | if (isa<CatchPadInst>(Val: ExitedPad)) |
367 | continue; |
368 | MemoMap[ExitedPad] = UnwindDestToken; |
369 | ExitedOriginalPad |= (ExitedPad == EHPad); |
370 | } |
371 | |
372 | if (ExitedOriginalPad) |
373 | return UnwindDestToken; |
374 | |
375 | // Continue the search. |
376 | } |
377 | |
378 | // No definitive information is contained within this funclet. |
379 | return nullptr; |
380 | } |
381 | |
382 | /// Given an EH pad, find where it unwinds. If it unwinds to an EH pad, |
383 | /// return that pad instruction. If it unwinds to caller, return |
384 | /// ConstantTokenNone. If it does not have a definitive unwind destination, |
385 | /// return nullptr. |
386 | /// |
387 | /// This routine gets invoked for calls in funclets in inlinees when inlining |
388 | /// an invoke. Since many funclets don't have calls inside them, it's queried |
389 | /// on-demand rather than building a map of pads to unwind dests up front. |
390 | /// Determining a funclet's unwind dest may require recursively searching its |
391 | /// descendants, and also ancestors and cousins if the descendants don't provide |
392 | /// an answer. Since most funclets will have their unwind dest immediately |
393 | /// available as the unwind dest of a catchswitch or cleanupret, this routine |
394 | /// searches top-down from the given pad and then up. To avoid worst-case |
395 | /// quadratic run-time given that approach, it uses a memo map to avoid |
396 | /// re-processing funclet trees. The callers that rewrite the IR as they go |
397 | /// take advantage of this, for correctness, by checking/forcing rewritten |
398 | /// pads' entries to match the original callee view. |
399 | static Value *getUnwindDestToken(Instruction *EHPad, |
400 | UnwindDestMemoTy &MemoMap) { |
401 | // Catchpads unwind to the same place as their catchswitch; |
402 | // redirct any queries on catchpads so the code below can |
403 | // deal with just catchswitches and cleanuppads. |
404 | if (auto *CPI = dyn_cast<CatchPadInst>(Val: EHPad)) |
405 | EHPad = CPI->getCatchSwitch(); |
406 | |
407 | // Check if we've already determined the unwind dest for this pad. |
408 | auto Memo = MemoMap.find(Val: EHPad); |
409 | if (Memo != MemoMap.end()) |
410 | return Memo->second; |
411 | |
412 | // Search EHPad and, if necessary, its descendants. |
413 | Value *UnwindDestToken = getUnwindDestTokenHelper(EHPad, MemoMap); |
414 | assert((UnwindDestToken == nullptr) != (MemoMap.count(EHPad) != 0)); |
415 | if (UnwindDestToken) |
416 | return UnwindDestToken; |
417 | |
418 | // No information is available for this EHPad from itself or any of its |
419 | // descendants. An unwind all the way out to a pad in the caller would |
420 | // need also to agree with the unwind dest of the parent funclet, so |
421 | // search up the chain to try to find a funclet with information. Put |
422 | // null entries in the memo map to avoid re-processing as we go up. |
423 | MemoMap[EHPad] = nullptr; |
424 | #ifndef NDEBUG |
425 | SmallPtrSet<Instruction *, 4> TempMemos; |
426 | TempMemos.insert(EHPad); |
427 | #endif |
428 | Instruction *LastUselessPad = EHPad; |
429 | Value *AncestorToken; |
430 | for (AncestorToken = getParentPad(EHPad); |
431 | auto *AncestorPad = dyn_cast<Instruction>(Val: AncestorToken); |
432 | AncestorToken = getParentPad(EHPad: AncestorToken)) { |
433 | // Skip over catchpads since they just follow their catchswitches. |
434 | if (isa<CatchPadInst>(Val: AncestorPad)) |
435 | continue; |
436 | // If the MemoMap had an entry mapping AncestorPad to nullptr, since we |
437 | // haven't yet called getUnwindDestTokenHelper for AncestorPad in this |
438 | // call to getUnwindDestToken, that would mean that AncestorPad had no |
439 | // information in itself, its descendants, or its ancestors. If that |
440 | // were the case, then we should also have recorded the lack of information |
441 | // for the descendant that we're coming from. So assert that we don't |
442 | // find a null entry in the MemoMap for AncestorPad. |
443 | assert(!MemoMap.count(AncestorPad) || MemoMap[AncestorPad]); |
444 | auto AncestorMemo = MemoMap.find(Val: AncestorPad); |
445 | if (AncestorMemo == MemoMap.end()) { |
446 | UnwindDestToken = getUnwindDestTokenHelper(EHPad: AncestorPad, MemoMap); |
447 | } else { |
448 | UnwindDestToken = AncestorMemo->second; |
449 | } |
450 | if (UnwindDestToken) |
451 | break; |
452 | LastUselessPad = AncestorPad; |
453 | MemoMap[LastUselessPad] = nullptr; |
454 | #ifndef NDEBUG |
455 | TempMemos.insert(LastUselessPad); |
456 | #endif |
457 | } |
458 | |
459 | // We know that getUnwindDestTokenHelper was called on LastUselessPad and |
460 | // returned nullptr (and likewise for EHPad and any of its ancestors up to |
461 | // LastUselessPad), so LastUselessPad has no information from below. Since |
462 | // getUnwindDestTokenHelper must investigate all downward paths through |
463 | // no-information nodes to prove that a node has no information like this, |
464 | // and since any time it finds information it records it in the MemoMap for |
465 | // not just the immediately-containing funclet but also any ancestors also |
466 | // exited, it must be the case that, walking downward from LastUselessPad, |
467 | // visiting just those nodes which have not been mapped to an unwind dest |
468 | // by getUnwindDestTokenHelper (the nullptr TempMemos notwithstanding, since |
469 | // they are just used to keep getUnwindDestTokenHelper from repeating work), |
470 | // any node visited must have been exhaustively searched with no information |
471 | // for it found. |
472 | SmallVector<Instruction *, 8> Worklist(1, LastUselessPad); |
473 | while (!Worklist.empty()) { |
474 | Instruction *UselessPad = Worklist.pop_back_val(); |
475 | auto Memo = MemoMap.find(Val: UselessPad); |
476 | if (Memo != MemoMap.end() && Memo->second) { |
477 | // Here the name 'UselessPad' is a bit of a misnomer, because we've found |
478 | // that it is a funclet that does have information about unwinding to |
479 | // a particular destination; its parent was a useless pad. |
480 | // Since its parent has no information, the unwind edge must not escape |
481 | // the parent, and must target a sibling of this pad. This local unwind |
482 | // gives us no information about EHPad. Leave it and the subtree rooted |
483 | // at it alone. |
484 | assert(getParentPad(Memo->second) == getParentPad(UselessPad)); |
485 | continue; |
486 | } |
487 | // We know we don't have information for UselesPad. If it has an entry in |
488 | // the MemoMap (mapping it to nullptr), it must be one of the TempMemos |
489 | // added on this invocation of getUnwindDestToken; if a previous invocation |
490 | // recorded nullptr, it would have had to prove that the ancestors of |
491 | // UselessPad, which include LastUselessPad, had no information, and that |
492 | // in turn would have required proving that the descendants of |
493 | // LastUselesPad, which include EHPad, have no information about |
494 | // LastUselessPad, which would imply that EHPad was mapped to nullptr in |
495 | // the MemoMap on that invocation, which isn't the case if we got here. |
496 | assert(!MemoMap.count(UselessPad) || TempMemos.count(UselessPad)); |
497 | // Assert as we enumerate users that 'UselessPad' doesn't have any unwind |
498 | // information that we'd be contradicting by making a map entry for it |
499 | // (which is something that getUnwindDestTokenHelper must have proved for |
500 | // us to get here). Just assert on is direct users here; the checks in |
501 | // this downward walk at its descendants will verify that they don't have |
502 | // any unwind edges that exit 'UselessPad' either (i.e. they either have no |
503 | // unwind edges or unwind to a sibling). |
504 | MemoMap[UselessPad] = UnwindDestToken; |
505 | if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Val: UselessPad)) { |
506 | assert(CatchSwitch->getUnwindDest() == nullptr && "Expected useless pad" ); |
507 | for (BasicBlock *HandlerBlock : CatchSwitch->handlers()) { |
508 | auto *CatchPad = HandlerBlock->getFirstNonPHI(); |
509 | for (User *U : CatchPad->users()) { |
510 | assert( |
511 | (!isa<InvokeInst>(U) || |
512 | (getParentPad( |
513 | cast<InvokeInst>(U)->getUnwindDest()->getFirstNonPHI()) == |
514 | CatchPad)) && |
515 | "Expected useless pad" ); |
516 | if (isa<CatchSwitchInst>(Val: U) || isa<CleanupPadInst>(Val: U)) |
517 | Worklist.push_back(Elt: cast<Instruction>(Val: U)); |
518 | } |
519 | } |
520 | } else { |
521 | assert(isa<CleanupPadInst>(UselessPad)); |
522 | for (User *U : UselessPad->users()) { |
523 | assert(!isa<CleanupReturnInst>(U) && "Expected useless pad" ); |
524 | assert((!isa<InvokeInst>(U) || |
525 | (getParentPad( |
526 | cast<InvokeInst>(U)->getUnwindDest()->getFirstNonPHI()) == |
527 | UselessPad)) && |
528 | "Expected useless pad" ); |
529 | if (isa<CatchSwitchInst>(Val: U) || isa<CleanupPadInst>(Val: U)) |
530 | Worklist.push_back(Elt: cast<Instruction>(Val: U)); |
531 | } |
532 | } |
533 | } |
534 | |
535 | return UnwindDestToken; |
536 | } |
537 | |
538 | /// When we inline a basic block into an invoke, |
539 | /// we have to turn all of the calls that can throw into invokes. |
540 | /// This function analyze BB to see if there are any calls, and if so, |
541 | /// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI |
542 | /// nodes in that block with the values specified in InvokeDestPHIValues. |
543 | static BasicBlock *HandleCallsInBlockInlinedThroughInvoke( |
544 | BasicBlock *BB, BasicBlock *UnwindEdge, |
545 | UnwindDestMemoTy *FuncletUnwindMap = nullptr) { |
546 | for (Instruction &I : llvm::make_early_inc_range(Range&: *BB)) { |
547 | // We only need to check for function calls: inlined invoke |
548 | // instructions require no special handling. |
549 | CallInst *CI = dyn_cast<CallInst>(Val: &I); |
550 | |
551 | if (!CI || CI->doesNotThrow()) |
552 | continue; |
553 | |
554 | // We do not need to (and in fact, cannot) convert possibly throwing calls |
555 | // to @llvm.experimental_deoptimize (resp. @llvm.experimental.guard) into |
556 | // invokes. The caller's "segment" of the deoptimization continuation |
557 | // attached to the newly inlined @llvm.experimental_deoptimize |
558 | // (resp. @llvm.experimental.guard) call should contain the exception |
559 | // handling logic, if any. |
560 | if (auto *F = CI->getCalledFunction()) |
561 | if (F->getIntrinsicID() == Intrinsic::experimental_deoptimize || |
562 | F->getIntrinsicID() == Intrinsic::experimental_guard) |
563 | continue; |
564 | |
565 | if (auto FuncletBundle = CI->getOperandBundle(ID: LLVMContext::OB_funclet)) { |
566 | // This call is nested inside a funclet. If that funclet has an unwind |
567 | // destination within the inlinee, then unwinding out of this call would |
568 | // be UB. Rewriting this call to an invoke which targets the inlined |
569 | // invoke's unwind dest would give the call's parent funclet multiple |
570 | // unwind destinations, which is something that subsequent EH table |
571 | // generation can't handle and that the veirifer rejects. So when we |
572 | // see such a call, leave it as a call. |
573 | auto *FuncletPad = cast<Instruction>(Val: FuncletBundle->Inputs[0]); |
574 | Value *UnwindDestToken = |
575 | getUnwindDestToken(EHPad: FuncletPad, MemoMap&: *FuncletUnwindMap); |
576 | if (UnwindDestToken && !isa<ConstantTokenNone>(Val: UnwindDestToken)) |
577 | continue; |
578 | #ifndef NDEBUG |
579 | Instruction *MemoKey; |
580 | if (auto *CatchPad = dyn_cast<CatchPadInst>(FuncletPad)) |
581 | MemoKey = CatchPad->getCatchSwitch(); |
582 | else |
583 | MemoKey = FuncletPad; |
584 | assert(FuncletUnwindMap->count(MemoKey) && |
585 | (*FuncletUnwindMap)[MemoKey] == UnwindDestToken && |
586 | "must get memoized to avoid confusing later searches" ); |
587 | #endif // NDEBUG |
588 | } |
589 | |
590 | changeToInvokeAndSplitBasicBlock(CI, UnwindEdge); |
591 | return BB; |
592 | } |
593 | return nullptr; |
594 | } |
595 | |
596 | /// If we inlined an invoke site, we need to convert calls |
597 | /// in the body of the inlined function into invokes. |
598 | /// |
599 | /// II is the invoke instruction being inlined. FirstNewBlock is the first |
600 | /// block of the inlined code (the last block is the end of the function), |
601 | /// and InlineCodeInfo is information about the code that got inlined. |
602 | static void HandleInlinedLandingPad(InvokeInst *II, BasicBlock *FirstNewBlock, |
603 | ClonedCodeInfo &InlinedCodeInfo) { |
604 | BasicBlock *InvokeDest = II->getUnwindDest(); |
605 | |
606 | Function *Caller = FirstNewBlock->getParent(); |
607 | |
608 | // The inlined code is currently at the end of the function, scan from the |
609 | // start of the inlined code to its end, checking for stuff we need to |
610 | // rewrite. |
611 | LandingPadInliningInfo Invoke(II); |
612 | |
613 | // Get all of the inlined landing pad instructions. |
614 | SmallPtrSet<LandingPadInst*, 16> InlinedLPads; |
615 | for (Function::iterator I = FirstNewBlock->getIterator(), E = Caller->end(); |
616 | I != E; ++I) |
617 | if (InvokeInst *II = dyn_cast<InvokeInst>(Val: I->getTerminator())) |
618 | InlinedLPads.insert(Ptr: II->getLandingPadInst()); |
619 | |
620 | // Append the clauses from the outer landing pad instruction into the inlined |
621 | // landing pad instructions. |
622 | LandingPadInst *OuterLPad = Invoke.getLandingPadInst(); |
623 | for (LandingPadInst *InlinedLPad : InlinedLPads) { |
624 | unsigned OuterNum = OuterLPad->getNumClauses(); |
625 | InlinedLPad->reserveClauses(Size: OuterNum); |
626 | for (unsigned OuterIdx = 0; OuterIdx != OuterNum; ++OuterIdx) |
627 | InlinedLPad->addClause(ClauseVal: OuterLPad->getClause(Idx: OuterIdx)); |
628 | if (OuterLPad->isCleanup()) |
629 | InlinedLPad->setCleanup(true); |
630 | } |
631 | |
632 | for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end(); |
633 | BB != E; ++BB) { |
634 | if (InlinedCodeInfo.ContainsCalls) |
635 | if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke( |
636 | BB: &*BB, UnwindEdge: Invoke.getOuterResumeDest())) |
637 | // Update any PHI nodes in the exceptional block to indicate that there |
638 | // is now a new entry in them. |
639 | Invoke.addIncomingPHIValuesFor(BB: NewBB); |
640 | |
641 | // Forward any resumes that are remaining here. |
642 | if (ResumeInst *RI = dyn_cast<ResumeInst>(Val: BB->getTerminator())) |
643 | Invoke.forwardResume(RI, InlinedLPads); |
644 | } |
645 | |
646 | // Now that everything is happy, we have one final detail. The PHI nodes in |
647 | // the exception destination block still have entries due to the original |
648 | // invoke instruction. Eliminate these entries (which might even delete the |
649 | // PHI node) now. |
650 | InvokeDest->removePredecessor(Pred: II->getParent()); |
651 | } |
652 | |
653 | /// If we inlined an invoke site, we need to convert calls |
654 | /// in the body of the inlined function into invokes. |
655 | /// |
656 | /// II is the invoke instruction being inlined. FirstNewBlock is the first |
657 | /// block of the inlined code (the last block is the end of the function), |
658 | /// and InlineCodeInfo is information about the code that got inlined. |
659 | static void HandleInlinedEHPad(InvokeInst *II, BasicBlock *FirstNewBlock, |
660 | ClonedCodeInfo &InlinedCodeInfo) { |
661 | BasicBlock *UnwindDest = II->getUnwindDest(); |
662 | Function *Caller = FirstNewBlock->getParent(); |
663 | |
664 | assert(UnwindDest->getFirstNonPHI()->isEHPad() && "unexpected BasicBlock!" ); |
665 | |
666 | // If there are PHI nodes in the unwind destination block, we need to keep |
667 | // track of which values came into them from the invoke before removing the |
668 | // edge from this block. |
669 | SmallVector<Value *, 8> UnwindDestPHIValues; |
670 | BasicBlock *InvokeBB = II->getParent(); |
671 | for (PHINode &PHI : UnwindDest->phis()) { |
672 | // Save the value to use for this edge. |
673 | UnwindDestPHIValues.push_back(Elt: PHI.getIncomingValueForBlock(BB: InvokeBB)); |
674 | } |
675 | |
676 | // Add incoming-PHI values to the unwind destination block for the given basic |
677 | // block, using the values for the original invoke's source block. |
678 | auto UpdatePHINodes = [&](BasicBlock *Src) { |
679 | BasicBlock::iterator I = UnwindDest->begin(); |
680 | for (Value *V : UnwindDestPHIValues) { |
681 | PHINode *PHI = cast<PHINode>(Val&: I); |
682 | PHI->addIncoming(V, BB: Src); |
683 | ++I; |
684 | } |
685 | }; |
686 | |
687 | // This connects all the instructions which 'unwind to caller' to the invoke |
688 | // destination. |
689 | UnwindDestMemoTy FuncletUnwindMap; |
690 | for (Function::iterator BB = FirstNewBlock->getIterator(), E = Caller->end(); |
691 | BB != E; ++BB) { |
692 | if (auto *CRI = dyn_cast<CleanupReturnInst>(Val: BB->getTerminator())) { |
693 | if (CRI->unwindsToCaller()) { |
694 | auto *CleanupPad = CRI->getCleanupPad(); |
695 | CleanupReturnInst::Create(CleanupPad, UnwindBB: UnwindDest, InsertBefore: CRI->getIterator()); |
696 | CRI->eraseFromParent(); |
697 | UpdatePHINodes(&*BB); |
698 | // Finding a cleanupret with an unwind destination would confuse |
699 | // subsequent calls to getUnwindDestToken, so map the cleanuppad |
700 | // to short-circuit any such calls and recognize this as an "unwind |
701 | // to caller" cleanup. |
702 | assert(!FuncletUnwindMap.count(CleanupPad) || |
703 | isa<ConstantTokenNone>(FuncletUnwindMap[CleanupPad])); |
704 | FuncletUnwindMap[CleanupPad] = |
705 | ConstantTokenNone::get(Context&: Caller->getContext()); |
706 | } |
707 | } |
708 | |
709 | Instruction *I = BB->getFirstNonPHI(); |
710 | if (!I->isEHPad()) |
711 | continue; |
712 | |
713 | Instruction *Replacement = nullptr; |
714 | if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Val: I)) { |
715 | if (CatchSwitch->unwindsToCaller()) { |
716 | Value *UnwindDestToken; |
717 | if (auto *ParentPad = |
718 | dyn_cast<Instruction>(Val: CatchSwitch->getParentPad())) { |
719 | // This catchswitch is nested inside another funclet. If that |
720 | // funclet has an unwind destination within the inlinee, then |
721 | // unwinding out of this catchswitch would be UB. Rewriting this |
722 | // catchswitch to unwind to the inlined invoke's unwind dest would |
723 | // give the parent funclet multiple unwind destinations, which is |
724 | // something that subsequent EH table generation can't handle and |
725 | // that the veirifer rejects. So when we see such a call, leave it |
726 | // as "unwind to caller". |
727 | UnwindDestToken = getUnwindDestToken(EHPad: ParentPad, MemoMap&: FuncletUnwindMap); |
728 | if (UnwindDestToken && !isa<ConstantTokenNone>(Val: UnwindDestToken)) |
729 | continue; |
730 | } else { |
731 | // This catchswitch has no parent to inherit constraints from, and |
732 | // none of its descendants can have an unwind edge that exits it and |
733 | // targets another funclet in the inlinee. It may or may not have a |
734 | // descendant that definitively has an unwind to caller. In either |
735 | // case, we'll have to assume that any unwinds out of it may need to |
736 | // be routed to the caller, so treat it as though it has a definitive |
737 | // unwind to caller. |
738 | UnwindDestToken = ConstantTokenNone::get(Context&: Caller->getContext()); |
739 | } |
740 | auto *NewCatchSwitch = CatchSwitchInst::Create( |
741 | ParentPad: CatchSwitch->getParentPad(), UnwindDest, |
742 | NumHandlers: CatchSwitch->getNumHandlers(), NameStr: CatchSwitch->getName(), |
743 | InsertBefore: CatchSwitch->getIterator()); |
744 | for (BasicBlock *PadBB : CatchSwitch->handlers()) |
745 | NewCatchSwitch->addHandler(Dest: PadBB); |
746 | // Propagate info for the old catchswitch over to the new one in |
747 | // the unwind map. This also serves to short-circuit any subsequent |
748 | // checks for the unwind dest of this catchswitch, which would get |
749 | // confused if they found the outer handler in the callee. |
750 | FuncletUnwindMap[NewCatchSwitch] = UnwindDestToken; |
751 | Replacement = NewCatchSwitch; |
752 | } |
753 | } else if (!isa<FuncletPadInst>(Val: I)) { |
754 | llvm_unreachable("unexpected EHPad!" ); |
755 | } |
756 | |
757 | if (Replacement) { |
758 | Replacement->takeName(V: I); |
759 | I->replaceAllUsesWith(V: Replacement); |
760 | I->eraseFromParent(); |
761 | UpdatePHINodes(&*BB); |
762 | } |
763 | } |
764 | |
765 | if (InlinedCodeInfo.ContainsCalls) |
766 | for (Function::iterator BB = FirstNewBlock->getIterator(), |
767 | E = Caller->end(); |
768 | BB != E; ++BB) |
769 | if (BasicBlock *NewBB = HandleCallsInBlockInlinedThroughInvoke( |
770 | BB: &*BB, UnwindEdge: UnwindDest, FuncletUnwindMap: &FuncletUnwindMap)) |
771 | // Update any PHI nodes in the exceptional block to indicate that there |
772 | // is now a new entry in them. |
773 | UpdatePHINodes(NewBB); |
774 | |
775 | // Now that everything is happy, we have one final detail. The PHI nodes in |
776 | // the exception destination block still have entries due to the original |
777 | // invoke instruction. Eliminate these entries (which might even delete the |
778 | // PHI node) now. |
779 | UnwindDest->removePredecessor(Pred: InvokeBB); |
780 | } |
781 | |
782 | static bool haveCommonPrefix(MDNode *MIBStackContext, |
783 | MDNode *CallsiteStackContext) { |
784 | assert(MIBStackContext->getNumOperands() > 0 && |
785 | CallsiteStackContext->getNumOperands() > 0); |
786 | // Because of the context trimming performed during matching, the callsite |
787 | // context could have more stack ids than the MIB. We match up to the end of |
788 | // the shortest stack context. |
789 | for (auto MIBStackIter = MIBStackContext->op_begin(), |
790 | CallsiteStackIter = CallsiteStackContext->op_begin(); |
791 | MIBStackIter != MIBStackContext->op_end() && |
792 | CallsiteStackIter != CallsiteStackContext->op_end(); |
793 | MIBStackIter++, CallsiteStackIter++) { |
794 | auto *Val1 = mdconst::dyn_extract<ConstantInt>(MD: *MIBStackIter); |
795 | auto *Val2 = mdconst::dyn_extract<ConstantInt>(MD: *CallsiteStackIter); |
796 | assert(Val1 && Val2); |
797 | if (Val1->getZExtValue() != Val2->getZExtValue()) |
798 | return false; |
799 | } |
800 | return true; |
801 | } |
802 | |
803 | static void removeMemProfMetadata(CallBase *Call) { |
804 | Call->setMetadata(KindID: LLVMContext::MD_memprof, Node: nullptr); |
805 | } |
806 | |
807 | static void removeCallsiteMetadata(CallBase *Call) { |
808 | Call->setMetadata(KindID: LLVMContext::MD_callsite, Node: nullptr); |
809 | } |
810 | |
811 | static void updateMemprofMetadata(CallBase *CI, |
812 | const std::vector<Metadata *> &MIBList) { |
813 | assert(!MIBList.empty()); |
814 | // Remove existing memprof, which will either be replaced or may not be needed |
815 | // if we are able to use a single allocation type function attribute. |
816 | removeMemProfMetadata(Call: CI); |
817 | CallStackTrie CallStack; |
818 | for (Metadata *MIB : MIBList) |
819 | CallStack.addCallStack(MIB: cast<MDNode>(Val: MIB)); |
820 | bool MemprofMDAttached = CallStack.buildAndAttachMIBMetadata(CI); |
821 | assert(MemprofMDAttached == CI->hasMetadata(LLVMContext::MD_memprof)); |
822 | if (!MemprofMDAttached) |
823 | // If we used a function attribute remove the callsite metadata as well. |
824 | removeCallsiteMetadata(Call: CI); |
825 | } |
826 | |
827 | // Update the metadata on the inlined copy ClonedCall of a call OrigCall in the |
828 | // inlined callee body, based on the callsite metadata InlinedCallsiteMD from |
829 | // the call that was inlined. |
830 | static void propagateMemProfHelper(const CallBase *OrigCall, |
831 | CallBase *ClonedCall, |
832 | MDNode *InlinedCallsiteMD) { |
833 | MDNode *OrigCallsiteMD = ClonedCall->getMetadata(KindID: LLVMContext::MD_callsite); |
834 | MDNode *ClonedCallsiteMD = nullptr; |
835 | // Check if the call originally had callsite metadata, and update it for the |
836 | // new call in the inlined body. |
837 | if (OrigCallsiteMD) { |
838 | // The cloned call's context is now the concatenation of the original call's |
839 | // callsite metadata and the callsite metadata on the call where it was |
840 | // inlined. |
841 | ClonedCallsiteMD = MDNode::concatenate(A: OrigCallsiteMD, B: InlinedCallsiteMD); |
842 | ClonedCall->setMetadata(KindID: LLVMContext::MD_callsite, Node: ClonedCallsiteMD); |
843 | } |
844 | |
845 | // Update any memprof metadata on the cloned call. |
846 | MDNode *OrigMemProfMD = ClonedCall->getMetadata(KindID: LLVMContext::MD_memprof); |
847 | if (!OrigMemProfMD) |
848 | return; |
849 | // We currently expect that allocations with memprof metadata also have |
850 | // callsite metadata for the allocation's part of the context. |
851 | assert(OrigCallsiteMD); |
852 | |
853 | // New call's MIB list. |
854 | std::vector<Metadata *> NewMIBList; |
855 | |
856 | // For each MIB metadata, check if its call stack context starts with the |
857 | // new clone's callsite metadata. If so, that MIB goes onto the cloned call in |
858 | // the inlined body. If not, it stays on the out-of-line original call. |
859 | for (auto &MIBOp : OrigMemProfMD->operands()) { |
860 | MDNode *MIB = dyn_cast<MDNode>(Val: MIBOp); |
861 | // Stack is first operand of MIB. |
862 | MDNode *StackMD = getMIBStackNode(MIB); |
863 | assert(StackMD); |
864 | // See if the new cloned callsite context matches this profiled context. |
865 | if (haveCommonPrefix(MIBStackContext: StackMD, CallsiteStackContext: ClonedCallsiteMD)) |
866 | // Add it to the cloned call's MIB list. |
867 | NewMIBList.push_back(x: MIB); |
868 | } |
869 | if (NewMIBList.empty()) { |
870 | removeMemProfMetadata(Call: ClonedCall); |
871 | removeCallsiteMetadata(Call: ClonedCall); |
872 | return; |
873 | } |
874 | if (NewMIBList.size() < OrigMemProfMD->getNumOperands()) |
875 | updateMemprofMetadata(CI: ClonedCall, MIBList: NewMIBList); |
876 | } |
877 | |
878 | // Update memprof related metadata (!memprof and !callsite) based on the |
879 | // inlining of Callee into the callsite at CB. The updates include merging the |
880 | // inlined callee's callsite metadata with that of the inlined call, |
881 | // and moving the subset of any memprof contexts to the inlined callee |
882 | // allocations if they match the new inlined call stack. |
883 | static void |
884 | propagateMemProfMetadata(Function *Callee, CallBase &CB, |
885 | bool ContainsMemProfMetadata, |
886 | const ValueMap<const Value *, WeakTrackingVH> &VMap) { |
887 | MDNode *CallsiteMD = CB.getMetadata(KindID: LLVMContext::MD_callsite); |
888 | // Only need to update if the inlined callsite had callsite metadata, or if |
889 | // there was any memprof metadata inlined. |
890 | if (!CallsiteMD && !ContainsMemProfMetadata) |
891 | return; |
892 | |
893 | // Propagate metadata onto the cloned calls in the inlined callee. |
894 | for (const auto &Entry : VMap) { |
895 | // See if this is a call that has been inlined and remapped, and not |
896 | // simplified away in the process. |
897 | auto *OrigCall = dyn_cast_or_null<CallBase>(Val: Entry.first); |
898 | auto *ClonedCall = dyn_cast_or_null<CallBase>(Val: Entry.second); |
899 | if (!OrigCall || !ClonedCall) |
900 | continue; |
901 | // If the inlined callsite did not have any callsite metadata, then it isn't |
902 | // involved in any profiled call contexts, and we can remove any memprof |
903 | // metadata on the cloned call. |
904 | if (!CallsiteMD) { |
905 | removeMemProfMetadata(Call: ClonedCall); |
906 | removeCallsiteMetadata(Call: ClonedCall); |
907 | continue; |
908 | } |
909 | propagateMemProfHelper(OrigCall, ClonedCall, InlinedCallsiteMD: CallsiteMD); |
910 | } |
911 | } |
912 | |
913 | /// When inlining a call site that has !llvm.mem.parallel_loop_access, |
914 | /// !llvm.access.group, !alias.scope or !noalias metadata, that metadata should |
915 | /// be propagated to all memory-accessing cloned instructions. |
916 | static void PropagateCallSiteMetadata(CallBase &CB, Function::iterator FStart, |
917 | Function::iterator FEnd) { |
918 | MDNode *MemParallelLoopAccess = |
919 | CB.getMetadata(KindID: LLVMContext::MD_mem_parallel_loop_access); |
920 | MDNode *AccessGroup = CB.getMetadata(KindID: LLVMContext::MD_access_group); |
921 | MDNode *AliasScope = CB.getMetadata(KindID: LLVMContext::MD_alias_scope); |
922 | MDNode *NoAlias = CB.getMetadata(KindID: LLVMContext::MD_noalias); |
923 | if (!MemParallelLoopAccess && !AccessGroup && !AliasScope && !NoAlias) |
924 | return; |
925 | |
926 | for (BasicBlock &BB : make_range(x: FStart, y: FEnd)) { |
927 | for (Instruction &I : BB) { |
928 | // This metadata is only relevant for instructions that access memory. |
929 | if (!I.mayReadOrWriteMemory()) |
930 | continue; |
931 | |
932 | if (MemParallelLoopAccess) { |
933 | // TODO: This probably should not overwrite MemParalleLoopAccess. |
934 | MemParallelLoopAccess = MDNode::concatenate( |
935 | A: I.getMetadata(KindID: LLVMContext::MD_mem_parallel_loop_access), |
936 | B: MemParallelLoopAccess); |
937 | I.setMetadata(KindID: LLVMContext::MD_mem_parallel_loop_access, |
938 | Node: MemParallelLoopAccess); |
939 | } |
940 | |
941 | if (AccessGroup) |
942 | I.setMetadata(KindID: LLVMContext::MD_access_group, Node: uniteAccessGroups( |
943 | AccGroups1: I.getMetadata(KindID: LLVMContext::MD_access_group), AccGroups2: AccessGroup)); |
944 | |
945 | if (AliasScope) |
946 | I.setMetadata(KindID: LLVMContext::MD_alias_scope, Node: MDNode::concatenate( |
947 | A: I.getMetadata(KindID: LLVMContext::MD_alias_scope), B: AliasScope)); |
948 | |
949 | if (NoAlias) |
950 | I.setMetadata(KindID: LLVMContext::MD_noalias, Node: MDNode::concatenate( |
951 | A: I.getMetadata(KindID: LLVMContext::MD_noalias), B: NoAlias)); |
952 | } |
953 | } |
954 | } |
955 | |
956 | /// Bundle operands of the inlined function must be added to inlined call sites. |
957 | static void PropagateOperandBundles(Function::iterator InlinedBB, |
958 | Instruction *CallSiteEHPad) { |
959 | for (Instruction &II : llvm::make_early_inc_range(Range&: *InlinedBB)) { |
960 | CallBase *I = dyn_cast<CallBase>(Val: &II); |
961 | if (!I) |
962 | continue; |
963 | // Skip call sites which already have a "funclet" bundle. |
964 | if (I->getOperandBundle(ID: LLVMContext::OB_funclet)) |
965 | continue; |
966 | // Skip call sites which are nounwind intrinsics (as long as they don't |
967 | // lower into regular function calls in the course of IR transformations). |
968 | auto *CalledFn = |
969 | dyn_cast<Function>(Val: I->getCalledOperand()->stripPointerCasts()); |
970 | if (CalledFn && CalledFn->isIntrinsic() && I->doesNotThrow() && |
971 | !IntrinsicInst::mayLowerToFunctionCall(IID: CalledFn->getIntrinsicID())) |
972 | continue; |
973 | |
974 | SmallVector<OperandBundleDef, 1> OpBundles; |
975 | I->getOperandBundlesAsDefs(Defs&: OpBundles); |
976 | OpBundles.emplace_back(Args: "funclet" , Args&: CallSiteEHPad); |
977 | |
978 | Instruction *NewInst = CallBase::Create(CB: I, Bundles: OpBundles, InsertPt: I->getIterator()); |
979 | NewInst->takeName(V: I); |
980 | I->replaceAllUsesWith(V: NewInst); |
981 | I->eraseFromParent(); |
982 | } |
983 | } |
984 | |
985 | namespace { |
986 | /// Utility for cloning !noalias and !alias.scope metadata. When a code region |
987 | /// using scoped alias metadata is inlined, the aliasing relationships may not |
988 | /// hold between the two version. It is necessary to create a deep clone of the |
989 | /// metadata, putting the two versions in separate scope domains. |
990 | class ScopedAliasMetadataDeepCloner { |
991 | using MetadataMap = DenseMap<const MDNode *, TrackingMDNodeRef>; |
992 | SetVector<const MDNode *> MD; |
993 | MetadataMap MDMap; |
994 | void addRecursiveMetadataUses(); |
995 | |
996 | public: |
997 | ScopedAliasMetadataDeepCloner(const Function *F); |
998 | |
999 | /// Create a new clone of the scoped alias metadata, which will be used by |
1000 | /// subsequent remap() calls. |
1001 | void clone(); |
1002 | |
1003 | /// Remap instructions in the given range from the original to the cloned |
1004 | /// metadata. |
1005 | void remap(Function::iterator FStart, Function::iterator FEnd); |
1006 | }; |
1007 | } // namespace |
1008 | |
1009 | ScopedAliasMetadataDeepCloner::ScopedAliasMetadataDeepCloner( |
1010 | const Function *F) { |
1011 | for (const BasicBlock &BB : *F) { |
1012 | for (const Instruction &I : BB) { |
1013 | if (const MDNode *M = I.getMetadata(KindID: LLVMContext::MD_alias_scope)) |
1014 | MD.insert(X: M); |
1015 | if (const MDNode *M = I.getMetadata(KindID: LLVMContext::MD_noalias)) |
1016 | MD.insert(X: M); |
1017 | |
1018 | // We also need to clone the metadata in noalias intrinsics. |
1019 | if (const auto *Decl = dyn_cast<NoAliasScopeDeclInst>(Val: &I)) |
1020 | MD.insert(X: Decl->getScopeList()); |
1021 | } |
1022 | } |
1023 | addRecursiveMetadataUses(); |
1024 | } |
1025 | |
1026 | void ScopedAliasMetadataDeepCloner::addRecursiveMetadataUses() { |
1027 | SmallVector<const Metadata *, 16> Queue(MD.begin(), MD.end()); |
1028 | while (!Queue.empty()) { |
1029 | const MDNode *M = cast<MDNode>(Val: Queue.pop_back_val()); |
1030 | for (const Metadata *Op : M->operands()) |
1031 | if (const MDNode *OpMD = dyn_cast<MDNode>(Val: Op)) |
1032 | if (MD.insert(X: OpMD)) |
1033 | Queue.push_back(Elt: OpMD); |
1034 | } |
1035 | } |
1036 | |
1037 | void ScopedAliasMetadataDeepCloner::clone() { |
1038 | assert(MDMap.empty() && "clone() already called ?" ); |
1039 | |
1040 | SmallVector<TempMDTuple, 16> DummyNodes; |
1041 | for (const MDNode *I : MD) { |
1042 | DummyNodes.push_back(Elt: MDTuple::getTemporary(Context&: I->getContext(), MDs: std::nullopt)); |
1043 | MDMap[I].reset(MD: DummyNodes.back().get()); |
1044 | } |
1045 | |
1046 | // Create new metadata nodes to replace the dummy nodes, replacing old |
1047 | // metadata references with either a dummy node or an already-created new |
1048 | // node. |
1049 | SmallVector<Metadata *, 4> NewOps; |
1050 | for (const MDNode *I : MD) { |
1051 | for (const Metadata *Op : I->operands()) { |
1052 | if (const MDNode *M = dyn_cast<MDNode>(Val: Op)) |
1053 | NewOps.push_back(Elt: MDMap[M]); |
1054 | else |
1055 | NewOps.push_back(Elt: const_cast<Metadata *>(Op)); |
1056 | } |
1057 | |
1058 | MDNode *NewM = MDNode::get(Context&: I->getContext(), MDs: NewOps); |
1059 | MDTuple *TempM = cast<MDTuple>(Val&: MDMap[I]); |
1060 | assert(TempM->isTemporary() && "Expected temporary node" ); |
1061 | |
1062 | TempM->replaceAllUsesWith(MD: NewM); |
1063 | NewOps.clear(); |
1064 | } |
1065 | } |
1066 | |
1067 | void ScopedAliasMetadataDeepCloner::remap(Function::iterator FStart, |
1068 | Function::iterator FEnd) { |
1069 | if (MDMap.empty()) |
1070 | return; // Nothing to do. |
1071 | |
1072 | for (BasicBlock &BB : make_range(x: FStart, y: FEnd)) { |
1073 | for (Instruction &I : BB) { |
1074 | // TODO: The null checks for the MDMap.lookup() results should no longer |
1075 | // be necessary. |
1076 | if (MDNode *M = I.getMetadata(KindID: LLVMContext::MD_alias_scope)) |
1077 | if (MDNode *MNew = MDMap.lookup(Val: M)) |
1078 | I.setMetadata(KindID: LLVMContext::MD_alias_scope, Node: MNew); |
1079 | |
1080 | if (MDNode *M = I.getMetadata(KindID: LLVMContext::MD_noalias)) |
1081 | if (MDNode *MNew = MDMap.lookup(Val: M)) |
1082 | I.setMetadata(KindID: LLVMContext::MD_noalias, Node: MNew); |
1083 | |
1084 | if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(Val: &I)) |
1085 | if (MDNode *MNew = MDMap.lookup(Val: Decl->getScopeList())) |
1086 | Decl->setScopeList(MNew); |
1087 | } |
1088 | } |
1089 | } |
1090 | |
1091 | /// If the inlined function has noalias arguments, |
1092 | /// then add new alias scopes for each noalias argument, tag the mapped noalias |
1093 | /// parameters with noalias metadata specifying the new scope, and tag all |
1094 | /// non-derived loads, stores and memory intrinsics with the new alias scopes. |
1095 | static void AddAliasScopeMetadata(CallBase &CB, ValueToValueMapTy &VMap, |
1096 | const DataLayout &DL, AAResults *CalleeAAR, |
1097 | ClonedCodeInfo &InlinedFunctionInfo) { |
1098 | if (!EnableNoAliasConversion) |
1099 | return; |
1100 | |
1101 | const Function *CalledFunc = CB.getCalledFunction(); |
1102 | SmallVector<const Argument *, 4> NoAliasArgs; |
1103 | |
1104 | for (const Argument &Arg : CalledFunc->args()) |
1105 | if (CB.paramHasAttr(ArgNo: Arg.getArgNo(), Kind: Attribute::NoAlias) && !Arg.use_empty()) |
1106 | NoAliasArgs.push_back(Elt: &Arg); |
1107 | |
1108 | if (NoAliasArgs.empty()) |
1109 | return; |
1110 | |
1111 | // To do a good job, if a noalias variable is captured, we need to know if |
1112 | // the capture point dominates the particular use we're considering. |
1113 | DominatorTree DT; |
1114 | DT.recalculate(Func&: const_cast<Function&>(*CalledFunc)); |
1115 | |
1116 | // noalias indicates that pointer values based on the argument do not alias |
1117 | // pointer values which are not based on it. So we add a new "scope" for each |
1118 | // noalias function argument. Accesses using pointers based on that argument |
1119 | // become part of that alias scope, accesses using pointers not based on that |
1120 | // argument are tagged as noalias with that scope. |
1121 | |
1122 | DenseMap<const Argument *, MDNode *> NewScopes; |
1123 | MDBuilder MDB(CalledFunc->getContext()); |
1124 | |
1125 | // Create a new scope domain for this function. |
1126 | MDNode *NewDomain = |
1127 | MDB.createAnonymousAliasScopeDomain(Name: CalledFunc->getName()); |
1128 | for (unsigned i = 0, e = NoAliasArgs.size(); i != e; ++i) { |
1129 | const Argument *A = NoAliasArgs[i]; |
1130 | |
1131 | std::string Name = std::string(CalledFunc->getName()); |
1132 | if (A->hasName()) { |
1133 | Name += ": %" ; |
1134 | Name += A->getName(); |
1135 | } else { |
1136 | Name += ": argument " ; |
1137 | Name += utostr(X: i); |
1138 | } |
1139 | |
1140 | // Note: We always create a new anonymous root here. This is true regardless |
1141 | // of the linkage of the callee because the aliasing "scope" is not just a |
1142 | // property of the callee, but also all control dependencies in the caller. |
1143 | MDNode *NewScope = MDB.createAnonymousAliasScope(Domain: NewDomain, Name); |
1144 | NewScopes.insert(KV: std::make_pair(x&: A, y&: NewScope)); |
1145 | |
1146 | if (UseNoAliasIntrinsic) { |
1147 | // Introduce a llvm.experimental.noalias.scope.decl for the noalias |
1148 | // argument. |
1149 | MDNode *AScopeList = MDNode::get(Context&: CalledFunc->getContext(), MDs: NewScope); |
1150 | auto *NoAliasDecl = |
1151 | IRBuilder<>(&CB).CreateNoAliasScopeDeclaration(ScopeTag: AScopeList); |
1152 | // Ignore the result for now. The result will be used when the |
1153 | // llvm.noalias intrinsic is introduced. |
1154 | (void)NoAliasDecl; |
1155 | } |
1156 | } |
1157 | |
1158 | // Iterate over all new instructions in the map; for all memory-access |
1159 | // instructions, add the alias scope metadata. |
1160 | for (ValueToValueMapTy::iterator VMI = VMap.begin(), VMIE = VMap.end(); |
1161 | VMI != VMIE; ++VMI) { |
1162 | if (const Instruction *I = dyn_cast<Instruction>(Val: VMI->first)) { |
1163 | if (!VMI->second) |
1164 | continue; |
1165 | |
1166 | Instruction *NI = dyn_cast<Instruction>(Val&: VMI->second); |
1167 | if (!NI || InlinedFunctionInfo.isSimplified(From: I, To: NI)) |
1168 | continue; |
1169 | |
1170 | bool IsArgMemOnlyCall = false, IsFuncCall = false; |
1171 | SmallVector<const Value *, 2> PtrArgs; |
1172 | |
1173 | if (const LoadInst *LI = dyn_cast<LoadInst>(Val: I)) |
1174 | PtrArgs.push_back(Elt: LI->getPointerOperand()); |
1175 | else if (const StoreInst *SI = dyn_cast<StoreInst>(Val: I)) |
1176 | PtrArgs.push_back(Elt: SI->getPointerOperand()); |
1177 | else if (const VAArgInst *VAAI = dyn_cast<VAArgInst>(Val: I)) |
1178 | PtrArgs.push_back(Elt: VAAI->getPointerOperand()); |
1179 | else if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(Val: I)) |
1180 | PtrArgs.push_back(Elt: CXI->getPointerOperand()); |
1181 | else if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(Val: I)) |
1182 | PtrArgs.push_back(Elt: RMWI->getPointerOperand()); |
1183 | else if (const auto *Call = dyn_cast<CallBase>(Val: I)) { |
1184 | // If we know that the call does not access memory, then we'll still |
1185 | // know that about the inlined clone of this call site, and we don't |
1186 | // need to add metadata. |
1187 | if (Call->doesNotAccessMemory()) |
1188 | continue; |
1189 | |
1190 | IsFuncCall = true; |
1191 | if (CalleeAAR) { |
1192 | MemoryEffects ME = CalleeAAR->getMemoryEffects(Call); |
1193 | |
1194 | // We'll retain this knowledge without additional metadata. |
1195 | if (ME.onlyAccessesInaccessibleMem()) |
1196 | continue; |
1197 | |
1198 | if (ME.onlyAccessesArgPointees()) |
1199 | IsArgMemOnlyCall = true; |
1200 | } |
1201 | |
1202 | for (Value *Arg : Call->args()) { |
1203 | // Only care about pointer arguments. If a noalias argument is |
1204 | // accessed through a non-pointer argument, it must be captured |
1205 | // first (e.g. via ptrtoint), and we protect against captures below. |
1206 | if (!Arg->getType()->isPointerTy()) |
1207 | continue; |
1208 | |
1209 | PtrArgs.push_back(Elt: Arg); |
1210 | } |
1211 | } |
1212 | |
1213 | // If we found no pointers, then this instruction is not suitable for |
1214 | // pairing with an instruction to receive aliasing metadata. |
1215 | // However, if this is a call, this we might just alias with none of the |
1216 | // noalias arguments. |
1217 | if (PtrArgs.empty() && !IsFuncCall) |
1218 | continue; |
1219 | |
1220 | // It is possible that there is only one underlying object, but you |
1221 | // need to go through several PHIs to see it, and thus could be |
1222 | // repeated in the Objects list. |
1223 | SmallPtrSet<const Value *, 4> ObjSet; |
1224 | SmallVector<Metadata *, 4> Scopes, NoAliases; |
1225 | |
1226 | for (const Value *V : PtrArgs) { |
1227 | SmallVector<const Value *, 4> Objects; |
1228 | getUnderlyingObjects(V, Objects, /* LI = */ nullptr); |
1229 | |
1230 | for (const Value *O : Objects) |
1231 | ObjSet.insert(Ptr: O); |
1232 | } |
1233 | |
1234 | // Figure out if we're derived from anything that is not a noalias |
1235 | // argument. |
1236 | bool RequiresNoCaptureBefore = false, UsesAliasingPtr = false, |
1237 | UsesUnknownObject = false; |
1238 | for (const Value *V : ObjSet) { |
1239 | // Is this value a constant that cannot be derived from any pointer |
1240 | // value (we need to exclude constant expressions, for example, that |
1241 | // are formed from arithmetic on global symbols). |
1242 | bool IsNonPtrConst = isa<ConstantInt>(Val: V) || isa<ConstantFP>(Val: V) || |
1243 | isa<ConstantPointerNull>(Val: V) || |
1244 | isa<ConstantDataVector>(Val: V) || isa<UndefValue>(Val: V); |
1245 | if (IsNonPtrConst) |
1246 | continue; |
1247 | |
1248 | // If this is anything other than a noalias argument, then we cannot |
1249 | // completely describe the aliasing properties using alias.scope |
1250 | // metadata (and, thus, won't add any). |
1251 | if (const Argument *A = dyn_cast<Argument>(Val: V)) { |
1252 | if (!CB.paramHasAttr(ArgNo: A->getArgNo(), Kind: Attribute::NoAlias)) |
1253 | UsesAliasingPtr = true; |
1254 | } else { |
1255 | UsesAliasingPtr = true; |
1256 | } |
1257 | |
1258 | if (isEscapeSource(V)) { |
1259 | // An escape source can only alias with a noalias argument if it has |
1260 | // been captured beforehand. |
1261 | RequiresNoCaptureBefore = true; |
1262 | } else if (!isa<Argument>(Val: V) && !isIdentifiedObject(V)) { |
1263 | // If this is neither an escape source, nor some identified object |
1264 | // (which cannot directly alias a noalias argument), nor some other |
1265 | // argument (which, by definition, also cannot alias a noalias |
1266 | // argument), conservatively do not make any assumptions. |
1267 | UsesUnknownObject = true; |
1268 | } |
1269 | } |
1270 | |
1271 | // Nothing we can do if the used underlying object cannot be reliably |
1272 | // determined. |
1273 | if (UsesUnknownObject) |
1274 | continue; |
1275 | |
1276 | // A function call can always get captured noalias pointers (via other |
1277 | // parameters, globals, etc.). |
1278 | if (IsFuncCall && !IsArgMemOnlyCall) |
1279 | RequiresNoCaptureBefore = true; |
1280 | |
1281 | // First, we want to figure out all of the sets with which we definitely |
1282 | // don't alias. Iterate over all noalias set, and add those for which: |
1283 | // 1. The noalias argument is not in the set of objects from which we |
1284 | // definitely derive. |
1285 | // 2. The noalias argument has not yet been captured. |
1286 | // An arbitrary function that might load pointers could see captured |
1287 | // noalias arguments via other noalias arguments or globals, and so we |
1288 | // must always check for prior capture. |
1289 | for (const Argument *A : NoAliasArgs) { |
1290 | if (ObjSet.contains(Ptr: A)) |
1291 | continue; // May be based on a noalias argument. |
1292 | |
1293 | // It might be tempting to skip the PointerMayBeCapturedBefore check if |
1294 | // A->hasNoCaptureAttr() is true, but this is incorrect because |
1295 | // nocapture only guarantees that no copies outlive the function, not |
1296 | // that the value cannot be locally captured. |
1297 | if (!RequiresNoCaptureBefore || |
1298 | !PointerMayBeCapturedBefore(V: A, /* ReturnCaptures */ false, |
1299 | /* StoreCaptures */ false, I, DT: &DT)) |
1300 | NoAliases.push_back(Elt: NewScopes[A]); |
1301 | } |
1302 | |
1303 | if (!NoAliases.empty()) |
1304 | NI->setMetadata(KindID: LLVMContext::MD_noalias, |
1305 | Node: MDNode::concatenate( |
1306 | A: NI->getMetadata(KindID: LLVMContext::MD_noalias), |
1307 | B: MDNode::get(Context&: CalledFunc->getContext(), MDs: NoAliases))); |
1308 | |
1309 | // Next, we want to figure out all of the sets to which we might belong. |
1310 | // We might belong to a set if the noalias argument is in the set of |
1311 | // underlying objects. If there is some non-noalias argument in our list |
1312 | // of underlying objects, then we cannot add a scope because the fact |
1313 | // that some access does not alias with any set of our noalias arguments |
1314 | // cannot itself guarantee that it does not alias with this access |
1315 | // (because there is some pointer of unknown origin involved and the |
1316 | // other access might also depend on this pointer). We also cannot add |
1317 | // scopes to arbitrary functions unless we know they don't access any |
1318 | // non-parameter pointer-values. |
1319 | bool CanAddScopes = !UsesAliasingPtr; |
1320 | if (CanAddScopes && IsFuncCall) |
1321 | CanAddScopes = IsArgMemOnlyCall; |
1322 | |
1323 | if (CanAddScopes) |
1324 | for (const Argument *A : NoAliasArgs) { |
1325 | if (ObjSet.count(Ptr: A)) |
1326 | Scopes.push_back(Elt: NewScopes[A]); |
1327 | } |
1328 | |
1329 | if (!Scopes.empty()) |
1330 | NI->setMetadata( |
1331 | KindID: LLVMContext::MD_alias_scope, |
1332 | Node: MDNode::concatenate(A: NI->getMetadata(KindID: LLVMContext::MD_alias_scope), |
1333 | B: MDNode::get(Context&: CalledFunc->getContext(), MDs: Scopes))); |
1334 | } |
1335 | } |
1336 | } |
1337 | |
1338 | static bool MayContainThrowingOrExitingCallAfterCB(CallBase *Begin, |
1339 | ReturnInst *End) { |
1340 | |
1341 | assert(Begin->getParent() == End->getParent() && |
1342 | "Expected to be in same basic block!" ); |
1343 | auto BeginIt = Begin->getIterator(); |
1344 | assert(BeginIt != End->getIterator() && "Non-empty BB has empty iterator" ); |
1345 | return !llvm::isGuaranteedToTransferExecutionToSuccessor( |
1346 | Begin: ++BeginIt, End: End->getIterator(), ScanLimit: InlinerAttributeWindow + 1); |
1347 | } |
1348 | |
1349 | // Add attributes from CB params and Fn attributes that can always be propagated |
1350 | // to the corresponding argument / inner callbases. |
1351 | static void AddParamAndFnBasicAttributes(const CallBase &CB, |
1352 | ValueToValueMapTy &VMap) { |
1353 | auto *CalledFunction = CB.getCalledFunction(); |
1354 | auto &Context = CalledFunction->getContext(); |
1355 | |
1356 | // Collect valid attributes for all params. |
1357 | SmallVector<AttrBuilder> ValidParamAttrs; |
1358 | bool HasAttrToPropagate = false; |
1359 | |
1360 | for (unsigned I = 0, E = CB.arg_size(); I < E; ++I) { |
1361 | ValidParamAttrs.emplace_back(Args: AttrBuilder{CB.getContext()}); |
1362 | // Access attributes can be propagated to any param with the same underlying |
1363 | // object as the argument. |
1364 | if (CB.paramHasAttr(ArgNo: I, Kind: Attribute::ReadNone)) |
1365 | ValidParamAttrs.back().addAttribute(Val: Attribute::ReadNone); |
1366 | if (CB.paramHasAttr(ArgNo: I, Kind: Attribute::ReadOnly)) |
1367 | ValidParamAttrs.back().addAttribute(Val: Attribute::ReadOnly); |
1368 | HasAttrToPropagate |= ValidParamAttrs.back().hasAttributes(); |
1369 | } |
1370 | |
1371 | // Won't be able to propagate anything. |
1372 | if (!HasAttrToPropagate) |
1373 | return; |
1374 | |
1375 | for (BasicBlock &BB : *CalledFunction) { |
1376 | for (Instruction &Ins : BB) { |
1377 | const auto *InnerCB = dyn_cast<CallBase>(Val: &Ins); |
1378 | if (!InnerCB) |
1379 | continue; |
1380 | auto *NewInnerCB = dyn_cast_or_null<CallBase>(Val: VMap.lookup(Val: InnerCB)); |
1381 | if (!NewInnerCB) |
1382 | continue; |
1383 | AttributeList AL = NewInnerCB->getAttributes(); |
1384 | for (unsigned I = 0, E = InnerCB->arg_size(); I < E; ++I) { |
1385 | // Check if the underlying value for the parameter is an argument. |
1386 | const Value *UnderlyingV = |
1387 | getUnderlyingObject(V: InnerCB->getArgOperand(i: I)); |
1388 | const Argument *Arg = dyn_cast<Argument>(Val: UnderlyingV); |
1389 | if (!Arg) |
1390 | continue; |
1391 | |
1392 | if (AL.hasParamAttr(ArgNo: I, Kind: Attribute::ByVal)) |
1393 | // It's unsound to propagate memory attributes to byval arguments. |
1394 | // Even if CalledFunction doesn't e.g. write to the argument, |
1395 | // the call to NewInnerCB may write to its by-value copy. |
1396 | continue; |
1397 | |
1398 | unsigned ArgNo = Arg->getArgNo(); |
1399 | // If so, propagate its access attributes. |
1400 | AL = AL.addParamAttributes(C&: Context, ArgNo: I, B: ValidParamAttrs[ArgNo]); |
1401 | // We can have conflicting attributes from the inner callsite and |
1402 | // to-be-inlined callsite. In that case, choose the most |
1403 | // restrictive. |
1404 | |
1405 | // readonly + writeonly means we can never deref so make readnone. |
1406 | if (AL.hasParamAttr(ArgNo: I, Kind: Attribute::ReadOnly) && |
1407 | AL.hasParamAttr(ArgNo: I, Kind: Attribute::WriteOnly)) |
1408 | AL = AL.addParamAttribute(C&: Context, ArgNo: I, Kind: Attribute::ReadNone); |
1409 | |
1410 | // If have readnone, need to clear readonly/writeonly |
1411 | if (AL.hasParamAttr(ArgNo: I, Kind: Attribute::ReadNone)) { |
1412 | AL = AL.removeParamAttribute(C&: Context, ArgNo: I, Kind: Attribute::ReadOnly); |
1413 | AL = AL.removeParamAttribute(C&: Context, ArgNo: I, Kind: Attribute::WriteOnly); |
1414 | } |
1415 | |
1416 | // Writable cannot exist in conjunction w/ readonly/readnone |
1417 | if (AL.hasParamAttr(ArgNo: I, Kind: Attribute::ReadOnly) || |
1418 | AL.hasParamAttr(ArgNo: I, Kind: Attribute::ReadNone)) |
1419 | AL = AL.removeParamAttribute(C&: Context, ArgNo: I, Kind: Attribute::Writable); |
1420 | } |
1421 | NewInnerCB->setAttributes(AL); |
1422 | } |
1423 | } |
1424 | } |
1425 | |
1426 | // Only allow these white listed attributes to be propagated back to the |
1427 | // callee. This is because other attributes may only be valid on the call |
1428 | // itself, i.e. attributes such as signext and zeroext. |
1429 | |
1430 | // Attributes that are always okay to propagate as if they are violated its |
1431 | // immediate UB. |
1432 | static AttrBuilder IdentifyValidUBGeneratingAttributes(CallBase &CB) { |
1433 | AttrBuilder Valid(CB.getContext()); |
1434 | if (auto DerefBytes = CB.getRetDereferenceableBytes()) |
1435 | Valid.addDereferenceableAttr(Bytes: DerefBytes); |
1436 | if (auto DerefOrNullBytes = CB.getRetDereferenceableOrNullBytes()) |
1437 | Valid.addDereferenceableOrNullAttr(Bytes: DerefOrNullBytes); |
1438 | if (CB.hasRetAttr(Kind: Attribute::NoAlias)) |
1439 | Valid.addAttribute(Val: Attribute::NoAlias); |
1440 | if (CB.hasRetAttr(Kind: Attribute::NoUndef)) |
1441 | Valid.addAttribute(Val: Attribute::NoUndef); |
1442 | return Valid; |
1443 | } |
1444 | |
1445 | // Attributes that need additional checks as propagating them may change |
1446 | // behavior or cause new UB. |
1447 | static AttrBuilder IdentifyValidPoisonGeneratingAttributes(CallBase &CB) { |
1448 | AttrBuilder Valid(CB.getContext()); |
1449 | if (CB.hasRetAttr(Kind: Attribute::NonNull)) |
1450 | Valid.addAttribute(Val: Attribute::NonNull); |
1451 | if (CB.hasRetAttr(Kind: Attribute::Alignment)) |
1452 | Valid.addAlignmentAttr(Align: CB.getRetAlign()); |
1453 | if (std::optional<ConstantRange> Range = CB.getRange()) |
1454 | Valid.addRangeAttr(CR: *Range); |
1455 | return Valid; |
1456 | } |
1457 | |
1458 | static void AddReturnAttributes(CallBase &CB, ValueToValueMapTy &VMap) { |
1459 | AttrBuilder ValidUB = IdentifyValidUBGeneratingAttributes(CB); |
1460 | AttrBuilder ValidPG = IdentifyValidPoisonGeneratingAttributes(CB); |
1461 | if (!ValidUB.hasAttributes() && !ValidPG.hasAttributes()) |
1462 | return; |
1463 | auto *CalledFunction = CB.getCalledFunction(); |
1464 | auto &Context = CalledFunction->getContext(); |
1465 | |
1466 | for (auto &BB : *CalledFunction) { |
1467 | auto *RI = dyn_cast<ReturnInst>(Val: BB.getTerminator()); |
1468 | if (!RI || !isa<CallBase>(Val: RI->getOperand(i_nocapture: 0))) |
1469 | continue; |
1470 | auto *RetVal = cast<CallBase>(Val: RI->getOperand(i_nocapture: 0)); |
1471 | // Check that the cloned RetVal exists and is a call, otherwise we cannot |
1472 | // add the attributes on the cloned RetVal. Simplification during inlining |
1473 | // could have transformed the cloned instruction. |
1474 | auto *NewRetVal = dyn_cast_or_null<CallBase>(Val: VMap.lookup(Val: RetVal)); |
1475 | if (!NewRetVal) |
1476 | continue; |
1477 | // Backward propagation of attributes to the returned value may be incorrect |
1478 | // if it is control flow dependent. |
1479 | // Consider: |
1480 | // @callee { |
1481 | // %rv = call @foo() |
1482 | // %rv2 = call @bar() |
1483 | // if (%rv2 != null) |
1484 | // return %rv2 |
1485 | // if (%rv == null) |
1486 | // exit() |
1487 | // return %rv |
1488 | // } |
1489 | // caller() { |
1490 | // %val = call nonnull @callee() |
1491 | // } |
1492 | // Here we cannot add the nonnull attribute on either foo or bar. So, we |
1493 | // limit the check to both RetVal and RI are in the same basic block and |
1494 | // there are no throwing/exiting instructions between these instructions. |
1495 | if (RI->getParent() != RetVal->getParent() || |
1496 | MayContainThrowingOrExitingCallAfterCB(Begin: RetVal, End: RI)) |
1497 | continue; |
1498 | // Add to the existing attributes of NewRetVal, i.e. the cloned call |
1499 | // instruction. |
1500 | // NB! When we have the same attribute already existing on NewRetVal, but |
1501 | // with a differing value, the AttributeList's merge API honours the already |
1502 | // existing attribute value (i.e. attributes such as dereferenceable, |
1503 | // dereferenceable_or_null etc). See AttrBuilder::merge for more details. |
1504 | AttributeList AL = NewRetVal->getAttributes(); |
1505 | if (ValidUB.getDereferenceableBytes() < AL.getRetDereferenceableBytes()) |
1506 | ValidUB.removeAttribute(Val: Attribute::Dereferenceable); |
1507 | if (ValidUB.getDereferenceableOrNullBytes() < |
1508 | AL.getRetDereferenceableOrNullBytes()) |
1509 | ValidUB.removeAttribute(Val: Attribute::DereferenceableOrNull); |
1510 | AttributeList NewAL = AL.addRetAttributes(C&: Context, B: ValidUB); |
1511 | // Attributes that may generate poison returns are a bit tricky. If we |
1512 | // propagate them, other uses of the callsite might have their behavior |
1513 | // change or cause UB (if they have noundef) b.c of the new potential |
1514 | // poison. |
1515 | // Take the following three cases: |
1516 | // |
1517 | // 1) |
1518 | // define nonnull ptr @foo() { |
1519 | // %p = call ptr @bar() |
1520 | // call void @use(ptr %p) willreturn nounwind |
1521 | // ret ptr %p |
1522 | // } |
1523 | // |
1524 | // 2) |
1525 | // define noundef nonnull ptr @foo() { |
1526 | // %p = call ptr @bar() |
1527 | // call void @use(ptr %p) willreturn nounwind |
1528 | // ret ptr %p |
1529 | // } |
1530 | // |
1531 | // 3) |
1532 | // define nonnull ptr @foo() { |
1533 | // %p = call noundef ptr @bar() |
1534 | // ret ptr %p |
1535 | // } |
1536 | // |
1537 | // In case 1, we can't propagate nonnull because poison value in @use may |
1538 | // change behavior or trigger UB. |
1539 | // In case 2, we don't need to be concerned about propagating nonnull, as |
1540 | // any new poison at @use will trigger UB anyways. |
1541 | // In case 3, we can never propagate nonnull because it may create UB due to |
1542 | // the noundef on @bar. |
1543 | if (ValidPG.getAlignment().valueOrOne() < AL.getRetAlignment().valueOrOne()) |
1544 | ValidPG.removeAttribute(Val: Attribute::Alignment); |
1545 | if (ValidPG.hasAttributes()) { |
1546 | Attribute CBRange = ValidPG.getAttribute(Kind: Attribute::Range); |
1547 | if (CBRange.isValid()) { |
1548 | Attribute NewRange = AL.getRetAttr(Kind: Attribute::Range); |
1549 | if (NewRange.isValid()) { |
1550 | ValidPG.addRangeAttr( |
1551 | CR: CBRange.getRange().intersectWith(CR: NewRange.getRange())); |
1552 | } |
1553 | } |
1554 | // Three checks. |
1555 | // If the callsite has `noundef`, then a poison due to violating the |
1556 | // return attribute will create UB anyways so we can always propagate. |
1557 | // Otherwise, if the return value (callee to be inlined) has `noundef`, we |
1558 | // can't propagate as a new poison return will cause UB. |
1559 | // Finally, check if the return value has no uses whose behavior may |
1560 | // change/may cause UB if we potentially return poison. At the moment this |
1561 | // is implemented overly conservatively with a single-use check. |
1562 | // TODO: Update the single-use check to iterate through uses and only bail |
1563 | // if we have a potentially dangerous use. |
1564 | |
1565 | if (CB.hasRetAttr(Kind: Attribute::NoUndef) || |
1566 | (RetVal->hasOneUse() && !RetVal->hasRetAttr(Kind: Attribute::NoUndef))) |
1567 | NewAL = NewAL.addRetAttributes(C&: Context, B: ValidPG); |
1568 | } |
1569 | NewRetVal->setAttributes(NewAL); |
1570 | } |
1571 | } |
1572 | |
1573 | /// If the inlined function has non-byval align arguments, then |
1574 | /// add @llvm.assume-based alignment assumptions to preserve this information. |
1575 | static void AddAlignmentAssumptions(CallBase &CB, InlineFunctionInfo &IFI) { |
1576 | if (!PreserveAlignmentAssumptions || !IFI.GetAssumptionCache) |
1577 | return; |
1578 | |
1579 | AssumptionCache *AC = &IFI.GetAssumptionCache(*CB.getCaller()); |
1580 | auto &DL = CB.getDataLayout(); |
1581 | |
1582 | // To avoid inserting redundant assumptions, we should check for assumptions |
1583 | // already in the caller. To do this, we might need a DT of the caller. |
1584 | DominatorTree DT; |
1585 | bool DTCalculated = false; |
1586 | |
1587 | Function *CalledFunc = CB.getCalledFunction(); |
1588 | for (Argument &Arg : CalledFunc->args()) { |
1589 | if (!Arg.getType()->isPointerTy() || Arg.hasPassPointeeByValueCopyAttr() || |
1590 | Arg.hasNUses(N: 0)) |
1591 | continue; |
1592 | MaybeAlign Alignment = Arg.getParamAlign(); |
1593 | if (!Alignment) |
1594 | continue; |
1595 | |
1596 | if (!DTCalculated) { |
1597 | DT.recalculate(Func&: *CB.getCaller()); |
1598 | DTCalculated = true; |
1599 | } |
1600 | // If we can already prove the asserted alignment in the context of the |
1601 | // caller, then don't bother inserting the assumption. |
1602 | Value *ArgVal = CB.getArgOperand(i: Arg.getArgNo()); |
1603 | if (getKnownAlignment(V: ArgVal, DL, CxtI: &CB, AC, DT: &DT) >= *Alignment) |
1604 | continue; |
1605 | |
1606 | CallInst *NewAsmp = IRBuilder<>(&CB).CreateAlignmentAssumption( |
1607 | DL, PtrValue: ArgVal, Alignment: Alignment->value()); |
1608 | AC->registerAssumption(CI: cast<AssumeInst>(Val: NewAsmp)); |
1609 | } |
1610 | } |
1611 | |
1612 | static void HandleByValArgumentInit(Type *ByValType, Value *Dst, Value *Src, |
1613 | Module *M, BasicBlock *InsertBlock, |
1614 | InlineFunctionInfo &IFI, |
1615 | Function *CalledFunc) { |
1616 | IRBuilder<> Builder(InsertBlock, InsertBlock->begin()); |
1617 | |
1618 | Value *Size = |
1619 | Builder.getInt64(C: M->getDataLayout().getTypeStoreSize(Ty: ByValType)); |
1620 | |
1621 | // Always generate a memcpy of alignment 1 here because we don't know |
1622 | // the alignment of the src pointer. Other optimizations can infer |
1623 | // better alignment. |
1624 | CallInst *CI = Builder.CreateMemCpy(Dst, /*DstAlign*/ Align(1), Src, |
1625 | /*SrcAlign*/ Align(1), Size); |
1626 | |
1627 | // The verifier requires that all calls of debug-info-bearing functions |
1628 | // from debug-info-bearing functions have a debug location (for inlining |
1629 | // purposes). Assign a dummy location to satisfy the constraint. |
1630 | if (!CI->getDebugLoc() && InsertBlock->getParent()->getSubprogram()) |
1631 | if (DISubprogram *SP = CalledFunc->getSubprogram()) |
1632 | CI->setDebugLoc(DILocation::get(Context&: SP->getContext(), Line: 0, Column: 0, Scope: SP)); |
1633 | } |
1634 | |
1635 | /// When inlining a call site that has a byval argument, |
1636 | /// we have to make the implicit memcpy explicit by adding it. |
1637 | static Value *HandleByValArgument(Type *ByValType, Value *Arg, |
1638 | Instruction *TheCall, |
1639 | const Function *CalledFunc, |
1640 | InlineFunctionInfo &IFI, |
1641 | MaybeAlign ByValAlignment) { |
1642 | Function *Caller = TheCall->getFunction(); |
1643 | const DataLayout &DL = Caller->getDataLayout(); |
1644 | |
1645 | // If the called function is readonly, then it could not mutate the caller's |
1646 | // copy of the byval'd memory. In this case, it is safe to elide the copy and |
1647 | // temporary. |
1648 | if (CalledFunc->onlyReadsMemory()) { |
1649 | // If the byval argument has a specified alignment that is greater than the |
1650 | // passed in pointer, then we either have to round up the input pointer or |
1651 | // give up on this transformation. |
1652 | if (ByValAlignment.valueOrOne() == 1) |
1653 | return Arg; |
1654 | |
1655 | AssumptionCache *AC = |
1656 | IFI.GetAssumptionCache ? &IFI.GetAssumptionCache(*Caller) : nullptr; |
1657 | |
1658 | // If the pointer is already known to be sufficiently aligned, or if we can |
1659 | // round it up to a larger alignment, then we don't need a temporary. |
1660 | if (getOrEnforceKnownAlignment(V: Arg, PrefAlign: *ByValAlignment, DL, CxtI: TheCall, AC) >= |
1661 | *ByValAlignment) |
1662 | return Arg; |
1663 | |
1664 | // Otherwise, we have to make a memcpy to get a safe alignment. This is bad |
1665 | // for code quality, but rarely happens and is required for correctness. |
1666 | } |
1667 | |
1668 | // Create the alloca. If we have DataLayout, use nice alignment. |
1669 | Align Alignment = DL.getPrefTypeAlign(Ty: ByValType); |
1670 | |
1671 | // If the byval had an alignment specified, we *must* use at least that |
1672 | // alignment, as it is required by the byval argument (and uses of the |
1673 | // pointer inside the callee). |
1674 | if (ByValAlignment) |
1675 | Alignment = std::max(a: Alignment, b: *ByValAlignment); |
1676 | |
1677 | AllocaInst *NewAlloca = |
1678 | new AllocaInst(ByValType, Arg->getType()->getPointerAddressSpace(), |
1679 | nullptr, Alignment, Arg->getName()); |
1680 | NewAlloca->insertBefore(InsertPos: Caller->begin()->begin()); |
1681 | IFI.StaticAllocas.push_back(Elt: NewAlloca); |
1682 | |
1683 | // Uses of the argument in the function should use our new alloca |
1684 | // instead. |
1685 | return NewAlloca; |
1686 | } |
1687 | |
1688 | // Check whether this Value is used by a lifetime intrinsic. |
1689 | static bool isUsedByLifetimeMarker(Value *V) { |
1690 | for (User *U : V->users()) |
1691 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: U)) |
1692 | if (II->isLifetimeStartOrEnd()) |
1693 | return true; |
1694 | return false; |
1695 | } |
1696 | |
1697 | // Check whether the given alloca already has |
1698 | // lifetime.start or lifetime.end intrinsics. |
1699 | static bool hasLifetimeMarkers(AllocaInst *AI) { |
1700 | Type *Ty = AI->getType(); |
1701 | Type *Int8PtrTy = |
1702 | PointerType::get(C&: Ty->getContext(), AddressSpace: Ty->getPointerAddressSpace()); |
1703 | if (Ty == Int8PtrTy) |
1704 | return isUsedByLifetimeMarker(V: AI); |
1705 | |
1706 | // Do a scan to find all the casts to i8*. |
1707 | for (User *U : AI->users()) { |
1708 | if (U->getType() != Int8PtrTy) continue; |
1709 | if (U->stripPointerCasts() != AI) continue; |
1710 | if (isUsedByLifetimeMarker(V: U)) |
1711 | return true; |
1712 | } |
1713 | return false; |
1714 | } |
1715 | |
1716 | /// Return the result of AI->isStaticAlloca() if AI were moved to the entry |
1717 | /// block. Allocas used in inalloca calls and allocas of dynamic array size |
1718 | /// cannot be static. |
1719 | static bool allocaWouldBeStaticInEntry(const AllocaInst *AI ) { |
1720 | return isa<Constant>(Val: AI->getArraySize()) && !AI->isUsedWithInAlloca(); |
1721 | } |
1722 | |
1723 | /// Returns a DebugLoc for a new DILocation which is a clone of \p OrigDL |
1724 | /// inlined at \p InlinedAt. \p IANodes is an inlined-at cache. |
1725 | static DebugLoc inlineDebugLoc(DebugLoc OrigDL, DILocation *InlinedAt, |
1726 | LLVMContext &Ctx, |
1727 | DenseMap<const MDNode *, MDNode *> &IANodes) { |
1728 | auto IA = DebugLoc::appendInlinedAt(DL: OrigDL, InlinedAt, Ctx, Cache&: IANodes); |
1729 | return DILocation::get(Context&: Ctx, Line: OrigDL.getLine(), Column: OrigDL.getCol(), |
1730 | Scope: OrigDL.getScope(), InlinedAt: IA); |
1731 | } |
1732 | |
1733 | /// Update inlined instructions' line numbers to |
1734 | /// to encode location where these instructions are inlined. |
1735 | static void fixupLineNumbers(Function *Fn, Function::iterator FI, |
1736 | Instruction *TheCall, bool CalleeHasDebugInfo) { |
1737 | const DebugLoc &TheCallDL = TheCall->getDebugLoc(); |
1738 | if (!TheCallDL) |
1739 | return; |
1740 | |
1741 | auto &Ctx = Fn->getContext(); |
1742 | DILocation *InlinedAtNode = TheCallDL; |
1743 | |
1744 | // Create a unique call site, not to be confused with any other call from the |
1745 | // same location. |
1746 | InlinedAtNode = DILocation::getDistinct( |
1747 | Context&: Ctx, Line: InlinedAtNode->getLine(), Column: InlinedAtNode->getColumn(), |
1748 | Scope: InlinedAtNode->getScope(), InlinedAt: InlinedAtNode->getInlinedAt()); |
1749 | |
1750 | // Cache the inlined-at nodes as they're built so they are reused, without |
1751 | // this every instruction's inlined-at chain would become distinct from each |
1752 | // other. |
1753 | DenseMap<const MDNode *, MDNode *> IANodes; |
1754 | |
1755 | // Check if we are not generating inline line tables and want to use |
1756 | // the call site location instead. |
1757 | bool NoInlineLineTables = Fn->hasFnAttribute(Kind: "no-inline-line-tables" ); |
1758 | |
1759 | // Helper-util for updating the metadata attached to an instruction. |
1760 | auto UpdateInst = [&](Instruction &I) { |
1761 | // Loop metadata needs to be updated so that the start and end locs |
1762 | // reference inlined-at locations. |
1763 | auto updateLoopInfoLoc = [&Ctx, &InlinedAtNode, |
1764 | &IANodes](Metadata *MD) -> Metadata * { |
1765 | if (auto *Loc = dyn_cast_or_null<DILocation>(Val: MD)) |
1766 | return inlineDebugLoc(OrigDL: Loc, InlinedAt: InlinedAtNode, Ctx, IANodes).get(); |
1767 | return MD; |
1768 | }; |
1769 | updateLoopMetadataDebugLocations(I, Updater: updateLoopInfoLoc); |
1770 | |
1771 | if (!NoInlineLineTables) |
1772 | if (DebugLoc DL = I.getDebugLoc()) { |
1773 | DebugLoc IDL = |
1774 | inlineDebugLoc(OrigDL: DL, InlinedAt: InlinedAtNode, Ctx&: I.getContext(), IANodes); |
1775 | I.setDebugLoc(IDL); |
1776 | return; |
1777 | } |
1778 | |
1779 | if (CalleeHasDebugInfo && !NoInlineLineTables) |
1780 | return; |
1781 | |
1782 | // If the inlined instruction has no line number, or if inline info |
1783 | // is not being generated, make it look as if it originates from the call |
1784 | // location. This is important for ((__always_inline, __nodebug__)) |
1785 | // functions which must use caller location for all instructions in their |
1786 | // function body. |
1787 | |
1788 | // Don't update static allocas, as they may get moved later. |
1789 | if (auto *AI = dyn_cast<AllocaInst>(Val: &I)) |
1790 | if (allocaWouldBeStaticInEntry(AI)) |
1791 | return; |
1792 | |
1793 | // Do not force a debug loc for pseudo probes, since they do not need to |
1794 | // be debuggable, and also they are expected to have a zero/null dwarf |
1795 | // discriminator at this point which could be violated otherwise. |
1796 | if (isa<PseudoProbeInst>(Val: I)) |
1797 | return; |
1798 | |
1799 | I.setDebugLoc(TheCallDL); |
1800 | }; |
1801 | |
1802 | // Helper-util for updating debug-info records attached to instructions. |
1803 | auto UpdateDVR = [&](DbgRecord *DVR) { |
1804 | assert(DVR->getDebugLoc() && "Debug Value must have debug loc" ); |
1805 | if (NoInlineLineTables) { |
1806 | DVR->setDebugLoc(TheCallDL); |
1807 | return; |
1808 | } |
1809 | DebugLoc DL = DVR->getDebugLoc(); |
1810 | DebugLoc IDL = |
1811 | inlineDebugLoc(OrigDL: DL, InlinedAt: InlinedAtNode, |
1812 | Ctx&: DVR->getMarker()->getParent()->getContext(), IANodes); |
1813 | DVR->setDebugLoc(IDL); |
1814 | }; |
1815 | |
1816 | // Iterate over all instructions, updating metadata and debug-info records. |
1817 | for (; FI != Fn->end(); ++FI) { |
1818 | for (Instruction &I : *FI) { |
1819 | UpdateInst(I); |
1820 | for (DbgRecord &DVR : I.getDbgRecordRange()) { |
1821 | UpdateDVR(&DVR); |
1822 | } |
1823 | } |
1824 | |
1825 | // Remove debug info intrinsics if we're not keeping inline info. |
1826 | if (NoInlineLineTables) { |
1827 | BasicBlock::iterator BI = FI->begin(); |
1828 | while (BI != FI->end()) { |
1829 | if (isa<DbgInfoIntrinsic>(Val: BI)) { |
1830 | BI = BI->eraseFromParent(); |
1831 | continue; |
1832 | } else { |
1833 | BI->dropDbgRecords(); |
1834 | } |
1835 | ++BI; |
1836 | } |
1837 | } |
1838 | } |
1839 | } |
1840 | |
1841 | #undef DEBUG_TYPE |
1842 | #define DEBUG_TYPE "assignment-tracking" |
1843 | /// Find Alloca and linked DbgAssignIntrinsic for locals escaped by \p CB. |
1844 | static at::StorageToVarsMap collectEscapedLocals(const DataLayout &DL, |
1845 | const CallBase &CB) { |
1846 | at::StorageToVarsMap EscapedLocals; |
1847 | SmallPtrSet<const Value *, 4> SeenBases; |
1848 | |
1849 | LLVM_DEBUG( |
1850 | errs() << "# Finding caller local variables escaped by callee\n" ); |
1851 | for (const Value *Arg : CB.args()) { |
1852 | LLVM_DEBUG(errs() << "INSPECT: " << *Arg << "\n" ); |
1853 | if (!Arg->getType()->isPointerTy()) { |
1854 | LLVM_DEBUG(errs() << " | SKIP: Not a pointer\n" ); |
1855 | continue; |
1856 | } |
1857 | |
1858 | const Instruction *I = dyn_cast<Instruction>(Val: Arg); |
1859 | if (!I) { |
1860 | LLVM_DEBUG(errs() << " | SKIP: Not result of instruction\n" ); |
1861 | continue; |
1862 | } |
1863 | |
1864 | // Walk back to the base storage. |
1865 | assert(Arg->getType()->isPtrOrPtrVectorTy()); |
1866 | APInt TmpOffset(DL.getIndexTypeSizeInBits(Ty: Arg->getType()), 0, false); |
1867 | const AllocaInst *Base = dyn_cast<AllocaInst>( |
1868 | Val: Arg->stripAndAccumulateConstantOffsets(DL, Offset&: TmpOffset, AllowNonInbounds: true)); |
1869 | if (!Base) { |
1870 | LLVM_DEBUG(errs() << " | SKIP: Couldn't walk back to base storage\n" ); |
1871 | continue; |
1872 | } |
1873 | |
1874 | assert(Base); |
1875 | LLVM_DEBUG(errs() << " | BASE: " << *Base << "\n" ); |
1876 | // We only need to process each base address once - skip any duplicates. |
1877 | if (!SeenBases.insert(Ptr: Base).second) |
1878 | continue; |
1879 | |
1880 | // Find all local variables associated with the backing storage. |
1881 | auto CollectAssignsForStorage = [&](auto *DbgAssign) { |
1882 | // Skip variables from inlined functions - they are not local variables. |
1883 | if (DbgAssign->getDebugLoc().getInlinedAt()) |
1884 | return; |
1885 | LLVM_DEBUG(errs() << " > DEF : " << *DbgAssign << "\n" ); |
1886 | EscapedLocals[Base].insert(X: at::VarRecord(DbgAssign)); |
1887 | }; |
1888 | for_each(Range: at::getAssignmentMarkers(Inst: Base), F: CollectAssignsForStorage); |
1889 | for_each(Range: at::getDVRAssignmentMarkers(Inst: Base), F: CollectAssignsForStorage); |
1890 | } |
1891 | return EscapedLocals; |
1892 | } |
1893 | |
1894 | static void trackInlinedStores(Function::iterator Start, Function::iterator End, |
1895 | const CallBase &CB) { |
1896 | LLVM_DEBUG(errs() << "trackInlinedStores into " |
1897 | << Start->getParent()->getName() << " from " |
1898 | << CB.getCalledFunction()->getName() << "\n" ); |
1899 | std::unique_ptr<DataLayout> DL = std::make_unique<DataLayout>(args: CB.getModule()); |
1900 | at::trackAssignments(Start, End, Vars: collectEscapedLocals(DL: *DL, CB), DL: *DL); |
1901 | } |
1902 | |
1903 | /// Update inlined instructions' DIAssignID metadata. We need to do this |
1904 | /// otherwise a function inlined more than once into the same function |
1905 | /// will cause DIAssignID to be shared by many instructions. |
1906 | static void fixupAssignments(Function::iterator Start, Function::iterator End) { |
1907 | DenseMap<DIAssignID *, DIAssignID *> Map; |
1908 | // Loop over all the inlined instructions. If we find a DIAssignID |
1909 | // attachment or use, replace it with a new version. |
1910 | for (auto BBI = Start; BBI != End; ++BBI) { |
1911 | for (Instruction &I : *BBI) |
1912 | at::remapAssignID(Map, I); |
1913 | } |
1914 | } |
1915 | #undef DEBUG_TYPE |
1916 | #define DEBUG_TYPE "inline-function" |
1917 | |
1918 | /// Update the block frequencies of the caller after a callee has been inlined. |
1919 | /// |
1920 | /// Each block cloned into the caller has its block frequency scaled by the |
1921 | /// ratio of CallSiteFreq/CalleeEntryFreq. This ensures that the cloned copy of |
1922 | /// callee's entry block gets the same frequency as the callsite block and the |
1923 | /// relative frequencies of all cloned blocks remain the same after cloning. |
1924 | static void updateCallerBFI(BasicBlock *CallSiteBlock, |
1925 | const ValueToValueMapTy &VMap, |
1926 | BlockFrequencyInfo *CallerBFI, |
1927 | BlockFrequencyInfo *CalleeBFI, |
1928 | const BasicBlock &CalleeEntryBlock) { |
1929 | SmallPtrSet<BasicBlock *, 16> ClonedBBs; |
1930 | for (auto Entry : VMap) { |
1931 | if (!isa<BasicBlock>(Val: Entry.first) || !Entry.second) |
1932 | continue; |
1933 | auto *OrigBB = cast<BasicBlock>(Val: Entry.first); |
1934 | auto *ClonedBB = cast<BasicBlock>(Val: Entry.second); |
1935 | BlockFrequency Freq = CalleeBFI->getBlockFreq(BB: OrigBB); |
1936 | if (!ClonedBBs.insert(Ptr: ClonedBB).second) { |
1937 | // Multiple blocks in the callee might get mapped to one cloned block in |
1938 | // the caller since we prune the callee as we clone it. When that happens, |
1939 | // we want to use the maximum among the original blocks' frequencies. |
1940 | BlockFrequency NewFreq = CallerBFI->getBlockFreq(BB: ClonedBB); |
1941 | if (NewFreq > Freq) |
1942 | Freq = NewFreq; |
1943 | } |
1944 | CallerBFI->setBlockFreq(BB: ClonedBB, Freq); |
1945 | } |
1946 | BasicBlock *EntryClone = cast<BasicBlock>(Val: VMap.lookup(Val: &CalleeEntryBlock)); |
1947 | CallerBFI->setBlockFreqAndScale( |
1948 | ReferenceBB: EntryClone, Freq: CallerBFI->getBlockFreq(BB: CallSiteBlock), BlocksToScale&: ClonedBBs); |
1949 | } |
1950 | |
1951 | /// Update the branch metadata for cloned call instructions. |
1952 | static void updateCallProfile(Function *Callee, const ValueToValueMapTy &VMap, |
1953 | const ProfileCount &CalleeEntryCount, |
1954 | const CallBase &TheCall, ProfileSummaryInfo *PSI, |
1955 | BlockFrequencyInfo *CallerBFI) { |
1956 | if (CalleeEntryCount.isSynthetic() || CalleeEntryCount.getCount() < 1) |
1957 | return; |
1958 | auto CallSiteCount = |
1959 | PSI ? PSI->getProfileCount(CallInst: TheCall, BFI: CallerBFI) : std::nullopt; |
1960 | int64_t CallCount = |
1961 | std::min(a: CallSiteCount.value_or(u: 0), b: CalleeEntryCount.getCount()); |
1962 | updateProfileCallee(Callee, EntryDelta: -CallCount, VMap: &VMap); |
1963 | } |
1964 | |
1965 | void llvm::updateProfileCallee( |
1966 | Function *Callee, int64_t EntryDelta, |
1967 | const ValueMap<const Value *, WeakTrackingVH> *VMap) { |
1968 | auto CalleeCount = Callee->getEntryCount(); |
1969 | if (!CalleeCount) |
1970 | return; |
1971 | |
1972 | const uint64_t PriorEntryCount = CalleeCount->getCount(); |
1973 | |
1974 | // Since CallSiteCount is an estimate, it could exceed the original callee |
1975 | // count and has to be set to 0 so guard against underflow. |
1976 | const uint64_t NewEntryCount = |
1977 | (EntryDelta < 0 && static_cast<uint64_t>(-EntryDelta) > PriorEntryCount) |
1978 | ? 0 |
1979 | : PriorEntryCount + EntryDelta; |
1980 | |
1981 | auto updateVTableProfWeight = [](CallBase *CB, const uint64_t NewEntryCount, |
1982 | const uint64_t PriorEntryCount) { |
1983 | Instruction *VPtr = PGOIndirectCallVisitor::tryGetVTableInstruction(CB); |
1984 | if (VPtr) |
1985 | scaleProfData(I&: *VPtr, S: NewEntryCount, T: PriorEntryCount); |
1986 | }; |
1987 | |
1988 | // During inlining ? |
1989 | if (VMap) { |
1990 | uint64_t CloneEntryCount = PriorEntryCount - NewEntryCount; |
1991 | for (auto Entry : *VMap) { |
1992 | if (isa<CallInst>(Val: Entry.first)) |
1993 | if (auto *CI = dyn_cast_or_null<CallInst>(Val: Entry.second)) { |
1994 | CI->updateProfWeight(S: CloneEntryCount, T: PriorEntryCount); |
1995 | updateVTableProfWeight(CI, CloneEntryCount, PriorEntryCount); |
1996 | } |
1997 | |
1998 | if (isa<InvokeInst>(Val: Entry.first)) |
1999 | if (auto *II = dyn_cast_or_null<InvokeInst>(Val: Entry.second)) { |
2000 | II->updateProfWeight(S: CloneEntryCount, T: PriorEntryCount); |
2001 | updateVTableProfWeight(II, CloneEntryCount, PriorEntryCount); |
2002 | } |
2003 | } |
2004 | } |
2005 | |
2006 | if (EntryDelta) { |
2007 | Callee->setEntryCount(Count: NewEntryCount); |
2008 | |
2009 | for (BasicBlock &BB : *Callee) |
2010 | // No need to update the callsite if it is pruned during inlining. |
2011 | if (!VMap || VMap->count(Val: &BB)) |
2012 | for (Instruction &I : BB) { |
2013 | if (CallInst *CI = dyn_cast<CallInst>(Val: &I)) { |
2014 | CI->updateProfWeight(S: NewEntryCount, T: PriorEntryCount); |
2015 | updateVTableProfWeight(CI, NewEntryCount, PriorEntryCount); |
2016 | } |
2017 | if (InvokeInst *II = dyn_cast<InvokeInst>(Val: &I)) { |
2018 | II->updateProfWeight(S: NewEntryCount, T: PriorEntryCount); |
2019 | updateVTableProfWeight(II, NewEntryCount, PriorEntryCount); |
2020 | } |
2021 | } |
2022 | } |
2023 | } |
2024 | |
2025 | /// An operand bundle "clang.arc.attachedcall" on a call indicates the call |
2026 | /// result is implicitly consumed by a call to retainRV or claimRV immediately |
2027 | /// after the call. This function inlines the retainRV/claimRV calls. |
2028 | /// |
2029 | /// There are three cases to consider: |
2030 | /// |
2031 | /// 1. If there is a call to autoreleaseRV that takes a pointer to the returned |
2032 | /// object in the callee return block, the autoreleaseRV call and the |
2033 | /// retainRV/claimRV call in the caller cancel out. If the call in the caller |
2034 | /// is a claimRV call, a call to objc_release is emitted. |
2035 | /// |
2036 | /// 2. If there is a call in the callee return block that doesn't have operand |
2037 | /// bundle "clang.arc.attachedcall", the operand bundle on the original call |
2038 | /// is transferred to the call in the callee. |
2039 | /// |
2040 | /// 3. Otherwise, a call to objc_retain is inserted if the call in the caller is |
2041 | /// a retainRV call. |
2042 | static void |
2043 | inlineRetainOrClaimRVCalls(CallBase &CB, objcarc::ARCInstKind RVCallKind, |
2044 | const SmallVectorImpl<ReturnInst *> &Returns) { |
2045 | Module *Mod = CB.getModule(); |
2046 | assert(objcarc::isRetainOrClaimRV(RVCallKind) && "unexpected ARC function" ); |
2047 | bool IsRetainRV = RVCallKind == objcarc::ARCInstKind::RetainRV, |
2048 | IsUnsafeClaimRV = !IsRetainRV; |
2049 | |
2050 | for (auto *RI : Returns) { |
2051 | Value *RetOpnd = objcarc::GetRCIdentityRoot(V: RI->getOperand(i_nocapture: 0)); |
2052 | bool InsertRetainCall = IsRetainRV; |
2053 | IRBuilder<> Builder(RI->getContext()); |
2054 | |
2055 | // Walk backwards through the basic block looking for either a matching |
2056 | // autoreleaseRV call or an unannotated call. |
2057 | auto InstRange = llvm::make_range(x: ++(RI->getIterator().getReverse()), |
2058 | y: RI->getParent()->rend()); |
2059 | for (Instruction &I : llvm::make_early_inc_range(Range&: InstRange)) { |
2060 | // Ignore casts. |
2061 | if (isa<CastInst>(Val: I)) |
2062 | continue; |
2063 | |
2064 | if (auto *II = dyn_cast<IntrinsicInst>(Val: &I)) { |
2065 | if (II->getIntrinsicID() != Intrinsic::objc_autoreleaseReturnValue || |
2066 | !II->hasNUses(N: 0) || |
2067 | objcarc::GetRCIdentityRoot(V: II->getOperand(i_nocapture: 0)) != RetOpnd) |
2068 | break; |
2069 | |
2070 | // If we've found a matching authoreleaseRV call: |
2071 | // - If claimRV is attached to the call, insert a call to objc_release |
2072 | // and erase the autoreleaseRV call. |
2073 | // - If retainRV is attached to the call, just erase the autoreleaseRV |
2074 | // call. |
2075 | if (IsUnsafeClaimRV) { |
2076 | Builder.SetInsertPoint(II); |
2077 | Function *IFn = |
2078 | Intrinsic::getDeclaration(M: Mod, id: Intrinsic::objc_release); |
2079 | Builder.CreateCall(Callee: IFn, Args: RetOpnd, Name: "" ); |
2080 | } |
2081 | II->eraseFromParent(); |
2082 | InsertRetainCall = false; |
2083 | break; |
2084 | } |
2085 | |
2086 | auto *CI = dyn_cast<CallInst>(Val: &I); |
2087 | |
2088 | if (!CI) |
2089 | break; |
2090 | |
2091 | if (objcarc::GetRCIdentityRoot(V: CI) != RetOpnd || |
2092 | objcarc::hasAttachedCallOpBundle(CB: CI)) |
2093 | break; |
2094 | |
2095 | // If we've found an unannotated call that defines RetOpnd, add a |
2096 | // "clang.arc.attachedcall" operand bundle. |
2097 | Value *BundleArgs[] = {*objcarc::getAttachedARCFunction(CB: &CB)}; |
2098 | OperandBundleDef OB("clang.arc.attachedcall" , BundleArgs); |
2099 | auto *NewCall = CallBase::addOperandBundle( |
2100 | CB: CI, ID: LLVMContext::OB_clang_arc_attachedcall, OB, InsertPt: CI->getIterator()); |
2101 | NewCall->copyMetadata(SrcInst: *CI); |
2102 | CI->replaceAllUsesWith(V: NewCall); |
2103 | CI->eraseFromParent(); |
2104 | InsertRetainCall = false; |
2105 | break; |
2106 | } |
2107 | |
2108 | if (InsertRetainCall) { |
2109 | // The retainRV is attached to the call and we've failed to find a |
2110 | // matching autoreleaseRV or an annotated call in the callee. Emit a call |
2111 | // to objc_retain. |
2112 | Builder.SetInsertPoint(RI); |
2113 | Function *IFn = Intrinsic::getDeclaration(M: Mod, id: Intrinsic::objc_retain); |
2114 | Builder.CreateCall(Callee: IFn, Args: RetOpnd, Name: "" ); |
2115 | } |
2116 | } |
2117 | } |
2118 | |
2119 | /// This function inlines the called function into the basic block of the |
2120 | /// caller. This returns false if it is not possible to inline this call. |
2121 | /// The program is still in a well defined state if this occurs though. |
2122 | /// |
2123 | /// Note that this only does one level of inlining. For example, if the |
2124 | /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now |
2125 | /// exists in the instruction stream. Similarly this will inline a recursive |
2126 | /// function by one level. |
2127 | llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI, |
2128 | bool MergeAttributes, |
2129 | AAResults *CalleeAAR, |
2130 | bool InsertLifetime, |
2131 | Function *ForwardVarArgsTo) { |
2132 | assert(CB.getParent() && CB.getFunction() && "Instruction not in function!" ); |
2133 | |
2134 | // FIXME: we don't inline callbr yet. |
2135 | if (isa<CallBrInst>(Val: CB)) |
2136 | return InlineResult::failure(Reason: "We don't inline callbr yet." ); |
2137 | |
2138 | // If IFI has any state in it, zap it before we fill it in. |
2139 | IFI.reset(); |
2140 | |
2141 | Function *CalledFunc = CB.getCalledFunction(); |
2142 | if (!CalledFunc || // Can't inline external function or indirect |
2143 | CalledFunc->isDeclaration()) // call! |
2144 | return InlineResult::failure(Reason: "external or indirect" ); |
2145 | |
2146 | // The inliner does not know how to inline through calls with operand bundles |
2147 | // in general ... |
2148 | Value *ConvergenceControlToken = nullptr; |
2149 | if (CB.hasOperandBundles()) { |
2150 | for (int i = 0, e = CB.getNumOperandBundles(); i != e; ++i) { |
2151 | auto OBUse = CB.getOperandBundleAt(Index: i); |
2152 | uint32_t Tag = OBUse.getTagID(); |
2153 | // ... but it knows how to inline through "deopt" operand bundles ... |
2154 | if (Tag == LLVMContext::OB_deopt) |
2155 | continue; |
2156 | // ... and "funclet" operand bundles. |
2157 | if (Tag == LLVMContext::OB_funclet) |
2158 | continue; |
2159 | if (Tag == LLVMContext::OB_clang_arc_attachedcall) |
2160 | continue; |
2161 | if (Tag == LLVMContext::OB_kcfi) |
2162 | continue; |
2163 | if (Tag == LLVMContext::OB_convergencectrl) { |
2164 | ConvergenceControlToken = OBUse.Inputs[0].get(); |
2165 | continue; |
2166 | } |
2167 | |
2168 | return InlineResult::failure(Reason: "unsupported operand bundle" ); |
2169 | } |
2170 | } |
2171 | |
2172 | // FIXME: The check below is redundant and incomplete. According to spec, if a |
2173 | // convergent call is missing a token, then the caller is using uncontrolled |
2174 | // convergence. If the callee has an entry intrinsic, then the callee is using |
2175 | // controlled convergence, and the call cannot be inlined. A proper |
2176 | // implemenation of this check requires a whole new analysis that identifies |
2177 | // convergence in every function. For now, we skip that and just do this one |
2178 | // cursory check. The underlying assumption is that in a compiler flow that |
2179 | // fully implements convergence control tokens, there is no mixing of |
2180 | // controlled and uncontrolled convergent operations in the whole program. |
2181 | if (CB.isConvergent()) { |
2182 | auto *I = CalledFunc->getEntryBlock().getFirstNonPHI(); |
2183 | if (auto *IntrinsicCall = dyn_cast<IntrinsicInst>(Val: I)) { |
2184 | if (IntrinsicCall->getIntrinsicID() == |
2185 | Intrinsic::experimental_convergence_entry) { |
2186 | if (!ConvergenceControlToken) { |
2187 | return InlineResult::failure( |
2188 | Reason: "convergent call needs convergencectrl operand" ); |
2189 | } |
2190 | } |
2191 | } |
2192 | } |
2193 | |
2194 | // If the call to the callee cannot throw, set the 'nounwind' flag on any |
2195 | // calls that we inline. |
2196 | bool MarkNoUnwind = CB.doesNotThrow(); |
2197 | |
2198 | BasicBlock *OrigBB = CB.getParent(); |
2199 | Function *Caller = OrigBB->getParent(); |
2200 | |
2201 | // GC poses two hazards to inlining, which only occur when the callee has GC: |
2202 | // 1. If the caller has no GC, then the callee's GC must be propagated to the |
2203 | // caller. |
2204 | // 2. If the caller has a differing GC, it is invalid to inline. |
2205 | if (CalledFunc->hasGC()) { |
2206 | if (!Caller->hasGC()) |
2207 | Caller->setGC(CalledFunc->getGC()); |
2208 | else if (CalledFunc->getGC() != Caller->getGC()) |
2209 | return InlineResult::failure(Reason: "incompatible GC" ); |
2210 | } |
2211 | |
2212 | // Get the personality function from the callee if it contains a landing pad. |
2213 | Constant *CalledPersonality = |
2214 | CalledFunc->hasPersonalityFn() |
2215 | ? CalledFunc->getPersonalityFn()->stripPointerCasts() |
2216 | : nullptr; |
2217 | |
2218 | // Find the personality function used by the landing pads of the caller. If it |
2219 | // exists, then check to see that it matches the personality function used in |
2220 | // the callee. |
2221 | Constant *CallerPersonality = |
2222 | Caller->hasPersonalityFn() |
2223 | ? Caller->getPersonalityFn()->stripPointerCasts() |
2224 | : nullptr; |
2225 | if (CalledPersonality) { |
2226 | if (!CallerPersonality) |
2227 | Caller->setPersonalityFn(CalledPersonality); |
2228 | // If the personality functions match, then we can perform the |
2229 | // inlining. Otherwise, we can't inline. |
2230 | // TODO: This isn't 100% true. Some personality functions are proper |
2231 | // supersets of others and can be used in place of the other. |
2232 | else if (CalledPersonality != CallerPersonality) |
2233 | return InlineResult::failure(Reason: "incompatible personality" ); |
2234 | } |
2235 | |
2236 | // We need to figure out which funclet the callsite was in so that we may |
2237 | // properly nest the callee. |
2238 | Instruction *CallSiteEHPad = nullptr; |
2239 | if (CallerPersonality) { |
2240 | EHPersonality Personality = classifyEHPersonality(Pers: CallerPersonality); |
2241 | if (isScopedEHPersonality(Pers: Personality)) { |
2242 | std::optional<OperandBundleUse> ParentFunclet = |
2243 | CB.getOperandBundle(ID: LLVMContext::OB_funclet); |
2244 | if (ParentFunclet) |
2245 | CallSiteEHPad = cast<FuncletPadInst>(Val: ParentFunclet->Inputs.front()); |
2246 | |
2247 | // OK, the inlining site is legal. What about the target function? |
2248 | |
2249 | if (CallSiteEHPad) { |
2250 | if (Personality == EHPersonality::MSVC_CXX) { |
2251 | // The MSVC personality cannot tolerate catches getting inlined into |
2252 | // cleanup funclets. |
2253 | if (isa<CleanupPadInst>(Val: CallSiteEHPad)) { |
2254 | // Ok, the call site is within a cleanuppad. Let's check the callee |
2255 | // for catchpads. |
2256 | for (const BasicBlock &CalledBB : *CalledFunc) { |
2257 | if (isa<CatchSwitchInst>(Val: CalledBB.getFirstNonPHI())) |
2258 | return InlineResult::failure(Reason: "catch in cleanup funclet" ); |
2259 | } |
2260 | } |
2261 | } else if (isAsynchronousEHPersonality(Pers: Personality)) { |
2262 | // SEH is even less tolerant, there may not be any sort of exceptional |
2263 | // funclet in the callee. |
2264 | for (const BasicBlock &CalledBB : *CalledFunc) { |
2265 | if (CalledBB.isEHPad()) |
2266 | return InlineResult::failure(Reason: "SEH in cleanup funclet" ); |
2267 | } |
2268 | } |
2269 | } |
2270 | } |
2271 | } |
2272 | |
2273 | // Determine if we are dealing with a call in an EHPad which does not unwind |
2274 | // to caller. |
2275 | bool EHPadForCallUnwindsLocally = false; |
2276 | if (CallSiteEHPad && isa<CallInst>(Val: CB)) { |
2277 | UnwindDestMemoTy FuncletUnwindMap; |
2278 | Value *CallSiteUnwindDestToken = |
2279 | getUnwindDestToken(EHPad: CallSiteEHPad, MemoMap&: FuncletUnwindMap); |
2280 | |
2281 | EHPadForCallUnwindsLocally = |
2282 | CallSiteUnwindDestToken && |
2283 | !isa<ConstantTokenNone>(Val: CallSiteUnwindDestToken); |
2284 | } |
2285 | |
2286 | // Get an iterator to the last basic block in the function, which will have |
2287 | // the new function inlined after it. |
2288 | Function::iterator LastBlock = --Caller->end(); |
2289 | |
2290 | // Make sure to capture all of the return instructions from the cloned |
2291 | // function. |
2292 | SmallVector<ReturnInst*, 8> Returns; |
2293 | ClonedCodeInfo InlinedFunctionInfo; |
2294 | Function::iterator FirstNewBlock; |
2295 | |
2296 | { // Scope to destroy VMap after cloning. |
2297 | ValueToValueMapTy VMap; |
2298 | struct ByValInit { |
2299 | Value *Dst; |
2300 | Value *Src; |
2301 | Type *Ty; |
2302 | }; |
2303 | // Keep a list of pair (dst, src) to emit byval initializations. |
2304 | SmallVector<ByValInit, 4> ByValInits; |
2305 | |
2306 | // When inlining a function that contains noalias scope metadata, |
2307 | // this metadata needs to be cloned so that the inlined blocks |
2308 | // have different "unique scopes" at every call site. |
2309 | // Track the metadata that must be cloned. Do this before other changes to |
2310 | // the function, so that we do not get in trouble when inlining caller == |
2311 | // callee. |
2312 | ScopedAliasMetadataDeepCloner SAMetadataCloner(CB.getCalledFunction()); |
2313 | |
2314 | auto &DL = Caller->getDataLayout(); |
2315 | |
2316 | // Calculate the vector of arguments to pass into the function cloner, which |
2317 | // matches up the formal to the actual argument values. |
2318 | auto AI = CB.arg_begin(); |
2319 | unsigned ArgNo = 0; |
2320 | for (Function::arg_iterator I = CalledFunc->arg_begin(), |
2321 | E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { |
2322 | Value *ActualArg = *AI; |
2323 | |
2324 | // When byval arguments actually inlined, we need to make the copy implied |
2325 | // by them explicit. However, we don't do this if the callee is readonly |
2326 | // or readnone, because the copy would be unneeded: the callee doesn't |
2327 | // modify the struct. |
2328 | if (CB.isByValArgument(ArgNo)) { |
2329 | ActualArg = HandleByValArgument(ByValType: CB.getParamByValType(ArgNo), Arg: ActualArg, |
2330 | TheCall: &CB, CalledFunc, IFI, |
2331 | ByValAlignment: CalledFunc->getParamAlign(ArgNo)); |
2332 | if (ActualArg != *AI) |
2333 | ByValInits.push_back( |
2334 | Elt: {.Dst: ActualArg, .Src: (Value *)*AI, .Ty: CB.getParamByValType(ArgNo)}); |
2335 | } |
2336 | |
2337 | VMap[&*I] = ActualArg; |
2338 | } |
2339 | |
2340 | // TODO: Remove this when users have been updated to the assume bundles. |
2341 | // Add alignment assumptions if necessary. We do this before the inlined |
2342 | // instructions are actually cloned into the caller so that we can easily |
2343 | // check what will be known at the start of the inlined code. |
2344 | AddAlignmentAssumptions(CB, IFI); |
2345 | |
2346 | AssumptionCache *AC = |
2347 | IFI.GetAssumptionCache ? &IFI.GetAssumptionCache(*Caller) : nullptr; |
2348 | |
2349 | /// Preserve all attributes on of the call and its parameters. |
2350 | salvageKnowledge(I: &CB, AC); |
2351 | |
2352 | // We want the inliner to prune the code as it copies. We would LOVE to |
2353 | // have no dead or constant instructions leftover after inlining occurs |
2354 | // (which can happen, e.g., because an argument was constant), but we'll be |
2355 | // happy with whatever the cloner can do. |
2356 | CloneAndPruneFunctionInto(NewFunc: Caller, OldFunc: CalledFunc, VMap, |
2357 | /*ModuleLevelChanges=*/false, Returns, NameSuffix: ".i" , |
2358 | CodeInfo: &InlinedFunctionInfo); |
2359 | // Remember the first block that is newly cloned over. |
2360 | FirstNewBlock = LastBlock; ++FirstNewBlock; |
2361 | |
2362 | // Insert retainRV/clainRV runtime calls. |
2363 | objcarc::ARCInstKind RVCallKind = objcarc::getAttachedARCFunctionKind(CB: &CB); |
2364 | if (RVCallKind != objcarc::ARCInstKind::None) |
2365 | inlineRetainOrClaimRVCalls(CB, RVCallKind, Returns); |
2366 | |
2367 | // Updated caller/callee profiles only when requested. For sample loader |
2368 | // inlining, the context-sensitive inlinee profile doesn't need to be |
2369 | // subtracted from callee profile, and the inlined clone also doesn't need |
2370 | // to be scaled based on call site count. |
2371 | if (IFI.UpdateProfile) { |
2372 | if (IFI.CallerBFI != nullptr && IFI.CalleeBFI != nullptr) |
2373 | // Update the BFI of blocks cloned into the caller. |
2374 | updateCallerBFI(CallSiteBlock: OrigBB, VMap, CallerBFI: IFI.CallerBFI, CalleeBFI: IFI.CalleeBFI, |
2375 | CalleeEntryBlock: CalledFunc->front()); |
2376 | |
2377 | if (auto Profile = CalledFunc->getEntryCount()) |
2378 | updateCallProfile(Callee: CalledFunc, VMap, CalleeEntryCount: *Profile, TheCall: CB, PSI: IFI.PSI, |
2379 | CallerBFI: IFI.CallerBFI); |
2380 | } |
2381 | |
2382 | // Inject byval arguments initialization. |
2383 | for (ByValInit &Init : ByValInits) |
2384 | HandleByValArgumentInit(ByValType: Init.Ty, Dst: Init.Dst, Src: Init.Src, M: Caller->getParent(), |
2385 | InsertBlock: &*FirstNewBlock, IFI, CalledFunc); |
2386 | |
2387 | std::optional<OperandBundleUse> ParentDeopt = |
2388 | CB.getOperandBundle(ID: LLVMContext::OB_deopt); |
2389 | if (ParentDeopt) { |
2390 | SmallVector<OperandBundleDef, 2> OpDefs; |
2391 | |
2392 | for (auto &VH : InlinedFunctionInfo.OperandBundleCallSites) { |
2393 | CallBase *ICS = dyn_cast_or_null<CallBase>(Val&: VH); |
2394 | if (!ICS) |
2395 | continue; // instruction was DCE'd or RAUW'ed to undef |
2396 | |
2397 | OpDefs.clear(); |
2398 | |
2399 | OpDefs.reserve(N: ICS->getNumOperandBundles()); |
2400 | |
2401 | for (unsigned COBi = 0, COBe = ICS->getNumOperandBundles(); COBi < COBe; |
2402 | ++COBi) { |
2403 | auto ChildOB = ICS->getOperandBundleAt(Index: COBi); |
2404 | if (ChildOB.getTagID() != LLVMContext::OB_deopt) { |
2405 | // If the inlined call has other operand bundles, let them be |
2406 | OpDefs.emplace_back(Args&: ChildOB); |
2407 | continue; |
2408 | } |
2409 | |
2410 | // It may be useful to separate this logic (of handling operand |
2411 | // bundles) out to a separate "policy" component if this gets crowded. |
2412 | // Prepend the parent's deoptimization continuation to the newly |
2413 | // inlined call's deoptimization continuation. |
2414 | std::vector<Value *> MergedDeoptArgs; |
2415 | MergedDeoptArgs.reserve(n: ParentDeopt->Inputs.size() + |
2416 | ChildOB.Inputs.size()); |
2417 | |
2418 | llvm::append_range(C&: MergedDeoptArgs, R&: ParentDeopt->Inputs); |
2419 | llvm::append_range(C&: MergedDeoptArgs, R&: ChildOB.Inputs); |
2420 | |
2421 | OpDefs.emplace_back(Args: "deopt" , Args: std::move(MergedDeoptArgs)); |
2422 | } |
2423 | |
2424 | Instruction *NewI = CallBase::Create(CB: ICS, Bundles: OpDefs, InsertPt: ICS->getIterator()); |
2425 | |
2426 | // Note: the RAUW does the appropriate fixup in VMap, so we need to do |
2427 | // this even if the call returns void. |
2428 | ICS->replaceAllUsesWith(V: NewI); |
2429 | |
2430 | VH = nullptr; |
2431 | ICS->eraseFromParent(); |
2432 | } |
2433 | } |
2434 | |
2435 | // For 'nodebug' functions, the associated DISubprogram is always null. |
2436 | // Conservatively avoid propagating the callsite debug location to |
2437 | // instructions inlined from a function whose DISubprogram is not null. |
2438 | fixupLineNumbers(Fn: Caller, FI: FirstNewBlock, TheCall: &CB, |
2439 | CalleeHasDebugInfo: CalledFunc->getSubprogram() != nullptr); |
2440 | |
2441 | if (isAssignmentTrackingEnabled(M: *Caller->getParent())) { |
2442 | // Interpret inlined stores to caller-local variables as assignments. |
2443 | trackInlinedStores(Start: FirstNewBlock, End: Caller->end(), CB); |
2444 | |
2445 | // Update DIAssignID metadata attachments and uses so that they are |
2446 | // unique to this inlined instance. |
2447 | fixupAssignments(Start: FirstNewBlock, End: Caller->end()); |
2448 | } |
2449 | |
2450 | // Now clone the inlined noalias scope metadata. |
2451 | SAMetadataCloner.clone(); |
2452 | SAMetadataCloner.remap(FStart: FirstNewBlock, FEnd: Caller->end()); |
2453 | |
2454 | // Add noalias metadata if necessary. |
2455 | AddAliasScopeMetadata(CB, VMap, DL, CalleeAAR, InlinedFunctionInfo); |
2456 | |
2457 | // Clone return attributes on the callsite into the calls within the inlined |
2458 | // function which feed into its return value. |
2459 | AddReturnAttributes(CB, VMap); |
2460 | |
2461 | // Clone attributes on the params of the callsite to calls within the |
2462 | // inlined function which use the same param. |
2463 | AddParamAndFnBasicAttributes(CB, VMap); |
2464 | |
2465 | propagateMemProfMetadata(Callee: CalledFunc, CB, |
2466 | ContainsMemProfMetadata: InlinedFunctionInfo.ContainsMemProfMetadata, VMap); |
2467 | |
2468 | // Propagate metadata on the callsite if necessary. |
2469 | PropagateCallSiteMetadata(CB, FStart: FirstNewBlock, FEnd: Caller->end()); |
2470 | |
2471 | // Register any cloned assumptions. |
2472 | if (IFI.GetAssumptionCache) |
2473 | for (BasicBlock &NewBlock : |
2474 | make_range(x: FirstNewBlock->getIterator(), y: Caller->end())) |
2475 | for (Instruction &I : NewBlock) |
2476 | if (auto *II = dyn_cast<AssumeInst>(Val: &I)) |
2477 | IFI.GetAssumptionCache(*Caller).registerAssumption(CI: II); |
2478 | } |
2479 | |
2480 | if (ConvergenceControlToken) { |
2481 | auto *I = FirstNewBlock->getFirstNonPHI(); |
2482 | if (auto *IntrinsicCall = dyn_cast<IntrinsicInst>(Val: I)) { |
2483 | if (IntrinsicCall->getIntrinsicID() == |
2484 | Intrinsic::experimental_convergence_entry) { |
2485 | IntrinsicCall->replaceAllUsesWith(V: ConvergenceControlToken); |
2486 | IntrinsicCall->eraseFromParent(); |
2487 | } |
2488 | } |
2489 | } |
2490 | |
2491 | // If there are any alloca instructions in the block that used to be the entry |
2492 | // block for the callee, move them to the entry block of the caller. First |
2493 | // calculate which instruction they should be inserted before. We insert the |
2494 | // instructions at the end of the current alloca list. |
2495 | { |
2496 | BasicBlock::iterator InsertPoint = Caller->begin()->begin(); |
2497 | for (BasicBlock::iterator I = FirstNewBlock->begin(), |
2498 | E = FirstNewBlock->end(); I != E; ) { |
2499 | AllocaInst *AI = dyn_cast<AllocaInst>(Val: I++); |
2500 | if (!AI) continue; |
2501 | |
2502 | // If the alloca is now dead, remove it. This often occurs due to code |
2503 | // specialization. |
2504 | if (AI->use_empty()) { |
2505 | AI->eraseFromParent(); |
2506 | continue; |
2507 | } |
2508 | |
2509 | if (!allocaWouldBeStaticInEntry(AI)) |
2510 | continue; |
2511 | |
2512 | // Keep track of the static allocas that we inline into the caller. |
2513 | IFI.StaticAllocas.push_back(Elt: AI); |
2514 | |
2515 | // Scan for the block of allocas that we can move over, and move them |
2516 | // all at once. |
2517 | while (isa<AllocaInst>(Val: I) && |
2518 | !cast<AllocaInst>(Val&: I)->use_empty() && |
2519 | allocaWouldBeStaticInEntry(AI: cast<AllocaInst>(Val&: I))) { |
2520 | IFI.StaticAllocas.push_back(Elt: cast<AllocaInst>(Val&: I)); |
2521 | ++I; |
2522 | } |
2523 | |
2524 | // Transfer all of the allocas over in a block. Using splice means |
2525 | // that the instructions aren't removed from the symbol table, then |
2526 | // reinserted. |
2527 | I.setTailBit(true); |
2528 | Caller->getEntryBlock().splice(ToIt: InsertPoint, FromBB: &*FirstNewBlock, |
2529 | FromBeginIt: AI->getIterator(), FromEndIt: I); |
2530 | } |
2531 | } |
2532 | |
2533 | SmallVector<Value*,4> VarArgsToForward; |
2534 | SmallVector<AttributeSet, 4> VarArgsAttrs; |
2535 | for (unsigned i = CalledFunc->getFunctionType()->getNumParams(); |
2536 | i < CB.arg_size(); i++) { |
2537 | VarArgsToForward.push_back(Elt: CB.getArgOperand(i)); |
2538 | VarArgsAttrs.push_back(Elt: CB.getAttributes().getParamAttrs(ArgNo: i)); |
2539 | } |
2540 | |
2541 | bool InlinedMustTailCalls = false, InlinedDeoptimizeCalls = false; |
2542 | if (InlinedFunctionInfo.ContainsCalls) { |
2543 | CallInst::TailCallKind CallSiteTailKind = CallInst::TCK_None; |
2544 | if (CallInst *CI = dyn_cast<CallInst>(Val: &CB)) |
2545 | CallSiteTailKind = CI->getTailCallKind(); |
2546 | |
2547 | // For inlining purposes, the "notail" marker is the same as no marker. |
2548 | if (CallSiteTailKind == CallInst::TCK_NoTail) |
2549 | CallSiteTailKind = CallInst::TCK_None; |
2550 | |
2551 | for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; |
2552 | ++BB) { |
2553 | for (Instruction &I : llvm::make_early_inc_range(Range&: *BB)) { |
2554 | CallInst *CI = dyn_cast<CallInst>(Val: &I); |
2555 | if (!CI) |
2556 | continue; |
2557 | |
2558 | // Forward varargs from inlined call site to calls to the |
2559 | // ForwardVarArgsTo function, if requested, and to musttail calls. |
2560 | if (!VarArgsToForward.empty() && |
2561 | ((ForwardVarArgsTo && |
2562 | CI->getCalledFunction() == ForwardVarArgsTo) || |
2563 | CI->isMustTailCall())) { |
2564 | // Collect attributes for non-vararg parameters. |
2565 | AttributeList Attrs = CI->getAttributes(); |
2566 | SmallVector<AttributeSet, 8> ArgAttrs; |
2567 | if (!Attrs.isEmpty() || !VarArgsAttrs.empty()) { |
2568 | for (unsigned ArgNo = 0; |
2569 | ArgNo < CI->getFunctionType()->getNumParams(); ++ArgNo) |
2570 | ArgAttrs.push_back(Elt: Attrs.getParamAttrs(ArgNo)); |
2571 | } |
2572 | |
2573 | // Add VarArg attributes. |
2574 | ArgAttrs.append(in_start: VarArgsAttrs.begin(), in_end: VarArgsAttrs.end()); |
2575 | Attrs = AttributeList::get(C&: CI->getContext(), FnAttrs: Attrs.getFnAttrs(), |
2576 | RetAttrs: Attrs.getRetAttrs(), ArgAttrs); |
2577 | // Add VarArgs to existing parameters. |
2578 | SmallVector<Value *, 6> Params(CI->args()); |
2579 | Params.append(in_start: VarArgsToForward.begin(), in_end: VarArgsToForward.end()); |
2580 | CallInst *NewCI = CallInst::Create( |
2581 | Ty: CI->getFunctionType(), Func: CI->getCalledOperand(), Args: Params, NameStr: "" , InsertBefore: CI->getIterator()); |
2582 | NewCI->setDebugLoc(CI->getDebugLoc()); |
2583 | NewCI->setAttributes(Attrs); |
2584 | NewCI->setCallingConv(CI->getCallingConv()); |
2585 | CI->replaceAllUsesWith(V: NewCI); |
2586 | CI->eraseFromParent(); |
2587 | CI = NewCI; |
2588 | } |
2589 | |
2590 | if (Function *F = CI->getCalledFunction()) |
2591 | InlinedDeoptimizeCalls |= |
2592 | F->getIntrinsicID() == Intrinsic::experimental_deoptimize; |
2593 | |
2594 | // We need to reduce the strength of any inlined tail calls. For |
2595 | // musttail, we have to avoid introducing potential unbounded stack |
2596 | // growth. For example, if functions 'f' and 'g' are mutually recursive |
2597 | // with musttail, we can inline 'g' into 'f' so long as we preserve |
2598 | // musttail on the cloned call to 'f'. If either the inlined call site |
2599 | // or the cloned call site is *not* musttail, the program already has |
2600 | // one frame of stack growth, so it's safe to remove musttail. Here is |
2601 | // a table of example transformations: |
2602 | // |
2603 | // f -> musttail g -> musttail f ==> f -> musttail f |
2604 | // f -> musttail g -> tail f ==> f -> tail f |
2605 | // f -> g -> musttail f ==> f -> f |
2606 | // f -> g -> tail f ==> f -> f |
2607 | // |
2608 | // Inlined notail calls should remain notail calls. |
2609 | CallInst::TailCallKind ChildTCK = CI->getTailCallKind(); |
2610 | if (ChildTCK != CallInst::TCK_NoTail) |
2611 | ChildTCK = std::min(a: CallSiteTailKind, b: ChildTCK); |
2612 | CI->setTailCallKind(ChildTCK); |
2613 | InlinedMustTailCalls |= CI->isMustTailCall(); |
2614 | |
2615 | // Call sites inlined through a 'nounwind' call site should be |
2616 | // 'nounwind' as well. However, avoid marking call sites explicitly |
2617 | // where possible. This helps expose more opportunities for CSE after |
2618 | // inlining, commonly when the callee is an intrinsic. |
2619 | if (MarkNoUnwind && !CI->doesNotThrow()) |
2620 | CI->setDoesNotThrow(); |
2621 | } |
2622 | } |
2623 | } |
2624 | |
2625 | // Leave lifetime markers for the static alloca's, scoping them to the |
2626 | // function we just inlined. |
2627 | // We need to insert lifetime intrinsics even at O0 to avoid invalid |
2628 | // access caused by multithreaded coroutines. The check |
2629 | // `Caller->isPresplitCoroutine()` would affect AlwaysInliner at O0 only. |
2630 | if ((InsertLifetime || Caller->isPresplitCoroutine()) && |
2631 | !IFI.StaticAllocas.empty()) { |
2632 | IRBuilder<> builder(&*FirstNewBlock, FirstNewBlock->begin()); |
2633 | for (AllocaInst *AI : IFI.StaticAllocas) { |
2634 | // Don't mark swifterror allocas. They can't have bitcast uses. |
2635 | if (AI->isSwiftError()) |
2636 | continue; |
2637 | |
2638 | // If the alloca is already scoped to something smaller than the whole |
2639 | // function then there's no need to add redundant, less accurate markers. |
2640 | if (hasLifetimeMarkers(AI)) |
2641 | continue; |
2642 | |
2643 | // Try to determine the size of the allocation. |
2644 | ConstantInt *AllocaSize = nullptr; |
2645 | if (ConstantInt *AIArraySize = |
2646 | dyn_cast<ConstantInt>(Val: AI->getArraySize())) { |
2647 | auto &DL = Caller->getDataLayout(); |
2648 | Type *AllocaType = AI->getAllocatedType(); |
2649 | TypeSize AllocaTypeSize = DL.getTypeAllocSize(Ty: AllocaType); |
2650 | uint64_t AllocaArraySize = AIArraySize->getLimitedValue(); |
2651 | |
2652 | // Don't add markers for zero-sized allocas. |
2653 | if (AllocaArraySize == 0) |
2654 | continue; |
2655 | |
2656 | // Check that array size doesn't saturate uint64_t and doesn't |
2657 | // overflow when it's multiplied by type size. |
2658 | if (!AllocaTypeSize.isScalable() && |
2659 | AllocaArraySize != std::numeric_limits<uint64_t>::max() && |
2660 | std::numeric_limits<uint64_t>::max() / AllocaArraySize >= |
2661 | AllocaTypeSize.getFixedValue()) { |
2662 | AllocaSize = ConstantInt::get(Ty: Type::getInt64Ty(C&: AI->getContext()), |
2663 | V: AllocaArraySize * AllocaTypeSize); |
2664 | } |
2665 | } |
2666 | |
2667 | builder.CreateLifetimeStart(Ptr: AI, Size: AllocaSize); |
2668 | for (ReturnInst *RI : Returns) { |
2669 | // Don't insert llvm.lifetime.end calls between a musttail or deoptimize |
2670 | // call and a return. The return kills all local allocas. |
2671 | if (InlinedMustTailCalls && |
2672 | RI->getParent()->getTerminatingMustTailCall()) |
2673 | continue; |
2674 | if (InlinedDeoptimizeCalls && |
2675 | RI->getParent()->getTerminatingDeoptimizeCall()) |
2676 | continue; |
2677 | IRBuilder<>(RI).CreateLifetimeEnd(Ptr: AI, Size: AllocaSize); |
2678 | } |
2679 | } |
2680 | } |
2681 | |
2682 | // If the inlined code contained dynamic alloca instructions, wrap the inlined |
2683 | // code with llvm.stacksave/llvm.stackrestore intrinsics. |
2684 | if (InlinedFunctionInfo.ContainsDynamicAllocas) { |
2685 | // Insert the llvm.stacksave. |
2686 | CallInst *SavedPtr = IRBuilder<>(&*FirstNewBlock, FirstNewBlock->begin()) |
2687 | .CreateStackSave(Name: "savedstack" ); |
2688 | |
2689 | // Insert a call to llvm.stackrestore before any return instructions in the |
2690 | // inlined function. |
2691 | for (ReturnInst *RI : Returns) { |
2692 | // Don't insert llvm.stackrestore calls between a musttail or deoptimize |
2693 | // call and a return. The return will restore the stack pointer. |
2694 | if (InlinedMustTailCalls && RI->getParent()->getTerminatingMustTailCall()) |
2695 | continue; |
2696 | if (InlinedDeoptimizeCalls && RI->getParent()->getTerminatingDeoptimizeCall()) |
2697 | continue; |
2698 | IRBuilder<>(RI).CreateStackRestore(Ptr: SavedPtr); |
2699 | } |
2700 | } |
2701 | |
2702 | // If we are inlining for an invoke instruction, we must make sure to rewrite |
2703 | // any call instructions into invoke instructions. This is sensitive to which |
2704 | // funclet pads were top-level in the inlinee, so must be done before |
2705 | // rewriting the "parent pad" links. |
2706 | if (auto *II = dyn_cast<InvokeInst>(Val: &CB)) { |
2707 | BasicBlock *UnwindDest = II->getUnwindDest(); |
2708 | Instruction *FirstNonPHI = UnwindDest->getFirstNonPHI(); |
2709 | if (isa<LandingPadInst>(Val: FirstNonPHI)) { |
2710 | HandleInlinedLandingPad(II, FirstNewBlock: &*FirstNewBlock, InlinedCodeInfo&: InlinedFunctionInfo); |
2711 | } else { |
2712 | HandleInlinedEHPad(II, FirstNewBlock: &*FirstNewBlock, InlinedCodeInfo&: InlinedFunctionInfo); |
2713 | } |
2714 | } |
2715 | |
2716 | // Update the lexical scopes of the new funclets and callsites. |
2717 | // Anything that had 'none' as its parent is now nested inside the callsite's |
2718 | // EHPad. |
2719 | if (CallSiteEHPad) { |
2720 | for (Function::iterator BB = FirstNewBlock->getIterator(), |
2721 | E = Caller->end(); |
2722 | BB != E; ++BB) { |
2723 | // Add bundle operands to inlined call sites. |
2724 | PropagateOperandBundles(InlinedBB: BB, CallSiteEHPad); |
2725 | |
2726 | // It is problematic if the inlinee has a cleanupret which unwinds to |
2727 | // caller and we inline it into a call site which doesn't unwind but into |
2728 | // an EH pad that does. Such an edge must be dynamically unreachable. |
2729 | // As such, we replace the cleanupret with unreachable. |
2730 | if (auto *CleanupRet = dyn_cast<CleanupReturnInst>(Val: BB->getTerminator())) |
2731 | if (CleanupRet->unwindsToCaller() && EHPadForCallUnwindsLocally) |
2732 | changeToUnreachable(I: CleanupRet); |
2733 | |
2734 | Instruction *I = BB->getFirstNonPHI(); |
2735 | if (!I->isEHPad()) |
2736 | continue; |
2737 | |
2738 | if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(Val: I)) { |
2739 | if (isa<ConstantTokenNone>(Val: CatchSwitch->getParentPad())) |
2740 | CatchSwitch->setParentPad(CallSiteEHPad); |
2741 | } else { |
2742 | auto *FPI = cast<FuncletPadInst>(Val: I); |
2743 | if (isa<ConstantTokenNone>(Val: FPI->getParentPad())) |
2744 | FPI->setParentPad(CallSiteEHPad); |
2745 | } |
2746 | } |
2747 | } |
2748 | |
2749 | if (InlinedDeoptimizeCalls) { |
2750 | // We need to at least remove the deoptimizing returns from the Return set, |
2751 | // so that the control flow from those returns does not get merged into the |
2752 | // caller (but terminate it instead). If the caller's return type does not |
2753 | // match the callee's return type, we also need to change the return type of |
2754 | // the intrinsic. |
2755 | if (Caller->getReturnType() == CB.getType()) { |
2756 | llvm::erase_if(C&: Returns, P: [](ReturnInst *RI) { |
2757 | return RI->getParent()->getTerminatingDeoptimizeCall() != nullptr; |
2758 | }); |
2759 | } else { |
2760 | SmallVector<ReturnInst *, 8> NormalReturns; |
2761 | Function *NewDeoptIntrinsic = Intrinsic::getDeclaration( |
2762 | M: Caller->getParent(), id: Intrinsic::experimental_deoptimize, |
2763 | Tys: {Caller->getReturnType()}); |
2764 | |
2765 | for (ReturnInst *RI : Returns) { |
2766 | CallInst *DeoptCall = RI->getParent()->getTerminatingDeoptimizeCall(); |
2767 | if (!DeoptCall) { |
2768 | NormalReturns.push_back(Elt: RI); |
2769 | continue; |
2770 | } |
2771 | |
2772 | // The calling convention on the deoptimize call itself may be bogus, |
2773 | // since the code we're inlining may have undefined behavior (and may |
2774 | // never actually execute at runtime); but all |
2775 | // @llvm.experimental.deoptimize declarations have to have the same |
2776 | // calling convention in a well-formed module. |
2777 | auto CallingConv = DeoptCall->getCalledFunction()->getCallingConv(); |
2778 | NewDeoptIntrinsic->setCallingConv(CallingConv); |
2779 | auto *CurBB = RI->getParent(); |
2780 | RI->eraseFromParent(); |
2781 | |
2782 | SmallVector<Value *, 4> CallArgs(DeoptCall->args()); |
2783 | |
2784 | SmallVector<OperandBundleDef, 1> OpBundles; |
2785 | DeoptCall->getOperandBundlesAsDefs(Defs&: OpBundles); |
2786 | auto DeoptAttributes = DeoptCall->getAttributes(); |
2787 | DeoptCall->eraseFromParent(); |
2788 | assert(!OpBundles.empty() && |
2789 | "Expected at least the deopt operand bundle" ); |
2790 | |
2791 | IRBuilder<> Builder(CurBB); |
2792 | CallInst *NewDeoptCall = |
2793 | Builder.CreateCall(Callee: NewDeoptIntrinsic, Args: CallArgs, OpBundles); |
2794 | NewDeoptCall->setCallingConv(CallingConv); |
2795 | NewDeoptCall->setAttributes(DeoptAttributes); |
2796 | if (NewDeoptCall->getType()->isVoidTy()) |
2797 | Builder.CreateRetVoid(); |
2798 | else |
2799 | Builder.CreateRet(V: NewDeoptCall); |
2800 | // Since the ret type is changed, remove the incompatible attributes. |
2801 | NewDeoptCall->removeRetAttrs( |
2802 | AttrsToRemove: AttributeFuncs::typeIncompatible(Ty: NewDeoptCall->getType())); |
2803 | } |
2804 | |
2805 | // Leave behind the normal returns so we can merge control flow. |
2806 | std::swap(LHS&: Returns, RHS&: NormalReturns); |
2807 | } |
2808 | } |
2809 | |
2810 | // Handle any inlined musttail call sites. In order for a new call site to be |
2811 | // musttail, the source of the clone and the inlined call site must have been |
2812 | // musttail. Therefore it's safe to return without merging control into the |
2813 | // phi below. |
2814 | if (InlinedMustTailCalls) { |
2815 | // Check if we need to bitcast the result of any musttail calls. |
2816 | Type *NewRetTy = Caller->getReturnType(); |
2817 | bool NeedBitCast = !CB.use_empty() && CB.getType() != NewRetTy; |
2818 | |
2819 | // Handle the returns preceded by musttail calls separately. |
2820 | SmallVector<ReturnInst *, 8> NormalReturns; |
2821 | for (ReturnInst *RI : Returns) { |
2822 | CallInst *ReturnedMustTail = |
2823 | RI->getParent()->getTerminatingMustTailCall(); |
2824 | if (!ReturnedMustTail) { |
2825 | NormalReturns.push_back(Elt: RI); |
2826 | continue; |
2827 | } |
2828 | if (!NeedBitCast) |
2829 | continue; |
2830 | |
2831 | // Delete the old return and any preceding bitcast. |
2832 | BasicBlock *CurBB = RI->getParent(); |
2833 | auto *OldCast = dyn_cast_or_null<BitCastInst>(Val: RI->getReturnValue()); |
2834 | RI->eraseFromParent(); |
2835 | if (OldCast) |
2836 | OldCast->eraseFromParent(); |
2837 | |
2838 | // Insert a new bitcast and return with the right type. |
2839 | IRBuilder<> Builder(CurBB); |
2840 | Builder.CreateRet(V: Builder.CreateBitCast(V: ReturnedMustTail, DestTy: NewRetTy)); |
2841 | } |
2842 | |
2843 | // Leave behind the normal returns so we can merge control flow. |
2844 | std::swap(LHS&: Returns, RHS&: NormalReturns); |
2845 | } |
2846 | |
2847 | // Now that all of the transforms on the inlined code have taken place but |
2848 | // before we splice the inlined code into the CFG and lose track of which |
2849 | // blocks were actually inlined, collect the call sites. We only do this if |
2850 | // call graph updates weren't requested, as those provide value handle based |
2851 | // tracking of inlined call sites instead. Calls to intrinsics are not |
2852 | // collected because they are not inlineable. |
2853 | if (InlinedFunctionInfo.ContainsCalls) { |
2854 | // Otherwise just collect the raw call sites that were inlined. |
2855 | for (BasicBlock &NewBB : |
2856 | make_range(x: FirstNewBlock->getIterator(), y: Caller->end())) |
2857 | for (Instruction &I : NewBB) |
2858 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) |
2859 | if (!(CB->getCalledFunction() && |
2860 | CB->getCalledFunction()->isIntrinsic())) |
2861 | IFI.InlinedCallSites.push_back(Elt: CB); |
2862 | } |
2863 | |
2864 | // If we cloned in _exactly one_ basic block, and if that block ends in a |
2865 | // return instruction, we splice the body of the inlined callee directly into |
2866 | // the calling basic block. |
2867 | if (Returns.size() == 1 && std::distance(first: FirstNewBlock, last: Caller->end()) == 1) { |
2868 | // Move all of the instructions right before the call. |
2869 | OrigBB->splice(ToIt: CB.getIterator(), FromBB: &*FirstNewBlock, FromBeginIt: FirstNewBlock->begin(), |
2870 | FromEndIt: FirstNewBlock->end()); |
2871 | // Remove the cloned basic block. |
2872 | Caller->back().eraseFromParent(); |
2873 | |
2874 | // If the call site was an invoke instruction, add a branch to the normal |
2875 | // destination. |
2876 | if (InvokeInst *II = dyn_cast<InvokeInst>(Val: &CB)) { |
2877 | BranchInst *NewBr = BranchInst::Create(IfTrue: II->getNormalDest(), InsertBefore: CB.getIterator()); |
2878 | NewBr->setDebugLoc(Returns[0]->getDebugLoc()); |
2879 | } |
2880 | |
2881 | // If the return instruction returned a value, replace uses of the call with |
2882 | // uses of the returned value. |
2883 | if (!CB.use_empty()) { |
2884 | ReturnInst *R = Returns[0]; |
2885 | if (&CB == R->getReturnValue()) |
2886 | CB.replaceAllUsesWith(V: PoisonValue::get(T: CB.getType())); |
2887 | else |
2888 | CB.replaceAllUsesWith(V: R->getReturnValue()); |
2889 | } |
2890 | // Since we are now done with the Call/Invoke, we can delete it. |
2891 | CB.eraseFromParent(); |
2892 | |
2893 | // Since we are now done with the return instruction, delete it also. |
2894 | Returns[0]->eraseFromParent(); |
2895 | |
2896 | if (MergeAttributes) |
2897 | AttributeFuncs::mergeAttributesForInlining(Caller&: *Caller, Callee: *CalledFunc); |
2898 | |
2899 | // We are now done with the inlining. |
2900 | return InlineResult::success(); |
2901 | } |
2902 | |
2903 | // Otherwise, we have the normal case, of more than one block to inline or |
2904 | // multiple return sites. |
2905 | |
2906 | // We want to clone the entire callee function into the hole between the |
2907 | // "starter" and "ender" blocks. How we accomplish this depends on whether |
2908 | // this is an invoke instruction or a call instruction. |
2909 | BasicBlock *AfterCallBB; |
2910 | BranchInst *CreatedBranchToNormalDest = nullptr; |
2911 | if (InvokeInst *II = dyn_cast<InvokeInst>(Val: &CB)) { |
2912 | |
2913 | // Add an unconditional branch to make this look like the CallInst case... |
2914 | CreatedBranchToNormalDest = BranchInst::Create(IfTrue: II->getNormalDest(), InsertBefore: CB.getIterator()); |
2915 | |
2916 | // Split the basic block. This guarantees that no PHI nodes will have to be |
2917 | // updated due to new incoming edges, and make the invoke case more |
2918 | // symmetric to the call case. |
2919 | AfterCallBB = |
2920 | OrigBB->splitBasicBlock(I: CreatedBranchToNormalDest->getIterator(), |
2921 | BBName: CalledFunc->getName() + ".exit" ); |
2922 | |
2923 | } else { // It's a call |
2924 | // If this is a call instruction, we need to split the basic block that |
2925 | // the call lives in. |
2926 | // |
2927 | AfterCallBB = OrigBB->splitBasicBlock(I: CB.getIterator(), |
2928 | BBName: CalledFunc->getName() + ".exit" ); |
2929 | } |
2930 | |
2931 | if (IFI.CallerBFI) { |
2932 | // Copy original BB's block frequency to AfterCallBB |
2933 | IFI.CallerBFI->setBlockFreq(BB: AfterCallBB, |
2934 | Freq: IFI.CallerBFI->getBlockFreq(BB: OrigBB)); |
2935 | } |
2936 | |
2937 | // Change the branch that used to go to AfterCallBB to branch to the first |
2938 | // basic block of the inlined function. |
2939 | // |
2940 | Instruction *Br = OrigBB->getTerminator(); |
2941 | assert(Br && Br->getOpcode() == Instruction::Br && |
2942 | "splitBasicBlock broken!" ); |
2943 | Br->setOperand(i: 0, Val: &*FirstNewBlock); |
2944 | |
2945 | // Now that the function is correct, make it a little bit nicer. In |
2946 | // particular, move the basic blocks inserted from the end of the function |
2947 | // into the space made by splitting the source basic block. |
2948 | Caller->splice(ToIt: AfterCallBB->getIterator(), FromF: Caller, FromBeginIt: FirstNewBlock, |
2949 | FromEndIt: Caller->end()); |
2950 | |
2951 | // Handle all of the return instructions that we just cloned in, and eliminate |
2952 | // any users of the original call/invoke instruction. |
2953 | Type *RTy = CalledFunc->getReturnType(); |
2954 | |
2955 | PHINode *PHI = nullptr; |
2956 | if (Returns.size() > 1) { |
2957 | // The PHI node should go at the front of the new basic block to merge all |
2958 | // possible incoming values. |
2959 | if (!CB.use_empty()) { |
2960 | PHI = PHINode::Create(Ty: RTy, NumReservedValues: Returns.size(), NameStr: CB.getName()); |
2961 | PHI->insertBefore(InsertPos: AfterCallBB->begin()); |
2962 | // Anything that used the result of the function call should now use the |
2963 | // PHI node as their operand. |
2964 | CB.replaceAllUsesWith(V: PHI); |
2965 | } |
2966 | |
2967 | // Loop over all of the return instructions adding entries to the PHI node |
2968 | // as appropriate. |
2969 | if (PHI) { |
2970 | for (ReturnInst *RI : Returns) { |
2971 | assert(RI->getReturnValue()->getType() == PHI->getType() && |
2972 | "Ret value not consistent in function!" ); |
2973 | PHI->addIncoming(V: RI->getReturnValue(), BB: RI->getParent()); |
2974 | } |
2975 | } |
2976 | |
2977 | // Add a branch to the merge points and remove return instructions. |
2978 | DebugLoc Loc; |
2979 | for (ReturnInst *RI : Returns) { |
2980 | BranchInst *BI = BranchInst::Create(IfTrue: AfterCallBB, InsertBefore: RI->getIterator()); |
2981 | Loc = RI->getDebugLoc(); |
2982 | BI->setDebugLoc(Loc); |
2983 | RI->eraseFromParent(); |
2984 | } |
2985 | // We need to set the debug location to *somewhere* inside the |
2986 | // inlined function. The line number may be nonsensical, but the |
2987 | // instruction will at least be associated with the right |
2988 | // function. |
2989 | if (CreatedBranchToNormalDest) |
2990 | CreatedBranchToNormalDest->setDebugLoc(Loc); |
2991 | } else if (!Returns.empty()) { |
2992 | // Otherwise, if there is exactly one return value, just replace anything |
2993 | // using the return value of the call with the computed value. |
2994 | if (!CB.use_empty()) { |
2995 | if (&CB == Returns[0]->getReturnValue()) |
2996 | CB.replaceAllUsesWith(V: PoisonValue::get(T: CB.getType())); |
2997 | else |
2998 | CB.replaceAllUsesWith(V: Returns[0]->getReturnValue()); |
2999 | } |
3000 | |
3001 | // Update PHI nodes that use the ReturnBB to use the AfterCallBB. |
3002 | BasicBlock *ReturnBB = Returns[0]->getParent(); |
3003 | ReturnBB->replaceAllUsesWith(V: AfterCallBB); |
3004 | |
3005 | // Splice the code from the return block into the block that it will return |
3006 | // to, which contains the code that was after the call. |
3007 | AfterCallBB->splice(ToIt: AfterCallBB->begin(), FromBB: ReturnBB); |
3008 | |
3009 | if (CreatedBranchToNormalDest) |
3010 | CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc()); |
3011 | |
3012 | // Delete the return instruction now and empty ReturnBB now. |
3013 | Returns[0]->eraseFromParent(); |
3014 | ReturnBB->eraseFromParent(); |
3015 | } else if (!CB.use_empty()) { |
3016 | // No returns, but something is using the return value of the call. Just |
3017 | // nuke the result. |
3018 | CB.replaceAllUsesWith(V: PoisonValue::get(T: CB.getType())); |
3019 | } |
3020 | |
3021 | // Since we are now done with the Call/Invoke, we can delete it. |
3022 | CB.eraseFromParent(); |
3023 | |
3024 | // If we inlined any musttail calls and the original return is now |
3025 | // unreachable, delete it. It can only contain a bitcast and ret. |
3026 | if (InlinedMustTailCalls && pred_empty(BB: AfterCallBB)) |
3027 | AfterCallBB->eraseFromParent(); |
3028 | |
3029 | // We should always be able to fold the entry block of the function into the |
3030 | // single predecessor of the block... |
3031 | assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!" ); |
3032 | BasicBlock *CalleeEntry = cast<BranchInst>(Val: Br)->getSuccessor(i: 0); |
3033 | |
3034 | // Splice the code entry block into calling block, right before the |
3035 | // unconditional branch. |
3036 | CalleeEntry->replaceAllUsesWith(V: OrigBB); // Update PHI nodes |
3037 | OrigBB->splice(ToIt: Br->getIterator(), FromBB: CalleeEntry); |
3038 | |
3039 | // Remove the unconditional branch. |
3040 | Br->eraseFromParent(); |
3041 | |
3042 | // Now we can remove the CalleeEntry block, which is now empty. |
3043 | CalleeEntry->eraseFromParent(); |
3044 | |
3045 | // If we inserted a phi node, check to see if it has a single value (e.g. all |
3046 | // the entries are the same or undef). If so, remove the PHI so it doesn't |
3047 | // block other optimizations. |
3048 | if (PHI) { |
3049 | AssumptionCache *AC = |
3050 | IFI.GetAssumptionCache ? &IFI.GetAssumptionCache(*Caller) : nullptr; |
3051 | auto &DL = Caller->getDataLayout(); |
3052 | if (Value *V = simplifyInstruction(I: PHI, Q: {DL, nullptr, nullptr, AC})) { |
3053 | PHI->replaceAllUsesWith(V); |
3054 | PHI->eraseFromParent(); |
3055 | } |
3056 | } |
3057 | |
3058 | if (MergeAttributes) |
3059 | AttributeFuncs::mergeAttributesForInlining(Caller&: *Caller, Callee: *CalledFunc); |
3060 | |
3061 | return InlineResult::success(); |
3062 | } |
3063 | |