| 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/SmallVector.h" |
| 24 | #include "llvm/ADT/Statistic.h" |
| 25 | #include "llvm/Analysis/AssumptionCache.h" |
| 26 | #include "llvm/Analysis/BasicAliasAnalysis.h" |
| 27 | #include "llvm/Analysis/CFG.h" |
| 28 | #include "llvm/Analysis/CGSCCPassManager.h" |
| 29 | #include "llvm/Analysis/CallGraph.h" |
| 30 | #include "llvm/Analysis/CallGraphSCCPass.h" |
| 31 | #include "llvm/Analysis/CaptureTracking.h" |
| 32 | #include "llvm/Analysis/LazyCallGraph.h" |
| 33 | #include "llvm/Analysis/MemoryLocation.h" |
| 34 | #include "llvm/Analysis/ValueTracking.h" |
| 35 | #include "llvm/IR/Argument.h" |
| 36 | #include "llvm/IR/Attributes.h" |
| 37 | #include "llvm/IR/BasicBlock.h" |
| 38 | #include "llvm/IR/Constant.h" |
| 39 | #include "llvm/IR/ConstantRangeList.h" |
| 40 | #include "llvm/IR/Constants.h" |
| 41 | #include "llvm/IR/Function.h" |
| 42 | #include "llvm/IR/InstIterator.h" |
| 43 | #include "llvm/IR/InstrTypes.h" |
| 44 | #include "llvm/IR/Instruction.h" |
| 45 | #include "llvm/IR/Instructions.h" |
| 46 | #include "llvm/IR/IntrinsicInst.h" |
| 47 | #include "llvm/IR/Metadata.h" |
| 48 | #include "llvm/IR/ModuleSummaryIndex.h" |
| 49 | #include "llvm/IR/PassManager.h" |
| 50 | #include "llvm/IR/Type.h" |
| 51 | #include "llvm/IR/Use.h" |
| 52 | #include "llvm/IR/User.h" |
| 53 | #include "llvm/IR/Value.h" |
| 54 | #include "llvm/Support/Casting.h" |
| 55 | #include "llvm/Support/CommandLine.h" |
| 56 | #include "llvm/Support/Compiler.h" |
| 57 | #include "llvm/Support/Debug.h" |
| 58 | #include "llvm/Support/ErrorHandling.h" |
| 59 | #include "llvm/Support/raw_ostream.h" |
| 60 | #include "llvm/Transforms/IPO.h" |
| 61 | #include "llvm/Transforms/Utils/Local.h" |
| 62 | #include <cassert> |
| 63 | #include <iterator> |
| 64 | #include <map> |
| 65 | #include <optional> |
| 66 | #include <vector> |
| 67 | |
| 68 | using namespace llvm; |
| 69 | |
| 70 | #define DEBUG_TYPE "function-attrs" |
| 71 | |
| 72 | STATISTIC(NumMemoryAttr, "Number of functions with improved memory attribute" ); |
| 73 | STATISTIC(NumCapturesNone, "Number of arguments marked captures(none)" ); |
| 74 | STATISTIC(NumCapturesPartial, "Number of arguments marked with captures " |
| 75 | "attribute other than captures(none)" ); |
| 76 | STATISTIC(NumReturned, "Number of arguments marked returned" ); |
| 77 | STATISTIC(NumReadNoneArg, "Number of arguments marked readnone" ); |
| 78 | STATISTIC(NumReadOnlyArg, "Number of arguments marked readonly" ); |
| 79 | STATISTIC(NumWriteOnlyArg, "Number of arguments marked writeonly" ); |
| 80 | STATISTIC(NumNoAlias, "Number of function returns marked noalias" ); |
| 81 | STATISTIC(NumNonNullReturn, "Number of function returns marked nonnull" ); |
| 82 | STATISTIC(NumNoUndefReturn, "Number of function returns marked noundef" ); |
| 83 | STATISTIC(NumNoRecurse, "Number of functions marked as norecurse" ); |
| 84 | STATISTIC(NumNoUnwind, "Number of functions marked as nounwind" ); |
| 85 | STATISTIC(NumNoFree, "Number of functions marked as nofree" ); |
| 86 | STATISTIC(NumWillReturn, "Number of functions marked as willreturn" ); |
| 87 | STATISTIC(NumNoSync, "Number of functions marked as nosync" ); |
| 88 | STATISTIC(NumCold, "Number of functions marked as cold" ); |
| 89 | |
| 90 | STATISTIC(NumThinLinkNoRecurse, |
| 91 | "Number of functions marked as norecurse during thinlink" ); |
| 92 | STATISTIC(NumThinLinkNoUnwind, |
| 93 | "Number of functions marked as nounwind during thinlink" ); |
| 94 | |
| 95 | static cl::opt<bool> EnableNonnullArgPropagation( |
| 96 | "enable-nonnull-arg-prop" , cl::init(Val: true), cl::Hidden, |
| 97 | cl::desc("Try to propagate nonnull argument attributes from callsites to " |
| 98 | "caller functions." )); |
| 99 | |
| 100 | static cl::opt<bool> DisableNoUnwindInference( |
| 101 | "disable-nounwind-inference" , cl::Hidden, |
| 102 | cl::desc("Stop inferring nounwind attribute during function-attrs pass" )); |
| 103 | |
| 104 | static cl::opt<bool> DisableNoFreeInference( |
| 105 | "disable-nofree-inference" , cl::Hidden, |
| 106 | cl::desc("Stop inferring nofree attribute during function-attrs pass" )); |
| 107 | |
| 108 | static cl::opt<bool> DisableThinLTOPropagation( |
| 109 | "disable-thinlto-funcattrs" , cl::init(Val: true), cl::Hidden, |
| 110 | cl::desc("Don't propagate function-attrs in thinLTO" )); |
| 111 | |
| 112 | static void addCapturesStat(CaptureInfo CI) { |
| 113 | if (capturesNothing(CC: CI)) |
| 114 | ++NumCapturesNone; |
| 115 | else |
| 116 | ++NumCapturesPartial; |
| 117 | } |
| 118 | |
| 119 | namespace { |
| 120 | |
| 121 | using SCCNodeSet = SmallSetVector<Function *, 8>; |
| 122 | |
| 123 | } // end anonymous namespace |
| 124 | |
| 125 | static void addLocAccess(MemoryEffects &ME, const MemoryLocation &Loc, |
| 126 | ModRefInfo MR, AAResults &AAR) { |
| 127 | // Ignore accesses to known-invariant or local memory. |
| 128 | MR &= AAR.getModRefInfoMask(Loc, /*IgnoreLocal=*/IgnoreLocals: true); |
| 129 | if (isNoModRef(MRI: MR)) |
| 130 | return; |
| 131 | |
| 132 | const Value *UO = getUnderlyingObjectAggressive(V: Loc.Ptr); |
| 133 | if (isa<AllocaInst>(Val: UO)) |
| 134 | return; |
| 135 | if (isa<Argument>(Val: UO)) { |
| 136 | ME |= MemoryEffects::argMemOnly(MR); |
| 137 | return; |
| 138 | } |
| 139 | |
| 140 | // If it's not an identified object, it might be an argument. |
| 141 | if (!isIdentifiedObject(V: UO)) |
| 142 | ME |= MemoryEffects::argMemOnly(MR); |
| 143 | ME |= MemoryEffects(IRMemLocation::ErrnoMem, MR); |
| 144 | ME |= MemoryEffects(IRMemLocation::Other, MR); |
| 145 | } |
| 146 | |
| 147 | static void addArgLocs(MemoryEffects &ME, const CallBase *Call, |
| 148 | ModRefInfo ArgMR, AAResults &AAR) { |
| 149 | for (const Value *Arg : Call->args()) { |
| 150 | if (!Arg->getType()->isPtrOrPtrVectorTy()) |
| 151 | continue; |
| 152 | |
| 153 | addLocAccess(ME, |
| 154 | Loc: MemoryLocation::getBeforeOrAfter(Ptr: Arg, AATags: Call->getAAMetadata()), |
| 155 | MR: ArgMR, AAR); |
| 156 | } |
| 157 | } |
| 158 | |
| 159 | /// Returns the memory access attribute for function F using AAR for AA results, |
| 160 | /// where SCCNodes is the current SCC. |
| 161 | /// |
| 162 | /// If ThisBody is true, this function may examine the function body and will |
| 163 | /// return a result pertaining to this copy of the function. If it is false, the |
| 164 | /// result will be based only on AA results for the function declaration; it |
| 165 | /// will be assumed that some other (perhaps less optimized) version of the |
| 166 | /// function may be selected at link time. |
| 167 | /// |
| 168 | /// The return value is split into two parts: Memory effects that always apply, |
| 169 | /// and additional memory effects that apply if any of the functions in the SCC |
| 170 | /// can access argmem. |
| 171 | static std::pair<MemoryEffects, MemoryEffects> |
| 172 | checkFunctionMemoryAccess(Function &F, bool ThisBody, AAResults &AAR, |
| 173 | const SCCNodeSet &SCCNodes) { |
| 174 | MemoryEffects OrigME = AAR.getMemoryEffects(F: &F); |
| 175 | if (OrigME.doesNotAccessMemory()) |
| 176 | // Already perfect! |
| 177 | return {OrigME, MemoryEffects::none()}; |
| 178 | |
| 179 | if (!ThisBody) |
| 180 | return {OrigME, MemoryEffects::none()}; |
| 181 | |
| 182 | MemoryEffects ME = MemoryEffects::none(); |
| 183 | // Additional locations accessed if the SCC accesses argmem. |
| 184 | MemoryEffects RecursiveArgME = MemoryEffects::none(); |
| 185 | |
| 186 | // Inalloca and preallocated arguments are always clobbered by the call. |
| 187 | if (F.getAttributes().hasAttrSomewhere(Kind: Attribute::InAlloca) || |
| 188 | F.getAttributes().hasAttrSomewhere(Kind: Attribute::Preallocated)) |
| 189 | ME |= MemoryEffects::argMemOnly(MR: ModRefInfo::ModRef); |
| 190 | |
| 191 | // Scan the function body for instructions that may read or write memory. |
| 192 | for (Instruction &I : instructions(F)) { |
| 193 | // Some instructions can be ignored even if they read or write memory. |
| 194 | // Detect these now, skipping to the next instruction if one is found. |
| 195 | if (auto *Call = dyn_cast<CallBase>(Val: &I)) { |
| 196 | // We can optimistically ignore calls to functions in the same SCC, with |
| 197 | // two caveats: |
| 198 | // * Calls with operand bundles may have additional effects. |
| 199 | // * Argument memory accesses may imply additional effects depending on |
| 200 | // what the argument location is. |
| 201 | if (!Call->hasOperandBundles() && Call->getCalledFunction() && |
| 202 | SCCNodes.count(key: Call->getCalledFunction())) { |
| 203 | // Keep track of which additional locations are accessed if the SCC |
| 204 | // turns out to access argmem. |
| 205 | addArgLocs(ME&: RecursiveArgME, Call, ArgMR: ModRefInfo::ModRef, AAR); |
| 206 | continue; |
| 207 | } |
| 208 | |
| 209 | MemoryEffects CallME = AAR.getMemoryEffects(Call); |
| 210 | |
| 211 | // If the call doesn't access memory, we're done. |
| 212 | if (CallME.doesNotAccessMemory()) |
| 213 | continue; |
| 214 | |
| 215 | // A pseudo probe call shouldn't change any function attribute since it |
| 216 | // doesn't translate to a real instruction. It comes with a memory access |
| 217 | // tag to prevent itself being removed by optimizations and not block |
| 218 | // other instructions being optimized. |
| 219 | if (isa<PseudoProbeInst>(Val: I)) |
| 220 | continue; |
| 221 | |
| 222 | // Merge callee's memory effects into caller's ones, including |
| 223 | // inaccessible and errno memory, but excluding argument memory, which is |
| 224 | // handled separately. |
| 225 | ME |= CallME.getWithoutLoc(Loc: IRMemLocation::ArgMem); |
| 226 | |
| 227 | // If the call accesses captured memory (currently part of "other") and |
| 228 | // an argument is captured (currently not tracked), then it may also |
| 229 | // access argument memory. |
| 230 | ModRefInfo OtherMR = CallME.getModRef(Loc: IRMemLocation::Other); |
| 231 | ME |= MemoryEffects::argMemOnly(MR: OtherMR); |
| 232 | |
| 233 | // Check whether all pointer arguments point to local memory, and |
| 234 | // ignore calls that only access local memory. |
| 235 | ModRefInfo ArgMR = CallME.getModRef(Loc: IRMemLocation::ArgMem); |
| 236 | if (ArgMR != ModRefInfo::NoModRef) |
| 237 | addArgLocs(ME, Call, ArgMR, AAR); |
| 238 | continue; |
| 239 | } |
| 240 | |
| 241 | ModRefInfo MR = ModRefInfo::NoModRef; |
| 242 | if (I.mayWriteToMemory()) |
| 243 | MR |= ModRefInfo::Mod; |
| 244 | if (I.mayReadFromMemory()) |
| 245 | MR |= ModRefInfo::Ref; |
| 246 | if (MR == ModRefInfo::NoModRef) |
| 247 | continue; |
| 248 | |
| 249 | std::optional<MemoryLocation> Loc = MemoryLocation::getOrNone(Inst: &I); |
| 250 | if (!Loc) { |
| 251 | // If no location is known, conservatively assume anything can be |
| 252 | // accessed. |
| 253 | ME |= MemoryEffects(MR); |
| 254 | continue; |
| 255 | } |
| 256 | |
| 257 | // Volatile operations may access inaccessible memory. |
| 258 | if (I.isVolatile()) |
| 259 | ME |= MemoryEffects::inaccessibleMemOnly(MR); |
| 260 | |
| 261 | addLocAccess(ME, Loc: *Loc, MR, AAR); |
| 262 | } |
| 263 | |
| 264 | return {OrigME & ME, RecursiveArgME}; |
| 265 | } |
| 266 | |
| 267 | MemoryEffects llvm::computeFunctionBodyMemoryAccess(Function &F, |
| 268 | AAResults &AAR) { |
| 269 | return checkFunctionMemoryAccess(F, /*ThisBody=*/true, AAR, SCCNodes: {}).first; |
| 270 | } |
| 271 | |
| 272 | /// Deduce readonly/readnone/writeonly attributes for the SCC. |
| 273 | template <typename AARGetterT> |
| 274 | static void addMemoryAttrs(const SCCNodeSet &SCCNodes, AARGetterT &&AARGetter, |
| 275 | SmallPtrSet<Function *, 8> &Changed) { |
| 276 | MemoryEffects ME = MemoryEffects::none(); |
| 277 | MemoryEffects RecursiveArgME = MemoryEffects::none(); |
| 278 | for (Function *F : SCCNodes) { |
| 279 | // Call the callable parameter to look up AA results for this function. |
| 280 | AAResults &AAR = AARGetter(*F); |
| 281 | // Non-exact function definitions may not be selected at link time, and an |
| 282 | // alternative version that writes to memory may be selected. See the |
| 283 | // comment on GlobalValue::isDefinitionExact for more details. |
| 284 | auto [FnME, FnRecursiveArgME] = |
| 285 | checkFunctionMemoryAccess(F&: *F, ThisBody: F->hasExactDefinition(), AAR, SCCNodes); |
| 286 | ME |= FnME; |
| 287 | RecursiveArgME |= FnRecursiveArgME; |
| 288 | // Reached bottom of the lattice, we will not be able to improve the result. |
| 289 | if (ME == MemoryEffects::unknown()) |
| 290 | return; |
| 291 | } |
| 292 | |
| 293 | // If the SCC accesses argmem, add recursive accesses resulting from that. |
| 294 | ModRefInfo ArgMR = ME.getModRef(Loc: IRMemLocation::ArgMem); |
| 295 | if (ArgMR != ModRefInfo::NoModRef) |
| 296 | ME |= RecursiveArgME & MemoryEffects(ArgMR); |
| 297 | |
| 298 | for (Function *F : SCCNodes) { |
| 299 | MemoryEffects OldME = F->getMemoryEffects(); |
| 300 | MemoryEffects NewME = ME & OldME; |
| 301 | if (NewME != OldME) { |
| 302 | ++NumMemoryAttr; |
| 303 | F->setMemoryEffects(NewME); |
| 304 | // Remove conflicting writable attributes. |
| 305 | if (!isModSet(MRI: NewME.getModRef(Loc: IRMemLocation::ArgMem))) |
| 306 | for (Argument &A : F->args()) |
| 307 | A.removeAttr(Kind: Attribute::Writable); |
| 308 | Changed.insert(Ptr: F); |
| 309 | } |
| 310 | } |
| 311 | } |
| 312 | |
| 313 | // Compute definitive function attributes for a function taking into account |
| 314 | // prevailing definitions and linkage types |
| 315 | static FunctionSummary *calculatePrevailingSummary( |
| 316 | ValueInfo VI, |
| 317 | DenseMap<ValueInfo, FunctionSummary *> &CachedPrevailingSummary, |
| 318 | function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
| 319 | IsPrevailing) { |
| 320 | |
| 321 | auto [It, Inserted] = CachedPrevailingSummary.try_emplace(Key: VI); |
| 322 | if (!Inserted) |
| 323 | return It->second; |
| 324 | |
| 325 | /// At this point, prevailing symbols have been resolved. The following leads |
| 326 | /// to returning a conservative result: |
| 327 | /// - Multiple instances with local linkage. Normally local linkage would be |
| 328 | /// unique per module |
| 329 | /// as the GUID includes the module path. We could have a guid alias if |
| 330 | /// there wasn't any distinguishing path when each file was compiled, but |
| 331 | /// that should be rare so we'll punt on those. |
| 332 | |
| 333 | /// These next 2 cases should not happen and will assert: |
| 334 | /// - Multiple instances with external linkage. This should be caught in |
| 335 | /// symbol resolution |
| 336 | /// - Non-existent FunctionSummary for Aliasee. This presents a hole in our |
| 337 | /// knowledge meaning we have to go conservative. |
| 338 | |
| 339 | /// Otherwise, we calculate attributes for a function as: |
| 340 | /// 1. If we have a local linkage, take its attributes. If there's somehow |
| 341 | /// multiple, bail and go conservative. |
| 342 | /// 2. If we have an external/WeakODR/LinkOnceODR linkage check that it is |
| 343 | /// prevailing, take its attributes. |
| 344 | /// 3. If we have a Weak/LinkOnce linkage the copies can have semantic |
| 345 | /// differences. However, if the prevailing copy is known it will be used |
| 346 | /// so take its attributes. If the prevailing copy is in a native file |
| 347 | /// all IR copies will be dead and propagation will go conservative. |
| 348 | /// 4. AvailableExternally summaries without a prevailing copy are known to |
| 349 | /// occur in a couple of circumstances: |
| 350 | /// a. An internal function gets imported due to its caller getting |
| 351 | /// imported, it becomes AvailableExternally but no prevailing |
| 352 | /// definition exists. Because it has to get imported along with its |
| 353 | /// caller the attributes will be captured by propagating on its |
| 354 | /// caller. |
| 355 | /// b. C++11 [temp.explicit]p10 can generate AvailableExternally |
| 356 | /// definitions of explicitly instanced template declarations |
| 357 | /// for inlining which are ultimately dropped from the TU. Since this |
| 358 | /// is localized to the TU the attributes will have already made it to |
| 359 | /// the callers. |
| 360 | /// These are edge cases and already captured by their callers so we |
| 361 | /// ignore these for now. If they become relevant to optimize in the |
| 362 | /// future this can be revisited. |
| 363 | /// 5. Otherwise, go conservative. |
| 364 | |
| 365 | FunctionSummary *Local = nullptr; |
| 366 | FunctionSummary *Prevailing = nullptr; |
| 367 | |
| 368 | for (const auto &GVS : VI.getSummaryList()) { |
| 369 | if (!GVS->isLive()) |
| 370 | continue; |
| 371 | |
| 372 | FunctionSummary *FS = dyn_cast<FunctionSummary>(Val: GVS->getBaseObject()); |
| 373 | // Virtual and Unknown (e.g. indirect) calls require going conservative |
| 374 | if (!FS || FS->fflags().HasUnknownCall) |
| 375 | return nullptr; |
| 376 | |
| 377 | const auto &Linkage = GVS->linkage(); |
| 378 | if (GlobalValue::isLocalLinkage(Linkage)) { |
| 379 | if (Local) { |
| 380 | LLVM_DEBUG( |
| 381 | dbgs() |
| 382 | << "ThinLTO FunctionAttrs: Multiple Local Linkage, bailing on " |
| 383 | "function " |
| 384 | << VI.name() << " from " << FS->modulePath() << ". Previous module " |
| 385 | << Local->modulePath() << "\n" ); |
| 386 | return nullptr; |
| 387 | } |
| 388 | Local = FS; |
| 389 | } else if (GlobalValue::isExternalLinkage(Linkage)) { |
| 390 | assert(IsPrevailing(VI.getGUID(), GVS.get())); |
| 391 | Prevailing = FS; |
| 392 | break; |
| 393 | } else if (GlobalValue::isWeakODRLinkage(Linkage) || |
| 394 | GlobalValue::isLinkOnceODRLinkage(Linkage) || |
| 395 | GlobalValue::isWeakAnyLinkage(Linkage) || |
| 396 | GlobalValue::isLinkOnceAnyLinkage(Linkage)) { |
| 397 | if (IsPrevailing(VI.getGUID(), GVS.get())) { |
| 398 | Prevailing = FS; |
| 399 | break; |
| 400 | } |
| 401 | } else if (GlobalValue::isAvailableExternallyLinkage(Linkage)) { |
| 402 | // TODO: Handle these cases if they become meaningful |
| 403 | continue; |
| 404 | } |
| 405 | } |
| 406 | |
| 407 | auto &CPS = CachedPrevailingSummary[VI]; |
| 408 | if (Local) { |
| 409 | assert(!Prevailing); |
| 410 | CPS = Local; |
| 411 | } else if (Prevailing) { |
| 412 | assert(!Local); |
| 413 | CPS = Prevailing; |
| 414 | } |
| 415 | |
| 416 | return CPS; |
| 417 | } |
| 418 | |
| 419 | bool llvm::thinLTOPropagateFunctionAttrs( |
| 420 | ModuleSummaryIndex &Index, |
| 421 | function_ref<bool(GlobalValue::GUID, const GlobalValueSummary *)> |
| 422 | IsPrevailing) { |
| 423 | // TODO: implement addNoAliasAttrs once |
| 424 | // there's more information about the return type in the summary |
| 425 | if (DisableThinLTOPropagation) |
| 426 | return false; |
| 427 | |
| 428 | DenseMap<ValueInfo, FunctionSummary *> CachedPrevailingSummary; |
| 429 | bool Changed = false; |
| 430 | |
| 431 | auto PropagateAttributes = [&](std::vector<ValueInfo> &SCCNodes) { |
| 432 | // Assume we can propagate unless we discover otherwise |
| 433 | FunctionSummary::FFlags InferredFlags; |
| 434 | InferredFlags.NoRecurse = (SCCNodes.size() == 1); |
| 435 | InferredFlags.NoUnwind = true; |
| 436 | |
| 437 | for (auto &V : SCCNodes) { |
| 438 | FunctionSummary *CallerSummary = |
| 439 | calculatePrevailingSummary(VI: V, CachedPrevailingSummary, IsPrevailing); |
| 440 | |
| 441 | // Function summaries can fail to contain information such as declarations |
| 442 | if (!CallerSummary) |
| 443 | return; |
| 444 | |
| 445 | if (CallerSummary->fflags().MayThrow) |
| 446 | InferredFlags.NoUnwind = false; |
| 447 | |
| 448 | for (const auto &Callee : CallerSummary->calls()) { |
| 449 | FunctionSummary *CalleeSummary = calculatePrevailingSummary( |
| 450 | VI: Callee.first, CachedPrevailingSummary, IsPrevailing); |
| 451 | |
| 452 | if (!CalleeSummary) |
| 453 | return; |
| 454 | |
| 455 | if (!CalleeSummary->fflags().NoRecurse) |
| 456 | InferredFlags.NoRecurse = false; |
| 457 | |
| 458 | if (!CalleeSummary->fflags().NoUnwind) |
| 459 | InferredFlags.NoUnwind = false; |
| 460 | |
| 461 | if (!InferredFlags.NoUnwind && !InferredFlags.NoRecurse) |
| 462 | break; |
| 463 | } |
| 464 | } |
| 465 | |
| 466 | if (InferredFlags.NoUnwind || InferredFlags.NoRecurse) { |
| 467 | Changed = true; |
| 468 | for (auto &V : SCCNodes) { |
| 469 | if (InferredFlags.NoRecurse) { |
| 470 | LLVM_DEBUG(dbgs() << "ThinLTO FunctionAttrs: Propagated NoRecurse to " |
| 471 | << V.name() << "\n" ); |
| 472 | ++NumThinLinkNoRecurse; |
| 473 | } |
| 474 | |
| 475 | if (InferredFlags.NoUnwind) { |
| 476 | LLVM_DEBUG(dbgs() << "ThinLTO FunctionAttrs: Propagated NoUnwind to " |
| 477 | << V.name() << "\n" ); |
| 478 | ++NumThinLinkNoUnwind; |
| 479 | } |
| 480 | |
| 481 | for (const auto &S : V.getSummaryList()) { |
| 482 | if (auto *FS = dyn_cast<FunctionSummary>(Val: S.get())) { |
| 483 | if (InferredFlags.NoRecurse) |
| 484 | FS->setNoRecurse(); |
| 485 | |
| 486 | if (InferredFlags.NoUnwind) |
| 487 | FS->setNoUnwind(); |
| 488 | } |
| 489 | } |
| 490 | } |
| 491 | } |
| 492 | }; |
| 493 | |
| 494 | // Call propagation functions on each SCC in the Index |
| 495 | for (scc_iterator<ModuleSummaryIndex *> I = scc_begin(G: &Index); !I.isAtEnd(); |
| 496 | ++I) { |
| 497 | std::vector<ValueInfo> Nodes(*I); |
| 498 | PropagateAttributes(Nodes); |
| 499 | } |
| 500 | return Changed; |
| 501 | } |
| 502 | |
| 503 | namespace { |
| 504 | |
| 505 | /// For a given pointer Argument, this retains a list of Arguments of functions |
| 506 | /// in the same SCC that the pointer data flows into. We use this to build an |
| 507 | /// SCC of the arguments. |
| 508 | struct ArgumentGraphNode { |
| 509 | Argument *Definition; |
| 510 | /// CaptureComponents for this argument, excluding captures via Uses. |
| 511 | /// We don't distinguish between other/return captures here. |
| 512 | CaptureComponents CC = CaptureComponents::None; |
| 513 | SmallVector<ArgumentGraphNode *, 4> Uses; |
| 514 | }; |
| 515 | |
| 516 | class ArgumentGraph { |
| 517 | // We store pointers to ArgumentGraphNode objects, so it's important that |
| 518 | // that they not move around upon insert. |
| 519 | using ArgumentMapTy = std::map<Argument *, ArgumentGraphNode>; |
| 520 | |
| 521 | ArgumentMapTy ArgumentMap; |
| 522 | |
| 523 | // There is no root node for the argument graph, in fact: |
| 524 | // void f(int *x, int *y) { if (...) f(x, y); } |
| 525 | // is an example where the graph is disconnected. The SCCIterator requires a |
| 526 | // single entry point, so we maintain a fake ("synthetic") root node that |
| 527 | // uses every node. Because the graph is directed and nothing points into |
| 528 | // the root, it will not participate in any SCCs (except for its own). |
| 529 | ArgumentGraphNode SyntheticRoot; |
| 530 | |
| 531 | public: |
| 532 | ArgumentGraph() { SyntheticRoot.Definition = nullptr; } |
| 533 | |
| 534 | using iterator = SmallVectorImpl<ArgumentGraphNode *>::iterator; |
| 535 | |
| 536 | iterator begin() { return SyntheticRoot.Uses.begin(); } |
| 537 | iterator end() { return SyntheticRoot.Uses.end(); } |
| 538 | ArgumentGraphNode *getEntryNode() { return &SyntheticRoot; } |
| 539 | |
| 540 | ArgumentGraphNode *operator[](Argument *A) { |
| 541 | ArgumentGraphNode &Node = ArgumentMap[A]; |
| 542 | Node.Definition = A; |
| 543 | SyntheticRoot.Uses.push_back(Elt: &Node); |
| 544 | return &Node; |
| 545 | } |
| 546 | }; |
| 547 | |
| 548 | /// This tracker checks whether callees are in the SCC, and if so it does not |
| 549 | /// consider that a capture, instead adding it to the "Uses" list and |
| 550 | /// continuing with the analysis. |
| 551 | struct ArgumentUsesTracker : public CaptureTracker { |
| 552 | ArgumentUsesTracker(const SCCNodeSet &SCCNodes) : SCCNodes(SCCNodes) {} |
| 553 | |
| 554 | void tooManyUses() override { CI = CaptureInfo::all(); } |
| 555 | |
| 556 | Action captured(const Use *U, UseCaptureInfo UseCI) override { |
| 557 | if (updateCaptureInfo(U, CC: UseCI.UseCC)) { |
| 558 | // Don't bother continuing if we already capture everything. |
| 559 | if (capturesAll(CC: CI.getOtherComponents())) |
| 560 | return Stop; |
| 561 | return Continue; |
| 562 | } |
| 563 | |
| 564 | // For SCC argument tracking, we're not going to analyze other/ret |
| 565 | // components separately, so don't follow the return value. |
| 566 | return ContinueIgnoringReturn; |
| 567 | } |
| 568 | |
| 569 | bool updateCaptureInfo(const Use *U, CaptureComponents CC) { |
| 570 | CallBase *CB = dyn_cast<CallBase>(Val: U->getUser()); |
| 571 | if (!CB) { |
| 572 | if (isa<ReturnInst>(Val: U->getUser())) |
| 573 | CI |= CaptureInfo::retOnly(RetComponents: CC); |
| 574 | else |
| 575 | // Conservatively assume that the captured value might make its way |
| 576 | // into the return value as well. This could be made more precise. |
| 577 | CI |= CaptureInfo(CC); |
| 578 | return true; |
| 579 | } |
| 580 | |
| 581 | Function *F = CB->getCalledFunction(); |
| 582 | if (!F || !F->hasExactDefinition() || !SCCNodes.count(key: F)) { |
| 583 | CI |= CaptureInfo(CC); |
| 584 | return true; |
| 585 | } |
| 586 | |
| 587 | assert(!CB->isCallee(U) && "callee operand reported captured?" ); |
| 588 | const unsigned UseIndex = CB->getDataOperandNo(U); |
| 589 | if (UseIndex >= CB->arg_size()) { |
| 590 | // Data operand, but not a argument operand -- must be a bundle operand |
| 591 | assert(CB->hasOperandBundles() && "Must be!" ); |
| 592 | |
| 593 | // CaptureTracking told us that we're being captured by an operand bundle |
| 594 | // use. In this case it does not matter if the callee is within our SCC |
| 595 | // or not -- we've been captured in some unknown way, and we have to be |
| 596 | // conservative. |
| 597 | CI |= CaptureInfo(CC); |
| 598 | return true; |
| 599 | } |
| 600 | |
| 601 | if (UseIndex >= F->arg_size()) { |
| 602 | assert(F->isVarArg() && "More params than args in non-varargs call" ); |
| 603 | CI |= CaptureInfo(CC); |
| 604 | return true; |
| 605 | } |
| 606 | |
| 607 | // TODO(captures): Could improve precision by remembering maximum |
| 608 | // capture components for the argument. |
| 609 | Uses.push_back(Elt: &*std::next(x: F->arg_begin(), n: UseIndex)); |
| 610 | return false; |
| 611 | } |
| 612 | |
| 613 | // Does not include potential captures via Uses in the SCC. |
| 614 | CaptureInfo CI = CaptureInfo::none(); |
| 615 | |
| 616 | // Uses within our SCC. |
| 617 | SmallVector<Argument *, 4> Uses; |
| 618 | |
| 619 | const SCCNodeSet &SCCNodes; |
| 620 | }; |
| 621 | |
| 622 | /// A struct of argument use: a Use and the offset it accesses. This struct |
| 623 | /// is to track uses inside function via GEP. If GEP has a non-constant index, |
| 624 | /// the Offset field is nullopt. |
| 625 | struct ArgumentUse { |
| 626 | Use *U; |
| 627 | std::optional<int64_t> Offset; |
| 628 | }; |
| 629 | |
| 630 | /// A struct of argument access info. "Unknown" accesses are the cases like |
| 631 | /// unrecognized instructions, instructions that have more than one use of |
| 632 | /// the argument, or volatile memory accesses. "WriteWithSideEffect" are call |
| 633 | /// instructions that not only write an argument but also capture it. |
| 634 | struct ArgumentAccessInfo { |
| 635 | enum class AccessType : uint8_t { Write, WriteWithSideEffect, Read, Unknown }; |
| 636 | AccessType ArgAccessType; |
| 637 | ConstantRangeList AccessRanges; |
| 638 | }; |
| 639 | |
| 640 | /// A struct to wrap the argument use info per block. |
| 641 | struct UsesPerBlockInfo { |
| 642 | SmallDenseMap<Instruction *, ArgumentAccessInfo, 4> Insts; |
| 643 | bool HasWrites = false; |
| 644 | bool HasUnknownAccess = false; |
| 645 | }; |
| 646 | |
| 647 | /// A struct to summarize the argument use info in a function. |
| 648 | struct ArgumentUsesSummary { |
| 649 | bool HasAnyWrite = false; |
| 650 | bool HasWriteOutsideEntryBB = false; |
| 651 | SmallDenseMap<const BasicBlock *, UsesPerBlockInfo, 16> UsesPerBlock; |
| 652 | }; |
| 653 | |
| 654 | ArgumentAccessInfo getArgumentAccessInfo(const Instruction *I, |
| 655 | const ArgumentUse &ArgUse, |
| 656 | const DataLayout &DL) { |
| 657 | auto GetTypeAccessRange = |
| 658 | [&DL](Type *Ty, |
| 659 | std::optional<int64_t> Offset) -> std::optional<ConstantRange> { |
| 660 | auto TypeSize = DL.getTypeStoreSize(Ty); |
| 661 | if (!TypeSize.isScalable() && Offset) { |
| 662 | int64_t Size = TypeSize.getFixedValue(); |
| 663 | APInt Low(64, *Offset, true); |
| 664 | bool Overflow; |
| 665 | APInt High = Low.sadd_ov(RHS: APInt(64, Size, true), Overflow); |
| 666 | // Bail if the range overflows signed 64-bit int. |
| 667 | if (Overflow) |
| 668 | return std::nullopt; |
| 669 | return ConstantRange(Low, High); |
| 670 | } |
| 671 | return std::nullopt; |
| 672 | }; |
| 673 | auto GetConstantIntRange = |
| 674 | [](Value *Length, |
| 675 | std::optional<int64_t> Offset) -> std::optional<ConstantRange> { |
| 676 | auto *ConstantLength = dyn_cast<ConstantInt>(Val: Length); |
| 677 | if (ConstantLength && Offset) { |
| 678 | int64_t Len = ConstantLength->getSExtValue(); |
| 679 | |
| 680 | // Reject zero or negative lengths |
| 681 | if (Len <= 0) |
| 682 | return std::nullopt; |
| 683 | |
| 684 | APInt Low(64, *Offset, true); |
| 685 | bool Overflow; |
| 686 | APInt High = Low.sadd_ov(RHS: APInt(64, Len, true), Overflow); |
| 687 | if (Overflow) |
| 688 | return std::nullopt; |
| 689 | |
| 690 | return ConstantRange(Low, High); |
| 691 | } |
| 692 | return std::nullopt; |
| 693 | }; |
| 694 | |
| 695 | if (auto *SI = dyn_cast<StoreInst>(Val: I)) { |
| 696 | if (SI->isSimple() && &SI->getOperandUse(i: 1) == ArgUse.U) { |
| 697 | // Get the fixed type size of "SI". Since the access range of a write |
| 698 | // will be unioned, if "SI" doesn't have a fixed type size, we just set |
| 699 | // the access range to empty. |
| 700 | ConstantRangeList AccessRanges; |
| 701 | if (auto TypeAccessRange = |
| 702 | GetTypeAccessRange(SI->getAccessType(), ArgUse.Offset)) |
| 703 | AccessRanges.insert(NewRange: *TypeAccessRange); |
| 704 | return {.ArgAccessType: ArgumentAccessInfo::AccessType::Write, .AccessRanges: std::move(AccessRanges)}; |
| 705 | } |
| 706 | } else if (auto *LI = dyn_cast<LoadInst>(Val: I)) { |
| 707 | if (LI->isSimple()) { |
| 708 | assert(&LI->getOperandUse(0) == ArgUse.U); |
| 709 | // Get the fixed type size of "LI". Different from Write, if "LI" |
| 710 | // doesn't have a fixed type size, we conservatively set as a clobber |
| 711 | // with an empty access range. |
| 712 | if (auto TypeAccessRange = |
| 713 | GetTypeAccessRange(LI->getAccessType(), ArgUse.Offset)) |
| 714 | return {.ArgAccessType: ArgumentAccessInfo::AccessType::Read, .AccessRanges: {*TypeAccessRange}}; |
| 715 | } |
| 716 | } else if (auto *MemSet = dyn_cast<MemSetInst>(Val: I)) { |
| 717 | if (!MemSet->isVolatile()) { |
| 718 | ConstantRangeList AccessRanges; |
| 719 | if (auto AccessRange = |
| 720 | GetConstantIntRange(MemSet->getLength(), ArgUse.Offset)) |
| 721 | AccessRanges.insert(NewRange: *AccessRange); |
| 722 | return {.ArgAccessType: ArgumentAccessInfo::AccessType::Write, .AccessRanges: AccessRanges}; |
| 723 | } |
| 724 | } else if (auto *MTI = dyn_cast<MemTransferInst>(Val: I)) { |
| 725 | if (!MTI->isVolatile()) { |
| 726 | if (&MTI->getOperandUse(i: 0) == ArgUse.U) { |
| 727 | ConstantRangeList AccessRanges; |
| 728 | if (auto AccessRange = |
| 729 | GetConstantIntRange(MTI->getLength(), ArgUse.Offset)) |
| 730 | AccessRanges.insert(NewRange: *AccessRange); |
| 731 | return {.ArgAccessType: ArgumentAccessInfo::AccessType::Write, .AccessRanges: AccessRanges}; |
| 732 | } else if (&MTI->getOperandUse(i: 1) == ArgUse.U) { |
| 733 | if (auto AccessRange = |
| 734 | GetConstantIntRange(MTI->getLength(), ArgUse.Offset)) |
| 735 | return {.ArgAccessType: ArgumentAccessInfo::AccessType::Read, .AccessRanges: {*AccessRange}}; |
| 736 | } |
| 737 | } |
| 738 | } else if (auto *CB = dyn_cast<CallBase>(Val: I)) { |
| 739 | if (CB->isArgOperand(U: ArgUse.U) && |
| 740 | !CB->isByValArgument(ArgNo: CB->getArgOperandNo(U: ArgUse.U))) { |
| 741 | unsigned ArgNo = CB->getArgOperandNo(U: ArgUse.U); |
| 742 | bool IsInitialize = CB->paramHasAttr(ArgNo, Kind: Attribute::Initializes); |
| 743 | if (IsInitialize && ArgUse.Offset) { |
| 744 | // Argument is a Write when parameter is writeonly/readnone |
| 745 | // and nocapture. Otherwise, it's a WriteWithSideEffect. |
| 746 | auto Access = CB->onlyWritesMemory(OpNo: ArgNo) && CB->doesNotCapture(OpNo: ArgNo) |
| 747 | ? ArgumentAccessInfo::AccessType::Write |
| 748 | : ArgumentAccessInfo::AccessType::WriteWithSideEffect; |
| 749 | ConstantRangeList AccessRanges; |
| 750 | Attribute Attr = CB->getParamAttr(ArgNo, Kind: Attribute::Initializes); |
| 751 | ConstantRangeList CBCRL = Attr.getValueAsConstantRangeList(); |
| 752 | for (ConstantRange &CR : CBCRL) |
| 753 | AccessRanges.insert(NewRange: ConstantRange(CR.getLower() + *ArgUse.Offset, |
| 754 | CR.getUpper() + *ArgUse.Offset)); |
| 755 | return {.ArgAccessType: Access, .AccessRanges: AccessRanges}; |
| 756 | } |
| 757 | } |
| 758 | } |
| 759 | // Other unrecognized instructions are considered as unknown. |
| 760 | return {.ArgAccessType: ArgumentAccessInfo::AccessType::Unknown, .AccessRanges: {}}; |
| 761 | } |
| 762 | |
| 763 | // Collect the uses of argument "A" in "F". |
| 764 | ArgumentUsesSummary collectArgumentUsesPerBlock(Argument &A, Function &F) { |
| 765 | auto &DL = F.getParent()->getDataLayout(); |
| 766 | unsigned PointerSize = |
| 767 | DL.getIndexSizeInBits(AS: A.getType()->getPointerAddressSpace()); |
| 768 | ArgumentUsesSummary Result; |
| 769 | |
| 770 | BasicBlock &EntryBB = F.getEntryBlock(); |
| 771 | SmallVector<ArgumentUse, 4> Worklist; |
| 772 | for (Use &U : A.uses()) |
| 773 | Worklist.push_back(Elt: {.U: &U, .Offset: 0}); |
| 774 | |
| 775 | // Update "UsesPerBlock" with the block of "I" as key and "Info" as value. |
| 776 | // Return true if the block of "I" has write accesses after updating. |
| 777 | auto UpdateUseInfo = [&Result](Instruction *I, ArgumentAccessInfo Info) { |
| 778 | auto *BB = I->getParent(); |
| 779 | auto &BBInfo = Result.UsesPerBlock[BB]; |
| 780 | auto [It, Inserted] = BBInfo.Insts.try_emplace(Key: I); |
| 781 | auto &IInfo = It->second; |
| 782 | |
| 783 | // Instructions that have more than one use of the argument are considered |
| 784 | // as clobbers. |
| 785 | if (!Inserted) { |
| 786 | IInfo = {.ArgAccessType: ArgumentAccessInfo::AccessType::Unknown, .AccessRanges: {}}; |
| 787 | BBInfo.HasUnknownAccess = true; |
| 788 | return false; |
| 789 | } |
| 790 | |
| 791 | IInfo = std::move(Info); |
| 792 | BBInfo.HasUnknownAccess |= |
| 793 | IInfo.ArgAccessType == ArgumentAccessInfo::AccessType::Unknown; |
| 794 | bool InfoHasWrites = |
| 795 | (IInfo.ArgAccessType == ArgumentAccessInfo::AccessType::Write || |
| 796 | IInfo.ArgAccessType == |
| 797 | ArgumentAccessInfo::AccessType::WriteWithSideEffect) && |
| 798 | !IInfo.AccessRanges.empty(); |
| 799 | BBInfo.HasWrites |= InfoHasWrites; |
| 800 | return InfoHasWrites; |
| 801 | }; |
| 802 | |
| 803 | // No need for a visited set because we don't look through phis, so there are |
| 804 | // no cycles. |
| 805 | while (!Worklist.empty()) { |
| 806 | ArgumentUse ArgUse = Worklist.pop_back_val(); |
| 807 | User *U = ArgUse.U->getUser(); |
| 808 | // Add GEP uses to worklist. |
| 809 | // If the GEP is not a constant GEP, set the ArgumentUse::Offset to nullopt. |
| 810 | if (auto *GEP = dyn_cast<GEPOperator>(Val: U)) { |
| 811 | std::optional<int64_t> NewOffset = std::nullopt; |
| 812 | if (ArgUse.Offset) { |
| 813 | APInt Offset(PointerSize, 0); |
| 814 | if (GEP->accumulateConstantOffset(DL, Offset)) |
| 815 | NewOffset = *ArgUse.Offset + Offset.getSExtValue(); |
| 816 | } |
| 817 | for (Use &U : GEP->uses()) |
| 818 | Worklist.push_back(Elt: {.U: &U, .Offset: NewOffset}); |
| 819 | continue; |
| 820 | } |
| 821 | |
| 822 | auto *I = cast<Instruction>(Val: U); |
| 823 | bool HasWrite = UpdateUseInfo(I, getArgumentAccessInfo(I, ArgUse, DL)); |
| 824 | |
| 825 | Result.HasAnyWrite |= HasWrite; |
| 826 | |
| 827 | if (HasWrite && I->getParent() != &EntryBB) |
| 828 | Result.HasWriteOutsideEntryBB = true; |
| 829 | } |
| 830 | return Result; |
| 831 | } |
| 832 | |
| 833 | } // end anonymous namespace |
| 834 | |
| 835 | namespace llvm { |
| 836 | |
| 837 | template <> struct GraphTraits<ArgumentGraphNode *> { |
| 838 | using NodeRef = ArgumentGraphNode *; |
| 839 | using ChildIteratorType = SmallVectorImpl<ArgumentGraphNode *>::iterator; |
| 840 | |
| 841 | static NodeRef getEntryNode(NodeRef A) { return A; } |
| 842 | static ChildIteratorType child_begin(NodeRef N) { return N->Uses.begin(); } |
| 843 | static ChildIteratorType child_end(NodeRef N) { return N->Uses.end(); } |
| 844 | }; |
| 845 | |
| 846 | template <> |
| 847 | struct GraphTraits<ArgumentGraph *> : public GraphTraits<ArgumentGraphNode *> { |
| 848 | static NodeRef getEntryNode(ArgumentGraph *AG) { return AG->getEntryNode(); } |
| 849 | |
| 850 | static ChildIteratorType nodes_begin(ArgumentGraph *AG) { |
| 851 | return AG->begin(); |
| 852 | } |
| 853 | |
| 854 | static ChildIteratorType nodes_end(ArgumentGraph *AG) { return AG->end(); } |
| 855 | }; |
| 856 | |
| 857 | } // end namespace llvm |
| 858 | |
| 859 | /// Returns Attribute::None, Attribute::ReadOnly or Attribute::ReadNone. |
| 860 | static Attribute::AttrKind |
| 861 | determinePointerAccessAttrs(Argument *A, |
| 862 | const SmallPtrSet<Argument *, 8> &SCCNodes) { |
| 863 | SmallVector<Use *, 32> Worklist; |
| 864 | SmallPtrSet<Use *, 32> Visited; |
| 865 | |
| 866 | // inalloca arguments are always clobbered by the call. |
| 867 | if (A->hasInAllocaAttr() || A->hasPreallocatedAttr()) |
| 868 | return Attribute::None; |
| 869 | |
| 870 | bool IsRead = false; |
| 871 | bool IsWrite = false; |
| 872 | |
| 873 | for (Use &U : A->uses()) { |
| 874 | Visited.insert(Ptr: &U); |
| 875 | Worklist.push_back(Elt: &U); |
| 876 | } |
| 877 | |
| 878 | while (!Worklist.empty()) { |
| 879 | if (IsWrite && IsRead) |
| 880 | // No point in searching further.. |
| 881 | return Attribute::None; |
| 882 | |
| 883 | Use *U = Worklist.pop_back_val(); |
| 884 | Instruction *I = cast<Instruction>(Val: U->getUser()); |
| 885 | |
| 886 | switch (I->getOpcode()) { |
| 887 | case Instruction::BitCast: |
| 888 | case Instruction::GetElementPtr: |
| 889 | case Instruction::PHI: |
| 890 | case Instruction::Select: |
| 891 | case Instruction::AddrSpaceCast: |
| 892 | // The original value is not read/written via this if the new value isn't. |
| 893 | for (Use &UU : I->uses()) |
| 894 | if (Visited.insert(Ptr: &UU).second) |
| 895 | Worklist.push_back(Elt: &UU); |
| 896 | break; |
| 897 | |
| 898 | case Instruction::Call: |
| 899 | case Instruction::Invoke: { |
| 900 | CallBase &CB = cast<CallBase>(Val&: *I); |
| 901 | if (CB.isCallee(U)) { |
| 902 | IsRead = true; |
| 903 | // Note that indirect calls do not capture, see comment in |
| 904 | // CaptureTracking for context |
| 905 | continue; |
| 906 | } |
| 907 | |
| 908 | // Given we've explicitly handled the callee operand above, what's left |
| 909 | // must be a data operand (e.g. argument or operand bundle) |
| 910 | const unsigned UseIndex = CB.getDataOperandNo(U); |
| 911 | |
| 912 | // Some intrinsics (for instance ptrmask) do not capture their results, |
| 913 | // but return results thas alias their pointer argument, and thus should |
| 914 | // be handled like GEP or addrspacecast above. |
| 915 | if (isIntrinsicReturningPointerAliasingArgumentWithoutCapturing( |
| 916 | Call: &CB, /*MustPreserveNullness=*/false)) { |
| 917 | for (Use &UU : CB.uses()) |
| 918 | if (Visited.insert(Ptr: &UU).second) |
| 919 | Worklist.push_back(Elt: &UU); |
| 920 | } else if (capturesAnyProvenance(CC: CB.getCaptureInfo(OpNo: UseIndex))) { |
| 921 | if (!CB.onlyReadsMemory()) |
| 922 | // If the callee can save a copy into other memory, then simply |
| 923 | // scanning uses of the call is insufficient. We have no way |
| 924 | // of tracking copies of the pointer through memory to see |
| 925 | // if a reloaded copy is written to, thus we must give up. |
| 926 | return Attribute::None; |
| 927 | // Push users for processing once we finish this one |
| 928 | if (!I->getType()->isVoidTy()) |
| 929 | for (Use &UU : I->uses()) |
| 930 | if (Visited.insert(Ptr: &UU).second) |
| 931 | Worklist.push_back(Elt: &UU); |
| 932 | } |
| 933 | |
| 934 | ModRefInfo ArgMR = CB.getMemoryEffects().getModRef(Loc: IRMemLocation::ArgMem); |
| 935 | if (isNoModRef(MRI: ArgMR)) |
| 936 | continue; |
| 937 | |
| 938 | if (Function *F = CB.getCalledFunction()) |
| 939 | if (CB.isArgOperand(U) && UseIndex < F->arg_size() && |
| 940 | SCCNodes.count(Ptr: F->getArg(i: UseIndex))) |
| 941 | // This is an argument which is part of the speculative SCC. Note |
| 942 | // that only operands corresponding to formal arguments of the callee |
| 943 | // can participate in the speculation. |
| 944 | break; |
| 945 | |
| 946 | // The accessors used on call site here do the right thing for calls and |
| 947 | // invokes with operand bundles. |
| 948 | if (CB.doesNotAccessMemory(OpNo: UseIndex)) { |
| 949 | /* nop */ |
| 950 | } else if (!isModSet(MRI: ArgMR) || CB.onlyReadsMemory(OpNo: UseIndex)) { |
| 951 | IsRead = true; |
| 952 | } else if (!isRefSet(MRI: ArgMR) || |
| 953 | CB.dataOperandHasImpliedAttr(i: UseIndex, Kind: Attribute::WriteOnly)) { |
| 954 | IsWrite = true; |
| 955 | } else { |
| 956 | return Attribute::None; |
| 957 | } |
| 958 | break; |
| 959 | } |
| 960 | |
| 961 | case Instruction::Load: |
| 962 | // A volatile load has side effects beyond what readonly can be relied |
| 963 | // upon. |
| 964 | if (cast<LoadInst>(Val: I)->isVolatile()) |
| 965 | return Attribute::None; |
| 966 | |
| 967 | IsRead = true; |
| 968 | break; |
| 969 | |
| 970 | case Instruction::Store: |
| 971 | if (cast<StoreInst>(Val: I)->getValueOperand() == *U) |
| 972 | // untrackable capture |
| 973 | return Attribute::None; |
| 974 | |
| 975 | // A volatile store has side effects beyond what writeonly can be relied |
| 976 | // upon. |
| 977 | if (cast<StoreInst>(Val: I)->isVolatile()) |
| 978 | return Attribute::None; |
| 979 | |
| 980 | IsWrite = true; |
| 981 | break; |
| 982 | |
| 983 | case Instruction::ICmp: |
| 984 | case Instruction::Ret: |
| 985 | break; |
| 986 | |
| 987 | default: |
| 988 | return Attribute::None; |
| 989 | } |
| 990 | } |
| 991 | |
| 992 | if (IsWrite && IsRead) |
| 993 | return Attribute::None; |
| 994 | else if (IsRead) |
| 995 | return Attribute::ReadOnly; |
| 996 | else if (IsWrite) |
| 997 | return Attribute::WriteOnly; |
| 998 | else |
| 999 | return Attribute::ReadNone; |
| 1000 | } |
| 1001 | |
| 1002 | /// Deduce returned attributes for the SCC. |
| 1003 | static void addArgumentReturnedAttrs(const SCCNodeSet &SCCNodes, |
| 1004 | SmallPtrSet<Function *, 8> &Changed) { |
| 1005 | // Check each function in turn, determining if an argument is always returned. |
| 1006 | for (Function *F : SCCNodes) { |
| 1007 | // We can infer and propagate function attributes only when we know that the |
| 1008 | // definition we'll get at link time is *exactly* the definition we see now. |
| 1009 | // For more details, see GlobalValue::mayBeDerefined. |
| 1010 | if (!F->hasExactDefinition()) |
| 1011 | continue; |
| 1012 | |
| 1013 | if (F->getReturnType()->isVoidTy()) |
| 1014 | continue; |
| 1015 | |
| 1016 | // There is nothing to do if an argument is already marked as 'returned'. |
| 1017 | if (F->getAttributes().hasAttrSomewhere(Kind: Attribute::Returned)) |
| 1018 | continue; |
| 1019 | |
| 1020 | auto FindRetArg = [&]() -> Argument * { |
| 1021 | Argument *RetArg = nullptr; |
| 1022 | for (BasicBlock &BB : *F) |
| 1023 | if (auto *Ret = dyn_cast<ReturnInst>(Val: BB.getTerminator())) { |
| 1024 | // Note that stripPointerCasts should look through functions with |
| 1025 | // returned arguments. |
| 1026 | auto *RetVal = |
| 1027 | dyn_cast<Argument>(Val: Ret->getReturnValue()->stripPointerCasts()); |
| 1028 | if (!RetVal || RetVal->getType() != F->getReturnType()) |
| 1029 | return nullptr; |
| 1030 | |
| 1031 | if (!RetArg) |
| 1032 | RetArg = RetVal; |
| 1033 | else if (RetArg != RetVal) |
| 1034 | return nullptr; |
| 1035 | } |
| 1036 | |
| 1037 | return RetArg; |
| 1038 | }; |
| 1039 | |
| 1040 | if (Argument *RetArg = FindRetArg()) { |
| 1041 | RetArg->addAttr(Kind: Attribute::Returned); |
| 1042 | ++NumReturned; |
| 1043 | Changed.insert(Ptr: F); |
| 1044 | } |
| 1045 | } |
| 1046 | } |
| 1047 | |
| 1048 | /// If a callsite has arguments that are also arguments to the parent function, |
| 1049 | /// try to propagate attributes from the callsite's arguments to the parent's |
| 1050 | /// arguments. This may be important because inlining can cause information loss |
| 1051 | /// when attribute knowledge disappears with the inlined call. |
| 1052 | static bool addArgumentAttrsFromCallsites(Function &F) { |
| 1053 | if (!EnableNonnullArgPropagation) |
| 1054 | return false; |
| 1055 | |
| 1056 | bool Changed = false; |
| 1057 | |
| 1058 | // For an argument attribute to transfer from a callsite to the parent, the |
| 1059 | // call must be guaranteed to execute every time the parent is called. |
| 1060 | // Conservatively, just check for calls in the entry block that are guaranteed |
| 1061 | // to execute. |
| 1062 | // TODO: This could be enhanced by testing if the callsite post-dominates the |
| 1063 | // entry block or by doing simple forward walks or backward walks to the |
| 1064 | // callsite. |
| 1065 | BasicBlock &Entry = F.getEntryBlock(); |
| 1066 | for (Instruction &I : Entry) { |
| 1067 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) { |
| 1068 | if (auto *CalledFunc = CB->getCalledFunction()) { |
| 1069 | for (auto &CSArg : CalledFunc->args()) { |
| 1070 | if (!CSArg.hasNonNullAttr(/* AllowUndefOrPoison */ false)) |
| 1071 | continue; |
| 1072 | |
| 1073 | // If the non-null callsite argument operand is an argument to 'F' |
| 1074 | // (the caller) and the call is guaranteed to execute, then the value |
| 1075 | // must be non-null throughout 'F'. |
| 1076 | auto *FArg = dyn_cast<Argument>(Val: CB->getArgOperand(i: CSArg.getArgNo())); |
| 1077 | if (FArg && !FArg->hasNonNullAttr()) { |
| 1078 | FArg->addAttr(Kind: Attribute::NonNull); |
| 1079 | Changed = true; |
| 1080 | } |
| 1081 | } |
| 1082 | } |
| 1083 | } |
| 1084 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &I)) |
| 1085 | break; |
| 1086 | } |
| 1087 | |
| 1088 | return Changed; |
| 1089 | } |
| 1090 | |
| 1091 | static bool addAccessAttr(Argument *A, Attribute::AttrKind R) { |
| 1092 | assert((R == Attribute::ReadOnly || R == Attribute::ReadNone || |
| 1093 | R == Attribute::WriteOnly) |
| 1094 | && "Must be an access attribute." ); |
| 1095 | assert(A && "Argument must not be null." ); |
| 1096 | |
| 1097 | // If the argument already has the attribute, nothing needs to be done. |
| 1098 | if (A->hasAttribute(Kind: R)) |
| 1099 | return false; |
| 1100 | |
| 1101 | // Otherwise, remove potentially conflicting attribute, add the new one, |
| 1102 | // and update statistics. |
| 1103 | A->removeAttr(Kind: Attribute::WriteOnly); |
| 1104 | A->removeAttr(Kind: Attribute::ReadOnly); |
| 1105 | A->removeAttr(Kind: Attribute::ReadNone); |
| 1106 | // Remove conflicting writable attribute. |
| 1107 | if (R == Attribute::ReadNone || R == Attribute::ReadOnly) |
| 1108 | A->removeAttr(Kind: Attribute::Writable); |
| 1109 | A->addAttr(Kind: R); |
| 1110 | if (R == Attribute::ReadOnly) |
| 1111 | ++NumReadOnlyArg; |
| 1112 | else if (R == Attribute::WriteOnly) |
| 1113 | ++NumWriteOnlyArg; |
| 1114 | else |
| 1115 | ++NumReadNoneArg; |
| 1116 | return true; |
| 1117 | } |
| 1118 | |
| 1119 | static bool inferInitializes(Argument &A, Function &F) { |
| 1120 | auto ArgumentUses = collectArgumentUsesPerBlock(A, F); |
| 1121 | // No write anywhere in the function, bail. |
| 1122 | if (!ArgumentUses.HasAnyWrite) |
| 1123 | return false; |
| 1124 | |
| 1125 | auto &UsesPerBlock = ArgumentUses.UsesPerBlock; |
| 1126 | BasicBlock &EntryBB = F.getEntryBlock(); |
| 1127 | // A map to store the argument ranges initialized by a BasicBlock (including |
| 1128 | // its successors). |
| 1129 | DenseMap<const BasicBlock *, ConstantRangeList> Initialized; |
| 1130 | // Visit the successors of "BB" block and the instructions in BB (post-order) |
| 1131 | // to get the argument ranges initialized by "BB" (including its successors). |
| 1132 | // The result will be cached in "Initialized". |
| 1133 | auto VisitBlock = [&](const BasicBlock *BB) -> ConstantRangeList { |
| 1134 | auto UPB = UsesPerBlock.find(Val: BB); |
| 1135 | ConstantRangeList CRL; |
| 1136 | |
| 1137 | // Start with intersection of successors. |
| 1138 | // If this block has any clobbering use, we're going to clear out the |
| 1139 | // ranges at some point in this block anyway, so don't bother looking at |
| 1140 | // successors. |
| 1141 | if (UPB == UsesPerBlock.end() || !UPB->second.HasUnknownAccess) { |
| 1142 | bool HasAddedSuccessor = false; |
| 1143 | for (auto *Succ : successors(BB)) { |
| 1144 | if (auto SuccI = Initialized.find(Val: Succ); SuccI != Initialized.end()) { |
| 1145 | if (HasAddedSuccessor) { |
| 1146 | CRL = CRL.intersectWith(CRL: SuccI->second); |
| 1147 | } else { |
| 1148 | CRL = SuccI->second; |
| 1149 | HasAddedSuccessor = true; |
| 1150 | } |
| 1151 | } else { |
| 1152 | CRL = ConstantRangeList(); |
| 1153 | break; |
| 1154 | } |
| 1155 | } |
| 1156 | } |
| 1157 | |
| 1158 | if (UPB != UsesPerBlock.end()) { |
| 1159 | // Sort uses in this block by instruction order. |
| 1160 | SmallVector<std::pair<Instruction *, ArgumentAccessInfo>, 2> Insts; |
| 1161 | append_range(C&: Insts, R&: UPB->second.Insts); |
| 1162 | sort(C&: Insts, Comp: [](std::pair<Instruction *, ArgumentAccessInfo> &LHS, |
| 1163 | std::pair<Instruction *, ArgumentAccessInfo> &RHS) { |
| 1164 | return LHS.first->comesBefore(Other: RHS.first); |
| 1165 | }); |
| 1166 | |
| 1167 | // From the end of the block to the beginning of the block, set |
| 1168 | // initializes ranges. |
| 1169 | for (auto &[_, Info] : reverse(C&: Insts)) { |
| 1170 | if (Info.ArgAccessType == ArgumentAccessInfo::AccessType::Unknown || |
| 1171 | Info.ArgAccessType == |
| 1172 | ArgumentAccessInfo::AccessType::WriteWithSideEffect) |
| 1173 | CRL = ConstantRangeList(); |
| 1174 | if (!Info.AccessRanges.empty()) { |
| 1175 | if (Info.ArgAccessType == ArgumentAccessInfo::AccessType::Write || |
| 1176 | Info.ArgAccessType == |
| 1177 | ArgumentAccessInfo::AccessType::WriteWithSideEffect) { |
| 1178 | CRL = CRL.unionWith(CRL: Info.AccessRanges); |
| 1179 | } else { |
| 1180 | assert(Info.ArgAccessType == ArgumentAccessInfo::AccessType::Read); |
| 1181 | for (const auto &ReadRange : Info.AccessRanges) |
| 1182 | CRL.subtract(SubRange: ReadRange); |
| 1183 | } |
| 1184 | } |
| 1185 | } |
| 1186 | } |
| 1187 | return CRL; |
| 1188 | }; |
| 1189 | |
| 1190 | ConstantRangeList EntryCRL; |
| 1191 | // If all write instructions are in the EntryBB, or if the EntryBB has |
| 1192 | // a clobbering use, we only need to look at EntryBB. |
| 1193 | bool OnlyScanEntryBlock = !ArgumentUses.HasWriteOutsideEntryBB; |
| 1194 | if (!OnlyScanEntryBlock) |
| 1195 | if (auto EntryUPB = UsesPerBlock.find(Val: &EntryBB); |
| 1196 | EntryUPB != UsesPerBlock.end()) |
| 1197 | OnlyScanEntryBlock = EntryUPB->second.HasUnknownAccess; |
| 1198 | if (OnlyScanEntryBlock) { |
| 1199 | EntryCRL = VisitBlock(&EntryBB); |
| 1200 | if (EntryCRL.empty()) |
| 1201 | return false; |
| 1202 | } else { |
| 1203 | // Now we have to go through CFG to get the initialized argument ranges |
| 1204 | // across blocks. With dominance and post-dominance, the initialized ranges |
| 1205 | // by a block include both accesses inside this block and accesses in its |
| 1206 | // (transitive) successors. So visit successors before predecessors with a |
| 1207 | // post-order walk of the blocks and memorize the results in "Initialized". |
| 1208 | for (const BasicBlock *BB : post_order(G: &F)) { |
| 1209 | ConstantRangeList CRL = VisitBlock(BB); |
| 1210 | if (!CRL.empty()) |
| 1211 | Initialized[BB] = CRL; |
| 1212 | } |
| 1213 | |
| 1214 | auto EntryCRLI = Initialized.find(Val: &EntryBB); |
| 1215 | if (EntryCRLI == Initialized.end()) |
| 1216 | return false; |
| 1217 | |
| 1218 | EntryCRL = EntryCRLI->second; |
| 1219 | } |
| 1220 | |
| 1221 | assert(!EntryCRL.empty() && |
| 1222 | "should have bailed already if EntryCRL is empty" ); |
| 1223 | |
| 1224 | if (A.hasAttribute(Kind: Attribute::Initializes)) { |
| 1225 | ConstantRangeList PreviousCRL = |
| 1226 | A.getAttribute(Kind: Attribute::Initializes).getValueAsConstantRangeList(); |
| 1227 | if (PreviousCRL == EntryCRL) |
| 1228 | return false; |
| 1229 | EntryCRL = EntryCRL.unionWith(CRL: PreviousCRL); |
| 1230 | } |
| 1231 | |
| 1232 | A.addAttr(Attr: Attribute::get(Context&: A.getContext(), Kind: Attribute::Initializes, |
| 1233 | Val: EntryCRL.rangesRef())); |
| 1234 | |
| 1235 | return true; |
| 1236 | } |
| 1237 | |
| 1238 | /// Deduce nocapture attributes for the SCC. |
| 1239 | static void addArgumentAttrs(const SCCNodeSet &SCCNodes, |
| 1240 | SmallPtrSet<Function *, 8> &Changed, |
| 1241 | bool SkipInitializes) { |
| 1242 | ArgumentGraph AG; |
| 1243 | |
| 1244 | auto DetermineAccessAttrsForSingleton = [](Argument *A) { |
| 1245 | SmallPtrSet<Argument *, 8> Self; |
| 1246 | Self.insert(Ptr: A); |
| 1247 | Attribute::AttrKind R = determinePointerAccessAttrs(A, SCCNodes: Self); |
| 1248 | if (R != Attribute::None) |
| 1249 | return addAccessAttr(A, R); |
| 1250 | return false; |
| 1251 | }; |
| 1252 | |
| 1253 | // Check each function in turn, determining which pointer arguments are not |
| 1254 | // captured. |
| 1255 | for (Function *F : SCCNodes) { |
| 1256 | // We can infer and propagate function attributes only when we know that the |
| 1257 | // definition we'll get at link time is *exactly* the definition we see now. |
| 1258 | // For more details, see GlobalValue::mayBeDerefined. |
| 1259 | if (!F->hasExactDefinition()) |
| 1260 | continue; |
| 1261 | |
| 1262 | if (addArgumentAttrsFromCallsites(F&: *F)) |
| 1263 | Changed.insert(Ptr: F); |
| 1264 | |
| 1265 | // Functions that are readonly (or readnone) and nounwind and don't return |
| 1266 | // a value can't capture arguments. Don't analyze them. |
| 1267 | if (F->onlyReadsMemory() && F->doesNotThrow() && F->willReturn() && |
| 1268 | F->getReturnType()->isVoidTy()) { |
| 1269 | for (Argument &A : F->args()) { |
| 1270 | if (A.getType()->isPointerTy() && !A.hasNoCaptureAttr()) { |
| 1271 | A.addAttr(Attr: Attribute::getWithCaptureInfo(Context&: A.getContext(), |
| 1272 | CI: CaptureInfo::none())); |
| 1273 | ++NumCapturesNone; |
| 1274 | Changed.insert(Ptr: F); |
| 1275 | } |
| 1276 | } |
| 1277 | continue; |
| 1278 | } |
| 1279 | |
| 1280 | for (Argument &A : F->args()) { |
| 1281 | if (!A.getType()->isPointerTy()) |
| 1282 | continue; |
| 1283 | bool HasNonLocalUses = false; |
| 1284 | CaptureInfo OrigCI = A.getAttributes().getCaptureInfo(); |
| 1285 | if (!capturesNothing(CC: OrigCI)) { |
| 1286 | ArgumentUsesTracker Tracker(SCCNodes); |
| 1287 | PointerMayBeCaptured(V: &A, Tracker: &Tracker); |
| 1288 | CaptureInfo NewCI = Tracker.CI & OrigCI; |
| 1289 | if (NewCI != OrigCI) { |
| 1290 | if (Tracker.Uses.empty()) { |
| 1291 | // If the information is complete, add the attribute now. |
| 1292 | A.addAttr(Attr: Attribute::getWithCaptureInfo(Context&: A.getContext(), CI: NewCI)); |
| 1293 | addCapturesStat(CI: NewCI); |
| 1294 | Changed.insert(Ptr: F); |
| 1295 | } else { |
| 1296 | // If it's not trivially captured and not trivially not captured, |
| 1297 | // then it must be calling into another function in our SCC. Save |
| 1298 | // its particulars for Argument-SCC analysis later. |
| 1299 | ArgumentGraphNode *Node = AG[&A]; |
| 1300 | Node->CC = CaptureComponents(NewCI); |
| 1301 | for (Argument *Use : Tracker.Uses) { |
| 1302 | Node->Uses.push_back(Elt: AG[Use]); |
| 1303 | if (Use != &A) |
| 1304 | HasNonLocalUses = true; |
| 1305 | } |
| 1306 | } |
| 1307 | } |
| 1308 | // Otherwise, it's captured. Don't bother doing SCC analysis on it. |
| 1309 | } |
| 1310 | if (!HasNonLocalUses && !A.onlyReadsMemory()) { |
| 1311 | // Can we determine that it's readonly/readnone/writeonly without doing |
| 1312 | // an SCC? Note that we don't allow any calls at all here, or else our |
| 1313 | // result will be dependent on the iteration order through the |
| 1314 | // functions in the SCC. |
| 1315 | if (DetermineAccessAttrsForSingleton(&A)) |
| 1316 | Changed.insert(Ptr: F); |
| 1317 | } |
| 1318 | if (!SkipInitializes && !A.onlyReadsMemory()) { |
| 1319 | if (inferInitializes(A, F&: *F)) |
| 1320 | Changed.insert(Ptr: F); |
| 1321 | } |
| 1322 | } |
| 1323 | } |
| 1324 | |
| 1325 | // The graph we've collected is partial because we stopped scanning for |
| 1326 | // argument uses once we solved the argument trivially. These partial nodes |
| 1327 | // show up as ArgumentGraphNode objects with an empty Uses list, and for |
| 1328 | // these nodes the final decision about whether they capture has already been |
| 1329 | // made. If the definition doesn't have a 'nocapture' attribute by now, it |
| 1330 | // captures. |
| 1331 | |
| 1332 | for (scc_iterator<ArgumentGraph *> I = scc_begin(G: &AG); !I.isAtEnd(); ++I) { |
| 1333 | const std::vector<ArgumentGraphNode *> &ArgumentSCC = *I; |
| 1334 | if (ArgumentSCC.size() == 1) { |
| 1335 | if (!ArgumentSCC[0]->Definition) |
| 1336 | continue; // synthetic root node |
| 1337 | |
| 1338 | // eg. "void f(int* x) { if (...) f(x); }" |
| 1339 | if (ArgumentSCC[0]->Uses.size() == 1 && |
| 1340 | ArgumentSCC[0]->Uses[0] == ArgumentSCC[0]) { |
| 1341 | Argument *A = ArgumentSCC[0]->Definition; |
| 1342 | CaptureInfo OrigCI = A->getAttributes().getCaptureInfo(); |
| 1343 | CaptureInfo NewCI = CaptureInfo(ArgumentSCC[0]->CC) & OrigCI; |
| 1344 | if (NewCI != OrigCI) { |
| 1345 | A->addAttr(Attr: Attribute::getWithCaptureInfo(Context&: A->getContext(), CI: NewCI)); |
| 1346 | addCapturesStat(CI: NewCI); |
| 1347 | Changed.insert(Ptr: A->getParent()); |
| 1348 | } |
| 1349 | |
| 1350 | // Infer the access attributes given the new captures one |
| 1351 | if (DetermineAccessAttrsForSingleton(A)) |
| 1352 | Changed.insert(Ptr: A->getParent()); |
| 1353 | } |
| 1354 | continue; |
| 1355 | } |
| 1356 | |
| 1357 | SmallPtrSet<Argument *, 8> ArgumentSCCNodes; |
| 1358 | // Fill ArgumentSCCNodes with the elements of the ArgumentSCC. Used for |
| 1359 | // quickly looking up whether a given Argument is in this ArgumentSCC. |
| 1360 | for (ArgumentGraphNode *I : ArgumentSCC) { |
| 1361 | ArgumentSCCNodes.insert(Ptr: I->Definition); |
| 1362 | } |
| 1363 | |
| 1364 | // At the SCC level, only track merged CaptureComponents. We're not |
| 1365 | // currently prepared to handle propagation of return-only captures across |
| 1366 | // the SCC. |
| 1367 | CaptureComponents CC = CaptureComponents::None; |
| 1368 | for (ArgumentGraphNode *N : ArgumentSCC) { |
| 1369 | for (ArgumentGraphNode *Use : N->Uses) { |
| 1370 | Argument *A = Use->Definition; |
| 1371 | if (ArgumentSCCNodes.count(Ptr: A)) |
| 1372 | CC |= Use->CC; |
| 1373 | else |
| 1374 | CC |= CaptureComponents(A->getAttributes().getCaptureInfo()); |
| 1375 | break; |
| 1376 | } |
| 1377 | if (capturesAll(CC)) |
| 1378 | break; |
| 1379 | } |
| 1380 | |
| 1381 | if (!capturesAll(CC)) { |
| 1382 | for (ArgumentGraphNode *N : ArgumentSCC) { |
| 1383 | Argument *A = N->Definition; |
| 1384 | CaptureInfo OrigCI = A->getAttributes().getCaptureInfo(); |
| 1385 | CaptureInfo NewCI = CaptureInfo(N->CC | CC) & OrigCI; |
| 1386 | if (NewCI != OrigCI) { |
| 1387 | A->addAttr(Attr: Attribute::getWithCaptureInfo(Context&: A->getContext(), CI: NewCI)); |
| 1388 | addCapturesStat(CI: NewCI); |
| 1389 | Changed.insert(Ptr: A->getParent()); |
| 1390 | } |
| 1391 | } |
| 1392 | } |
| 1393 | |
| 1394 | if (capturesAnyProvenance(CC)) { |
| 1395 | // As the pointer provenance may be captured, determine the pointer |
| 1396 | // attributes looking at each argument individually. |
| 1397 | for (ArgumentGraphNode *N : ArgumentSCC) { |
| 1398 | if (DetermineAccessAttrsForSingleton(N->Definition)) |
| 1399 | Changed.insert(Ptr: N->Definition->getParent()); |
| 1400 | } |
| 1401 | continue; |
| 1402 | } |
| 1403 | |
| 1404 | // We also want to compute readonly/readnone/writeonly. With a small number |
| 1405 | // of false negatives, we can assume that any pointer which is captured |
| 1406 | // isn't going to be provably readonly or readnone, since by definition |
| 1407 | // we can't analyze all uses of a captured pointer. |
| 1408 | // |
| 1409 | // The false negatives happen when the pointer is captured by a function |
| 1410 | // that promises readonly/readnone behaviour on the pointer, then the |
| 1411 | // pointer's lifetime ends before anything that writes to arbitrary memory. |
| 1412 | // Also, a readonly/readnone pointer may be returned, but returning a |
| 1413 | // pointer is capturing it. |
| 1414 | |
| 1415 | auto meetAccessAttr = [](Attribute::AttrKind A, Attribute::AttrKind B) { |
| 1416 | if (A == B) |
| 1417 | return A; |
| 1418 | if (A == Attribute::ReadNone) |
| 1419 | return B; |
| 1420 | if (B == Attribute::ReadNone) |
| 1421 | return A; |
| 1422 | return Attribute::None; |
| 1423 | }; |
| 1424 | |
| 1425 | Attribute::AttrKind AccessAttr = Attribute::ReadNone; |
| 1426 | for (ArgumentGraphNode *N : ArgumentSCC) { |
| 1427 | Argument *A = N->Definition; |
| 1428 | Attribute::AttrKind K = determinePointerAccessAttrs(A, SCCNodes: ArgumentSCCNodes); |
| 1429 | AccessAttr = meetAccessAttr(AccessAttr, K); |
| 1430 | if (AccessAttr == Attribute::None) |
| 1431 | break; |
| 1432 | } |
| 1433 | |
| 1434 | if (AccessAttr != Attribute::None) { |
| 1435 | for (ArgumentGraphNode *N : ArgumentSCC) { |
| 1436 | Argument *A = N->Definition; |
| 1437 | if (addAccessAttr(A, R: AccessAttr)) |
| 1438 | Changed.insert(Ptr: A->getParent()); |
| 1439 | } |
| 1440 | } |
| 1441 | } |
| 1442 | } |
| 1443 | |
| 1444 | /// Tests whether a function is "malloc-like". |
| 1445 | /// |
| 1446 | /// A function is "malloc-like" if it returns either null or a pointer that |
| 1447 | /// doesn't alias any other pointer visible to the caller. |
| 1448 | static bool isFunctionMallocLike(Function *F, const SCCNodeSet &SCCNodes) { |
| 1449 | SmallSetVector<Value *, 8> FlowsToReturn; |
| 1450 | for (BasicBlock &BB : *F) |
| 1451 | if (ReturnInst *Ret = dyn_cast<ReturnInst>(Val: BB.getTerminator())) |
| 1452 | FlowsToReturn.insert(X: Ret->getReturnValue()); |
| 1453 | |
| 1454 | for (unsigned i = 0; i != FlowsToReturn.size(); ++i) { |
| 1455 | Value *RetVal = FlowsToReturn[i]; |
| 1456 | |
| 1457 | if (Constant *C = dyn_cast<Constant>(Val: RetVal)) { |
| 1458 | if (!C->isNullValue() && !isa<UndefValue>(Val: C)) |
| 1459 | return false; |
| 1460 | |
| 1461 | continue; |
| 1462 | } |
| 1463 | |
| 1464 | if (isa<Argument>(Val: RetVal)) |
| 1465 | return false; |
| 1466 | |
| 1467 | if (Instruction *RVI = dyn_cast<Instruction>(Val: RetVal)) |
| 1468 | switch (RVI->getOpcode()) { |
| 1469 | // Extend the analysis by looking upwards. |
| 1470 | case Instruction::BitCast: |
| 1471 | case Instruction::GetElementPtr: |
| 1472 | case Instruction::AddrSpaceCast: |
| 1473 | FlowsToReturn.insert(X: RVI->getOperand(i: 0)); |
| 1474 | continue; |
| 1475 | case Instruction::Select: { |
| 1476 | SelectInst *SI = cast<SelectInst>(Val: RVI); |
| 1477 | FlowsToReturn.insert(X: SI->getTrueValue()); |
| 1478 | FlowsToReturn.insert(X: SI->getFalseValue()); |
| 1479 | continue; |
| 1480 | } |
| 1481 | case Instruction::PHI: { |
| 1482 | PHINode *PN = cast<PHINode>(Val: RVI); |
| 1483 | FlowsToReturn.insert_range(R: PN->incoming_values()); |
| 1484 | continue; |
| 1485 | } |
| 1486 | |
| 1487 | // Check whether the pointer came from an allocation. |
| 1488 | case Instruction::Alloca: |
| 1489 | break; |
| 1490 | case Instruction::Call: |
| 1491 | case Instruction::Invoke: { |
| 1492 | CallBase &CB = cast<CallBase>(Val&: *RVI); |
| 1493 | if (CB.hasRetAttr(Kind: Attribute::NoAlias)) |
| 1494 | break; |
| 1495 | if (CB.getCalledFunction() && SCCNodes.count(key: CB.getCalledFunction())) |
| 1496 | break; |
| 1497 | [[fallthrough]]; |
| 1498 | } |
| 1499 | default: |
| 1500 | return false; // Did not come from an allocation. |
| 1501 | } |
| 1502 | |
| 1503 | if (PointerMayBeCaptured(V: RetVal, /*ReturnCaptures=*/false)) |
| 1504 | return false; |
| 1505 | } |
| 1506 | |
| 1507 | return true; |
| 1508 | } |
| 1509 | |
| 1510 | /// Deduce noalias attributes for the SCC. |
| 1511 | static void addNoAliasAttrs(const SCCNodeSet &SCCNodes, |
| 1512 | SmallPtrSet<Function *, 8> &Changed) { |
| 1513 | // Check each function in turn, determining which functions return noalias |
| 1514 | // pointers. |
| 1515 | for (Function *F : SCCNodes) { |
| 1516 | // Already noalias. |
| 1517 | if (F->returnDoesNotAlias()) |
| 1518 | continue; |
| 1519 | |
| 1520 | // We can infer and propagate function attributes only when we know that the |
| 1521 | // definition we'll get at link time is *exactly* the definition we see now. |
| 1522 | // For more details, see GlobalValue::mayBeDerefined. |
| 1523 | if (!F->hasExactDefinition()) |
| 1524 | return; |
| 1525 | |
| 1526 | // We annotate noalias return values, which are only applicable to |
| 1527 | // pointer types. |
| 1528 | if (!F->getReturnType()->isPointerTy()) |
| 1529 | continue; |
| 1530 | |
| 1531 | if (!isFunctionMallocLike(F, SCCNodes)) |
| 1532 | return; |
| 1533 | } |
| 1534 | |
| 1535 | for (Function *F : SCCNodes) { |
| 1536 | if (F->returnDoesNotAlias() || |
| 1537 | !F->getReturnType()->isPointerTy()) |
| 1538 | continue; |
| 1539 | |
| 1540 | F->setReturnDoesNotAlias(); |
| 1541 | ++NumNoAlias; |
| 1542 | Changed.insert(Ptr: F); |
| 1543 | } |
| 1544 | } |
| 1545 | |
| 1546 | /// Tests whether this function is known to not return null. |
| 1547 | /// |
| 1548 | /// Requires that the function returns a pointer. |
| 1549 | /// |
| 1550 | /// Returns true if it believes the function will not return a null, and sets |
| 1551 | /// \p Speculative based on whether the returned conclusion is a speculative |
| 1552 | /// conclusion due to SCC calls. |
| 1553 | static bool isReturnNonNull(Function *F, const SCCNodeSet &SCCNodes, |
| 1554 | bool &Speculative) { |
| 1555 | assert(F->getReturnType()->isPointerTy() && |
| 1556 | "nonnull only meaningful on pointer types" ); |
| 1557 | Speculative = false; |
| 1558 | |
| 1559 | SmallSetVector<Value *, 8> FlowsToReturn; |
| 1560 | for (BasicBlock &BB : *F) |
| 1561 | if (auto *Ret = dyn_cast<ReturnInst>(Val: BB.getTerminator())) |
| 1562 | FlowsToReturn.insert(X: Ret->getReturnValue()); |
| 1563 | |
| 1564 | auto &DL = F->getDataLayout(); |
| 1565 | |
| 1566 | for (unsigned i = 0; i != FlowsToReturn.size(); ++i) { |
| 1567 | Value *RetVal = FlowsToReturn[i]; |
| 1568 | |
| 1569 | // If this value is locally known to be non-null, we're good |
| 1570 | if (isKnownNonZero(V: RetVal, Q: DL)) |
| 1571 | continue; |
| 1572 | |
| 1573 | // Otherwise, we need to look upwards since we can't make any local |
| 1574 | // conclusions. |
| 1575 | Instruction *RVI = dyn_cast<Instruction>(Val: RetVal); |
| 1576 | if (!RVI) |
| 1577 | return false; |
| 1578 | switch (RVI->getOpcode()) { |
| 1579 | // Extend the analysis by looking upwards. |
| 1580 | case Instruction::BitCast: |
| 1581 | case Instruction::AddrSpaceCast: |
| 1582 | FlowsToReturn.insert(X: RVI->getOperand(i: 0)); |
| 1583 | continue; |
| 1584 | case Instruction::GetElementPtr: |
| 1585 | if (cast<GEPOperator>(Val: RVI)->isInBounds()) { |
| 1586 | FlowsToReturn.insert(X: RVI->getOperand(i: 0)); |
| 1587 | continue; |
| 1588 | } |
| 1589 | return false; |
| 1590 | case Instruction::Select: { |
| 1591 | SelectInst *SI = cast<SelectInst>(Val: RVI); |
| 1592 | FlowsToReturn.insert(X: SI->getTrueValue()); |
| 1593 | FlowsToReturn.insert(X: SI->getFalseValue()); |
| 1594 | continue; |
| 1595 | } |
| 1596 | case Instruction::PHI: { |
| 1597 | PHINode *PN = cast<PHINode>(Val: RVI); |
| 1598 | for (int i = 0, e = PN->getNumIncomingValues(); i != e; ++i) |
| 1599 | FlowsToReturn.insert(X: PN->getIncomingValue(i)); |
| 1600 | continue; |
| 1601 | } |
| 1602 | case Instruction::Call: |
| 1603 | case Instruction::Invoke: { |
| 1604 | CallBase &CB = cast<CallBase>(Val&: *RVI); |
| 1605 | Function *Callee = CB.getCalledFunction(); |
| 1606 | // A call to a node within the SCC is assumed to return null until |
| 1607 | // proven otherwise |
| 1608 | if (Callee && SCCNodes.count(key: Callee)) { |
| 1609 | Speculative = true; |
| 1610 | continue; |
| 1611 | } |
| 1612 | return false; |
| 1613 | } |
| 1614 | default: |
| 1615 | return false; // Unknown source, may be null |
| 1616 | }; |
| 1617 | llvm_unreachable("should have either continued or returned" ); |
| 1618 | } |
| 1619 | |
| 1620 | return true; |
| 1621 | } |
| 1622 | |
| 1623 | /// Deduce nonnull attributes for the SCC. |
| 1624 | static void addNonNullAttrs(const SCCNodeSet &SCCNodes, |
| 1625 | SmallPtrSet<Function *, 8> &Changed) { |
| 1626 | // Speculative that all functions in the SCC return only nonnull |
| 1627 | // pointers. We may refute this as we analyze functions. |
| 1628 | bool SCCReturnsNonNull = true; |
| 1629 | |
| 1630 | // Check each function in turn, determining which functions return nonnull |
| 1631 | // pointers. |
| 1632 | for (Function *F : SCCNodes) { |
| 1633 | // Already nonnull. |
| 1634 | if (F->getAttributes().hasRetAttr(Kind: Attribute::NonNull)) |
| 1635 | continue; |
| 1636 | |
| 1637 | // We can infer and propagate function attributes only when we know that the |
| 1638 | // definition we'll get at link time is *exactly* the definition we see now. |
| 1639 | // For more details, see GlobalValue::mayBeDerefined. |
| 1640 | if (!F->hasExactDefinition()) |
| 1641 | return; |
| 1642 | |
| 1643 | // We annotate nonnull return values, which are only applicable to |
| 1644 | // pointer types. |
| 1645 | if (!F->getReturnType()->isPointerTy()) |
| 1646 | continue; |
| 1647 | |
| 1648 | bool Speculative = false; |
| 1649 | if (isReturnNonNull(F, SCCNodes, Speculative)) { |
| 1650 | if (!Speculative) { |
| 1651 | // Mark the function eagerly since we may discover a function |
| 1652 | // which prevents us from speculating about the entire SCC |
| 1653 | LLVM_DEBUG(dbgs() << "Eagerly marking " << F->getName() |
| 1654 | << " as nonnull\n" ); |
| 1655 | F->addRetAttr(Kind: Attribute::NonNull); |
| 1656 | ++NumNonNullReturn; |
| 1657 | Changed.insert(Ptr: F); |
| 1658 | } |
| 1659 | continue; |
| 1660 | } |
| 1661 | // At least one function returns something which could be null, can't |
| 1662 | // speculate any more. |
| 1663 | SCCReturnsNonNull = false; |
| 1664 | } |
| 1665 | |
| 1666 | if (SCCReturnsNonNull) { |
| 1667 | for (Function *F : SCCNodes) { |
| 1668 | if (F->getAttributes().hasRetAttr(Kind: Attribute::NonNull) || |
| 1669 | !F->getReturnType()->isPointerTy()) |
| 1670 | continue; |
| 1671 | |
| 1672 | LLVM_DEBUG(dbgs() << "SCC marking " << F->getName() << " as nonnull\n" ); |
| 1673 | F->addRetAttr(Kind: Attribute::NonNull); |
| 1674 | ++NumNonNullReturn; |
| 1675 | Changed.insert(Ptr: F); |
| 1676 | } |
| 1677 | } |
| 1678 | } |
| 1679 | |
| 1680 | /// Deduce noundef attributes for the SCC. |
| 1681 | static void addNoUndefAttrs(const SCCNodeSet &SCCNodes, |
| 1682 | SmallPtrSet<Function *, 8> &Changed) { |
| 1683 | // Check each function in turn, determining which functions return noundef |
| 1684 | // values. |
| 1685 | for (Function *F : SCCNodes) { |
| 1686 | // Already noundef. |
| 1687 | AttributeList Attrs = F->getAttributes(); |
| 1688 | if (Attrs.hasRetAttr(Kind: Attribute::NoUndef)) |
| 1689 | continue; |
| 1690 | |
| 1691 | // We can infer and propagate function attributes only when we know that the |
| 1692 | // definition we'll get at link time is *exactly* the definition we see now. |
| 1693 | // For more details, see GlobalValue::mayBeDerefined. |
| 1694 | if (!F->hasExactDefinition()) |
| 1695 | return; |
| 1696 | |
| 1697 | // MemorySanitizer assumes that the definition and declaration of a |
| 1698 | // function will be consistent. A function with sanitize_memory attribute |
| 1699 | // should be skipped from inference. |
| 1700 | if (F->hasFnAttribute(Kind: Attribute::SanitizeMemory)) |
| 1701 | continue; |
| 1702 | |
| 1703 | if (F->getReturnType()->isVoidTy()) |
| 1704 | continue; |
| 1705 | |
| 1706 | const DataLayout &DL = F->getDataLayout(); |
| 1707 | if (all_of(Range&: *F, P: [&](BasicBlock &BB) { |
| 1708 | if (auto *Ret = dyn_cast<ReturnInst>(Val: BB.getTerminator())) { |
| 1709 | // TODO: perform context-sensitive analysis? |
| 1710 | Value *RetVal = Ret->getReturnValue(); |
| 1711 | if (!isGuaranteedNotToBeUndefOrPoison(V: RetVal)) |
| 1712 | return false; |
| 1713 | |
| 1714 | // We know the original return value is not poison now, but it |
| 1715 | // could still be converted to poison by another return attribute. |
| 1716 | // Try to explicitly re-prove the relevant attributes. |
| 1717 | if (Attrs.hasRetAttr(Kind: Attribute::NonNull) && |
| 1718 | !isKnownNonZero(V: RetVal, Q: DL)) |
| 1719 | return false; |
| 1720 | |
| 1721 | if (MaybeAlign Align = Attrs.getRetAlignment()) |
| 1722 | if (RetVal->getPointerAlignment(DL) < *Align) |
| 1723 | return false; |
| 1724 | |
| 1725 | Attribute Attr = Attrs.getRetAttr(Kind: Attribute::Range); |
| 1726 | if (Attr.isValid() && |
| 1727 | !Attr.getRange().contains( |
| 1728 | CR: computeConstantRange(V: RetVal, /*ForSigned=*/false))) |
| 1729 | return false; |
| 1730 | } |
| 1731 | return true; |
| 1732 | })) { |
| 1733 | F->addRetAttr(Kind: Attribute::NoUndef); |
| 1734 | ++NumNoUndefReturn; |
| 1735 | Changed.insert(Ptr: F); |
| 1736 | } |
| 1737 | } |
| 1738 | } |
| 1739 | |
| 1740 | namespace { |
| 1741 | |
| 1742 | /// Collects a set of attribute inference requests and performs them all in one |
| 1743 | /// go on a single SCC Node. Inference involves scanning function bodies |
| 1744 | /// looking for instructions that violate attribute assumptions. |
| 1745 | /// As soon as all the bodies are fine we are free to set the attribute. |
| 1746 | /// Customization of inference for individual attributes is performed by |
| 1747 | /// providing a handful of predicates for each attribute. |
| 1748 | class AttributeInferer { |
| 1749 | public: |
| 1750 | /// Describes a request for inference of a single attribute. |
| 1751 | struct InferenceDescriptor { |
| 1752 | |
| 1753 | /// Returns true if this function does not have to be handled. |
| 1754 | /// General intent for this predicate is to provide an optimization |
| 1755 | /// for functions that do not need this attribute inference at all |
| 1756 | /// (say, for functions that already have the attribute). |
| 1757 | std::function<bool(const Function &)> SkipFunction; |
| 1758 | |
| 1759 | /// Returns true if this instruction violates attribute assumptions. |
| 1760 | std::function<bool(Instruction &)> InstrBreaksAttribute; |
| 1761 | |
| 1762 | /// Sets the inferred attribute for this function. |
| 1763 | std::function<void(Function &)> SetAttribute; |
| 1764 | |
| 1765 | /// Attribute we derive. |
| 1766 | Attribute::AttrKind AKind; |
| 1767 | |
| 1768 | /// If true, only "exact" definitions can be used to infer this attribute. |
| 1769 | /// See GlobalValue::isDefinitionExact. |
| 1770 | bool RequiresExactDefinition; |
| 1771 | |
| 1772 | InferenceDescriptor(Attribute::AttrKind AK, |
| 1773 | std::function<bool(const Function &)> SkipFunc, |
| 1774 | std::function<bool(Instruction &)> InstrScan, |
| 1775 | std::function<void(Function &)> SetAttr, |
| 1776 | bool ReqExactDef) |
| 1777 | : SkipFunction(SkipFunc), InstrBreaksAttribute(InstrScan), |
| 1778 | SetAttribute(SetAttr), AKind(AK), |
| 1779 | RequiresExactDefinition(ReqExactDef) {} |
| 1780 | }; |
| 1781 | |
| 1782 | private: |
| 1783 | SmallVector<InferenceDescriptor, 4> InferenceDescriptors; |
| 1784 | |
| 1785 | public: |
| 1786 | void registerAttrInference(InferenceDescriptor AttrInference) { |
| 1787 | InferenceDescriptors.push_back(Elt: AttrInference); |
| 1788 | } |
| 1789 | |
| 1790 | void run(const SCCNodeSet &SCCNodes, SmallPtrSet<Function *, 8> &Changed); |
| 1791 | }; |
| 1792 | |
| 1793 | /// Perform all the requested attribute inference actions according to the |
| 1794 | /// attribute predicates stored before. |
| 1795 | void AttributeInferer::run(const SCCNodeSet &SCCNodes, |
| 1796 | SmallPtrSet<Function *, 8> &Changed) { |
| 1797 | SmallVector<InferenceDescriptor, 4> InferInSCC = InferenceDescriptors; |
| 1798 | // Go through all the functions in SCC and check corresponding attribute |
| 1799 | // assumptions for each of them. Attributes that are invalid for this SCC |
| 1800 | // will be removed from InferInSCC. |
| 1801 | for (Function *F : SCCNodes) { |
| 1802 | |
| 1803 | // No attributes whose assumptions are still valid - done. |
| 1804 | if (InferInSCC.empty()) |
| 1805 | return; |
| 1806 | |
| 1807 | // Check if our attributes ever need scanning/can be scanned. |
| 1808 | llvm::erase_if(C&: InferInSCC, P: [F](const InferenceDescriptor &ID) { |
| 1809 | if (ID.SkipFunction(*F)) |
| 1810 | return false; |
| 1811 | |
| 1812 | // Remove from further inference (invalidate) when visiting a function |
| 1813 | // that has no instructions to scan/has an unsuitable definition. |
| 1814 | return F->isDeclaration() || |
| 1815 | (ID.RequiresExactDefinition && !F->hasExactDefinition()); |
| 1816 | }); |
| 1817 | |
| 1818 | // For each attribute still in InferInSCC that doesn't explicitly skip F, |
| 1819 | // set up the F instructions scan to verify assumptions of the attribute. |
| 1820 | SmallVector<InferenceDescriptor, 4> InferInThisFunc; |
| 1821 | llvm::copy_if( |
| 1822 | Range&: InferInSCC, Out: std::back_inserter(x&: InferInThisFunc), |
| 1823 | P: [F](const InferenceDescriptor &ID) { return !ID.SkipFunction(*F); }); |
| 1824 | |
| 1825 | if (InferInThisFunc.empty()) |
| 1826 | continue; |
| 1827 | |
| 1828 | // Start instruction scan. |
| 1829 | for (Instruction &I : instructions(F&: *F)) { |
| 1830 | llvm::erase_if(C&: InferInThisFunc, P: [&](const InferenceDescriptor &ID) { |
| 1831 | if (!ID.InstrBreaksAttribute(I)) |
| 1832 | return false; |
| 1833 | // Remove attribute from further inference on any other functions |
| 1834 | // because attribute assumptions have just been violated. |
| 1835 | llvm::erase_if(C&: InferInSCC, P: [&ID](const InferenceDescriptor &D) { |
| 1836 | return D.AKind == ID.AKind; |
| 1837 | }); |
| 1838 | // Remove attribute from the rest of current instruction scan. |
| 1839 | return true; |
| 1840 | }); |
| 1841 | |
| 1842 | if (InferInThisFunc.empty()) |
| 1843 | break; |
| 1844 | } |
| 1845 | } |
| 1846 | |
| 1847 | if (InferInSCC.empty()) |
| 1848 | return; |
| 1849 | |
| 1850 | for (Function *F : SCCNodes) |
| 1851 | // At this point InferInSCC contains only functions that were either: |
| 1852 | // - explicitly skipped from scan/inference, or |
| 1853 | // - verified to have no instructions that break attribute assumptions. |
| 1854 | // Hence we just go and force the attribute for all non-skipped functions. |
| 1855 | for (auto &ID : InferInSCC) { |
| 1856 | if (ID.SkipFunction(*F)) |
| 1857 | continue; |
| 1858 | Changed.insert(Ptr: F); |
| 1859 | ID.SetAttribute(*F); |
| 1860 | } |
| 1861 | } |
| 1862 | |
| 1863 | struct SCCNodesResult { |
| 1864 | SCCNodeSet SCCNodes; |
| 1865 | }; |
| 1866 | |
| 1867 | } // end anonymous namespace |
| 1868 | |
| 1869 | /// Helper for non-Convergent inference predicate InstrBreaksAttribute. |
| 1870 | static bool InstrBreaksNonConvergent(Instruction &I, |
| 1871 | const SCCNodeSet &SCCNodes) { |
| 1872 | const CallBase *CB = dyn_cast<CallBase>(Val: &I); |
| 1873 | // Breaks non-convergent assumption if CS is a convergent call to a function |
| 1874 | // not in the SCC. |
| 1875 | return CB && CB->isConvergent() && |
| 1876 | !SCCNodes.contains(key: CB->getCalledFunction()); |
| 1877 | } |
| 1878 | |
| 1879 | /// Helper for NoUnwind inference predicate InstrBreaksAttribute. |
| 1880 | static bool InstrBreaksNonThrowing(Instruction &I, const SCCNodeSet &SCCNodes) { |
| 1881 | if (!I.mayThrow(/* IncludePhaseOneUnwind */ true)) |
| 1882 | return false; |
| 1883 | if (const auto *CI = dyn_cast<CallInst>(Val: &I)) { |
| 1884 | if (Function *Callee = CI->getCalledFunction()) { |
| 1885 | // I is a may-throw call to a function inside our SCC. This doesn't |
| 1886 | // invalidate our current working assumption that the SCC is no-throw; we |
| 1887 | // just have to scan that other function. |
| 1888 | if (SCCNodes.contains(key: Callee)) |
| 1889 | return false; |
| 1890 | } |
| 1891 | } |
| 1892 | return true; |
| 1893 | } |
| 1894 | |
| 1895 | /// Helper for NoFree inference predicate InstrBreaksAttribute. |
| 1896 | static bool InstrBreaksNoFree(Instruction &I, const SCCNodeSet &SCCNodes) { |
| 1897 | CallBase *CB = dyn_cast<CallBase>(Val: &I); |
| 1898 | if (!CB) |
| 1899 | return false; |
| 1900 | |
| 1901 | if (CB->hasFnAttr(Kind: Attribute::NoFree)) |
| 1902 | return false; |
| 1903 | |
| 1904 | // Speculatively assume in SCC. |
| 1905 | if (Function *Callee = CB->getCalledFunction()) |
| 1906 | if (SCCNodes.contains(key: Callee)) |
| 1907 | return false; |
| 1908 | |
| 1909 | return true; |
| 1910 | } |
| 1911 | |
| 1912 | // Return true if this is an atomic which has an ordering stronger than |
| 1913 | // unordered. Note that this is different than the predicate we use in |
| 1914 | // Attributor. Here we chose to be conservative and consider monotonic |
| 1915 | // operations potentially synchronizing. We generally don't do much with |
| 1916 | // monotonic operations, so this is simply risk reduction. |
| 1917 | static bool isOrderedAtomic(Instruction *I) { |
| 1918 | if (!I->isAtomic()) |
| 1919 | return false; |
| 1920 | |
| 1921 | if (auto *FI = dyn_cast<FenceInst>(Val: I)) |
| 1922 | // All legal orderings for fence are stronger than monotonic. |
| 1923 | return FI->getSyncScopeID() != SyncScope::SingleThread; |
| 1924 | else if (isa<AtomicCmpXchgInst>(Val: I) || isa<AtomicRMWInst>(Val: I)) |
| 1925 | return true; |
| 1926 | else if (auto *SI = dyn_cast<StoreInst>(Val: I)) |
| 1927 | return !SI->isUnordered(); |
| 1928 | else if (auto *LI = dyn_cast<LoadInst>(Val: I)) |
| 1929 | return !LI->isUnordered(); |
| 1930 | else { |
| 1931 | llvm_unreachable("unknown atomic instruction?" ); |
| 1932 | } |
| 1933 | } |
| 1934 | |
| 1935 | static bool InstrBreaksNoSync(Instruction &I, const SCCNodeSet &SCCNodes) { |
| 1936 | // Volatile may synchronize |
| 1937 | if (I.isVolatile()) |
| 1938 | return true; |
| 1939 | |
| 1940 | // An ordered atomic may synchronize. (See comment about on monotonic.) |
| 1941 | if (isOrderedAtomic(I: &I)) |
| 1942 | return true; |
| 1943 | |
| 1944 | auto *CB = dyn_cast<CallBase>(Val: &I); |
| 1945 | if (!CB) |
| 1946 | // Non call site cases covered by the two checks above |
| 1947 | return false; |
| 1948 | |
| 1949 | if (CB->hasFnAttr(Kind: Attribute::NoSync)) |
| 1950 | return false; |
| 1951 | |
| 1952 | // Non volatile memset/memcpy/memmoves are nosync |
| 1953 | // NOTE: Only intrinsics with volatile flags should be handled here. All |
| 1954 | // others should be marked in Intrinsics.td. |
| 1955 | if (auto *MI = dyn_cast<MemIntrinsic>(Val: &I)) |
| 1956 | if (!MI->isVolatile()) |
| 1957 | return false; |
| 1958 | |
| 1959 | // Speculatively assume in SCC. |
| 1960 | if (Function *Callee = CB->getCalledFunction()) |
| 1961 | if (SCCNodes.contains(key: Callee)) |
| 1962 | return false; |
| 1963 | |
| 1964 | return true; |
| 1965 | } |
| 1966 | |
| 1967 | /// Attempt to remove convergent function attribute when possible. |
| 1968 | /// |
| 1969 | /// Returns true if any changes to function attributes were made. |
| 1970 | static void inferConvergent(const SCCNodeSet &SCCNodes, |
| 1971 | SmallPtrSet<Function *, 8> &Changed) { |
| 1972 | AttributeInferer AI; |
| 1973 | |
| 1974 | // Request to remove the convergent attribute from all functions in the SCC |
| 1975 | // if every callsite within the SCC is not convergent (except for calls |
| 1976 | // to functions within the SCC). |
| 1977 | // Note: Removal of the attr from the callsites will happen in |
| 1978 | // InstCombineCalls separately. |
| 1979 | AI.registerAttrInference(AttrInference: AttributeInferer::InferenceDescriptor{ |
| 1980 | Attribute::Convergent, |
| 1981 | // Skip non-convergent functions. |
| 1982 | [](const Function &F) { return !F.isConvergent(); }, |
| 1983 | // Instructions that break non-convergent assumption. |
| 1984 | [SCCNodes](Instruction &I) { |
| 1985 | return InstrBreaksNonConvergent(I, SCCNodes); |
| 1986 | }, |
| 1987 | [](Function &F) { |
| 1988 | LLVM_DEBUG(dbgs() << "Removing convergent attr from fn " << F.getName() |
| 1989 | << "\n" ); |
| 1990 | F.setNotConvergent(); |
| 1991 | }, |
| 1992 | /* RequiresExactDefinition= */ false}); |
| 1993 | // Perform all the requested attribute inference actions. |
| 1994 | AI.run(SCCNodes, Changed); |
| 1995 | } |
| 1996 | |
| 1997 | /// Infer attributes from all functions in the SCC by scanning every |
| 1998 | /// instruction for compliance to the attribute assumptions. |
| 1999 | /// |
| 2000 | /// Returns true if any changes to function attributes were made. |
| 2001 | static void inferAttrsFromFunctionBodies(const SCCNodeSet &SCCNodes, |
| 2002 | SmallPtrSet<Function *, 8> &Changed) { |
| 2003 | AttributeInferer AI; |
| 2004 | |
| 2005 | if (!DisableNoUnwindInference) |
| 2006 | // Request to infer nounwind attribute for all the functions in the SCC if |
| 2007 | // every callsite within the SCC is not throwing (except for calls to |
| 2008 | // functions within the SCC). Note that nounwind attribute suffers from |
| 2009 | // derefinement - results may change depending on how functions are |
| 2010 | // optimized. Thus it can be inferred only from exact definitions. |
| 2011 | AI.registerAttrInference(AttrInference: AttributeInferer::InferenceDescriptor{ |
| 2012 | Attribute::NoUnwind, |
| 2013 | // Skip non-throwing functions. |
| 2014 | [](const Function &F) { return F.doesNotThrow(); }, |
| 2015 | // Instructions that break non-throwing assumption. |
| 2016 | [&SCCNodes](Instruction &I) { |
| 2017 | return InstrBreaksNonThrowing(I, SCCNodes); |
| 2018 | }, |
| 2019 | [](Function &F) { |
| 2020 | LLVM_DEBUG(dbgs() |
| 2021 | << "Adding nounwind attr to fn " << F.getName() << "\n" ); |
| 2022 | F.setDoesNotThrow(); |
| 2023 | ++NumNoUnwind; |
| 2024 | }, |
| 2025 | /* RequiresExactDefinition= */ true}); |
| 2026 | |
| 2027 | if (!DisableNoFreeInference) |
| 2028 | // Request to infer nofree attribute for all the functions in the SCC if |
| 2029 | // every callsite within the SCC does not directly or indirectly free |
| 2030 | // memory (except for calls to functions within the SCC). Note that nofree |
| 2031 | // attribute suffers from derefinement - results may change depending on |
| 2032 | // how functions are optimized. Thus it can be inferred only from exact |
| 2033 | // definitions. |
| 2034 | AI.registerAttrInference(AttrInference: AttributeInferer::InferenceDescriptor{ |
| 2035 | Attribute::NoFree, |
| 2036 | // Skip functions known not to free memory. |
| 2037 | [](const Function &F) { return F.doesNotFreeMemory(); }, |
| 2038 | // Instructions that break non-deallocating assumption. |
| 2039 | [&SCCNodes](Instruction &I) { |
| 2040 | return InstrBreaksNoFree(I, SCCNodes); |
| 2041 | }, |
| 2042 | [](Function &F) { |
| 2043 | LLVM_DEBUG(dbgs() |
| 2044 | << "Adding nofree attr to fn " << F.getName() << "\n" ); |
| 2045 | F.setDoesNotFreeMemory(); |
| 2046 | ++NumNoFree; |
| 2047 | }, |
| 2048 | /* RequiresExactDefinition= */ true}); |
| 2049 | |
| 2050 | AI.registerAttrInference(AttrInference: AttributeInferer::InferenceDescriptor{ |
| 2051 | Attribute::NoSync, |
| 2052 | // Skip already marked functions. |
| 2053 | [](const Function &F) { return F.hasNoSync(); }, |
| 2054 | // Instructions that break nosync assumption. |
| 2055 | [&SCCNodes](Instruction &I) { |
| 2056 | return InstrBreaksNoSync(I, SCCNodes); |
| 2057 | }, |
| 2058 | [](Function &F) { |
| 2059 | LLVM_DEBUG(dbgs() |
| 2060 | << "Adding nosync attr to fn " << F.getName() << "\n" ); |
| 2061 | F.setNoSync(); |
| 2062 | ++NumNoSync; |
| 2063 | }, |
| 2064 | /* RequiresExactDefinition= */ true}); |
| 2065 | |
| 2066 | // Perform all the requested attribute inference actions. |
| 2067 | AI.run(SCCNodes, Changed); |
| 2068 | } |
| 2069 | |
| 2070 | // Determines if the function 'F' can be marked 'norecurse'. |
| 2071 | // It returns true if any call within 'F' could lead to a recursive |
| 2072 | // call back to 'F', and false otherwise. |
| 2073 | // The 'AnyFunctionsAddressIsTaken' parameter is a module-wide flag |
| 2074 | // that is true if any function's address is taken, or if any function |
| 2075 | // has external linkage. This is used to determine the safety of |
| 2076 | // external/library calls. |
| 2077 | static bool mayHaveRecursiveCallee(Function &F, |
| 2078 | bool AnyFunctionsAddressIsTaken = true) { |
| 2079 | for (const auto &BB : F) { |
| 2080 | for (const auto &I : BB.instructionsWithoutDebug()) { |
| 2081 | if (const auto *CB = dyn_cast<CallBase>(Val: &I)) { |
| 2082 | const Function *Callee = CB->getCalledFunction(); |
| 2083 | if (!Callee || Callee == &F) |
| 2084 | return true; |
| 2085 | |
| 2086 | if (Callee->doesNotRecurse()) |
| 2087 | continue; |
| 2088 | |
| 2089 | if (!AnyFunctionsAddressIsTaken || |
| 2090 | (Callee->isDeclaration() && |
| 2091 | Callee->hasFnAttribute(Kind: Attribute::NoCallback))) |
| 2092 | continue; |
| 2093 | return true; |
| 2094 | } |
| 2095 | } |
| 2096 | } |
| 2097 | return false; |
| 2098 | } |
| 2099 | |
| 2100 | static void addNoRecurseAttrs(const SCCNodeSet &SCCNodes, |
| 2101 | SmallPtrSet<Function *, 8> &Changed) { |
| 2102 | // Try and identify functions that do not recurse. |
| 2103 | |
| 2104 | // If the SCC contains multiple nodes we know for sure there is recursion. |
| 2105 | if (SCCNodes.size() != 1) |
| 2106 | return; |
| 2107 | |
| 2108 | Function *F = *SCCNodes.begin(); |
| 2109 | if (!F || !F->hasExactDefinition() || F->doesNotRecurse()) |
| 2110 | return; |
| 2111 | if (!mayHaveRecursiveCallee(F&: *F)) { |
| 2112 | // Every call was to a non-recursive function other than this function, and |
| 2113 | // we have no indirect recursion as the SCC size is one. This function |
| 2114 | // cannot recurse. |
| 2115 | F->setDoesNotRecurse(); |
| 2116 | ++NumNoRecurse; |
| 2117 | Changed.insert(Ptr: F); |
| 2118 | } |
| 2119 | } |
| 2120 | |
| 2121 | // Set the noreturn function attribute if possible. |
| 2122 | static void addNoReturnAttrs(const SCCNodeSet &SCCNodes, |
| 2123 | SmallPtrSet<Function *, 8> &Changed) { |
| 2124 | for (Function *F : SCCNodes) { |
| 2125 | if (!F || !F->hasExactDefinition() || F->hasFnAttribute(Kind: Attribute::Naked) || |
| 2126 | F->doesNotReturn()) |
| 2127 | continue; |
| 2128 | |
| 2129 | if (!canReturn(F: *F)) { |
| 2130 | F->setDoesNotReturn(); |
| 2131 | Changed.insert(Ptr: F); |
| 2132 | } |
| 2133 | } |
| 2134 | } |
| 2135 | |
| 2136 | static bool allPathsGoThroughCold(Function &F) { |
| 2137 | SmallDenseMap<BasicBlock *, bool, 16> ColdPaths; |
| 2138 | ColdPaths[&F.front()] = false; |
| 2139 | SmallVector<BasicBlock *> Jobs; |
| 2140 | Jobs.push_back(Elt: &F.front()); |
| 2141 | |
| 2142 | while (!Jobs.empty()) { |
| 2143 | BasicBlock *BB = Jobs.pop_back_val(); |
| 2144 | |
| 2145 | // If block contains a cold callsite this path through the CG is cold. |
| 2146 | // Ignore whether the instructions actually are guaranteed to transfer |
| 2147 | // execution. Divergent behavior is considered unlikely. |
| 2148 | if (any_of(Range&: *BB, P: [](Instruction &I) { |
| 2149 | if (auto *CB = dyn_cast<CallBase>(Val: &I)) |
| 2150 | return CB->hasFnAttr(Kind: Attribute::Cold); |
| 2151 | return false; |
| 2152 | })) { |
| 2153 | ColdPaths[BB] = true; |
| 2154 | continue; |
| 2155 | } |
| 2156 | |
| 2157 | auto Succs = successors(BB); |
| 2158 | // We found a path that doesn't go through any cold callsite. |
| 2159 | if (Succs.empty()) |
| 2160 | return false; |
| 2161 | |
| 2162 | // We didn't find a cold callsite in this BB, so check that all successors |
| 2163 | // contain a cold callsite (or that their successors do). |
| 2164 | // Potential TODO: We could use static branch hints to assume certain |
| 2165 | // successor paths are inherently cold, irrespective of if they contain a |
| 2166 | // cold callsite. |
| 2167 | for (BasicBlock *Succ : Succs) { |
| 2168 | // Start with false, this is necessary to ensure we don't turn loops into |
| 2169 | // cold. |
| 2170 | auto [Iter, Inserted] = ColdPaths.try_emplace(Key: Succ, Args: false); |
| 2171 | if (!Inserted) { |
| 2172 | if (Iter->second) |
| 2173 | continue; |
| 2174 | return false; |
| 2175 | } |
| 2176 | Jobs.push_back(Elt: Succ); |
| 2177 | } |
| 2178 | } |
| 2179 | return true; |
| 2180 | } |
| 2181 | |
| 2182 | // Set the cold function attribute if possible. |
| 2183 | static void addColdAttrs(const SCCNodeSet &SCCNodes, |
| 2184 | SmallPtrSet<Function *, 8> &Changed) { |
| 2185 | for (Function *F : SCCNodes) { |
| 2186 | if (!F || !F->hasExactDefinition() || F->hasFnAttribute(Kind: Attribute::Naked) || |
| 2187 | F->hasFnAttribute(Kind: Attribute::Cold) || F->hasFnAttribute(Kind: Attribute::Hot)) |
| 2188 | continue; |
| 2189 | |
| 2190 | // Potential TODO: We could add attribute `cold` on functions with `coldcc`. |
| 2191 | if (allPathsGoThroughCold(F&: *F)) { |
| 2192 | F->addFnAttr(Kind: Attribute::Cold); |
| 2193 | ++NumCold; |
| 2194 | Changed.insert(Ptr: F); |
| 2195 | continue; |
| 2196 | } |
| 2197 | } |
| 2198 | } |
| 2199 | |
| 2200 | static bool functionWillReturn(const Function &F) { |
| 2201 | // We can infer and propagate function attributes only when we know that the |
| 2202 | // definition we'll get at link time is *exactly* the definition we see now. |
| 2203 | // For more details, see GlobalValue::mayBeDerefined. |
| 2204 | if (!F.hasExactDefinition()) |
| 2205 | return false; |
| 2206 | |
| 2207 | // Must-progress function without side-effects must return. |
| 2208 | if (F.mustProgress() && F.onlyReadsMemory()) |
| 2209 | return true; |
| 2210 | |
| 2211 | // Can only analyze functions with a definition. |
| 2212 | if (F.isDeclaration()) |
| 2213 | return false; |
| 2214 | |
| 2215 | // Functions with loops require more sophisticated analysis, as the loop |
| 2216 | // may be infinite. For now, don't try to handle them. |
| 2217 | SmallVector<std::pair<const BasicBlock *, const BasicBlock *>> Backedges; |
| 2218 | FindFunctionBackedges(F, Result&: Backedges); |
| 2219 | if (!Backedges.empty()) |
| 2220 | return false; |
| 2221 | |
| 2222 | // If there are no loops, then the function is willreturn if all calls in |
| 2223 | // it are willreturn. |
| 2224 | return all_of(Range: instructions(F), P: [](const Instruction &I) { |
| 2225 | return I.willReturn(); |
| 2226 | }); |
| 2227 | } |
| 2228 | |
| 2229 | // Set the willreturn function attribute if possible. |
| 2230 | static void addWillReturn(const SCCNodeSet &SCCNodes, |
| 2231 | SmallPtrSet<Function *, 8> &Changed) { |
| 2232 | for (Function *F : SCCNodes) { |
| 2233 | if (!F || F->willReturn() || !functionWillReturn(F: *F)) |
| 2234 | continue; |
| 2235 | |
| 2236 | F->setWillReturn(); |
| 2237 | NumWillReturn++; |
| 2238 | Changed.insert(Ptr: F); |
| 2239 | } |
| 2240 | } |
| 2241 | |
| 2242 | static SCCNodesResult createSCCNodeSet(ArrayRef<Function *> Functions) { |
| 2243 | SCCNodesResult Res; |
| 2244 | for (Function *F : Functions) { |
| 2245 | if (!F || F->hasOptNone() || F->hasFnAttribute(Kind: Attribute::Naked) || |
| 2246 | F->isPresplitCoroutine()) { |
| 2247 | // Omit any functions we're trying not to optimize from the set. |
| 2248 | continue; |
| 2249 | } |
| 2250 | |
| 2251 | Res.SCCNodes.insert(X: F); |
| 2252 | } |
| 2253 | return Res; |
| 2254 | } |
| 2255 | |
| 2256 | template <typename AARGetterT> |
| 2257 | static SmallPtrSet<Function *, 8> |
| 2258 | deriveAttrsInPostOrder(ArrayRef<Function *> Functions, AARGetterT &&AARGetter, |
| 2259 | bool ArgAttrsOnly) { |
| 2260 | SCCNodesResult Nodes = createSCCNodeSet(Functions); |
| 2261 | |
| 2262 | // Bail if the SCC only contains optnone functions. |
| 2263 | if (Nodes.SCCNodes.empty()) |
| 2264 | return {}; |
| 2265 | |
| 2266 | SmallPtrSet<Function *, 8> Changed; |
| 2267 | if (ArgAttrsOnly) { |
| 2268 | // ArgAttrsOnly means to only infer attributes that may aid optimizations |
| 2269 | // on the *current* function. "initializes" attribute is to aid |
| 2270 | // optimizations (like DSE) on the callers, so skip "initializes" here. |
| 2271 | addArgumentAttrs(SCCNodes: Nodes.SCCNodes, Changed, /*SkipInitializes=*/true); |
| 2272 | return Changed; |
| 2273 | } |
| 2274 | |
| 2275 | addArgumentReturnedAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
| 2276 | addMemoryAttrs(Nodes.SCCNodes, AARGetter, Changed); |
| 2277 | addArgumentAttrs(SCCNodes: Nodes.SCCNodes, Changed, /*SkipInitializes=*/false); |
| 2278 | inferConvergent(SCCNodes: Nodes.SCCNodes, Changed); |
| 2279 | addNoReturnAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
| 2280 | addColdAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
| 2281 | addWillReturn(SCCNodes: Nodes.SCCNodes, Changed); |
| 2282 | addNoUndefAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
| 2283 | addNoAliasAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
| 2284 | addNonNullAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
| 2285 | inferAttrsFromFunctionBodies(SCCNodes: Nodes.SCCNodes, Changed); |
| 2286 | addNoRecurseAttrs(SCCNodes: Nodes.SCCNodes, Changed); |
| 2287 | |
| 2288 | // Finally, infer the maximal set of attributes from the ones we've inferred |
| 2289 | // above. This is handling the cases where one attribute on a signature |
| 2290 | // implies another, but for implementation reasons the inference rule for |
| 2291 | // the later is missing (or simply less sophisticated). |
| 2292 | for (Function *F : Nodes.SCCNodes) |
| 2293 | if (F) |
| 2294 | if (inferAttributesFromOthers(F&: *F)) |
| 2295 | Changed.insert(Ptr: F); |
| 2296 | |
| 2297 | return Changed; |
| 2298 | } |
| 2299 | |
| 2300 | PreservedAnalyses PostOrderFunctionAttrsPass::run(LazyCallGraph::SCC &C, |
| 2301 | CGSCCAnalysisManager &AM, |
| 2302 | LazyCallGraph &CG, |
| 2303 | CGSCCUpdateResult &) { |
| 2304 | // Skip non-recursive functions if requested. |
| 2305 | // Only infer argument attributes for non-recursive functions, because |
| 2306 | // it can affect optimization behavior in conjunction with noalias. |
| 2307 | bool ArgAttrsOnly = false; |
| 2308 | if (C.size() == 1 && SkipNonRecursive) { |
| 2309 | LazyCallGraph::Node &N = *C.begin(); |
| 2310 | if (!N->lookup(N)) |
| 2311 | ArgAttrsOnly = true; |
| 2312 | } |
| 2313 | |
| 2314 | FunctionAnalysisManager &FAM = |
| 2315 | AM.getResult<FunctionAnalysisManagerCGSCCProxy>(IR&: C, ExtraArgs&: CG).getManager(); |
| 2316 | |
| 2317 | // We pass a lambda into functions to wire them up to the analysis manager |
| 2318 | // for getting function analyses. |
| 2319 | auto AARGetter = [&](Function &F) -> AAResults & { |
| 2320 | return FAM.getResult<AAManager>(IR&: F); |
| 2321 | }; |
| 2322 | |
| 2323 | SmallVector<Function *, 8> Functions; |
| 2324 | for (LazyCallGraph::Node &N : C) { |
| 2325 | Functions.push_back(Elt: &N.getFunction()); |
| 2326 | } |
| 2327 | |
| 2328 | auto ChangedFunctions = |
| 2329 | deriveAttrsInPostOrder(Functions, AARGetter, ArgAttrsOnly); |
| 2330 | if (ChangedFunctions.empty()) |
| 2331 | return PreservedAnalyses::all(); |
| 2332 | |
| 2333 | // Invalidate analyses for modified functions so that we don't have to |
| 2334 | // invalidate all analyses for all functions in this SCC. |
| 2335 | PreservedAnalyses FuncPA; |
| 2336 | // We haven't changed the CFG for modified functions. |
| 2337 | FuncPA.preserveSet<CFGAnalyses>(); |
| 2338 | for (Function *Changed : ChangedFunctions) { |
| 2339 | FAM.invalidate(IR&: *Changed, PA: FuncPA); |
| 2340 | // Also invalidate any direct callers of changed functions since analyses |
| 2341 | // may care about attributes of direct callees. For example, MemorySSA cares |
| 2342 | // about whether or not a call's callee modifies memory and queries that |
| 2343 | // through function attributes. |
| 2344 | for (auto *U : Changed->users()) { |
| 2345 | if (auto *Call = dyn_cast<CallBase>(Val: U)) { |
| 2346 | if (Call->getCalledOperand() == Changed) |
| 2347 | FAM.invalidate(IR&: *Call->getFunction(), PA: FuncPA); |
| 2348 | } |
| 2349 | } |
| 2350 | } |
| 2351 | |
| 2352 | PreservedAnalyses PA; |
| 2353 | // We have not added or removed functions. |
| 2354 | PA.preserve<FunctionAnalysisManagerCGSCCProxy>(); |
| 2355 | // We already invalidated all relevant function analyses above. |
| 2356 | PA.preserveSet<AllAnalysesOn<Function>>(); |
| 2357 | return PA; |
| 2358 | } |
| 2359 | |
| 2360 | void PostOrderFunctionAttrsPass::printPipeline( |
| 2361 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
| 2362 | static_cast<PassInfoMixin<PostOrderFunctionAttrsPass> *>(this)->printPipeline( |
| 2363 | OS, MapClassName2PassName); |
| 2364 | if (SkipNonRecursive) |
| 2365 | OS << "<skip-non-recursive-function-attrs>" ; |
| 2366 | } |
| 2367 | |
| 2368 | template <typename AARGetterT> |
| 2369 | static bool runImpl(CallGraphSCC &SCC, AARGetterT AARGetter) { |
| 2370 | SmallVector<Function *, 8> Functions; |
| 2371 | for (CallGraphNode *I : SCC) { |
| 2372 | Functions.push_back(Elt: I->getFunction()); |
| 2373 | } |
| 2374 | |
| 2375 | return !deriveAttrsInPostOrder(Functions, AARGetter).empty(); |
| 2376 | } |
| 2377 | |
| 2378 | static bool addNoRecurseAttrsTopDown(Function &F) { |
| 2379 | // We check the preconditions for the function prior to calling this to avoid |
| 2380 | // the cost of building up a reversible post-order list. We assert them here |
| 2381 | // to make sure none of the invariants this relies on were violated. |
| 2382 | assert(!F.isDeclaration() && "Cannot deduce norecurse without a definition!" ); |
| 2383 | assert(!F.doesNotRecurse() && |
| 2384 | "This function has already been deduced as norecurs!" ); |
| 2385 | assert(F.hasInternalLinkage() && |
| 2386 | "Can only do top-down deduction for internal linkage functions!" ); |
| 2387 | |
| 2388 | // If F is internal and all of its uses are calls from a non-recursive |
| 2389 | // functions, then none of its calls could in fact recurse without going |
| 2390 | // through a function marked norecurse, and so we can mark this function too |
| 2391 | // as norecurse. Note that the uses must actually be calls -- otherwise |
| 2392 | // a pointer to this function could be returned from a norecurse function but |
| 2393 | // this function could be recursively (indirectly) called. Note that this |
| 2394 | // also detects if F is directly recursive as F is not yet marked as |
| 2395 | // a norecurse function. |
| 2396 | for (auto &U : F.uses()) { |
| 2397 | auto *I = dyn_cast<Instruction>(Val: U.getUser()); |
| 2398 | if (!I) |
| 2399 | return false; |
| 2400 | CallBase *CB = dyn_cast<CallBase>(Val: I); |
| 2401 | if (!CB || !CB->isCallee(U: &U) || |
| 2402 | !CB->getParent()->getParent()->doesNotRecurse()) |
| 2403 | return false; |
| 2404 | } |
| 2405 | F.setDoesNotRecurse(); |
| 2406 | ++NumNoRecurse; |
| 2407 | return true; |
| 2408 | } |
| 2409 | |
| 2410 | static bool deduceFunctionAttributeInRPO(Module &M, LazyCallGraph &CG) { |
| 2411 | // We only have a post-order SCC traversal (because SCCs are inherently |
| 2412 | // discovered in post-order), so we accumulate them in a vector and then walk |
| 2413 | // it in reverse. This is simpler than using the RPO iterator infrastructure |
| 2414 | // because we need to combine SCC detection and the PO walk of the call |
| 2415 | // graph. We can also cheat egregiously because we're primarily interested in |
| 2416 | // synthesizing norecurse and so we can only save the singular SCCs as SCCs |
| 2417 | // with multiple functions in them will clearly be recursive. |
| 2418 | |
| 2419 | SmallVector<Function *, 16> Worklist; |
| 2420 | CG.buildRefSCCs(); |
| 2421 | for (LazyCallGraph::RefSCC &RC : CG.postorder_ref_sccs()) { |
| 2422 | for (LazyCallGraph::SCC &SCC : RC) { |
| 2423 | if (SCC.size() != 1) |
| 2424 | continue; |
| 2425 | Function &F = SCC.begin()->getFunction(); |
| 2426 | if (!F.isDeclaration() && !F.doesNotRecurse() && F.hasInternalLinkage()) |
| 2427 | Worklist.push_back(Elt: &F); |
| 2428 | } |
| 2429 | } |
| 2430 | bool Changed = false; |
| 2431 | for (auto *F : llvm::reverse(C&: Worklist)) |
| 2432 | Changed |= addNoRecurseAttrsTopDown(F&: *F); |
| 2433 | |
| 2434 | return Changed; |
| 2435 | } |
| 2436 | |
| 2437 | PreservedAnalyses |
| 2438 | ReversePostOrderFunctionAttrsPass::run(Module &M, ModuleAnalysisManager &AM) { |
| 2439 | auto &CG = AM.getResult<LazyCallGraphAnalysis>(IR&: M); |
| 2440 | |
| 2441 | if (!deduceFunctionAttributeInRPO(M, CG)) |
| 2442 | return PreservedAnalyses::all(); |
| 2443 | |
| 2444 | PreservedAnalyses PA; |
| 2445 | PA.preserve<LazyCallGraphAnalysis>(); |
| 2446 | return PA; |
| 2447 | } |
| 2448 | |
| 2449 | PreservedAnalyses NoRecurseLTOInferencePass::run(Module &M, |
| 2450 | ModuleAnalysisManager &MAM) { |
| 2451 | |
| 2452 | // Check if any function in the whole program has its address taken or has |
| 2453 | // potentially external linkage. |
| 2454 | // We use this information when inferring norecurse attribute: If there is |
| 2455 | // no function whose address is taken and all functions have internal |
| 2456 | // linkage, there is no path for a callback to any user function. |
| 2457 | bool AnyFunctionsAddressIsTaken = false; |
| 2458 | for (Function &F : M) { |
| 2459 | if (F.isDeclaration() || F.doesNotRecurse()) |
| 2460 | continue; |
| 2461 | if (!F.hasLocalLinkage() || F.hasAddressTaken()) { |
| 2462 | AnyFunctionsAddressIsTaken = true; |
| 2463 | break; |
| 2464 | } |
| 2465 | } |
| 2466 | |
| 2467 | // Run norecurse inference on all RefSCCs in the LazyCallGraph for this |
| 2468 | // module. |
| 2469 | bool Changed = false; |
| 2470 | LazyCallGraph &CG = MAM.getResult<LazyCallGraphAnalysis>(IR&: M); |
| 2471 | CG.buildRefSCCs(); |
| 2472 | |
| 2473 | for (LazyCallGraph::RefSCC &RC : CG.postorder_ref_sccs()) { |
| 2474 | // Skip any RefSCC that is part of a call cycle. A RefSCC containing more |
| 2475 | // than one SCC indicates a recursive relationship involving indirect calls. |
| 2476 | if (RC.size() > 1) |
| 2477 | continue; |
| 2478 | |
| 2479 | // RefSCC contains a single-SCC. SCC size > 1 indicates mutually recursive |
| 2480 | // functions. Ex: foo1 -> foo2 -> foo3 -> foo1. |
| 2481 | LazyCallGraph::SCC &S = *RC.begin(); |
| 2482 | if (S.size() > 1) |
| 2483 | continue; |
| 2484 | |
| 2485 | // Get the single function from this SCC. |
| 2486 | Function &F = S.begin()->getFunction(); |
| 2487 | if (!F.hasExactDefinition() || F.doesNotRecurse()) |
| 2488 | continue; |
| 2489 | |
| 2490 | // If the analysis confirms that this function has no recursive calls |
| 2491 | // (either direct, indirect, or through external linkages), |
| 2492 | // we can safely apply the norecurse attribute. |
| 2493 | if (!mayHaveRecursiveCallee(F, AnyFunctionsAddressIsTaken)) { |
| 2494 | F.setDoesNotRecurse(); |
| 2495 | ++NumNoRecurse; |
| 2496 | Changed = true; |
| 2497 | } |
| 2498 | } |
| 2499 | |
| 2500 | PreservedAnalyses PA; |
| 2501 | if (Changed) |
| 2502 | PA.preserve<LazyCallGraphAnalysis>(); |
| 2503 | else |
| 2504 | PA = PreservedAnalyses::all(); |
| 2505 | return PA; |
| 2506 | } |
| 2507 | |