1 | //===- FunctionAttrs.cpp - Pass which marks functions attributes ----------===// |
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 | /// \file |
10 | /// This file implements interprocedural passes which walk the |
11 | /// call-graph deducing and/or propagating function attributes. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "llvm/Transforms/IPO/FunctionAttrs.h" |
16 | #include "llvm/ADT/ArrayRef.h" |
17 | #include "llvm/ADT/DenseMap.h" |
18 | #include "llvm/ADT/PostOrderIterator.h" |
19 | #include "llvm/ADT/SCCIterator.h" |
20 | #include "llvm/ADT/STLExtras.h" |
21 | #include "llvm/ADT/SetVector.h" |
22 | #include "llvm/ADT/SmallPtrSet.h" |
23 | #include "llvm/ADT/SmallSet.h" |
24 | #include "llvm/ADT/SmallVector.h" |
25 | #include "llvm/ADT/Statistic.h" |
26 | #include "llvm/Analysis/AssumptionCache.h" |
27 | #include "llvm/Analysis/BasicAliasAnalysis.h" |
28 | #include "llvm/Analysis/CFG.h" |
29 | #include "llvm/Analysis/CGSCCPassManager.h" |
30 | #include "llvm/Analysis/CallGraph.h" |
31 | #include "llvm/Analysis/CallGraphSCCPass.h" |
32 | #include "llvm/Analysis/CaptureTracking.h" |
33 | #include "llvm/Analysis/LazyCallGraph.h" |
34 | #include "llvm/Analysis/MemoryLocation.h" |
35 | #include "llvm/Analysis/ValueTracking.h" |
36 | #include "llvm/IR/Argument.h" |
37 | #include "llvm/IR/Attributes.h" |
38 | #include "llvm/IR/BasicBlock.h" |
39 | #include "llvm/IR/Constant.h" |
40 | #include "llvm/IR/ConstantRangeList.h" |
41 | #include "llvm/IR/Constants.h" |
42 | #include "llvm/IR/Function.h" |
43 | #include "llvm/IR/InstIterator.h" |
44 | #include "llvm/IR/InstrTypes.h" |
45 | #include "llvm/IR/Instruction.h" |
46 | #include "llvm/IR/Instructions.h" |
47 | #include "llvm/IR/IntrinsicInst.h" |
48 | #include "llvm/IR/Metadata.h" |
49 | #include "llvm/IR/ModuleSummaryIndex.h" |
50 | #include "llvm/IR/PassManager.h" |
51 | #include "llvm/IR/Type.h" |
52 | #include "llvm/IR/Use.h" |
53 | #include "llvm/IR/User.h" |
54 | #include "llvm/IR/Value.h" |
55 | #include "llvm/Support/Casting.h" |
56 | #include "llvm/Support/CommandLine.h" |
57 | #include "llvm/Support/Compiler.h" |
58 | #include "llvm/Support/Debug.h" |
59 | #include "llvm/Support/ErrorHandling.h" |
60 | #include "llvm/Support/raw_ostream.h" |
61 | #include "llvm/Transforms/IPO.h" |
62 | #include "llvm/Transforms/Utils/Local.h" |
63 | #include <cassert> |
64 | #include <iterator> |
65 | #include <map> |
66 | #include <optional> |
67 | #include <vector> |
68 | |
69 | using namespace llvm; |
70 | |
71 | #define DEBUG_TYPE "function-attrs" |
72 | |
73 | STATISTIC(NumMemoryAttr, "Number of functions with improved memory attribute" ); |
74 | STATISTIC(NumCapturesNone, "Number of arguments marked captures(none)" ); |
75 | STATISTIC(NumCapturesPartial, "Number of arguments marked with captures " |
76 | "attribute other than captures(none)" ); |
77 | STATISTIC(NumReturned, "Number of arguments marked returned" ); |
78 | STATISTIC(NumReadNoneArg, "Number of arguments marked readnone" ); |
79 | STATISTIC(NumReadOnlyArg, "Number of arguments marked readonly" ); |
80 | STATISTIC(NumWriteOnlyArg, "Number of arguments marked writeonly" ); |
81 | STATISTIC(NumNoAlias, "Number of function returns marked noalias" ); |
82 | STATISTIC(NumNonNullReturn, "Number of function returns marked nonnull" ); |
83 | STATISTIC(NumNoUndefReturn, "Number of function returns marked noundef" ); |
84 | STATISTIC(NumNoRecurse, "Number of functions marked as norecurse" ); |
85 | STATISTIC(NumNoUnwind, "Number of functions marked as nounwind" ); |
86 | STATISTIC(NumNoFree, "Number of functions marked as nofree" ); |
87 | STATISTIC(NumWillReturn, "Number of functions marked as willreturn" ); |
88 | STATISTIC(NumNoSync, "Number of functions marked as nosync" ); |
89 | STATISTIC(NumCold, "Number of functions marked as cold" ); |
90 | |
91 | STATISTIC(NumThinLinkNoRecurse, |
92 | "Number of functions marked as norecurse during thinlink" ); |
93 | STATISTIC(NumThinLinkNoUnwind, |
94 | "Number of functions marked as nounwind during thinlink" ); |
95 | |
96 | static cl::opt<bool> EnableNonnullArgPropagation( |
97 | "enable-nonnull-arg-prop" , cl::init(Val: true), cl::Hidden, |
98 | cl::desc("Try to propagate nonnull argument attributes from callsites to " |
99 | "caller functions." )); |
100 | |
101 | static cl::opt<bool> DisableNoUnwindInference( |
102 | "disable-nounwind-inference" , cl::Hidden, |
103 | cl::desc("Stop inferring nounwind attribute during function-attrs pass" )); |
104 | |
105 | static cl::opt<bool> DisableNoFreeInference( |
106 | "disable-nofree-inference" , cl::Hidden, |
107 | cl::desc("Stop inferring nofree attribute during function-attrs pass" )); |
108 | |
109 | static cl::opt<bool> DisableThinLTOPropagation( |
110 | "disable-thinlto-funcattrs" , cl::init(Val: true), cl::Hidden, |
111 | cl::desc("Don't propagate function-attrs in thinLTO" )); |
112 | |
113 | static void addCapturesStat(CaptureInfo CI) { |
114 | if (capturesNothing(CC: CI)) |
115 | ++NumCapturesNone; |
116 | else |
117 | ++NumCapturesPartial; |
118 | } |
119 | |
120 | namespace { |
121 | |
122 | using SCCNodeSet = SmallSetVector<Function *, 8>; |
123 | |
124 | } // end anonymous namespace |
125 | |
126 | static void addLocAccess(MemoryEffects &ME, const MemoryLocation &Loc, |
127 | ModRefInfo MR, AAResults &AAR) { |
128 | // Ignore accesses to known-invariant or local memory. |
129 | MR &= AAR.getModRefInfoMask(Loc, /*IgnoreLocal=*/IgnoreLocals: true); |
130 | if (isNoModRef(MRI: MR)) |
131 | return; |
132 | |
133 | const Value *UO = getUnderlyingObjectAggressive(V: Loc.Ptr); |
134 | if (isa<AllocaInst>(Val: UO)) |
135 | return; |
136 | if (isa<Argument>(Val: UO)) { |
137 | ME |= MemoryEffects::argMemOnly(MR); |
138 | return; |
139 | } |
140 | |
141 | // If it's not an identified object, it might be an argument. |
142 | if (!isIdentifiedObject(V: UO)) |
143 | ME |= MemoryEffects::argMemOnly(MR); |
144 | ME |= MemoryEffects(IRMemLocation::ErrnoMem, MR); |
145 | ME |= MemoryEffects(IRMemLocation::Other, MR); |
146 | } |
147 | |
148 | static void addArgLocs(MemoryEffects &ME, const CallBase *Call, |
149 | ModRefInfo ArgMR, AAResults &AAR) { |
150 | for (const Value *Arg : Call->args()) { |
151 | if (!Arg->getType()->isPtrOrPtrVectorTy()) |
152 | continue; |
153 | |
154 | addLocAccess(ME, |
155 | Loc: MemoryLocation::getBeforeOrAfter(Ptr: Arg, AATags: Call->getAAMetadata()), |
156 | MR: ArgMR, AAR); |
157 | } |
158 | } |
159 | |
160 | /// Returns the memory access attribute for function F using AAR for AA results, |
161 | /// where SCCNodes is the current SCC. |
162 | /// |
163 | /// If ThisBody is true, this function may examine the function body and will |
164 | /// return a result pertaining to this copy of the function. If it is false, the |
165 | /// result will be based only on AA results for the function declaration; it |
166 | /// will be assumed that some other (perhaps less optimized) version of the |
167 | /// function may be selected at link time. |
168 | /// |
169 | /// The return value is split into two parts: Memory effects that always apply, |
170 | /// and additional memory effects that apply if any of the functions in the SCC |
171 | /// can access argmem. |
172 | static std::pair<MemoryEffects, MemoryEffects> |
173 | checkFunctionMemoryAccess(Function &F, bool ThisBody, AAResults &AAR, |
174 | const SCCNodeSet &SCCNodes) { |
175 | MemoryEffects OrigME = AAR.getMemoryEffects(F: &F); |
176 | if (OrigME.doesNotAccessMemory()) |
177 | // Already perfect! |
178 | return {OrigME, MemoryEffects::none()}; |
179 | |
180 | if (!ThisBody) |
181 | return {OrigME, MemoryEffects::none()}; |
182 | |
183 | MemoryEffects ME = MemoryEffects::none(); |
184 | // Additional locations accessed if the SCC accesses argmem. |
185 | MemoryEffects RecursiveArgME = MemoryEffects::none(); |
186 | |
187 | // Inalloca and preallocated arguments are always clobbered by the call. |
188 | if (F.getAttributes().hasAttrSomewhere(Kind: Attribute::InAlloca) || |
189 | F.getAttributes().hasAttrSomewhere(Kind: Attribute::Preallocated)) |
190 | ME |= MemoryEffects::argMemOnly(MR: ModRefInfo::ModRef); |
191 | |
192 | // Scan the function body for instructions that may read or write memory. |
193 | for (Instruction &I : instructions(F)) { |
194 | // Some instructions can be ignored even if they read or write memory. |
195 | // Detect these now, skipping to the next instruction if one is found. |
196 | if (auto *Call = dyn_cast<CallBase>(Val: &I)) { |
197 | // We can optimistically ignore calls to functions in the same SCC, with |
198 | // two caveats: |
199 | // * Calls with operand bundles may have additional effects. |
200 | // * Argument memory accesses may imply additional effects depending on |
201 | // what the argument location is. |
202 | if (!Call->hasOperandBundles() && Call->getCalledFunction() && |
203 | SCCNodes.count(key: Call->getCalledFunction())) { |
204 | // Keep track of which additional locations are accessed if the SCC |
205 | // turns out to access argmem. |
206 | addArgLocs(ME&: RecursiveArgME, Call, ArgMR: ModRefInfo::ModRef, AAR); |
207 | continue; |
208 | } |
209 | |
210 | MemoryEffects CallME = AAR.getMemoryEffects(Call); |
211 | |
212 | // If the call doesn't access memory, we're done. |
213 | if (CallME.doesNotAccessMemory()) |
214 | continue; |
215 | |
216 | // A pseudo probe call shouldn't change any function attribute since it |
217 | // doesn't translate to a real instruction. It comes with a memory access |
218 | // tag to prevent itself being removed by optimizations and not block |
219 | // other instructions being optimized. |
220 | if (isa<PseudoProbeInst>(Val: I)) |
221 | continue; |
222 | |
223 | // Merge callee's memory effects into caller's ones, including |
224 | // inaccessible and errno memory, but excluding argument memory, which is |
225 | // handled separately. |
226 | ME |= CallME.getWithoutLoc(Loc: IRMemLocation::ArgMem); |
227 | |
228 | // If the call accesses captured memory (currently part of "other") and |
229 | // an argument is captured (currently not tracked), then it may also |
230 | // access argument memory. |
231 | ModRefInfo OtherMR = CallME.getModRef(Loc: IRMemLocation::Other); |
232 | ME |= MemoryEffects::argMemOnly(MR: OtherMR); |
233 | |
234 | // Check whether all pointer arguments point to local memory, and |
235 | // ignore calls that only access local memory. |
236 | ModRefInfo ArgMR = CallME.getModRef(Loc: IRMemLocation::ArgMem); |
237 | if (ArgMR != ModRefInfo::NoModRef) |
238 | addArgLocs(ME, Call, ArgMR, AAR); |
239 | continue; |
240 | } |
241 | |
242 | ModRefInfo MR = ModRefInfo::NoModRef; |
243 | if (I.mayWriteToMemory()) |
244 | MR |= ModRefInfo::Mod; |
245 | if (I.mayReadFromMemory()) |
246 | MR |= ModRefInfo::Ref; |
247 | if (MR == ModRefInfo::NoModRef) |
248 | continue; |
249 | |
250 | std::optional<MemoryLocation> Loc = MemoryLocation::getOrNone(Inst: &I); |
251 | if (!Loc) { |
252 | // If no location is known, conservatively assume anything can be |
253 | // accessed. |
254 | ME |= MemoryEffects(MR); |
255 | continue; |
256 | } |
257 | |
258 | // Volatile operations may access inaccessible memory. |
259 | if (I.isVolatile()) |
260 | ME |= MemoryEffects::inaccessibleMemOnly(MR); |
261 | |
262 | addLocAccess(ME, Loc: *Loc, MR, AAR); |
263 | } |
264 | |
265 | return {OrigME & ME, RecursiveArgME}; |
266 | } |
267 | |
268 | MemoryEffects llvm::computeFunctionBodyMemoryAccess(Function &F, |
269 | AAResults &AAR) { |
270 | return checkFunctionMemoryAccess(F, /*ThisBody=*/true, AAR, SCCNodes: {}).first; |
271 | } |
272 | |
273 | /// Deduce readonly/readnone/writeonly attributes for the SCC. |
274 | template <typename AARGetterT> |
275 | static void addMemoryAttrs(const SCCNodeSet &SCCNodes, AARGetterT &&AARGetter, |
276 | SmallSet<Function *, 8> &Changed) { |
277 | MemoryEffects ME = MemoryEffects::none(); |
278 | MemoryEffects RecursiveArgME = MemoryEffects::none(); |
279 | for (Function *F : SCCNodes) { |
280 | // Call the callable parameter to look up AA results for this function. |
281 | AAResults &AAR = AARGetter(*F); |
282 | // Non-exact function definitions may not be selected at link time, and an |
283 | // alternative version that writes to memory may be selected. See the |
284 | // comment on GlobalValue::isDefinitionExact for more details. |
285 | auto [FnME, FnRecursiveArgME] = |
286 | checkFunctionMemoryAccess(F&: *F, ThisBody: F->hasExactDefinition(), AAR, SCCNodes); |
287 | ME |= FnME; |
288 | RecursiveArgME |= FnRecursiveArgME; |
289 | // Reached bottom of the lattice, we will not be able to improve the result. |
290 | if (ME == MemoryEffects::unknown()) |
291 | return; |
292 | } |
293 | |
294 | // If the SCC accesses argmem, add recursive accesses resulting from that. |
295 | ModRefInfo ArgMR = ME.getModRef(Loc: IRMemLocation::ArgMem); |
296 | if (ArgMR != ModRefInfo::NoModRef) |
297 | ME |= RecursiveArgME & MemoryEffects(ArgMR); |
298 | |
299 | for (Function *F : SCCNodes) { |
300 | MemoryEffects OldME = F->getMemoryEffects(); |
301 | MemoryEffects NewME = ME & OldME; |
302 | if (NewME != OldME) { |
303 | ++NumMemoryAttr; |
304 | F->setMemoryEffects(NewME); |
305 | // Remove conflicting writable attributes. |
306 | if (!isModSet(MRI: NewME.getModRef(Loc: IRMemLocation::ArgMem))) |
307 | for (Argument &A : F->args()) |
308 | A.removeAttr(Kind: Attribute::Writable); |
309 | Changed.insert(Ptr: F); |
310 | } |
311 | } |
312 | } |
313 | |
314 | // Compute definitive function attributes for a function taking into account |
315 | // prevailing definitions and linkage types |
316 | static FunctionSummary *calculatePrevailingSummary( |
317 | ValueInfo VI, |
318 | DenseMap<ValueInfo, FunctionSummary *> &CachedPrevailingSummary, |
319 | function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
320 | IsPrevailing) { |
321 | |
322 | auto [It, Inserted] = CachedPrevailingSummary.try_emplace(Key: VI); |
323 | if (!Inserted) |
324 | return It->second; |
325 | |
326 | /// At this point, prevailing symbols have been resolved. The following leads |
327 | /// to returning a conservative result: |
328 | /// - Multiple instances with local linkage. Normally local linkage would be |
329 | /// unique per module |
330 | /// as the GUID includes the module path. We could have a guid alias if |
331 | /// there wasn't any distinguishing path when each file was compiled, but |
332 | /// that should be rare so we'll punt on those. |
333 | |
334 | /// These next 2 cases should not happen and will assert: |
335 | /// - Multiple instances with external linkage. This should be caught in |
336 | /// symbol resolution |
337 | /// - Non-existent FunctionSummary for Aliasee. This presents a hole in our |
338 | /// knowledge meaning we have to go conservative. |
339 | |
340 | /// Otherwise, we calculate attributes for a function as: |
341 | /// 1. If we have a local linkage, take its attributes. If there's somehow |
342 | /// multiple, bail and go conservative. |
343 | /// 2. If we have an external/WeakODR/LinkOnceODR linkage check that it is |
344 | /// prevailing, take its attributes. |
345 | /// 3. If we have a Weak/LinkOnce linkage the copies can have semantic |
346 | /// differences. However, if the prevailing copy is known it will be used |
347 | /// so take its attributes. If the prevailing copy is in a native file |
348 | /// all IR copies will be dead and propagation will go conservative. |
349 | /// 4. AvailableExternally summaries without a prevailing copy are known to |
350 | /// occur in a couple of circumstances: |
351 | /// a. An internal function gets imported due to its caller getting |
352 | /// imported, it becomes AvailableExternally but no prevailing |
353 | /// definition exists. Because it has to get imported along with its |
354 | /// caller the attributes will be captured by propagating on its |
355 | /// caller. |
356 | /// b. C++11 [temp.explicit]p10 can generate AvailableExternally |
357 | /// definitions of explicitly instanced template declarations |
358 | /// for inlining which are ultimately dropped from the TU. Since this |
359 | /// is localized to the TU the attributes will have already made it to |
360 | /// the callers. |
361 | /// These are edge cases and already captured by their callers so we |
362 | /// ignore these for now. If they become relevant to optimize in the |
363 | /// future this can be revisited. |
364 | /// 5. Otherwise, go conservative. |
365 | |
366 | FunctionSummary *Local = nullptr; |
367 | FunctionSummary *Prevailing = nullptr; |
368 | |
369 | for (const auto &GVS : VI.getSummaryList()) { |
370 | if (!GVS->isLive()) |
371 | continue; |
372 | |
373 | FunctionSummary *FS = dyn_cast<FunctionSummary>(Val: GVS->getBaseObject()); |
374 | // Virtual and Unknown (e.g. indirect) calls require going conservative |
375 | if (!FS || FS->fflags().HasUnknownCall) |
376 | return nullptr; |
377 | |
378 | const auto &Linkage = GVS->linkage(); |
379 | if (GlobalValue::isLocalLinkage(Linkage)) { |
380 | if (Local) { |
381 | LLVM_DEBUG( |
382 | dbgs() |
383 | << "ThinLTO FunctionAttrs: Multiple Local Linkage, bailing on " |
384 | "function " |
385 | << VI.name() << " from " << FS->modulePath() << ". Previous module " |
386 | << Local->modulePath() << "\n" ); |
387 | return nullptr; |
388 | } |
389 | Local = FS; |
390 | } else if (GlobalValue::isExternalLinkage(Linkage)) { |
391 | assert(IsPrevailing(VI.getGUID(), GVS.get())); |
392 | Prevailing = FS; |
393 | break; |
394 | } else if (GlobalValue::isWeakODRLinkage(Linkage) || |
395 | GlobalValue::isLinkOnceODRLinkage(Linkage) || |
396 | GlobalValue::isWeakAnyLinkage(Linkage) || |
397 | GlobalValue::isLinkOnceAnyLinkage(Linkage)) { |
398 | if (IsPrevailing(VI.getGUID(), GVS.get())) { |
399 | Prevailing = FS; |
400 | break; |
401 | } |
402 | } else if (GlobalValue::isAvailableExternallyLinkage(Linkage)) { |
403 | // TODO: Handle these cases if they become meaningful |
404 | continue; |
405 | } |
406 | } |
407 | |
408 | auto &CPS = CachedPrevailingSummary[VI]; |
409 | if (Local) { |
410 | assert(!Prevailing); |
411 | CPS = Local; |
412 | } else if (Prevailing) { |
413 | assert(!Local); |
414 | CPS = Prevailing; |
415 | } |
416 | |
417 | return CPS; |
418 | } |
419 | |
420 | bool llvm::thinLTOPropagateFunctionAttrs( |
421 | ModuleSummaryIndex &Index, |
422 | function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
423 | IsPrevailing) { |
424 | // TODO: implement addNoAliasAttrs once |
425 | // there's more information about the return type in the summary |
426 | if (DisableThinLTOPropagation) |
427 | return false; |
428 | |
429 | DenseMap<ValueInfo, FunctionSummary *> CachedPrevailingSummary; |
430 | bool Changed = false; |
431 | |
432 | auto PropagateAttributes = [&](std::vector<ValueInfo> &SCCNodes) { |
433 | // Assume we can propagate unless we discover otherwise |
434 | FunctionSummary::FFlags InferredFlags; |
435 | InferredFlags.NoRecurse = (SCCNodes.size() == 1); |
436 | InferredFlags.NoUnwind = true; |
437 | |
438 | for (auto &V : SCCNodes) { |
439 | FunctionSummary *CallerSummary = |
440 | calculatePrevailingSummary(VI: V, CachedPrevailingSummary, IsPrevailing); |
441 | |
442 | // Function summaries can fail to contain information such as declarations |
443 | if (!CallerSummary) |
444 | return; |
445 | |
446 | if (CallerSummary->fflags().MayThrow) |
447 | InferredFlags.NoUnwind = false; |
448 | |
449 | for (const auto &Callee : CallerSummary->calls()) { |
450 | FunctionSummary *CalleeSummary = calculatePrevailingSummary( |
451 | VI: Callee.first, CachedPrevailingSummary, IsPrevailing); |
452 | |
453 | if (!CalleeSummary) |
454 | return; |
455 | |
456 | if (!CalleeSummary->fflags().NoRecurse) |
457 | InferredFlags.NoRecurse = false; |
458 | |
459 | if (!CalleeSummary->fflags().NoUnwind) |
460 | InferredFlags.NoUnwind = false; |
461 | |
462 | if (!InferredFlags.NoUnwind && !InferredFlags.NoRecurse) |
463 | break; |
464 | } |
465 | } |
466 | |
467 | if (InferredFlags.NoUnwind || InferredFlags.NoRecurse) { |
468 | Changed = true; |
469 | for (auto &V : SCCNodes) { |
470 | if (InferredFlags.NoRecurse) { |
471 | LLVM_DEBUG(dbgs() << "ThinLTO FunctionAttrs: Propagated NoRecurse to " |
472 | << V.name() << "\n" ); |
473 | ++NumThinLinkNoRecurse; |
474 | } |
475 | |
476 | if (InferredFlags.NoUnwind) { |
477 | LLVM_DEBUG(dbgs() << "ThinLTO FunctionAttrs: Propagated NoUnwind to " |
478 | << V.name() << "\n" ); |
479 | ++NumThinLinkNoUnwind; |
480 | } |
481 | |
482 | for (const auto &S : V.getSummaryList()) { |
483 | if (auto *FS = dyn_cast<FunctionSummary>(Val: S.get())) { |
484 | if (InferredFlags.NoRecurse) |
485 | FS->setNoRecurse(); |
486 | |
487 | if (InferredFlags.NoUnwind) |
488 | FS->setNoUnwind(); |
489 | } |
490 | } |
491 | } |
492 | } |
493 | }; |
494 | |
495 | // Call propagation functions on each SCC in the Index |
496 | for (scc_iterator<ModuleSummaryIndex *> I = scc_begin(G: &Index); !I.isAtEnd(); |
497 | ++I) { |
498 | std::vector<ValueInfo> Nodes(*I); |
499 | PropagateAttributes(Nodes); |
500 | } |
501 | return Changed; |
502 | } |
503 | |
504 | namespace { |
505 | |
506 | /// For a given pointer Argument, this retains a list of Arguments of functions |
507 | /// in the same SCC that the pointer data flows into. We use this to build an |
508 | /// SCC of the arguments. |
509 | struct ArgumentGraphNode { |
510 | Argument *Definition; |
511 | /// CaptureComponents for this argument, excluding captures via Uses. |
512 | /// We don't distinguish between other/return captures here. |
513 | CaptureComponents CC = CaptureComponents::None; |
514 | SmallVector<ArgumentGraphNode *, 4> Uses; |
515 | }; |
516 | |
517 | class ArgumentGraph { |
518 | // We store pointers to ArgumentGraphNode objects, so it's important that |
519 | // that they not move around upon insert. |
520 | using ArgumentMapTy = std::map<Argument *, ArgumentGraphNode>; |
521 | |
522 | ArgumentMapTy ArgumentMap; |
523 | |
524 | // There is no root node for the argument graph, in fact: |
525 | // void f(int *x, int *y) { if (...) f(x, y); } |
526 | // is an example where the graph is disconnected. The SCCIterator requires a |
527 | // single entry point, so we maintain a fake ("synthetic") root node that |
528 | // uses every node. Because the graph is directed and nothing points into |
529 | // the root, it will not participate in any SCCs (except for its own). |
530 | ArgumentGraphNode SyntheticRoot; |
531 | |
532 | public: |
533 | ArgumentGraph() { SyntheticRoot.Definition = nullptr; } |
534 | |
535 | using iterator = SmallVectorImpl<ArgumentGraphNode *>::iterator; |
536 | |
537 | iterator begin() { return SyntheticRoot.Uses.begin(); } |
538 | iterator end() { return SyntheticRoot.Uses.end(); } |
539 | ArgumentGraphNode *getEntryNode() { return &SyntheticRoot; } |
540 | |
541 | ArgumentGraphNode *operator[](Argument *A) { |
542 | ArgumentGraphNode &Node = ArgumentMap[A]; |
543 | Node.Definition = A; |
544 | SyntheticRoot.Uses.push_back(Elt: &Node); |
545 | return &Node; |
546 | } |
547 | }; |
548 | |
549 | /// This tracker checks whether callees are in the SCC, and if so it does not |
550 | /// consider that a capture, instead adding it to the "Uses" list and |
551 | /// continuing with the analysis. |
552 | struct ArgumentUsesTracker : public CaptureTracker { |
553 | ArgumentUsesTracker(const SCCNodeSet &SCCNodes) : SCCNodes(SCCNodes) {} |
554 | |
555 | void tooManyUses() override { CI = CaptureInfo::all(); } |
556 | |
557 | Action captured(const Use *U, UseCaptureInfo UseCI) override { |
558 | if (updateCaptureInfo(U, CC: UseCI.UseCC)) { |
559 | // Don't bother continuing if we already capture everything. |
560 | if (capturesAll(CC: CI.getOtherComponents())) |
561 | return Stop; |
562 | return Continue; |
563 | } |
564 | |
565 | // For SCC argument tracking, we're not going to analyze other/ret |
566 | // components separately, so don't follow the return value. |
567 | return ContinueIgnoringReturn; |
568 | } |
569 | |
570 | bool updateCaptureInfo(const Use *U, CaptureComponents CC) { |
571 | CallBase *CB = dyn_cast<CallBase>(Val: U->getUser()); |
572 | if (!CB) { |
573 | if (isa<ReturnInst>(Val: U->getUser())) |
574 | CI |= CaptureInfo::retOnly(RetComponents: CC); |
575 | else |
576 | // Conservatively assume that the captured value might make its way |
577 | // into the return value as well. This could be made more precise. |
578 | CI |= CaptureInfo(CC); |
579 | return true; |
580 | } |
581 | |
582 | Function *F = CB->getCalledFunction(); |
583 | if (!F || !F->hasExactDefinition() || !SCCNodes.count(key: F)) { |
584 | CI |= CaptureInfo(CC); |
585 | return true; |
586 | } |
587 | |
588 | assert(!CB->isCallee(U) && "callee operand reported captured?" ); |
589 | const unsigned UseIndex = CB->getDataOperandNo(U); |
590 | if (UseIndex >= CB->arg_size()) { |
591 | // Data operand, but not a argument operand -- must be a bundle operand |
592 | assert(CB->hasOperandBundles() && "Must be!" ); |
593 | |
594 | // CaptureTracking told us that we're being captured by an operand bundle |
595 | // use. In this case it does not matter if the callee is within our SCC |
596 | // or not -- we've been captured in some unknown way, and we have to be |
597 | // conservative. |
598 | CI |= CaptureInfo(CC); |
599 | return true; |
600 | } |
601 | |
602 | if (UseIndex >= F->arg_size()) { |
603 | assert(F->isVarArg() && "More params than args in non-varargs call" ); |
604 | CI |= CaptureInfo(CC); |
605 | return true; |
606 | } |
607 | |
608 | // TODO(captures): Could improve precision by remembering maximum |
609 | // capture components for the argument. |
610 | Uses.push_back(Elt: &*std::next(x: F->arg_begin(), n: UseIndex)); |
611 | return false; |
612 | } |
613 | |
614 | // Does not include potential captures via Uses in the SCC. |
615 | CaptureInfo CI = CaptureInfo::none(); |
616 | |
617 | // Uses within our SCC. |
618 | SmallVector<Argument *, 4> Uses; |
619 | |
620 | const SCCNodeSet &SCCNodes; |
621 | }; |
622 | |
623 | /// A struct of argument use: a Use and the offset it accesses. This struct |
624 | /// is to track uses inside function via GEP. If GEP has a non-constant index, |
625 | /// the Offset field is nullopt. |
626 | struct ArgumentUse { |
627 | Use *U; |
628 | std::optional<int64_t> Offset; |
629 | }; |
630 | |
631 | /// A struct of argument access info. "Unknown" accesses are the cases like |
632 | /// unrecognized instructions, instructions that have more than one use of |
633 | /// the argument, or volatile memory accesses. "WriteWithSideEffect" are call |
634 | /// instructions that not only write an argument but also capture it. |
635 | struct ArgumentAccessInfo { |
636 | enum class AccessType : uint8_t { Write, WriteWithSideEffect, Read, Unknown }; |
637 | AccessType ArgAccessType; |
638 | ConstantRangeList AccessRanges; |
639 | }; |
640 | |
641 | /// A struct to wrap the argument use info per block. |
642 | struct UsesPerBlockInfo { |
643 | SmallDenseMap<Instruction *, ArgumentAccessInfo, 4> Insts; |
644 | bool HasWrites = false; |
645 | bool HasUnknownAccess = false; |
646 | }; |
647 | |
648 | /// A struct to summarize the argument use info in a function. |
649 | struct ArgumentUsesSummary { |
650 | bool HasAnyWrite = false; |
651 | bool HasWriteOutsideEntryBB = false; |
652 | SmallDenseMap<const BasicBlock *, UsesPerBlockInfo, 16> UsesPerBlock; |
653 | }; |
654 | |
655 | ArgumentAccessInfo getArgumentAccessInfo(const Instruction *I, |
656 | const ArgumentUse &ArgUse, |
657 | const DataLayout &DL) { |
658 | auto GetTypeAccessRange = |
659 | [&DL](Type *Ty, |
660 | std::optional<int64_t> Offset) -> std::optional<ConstantRange> { |
661 | auto TypeSize = DL.getTypeStoreSize(Ty); |
662 | if (!TypeSize.isScalable() && Offset) { |
663 | int64_t Size = TypeSize.getFixedValue(); |
664 | APInt Low(64, *Offset, true); |
665 | bool Overflow; |
666 | APInt High = Low.sadd_ov(RHS: APInt(64, Size, true), Overflow); |
667 | // Bail if the range overflows signed 64-bit int. |
668 | if (Overflow) |
669 | return std::nullopt; |
670 | return ConstantRange(Low, High); |
671 | } |
672 | return std::nullopt; |
673 | }; |
674 | auto GetConstantIntRange = |
675 | [](Value *Length, |
676 | std::optional<int64_t> Offset) -> std::optional<ConstantRange> { |
677 | auto *ConstantLength = dyn_cast<ConstantInt>(Val: Length); |
678 | if (ConstantLength && Offset) { |
679 | int64_t Len = ConstantLength->getSExtValue(); |
680 | |
681 | // Reject zero or negative lengths |
682 | if (Len <= 0) |
683 | return std::nullopt; |
684 | |
685 | APInt Low(64, *Offset, true); |
686 | bool Overflow; |
687 | APInt High = Low.sadd_ov(RHS: APInt(64, Len, true), Overflow); |
688 | if (Overflow) |
689 | return std::nullopt; |
690 | |
691 | return ConstantRange(Low, High); |
692 | } |
693 | return std::nullopt; |
694 | }; |
695 | |
696 | if (auto *SI = dyn_cast<StoreInst>(Val: I)) { |
697 | if (SI->isSimple() && &SI->getOperandUse(i: 1) == ArgUse.U) { |
698 | // Get the fixed type size of "SI". Since the access range of a write |
699 | // will be unioned, if "SI" doesn't have a fixed type size, we just set |
700 | // the access range to empty. |
701 | ConstantRangeList AccessRanges; |
702 | if (auto TypeAccessRange = |
703 | GetTypeAccessRange(SI->getAccessType(), ArgUse.Offset)) |
704 | AccessRanges.insert(NewRange: *TypeAccessRange); |
705 | return {.ArgAccessType: ArgumentAccessInfo::AccessType::Write, .AccessRanges: std::move(AccessRanges)}; |
706 | } |
707 | } else if (auto *LI = dyn_cast<LoadInst>(Val: I)) { |
708 | if (LI->isSimple()) { |
709 | assert(&LI->getOperandUse(0) == ArgUse.U); |
710 | // Get the fixed type size of "LI". Different from Write, if "LI" |
711 | // doesn't have a fixed type size, we conservatively set as a clobber |
712 | // with an empty access range. |
713 | if (auto TypeAccessRange = |
714 | GetTypeAccessRange(LI->getAccessType(), ArgUse.Offset)) |
715 | return {.ArgAccessType: ArgumentAccessInfo::AccessType::Read, .AccessRanges: {*TypeAccessRange}}; |
716 | } |
717 | } else if (auto *MemSet = dyn_cast<MemSetInst>(Val: I)) { |
718 | if (!MemSet->isVolatile()) { |
719 | ConstantRangeList AccessRanges; |
720 | if (auto AccessRange = |
721 | GetConstantIntRange(MemSet->getLength(), ArgUse.Offset)) |
722 | AccessRanges.insert(NewRange: *AccessRange); |
723 | return {.ArgAccessType: ArgumentAccessInfo::AccessType::Write, .AccessRanges: AccessRanges}; |
724 | } |
725 | } else if (auto *MTI = dyn_cast<MemTransferInst>(Val: I)) { |
726 | if (!MTI->isVolatile()) { |
727 | if (&MTI->getOperandUse(i: 0) == ArgUse.U) { |
728 | ConstantRangeList AccessRanges; |
729 | if (auto AccessRange = |
730 | GetConstantIntRange(MTI->getLength(), ArgUse.Offset)) |
731 | AccessRanges.insert(NewRange: *AccessRange); |
732 | return {.ArgAccessType: ArgumentAccessInfo::AccessType::Write, .AccessRanges: AccessRanges}; |
733 | } else if (&MTI->getOperandUse(i: 1) == ArgUse.U) { |
734 | if (auto AccessRange = |
735 | GetConstantIntRange(MTI->getLength(), ArgUse.Offset)) |
736 | return {.ArgAccessType: ArgumentAccessInfo::AccessType::Read, .AccessRanges: {*AccessRange}}; |
737 | } |
738 | } |
739 | } else if (auto *CB = dyn_cast<CallBase>(Val: I)) { |
740 | if (CB->isArgOperand(U: ArgUse.U) && |
741 | !CB->isByValArgument(ArgNo: CB->getArgOperandNo(U: ArgUse.U))) { |
742 | unsigned ArgNo = CB->getArgOperandNo(U: ArgUse.U); |
743 | bool IsInitialize = CB->paramHasAttr(ArgNo, Kind: Attribute::Initializes); |
744 | if (IsInitialize && ArgUse.Offset) { |
745 | // Argument is a Write when parameter is writeonly/readnone |
746 | // and nocapture. Otherwise, it's a WriteWithSideEffect. |
747 | auto Access = CB->onlyWritesMemory(OpNo: ArgNo) && CB->doesNotCapture(OpNo: ArgNo) |
748 | ? ArgumentAccessInfo::AccessType::Write |
749 | : ArgumentAccessInfo::AccessType::WriteWithSideEffect; |
750 | ConstantRangeList AccessRanges; |
751 | Attribute Attr = CB->getParamAttr(ArgNo, Kind: Attribute::Initializes); |
752 | ConstantRangeList CBCRL = Attr.getValueAsConstantRangeList(); |
753 | for (ConstantRange &CR : CBCRL) |
754 | AccessRanges.insert(NewRange: ConstantRange(CR.getLower() + *ArgUse.Offset, |
755 | CR.getUpper() + *ArgUse.Offset)); |
756 | return {.ArgAccessType: Access, .AccessRanges: AccessRanges}; |
757 | } |
758 | } |
759 | } |
760 | // Other unrecognized instructions are considered as unknown. |
761 | return {.ArgAccessType: ArgumentAccessInfo::AccessType::Unknown, .AccessRanges: {}}; |
762 | } |
763 | |
764 | // Collect the uses of argument "A" in "F". |
765 | ArgumentUsesSummary collectArgumentUsesPerBlock(Argument &A, Function &F) { |
766 | auto &DL = F.getParent()->getDataLayout(); |
767 | unsigned PointerSize = |
768 | DL.getIndexSizeInBits(AS: A.getType()->getPointerAddressSpace()); |
769 | ArgumentUsesSummary Result; |
770 | |
771 | BasicBlock &EntryBB = F.getEntryBlock(); |
772 | SmallVector<ArgumentUse, 4> Worklist; |
773 | for (Use &U : A.uses()) |
774 | Worklist.push_back(Elt: {.U: &U, .Offset: 0}); |
775 | |
776 | // Update "UsesPerBlock" with the block of "I" as key and "Info" as value. |
777 | // Return true if the block of "I" has write accesses after updating. |
778 | auto UpdateUseInfo = [&Result](Instruction *I, ArgumentAccessInfo Info) { |
779 | auto *BB = I->getParent(); |
780 | auto &BBInfo = Result.UsesPerBlock[BB]; |
781 | auto [It, Inserted] = BBInfo.Insts.try_emplace(Key: I); |
782 | auto &IInfo = It->second; |
783 | |
784 | // Instructions that have more than one use of the argument are considered |
785 | // as clobbers. |
786 | if (!Inserted) { |
787 | IInfo = {.ArgAccessType: ArgumentAccessInfo::AccessType::Unknown, .AccessRanges: {}}; |
788 | BBInfo.HasUnknownAccess = true; |
789 | return false; |
790 | } |
791 | |
792 | IInfo = std::move(Info); |
793 | BBInfo.HasUnknownAccess |= |
794 | IInfo.ArgAccessType == ArgumentAccessInfo::AccessType::Unknown; |
795 | bool InfoHasWrites = |
796 | (IInfo.ArgAccessType == ArgumentAccessInfo::AccessType::Write || |
797 | IInfo.ArgAccessType == |
798 | ArgumentAccessInfo::AccessType::WriteWithSideEffect) && |
799 | !IInfo.AccessRanges.empty(); |
800 | BBInfo.HasWrites |= InfoHasWrites; |
801 | return InfoHasWrites; |
802 | }; |
803 | |
804 | // No need for a visited set because we don't look through phis, so there are |
805 | // no cycles. |
806 | while (!Worklist.empty()) { |
807 | ArgumentUse ArgUse = Worklist.pop_back_val(); |
808 | User *U = ArgUse.U->getUser(); |
809 | // Add GEP uses to worklist. |
810 | // If the GEP is not a constant GEP, set the ArgumentUse::Offset to nullopt. |
811 | if (auto *GEP = dyn_cast<GEPOperator>(Val: U)) { |
812 | std::optional<int64_t> NewOffset = std::nullopt; |
813 | if (ArgUse.Offset) { |
814 | APInt Offset(PointerSize, 0); |
815 | if (GEP->accumulateConstantOffset(DL, Offset)) |
816 | NewOffset = *ArgUse.Offset + Offset.getSExtValue(); |
817 | } |
818 | for (Use &U : GEP->uses()) |
819 | Worklist.push_back(Elt: {.U: &U, .Offset: NewOffset}); |
820 | continue; |
821 | } |
822 | |
823 | auto *I = cast<Instruction>(Val: U); |
824 | bool HasWrite = UpdateUseInfo(I, getArgumentAccessInfo(I, ArgUse, DL)); |
825 | |
826 | Result.HasAnyWrite |= HasWrite; |
827 | |
828 | if (HasWrite && I->getParent() != &EntryBB) |
829 | Result.HasWriteOutsideEntryBB = true; |
830 | } |
831 | return Result; |
832 | } |
833 | |
834 | } // end anonymous namespace |
835 | |
836 | namespace llvm { |
837 | |
838 | template <> struct GraphTraits<ArgumentGraphNode *> { |
839 | using NodeRef = ArgumentGraphNode *; |
840 | using ChildIteratorType = SmallVectorImpl<ArgumentGraphNode *>::iterator; |
841 | |
842 | static NodeRef getEntryNode(NodeRef A) { return A; } |
843 | static ChildIteratorType child_begin(NodeRef N) { return N->Uses.begin(); } |
844 | static ChildIteratorType child_end(NodeRef N) { return N->Uses.end(); } |
845 | }; |
846 | |
847 | template <> |
848 | struct GraphTraits<ArgumentGraph *> : public GraphTraits<ArgumentGraphNode *> { |
849 | static NodeRef getEntryNode(ArgumentGraph *AG) { return AG->getEntryNode(); } |
850 | |
851 | static ChildIteratorType nodes_begin(ArgumentGraph *AG) { |
852 | return AG->begin(); |
853 | } |
854 | |
855 | static ChildIteratorType nodes_end(ArgumentGraph *AG) { return AG->end(); } |
856 | }; |
857 | |
858 | } // end namespace llvm |
859 | |
860 | /// Returns Attribute::None, Attribute::ReadOnly or Attribute::ReadNone. |
861 | static Attribute::AttrKind |
862 | determinePointerAccessAttrs(Argument *A, |
863 | const SmallPtrSet<Argument *, 8> &SCCNodes) { |
864 | SmallVector<Use *, 32> Worklist; |
865 | SmallPtrSet<Use *, 32> Visited; |
866 | |
867 | // inalloca arguments are always clobbered by the call. |
868 | if (A->hasInAllocaAttr() || A->hasPreallocatedAttr()) |
869 | return Attribute::None; |
870 | |
871 | bool IsRead = false; |
872 | bool IsWrite = false; |
873 | |
874 | for (Use &U : A->uses()) { |
875 | Visited.insert(Ptr: &U); |
876 | Worklist.push_back(Elt: &U); |
877 | } |
878 | |
879 | while (!Worklist.empty()) { |
880 | if (IsWrite && IsRead) |
881 | // No point in searching further.. |
882 | return Attribute::None; |
883 | |
884 | Use *U = Worklist.pop_back_val(); |
885 | Instruction *I = cast<Instruction>(Val: U->getUser()); |
886 | |
887 | switch (I->getOpcode()) { |
888 | case Instruction::BitCast: |
889 | case Instruction::GetElementPtr: |
890 | case Instruction::PHI: |
891 | case Instruction::Select: |
892 | case Instruction::AddrSpaceCast: |
893 | // The original value is not read/written via this if the new value isn't. |
894 | for (Use &UU : I->uses()) |
895 | if (Visited.insert(Ptr: &UU).second) |
896 | Worklist.push_back(Elt: &UU); |
897 | break; |
898 | |
899 | case Instruction::Call: |
900 | case Instruction::Invoke: { |
901 | CallBase &CB = cast<CallBase>(Val&: *I); |
902 | if (CB.isCallee(U)) { |
903 | IsRead = true; |
904 | // Note that indirect calls do not capture, see comment in |
905 | // CaptureTracking for context |
906 | continue; |
907 | } |
908 | |
909 | // Given we've explicitly handled the callee operand above, what's left |
910 | // must be a data operand (e.g. argument or operand bundle) |
911 | const unsigned UseIndex = CB.getDataOperandNo(U); |
912 | |
913 | // Some intrinsics (for instance ptrmask) do not capture their results, |
914 | // but return results thas alias their pointer argument, and thus should |
915 | // be handled like GEP or addrspacecast above. |
916 | if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( |
917 | Call: &CB, /*MustPreserveNullness=*/false)) { |
918 | for (Use &UU : CB.uses()) |
919 | if (Visited.insert(Ptr: &UU).second) |
920 | Worklist.push_back(Elt: &UU); |
921 | } else if (capturesAnyProvenance(CC: CB.getCaptureInfo(OpNo: UseIndex))) { |
922 | if (!CB.onlyReadsMemory()) |
923 | // If the callee can save a copy into other memory, then simply |
924 | // scanning uses of the call is insufficient. We have no way |
925 | // of tracking copies of the pointer through memory to see |
926 | // if a reloaded copy is written to, thus we must give up. |
927 | return Attribute::None; |
928 | // Push users for processing once we finish this one |
929 | if (!I->getType()->isVoidTy()) |
930 | for (Use &UU : I->uses()) |
931 | if (Visited.insert(Ptr: &UU).second) |
932 | Worklist.push_back(Elt: &UU); |
933 | } |
934 | |
935 | ModRefInfo ArgMR = CB.getMemoryEffects().getModRef(Loc: IRMemLocation::ArgMem); |
936 | if (isNoModRef(MRI: ArgMR)) |
937 | continue; |
938 | |
939 | if (Function *F = CB.getCalledFunction()) |
940 | if (CB.isArgOperand(U) && UseIndex < F->arg_size() && |
941 | SCCNodes.count(Ptr: F->getArg(i: UseIndex))) |
942 | // This is an argument which is part of the speculative SCC. Note |
943 | // that only operands corresponding to formal arguments of the callee |
944 | // can participate in the speculation. |
945 | break; |
946 | |
947 | // The accessors used on call site here do the right thing for calls and |
948 | // invokes with operand bundles. |
949 | if (CB.doesNotAccessMemory(OpNo: UseIndex)) { |
950 | /* nop */ |
951 | } else if (!isModSet(MRI: ArgMR) || CB.onlyReadsMemory(OpNo: UseIndex)) { |
952 | IsRead = true; |
953 | } else if (!isRefSet(MRI: ArgMR) || |
954 | CB.dataOperandHasImpliedAttr(i: UseIndex, Kind: Attribute::WriteOnly)) { |
955 | IsWrite = true; |
956 | } else { |
957 | return Attribute::None; |
958 | } |
959 | break; |
960 | } |
961 | |
962 | case Instruction::Load: |
963 | // A volatile load has side effects beyond what readonly can be relied |
964 | // upon. |
965 | if (cast<LoadInst>(Val: I)->isVolatile()) |
966 | return Attribute::None; |
967 | |
968 | IsRead = true; |
969 | break; |
970 | |
971 | case Instruction::Store: |
972 | if (cast<StoreInst>(Val: I)->getValueOperand() == *U) |
973 | // untrackable capture |
974 | return Attribute::None; |
975 | |
976 | // A volatile store has side effects beyond what writeonly can be relied |
977 | // upon. |
978 | if (cast<StoreInst>(Val: I)->isVolatile()) |
979 | return Attribute::None; |
980 | |
981 | IsWrite = true; |
982 | break; |
983 | |
984 | case Instruction::ICmp: |
985 | case Instruction::Ret: |
986 | break; |
987 | |
988 | default: |
989 | return Attribute::None; |
990 | } |
991 | } |
992 | |
993 | if (IsWrite && IsRead) |
994 | return Attribute::None; |
995 | else if (IsRead) |
996 | return Attribute::ReadOnly; |
997 | else if (IsWrite) |
998 | return Attribute::WriteOnly; |
999 | else |
1000 | return Attribute::ReadNone; |
1001 | } |
1002 | |
1003 | /// Deduce returned attributes for the SCC. |
1004 | static void addArgumentReturnedAttrs(const SCCNodeSet &SCCNodes, |
1005 | SmallSet<Function *, 8> &Changed) { |
1006 | // Check each function in turn, determining if an argument is always returned. |
1007 | for (Function *F : SCCNodes) { |
1008 | // We can infer and propagate function attributes only when we know that the |
1009 | // definition we'll get at link time is *exactly* the definition we see now. |
1010 | // For more details, see GlobalValue::mayBeDerefined. |
1011 | if (!F->hasExactDefinition()) |
1012 | continue; |
1013 | |
1014 | if (F->getReturnType()->isVoidTy()) |
1015 | continue; |
1016 | |
1017 | // There is nothing to do if an argument is already marked as 'returned'. |
1018 | if (F->getAttributes().hasAttrSomewhere(Kind: Attribute::Returned)) |
1019 | continue; |
1020 | |
1021 | auto FindRetArg = [&]() -> Argument * { |
1022 | Argument *RetArg = nullptr; |
1023 | for (BasicBlock &BB : *F) |
1024 | if (auto *Ret = dyn_cast<ReturnInst>(Val: BB.getTerminator())) { |
1025 | // Note that stripPointerCasts should look through functions with |
1026 | // returned arguments. |
1027 | auto *RetVal = |
1028 | dyn_cast<Argument>(Val: Ret->getReturnValue()->stripPointerCasts()); |
1029 | if (!RetVal || RetVal->getType() != F->getReturnType()) |
1030 | return nullptr; |
1031 | |
1032 | if (!RetArg) |
1033 | RetArg = RetVal; |
1034 | else if (RetArg != RetVal) |
1035 | return nullptr; |
1036 | } |
1037 | |
1038 | return RetArg; |
1039 | }; |
1040 | |
1041 | if (Argument *RetArg = FindRetArg()) { |
1042 | RetArg->addAttr(Kind: Attribute::Returned); |
1043 | ++NumReturned; |
1044 | Changed.insert(Ptr: F); |
1045 | } |
1046 | } |
1047 | } |
1048 | |
1049 | /// If a callsite has arguments that are also arguments to the parent function, |
1050 | /// try to propagate attributes from the callsite's arguments to the parent's |
1051 | /// arguments. This may be important because inlining can cause information loss |
1052 | /// when attribute knowledge disappears with the inlined call. |
1053 | static bool addArgumentAttrsFromCallsites(Function &F) { |
1054 | if (!EnableNonnullArgPropagation) |
1055 | return false; |
1056 | |
1057 | bool Changed = false; |
1058 | |
1059 | // For an argument attribute to transfer from a callsite to the parent, the |
1060 | // call must be guaranteed to execute every time the parent is called. |
1061 | // Conservatively, just check for calls in the entry block that are guaranteed |
1062 | // to execute. |
1063 | // TODO: This could be enhanced by testing if the callsite post-dominates the |
1064 | // entry block or by doing simple forward walks or backward walks to the |
1065 | // callsite. |
1066 | BasicBlock &Entry = F.getEntryBlock(); |
1067 | for (Instruction &I : Entry) { |
1068 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) { |
1069 | if (auto *CalledFunc = CB->getCalledFunction()) { |
1070 | for (auto &CSArg : CalledFunc->args()) { |
1071 | if (!CSArg.hasNonNullAttr(/* AllowUndefOrPoison */ false)) |
1072 | continue; |
1073 | |
1074 | // If the non-null callsite argument operand is an argument to 'F' |
1075 | // (the caller) and the call is guaranteed to execute, then the value |
1076 | // must be non-null throughout 'F'. |
1077 | auto *FArg = dyn_cast<Argument>(Val: CB->getArgOperand(i: CSArg.getArgNo())); |
1078 | if (FArg && !FArg->hasNonNullAttr()) { |
1079 | FArg->addAttr(Kind: Attribute::NonNull); |
1080 | Changed = true; |
1081 | } |
1082 | } |
1083 | } |
1084 | } |
1085 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
1086 | break; |
1087 | } |
1088 | |
1089 | return Changed; |
1090 | } |
1091 | |
1092 | static bool addAccessAttr(Argument *A, Attribute::AttrKind R) { |
1093 | assert((R == Attribute::ReadOnly || R == Attribute::ReadNone || |
1094 | R == Attribute::WriteOnly) |
1095 | && "Must be an access attribute." ); |
1096 | assert(A && "Argument must not be null." ); |
1097 | |
1098 | // If the argument already has the attribute, nothing needs to be done. |
1099 | if (A->hasAttribute(Kind: R)) |
1100 | return false; |
1101 | |
1102 | // Otherwise, remove potentially conflicting attribute, add the new one, |
1103 | // and update statistics. |
1104 | A->removeAttr(Kind: Attribute::WriteOnly); |
1105 | A->removeAttr(Kind: Attribute::ReadOnly); |
1106 | A->removeAttr(Kind: Attribute::ReadNone); |
1107 | // Remove conflicting writable attribute. |
1108 | if (R == Attribute::ReadNone || R == Attribute::ReadOnly) |
1109 | A->removeAttr(Kind: Attribute::Writable); |
1110 | A->addAttr(Kind: R); |
1111 | if (R == Attribute::ReadOnly) |
1112 | ++NumReadOnlyArg; |
1113 | else if (R == Attribute::WriteOnly) |
1114 | ++NumWriteOnlyArg; |
1115 | else |
1116 | ++NumReadNoneArg; |
1117 | return true; |
1118 | } |
1119 | |
1120 | static bool inferInitializes(Argument &A, Function &F) { |
1121 | auto ArgumentUses = collectArgumentUsesPerBlock(A, F); |
1122 | // No write anywhere in the function, bail. |
1123 | if (!ArgumentUses.HasAnyWrite) |
1124 | return false; |
1125 | |
1126 | auto &UsesPerBlock = ArgumentUses.UsesPerBlock; |
1127 | BasicBlock &EntryBB = F.getEntryBlock(); |
1128 | // A map to store the argument ranges initialized by a BasicBlock (including |
1129 | // its successors). |
1130 | DenseMap<const BasicBlock *, ConstantRangeList> Initialized; |
1131 | // Visit the successors of "BB" block and the instructions in BB (post-order) |
1132 | // to get the argument ranges initialized by "BB" (including its successors). |
1133 | // The result will be cached in "Initialized". |
1134 | auto VisitBlock = [&](const BasicBlock *BB) -> ConstantRangeList { |
1135 | auto UPB = UsesPerBlock.find(Val: BB); |
1136 | ConstantRangeList CRL; |
1137 | |
1138 | // Start with intersection of successors. |
1139 | // If this block has any clobbering use, we're going to clear out the |
1140 | // ranges at some point in this block anyway, so don't bother looking at |
1141 | // successors. |
1142 | if (UPB == UsesPerBlock.end() || !UPB->second.HasUnknownAccess) { |
1143 | bool HasAddedSuccessor = false; |
1144 | for (auto *Succ : successors(BB)) { |
1145 | if (auto SuccI = Initialized.find(Val: Succ); SuccI != Initialized.end()) { |
1146 | if (HasAddedSuccessor) { |
1147 | CRL = CRL.intersectWith(CRL: SuccI->second); |
1148 | } else { |
1149 | CRL = SuccI->second; |
1150 | HasAddedSuccessor = true; |
1151 | } |
1152 | } else { |
1153 | CRL = ConstantRangeList(); |
1154 | break; |
1155 | } |
1156 | } |
1157 | } |
1158 | |
1159 | if (UPB != UsesPerBlock.end()) { |
1160 | // Sort uses in this block by instruction order. |
1161 | SmallVector<std::pair<Instruction *, ArgumentAccessInfo>, 2> Insts; |
1162 | append_range(C&: Insts, R&: UPB->second.Insts); |
1163 | sort(C&: Insts, Comp: [](std::pair<Instruction *, ArgumentAccessInfo> &LHS, |
1164 | std::pair<Instruction *, ArgumentAccessInfo> &RHS) { |
1165 | return LHS.first->comesBefore(Other: RHS.first); |
1166 | }); |
1167 | |
1168 | // From the end of the block to the beginning of the block, set |
1169 | // initializes ranges. |
1170 | for (auto &[_, Info] : reverse(C&: Insts)) { |
1171 | if (Info.ArgAccessType == ArgumentAccessInfo::AccessType::Unknown || |
1172 | Info.ArgAccessType == |
1173 | ArgumentAccessInfo::AccessType::WriteWithSideEffect) |
1174 | CRL = ConstantRangeList(); |
1175 | if (!Info.AccessRanges.empty()) { |
1176 | if (Info.ArgAccessType == ArgumentAccessInfo::AccessType::Write || |
1177 | Info.ArgAccessType == |
1178 | ArgumentAccessInfo::AccessType::WriteWithSideEffect) { |
1179 | CRL = CRL.unionWith(CRL: Info.AccessRanges); |
1180 | } else { |
1181 | assert(Info.ArgAccessType == ArgumentAccessInfo::AccessType::Read); |
1182 | for (const auto &ReadRange : Info.AccessRanges) |
1183 | CRL.subtract(SubRange: ReadRange); |
1184 | } |
1185 | } |
1186 | } |
1187 | } |
1188 | return CRL; |
1189 | }; |
1190 | |
1191 | ConstantRangeList EntryCRL; |
1192 | // If all write instructions are in the EntryBB, or if the EntryBB has |
1193 | // a clobbering use, we only need to look at EntryBB. |
1194 | bool OnlyScanEntryBlock = !ArgumentUses.HasWriteOutsideEntryBB; |
1195 | if (!OnlyScanEntryBlock) |
1196 | if (auto EntryUPB = UsesPerBlock.find(Val: &EntryBB); |
1197 | EntryUPB != UsesPerBlock.end()) |
1198 | OnlyScanEntryBlock = EntryUPB->second.HasUnknownAccess; |
1199 | if (OnlyScanEntryBlock) { |
1200 | EntryCRL = VisitBlock(&EntryBB); |
1201 | if (EntryCRL.empty()) |
1202 | return false; |
1203 | } else { |
1204 | // Now we have to go through CFG to get the initialized argument ranges |
1205 | // across blocks. With dominance and post-dominance, the initialized ranges |
1206 | // by a block include both accesses inside this block and accesses in its |
1207 | // (transitive) successors. So visit successors before predecessors with a |
1208 | // post-order walk of the blocks and memorize the results in "Initialized". |
1209 | for (const BasicBlock *BB : post_order(G: &F)) { |
1210 | ConstantRangeList CRL = VisitBlock(BB); |
1211 | if (!CRL.empty()) |
1212 | Initialized[BB] = CRL; |
1213 | } |
1214 | |
1215 | auto EntryCRLI = Initialized.find(Val: &EntryBB); |
1216 | if (EntryCRLI == Initialized.end()) |
1217 | return false; |
1218 | |
1219 | EntryCRL = EntryCRLI->second; |
1220 | } |
1221 | |
1222 | assert(!EntryCRL.empty() && |
1223 | "should have bailed already if EntryCRL is empty" ); |
1224 | |
1225 | if (A.hasAttribute(Kind: Attribute::Initializes)) { |
1226 | ConstantRangeList PreviousCRL = |
1227 | A.getAttribute(Kind: Attribute::Initializes).getValueAsConstantRangeList(); |
1228 | if (PreviousCRL == EntryCRL) |
1229 | return false; |
1230 | EntryCRL = EntryCRL.unionWith(CRL: PreviousCRL); |
1231 | } |
1232 | |
1233 | A.addAttr(Attr: Attribute::get(Context&: A.getContext(), Kind: Attribute::Initializes, |
1234 | Val: EntryCRL.rangesRef())); |
1235 | |
1236 | return true; |
1237 | } |
1238 | |
1239 | /// Deduce nocapture attributes for the SCC. |
1240 | static void addArgumentAttrs(const SCCNodeSet &SCCNodes, |
1241 | SmallSet<Function *, 8> &Changed, |
1242 | bool SkipInitializes) { |
1243 | ArgumentGraph AG; |
1244 | |
1245 | auto DetermineAccessAttrsForSingleton = [](Argument *A) { |
1246 | SmallPtrSet<Argument *, 8> Self; |
1247 | Self.insert(Ptr: A); |
1248 | Attribute::AttrKind R = determinePointerAccessAttrs(A, SCCNodes: Self); |
1249 | if (R != Attribute::None) |
1250 | return addAccessAttr(A, R); |
1251 | return false; |
1252 | }; |
1253 | |
1254 | // Check each function in turn, determining which pointer arguments are not |
1255 | // captured. |
1256 | for (Function *F : SCCNodes) { |
1257 | // We can infer and propagate function attributes only when we know that the |
1258 | // definition we'll get at link time is *exactly* the definition we see now. |
1259 | // For more details, see GlobalValue::mayBeDerefined. |
1260 | if (!F->hasExactDefinition()) |
1261 | continue; |
1262 | |
1263 | if (addArgumentAttrsFromCallsites(F&: *F)) |
1264 | Changed.insert(Ptr: F); |
1265 | |
1266 | // Functions that are readonly (or readnone) and nounwind and don't return |
1267 | // a value can't capture arguments. Don't analyze them. |
1268 | if (F->onlyReadsMemory() && F->doesNotThrow() && F->willReturn() && |
1269 | F->getReturnType()->isVoidTy()) { |
1270 | for (Argument &A : F->args()) { |
1271 | if (A.getType()->isPointerTy() && !A.hasNoCaptureAttr()) { |
1272 | A.addAttr(Attr: Attribute::getWithCaptureInfo(Context&: A.getContext(), |
1273 | CI: CaptureInfo::none())); |
1274 | ++NumCapturesNone; |
1275 | Changed.insert(Ptr: F); |
1276 | } |
1277 | } |
1278 | continue; |
1279 | } |
1280 | |
1281 | for (Argument &A : F->args()) { |
1282 | if (!A.getType()->isPointerTy()) |
1283 | continue; |
1284 | bool HasNonLocalUses = false; |
1285 | CaptureInfo OrigCI = A.getAttributes().getCaptureInfo(); |
1286 | if (!capturesNothing(CC: OrigCI)) { |
1287 | ArgumentUsesTracker Tracker(SCCNodes); |
1288 | PointerMayBeCaptured(V: &A, Tracker: &Tracker); |
1289 | CaptureInfo NewCI = Tracker.CI & OrigCI; |
1290 | if (NewCI != OrigCI) { |
1291 | if (Tracker.Uses.empty()) { |
1292 | // If the information is complete, add the attribute now. |
1293 | A.addAttr(Attr: Attribute::getWithCaptureInfo(Context&: A.getContext(), CI: NewCI)); |
1294 | addCapturesStat(CI: NewCI); |
1295 | Changed.insert(Ptr: F); |
1296 | } else { |
1297 | // If it's not trivially captured and not trivially not captured, |
1298 | // then it must be calling into another function in our SCC. Save |
1299 | // its particulars for Argument-SCC analysis later. |
1300 | ArgumentGraphNode *Node = AG[&A]; |
1301 | Node->CC = CaptureComponents(NewCI); |
1302 | for (Argument *Use : Tracker.Uses) { |
1303 | Node->Uses.push_back(Elt: AG[Use]); |
1304 | if (Use != &A) |
1305 | HasNonLocalUses = true; |
1306 | } |
1307 | } |
1308 | } |
1309 | // Otherwise, it's captured. Don't bother doing SCC analysis on it. |
1310 | } |
1311 | if (!HasNonLocalUses && !A.onlyReadsMemory()) { |
1312 | // Can we determine that it's readonly/readnone/writeonly without doing |
1313 | // an SCC? Note that we don't allow any calls at all here, or else our |
1314 | // result will be dependent on the iteration order through the |
1315 | // functions in the SCC. |
1316 | if (DetermineAccessAttrsForSingleton(&A)) |
1317 | Changed.insert(Ptr: F); |
1318 | } |
1319 | if (!SkipInitializes && !A.onlyReadsMemory()) { |
1320 | if (inferInitializes(A, F&: *F)) |
1321 | Changed.insert(Ptr: F); |
1322 | } |
1323 | } |
1324 | } |
1325 | |
1326 | // The graph we've collected is partial because we stopped scanning for |
1327 | // argument uses once we solved the argument trivially. These partial nodes |
1328 | // show up as ArgumentGraphNode objects with an empty Uses list, and for |
1329 | // these nodes the final decision about whether they capture has already been |
1330 | // made. If the definition doesn't have a 'nocapture' attribute by now, it |
1331 | // captures. |
1332 | |
1333 | for (scc_iterator<ArgumentGraph *> I = scc_begin(G: &AG); !I.isAtEnd(); ++I) { |
1334 | const std::vector<ArgumentGraphNode *> &ArgumentSCC = *I; |
1335 | if (ArgumentSCC.size() == 1) { |
1336 | if (!ArgumentSCC[0]->Definition) |
1337 | continue; // synthetic root node |
1338 | |
1339 | // eg. "void f(int* x) { if (...) f(x); }" |
1340 | if (ArgumentSCC[0]->Uses.size() == 1 && |
1341 | ArgumentSCC[0]->Uses[0] == ArgumentSCC[0]) { |
1342 | Argument *A = ArgumentSCC[0]->Definition; |
1343 | CaptureInfo OrigCI = A->getAttributes().getCaptureInfo(); |
1344 | CaptureInfo NewCI = CaptureInfo(ArgumentSCC[0]->CC) & OrigCI; |
1345 | if (NewCI != OrigCI) { |
1346 | A->addAttr(Attr: Attribute::getWithCaptureInfo(Context&: A->getContext(), CI: NewCI)); |
1347 | addCapturesStat(CI: NewCI); |
1348 | Changed.insert(Ptr: A->getParent()); |
1349 | } |
1350 | |
1351 | // Infer the access attributes given the new captures one |
1352 | if (DetermineAccessAttrsForSingleton(A)) |
1353 | Changed.insert(Ptr: A->getParent()); |
1354 | } |
1355 | continue; |
1356 | } |
1357 | |
1358 | SmallPtrSet<Argument *, 8> ArgumentSCCNodes; |
1359 | // Fill ArgumentSCCNodes with the elements of the ArgumentSCC. Used for |
1360 | // quickly looking up whether a given Argument is in this ArgumentSCC. |
1361 | for (ArgumentGraphNode *I : ArgumentSCC) { |
1362 | ArgumentSCCNodes.insert(Ptr: I->Definition); |
1363 | } |
1364 | |
1365 | // At the SCC level, only track merged CaptureComponents. We're not |
1366 | // currently prepared to handle propagation of return-only captures across |
1367 | // the SCC. |
1368 | CaptureComponents CC = CaptureComponents::None; |
1369 | for (ArgumentGraphNode *N : ArgumentSCC) { |
1370 | for (ArgumentGraphNode *Use : N->Uses) { |
1371 | Argument *A = Use->Definition; |
1372 | if (ArgumentSCCNodes.count(Ptr: A)) |
1373 | CC |= Use->CC; |
1374 | else |
1375 | CC |= CaptureComponents(A->getAttributes().getCaptureInfo()); |
1376 | break; |
1377 | } |
1378 | if (capturesAll(CC)) |
1379 | break; |
1380 | } |
1381 | |
1382 | if (!capturesAll(CC)) { |
1383 | for (ArgumentGraphNode *N : ArgumentSCC) { |
1384 | Argument *A = N->Definition; |
1385 | CaptureInfo OrigCI = A->getAttributes().getCaptureInfo(); |
1386 | CaptureInfo NewCI = CaptureInfo(N->CC | CC) & OrigCI; |
1387 | if (NewCI != OrigCI) { |
1388 | A->addAttr(Attr: Attribute::getWithCaptureInfo(Context&: A->getContext(), CI: NewCI)); |
1389 | addCapturesStat(CI: NewCI); |
1390 | Changed.insert(Ptr: A->getParent()); |
1391 | } |
1392 | } |
1393 | } |
1394 | |
1395 | if (capturesAnyProvenance(CC)) { |
1396 | // As the pointer provenance may be captured, determine the pointer |
1397 | // attributes looking at each argument individually. |
1398 | for (ArgumentGraphNode *N : ArgumentSCC) { |
1399 | if (DetermineAccessAttrsForSingleton(N->Definition)) |
1400 | Changed.insert(Ptr: N->Definition->getParent()); |
1401 | } |
1402 | continue; |
1403 | } |
1404 | |
1405 | // We also want to compute readonly/readnone/writeonly. With a small number |
1406 | // of false negatives, we can assume that any pointer which is captured |
1407 | // isn't going to be provably readonly or readnone, since by definition |
1408 | // we can't analyze all uses of a captured pointer. |
1409 | // |
1410 | // The false negatives happen when the pointer is captured by a function |
1411 | // that promises readonly/readnone behaviour on the pointer, then the |
1412 | // pointer's lifetime ends before anything that writes to arbitrary memory. |
1413 | // Also, a readonly/readnone pointer may be returned, but returning a |
1414 | // pointer is capturing it. |
1415 | |
1416 | auto meetAccessAttr = [](Attribute::AttrKind A, Attribute::AttrKind B) { |
1417 | if (A == B) |
1418 | return A; |
1419 | if (A == Attribute::ReadNone) |
1420 | return B; |
1421 | if (B == Attribute::ReadNone) |
1422 | return A; |
1423 | return Attribute::None; |
1424 | }; |
1425 | |
1426 | Attribute::AttrKind AccessAttr = Attribute::ReadNone; |
1427 | for (ArgumentGraphNode *N : ArgumentSCC) { |
1428 | Argument *A = N->Definition; |
1429 | Attribute::AttrKind K = determinePointerAccessAttrs(A, SCCNodes: ArgumentSCCNodes); |
1430 | AccessAttr = meetAccessAttr(AccessAttr, K); |
1431 | if (AccessAttr == Attribute::None) |
1432 | break; |
1433 | } |
1434 | |
1435 | if (AccessAttr != Attribute::None) { |
1436 | for (ArgumentGraphNode *N : ArgumentSCC) { |
1437 | Argument *A = N->Definition; |
1438 | if (addAccessAttr(A, R: AccessAttr)) |
1439 | Changed.insert(Ptr: A->getParent()); |
1440 | } |
1441 | } |
1442 | } |
1443 | } |
1444 | |
1445 | /// Tests whether a function is "malloc-like". |
1446 | /// |
1447 | /// A function is "malloc-like" if it returns either null or a pointer that |
1448 | /// doesn't alias any other pointer visible to the caller. |
1449 | static bool isFunctionMallocLike(Function *F, const SCCNodeSet &SCCNodes) { |
1450 | SmallSetVector<Value *, 8> FlowsToReturn; |
1451 | for (BasicBlock &BB : *F) |
1452 | if (ReturnInst *Ret = dyn_cast<ReturnInst>(Val: BB.getTerminator())) |
1453 | FlowsToReturn.insert(X: Ret->getReturnValue()); |
1454 | |
1455 | for (unsigned i = 0; i != FlowsToReturn.size(); ++i) { |
1456 | Value *RetVal = FlowsToReturn[i]; |
1457 | |
1458 | if (Constant *C = dyn_cast<Constant>(Val: RetVal)) { |
1459 | if (!C->isNullValue() && !isa<UndefValue>(Val: C)) |
1460 | return false; |
1461 | |
1462 | continue; |
1463 | } |
1464 | |
1465 | if (isa<Argument>(Val: RetVal)) |
1466 | return false; |
1467 | |
1468 | if (Instruction *RVI = dyn_cast<Instruction>(Val: RetVal)) |
1469 | switch (RVI->getOpcode()) { |
1470 | // Extend the analysis by looking upwards. |
1471 | case Instruction::BitCast: |
1472 | case Instruction::GetElementPtr: |
1473 | case Instruction::AddrSpaceCast: |
1474 | FlowsToReturn.insert(X: RVI->getOperand(i: 0)); |
1475 | continue; |
1476 | case Instruction::Select: { |
1477 | SelectInst *SI = cast<SelectInst>(Val: RVI); |
1478 | FlowsToReturn.insert(X: SI->getTrueValue()); |
1479 | FlowsToReturn.insert(X: SI->getFalseValue()); |
1480 | continue; |
1481 | } |
1482 | case Instruction::PHI: { |
1483 | PHINode *PN = cast<PHINode>(Val: RVI); |
1484 | FlowsToReturn.insert_range(R: PN->incoming_values()); |
1485 | continue; |
1486 | } |
1487 | |
1488 | // Check whether the pointer came from an allocation. |
1489 | case Instruction::Alloca: |
1490 | break; |
1491 | case Instruction::Call: |
1492 | case Instruction::Invoke: { |
1493 | CallBase &CB = cast<CallBase>(Val&: *RVI); |
1494 | if (CB.hasRetAttr(Kind: Attribute::NoAlias)) |
1495 | break; |
1496 | if (CB.getCalledFunction() && SCCNodes.count(key: CB.getCalledFunction())) |
1497 | break; |
1498 | [[fallthrough]]; |
1499 | } |
1500 | default: |
1501 | return false; // Did not come from an allocation. |
1502 | } |
1503 | |
1504 | if (PointerMayBeCaptured(V: RetVal, /*ReturnCaptures=*/false)) |
1505 | return false; |
1506 | } |
1507 | |
1508 | return true; |
1509 | } |
1510 | |
1511 | /// Deduce noalias attributes for the SCC. |
1512 | static void addNoAliasAttrs(const SCCNodeSet &SCCNodes, |
1513 | SmallSet<Function *, 8> &Changed) { |
1514 | // Check each function in turn, determining which functions return noalias |
1515 | // pointers. |
1516 | for (Function *F : SCCNodes) { |
1517 | // Already noalias. |
1518 | if (F->returnDoesNotAlias()) |
1519 | continue; |
1520 | |
1521 | // We can infer and propagate function attributes only when we know that the |
1522 | // definition we'll get at link time is *exactly* the definition we see now. |
1523 | // For more details, see GlobalValue::mayBeDerefined. |
1524 | if (!F->hasExactDefinition()) |
1525 | return; |
1526 | |
1527 | // We annotate noalias return values, which are only applicable to |
1528 | // pointer types. |
1529 | if (!F->getReturnType()->isPointerTy()) |
1530 | continue; |
1531 | |
1532 | if (!isFunctionMallocLike(F, SCCNodes)) |
1533 | return; |
1534 | } |
1535 | |
1536 | for (Function *F : SCCNodes) { |
1537 | if (F->returnDoesNotAlias() || |
1538 | !F->getReturnType()->isPointerTy()) |
1539 | continue; |
1540 | |
1541 | F->setReturnDoesNotAlias(); |
1542 | ++NumNoAlias; |
1543 | Changed.insert(Ptr: F); |
1544 | } |
1545 | } |
1546 | |
1547 | /// Tests whether this function is known to not return null. |
1548 | /// |
1549 | /// Requires that the function returns a pointer. |
1550 | /// |
1551 | /// Returns true if it believes the function will not return a null, and sets |
1552 | /// \p Speculative based on whether the returned conclusion is a speculative |
1553 | /// conclusion due to SCC calls. |
1554 | static bool isReturnNonNull(Function *F, const SCCNodeSet &SCCNodes, |
1555 | bool &Speculative) { |
1556 | assert(F->getReturnType()->isPointerTy() && |
1557 | "nonnull only meaningful on pointer types" ); |
1558 | Speculative = false; |
1559 | |
1560 | SmallSetVector<Value *, 8> FlowsToReturn; |
1561 | for (BasicBlock &BB : *F) |
1562 | if (auto *Ret = dyn_cast<ReturnInst>(Val: BB.getTerminator())) |
1563 | FlowsToReturn.insert(X: Ret->getReturnValue()); |
1564 | |
1565 | auto &DL = F->getDataLayout(); |
1566 | |
1567 | for (unsigned i = 0; i != FlowsToReturn.size(); ++i) { |
1568 | Value *RetVal = FlowsToReturn[i]; |
1569 | |
1570 | // If this value is locally known to be non-null, we're good |
1571 | if (isKnownNonZero(V: RetVal, Q: DL)) |
1572 | continue; |
1573 | |
1574 | // Otherwise, we need to look upwards since we can't make any local |
1575 | // conclusions. |
1576 | Instruction *RVI = dyn_cast<Instruction>(Val: RetVal); |
1577 | if (!RVI) |
1578 | return false; |
1579 | switch (RVI->getOpcode()) { |
1580 | // Extend the analysis by looking upwards. |
1581 | case Instruction::BitCast: |
1582 | case Instruction::AddrSpaceCast: |
1583 | FlowsToReturn.insert(X: RVI->getOperand(i: 0)); |
1584 | continue; |
1585 | case Instruction::GetElementPtr: |
1586 | if (cast<GEPOperator>(Val: RVI)->isInBounds()) { |
1587 | FlowsToReturn.insert(X: RVI->getOperand(i: 0)); |
1588 | continue; |
1589 | } |
1590 | return false; |
1591 | case Instruction::Select: { |
1592 | SelectInst *SI = cast<SelectInst>(Val: RVI); |
1593 | FlowsToReturn.insert(X: SI->getTrueValue()); |
1594 | FlowsToReturn.insert(X: SI->getFalseValue()); |
1595 | continue; |
1596 | } |
1597 | case Instruction::PHI: { |
1598 | PHINode *PN = cast<PHINode>(Val: RVI); |
1599 | for (int i = 0, e = PN->getNumIncomingValues(); i != e; ++i) |
1600 | FlowsToReturn.insert(X: PN->getIncomingValue(i)); |
1601 | continue; |
1602 | } |
1603 | case Instruction::Call: |
1604 | case Instruction::Invoke: { |
1605 | CallBase &CB = cast<CallBase>(Val&: *RVI); |
1606 | Function *Callee = CB.getCalledFunction(); |
1607 | // A call to a node within the SCC is assumed to return null until |
1608 | // proven otherwise |
1609 | if (Callee && SCCNodes.count(key: Callee)) { |
1610 | Speculative = true; |
1611 | continue; |
1612 | } |
1613 | return false; |
1614 | } |
1615 | default: |
1616 | return false; // Unknown source, may be null |
1617 | }; |
1618 | llvm_unreachable("should have either continued or returned" ); |
1619 | } |
1620 | |
1621 | return true; |
1622 | } |
1623 | |
1624 | /// Deduce nonnull attributes for the SCC. |
1625 | static void addNonNullAttrs(const SCCNodeSet &SCCNodes, |
1626 | SmallSet<Function *, 8> &Changed) { |
1627 | // Speculative that all functions in the SCC return only nonnull |
1628 | // pointers. We may refute this as we analyze functions. |
1629 | bool SCCReturnsNonNull = true; |
1630 | |
1631 | // Check each function in turn, determining which functions return nonnull |
1632 | // pointers. |
1633 | for (Function *F : SCCNodes) { |
1634 | // Already nonnull. |
1635 | if (F->getAttributes().hasRetAttr(Kind: Attribute::NonNull)) |
1636 | continue; |
1637 | |
1638 | // We can infer and propagate function attributes only when we know that the |
1639 | // definition we'll get at link time is *exactly* the definition we see now. |
1640 | // For more details, see GlobalValue::mayBeDerefined. |
1641 | if (!F->hasExactDefinition()) |
1642 | return; |
1643 | |
1644 | // We annotate nonnull return values, which are only applicable to |
1645 | // pointer types. |
1646 | if (!F->getReturnType()->isPointerTy()) |
1647 | continue; |
1648 | |
1649 | bool Speculative = false; |
1650 | if (isReturnNonNull(F, SCCNodes, Speculative)) { |
1651 | if (!Speculative) { |
1652 | // Mark the function eagerly since we may discover a function |
1653 | // which prevents us from speculating about the entire SCC |
1654 | LLVM_DEBUG(dbgs() << "Eagerly marking " << F->getName() |
1655 | << " as nonnull\n" ); |
1656 | F->addRetAttr(Kind: Attribute::NonNull); |
1657 | ++NumNonNullReturn; |
1658 | Changed.insert(Ptr: F); |
1659 | } |
1660 | continue; |
1661 | } |
1662 | // At least one function returns something which could be null, can't |
1663 | // speculate any more. |
1664 | SCCReturnsNonNull = false; |
1665 | } |
1666 | |
1667 | if (SCCReturnsNonNull) { |
1668 | for (Function *F : SCCNodes) { |
1669 | if (F->getAttributes().hasRetAttr(Kind: Attribute::NonNull) || |
1670 | !F->getReturnType()->isPointerTy()) |
1671 | continue; |
1672 | |
1673 | LLVM_DEBUG(dbgs() << "SCC marking " << F->getName() << " as nonnull\n" ); |
1674 | F->addRetAttr(Kind: Attribute::NonNull); |
1675 | ++NumNonNullReturn; |
1676 | Changed.insert(Ptr: F); |
1677 | } |
1678 | } |
1679 | } |
1680 | |
1681 | /// Deduce noundef attributes for the SCC. |
1682 | static void addNoUndefAttrs(const SCCNodeSet &SCCNodes, |
1683 | SmallSet<Function *, 8> &Changed) { |
1684 | // Check each function in turn, determining which functions return noundef |
1685 | // values. |
1686 | for (Function *F : SCCNodes) { |
1687 | // Already noundef. |
1688 | AttributeList Attrs = F->getAttributes(); |
1689 | if (Attrs.hasRetAttr(Kind: Attribute::NoUndef)) |
1690 | continue; |
1691 | |
1692 | // We can infer and propagate function attributes only when we know that the |
1693 | // definition we'll get at link time is *exactly* the definition we see now. |
1694 | // For more details, see GlobalValue::mayBeDerefined. |
1695 | if (!F->hasExactDefinition()) |
1696 | return; |
1697 | |
1698 | // MemorySanitizer assumes that the definition and declaration of a |
1699 | // function will be consistent. A function with sanitize_memory attribute |
1700 | // should be skipped from inference. |
1701 | if (F->hasFnAttribute(Kind: Attribute::SanitizeMemory)) |
1702 | continue; |
1703 | |
1704 | if (F->getReturnType()->isVoidTy()) |
1705 | continue; |
1706 | |
1707 | const DataLayout &DL = F->getDataLayout(); |
1708 | if (all_of(Range&: *F, P: [&](BasicBlock &BB) { |
1709 | if (auto *Ret = dyn_cast<ReturnInst>(Val: BB.getTerminator())) { |
1710 | // TODO: perform context-sensitive analysis? |
1711 | Value *RetVal = Ret->getReturnValue(); |
1712 | if (!isGuaranteedNotToBeUndefOrPoison(V: RetVal)) |
1713 | return false; |
1714 | |
1715 | // We know the original return value is not poison now, but it |
1716 | // could still be converted to poison by another return attribute. |
1717 | // Try to explicitly re-prove the relevant attributes. |
1718 | if (Attrs.hasRetAttr(Kind: Attribute::NonNull) && |
1719 | !isKnownNonZero(V: RetVal, Q: DL)) |
1720 | return false; |
1721 | |
1722 | if (MaybeAlign Align = Attrs.getRetAlignment()) |
1723 | if (RetVal->getPointerAlignment(DL) < *Align) |
1724 | return false; |
1725 | |
1726 | Attribute Attr = Attrs.getRetAttr(Kind: Attribute::Range); |
1727 | if (Attr.isValid() && |
1728 | !Attr.getRange().contains( |
1729 | CR: computeConstantRange(V: RetVal, /*ForSigned=*/false))) |
1730 | return false; |
1731 | } |
1732 | return true; |
1733 | })) { |
1734 | F->addRetAttr(Kind: Attribute::NoUndef); |
1735 | ++NumNoUndefReturn; |
1736 | Changed.insert(Ptr: F); |
1737 | } |
1738 | } |
1739 | } |
1740 | |
1741 | namespace { |
1742 | |
1743 | /// Collects a set of attribute inference requests and performs them all in one |
1744 | /// go on a single SCC Node. Inference involves scanning function bodies |
1745 | /// looking for instructions that violate attribute assumptions. |
1746 | /// As soon as all the bodies are fine we are free to set the attribute. |
1747 | /// Customization of inference for individual attributes is performed by |
1748 | /// providing a handful of predicates for each attribute. |
1749 | class AttributeInferer { |
1750 | public: |
1751 | /// Describes a request for inference of a single attribute. |
1752 | struct InferenceDescriptor { |
1753 | |
1754 | /// Returns true if this function does not have to be handled. |
1755 | /// General intent for this predicate is to provide an optimization |
1756 | /// for functions that do not need this attribute inference at all |
1757 | /// (say, for functions that already have the attribute). |
1758 | std::function<bool(const Function &)> SkipFunction; |
1759 | |
1760 | /// Returns true if this instruction violates attribute assumptions. |
1761 | std::function<bool(Instruction &)> InstrBreaksAttribute; |
1762 | |
1763 | /// Sets the inferred attribute for this function. |
1764 | std::function<void(Function &)> SetAttribute; |
1765 | |
1766 | /// Attribute we derive. |
1767 | Attribute::AttrKind AKind; |
1768 | |
1769 | /// If true, only "exact" definitions can be used to infer this attribute. |
1770 | /// See GlobalValue::isDefinitionExact. |
1771 | bool RequiresExactDefinition; |
1772 | |
1773 | InferenceDescriptor(Attribute::AttrKind AK, |
1774 | std::function<bool(const Function &)> SkipFunc, |
1775 | std::function<bool(Instruction &)> InstrScan, |
1776 | std::function<void(Function &)> SetAttr, |
1777 | bool ReqExactDef) |
1778 | : SkipFunction(SkipFunc), InstrBreaksAttribute(InstrScan), |
1779 | SetAttribute(SetAttr), AKind(AK), |
1780 | RequiresExactDefinition(ReqExactDef) {} |
1781 | }; |
1782 | |
1783 | private: |
1784 | SmallVector<InferenceDescriptor, 4> InferenceDescriptors; |
1785 | |
1786 | public: |
1787 | void registerAttrInference(InferenceDescriptor AttrInference) { |
1788 | InferenceDescriptors.push_back(Elt: AttrInference); |
1789 | } |
1790 | |
1791 | void run(const SCCNodeSet &SCCNodes, SmallSet<Function *, 8> &Changed); |
1792 | }; |
1793 | |
1794 | /// Perform all the requested attribute inference actions according to the |
1795 | /// attribute predicates stored before. |
1796 | void AttributeInferer::run(const SCCNodeSet &SCCNodes, |
1797 | SmallSet<Function *, 8> &Changed) { |
1798 | SmallVector<InferenceDescriptor, 4> InferInSCC = InferenceDescriptors; |
1799 | // Go through all the functions in SCC and check corresponding attribute |
1800 | // assumptions for each of them. Attributes that are invalid for this SCC |
1801 | // will be removed from InferInSCC. |
1802 | for (Function *F : SCCNodes) { |
1803 | |
1804 | // No attributes whose assumptions are still valid - done. |
1805 | if (InferInSCC.empty()) |
1806 | return; |
1807 | |
1808 | // Check if our attributes ever need scanning/can be scanned. |
1809 | llvm::erase_if(C&: InferInSCC, P: [F](const InferenceDescriptor &ID) { |
1810 | if (ID.SkipFunction(*F)) |
1811 | return false; |
1812 | |
1813 | // Remove from further inference (invalidate) when visiting a function |
1814 | // that has no instructions to scan/has an unsuitable definition. |
1815 | return F->isDeclaration() || |
1816 | (ID.RequiresExactDefinition && !F->hasExactDefinition()); |
1817 | }); |
1818 | |
1819 | // For each attribute still in InferInSCC that doesn't explicitly skip F, |
1820 | // set up the F instructions scan to verify assumptions of the attribute. |
1821 | SmallVector<InferenceDescriptor, 4> InferInThisFunc; |
1822 | llvm::copy_if( |
1823 | Range&: InferInSCC, Out: std::back_inserter(x&: InferInThisFunc), |
1824 | P: [F](const InferenceDescriptor &ID) { return !ID.SkipFunction(*F); }); |
1825 | |
1826 | if (InferInThisFunc.empty()) |
1827 | continue; |
1828 | |
1829 | // Start instruction scan. |
1830 | for (Instruction &I : instructions(F&: *F)) { |
1831 | llvm::erase_if(C&: InferInThisFunc, P: [&](const InferenceDescriptor &ID) { |
1832 | if (!ID.InstrBreaksAttribute(I)) |
1833 | return false; |
1834 | // Remove attribute from further inference on any other functions |
1835 | // because attribute assumptions have just been violated. |
1836 | llvm::erase_if(C&: InferInSCC, P: [&ID](const InferenceDescriptor &D) { |
1837 | return D.AKind == ID.AKind; |
1838 | }); |
1839 | // Remove attribute from the rest of current instruction scan. |
1840 | return true; |
1841 | }); |
1842 | |
1843 | if (InferInThisFunc.empty()) |
1844 | break; |
1845 | } |
1846 | } |
1847 | |
1848 | if (InferInSCC.empty()) |
1849 | return; |
1850 | |
1851 | for (Function *F : SCCNodes) |
1852 | // At this point InferInSCC contains only functions that were either: |
1853 | // - explicitly skipped from scan/inference, or |
1854 | // - verified to have no instructions that break attribute assumptions. |
1855 | // Hence we just go and force the attribute for all non-skipped functions. |
1856 | for (auto &ID : InferInSCC) { |
1857 | if (ID.SkipFunction(*F)) |
1858 | continue; |
1859 | Changed.insert(Ptr: F); |
1860 | ID.SetAttribute(*F); |
1861 | } |
1862 | } |
1863 | |
1864 | struct SCCNodesResult { |
1865 | SCCNodeSet SCCNodes; |
1866 | bool HasUnknownCall; |
1867 | }; |
1868 | |
1869 | } // end anonymous namespace |
1870 | |
1871 | /// Helper for non-Convergent inference predicate InstrBreaksAttribute. |
1872 | static bool InstrBreaksNonConvergent(Instruction &I, |
1873 | const SCCNodeSet &SCCNodes) { |
1874 | const CallBase *CB = dyn_cast<CallBase>(Val: &I); |
1875 | // Breaks non-convergent assumption if CS is a convergent call to a function |
1876 | // not in the SCC. |
1877 | return CB && CB->isConvergent() && |
1878 | !SCCNodes.contains(key: CB->getCalledFunction()); |
1879 | } |
1880 | |
1881 | /// Helper for NoUnwind inference predicate InstrBreaksAttribute. |
1882 | static bool InstrBreaksNonThrowing(Instruction &I, const SCCNodeSet &SCCNodes) { |
1883 | if (!I.mayThrow(/* IncludePhaseOneUnwind */ true)) |
1884 | return false; |
1885 | if (const auto *CI = dyn_cast<CallInst>(Val: &I)) { |
1886 | if (Function *Callee = CI->getCalledFunction()) { |
1887 | // I is a may-throw call to a function inside our SCC. This doesn't |
1888 | // invalidate our current working assumption that the SCC is no-throw; we |
1889 | // just have to scan that other function. |
1890 | if (SCCNodes.contains(key: Callee)) |
1891 | return false; |
1892 | } |
1893 | } |
1894 | return true; |
1895 | } |
1896 | |
1897 | /// Helper for NoFree inference predicate InstrBreaksAttribute. |
1898 | static bool InstrBreaksNoFree(Instruction &I, const SCCNodeSet &SCCNodes) { |
1899 | CallBase *CB = dyn_cast<CallBase>(Val: &I); |
1900 | if (!CB) |
1901 | return false; |
1902 | |
1903 | if (CB->hasFnAttr(Kind: Attribute::NoFree)) |
1904 | return false; |
1905 | |
1906 | // Speculatively assume in SCC. |
1907 | if (Function *Callee = CB->getCalledFunction()) |
1908 | if (SCCNodes.contains(key: Callee)) |
1909 | return false; |
1910 | |
1911 | return true; |
1912 | } |
1913 | |
1914 | // Return true if this is an atomic which has an ordering stronger than |
1915 | // unordered. Note that this is different than the predicate we use in |
1916 | // Attributor. Here we chose to be conservative and consider monotonic |
1917 | // operations potentially synchronizing. We generally don't do much with |
1918 | // monotonic operations, so this is simply risk reduction. |
1919 | static bool isOrderedAtomic(Instruction *I) { |
1920 | if (!I->isAtomic()) |
1921 | return false; |
1922 | |
1923 | if (auto *FI = dyn_cast<FenceInst>(Val: I)) |
1924 | // All legal orderings for fence are stronger than monotonic. |
1925 | return FI->getSyncScopeID() != SyncScope::SingleThread; |
1926 | else if (isa<AtomicCmpXchgInst>(Val: I) || isa<AtomicRMWInst>(Val: I)) |
1927 | return true; |
1928 | else if (auto *SI = dyn_cast<StoreInst>(Val: I)) |
1929 | return !SI->isUnordered(); |
1930 | else if (auto *LI = dyn_cast<LoadInst>(Val: I)) |
1931 | return !LI->isUnordered(); |
1932 | else { |
1933 | llvm_unreachable("unknown atomic instruction?" ); |
1934 | } |
1935 | } |
1936 | |
1937 | static bool InstrBreaksNoSync(Instruction &I, const SCCNodeSet &SCCNodes) { |
1938 | // Volatile may synchronize |
1939 | if (I.isVolatile()) |
1940 | return true; |
1941 | |
1942 | // An ordered atomic may synchronize. (See comment about on monotonic.) |
1943 | if (isOrderedAtomic(I: &I)) |
1944 | return true; |
1945 | |
1946 | auto *CB = dyn_cast<CallBase>(Val: &I); |
1947 | if (!CB) |
1948 | // Non call site cases covered by the two checks above |
1949 | return false; |
1950 | |
1951 | if (CB->hasFnAttr(Kind: Attribute::NoSync)) |
1952 | return false; |
1953 | |
1954 | // Non volatile memset/memcpy/memmoves are nosync |
1955 | // NOTE: Only intrinsics with volatile flags should be handled here. All |
1956 | // others should be marked in Intrinsics.td. |
1957 | if (auto *MI = dyn_cast<MemIntrinsic>(Val: &I)) |
1958 | if (!MI->isVolatile()) |
1959 | return false; |
1960 | |
1961 | // Speculatively assume in SCC. |
1962 | if (Function *Callee = CB->getCalledFunction()) |
1963 | if (SCCNodes.contains(key: Callee)) |
1964 | return false; |
1965 | |
1966 | return true; |
1967 | } |
1968 | |
1969 | /// Attempt to remove convergent function attribute when possible. |
1970 | /// |
1971 | /// Returns true if any changes to function attributes were made. |
1972 | static void inferConvergent(const SCCNodeSet &SCCNodes, |
1973 | SmallSet<Function *, 8> &Changed) { |
1974 | AttributeInferer AI; |
1975 | |
1976 | // Request to remove the convergent attribute from all functions in the SCC |
1977 | // if every callsite within the SCC is not convergent (except for calls |
1978 | // to functions within the SCC). |
1979 | // Note: Removal of the attr from the callsites will happen in |
1980 | // InstCombineCalls separately. |
1981 | AI.registerAttrInference(AttrInference: AttributeInferer::InferenceDescriptor{ |
1982 | Attribute::Convergent, |
1983 | // Skip non-convergent functions. |
1984 | [](const Function &F) { return !F.isConvergent(); }, |
1985 | // Instructions that break non-convergent assumption. |
1986 | [SCCNodes](Instruction &I) { |
1987 | return InstrBreaksNonConvergent(I, SCCNodes); |
1988 | }, |
1989 | [](Function &F) { |
1990 | LLVM_DEBUG(dbgs() << "Removing convergent attr from fn " << F.getName() |
1991 | << "\n" ); |
1992 | F.setNotConvergent(); |
1993 | }, |
1994 | /* RequiresExactDefinition= */ false}); |
1995 | // Perform all the requested attribute inference actions. |
1996 | AI.run(SCCNodes, Changed); |
1997 | } |
1998 | |
1999 | /// Infer attributes from all functions in the SCC by scanning every |
2000 | /// instruction for compliance to the attribute assumptions. |
2001 | /// |
2002 | /// Returns true if any changes to function attributes were made. |
2003 | static void inferAttrsFromFunctionBodies(const SCCNodeSet &SCCNodes, |
2004 | SmallSet<Function *, 8> &Changed) { |
2005 | AttributeInferer AI; |
2006 | |
2007 | if (!DisableNoUnwindInference) |
2008 | // Request to infer nounwind attribute for all the functions in the SCC if |
2009 | // every callsite within the SCC is not throwing (except for calls to |
2010 | // functions within the SCC). Note that nounwind attribute suffers from |
2011 | // derefinement - results may change depending on how functions are |
2012 | // optimized. Thus it can be inferred only from exact definitions. |
2013 | AI.registerAttrInference(AttrInference: AttributeInferer::InferenceDescriptor{ |
2014 | Attribute::NoUnwind, |
2015 | // Skip non-throwing functions. |
2016 | [](const Function &F) { return F.doesNotThrow(); }, |
2017 | // Instructions that break non-throwing assumption. |
2018 | [&SCCNodes](Instruction &I) { |
2019 | return InstrBreaksNonThrowing(I, SCCNodes); |
2020 | }, |
2021 | [](Function &F) { |
2022 | LLVM_DEBUG(dbgs() |
2023 | << "Adding nounwind attr to fn " << F.getName() << "\n" ); |
2024 | F.setDoesNotThrow(); |
2025 | ++NumNoUnwind; |
2026 | }, |
2027 | /* RequiresExactDefinition= */ true}); |
2028 | |
2029 | if (!DisableNoFreeInference) |
2030 | // Request to infer nofree attribute for all the functions in the SCC if |
2031 | // every callsite within the SCC does not directly or indirectly free |
2032 | // memory (except for calls to functions within the SCC). Note that nofree |
2033 | // attribute suffers from derefinement - results may change depending on |
2034 | // how functions are optimized. Thus it can be inferred only from exact |
2035 | // definitions. |
2036 | AI.registerAttrInference(AttrInference: AttributeInferer::InferenceDescriptor{ |
2037 | Attribute::NoFree, |
2038 | // Skip functions known not to free memory. |
2039 | [](const Function &F) { return F.doesNotFreeMemory(); }, |
2040 | // Instructions that break non-deallocating assumption. |
2041 | [&SCCNodes](Instruction &I) { |
2042 | return InstrBreaksNoFree(I, SCCNodes); |
2043 | }, |
2044 | [](Function &F) { |
2045 | LLVM_DEBUG(dbgs() |
2046 | << "Adding nofree attr to fn " << F.getName() << "\n" ); |
2047 | F.setDoesNotFreeMemory(); |
2048 | ++NumNoFree; |
2049 | }, |
2050 | /* RequiresExactDefinition= */ true}); |
2051 | |
2052 | AI.registerAttrInference(AttrInference: AttributeInferer::InferenceDescriptor{ |
2053 | Attribute::NoSync, |
2054 | // Skip already marked functions. |
2055 | [](const Function &F) { return F.hasNoSync(); }, |
2056 | // Instructions that break nosync assumption. |
2057 | [&SCCNodes](Instruction &I) { |
2058 | return InstrBreaksNoSync(I, SCCNodes); |
2059 | }, |
2060 | [](Function &F) { |
2061 | LLVM_DEBUG(dbgs() |
2062 | << "Adding nosync attr to fn " << F.getName() << "\n" ); |
2063 | F.setNoSync(); |
2064 | ++NumNoSync; |
2065 | }, |
2066 | /* RequiresExactDefinition= */ true}); |
2067 | |
2068 | // Perform all the requested attribute inference actions. |
2069 | AI.run(SCCNodes, Changed); |
2070 | } |
2071 | |
2072 | static void addNoRecurseAttrs(const SCCNodeSet &SCCNodes, |
2073 | SmallSet<Function *, 8> &Changed) { |
2074 | // Try and identify functions that do not recurse. |
2075 | |
2076 | // If the SCC contains multiple nodes we know for sure there is recursion. |
2077 | if (SCCNodes.size() != 1) |
2078 | return; |
2079 | |
2080 | Function *F = *SCCNodes.begin(); |
2081 | if (!F || !F->hasExactDefinition() || F->doesNotRecurse()) |
2082 | return; |
2083 | |
2084 | // If all of the calls in F are identifiable and are to norecurse functions, F |
2085 | // is norecurse. This check also detects self-recursion as F is not currently |
2086 | // marked norecurse, so any called from F to F will not be marked norecurse. |
2087 | for (auto &BB : *F) |
2088 | for (auto &I : BB.instructionsWithoutDebug()) |
2089 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) { |
2090 | Function *Callee = CB->getCalledFunction(); |
2091 | if (!Callee || Callee == F || |
2092 | (!Callee->doesNotRecurse() && |
2093 | !(Callee->isDeclaration() && |
2094 | Callee->hasFnAttribute(Kind: Attribute::NoCallback)))) |
2095 | // Function calls a potentially recursive function. |
2096 | return; |
2097 | } |
2098 | |
2099 | // Every call was to a non-recursive function other than this function, and |
2100 | // we have no indirect recursion as the SCC size is one. This function cannot |
2101 | // recurse. |
2102 | F->setDoesNotRecurse(); |
2103 | ++NumNoRecurse; |
2104 | Changed.insert(Ptr: F); |
2105 | } |
2106 | |
2107 | // Set the noreturn function attribute if possible. |
2108 | static void addNoReturnAttrs(const SCCNodeSet &SCCNodes, |
2109 | SmallSet<Function *, 8> &Changed) { |
2110 | for (Function *F : SCCNodes) { |
2111 | if (!F || !F->hasExactDefinition() || F->hasFnAttribute(Kind: Attribute::Naked) || |
2112 | F->doesNotReturn()) |
2113 | continue; |
2114 | |
2115 | if (!canReturn(F: *F)) { |
2116 | F->setDoesNotReturn(); |
2117 | Changed.insert(Ptr: F); |
2118 | } |
2119 | } |
2120 | } |
2121 | |
2122 | static bool allPathsGoThroughCold(Function &F) { |
2123 | SmallDenseMap<BasicBlock *, bool, 16> ColdPaths; |
2124 | ColdPaths[&F.front()] = false; |
2125 | SmallVector<BasicBlock *> Jobs; |
2126 | Jobs.push_back(Elt: &F.front()); |
2127 | |
2128 | while (!Jobs.empty()) { |
2129 | BasicBlock *BB = Jobs.pop_back_val(); |
2130 | |
2131 | // If block contains a cold callsite this path through the CG is cold. |
2132 | // Ignore whether the instructions actually are guaranteed to transfer |
2133 | // execution. Divergent behavior is considered unlikely. |
2134 | if (any_of(Range&: *BB, P: [](Instruction &I) { |
2135 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) |
2136 | return CB->hasFnAttr(Kind: Attribute::Cold); |
2137 | return false; |
2138 | })) { |
2139 | ColdPaths[BB] = true; |
2140 | continue; |
2141 | } |
2142 | |
2143 | auto Succs = successors(BB); |
2144 | // We found a path that doesn't go through any cold callsite. |
2145 | if (Succs.empty()) |
2146 | return false; |
2147 | |
2148 | // We didn't find a cold callsite in this BB, so check that all successors |
2149 | // contain a cold callsite (or that their successors do). |
2150 | // Potential TODO: We could use static branch hints to assume certain |
2151 | // successor paths are inherently cold, irrespective of if they contain a |
2152 | // cold callsite. |
2153 | for (BasicBlock *Succ : Succs) { |
2154 | // Start with false, this is necessary to ensure we don't turn loops into |
2155 | // cold. |
2156 | auto [Iter, Inserted] = ColdPaths.try_emplace(Key: Succ, Args: false); |
2157 | if (!Inserted) { |
2158 | if (Iter->second) |
2159 | continue; |
2160 | return false; |
2161 | } |
2162 | Jobs.push_back(Elt: Succ); |
2163 | } |
2164 | } |
2165 | return true; |
2166 | } |
2167 | |
2168 | // Set the cold function attribute if possible. |
2169 | static void addColdAttrs(const SCCNodeSet &SCCNodes, |
2170 | SmallSet<Function *, 8> &Changed) { |
2171 | for (Function *F : SCCNodes) { |
2172 | if (!F || !F->hasExactDefinition() || F->hasFnAttribute(Kind: Attribute::Naked) || |
2173 | F->hasFnAttribute(Kind: Attribute::Cold) || F->hasFnAttribute(Kind: Attribute::Hot)) |
2174 | continue; |
2175 | |
2176 | // Potential TODO: We could add attribute `cold` on functions with `coldcc`. |
2177 | if (allPathsGoThroughCold(F&: *F)) { |
2178 | F->addFnAttr(Kind: Attribute::Cold); |
2179 | ++NumCold; |
2180 | Changed.insert(Ptr: F); |
2181 | continue; |
2182 | } |
2183 | } |
2184 | } |
2185 | |
2186 | static bool functionWillReturn(const Function &F) { |
2187 | // We can infer and propagate function attributes only when we know that the |
2188 | // definition we'll get at link time is *exactly* the definition we see now. |
2189 | // For more details, see GlobalValue::mayBeDerefined. |
2190 | if (!F.hasExactDefinition()) |
2191 | return false; |
2192 | |
2193 | // Must-progress function without side-effects must return. |
2194 | if (F.mustProgress() && F.onlyReadsMemory()) |
2195 | return true; |
2196 | |
2197 | // Can only analyze functions with a definition. |
2198 | if (F.isDeclaration()) |
2199 | return false; |
2200 | |
2201 | // Functions with loops require more sophisticated analysis, as the loop |
2202 | // may be infinite. For now, don't try to handle them. |
2203 | SmallVector<std::pair<const BasicBlock *, const BasicBlock *>> Backedges; |
2204 | FindFunctionBackedges(F, Result&: Backedges); |
2205 | if (!Backedges.empty()) |
2206 | return false; |
2207 | |
2208 | // If there are no loops, then the function is willreturn if all calls in |
2209 | // it are willreturn. |
2210 | return all_of(Range: instructions(F), P: [](const Instruction &I) { |
2211 | return I.willReturn(); |
2212 | }); |
2213 | } |
2214 | |
2215 | // Set the willreturn function attribute if possible. |
2216 | static void addWillReturn(const SCCNodeSet &SCCNodes, |
2217 | SmallSet<Function *, 8> &Changed) { |
2218 | for (Function *F : SCCNodes) { |
2219 | if (!F || F->willReturn() || !functionWillReturn(F: *F)) |
2220 | continue; |
2221 | |
2222 | F->setWillReturn(); |
2223 | NumWillReturn++; |
2224 | Changed.insert(Ptr: F); |
2225 | } |
2226 | } |
2227 | |
2228 | static SCCNodesResult createSCCNodeSet(ArrayRef<Function *> Functions) { |
2229 | SCCNodesResult Res; |
2230 | Res.HasUnknownCall = false; |
2231 | for (Function *F : Functions) { |
2232 | if (!F || F->hasOptNone() || F->hasFnAttribute(Kind: Attribute::Naked) || |
2233 | F->isPresplitCoroutine()) { |
2234 | // Treat any function we're trying not to optimize as if it were an |
2235 | // indirect call and omit it from the node set used below. |
2236 | Res.HasUnknownCall = true; |
2237 | continue; |
2238 | } |
2239 | // Track whether any functions in this SCC have an unknown call edge. |
2240 | // Note: if this is ever a performance hit, we can common it with |
2241 | // subsequent routines which also do scans over the instructions of the |
2242 | // function. |
2243 | if (!Res.HasUnknownCall) { |
2244 | for (Instruction &I : instructions(F&: *F)) { |
2245 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) { |
2246 | if (!CB->getCalledFunction()) { |
2247 | Res.HasUnknownCall = true; |
2248 | break; |
2249 | } |
2250 | } |
2251 | } |
2252 | } |
2253 | Res.SCCNodes.insert(X: F); |
2254 | } |
2255 | return Res; |
2256 | } |
2257 | |
2258 | template <typename AARGetterT> |
2259 | static SmallSet<Function *, 8> |
2260 | deriveAttrsInPostOrder(ArrayRef<Function *> Functions, AARGetterT &&AARGetter, |
2261 | bool ArgAttrsOnly) { |
2262 | SCCNodesResult Nodes = createSCCNodeSet(Functions); |
2263 | |
2264 | // Bail if the SCC only contains optnone functions. |
2265 | if (Nodes.SCCNodes.empty()) |
2266 | return {}; |
2267 | |
2268 | SmallSet<Function *, 8> Changed; |
2269 | if (ArgAttrsOnly) { |
2270 | // ArgAttrsOnly means to only infer attributes that may aid optimizations |
2271 | // on the *current* function. "initializes" attribute is to aid |
2272 | // optimizations (like DSE) on the callers, so skip "initializes" here. |
2273 | addArgumentAttrs(SCCNodes: Nodes.SCCNodes, Changed, /*SkipInitializes=*/true); |
2274 | return Changed; |
2275 | } |
2276 | |
2277 | addArgumentReturnedAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
2278 | addMemoryAttrs(Nodes.SCCNodes, AARGetter, Changed); |
2279 | addArgumentAttrs(SCCNodes: Nodes.SCCNodes, Changed, /*SkipInitializes=*/false); |
2280 | inferConvergent(SCCNodes: Nodes.SCCNodes, Changed); |
2281 | addNoReturnAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
2282 | addColdAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
2283 | addWillReturn(SCCNodes: Nodes.SCCNodes, Changed); |
2284 | addNoUndefAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
2285 | |
2286 | // If we have no external nodes participating in the SCC, we can deduce some |
2287 | // more precise attributes as well. |
2288 | if (!Nodes.HasUnknownCall) { |
2289 | addNoAliasAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
2290 | addNonNullAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
2291 | inferAttrsFromFunctionBodies(SCCNodes: Nodes.SCCNodes, Changed); |
2292 | addNoRecurseAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
2293 | } |
2294 | |
2295 | // Finally, infer the maximal set of attributes from the ones we've inferred |
2296 | // above. This is handling the cases where one attribute on a signature |
2297 | // implies another, but for implementation reasons the inference rule for |
2298 | // the later is missing (or simply less sophisticated). |
2299 | for (Function *F : Nodes.SCCNodes) |
2300 | if (F) |
2301 | if (inferAttributesFromOthers(F&: *F)) |
2302 | Changed.insert(Ptr: F); |
2303 | |
2304 | return Changed; |
2305 | } |
2306 | |
2307 | PreservedAnalyses PostOrderFunctionAttrsPass::run(LazyCallGraph::SCC &C, |
2308 | CGSCCAnalysisManager &AM, |
2309 | LazyCallGraph &CG, |
2310 | CGSCCUpdateResult &) { |
2311 | // Skip non-recursive functions if requested. |
2312 | // Only infer argument attributes for non-recursive functions, because |
2313 | // it can affect optimization behavior in conjunction with noalias. |
2314 | bool ArgAttrsOnly = false; |
2315 | if (C.size() == 1 && SkipNonRecursive) { |
2316 | LazyCallGraph::Node &N = *C.begin(); |
2317 | if (!N->lookup(N)) |
2318 | ArgAttrsOnly = true; |
2319 | } |
2320 | |
2321 | FunctionAnalysisManager &FAM = |
2322 | AM.getResult<FunctionAnalysisManagerCGSCCProxy>(IR&: C, ExtraArgs&: CG).getManager(); |
2323 | |
2324 | // We pass a lambda into functions to wire them up to the analysis manager |
2325 | // for getting function analyses. |
2326 | auto AARGetter = [&](Function &F) -> AAResults & { |
2327 | return FAM.getResult<AAManager>(IR&: F); |
2328 | }; |
2329 | |
2330 | SmallVector<Function *, 8> Functions; |
2331 | for (LazyCallGraph::Node &N : C) { |
2332 | Functions.push_back(Elt: &N.getFunction()); |
2333 | } |
2334 | |
2335 | auto ChangedFunctions = |
2336 | deriveAttrsInPostOrder(Functions, AARGetter, ArgAttrsOnly); |
2337 | if (ChangedFunctions.empty()) |
2338 | return PreservedAnalyses::all(); |
2339 | |
2340 | // Invalidate analyses for modified functions so that we don't have to |
2341 | // invalidate all analyses for all functions in this SCC. |
2342 | PreservedAnalyses FuncPA; |
2343 | // We haven't changed the CFG for modified functions. |
2344 | FuncPA.preserveSet<CFGAnalyses>(); |
2345 | for (Function *Changed : ChangedFunctions) { |
2346 | FAM.invalidate(IR&: *Changed, PA: FuncPA); |
2347 | // Also invalidate any direct callers of changed functions since analyses |
2348 | // may care about attributes of direct callees. For example, MemorySSA cares |
2349 | // about whether or not a call's callee modifies memory and queries that |
2350 | // through function attributes. |
2351 | for (auto *U : Changed->users()) { |
2352 | if (auto *Call = dyn_cast<CallBase>(Val: U)) { |
2353 | if (Call->getCalledFunction() == Changed) |
2354 | FAM.invalidate(IR&: *Call->getFunction(), PA: FuncPA); |
2355 | } |
2356 | } |
2357 | } |
2358 | |
2359 | PreservedAnalyses PA; |
2360 | // We have not added or removed functions. |
2361 | PA.preserve<FunctionAnalysisManagerCGSCCProxy>(); |
2362 | // We already invalidated all relevant function analyses above. |
2363 | PA.preserveSet<AllAnalysesOn<Function>>(); |
2364 | return PA; |
2365 | } |
2366 | |
2367 | void PostOrderFunctionAttrsPass::printPipeline( |
2368 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
2369 | static_cast<PassInfoMixin<PostOrderFunctionAttrsPass> *>(this)->printPipeline( |
2370 | OS, MapClassName2PassName); |
2371 | if (SkipNonRecursive) |
2372 | OS << "<skip-non-recursive-function-attrs>" ; |
2373 | } |
2374 | |
2375 | template <typename AARGetterT> |
2376 | static bool runImpl(CallGraphSCC &SCC, AARGetterT AARGetter) { |
2377 | SmallVector<Function *, 8> Functions; |
2378 | for (CallGraphNode *I : SCC) { |
2379 | Functions.push_back(Elt: I->getFunction()); |
2380 | } |
2381 | |
2382 | return !deriveAttrsInPostOrder(Functions, AARGetter).empty(); |
2383 | } |
2384 | |
2385 | static bool addNoRecurseAttrsTopDown(Function &F) { |
2386 | // We check the preconditions for the function prior to calling this to avoid |
2387 | // the cost of building up a reversible post-order list. We assert them here |
2388 | // to make sure none of the invariants this relies on were violated. |
2389 | assert(!F.isDeclaration() && "Cannot deduce norecurse without a definition!" ); |
2390 | assert(!F.doesNotRecurse() && |
2391 | "This function has already been deduced as norecurs!" ); |
2392 | assert(F.hasInternalLinkage() && |
2393 | "Can only do top-down deduction for internal linkage functions!" ); |
2394 | |
2395 | // If F is internal and all of its uses are calls from a non-recursive |
2396 | // functions, then none of its calls could in fact recurse without going |
2397 | // through a function marked norecurse, and so we can mark this function too |
2398 | // as norecurse. Note that the uses must actually be calls -- otherwise |
2399 | // a pointer to this function could be returned from a norecurse function but |
2400 | // this function could be recursively (indirectly) called. Note that this |
2401 | // also detects if F is directly recursive as F is not yet marked as |
2402 | // a norecurse function. |
2403 | for (auto &U : F.uses()) { |
2404 | auto *I = dyn_cast<Instruction>(Val: U.getUser()); |
2405 | if (!I) |
2406 | return false; |
2407 | CallBase *CB = dyn_cast<CallBase>(Val: I); |
2408 | if (!CB || !CB->isCallee(U: &U) || |
2409 | !CB->getParent()->getParent()->doesNotRecurse()) |
2410 | return false; |
2411 | } |
2412 | F.setDoesNotRecurse(); |
2413 | ++NumNoRecurse; |
2414 | return true; |
2415 | } |
2416 | |
2417 | static bool deduceFunctionAttributeInRPO(Module &M, LazyCallGraph &CG) { |
2418 | // We only have a post-order SCC traversal (because SCCs are inherently |
2419 | // discovered in post-order), so we accumulate them in a vector and then walk |
2420 | // it in reverse. This is simpler than using the RPO iterator infrastructure |
2421 | // because we need to combine SCC detection and the PO walk of the call |
2422 | // graph. We can also cheat egregiously because we're primarily interested in |
2423 | // synthesizing norecurse and so we can only save the singular SCCs as SCCs |
2424 | // with multiple functions in them will clearly be recursive. |
2425 | |
2426 | SmallVector<Function *, 16> Worklist; |
2427 | CG.buildRefSCCs(); |
2428 | for (LazyCallGraph::RefSCC &RC : CG.postorder_ref_sccs()) { |
2429 | for (LazyCallGraph::SCC &SCC : RC) { |
2430 | if (SCC.size() != 1) |
2431 | continue; |
2432 | Function &F = SCC.begin()->getFunction(); |
2433 | if (!F.isDeclaration() && !F.doesNotRecurse() && F.hasInternalLinkage()) |
2434 | Worklist.push_back(Elt: &F); |
2435 | } |
2436 | } |
2437 | bool Changed = false; |
2438 | for (auto *F : llvm::reverse(C&: Worklist)) |
2439 | Changed |= addNoRecurseAttrsTopDown(F&: *F); |
2440 | |
2441 | return Changed; |
2442 | } |
2443 | |
2444 | PreservedAnalyses |
2445 | ReversePostOrderFunctionAttrsPass::run(Module &M, ModuleAnalysisManager &AM) { |
2446 | auto &CG = AM.getResult<LazyCallGraphAnalysis>(IR&: M); |
2447 | |
2448 | if (!deduceFunctionAttributeInRPO(M, CG)) |
2449 | return PreservedAnalyses::all(); |
2450 | |
2451 | PreservedAnalyses PA; |
2452 | PA.preserve<LazyCallGraphAnalysis>(); |
2453 | return PA; |
2454 | } |
2455 | |