| 1 | //===--- CodeGenPGO.cpp - PGO Instrumentation for LLVM CodeGen --*- C++ -*-===// |
| 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 | // Instrumentation-based profile-guided optimization |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "CodeGenPGO.h" |
| 14 | #include "CGDebugInfo.h" |
| 15 | #include "CodeGenFunction.h" |
| 16 | #include "CoverageMappingGen.h" |
| 17 | #include "clang/AST/RecursiveASTVisitor.h" |
| 18 | #include "clang/AST/StmtVisitor.h" |
| 19 | #include "llvm/IR/Intrinsics.h" |
| 20 | #include "llvm/IR/MDBuilder.h" |
| 21 | #include "llvm/Support/CommandLine.h" |
| 22 | #include "llvm/Support/Endian.h" |
| 23 | #include "llvm/Support/MD5.h" |
| 24 | #include <optional> |
| 25 | |
| 26 | namespace llvm { |
| 27 | extern cl::opt<bool> EnableSingleByteCoverage; |
| 28 | } // namespace llvm |
| 29 | |
| 30 | static llvm::cl::opt<bool> |
| 31 | EnableValueProfiling("enable-value-profiling" , |
| 32 | llvm::cl::desc("Enable value profiling" ), |
| 33 | llvm::cl::Hidden, llvm::cl::init(Val: false)); |
| 34 | |
| 35 | using namespace clang; |
| 36 | using namespace CodeGen; |
| 37 | |
| 38 | void CodeGenPGO::setFuncName(StringRef Name, |
| 39 | llvm::GlobalValue::LinkageTypes Linkage) { |
| 40 | llvm::IndexedInstrProfReader *PGOReader = CGM.getPGOReader(); |
| 41 | FuncName = llvm::getPGOFuncName( |
| 42 | RawFuncName: Name, Linkage, FileName: CGM.getCodeGenOpts().MainFileName, |
| 43 | Version: PGOReader ? PGOReader->getVersion() : llvm::IndexedInstrProf::Version); |
| 44 | |
| 45 | // If we're generating a profile, create a variable for the name. |
| 46 | if (CGM.getCodeGenOpts().hasProfileClangInstr()) |
| 47 | FuncNameVar = llvm::createPGOFuncNameVar(M&: CGM.getModule(), Linkage, PGOFuncName: FuncName); |
| 48 | } |
| 49 | |
| 50 | void CodeGenPGO::setFuncName(llvm::Function *Fn) { |
| 51 | setFuncName(Name: Fn->getName(), Linkage: Fn->getLinkage()); |
| 52 | // Create PGOFuncName meta data. |
| 53 | llvm::createPGOFuncNameMetadata(F&: *Fn, PGOFuncName: FuncName); |
| 54 | } |
| 55 | |
| 56 | /// The version of the PGO hash algorithm. |
| 57 | enum PGOHashVersion : unsigned { |
| 58 | PGO_HASH_V1, |
| 59 | PGO_HASH_V2, |
| 60 | PGO_HASH_V3, |
| 61 | |
| 62 | // Keep this set to the latest hash version. |
| 63 | PGO_HASH_LATEST = PGO_HASH_V3 |
| 64 | }; |
| 65 | |
| 66 | namespace { |
| 67 | /// Stable hasher for PGO region counters. |
| 68 | /// |
| 69 | /// PGOHash produces a stable hash of a given function's control flow. |
| 70 | /// |
| 71 | /// Changing the output of this hash will invalidate all previously generated |
| 72 | /// profiles -- i.e., don't do it. |
| 73 | /// |
| 74 | /// \note When this hash does eventually change (years?), we still need to |
| 75 | /// support old hashes. We'll need to pull in the version number from the |
| 76 | /// profile data format and use the matching hash function. |
| 77 | class PGOHash { |
| 78 | uint64_t Working; |
| 79 | unsigned Count; |
| 80 | PGOHashVersion HashVersion; |
| 81 | llvm::MD5 MD5; |
| 82 | |
| 83 | static const int NumBitsPerType = 6; |
| 84 | static const unsigned NumTypesPerWord = sizeof(uint64_t) * 8 / NumBitsPerType; |
| 85 | static const unsigned TooBig = 1u << NumBitsPerType; |
| 86 | |
| 87 | public: |
| 88 | /// Hash values for AST nodes. |
| 89 | /// |
| 90 | /// Distinct values for AST nodes that have region counters attached. |
| 91 | /// |
| 92 | /// These values must be stable. All new members must be added at the end, |
| 93 | /// and no members should be removed. Changing the enumeration value for an |
| 94 | /// AST node will affect the hash of every function that contains that node. |
| 95 | enum HashType : unsigned char { |
| 96 | None = 0, |
| 97 | LabelStmt = 1, |
| 98 | WhileStmt, |
| 99 | DoStmt, |
| 100 | ForStmt, |
| 101 | CXXForRangeStmt, |
| 102 | ObjCForCollectionStmt, |
| 103 | SwitchStmt, |
| 104 | CaseStmt, |
| 105 | DefaultStmt, |
| 106 | IfStmt, |
| 107 | CXXTryStmt, |
| 108 | CXXCatchStmt, |
| 109 | ConditionalOperator, |
| 110 | BinaryOperatorLAnd, |
| 111 | BinaryOperatorLOr, |
| 112 | BinaryConditionalOperator, |
| 113 | // The preceding values are available with PGO_HASH_V1. |
| 114 | |
| 115 | EndOfScope, |
| 116 | IfThenBranch, |
| 117 | IfElseBranch, |
| 118 | GotoStmt, |
| 119 | IndirectGotoStmt, |
| 120 | BreakStmt, |
| 121 | ContinueStmt, |
| 122 | ReturnStmt, |
| 123 | ThrowExpr, |
| 124 | UnaryOperatorLNot, |
| 125 | BinaryOperatorLT, |
| 126 | BinaryOperatorGT, |
| 127 | BinaryOperatorLE, |
| 128 | BinaryOperatorGE, |
| 129 | BinaryOperatorEQ, |
| 130 | BinaryOperatorNE, |
| 131 | // The preceding values are available since PGO_HASH_V2. |
| 132 | |
| 133 | // Keep this last. It's for the static assert that follows. |
| 134 | LastHashType |
| 135 | }; |
| 136 | static_assert(LastHashType <= TooBig, "Too many types in HashType" ); |
| 137 | |
| 138 | PGOHash(PGOHashVersion HashVersion) |
| 139 | : Working(0), Count(0), HashVersion(HashVersion) {} |
| 140 | void combine(HashType Type); |
| 141 | uint64_t finalize(); |
| 142 | PGOHashVersion getHashVersion() const { return HashVersion; } |
| 143 | }; |
| 144 | const int PGOHash::NumBitsPerType; |
| 145 | const unsigned PGOHash::NumTypesPerWord; |
| 146 | const unsigned PGOHash::TooBig; |
| 147 | |
| 148 | /// Get the PGO hash version used in the given indexed profile. |
| 149 | static PGOHashVersion getPGOHashVersion(llvm::IndexedInstrProfReader *PGOReader, |
| 150 | CodeGenModule &CGM) { |
| 151 | if (PGOReader->getVersion() <= 4) |
| 152 | return PGO_HASH_V1; |
| 153 | if (PGOReader->getVersion() <= 5) |
| 154 | return PGO_HASH_V2; |
| 155 | return PGO_HASH_V3; |
| 156 | } |
| 157 | |
| 158 | /// A RecursiveASTVisitor that fills a map of statements to PGO counters. |
| 159 | struct MapRegionCounters : public RecursiveASTVisitor<MapRegionCounters> { |
| 160 | using Base = RecursiveASTVisitor<MapRegionCounters>; |
| 161 | |
| 162 | /// The next counter value to assign. |
| 163 | unsigned NextCounter; |
| 164 | /// The function hash. |
| 165 | PGOHash Hash; |
| 166 | /// The map of statements to counters. |
| 167 | llvm::DenseMap<const Stmt *, CounterPair> &CounterMap; |
| 168 | /// The state of MC/DC Coverage in this function. |
| 169 | MCDC::State &MCDCState; |
| 170 | /// Maximum number of supported MC/DC conditions in a boolean expression. |
| 171 | unsigned MCDCMaxCond; |
| 172 | /// The profile version. |
| 173 | uint64_t ProfileVersion; |
| 174 | /// Diagnostics Engine used to report warnings. |
| 175 | DiagnosticsEngine &Diag; |
| 176 | |
| 177 | MapRegionCounters(PGOHashVersion HashVersion, uint64_t ProfileVersion, |
| 178 | llvm::DenseMap<const Stmt *, CounterPair> &CounterMap, |
| 179 | MCDC::State &MCDCState, unsigned MCDCMaxCond, |
| 180 | DiagnosticsEngine &Diag) |
| 181 | : NextCounter(0), Hash(HashVersion), CounterMap(CounterMap), |
| 182 | MCDCState(MCDCState), MCDCMaxCond(MCDCMaxCond), |
| 183 | ProfileVersion(ProfileVersion), Diag(Diag) {} |
| 184 | |
| 185 | // Blocks and lambdas are handled as separate functions, so we need not |
| 186 | // traverse them in the parent context. |
| 187 | bool TraverseBlockExpr(BlockExpr *BE) { return true; } |
| 188 | bool TraverseLambdaExpr(LambdaExpr *LE) { |
| 189 | // Traverse the captures, but not the body. |
| 190 | for (auto C : zip(t: LE->captures(), u: LE->capture_inits())) |
| 191 | TraverseLambdaCapture(LE, C: &std::get<0>(t&: C), Init: std::get<1>(t&: C)); |
| 192 | return true; |
| 193 | } |
| 194 | bool TraverseCapturedStmt(CapturedStmt *CS) { return true; } |
| 195 | |
| 196 | bool VisitDecl(const Decl *D) { |
| 197 | switch (D->getKind()) { |
| 198 | default: |
| 199 | break; |
| 200 | case Decl::Function: |
| 201 | case Decl::CXXMethod: |
| 202 | case Decl::CXXConstructor: |
| 203 | case Decl::CXXDestructor: |
| 204 | case Decl::CXXConversion: |
| 205 | case Decl::ObjCMethod: |
| 206 | case Decl::Block: |
| 207 | case Decl::Captured: |
| 208 | CounterMap[D->getBody()] = NextCounter++; |
| 209 | break; |
| 210 | } |
| 211 | return true; |
| 212 | } |
| 213 | |
| 214 | /// If \p S gets a fresh counter, update the counter mappings. Return the |
| 215 | /// V1 hash of \p S. |
| 216 | PGOHash::HashType updateCounterMappings(Stmt *S) { |
| 217 | auto Type = getHashType(HashVersion: PGO_HASH_V1, S); |
| 218 | if (Type != PGOHash::None) |
| 219 | CounterMap[S] = NextCounter++; |
| 220 | return Type; |
| 221 | } |
| 222 | |
| 223 | /// The following stacks are used with dataTraverseStmtPre() and |
| 224 | /// dataTraverseStmtPost() to track the depth of nested logical operators in a |
| 225 | /// boolean expression in a function. The ultimate purpose is to keep track |
| 226 | /// of the number of leaf-level conditions in the boolean expression so that a |
| 227 | /// profile bitmap can be allocated based on that number. |
| 228 | /// |
| 229 | /// The stacks are also used to find error cases and notify the user. A |
| 230 | /// standard logical operator nest for a boolean expression could be in a form |
| 231 | /// similar to this: "x = a && b && c && (d || f)" |
| 232 | unsigned NumCond = 0; |
| 233 | bool SplitNestedLogicalOp = false; |
| 234 | SmallVector<const Stmt *, 16> NonLogOpStack; |
| 235 | SmallVector<const BinaryOperator *, 16> LogOpStack; |
| 236 | |
| 237 | // Hook: dataTraverseStmtPre() is invoked prior to visiting an AST Stmt node. |
| 238 | bool dataTraverseStmtPre(Stmt *S) { |
| 239 | /// If MC/DC is not enabled, MCDCMaxCond will be set to 0. Do nothing. |
| 240 | if (MCDCMaxCond == 0) |
| 241 | return true; |
| 242 | |
| 243 | /// At the top of the logical operator nest, reset the number of conditions, |
| 244 | /// also forget previously seen split nesting cases. |
| 245 | if (LogOpStack.empty()) { |
| 246 | NumCond = 0; |
| 247 | SplitNestedLogicalOp = false; |
| 248 | } |
| 249 | |
| 250 | if (const Expr *E = dyn_cast<Expr>(Val: S)) { |
| 251 | const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Val: E->IgnoreParens()); |
| 252 | if (BinOp && BinOp->isLogicalOp()) { |
| 253 | /// Check for "split-nested" logical operators. This happens when a new |
| 254 | /// boolean expression logical-op nest is encountered within an existing |
| 255 | /// boolean expression, separated by a non-logical operator. For |
| 256 | /// example, in "x = (a && b && c && foo(d && f))", the "d && f" case |
| 257 | /// starts a new boolean expression that is separated from the other |
| 258 | /// conditions by the operator foo(). Split-nested cases are not |
| 259 | /// supported by MC/DC. |
| 260 | SplitNestedLogicalOp = SplitNestedLogicalOp || !NonLogOpStack.empty(); |
| 261 | |
| 262 | LogOpStack.push_back(Elt: BinOp); |
| 263 | return true; |
| 264 | } |
| 265 | } |
| 266 | |
| 267 | /// Keep track of non-logical operators. These are OK as long as we don't |
| 268 | /// encounter a new logical operator after seeing one. |
| 269 | if (!LogOpStack.empty()) |
| 270 | NonLogOpStack.push_back(Elt: S); |
| 271 | |
| 272 | return true; |
| 273 | } |
| 274 | |
| 275 | // Hook: dataTraverseStmtPost() is invoked by the AST visitor after visiting |
| 276 | // an AST Stmt node. MC/DC will use it to to signal when the top of a |
| 277 | // logical operation (boolean expression) nest is encountered. |
| 278 | bool dataTraverseStmtPost(Stmt *S) { |
| 279 | /// If MC/DC is not enabled, MCDCMaxCond will be set to 0. Do nothing. |
| 280 | if (MCDCMaxCond == 0) |
| 281 | return true; |
| 282 | |
| 283 | if (const Expr *E = dyn_cast<Expr>(Val: S)) { |
| 284 | const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Val: E->IgnoreParens()); |
| 285 | if (BinOp && BinOp->isLogicalOp()) { |
| 286 | assert(LogOpStack.back() == BinOp); |
| 287 | LogOpStack.pop_back(); |
| 288 | |
| 289 | /// At the top of logical operator nest: |
| 290 | if (LogOpStack.empty()) { |
| 291 | /// Was the "split-nested" logical operator case encountered? |
| 292 | if (SplitNestedLogicalOp) { |
| 293 | unsigned DiagID = Diag.getCustomDiagID( |
| 294 | L: DiagnosticsEngine::Warning, |
| 295 | FormatString: "unsupported MC/DC boolean expression; " |
| 296 | "contains an operation with a nested boolean expression. " |
| 297 | "Expression will not be covered" ); |
| 298 | Diag.Report(Loc: S->getBeginLoc(), DiagID); |
| 299 | return true; |
| 300 | } |
| 301 | |
| 302 | /// Was the maximum number of conditions encountered? |
| 303 | if (NumCond > MCDCMaxCond) { |
| 304 | unsigned DiagID = Diag.getCustomDiagID( |
| 305 | L: DiagnosticsEngine::Warning, |
| 306 | FormatString: "unsupported MC/DC boolean expression; " |
| 307 | "number of conditions (%0) exceeds max (%1). " |
| 308 | "Expression will not be covered" ); |
| 309 | Diag.Report(Loc: S->getBeginLoc(), DiagID) << NumCond << MCDCMaxCond; |
| 310 | return true; |
| 311 | } |
| 312 | |
| 313 | // Otherwise, allocate the Decision. |
| 314 | MCDCState.DecisionByStmt[BinOp].BitmapIdx = 0; |
| 315 | } |
| 316 | return true; |
| 317 | } |
| 318 | } |
| 319 | |
| 320 | if (!LogOpStack.empty()) |
| 321 | NonLogOpStack.pop_back(); |
| 322 | |
| 323 | return true; |
| 324 | } |
| 325 | |
| 326 | /// The RHS of all logical operators gets a fresh counter in order to count |
| 327 | /// how many times the RHS evaluates to true or false, depending on the |
| 328 | /// semantics of the operator. This is only valid for ">= v7" of the profile |
| 329 | /// version so that we facilitate backward compatibility. In addition, in |
| 330 | /// order to use MC/DC, count the number of total LHS and RHS conditions. |
| 331 | bool VisitBinaryOperator(BinaryOperator *S) { |
| 332 | if (S->isLogicalOp()) { |
| 333 | if (CodeGenFunction::isInstrumentedCondition(C: S->getLHS())) |
| 334 | NumCond++; |
| 335 | |
| 336 | if (CodeGenFunction::isInstrumentedCondition(C: S->getRHS())) { |
| 337 | if (ProfileVersion >= llvm::IndexedInstrProf::Version7) |
| 338 | CounterMap[S->getRHS()] = NextCounter++; |
| 339 | |
| 340 | NumCond++; |
| 341 | } |
| 342 | } |
| 343 | return Base::VisitBinaryOperator(S); |
| 344 | } |
| 345 | |
| 346 | bool VisitConditionalOperator(ConditionalOperator *S) { |
| 347 | if (llvm::EnableSingleByteCoverage && S->getTrueExpr()) |
| 348 | CounterMap[S->getTrueExpr()] = NextCounter++; |
| 349 | if (llvm::EnableSingleByteCoverage && S->getFalseExpr()) |
| 350 | CounterMap[S->getFalseExpr()] = NextCounter++; |
| 351 | return Base::VisitConditionalOperator(S); |
| 352 | } |
| 353 | |
| 354 | /// Include \p S in the function hash. |
| 355 | bool VisitStmt(Stmt *S) { |
| 356 | auto Type = updateCounterMappings(S); |
| 357 | if (Hash.getHashVersion() != PGO_HASH_V1) |
| 358 | Type = getHashType(HashVersion: Hash.getHashVersion(), S); |
| 359 | if (Type != PGOHash::None) |
| 360 | Hash.combine(Type); |
| 361 | return true; |
| 362 | } |
| 363 | |
| 364 | bool TraverseIfStmt(IfStmt *If) { |
| 365 | // If we used the V1 hash, use the default traversal. |
| 366 | if (Hash.getHashVersion() == PGO_HASH_V1) |
| 367 | return Base::TraverseIfStmt(S: If); |
| 368 | |
| 369 | // When single byte coverage mode is enabled, add a counter to then and |
| 370 | // else. |
| 371 | bool NoSingleByteCoverage = !llvm::EnableSingleByteCoverage; |
| 372 | for (Stmt *CS : If->children()) { |
| 373 | if (!CS || NoSingleByteCoverage) |
| 374 | continue; |
| 375 | if (CS == If->getThen()) |
| 376 | CounterMap[If->getThen()] = NextCounter++; |
| 377 | else if (CS == If->getElse()) |
| 378 | CounterMap[If->getElse()] = NextCounter++; |
| 379 | } |
| 380 | |
| 381 | // Otherwise, keep track of which branch we're in while traversing. |
| 382 | VisitStmt(S: If); |
| 383 | |
| 384 | for (Stmt *CS : If->children()) { |
| 385 | if (!CS) |
| 386 | continue; |
| 387 | if (CS == If->getThen()) |
| 388 | Hash.combine(Type: PGOHash::IfThenBranch); |
| 389 | else if (CS == If->getElse()) |
| 390 | Hash.combine(Type: PGOHash::IfElseBranch); |
| 391 | TraverseStmt(S: CS); |
| 392 | } |
| 393 | Hash.combine(Type: PGOHash::EndOfScope); |
| 394 | return true; |
| 395 | } |
| 396 | |
| 397 | bool TraverseWhileStmt(WhileStmt *While) { |
| 398 | // When single byte coverage mode is enabled, add a counter to condition and |
| 399 | // body. |
| 400 | bool NoSingleByteCoverage = !llvm::EnableSingleByteCoverage; |
| 401 | for (Stmt *CS : While->children()) { |
| 402 | if (!CS || NoSingleByteCoverage) |
| 403 | continue; |
| 404 | if (CS == While->getCond()) |
| 405 | CounterMap[While->getCond()] = NextCounter++; |
| 406 | else if (CS == While->getBody()) |
| 407 | CounterMap[While->getBody()] = NextCounter++; |
| 408 | } |
| 409 | |
| 410 | Base::TraverseWhileStmt(S: While); |
| 411 | if (Hash.getHashVersion() != PGO_HASH_V1) |
| 412 | Hash.combine(Type: PGOHash::EndOfScope); |
| 413 | return true; |
| 414 | } |
| 415 | |
| 416 | bool TraverseDoStmt(DoStmt *Do) { |
| 417 | // When single byte coverage mode is enabled, add a counter to condition and |
| 418 | // body. |
| 419 | bool NoSingleByteCoverage = !llvm::EnableSingleByteCoverage; |
| 420 | for (Stmt *CS : Do->children()) { |
| 421 | if (!CS || NoSingleByteCoverage) |
| 422 | continue; |
| 423 | if (CS == Do->getCond()) |
| 424 | CounterMap[Do->getCond()] = NextCounter++; |
| 425 | else if (CS == Do->getBody()) |
| 426 | CounterMap[Do->getBody()] = NextCounter++; |
| 427 | } |
| 428 | |
| 429 | Base::TraverseDoStmt(S: Do); |
| 430 | if (Hash.getHashVersion() != PGO_HASH_V1) |
| 431 | Hash.combine(Type: PGOHash::EndOfScope); |
| 432 | return true; |
| 433 | } |
| 434 | |
| 435 | bool TraverseForStmt(ForStmt *For) { |
| 436 | // When single byte coverage mode is enabled, add a counter to condition, |
| 437 | // increment and body. |
| 438 | bool NoSingleByteCoverage = !llvm::EnableSingleByteCoverage; |
| 439 | for (Stmt *CS : For->children()) { |
| 440 | if (!CS || NoSingleByteCoverage) |
| 441 | continue; |
| 442 | if (CS == For->getCond()) |
| 443 | CounterMap[For->getCond()] = NextCounter++; |
| 444 | else if (CS == For->getInc()) |
| 445 | CounterMap[For->getInc()] = NextCounter++; |
| 446 | else if (CS == For->getBody()) |
| 447 | CounterMap[For->getBody()] = NextCounter++; |
| 448 | } |
| 449 | |
| 450 | Base::TraverseForStmt(S: For); |
| 451 | if (Hash.getHashVersion() != PGO_HASH_V1) |
| 452 | Hash.combine(Type: PGOHash::EndOfScope); |
| 453 | return true; |
| 454 | } |
| 455 | |
| 456 | bool TraverseCXXForRangeStmt(CXXForRangeStmt *ForRange) { |
| 457 | // When single byte coverage mode is enabled, add a counter to body. |
| 458 | bool NoSingleByteCoverage = !llvm::EnableSingleByteCoverage; |
| 459 | for (Stmt *CS : ForRange->children()) { |
| 460 | if (!CS || NoSingleByteCoverage) |
| 461 | continue; |
| 462 | if (CS == ForRange->getBody()) |
| 463 | CounterMap[ForRange->getBody()] = NextCounter++; |
| 464 | } |
| 465 | |
| 466 | Base::TraverseCXXForRangeStmt(S: ForRange); |
| 467 | if (Hash.getHashVersion() != PGO_HASH_V1) |
| 468 | Hash.combine(Type: PGOHash::EndOfScope); |
| 469 | return true; |
| 470 | } |
| 471 | |
| 472 | // If the statement type \p N is nestable, and its nesting impacts profile |
| 473 | // stability, define a custom traversal which tracks the end of the statement |
| 474 | // in the hash (provided we're not using the V1 hash). |
| 475 | #define DEFINE_NESTABLE_TRAVERSAL(N) \ |
| 476 | bool Traverse##N(N *S) { \ |
| 477 | Base::Traverse##N(S); \ |
| 478 | if (Hash.getHashVersion() != PGO_HASH_V1) \ |
| 479 | Hash.combine(PGOHash::EndOfScope); \ |
| 480 | return true; \ |
| 481 | } |
| 482 | |
| 483 | DEFINE_NESTABLE_TRAVERSAL(ObjCForCollectionStmt) |
| 484 | DEFINE_NESTABLE_TRAVERSAL(CXXTryStmt) |
| 485 | DEFINE_NESTABLE_TRAVERSAL(CXXCatchStmt) |
| 486 | |
| 487 | /// Get version \p HashVersion of the PGO hash for \p S. |
| 488 | PGOHash::HashType getHashType(PGOHashVersion HashVersion, const Stmt *S) { |
| 489 | switch (S->getStmtClass()) { |
| 490 | default: |
| 491 | break; |
| 492 | case Stmt::LabelStmtClass: |
| 493 | return PGOHash::LabelStmt; |
| 494 | case Stmt::WhileStmtClass: |
| 495 | return PGOHash::WhileStmt; |
| 496 | case Stmt::DoStmtClass: |
| 497 | return PGOHash::DoStmt; |
| 498 | case Stmt::ForStmtClass: |
| 499 | return PGOHash::ForStmt; |
| 500 | case Stmt::CXXForRangeStmtClass: |
| 501 | return PGOHash::CXXForRangeStmt; |
| 502 | case Stmt::ObjCForCollectionStmtClass: |
| 503 | return PGOHash::ObjCForCollectionStmt; |
| 504 | case Stmt::SwitchStmtClass: |
| 505 | return PGOHash::SwitchStmt; |
| 506 | case Stmt::CaseStmtClass: |
| 507 | return PGOHash::CaseStmt; |
| 508 | case Stmt::DefaultStmtClass: |
| 509 | return PGOHash::DefaultStmt; |
| 510 | case Stmt::IfStmtClass: |
| 511 | return PGOHash::IfStmt; |
| 512 | case Stmt::CXXTryStmtClass: |
| 513 | return PGOHash::CXXTryStmt; |
| 514 | case Stmt::CXXCatchStmtClass: |
| 515 | return PGOHash::CXXCatchStmt; |
| 516 | case Stmt::ConditionalOperatorClass: |
| 517 | return PGOHash::ConditionalOperator; |
| 518 | case Stmt::BinaryConditionalOperatorClass: |
| 519 | return PGOHash::BinaryConditionalOperator; |
| 520 | case Stmt::BinaryOperatorClass: { |
| 521 | const BinaryOperator *BO = cast<BinaryOperator>(Val: S); |
| 522 | if (BO->getOpcode() == BO_LAnd) |
| 523 | return PGOHash::BinaryOperatorLAnd; |
| 524 | if (BO->getOpcode() == BO_LOr) |
| 525 | return PGOHash::BinaryOperatorLOr; |
| 526 | if (HashVersion >= PGO_HASH_V2) { |
| 527 | switch (BO->getOpcode()) { |
| 528 | default: |
| 529 | break; |
| 530 | case BO_LT: |
| 531 | return PGOHash::BinaryOperatorLT; |
| 532 | case BO_GT: |
| 533 | return PGOHash::BinaryOperatorGT; |
| 534 | case BO_LE: |
| 535 | return PGOHash::BinaryOperatorLE; |
| 536 | case BO_GE: |
| 537 | return PGOHash::BinaryOperatorGE; |
| 538 | case BO_EQ: |
| 539 | return PGOHash::BinaryOperatorEQ; |
| 540 | case BO_NE: |
| 541 | return PGOHash::BinaryOperatorNE; |
| 542 | } |
| 543 | } |
| 544 | break; |
| 545 | } |
| 546 | } |
| 547 | |
| 548 | if (HashVersion >= PGO_HASH_V2) { |
| 549 | switch (S->getStmtClass()) { |
| 550 | default: |
| 551 | break; |
| 552 | case Stmt::GotoStmtClass: |
| 553 | return PGOHash::GotoStmt; |
| 554 | case Stmt::IndirectGotoStmtClass: |
| 555 | return PGOHash::IndirectGotoStmt; |
| 556 | case Stmt::BreakStmtClass: |
| 557 | return PGOHash::BreakStmt; |
| 558 | case Stmt::ContinueStmtClass: |
| 559 | return PGOHash::ContinueStmt; |
| 560 | case Stmt::ReturnStmtClass: |
| 561 | return PGOHash::ReturnStmt; |
| 562 | case Stmt::CXXThrowExprClass: |
| 563 | return PGOHash::ThrowExpr; |
| 564 | case Stmt::UnaryOperatorClass: { |
| 565 | const UnaryOperator *UO = cast<UnaryOperator>(Val: S); |
| 566 | if (UO->getOpcode() == UO_LNot) |
| 567 | return PGOHash::UnaryOperatorLNot; |
| 568 | break; |
| 569 | } |
| 570 | } |
| 571 | } |
| 572 | |
| 573 | return PGOHash::None; |
| 574 | } |
| 575 | }; |
| 576 | |
| 577 | /// A StmtVisitor that propagates the raw counts through the AST and |
| 578 | /// records the count at statements where the value may change. |
| 579 | struct ComputeRegionCounts : public ConstStmtVisitor<ComputeRegionCounts> { |
| 580 | /// PGO state. |
| 581 | CodeGenPGO &PGO; |
| 582 | |
| 583 | /// A flag that is set when the current count should be recorded on the |
| 584 | /// next statement, such as at the exit of a loop. |
| 585 | bool RecordNextStmtCount; |
| 586 | |
| 587 | /// The count at the current location in the traversal. |
| 588 | uint64_t CurrentCount; |
| 589 | |
| 590 | /// The map of statements to count values. |
| 591 | llvm::DenseMap<const Stmt *, uint64_t> &CountMap; |
| 592 | |
| 593 | /// BreakContinueStack - Keep counts of breaks and continues inside loops. |
| 594 | struct BreakContinue { |
| 595 | uint64_t BreakCount = 0; |
| 596 | uint64_t ContinueCount = 0; |
| 597 | BreakContinue() = default; |
| 598 | }; |
| 599 | SmallVector<BreakContinue, 8> BreakContinueStack; |
| 600 | |
| 601 | ComputeRegionCounts(llvm::DenseMap<const Stmt *, uint64_t> &CountMap, |
| 602 | CodeGenPGO &PGO) |
| 603 | : PGO(PGO), RecordNextStmtCount(false), CountMap(CountMap) {} |
| 604 | |
| 605 | void RecordStmtCount(const Stmt *S) { |
| 606 | if (RecordNextStmtCount) { |
| 607 | CountMap[S] = CurrentCount; |
| 608 | RecordNextStmtCount = false; |
| 609 | } |
| 610 | } |
| 611 | |
| 612 | /// Set and return the current count. |
| 613 | uint64_t setCount(uint64_t Count) { |
| 614 | CurrentCount = Count; |
| 615 | return Count; |
| 616 | } |
| 617 | |
| 618 | void VisitStmt(const Stmt *S) { |
| 619 | RecordStmtCount(S); |
| 620 | for (const Stmt *Child : S->children()) |
| 621 | if (Child) |
| 622 | this->Visit(S: Child); |
| 623 | } |
| 624 | |
| 625 | void VisitFunctionDecl(const FunctionDecl *D) { |
| 626 | // Counter tracks entry to the function body. |
| 627 | uint64_t BodyCount = setCount(PGO.getRegionCount(S: D->getBody())); |
| 628 | CountMap[D->getBody()] = BodyCount; |
| 629 | Visit(S: D->getBody()); |
| 630 | } |
| 631 | |
| 632 | // Skip lambda expressions. We visit these as FunctionDecls when we're |
| 633 | // generating them and aren't interested in the body when generating a |
| 634 | // parent context. |
| 635 | void VisitLambdaExpr(const LambdaExpr *LE) {} |
| 636 | |
| 637 | void VisitCapturedDecl(const CapturedDecl *D) { |
| 638 | // Counter tracks entry to the capture body. |
| 639 | uint64_t BodyCount = setCount(PGO.getRegionCount(S: D->getBody())); |
| 640 | CountMap[D->getBody()] = BodyCount; |
| 641 | Visit(S: D->getBody()); |
| 642 | } |
| 643 | |
| 644 | void VisitObjCMethodDecl(const ObjCMethodDecl *D) { |
| 645 | // Counter tracks entry to the method body. |
| 646 | uint64_t BodyCount = setCount(PGO.getRegionCount(S: D->getBody())); |
| 647 | CountMap[D->getBody()] = BodyCount; |
| 648 | Visit(S: D->getBody()); |
| 649 | } |
| 650 | |
| 651 | void VisitBlockDecl(const BlockDecl *D) { |
| 652 | // Counter tracks entry to the block body. |
| 653 | uint64_t BodyCount = setCount(PGO.getRegionCount(S: D->getBody())); |
| 654 | CountMap[D->getBody()] = BodyCount; |
| 655 | Visit(S: D->getBody()); |
| 656 | } |
| 657 | |
| 658 | void VisitReturnStmt(const ReturnStmt *S) { |
| 659 | RecordStmtCount(S); |
| 660 | if (S->getRetValue()) |
| 661 | Visit(S: S->getRetValue()); |
| 662 | CurrentCount = 0; |
| 663 | RecordNextStmtCount = true; |
| 664 | } |
| 665 | |
| 666 | void VisitCXXThrowExpr(const CXXThrowExpr *E) { |
| 667 | RecordStmtCount(S: E); |
| 668 | if (E->getSubExpr()) |
| 669 | Visit(S: E->getSubExpr()); |
| 670 | CurrentCount = 0; |
| 671 | RecordNextStmtCount = true; |
| 672 | } |
| 673 | |
| 674 | void VisitGotoStmt(const GotoStmt *S) { |
| 675 | RecordStmtCount(S); |
| 676 | CurrentCount = 0; |
| 677 | RecordNextStmtCount = true; |
| 678 | } |
| 679 | |
| 680 | void VisitLabelStmt(const LabelStmt *S) { |
| 681 | RecordNextStmtCount = false; |
| 682 | // Counter tracks the block following the label. |
| 683 | uint64_t BlockCount = setCount(PGO.getRegionCount(S)); |
| 684 | CountMap[S] = BlockCount; |
| 685 | Visit(S: S->getSubStmt()); |
| 686 | } |
| 687 | |
| 688 | void VisitBreakStmt(const BreakStmt *S) { |
| 689 | RecordStmtCount(S); |
| 690 | assert(!BreakContinueStack.empty() && "break not in a loop or switch!" ); |
| 691 | BreakContinueStack.back().BreakCount += CurrentCount; |
| 692 | CurrentCount = 0; |
| 693 | RecordNextStmtCount = true; |
| 694 | } |
| 695 | |
| 696 | void VisitContinueStmt(const ContinueStmt *S) { |
| 697 | RecordStmtCount(S); |
| 698 | assert(!BreakContinueStack.empty() && "continue stmt not in a loop!" ); |
| 699 | BreakContinueStack.back().ContinueCount += CurrentCount; |
| 700 | CurrentCount = 0; |
| 701 | RecordNextStmtCount = true; |
| 702 | } |
| 703 | |
| 704 | void VisitWhileStmt(const WhileStmt *S) { |
| 705 | RecordStmtCount(S); |
| 706 | uint64_t ParentCount = CurrentCount; |
| 707 | |
| 708 | BreakContinueStack.push_back(Elt: BreakContinue()); |
| 709 | // Visit the body region first so the break/continue adjustments can be |
| 710 | // included when visiting the condition. |
| 711 | uint64_t BodyCount = setCount(PGO.getRegionCount(S)); |
| 712 | CountMap[S->getBody()] = CurrentCount; |
| 713 | Visit(S: S->getBody()); |
| 714 | uint64_t BackedgeCount = CurrentCount; |
| 715 | |
| 716 | // ...then go back and propagate counts through the condition. The count |
| 717 | // at the start of the condition is the sum of the incoming edges, |
| 718 | // the backedge from the end of the loop body, and the edges from |
| 719 | // continue statements. |
| 720 | BreakContinue BC = BreakContinueStack.pop_back_val(); |
| 721 | uint64_t CondCount = |
| 722 | setCount(ParentCount + BackedgeCount + BC.ContinueCount); |
| 723 | CountMap[S->getCond()] = CondCount; |
| 724 | Visit(S: S->getCond()); |
| 725 | setCount(BC.BreakCount + CondCount - BodyCount); |
| 726 | RecordNextStmtCount = true; |
| 727 | } |
| 728 | |
| 729 | void VisitDoStmt(const DoStmt *S) { |
| 730 | RecordStmtCount(S); |
| 731 | uint64_t LoopCount = PGO.getRegionCount(S); |
| 732 | |
| 733 | BreakContinueStack.push_back(Elt: BreakContinue()); |
| 734 | // The count doesn't include the fallthrough from the parent scope. Add it. |
| 735 | uint64_t BodyCount = setCount(LoopCount + CurrentCount); |
| 736 | CountMap[S->getBody()] = BodyCount; |
| 737 | Visit(S: S->getBody()); |
| 738 | uint64_t BackedgeCount = CurrentCount; |
| 739 | |
| 740 | BreakContinue BC = BreakContinueStack.pop_back_val(); |
| 741 | // The count at the start of the condition is equal to the count at the |
| 742 | // end of the body, plus any continues. |
| 743 | uint64_t CondCount = setCount(BackedgeCount + BC.ContinueCount); |
| 744 | CountMap[S->getCond()] = CondCount; |
| 745 | Visit(S: S->getCond()); |
| 746 | setCount(BC.BreakCount + CondCount - LoopCount); |
| 747 | RecordNextStmtCount = true; |
| 748 | } |
| 749 | |
| 750 | void VisitForStmt(const ForStmt *S) { |
| 751 | RecordStmtCount(S); |
| 752 | if (S->getInit()) |
| 753 | Visit(S: S->getInit()); |
| 754 | |
| 755 | uint64_t ParentCount = CurrentCount; |
| 756 | |
| 757 | BreakContinueStack.push_back(Elt: BreakContinue()); |
| 758 | // Visit the body region first. (This is basically the same as a while |
| 759 | // loop; see further comments in VisitWhileStmt.) |
| 760 | uint64_t BodyCount = setCount(PGO.getRegionCount(S)); |
| 761 | CountMap[S->getBody()] = BodyCount; |
| 762 | Visit(S: S->getBody()); |
| 763 | uint64_t BackedgeCount = CurrentCount; |
| 764 | BreakContinue BC = BreakContinueStack.pop_back_val(); |
| 765 | |
| 766 | // The increment is essentially part of the body but it needs to include |
| 767 | // the count for all the continue statements. |
| 768 | if (S->getInc()) { |
| 769 | uint64_t IncCount = setCount(BackedgeCount + BC.ContinueCount); |
| 770 | CountMap[S->getInc()] = IncCount; |
| 771 | Visit(S: S->getInc()); |
| 772 | } |
| 773 | |
| 774 | // ...then go back and propagate counts through the condition. |
| 775 | uint64_t CondCount = |
| 776 | setCount(ParentCount + BackedgeCount + BC.ContinueCount); |
| 777 | if (S->getCond()) { |
| 778 | CountMap[S->getCond()] = CondCount; |
| 779 | Visit(S: S->getCond()); |
| 780 | } |
| 781 | setCount(BC.BreakCount + CondCount - BodyCount); |
| 782 | RecordNextStmtCount = true; |
| 783 | } |
| 784 | |
| 785 | void VisitCXXForRangeStmt(const CXXForRangeStmt *S) { |
| 786 | RecordStmtCount(S); |
| 787 | if (S->getInit()) |
| 788 | Visit(S: S->getInit()); |
| 789 | Visit(S: S->getLoopVarStmt()); |
| 790 | Visit(S: S->getRangeStmt()); |
| 791 | Visit(S: S->getBeginStmt()); |
| 792 | Visit(S: S->getEndStmt()); |
| 793 | |
| 794 | uint64_t ParentCount = CurrentCount; |
| 795 | BreakContinueStack.push_back(Elt: BreakContinue()); |
| 796 | // Visit the body region first. (This is basically the same as a while |
| 797 | // loop; see further comments in VisitWhileStmt.) |
| 798 | uint64_t BodyCount = setCount(PGO.getRegionCount(S)); |
| 799 | CountMap[S->getBody()] = BodyCount; |
| 800 | Visit(S: S->getBody()); |
| 801 | uint64_t BackedgeCount = CurrentCount; |
| 802 | BreakContinue BC = BreakContinueStack.pop_back_val(); |
| 803 | |
| 804 | // The increment is essentially part of the body but it needs to include |
| 805 | // the count for all the continue statements. |
| 806 | uint64_t IncCount = setCount(BackedgeCount + BC.ContinueCount); |
| 807 | CountMap[S->getInc()] = IncCount; |
| 808 | Visit(S: S->getInc()); |
| 809 | |
| 810 | // ...then go back and propagate counts through the condition. |
| 811 | uint64_t CondCount = |
| 812 | setCount(ParentCount + BackedgeCount + BC.ContinueCount); |
| 813 | CountMap[S->getCond()] = CondCount; |
| 814 | Visit(S: S->getCond()); |
| 815 | setCount(BC.BreakCount + CondCount - BodyCount); |
| 816 | RecordNextStmtCount = true; |
| 817 | } |
| 818 | |
| 819 | void VisitObjCForCollectionStmt(const ObjCForCollectionStmt *S) { |
| 820 | RecordStmtCount(S); |
| 821 | Visit(S: S->getElement()); |
| 822 | uint64_t ParentCount = CurrentCount; |
| 823 | BreakContinueStack.push_back(Elt: BreakContinue()); |
| 824 | // Counter tracks the body of the loop. |
| 825 | uint64_t BodyCount = setCount(PGO.getRegionCount(S)); |
| 826 | CountMap[S->getBody()] = BodyCount; |
| 827 | Visit(S: S->getBody()); |
| 828 | uint64_t BackedgeCount = CurrentCount; |
| 829 | BreakContinue BC = BreakContinueStack.pop_back_val(); |
| 830 | |
| 831 | setCount(BC.BreakCount + ParentCount + BackedgeCount + BC.ContinueCount - |
| 832 | BodyCount); |
| 833 | RecordNextStmtCount = true; |
| 834 | } |
| 835 | |
| 836 | void VisitSwitchStmt(const SwitchStmt *S) { |
| 837 | RecordStmtCount(S); |
| 838 | if (S->getInit()) |
| 839 | Visit(S: S->getInit()); |
| 840 | Visit(S: S->getCond()); |
| 841 | CurrentCount = 0; |
| 842 | BreakContinueStack.push_back(Elt: BreakContinue()); |
| 843 | Visit(S: S->getBody()); |
| 844 | // If the switch is inside a loop, add the continue counts. |
| 845 | BreakContinue BC = BreakContinueStack.pop_back_val(); |
| 846 | if (!BreakContinueStack.empty()) |
| 847 | BreakContinueStack.back().ContinueCount += BC.ContinueCount; |
| 848 | // Counter tracks the exit block of the switch. |
| 849 | setCount(PGO.getRegionCount(S)); |
| 850 | RecordNextStmtCount = true; |
| 851 | } |
| 852 | |
| 853 | void VisitSwitchCase(const SwitchCase *S) { |
| 854 | RecordNextStmtCount = false; |
| 855 | // Counter for this particular case. This counts only jumps from the |
| 856 | // switch header and does not include fallthrough from the case before |
| 857 | // this one. |
| 858 | uint64_t CaseCount = PGO.getRegionCount(S); |
| 859 | setCount(CurrentCount + CaseCount); |
| 860 | // We need the count without fallthrough in the mapping, so it's more useful |
| 861 | // for branch probabilities. |
| 862 | CountMap[S] = CaseCount; |
| 863 | RecordNextStmtCount = true; |
| 864 | Visit(S: S->getSubStmt()); |
| 865 | } |
| 866 | |
| 867 | void VisitIfStmt(const IfStmt *S) { |
| 868 | RecordStmtCount(S); |
| 869 | |
| 870 | if (S->isConsteval()) { |
| 871 | const Stmt *Stm = S->isNegatedConsteval() ? S->getThen() : S->getElse(); |
| 872 | if (Stm) |
| 873 | Visit(S: Stm); |
| 874 | return; |
| 875 | } |
| 876 | |
| 877 | uint64_t ParentCount = CurrentCount; |
| 878 | if (S->getInit()) |
| 879 | Visit(S: S->getInit()); |
| 880 | Visit(S: S->getCond()); |
| 881 | |
| 882 | // Counter tracks the "then" part of an if statement. The count for |
| 883 | // the "else" part, if it exists, will be calculated from this counter. |
| 884 | uint64_t ThenCount = setCount(PGO.getRegionCount(S)); |
| 885 | CountMap[S->getThen()] = ThenCount; |
| 886 | Visit(S: S->getThen()); |
| 887 | uint64_t OutCount = CurrentCount; |
| 888 | |
| 889 | uint64_t ElseCount = ParentCount - ThenCount; |
| 890 | if (S->getElse()) { |
| 891 | setCount(ElseCount); |
| 892 | CountMap[S->getElse()] = ElseCount; |
| 893 | Visit(S: S->getElse()); |
| 894 | OutCount += CurrentCount; |
| 895 | } else |
| 896 | OutCount += ElseCount; |
| 897 | setCount(OutCount); |
| 898 | RecordNextStmtCount = true; |
| 899 | } |
| 900 | |
| 901 | void VisitCXXTryStmt(const CXXTryStmt *S) { |
| 902 | RecordStmtCount(S); |
| 903 | Visit(S: S->getTryBlock()); |
| 904 | for (unsigned I = 0, E = S->getNumHandlers(); I < E; ++I) |
| 905 | Visit(S: S->getHandler(i: I)); |
| 906 | // Counter tracks the continuation block of the try statement. |
| 907 | setCount(PGO.getRegionCount(S)); |
| 908 | RecordNextStmtCount = true; |
| 909 | } |
| 910 | |
| 911 | void VisitCXXCatchStmt(const CXXCatchStmt *S) { |
| 912 | RecordNextStmtCount = false; |
| 913 | // Counter tracks the catch statement's handler block. |
| 914 | uint64_t CatchCount = setCount(PGO.getRegionCount(S)); |
| 915 | CountMap[S] = CatchCount; |
| 916 | Visit(S: S->getHandlerBlock()); |
| 917 | } |
| 918 | |
| 919 | void VisitAbstractConditionalOperator(const AbstractConditionalOperator *E) { |
| 920 | RecordStmtCount(S: E); |
| 921 | uint64_t ParentCount = CurrentCount; |
| 922 | Visit(S: E->getCond()); |
| 923 | |
| 924 | // Counter tracks the "true" part of a conditional operator. The |
| 925 | // count in the "false" part will be calculated from this counter. |
| 926 | uint64_t TrueCount = setCount(PGO.getRegionCount(S: E)); |
| 927 | CountMap[E->getTrueExpr()] = TrueCount; |
| 928 | Visit(S: E->getTrueExpr()); |
| 929 | uint64_t OutCount = CurrentCount; |
| 930 | |
| 931 | uint64_t FalseCount = setCount(ParentCount - TrueCount); |
| 932 | CountMap[E->getFalseExpr()] = FalseCount; |
| 933 | Visit(S: E->getFalseExpr()); |
| 934 | OutCount += CurrentCount; |
| 935 | |
| 936 | setCount(OutCount); |
| 937 | RecordNextStmtCount = true; |
| 938 | } |
| 939 | |
| 940 | void VisitBinLAnd(const BinaryOperator *E) { |
| 941 | RecordStmtCount(S: E); |
| 942 | uint64_t ParentCount = CurrentCount; |
| 943 | Visit(S: E->getLHS()); |
| 944 | // Counter tracks the right hand side of a logical and operator. |
| 945 | uint64_t RHSCount = setCount(PGO.getRegionCount(S: E)); |
| 946 | CountMap[E->getRHS()] = RHSCount; |
| 947 | Visit(S: E->getRHS()); |
| 948 | setCount(ParentCount + RHSCount - CurrentCount); |
| 949 | RecordNextStmtCount = true; |
| 950 | } |
| 951 | |
| 952 | void VisitBinLOr(const BinaryOperator *E) { |
| 953 | RecordStmtCount(S: E); |
| 954 | uint64_t ParentCount = CurrentCount; |
| 955 | Visit(S: E->getLHS()); |
| 956 | // Counter tracks the right hand side of a logical or operator. |
| 957 | uint64_t RHSCount = setCount(PGO.getRegionCount(S: E)); |
| 958 | CountMap[E->getRHS()] = RHSCount; |
| 959 | Visit(S: E->getRHS()); |
| 960 | setCount(ParentCount + RHSCount - CurrentCount); |
| 961 | RecordNextStmtCount = true; |
| 962 | } |
| 963 | }; |
| 964 | } // end anonymous namespace |
| 965 | |
| 966 | void PGOHash::combine(HashType Type) { |
| 967 | // Check that we never combine 0 and only have six bits. |
| 968 | assert(Type && "Hash is invalid: unexpected type 0" ); |
| 969 | assert(unsigned(Type) < TooBig && "Hash is invalid: too many types" ); |
| 970 | |
| 971 | // Pass through MD5 if enough work has built up. |
| 972 | if (Count && Count % NumTypesPerWord == 0) { |
| 973 | using namespace llvm::support; |
| 974 | uint64_t Swapped = |
| 975 | endian::byte_swap<uint64_t, llvm::endianness::little>(value: Working); |
| 976 | MD5.update(Data: llvm::ArrayRef((uint8_t *)&Swapped, sizeof(Swapped))); |
| 977 | Working = 0; |
| 978 | } |
| 979 | |
| 980 | // Accumulate the current type. |
| 981 | ++Count; |
| 982 | Working = Working << NumBitsPerType | Type; |
| 983 | } |
| 984 | |
| 985 | uint64_t PGOHash::finalize() { |
| 986 | // Use Working as the hash directly if we never used MD5. |
| 987 | if (Count <= NumTypesPerWord) |
| 988 | // No need to byte swap here, since none of the math was endian-dependent. |
| 989 | // This number will be byte-swapped as required on endianness transitions, |
| 990 | // so we will see the same value on the other side. |
| 991 | return Working; |
| 992 | |
| 993 | // Check for remaining work in Working. |
| 994 | if (Working) { |
| 995 | // Keep the buggy behavior from v1 and v2 for backward-compatibility. This |
| 996 | // is buggy because it converts a uint64_t into an array of uint8_t. |
| 997 | if (HashVersion < PGO_HASH_V3) { |
| 998 | MD5.update(Data: {(uint8_t)Working}); |
| 999 | } else { |
| 1000 | using namespace llvm::support; |
| 1001 | uint64_t Swapped = |
| 1002 | endian::byte_swap<uint64_t, llvm::endianness::little>(value: Working); |
| 1003 | MD5.update(Data: llvm::ArrayRef((uint8_t *)&Swapped, sizeof(Swapped))); |
| 1004 | } |
| 1005 | } |
| 1006 | |
| 1007 | // Finalize the MD5 and return the hash. |
| 1008 | llvm::MD5::MD5Result Result; |
| 1009 | MD5.final(Result); |
| 1010 | return Result.low(); |
| 1011 | } |
| 1012 | |
| 1013 | void CodeGenPGO::assignRegionCounters(GlobalDecl GD, llvm::Function *Fn) { |
| 1014 | const Decl *D = GD.getDecl(); |
| 1015 | if (!D->hasBody()) |
| 1016 | return; |
| 1017 | |
| 1018 | // Skip CUDA/HIP kernel launch stub functions. |
| 1019 | if (CGM.getLangOpts().CUDA && !CGM.getLangOpts().CUDAIsDevice && |
| 1020 | D->hasAttr<CUDAGlobalAttr>()) |
| 1021 | return; |
| 1022 | |
| 1023 | bool InstrumentRegions = CGM.getCodeGenOpts().hasProfileClangInstr(); |
| 1024 | llvm::IndexedInstrProfReader *PGOReader = CGM.getPGOReader(); |
| 1025 | if (!InstrumentRegions && !PGOReader) |
| 1026 | return; |
| 1027 | if (D->isImplicit()) |
| 1028 | return; |
| 1029 | // Constructors and destructors may be represented by several functions in IR. |
| 1030 | // If so, instrument only base variant, others are implemented by delegation |
| 1031 | // to the base one, it would be counted twice otherwise. |
| 1032 | if (CGM.getTarget().getCXXABI().hasConstructorVariants()) { |
| 1033 | if (const auto *CCD = dyn_cast<CXXConstructorDecl>(Val: D)) |
| 1034 | if (GD.getCtorType() != Ctor_Base && |
| 1035 | CodeGenFunction::IsConstructorDelegationValid(Ctor: CCD)) |
| 1036 | return; |
| 1037 | } |
| 1038 | if (isa<CXXDestructorDecl>(Val: D) && GD.getDtorType() != Dtor_Base) |
| 1039 | return; |
| 1040 | |
| 1041 | CGM.ClearUnusedCoverageMapping(D); |
| 1042 | if (Fn->hasFnAttribute(Kind: llvm::Attribute::NoProfile)) |
| 1043 | return; |
| 1044 | if (Fn->hasFnAttribute(Kind: llvm::Attribute::SkipProfile)) |
| 1045 | return; |
| 1046 | |
| 1047 | SourceManager &SM = CGM.getContext().getSourceManager(); |
| 1048 | if (!llvm::coverage::SystemHeadersCoverage && |
| 1049 | SM.isInSystemHeader(Loc: D->getLocation())) |
| 1050 | return; |
| 1051 | |
| 1052 | setFuncName(Fn); |
| 1053 | |
| 1054 | mapRegionCounters(D); |
| 1055 | if (CGM.getCodeGenOpts().CoverageMapping) |
| 1056 | emitCounterRegionMapping(D); |
| 1057 | if (PGOReader) { |
| 1058 | loadRegionCounts(PGOReader, IsInMainFile: SM.isInMainFile(Loc: D->getLocation())); |
| 1059 | computeRegionCounts(D); |
| 1060 | applyFunctionAttributes(PGOReader, Fn); |
| 1061 | } |
| 1062 | } |
| 1063 | |
| 1064 | void CodeGenPGO::mapRegionCounters(const Decl *D) { |
| 1065 | // Use the latest hash version when inserting instrumentation, but use the |
| 1066 | // version in the indexed profile if we're reading PGO data. |
| 1067 | PGOHashVersion HashVersion = PGO_HASH_LATEST; |
| 1068 | uint64_t ProfileVersion = llvm::IndexedInstrProf::Version; |
| 1069 | if (auto *PGOReader = CGM.getPGOReader()) { |
| 1070 | HashVersion = getPGOHashVersion(PGOReader, CGM); |
| 1071 | ProfileVersion = PGOReader->getVersion(); |
| 1072 | } |
| 1073 | |
| 1074 | // If MC/DC is enabled, set the MaxConditions to a preset value. Otherwise, |
| 1075 | // set it to zero. This value impacts the number of conditions accepted in a |
| 1076 | // given boolean expression, which impacts the size of the bitmap used to |
| 1077 | // track test vector execution for that boolean expression. Because the |
| 1078 | // bitmap scales exponentially (2^n) based on the number of conditions seen, |
| 1079 | // the maximum value is hard-coded at 6 conditions, which is more than enough |
| 1080 | // for most embedded applications. Setting a maximum value prevents the |
| 1081 | // bitmap footprint from growing too large without the user's knowledge. In |
| 1082 | // the future, this value could be adjusted with a command-line option. |
| 1083 | unsigned MCDCMaxConditions = |
| 1084 | (CGM.getCodeGenOpts().MCDCCoverage ? CGM.getCodeGenOpts().MCDCMaxConds |
| 1085 | : 0); |
| 1086 | |
| 1087 | RegionCounterMap.reset(p: new llvm::DenseMap<const Stmt *, CounterPair>); |
| 1088 | RegionMCDCState.reset(p: new MCDC::State); |
| 1089 | MapRegionCounters Walker(HashVersion, ProfileVersion, *RegionCounterMap, |
| 1090 | *RegionMCDCState, MCDCMaxConditions, CGM.getDiags()); |
| 1091 | if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: D)) |
| 1092 | Walker.TraverseDecl(D: const_cast<FunctionDecl *>(FD)); |
| 1093 | else if (const ObjCMethodDecl *MD = dyn_cast_or_null<ObjCMethodDecl>(Val: D)) |
| 1094 | Walker.TraverseDecl(D: const_cast<ObjCMethodDecl *>(MD)); |
| 1095 | else if (const BlockDecl *BD = dyn_cast_or_null<BlockDecl>(Val: D)) |
| 1096 | Walker.TraverseDecl(D: const_cast<BlockDecl *>(BD)); |
| 1097 | else if (const CapturedDecl *CD = dyn_cast_or_null<CapturedDecl>(Val: D)) |
| 1098 | Walker.TraverseDecl(D: const_cast<CapturedDecl *>(CD)); |
| 1099 | assert(Walker.NextCounter > 0 && "no entry counter mapped for decl" ); |
| 1100 | NumRegionCounters = Walker.NextCounter; |
| 1101 | FunctionHash = Walker.Hash.finalize(); |
| 1102 | } |
| 1103 | |
| 1104 | bool CodeGenPGO::skipRegionMappingForDecl(const Decl *D) { |
| 1105 | if (!D->getBody()) |
| 1106 | return true; |
| 1107 | |
| 1108 | // Skip host-only functions in the CUDA device compilation and device-only |
| 1109 | // functions in the host compilation. Just roughly filter them out based on |
| 1110 | // the function attributes. If there are effectively host-only or device-only |
| 1111 | // ones, their coverage mapping may still be generated. |
| 1112 | if (CGM.getLangOpts().CUDA && |
| 1113 | ((CGM.getLangOpts().CUDAIsDevice && !D->hasAttr<CUDADeviceAttr>() && |
| 1114 | !D->hasAttr<CUDAGlobalAttr>()) || |
| 1115 | (!CGM.getLangOpts().CUDAIsDevice && |
| 1116 | (D->hasAttr<CUDAGlobalAttr>() || |
| 1117 | (!D->hasAttr<CUDAHostAttr>() && D->hasAttr<CUDADeviceAttr>()))))) |
| 1118 | return true; |
| 1119 | |
| 1120 | // Don't map the functions in system headers. |
| 1121 | const auto &SM = CGM.getContext().getSourceManager(); |
| 1122 | auto Loc = D->getBody()->getBeginLoc(); |
| 1123 | return !llvm::coverage::SystemHeadersCoverage && SM.isInSystemHeader(Loc); |
| 1124 | } |
| 1125 | |
| 1126 | void CodeGenPGO::emitCounterRegionMapping(const Decl *D) { |
| 1127 | if (skipRegionMappingForDecl(D)) |
| 1128 | return; |
| 1129 | |
| 1130 | std::string CoverageMapping; |
| 1131 | llvm::raw_string_ostream OS(CoverageMapping); |
| 1132 | RegionMCDCState->BranchByStmt.clear(); |
| 1133 | CoverageMappingGen MappingGen( |
| 1134 | *CGM.getCoverageMapping(), CGM.getContext().getSourceManager(), |
| 1135 | CGM.getLangOpts(), RegionCounterMap.get(), RegionMCDCState.get()); |
| 1136 | MappingGen.emitCounterMapping(D, OS); |
| 1137 | |
| 1138 | if (CoverageMapping.empty()) |
| 1139 | return; |
| 1140 | |
| 1141 | CGM.getCoverageMapping()->addFunctionMappingRecord( |
| 1142 | FunctionName: FuncNameVar, FunctionNameValue: FuncName, FunctionHash, CoverageMapping); |
| 1143 | } |
| 1144 | |
| 1145 | void |
| 1146 | CodeGenPGO::emitEmptyCounterMapping(const Decl *D, StringRef Name, |
| 1147 | llvm::GlobalValue::LinkageTypes Linkage) { |
| 1148 | if (skipRegionMappingForDecl(D)) |
| 1149 | return; |
| 1150 | |
| 1151 | std::string CoverageMapping; |
| 1152 | llvm::raw_string_ostream OS(CoverageMapping); |
| 1153 | CoverageMappingGen MappingGen(*CGM.getCoverageMapping(), |
| 1154 | CGM.getContext().getSourceManager(), |
| 1155 | CGM.getLangOpts()); |
| 1156 | MappingGen.emitEmptyMapping(D, OS); |
| 1157 | |
| 1158 | if (CoverageMapping.empty()) |
| 1159 | return; |
| 1160 | |
| 1161 | setFuncName(Name, Linkage); |
| 1162 | CGM.getCoverageMapping()->addFunctionMappingRecord( |
| 1163 | FunctionName: FuncNameVar, FunctionNameValue: FuncName, FunctionHash, CoverageMapping, IsUsed: false); |
| 1164 | } |
| 1165 | |
| 1166 | void CodeGenPGO::computeRegionCounts(const Decl *D) { |
| 1167 | StmtCountMap.reset(p: new llvm::DenseMap<const Stmt *, uint64_t>); |
| 1168 | ComputeRegionCounts Walker(*StmtCountMap, *this); |
| 1169 | if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: D)) |
| 1170 | Walker.VisitFunctionDecl(D: FD); |
| 1171 | else if (const ObjCMethodDecl *MD = dyn_cast_or_null<ObjCMethodDecl>(Val: D)) |
| 1172 | Walker.VisitObjCMethodDecl(D: MD); |
| 1173 | else if (const BlockDecl *BD = dyn_cast_or_null<BlockDecl>(Val: D)) |
| 1174 | Walker.VisitBlockDecl(D: BD); |
| 1175 | else if (const CapturedDecl *CD = dyn_cast_or_null<CapturedDecl>(Val: D)) |
| 1176 | Walker.VisitCapturedDecl(D: const_cast<CapturedDecl *>(CD)); |
| 1177 | } |
| 1178 | |
| 1179 | void |
| 1180 | CodeGenPGO::applyFunctionAttributes(llvm::IndexedInstrProfReader *PGOReader, |
| 1181 | llvm::Function *Fn) { |
| 1182 | if (!haveRegionCounts()) |
| 1183 | return; |
| 1184 | |
| 1185 | uint64_t FunctionCount = getRegionCount(S: nullptr); |
| 1186 | Fn->setEntryCount(Count: FunctionCount); |
| 1187 | } |
| 1188 | |
| 1189 | std::pair<bool, bool> CodeGenPGO::getIsCounterPair(const Stmt *S) const { |
| 1190 | if (!RegionCounterMap) |
| 1191 | return {false, false}; |
| 1192 | |
| 1193 | auto I = RegionCounterMap->find(Val: S); |
| 1194 | if (I == RegionCounterMap->end()) |
| 1195 | return {false, false}; |
| 1196 | |
| 1197 | return {I->second.Executed.hasValue(), I->second.Skipped.hasValue()}; |
| 1198 | } |
| 1199 | |
| 1200 | void CodeGenPGO::emitCounterSetOrIncrement(CGBuilderTy &Builder, const Stmt *S, |
| 1201 | llvm::Value *StepV) { |
| 1202 | if (!RegionCounterMap || !Builder.GetInsertBlock()) |
| 1203 | return; |
| 1204 | |
| 1205 | unsigned Counter = (*RegionCounterMap)[S].Executed; |
| 1206 | |
| 1207 | // Make sure that pointer to global is passed in with zero addrspace |
| 1208 | // This is relevant during GPU profiling |
| 1209 | auto *NormalizedFuncNameVarPtr = |
| 1210 | llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast( |
| 1211 | C: FuncNameVar, Ty: llvm::PointerType::get(C&: CGM.getLLVMContext(), AddressSpace: 0)); |
| 1212 | |
| 1213 | llvm::Value *Args[] = { |
| 1214 | NormalizedFuncNameVarPtr, Builder.getInt64(C: FunctionHash), |
| 1215 | Builder.getInt32(C: NumRegionCounters), Builder.getInt32(C: Counter), StepV}; |
| 1216 | |
| 1217 | if (llvm::EnableSingleByteCoverage) |
| 1218 | Builder.CreateCall(Callee: CGM.getIntrinsic(IID: llvm::Intrinsic::instrprof_cover), |
| 1219 | Args: ArrayRef(Args, 4)); |
| 1220 | else if (!StepV) |
| 1221 | Builder.CreateCall(Callee: CGM.getIntrinsic(IID: llvm::Intrinsic::instrprof_increment), |
| 1222 | Args: ArrayRef(Args, 4)); |
| 1223 | else |
| 1224 | Builder.CreateCall( |
| 1225 | Callee: CGM.getIntrinsic(IID: llvm::Intrinsic::instrprof_increment_step), Args); |
| 1226 | } |
| 1227 | |
| 1228 | bool CodeGenPGO::canEmitMCDCCoverage(const CGBuilderTy &Builder) { |
| 1229 | return (CGM.getCodeGenOpts().hasProfileClangInstr() && |
| 1230 | CGM.getCodeGenOpts().MCDCCoverage && Builder.GetInsertBlock()); |
| 1231 | } |
| 1232 | |
| 1233 | void CodeGenPGO::emitMCDCParameters(CGBuilderTy &Builder) { |
| 1234 | if (!canEmitMCDCCoverage(Builder) || !RegionMCDCState) |
| 1235 | return; |
| 1236 | |
| 1237 | auto *I8PtrTy = llvm::PointerType::getUnqual(C&: CGM.getLLVMContext()); |
| 1238 | |
| 1239 | // Emit intrinsic representing MCDC bitmap parameters at function entry. |
| 1240 | // This is used by the instrumentation pass, but it isn't actually lowered to |
| 1241 | // anything. |
| 1242 | llvm::Value *Args[3] = {llvm::ConstantExpr::getBitCast(C: FuncNameVar, Ty: I8PtrTy), |
| 1243 | Builder.getInt64(C: FunctionHash), |
| 1244 | Builder.getInt32(C: RegionMCDCState->BitmapBits)}; |
| 1245 | Builder.CreateCall( |
| 1246 | Callee: CGM.getIntrinsic(IID: llvm::Intrinsic::instrprof_mcdc_parameters), Args); |
| 1247 | } |
| 1248 | |
| 1249 | void CodeGenPGO::emitMCDCTestVectorBitmapUpdate(CGBuilderTy &Builder, |
| 1250 | const Expr *S, |
| 1251 | Address MCDCCondBitmapAddr, |
| 1252 | CodeGenFunction &CGF) { |
| 1253 | if (!canEmitMCDCCoverage(Builder) || !RegionMCDCState) |
| 1254 | return; |
| 1255 | |
| 1256 | S = S->IgnoreParens(); |
| 1257 | |
| 1258 | auto DecisionStateIter = RegionMCDCState->DecisionByStmt.find(Val: S); |
| 1259 | if (DecisionStateIter == RegionMCDCState->DecisionByStmt.end()) |
| 1260 | return; |
| 1261 | |
| 1262 | // Don't create tvbitmap_update if the record is allocated but excluded. |
| 1263 | // Or `bitmap |= (1 << 0)` would be wrongly executed to the next bitmap. |
| 1264 | if (DecisionStateIter->second.Indices.size() == 0) |
| 1265 | return; |
| 1266 | |
| 1267 | // Extract the offset of the global bitmap associated with this expression. |
| 1268 | unsigned MCDCTestVectorBitmapOffset = DecisionStateIter->second.BitmapIdx; |
| 1269 | auto *I8PtrTy = llvm::PointerType::getUnqual(C&: CGM.getLLVMContext()); |
| 1270 | |
| 1271 | // Emit intrinsic responsible for updating the global bitmap corresponding to |
| 1272 | // a boolean expression. The index being set is based on the value loaded |
| 1273 | // from a pointer to a dedicated temporary value on the stack that is itself |
| 1274 | // updated via emitMCDCCondBitmapReset() and emitMCDCCondBitmapUpdate(). The |
| 1275 | // index represents an executed test vector. |
| 1276 | llvm::Value *Args[4] = {llvm::ConstantExpr::getBitCast(C: FuncNameVar, Ty: I8PtrTy), |
| 1277 | Builder.getInt64(C: FunctionHash), |
| 1278 | Builder.getInt32(C: MCDCTestVectorBitmapOffset), |
| 1279 | MCDCCondBitmapAddr.emitRawPointer(CGF)}; |
| 1280 | Builder.CreateCall( |
| 1281 | Callee: CGM.getIntrinsic(IID: llvm::Intrinsic::instrprof_mcdc_tvbitmap_update), Args); |
| 1282 | } |
| 1283 | |
| 1284 | void CodeGenPGO::emitMCDCCondBitmapReset(CGBuilderTy &Builder, const Expr *S, |
| 1285 | Address MCDCCondBitmapAddr) { |
| 1286 | if (!canEmitMCDCCoverage(Builder) || !RegionMCDCState) |
| 1287 | return; |
| 1288 | |
| 1289 | S = S->IgnoreParens(); |
| 1290 | |
| 1291 | if (!RegionMCDCState->DecisionByStmt.contains(Val: S)) |
| 1292 | return; |
| 1293 | |
| 1294 | // Emit intrinsic that resets a dedicated temporary value on the stack to 0. |
| 1295 | Builder.CreateStore(Val: Builder.getInt32(C: 0), Addr: MCDCCondBitmapAddr); |
| 1296 | } |
| 1297 | |
| 1298 | void CodeGenPGO::emitMCDCCondBitmapUpdate(CGBuilderTy &Builder, const Expr *S, |
| 1299 | Address MCDCCondBitmapAddr, |
| 1300 | llvm::Value *Val, |
| 1301 | CodeGenFunction &CGF) { |
| 1302 | if (!canEmitMCDCCoverage(Builder) || !RegionMCDCState) |
| 1303 | return; |
| 1304 | |
| 1305 | // Even though, for simplicity, parentheses and unary logical-NOT operators |
| 1306 | // are considered part of their underlying condition for both MC/DC and |
| 1307 | // branch coverage, the condition IDs themselves are assigned and tracked |
| 1308 | // using the underlying condition itself. This is done solely for |
| 1309 | // consistency since parentheses and logical-NOTs are ignored when checking |
| 1310 | // whether the condition is actually an instrumentable condition. This can |
| 1311 | // also make debugging a bit easier. |
| 1312 | S = CodeGenFunction::stripCond(C: S); |
| 1313 | |
| 1314 | auto BranchStateIter = RegionMCDCState->BranchByStmt.find(Val: S); |
| 1315 | if (BranchStateIter == RegionMCDCState->BranchByStmt.end()) |
| 1316 | return; |
| 1317 | |
| 1318 | // Extract the ID of the condition we are setting in the bitmap. |
| 1319 | const auto &Branch = BranchStateIter->second; |
| 1320 | assert(Branch.ID >= 0 && "Condition has no ID!" ); |
| 1321 | assert(Branch.DecisionStmt); |
| 1322 | |
| 1323 | // Cancel the emission if the Decision is erased after the allocation. |
| 1324 | const auto DecisionIter = |
| 1325 | RegionMCDCState->DecisionByStmt.find(Val: Branch.DecisionStmt); |
| 1326 | if (DecisionIter == RegionMCDCState->DecisionByStmt.end()) |
| 1327 | return; |
| 1328 | |
| 1329 | const auto &TVIdxs = DecisionIter->second.Indices[Branch.ID]; |
| 1330 | |
| 1331 | auto *CurTV = Builder.CreateLoad(Addr: MCDCCondBitmapAddr, |
| 1332 | Name: "mcdc." + Twine(Branch.ID + 1) + ".cur" ); |
| 1333 | auto *NewTV = Builder.CreateAdd(LHS: CurTV, RHS: Builder.getInt32(C: TVIdxs[true])); |
| 1334 | NewTV = Builder.CreateSelect( |
| 1335 | C: Val, True: NewTV, False: Builder.CreateAdd(LHS: CurTV, RHS: Builder.getInt32(C: TVIdxs[false]))); |
| 1336 | Builder.CreateStore(Val: NewTV, Addr: MCDCCondBitmapAddr); |
| 1337 | } |
| 1338 | |
| 1339 | void CodeGenPGO::setValueProfilingFlag(llvm::Module &M) { |
| 1340 | if (CGM.getCodeGenOpts().hasProfileClangInstr()) |
| 1341 | M.addModuleFlag(Behavior: llvm::Module::Warning, Key: "EnableValueProfiling" , |
| 1342 | Val: uint32_t(EnableValueProfiling)); |
| 1343 | } |
| 1344 | |
| 1345 | void CodeGenPGO::setProfileVersion(llvm::Module &M) { |
| 1346 | if (CGM.getCodeGenOpts().hasProfileClangInstr() && |
| 1347 | llvm::EnableSingleByteCoverage) { |
| 1348 | const StringRef VarName(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR)); |
| 1349 | llvm::Type *IntTy64 = llvm::Type::getInt64Ty(C&: M.getContext()); |
| 1350 | uint64_t ProfileVersion = |
| 1351 | (INSTR_PROF_RAW_VERSION | VARIANT_MASK_BYTE_COVERAGE); |
| 1352 | |
| 1353 | auto IRLevelVersionVariable = new llvm::GlobalVariable( |
| 1354 | M, IntTy64, true, llvm::GlobalValue::WeakAnyLinkage, |
| 1355 | llvm::Constant::getIntegerValue(Ty: IntTy64, |
| 1356 | V: llvm::APInt(64, ProfileVersion)), |
| 1357 | VarName); |
| 1358 | |
| 1359 | IRLevelVersionVariable->setVisibility(llvm::GlobalValue::HiddenVisibility); |
| 1360 | llvm::Triple TT(M.getTargetTriple()); |
| 1361 | if (TT.isGPU()) |
| 1362 | IRLevelVersionVariable->setVisibility( |
| 1363 | llvm::GlobalValue::ProtectedVisibility); |
| 1364 | if (TT.supportsCOMDAT()) { |
| 1365 | IRLevelVersionVariable->setLinkage(llvm::GlobalValue::ExternalLinkage); |
| 1366 | IRLevelVersionVariable->setComdat(M.getOrInsertComdat(Name: VarName)); |
| 1367 | } |
| 1368 | IRLevelVersionVariable->setDSOLocal(true); |
| 1369 | } |
| 1370 | } |
| 1371 | |
| 1372 | // This method either inserts a call to the profile run-time during |
| 1373 | // instrumentation or puts profile data into metadata for PGO use. |
| 1374 | void CodeGenPGO::valueProfile(CGBuilderTy &Builder, uint32_t ValueKind, |
| 1375 | llvm::Instruction *ValueSite, llvm::Value *ValuePtr) { |
| 1376 | |
| 1377 | if (!EnableValueProfiling) |
| 1378 | return; |
| 1379 | |
| 1380 | if (!ValuePtr || !ValueSite || !Builder.GetInsertBlock()) |
| 1381 | return; |
| 1382 | |
| 1383 | if (isa<llvm::Constant>(Val: ValuePtr)) |
| 1384 | return; |
| 1385 | |
| 1386 | bool InstrumentValueSites = CGM.getCodeGenOpts().hasProfileClangInstr(); |
| 1387 | if (InstrumentValueSites && RegionCounterMap) { |
| 1388 | auto BuilderInsertPoint = Builder.saveIP(); |
| 1389 | Builder.SetInsertPoint(ValueSite); |
| 1390 | llvm::Value *Args[5] = { |
| 1391 | FuncNameVar, |
| 1392 | Builder.getInt64(C: FunctionHash), |
| 1393 | Builder.CreatePtrToInt(V: ValuePtr, DestTy: Builder.getInt64Ty()), |
| 1394 | Builder.getInt32(C: ValueKind), |
| 1395 | Builder.getInt32(C: NumValueSites[ValueKind]++) |
| 1396 | }; |
| 1397 | Builder.CreateCall( |
| 1398 | Callee: CGM.getIntrinsic(IID: llvm::Intrinsic::instrprof_value_profile), Args); |
| 1399 | Builder.restoreIP(IP: BuilderInsertPoint); |
| 1400 | return; |
| 1401 | } |
| 1402 | |
| 1403 | llvm::IndexedInstrProfReader *PGOReader = CGM.getPGOReader(); |
| 1404 | if (PGOReader && haveRegionCounts()) { |
| 1405 | // We record the top most called three functions at each call site. |
| 1406 | // Profile metadata contains "VP" string identifying this metadata |
| 1407 | // as value profiling data, then a uint32_t value for the value profiling |
| 1408 | // kind, a uint64_t value for the total number of times the call is |
| 1409 | // executed, followed by the function hash and execution count (uint64_t) |
| 1410 | // pairs for each function. |
| 1411 | if (NumValueSites[ValueKind] >= ProfRecord->getNumValueSites(ValueKind)) |
| 1412 | return; |
| 1413 | |
| 1414 | llvm::annotateValueSite(M&: CGM.getModule(), Inst&: *ValueSite, InstrProfR: *ProfRecord, |
| 1415 | ValueKind: (llvm::InstrProfValueKind)ValueKind, |
| 1416 | SiteIndx: NumValueSites[ValueKind]); |
| 1417 | |
| 1418 | NumValueSites[ValueKind]++; |
| 1419 | } |
| 1420 | } |
| 1421 | |
| 1422 | void CodeGenPGO::loadRegionCounts(llvm::IndexedInstrProfReader *PGOReader, |
| 1423 | bool IsInMainFile) { |
| 1424 | CGM.getPGOStats().addVisited(MainFile: IsInMainFile); |
| 1425 | RegionCounts.clear(); |
| 1426 | auto RecordExpected = PGOReader->getInstrProfRecord(FuncName, FuncHash: FunctionHash); |
| 1427 | if (auto E = RecordExpected.takeError()) { |
| 1428 | auto IPE = std::get<0>(in: llvm::InstrProfError::take(E: std::move(E))); |
| 1429 | if (IPE == llvm::instrprof_error::unknown_function) |
| 1430 | CGM.getPGOStats().addMissing(MainFile: IsInMainFile); |
| 1431 | else if (IPE == llvm::instrprof_error::hash_mismatch) |
| 1432 | CGM.getPGOStats().addMismatched(MainFile: IsInMainFile); |
| 1433 | else if (IPE == llvm::instrprof_error::malformed) |
| 1434 | // TODO: Consider a more specific warning for this case. |
| 1435 | CGM.getPGOStats().addMismatched(MainFile: IsInMainFile); |
| 1436 | return; |
| 1437 | } |
| 1438 | ProfRecord = |
| 1439 | std::make_unique<llvm::InstrProfRecord>(args: std::move(RecordExpected.get())); |
| 1440 | RegionCounts = ProfRecord->Counts; |
| 1441 | } |
| 1442 | |
| 1443 | /// Calculate what to divide by to scale weights. |
| 1444 | /// |
| 1445 | /// Given the maximum weight, calculate a divisor that will scale all the |
| 1446 | /// weights to strictly less than UINT32_MAX. |
| 1447 | static uint64_t calculateWeightScale(uint64_t MaxWeight) { |
| 1448 | return MaxWeight < UINT32_MAX ? 1 : MaxWeight / UINT32_MAX + 1; |
| 1449 | } |
| 1450 | |
| 1451 | /// Scale an individual branch weight (and add 1). |
| 1452 | /// |
| 1453 | /// Scale a 64-bit weight down to 32-bits using \c Scale. |
| 1454 | /// |
| 1455 | /// According to Laplace's Rule of Succession, it is better to compute the |
| 1456 | /// weight based on the count plus 1, so universally add 1 to the value. |
| 1457 | /// |
| 1458 | /// \pre \c Scale was calculated by \a calculateWeightScale() with a weight no |
| 1459 | /// greater than \c Weight. |
| 1460 | static uint32_t scaleBranchWeight(uint64_t Weight, uint64_t Scale) { |
| 1461 | assert(Scale && "scale by 0?" ); |
| 1462 | uint64_t Scaled = Weight / Scale + 1; |
| 1463 | assert(Scaled <= UINT32_MAX && "overflow 32-bits" ); |
| 1464 | return Scaled; |
| 1465 | } |
| 1466 | |
| 1467 | llvm::MDNode *CodeGenFunction::createProfileWeights(uint64_t TrueCount, |
| 1468 | uint64_t FalseCount) const { |
| 1469 | // Check for empty weights. |
| 1470 | if (!TrueCount && !FalseCount) |
| 1471 | return nullptr; |
| 1472 | |
| 1473 | // Calculate how to scale down to 32-bits. |
| 1474 | uint64_t Scale = calculateWeightScale(MaxWeight: std::max(a: TrueCount, b: FalseCount)); |
| 1475 | |
| 1476 | llvm::MDBuilder MDHelper(CGM.getLLVMContext()); |
| 1477 | return MDHelper.createBranchWeights(TrueWeight: scaleBranchWeight(Weight: TrueCount, Scale), |
| 1478 | FalseWeight: scaleBranchWeight(Weight: FalseCount, Scale)); |
| 1479 | } |
| 1480 | |
| 1481 | llvm::MDNode * |
| 1482 | CodeGenFunction::createProfileWeights(ArrayRef<uint64_t> Weights) const { |
| 1483 | // We need at least two elements to create meaningful weights. |
| 1484 | if (Weights.size() < 2) |
| 1485 | return nullptr; |
| 1486 | |
| 1487 | // Check for empty weights. |
| 1488 | uint64_t MaxWeight = *llvm::max_element(Range&: Weights); |
| 1489 | if (MaxWeight == 0) |
| 1490 | return nullptr; |
| 1491 | |
| 1492 | // Calculate how to scale down to 32-bits. |
| 1493 | uint64_t Scale = calculateWeightScale(MaxWeight); |
| 1494 | |
| 1495 | SmallVector<uint32_t, 16> ScaledWeights; |
| 1496 | ScaledWeights.reserve(N: Weights.size()); |
| 1497 | for (uint64_t W : Weights) |
| 1498 | ScaledWeights.push_back(Elt: scaleBranchWeight(Weight: W, Scale)); |
| 1499 | |
| 1500 | llvm::MDBuilder MDHelper(CGM.getLLVMContext()); |
| 1501 | return MDHelper.createBranchWeights(Weights: ScaledWeights); |
| 1502 | } |
| 1503 | |
| 1504 | llvm::MDNode * |
| 1505 | CodeGenFunction::createProfileWeightsForLoop(const Stmt *Cond, |
| 1506 | uint64_t LoopCount) const { |
| 1507 | if (!PGO->haveRegionCounts()) |
| 1508 | return nullptr; |
| 1509 | std::optional<uint64_t> CondCount = PGO->getStmtCount(S: Cond); |
| 1510 | if (!CondCount || *CondCount == 0) |
| 1511 | return nullptr; |
| 1512 | return createProfileWeights(TrueCount: LoopCount, |
| 1513 | FalseCount: std::max(a: *CondCount, b: LoopCount) - LoopCount); |
| 1514 | } |
| 1515 | |
| 1516 | void CodeGenFunction::incrementProfileCounter(const Stmt *S, |
| 1517 | llvm::Value *StepV) { |
| 1518 | if (CGM.getCodeGenOpts().hasProfileClangInstr() && |
| 1519 | !CurFn->hasFnAttribute(Kind: llvm::Attribute::NoProfile) && |
| 1520 | !CurFn->hasFnAttribute(Kind: llvm::Attribute::SkipProfile)) { |
| 1521 | auto AL = ApplyDebugLocation::CreateArtificial(CGF&: *this); |
| 1522 | PGO->emitCounterSetOrIncrement(Builder, S, StepV); |
| 1523 | } |
| 1524 | PGO->setCurrentStmt(S); |
| 1525 | } |
| 1526 | |
| 1527 | std::pair<bool, bool> CodeGenFunction::getIsCounterPair(const Stmt *S) const { |
| 1528 | return PGO->getIsCounterPair(S); |
| 1529 | } |
| 1530 | void CodeGenFunction::markStmtAsUsed(bool Skipped, const Stmt *S) { |
| 1531 | PGO->markStmtAsUsed(Skipped, S); |
| 1532 | } |
| 1533 | void CodeGenFunction::markStmtMaybeUsed(const Stmt *S) { |
| 1534 | PGO->markStmtMaybeUsed(S); |
| 1535 | } |
| 1536 | |
| 1537 | void CodeGenFunction::maybeCreateMCDCCondBitmap() { |
| 1538 | if (isMCDCCoverageEnabled()) { |
| 1539 | PGO->emitMCDCParameters(Builder); |
| 1540 | MCDCCondBitmapAddr = CreateIRTemp(T: getContext().UnsignedIntTy, Name: "mcdc.addr" ); |
| 1541 | } |
| 1542 | } |
| 1543 | void CodeGenFunction::maybeResetMCDCCondBitmap(const Expr *E) { |
| 1544 | if (isMCDCCoverageEnabled() && isBinaryLogicalOp(E)) { |
| 1545 | PGO->emitMCDCCondBitmapReset(Builder, S: E, MCDCCondBitmapAddr); |
| 1546 | PGO->setCurrentStmt(E); |
| 1547 | } |
| 1548 | } |
| 1549 | void CodeGenFunction::maybeUpdateMCDCTestVectorBitmap(const Expr *E) { |
| 1550 | if (isMCDCCoverageEnabled() && isBinaryLogicalOp(E)) { |
| 1551 | PGO->emitMCDCTestVectorBitmapUpdate(Builder, S: E, MCDCCondBitmapAddr, CGF&: *this); |
| 1552 | PGO->setCurrentStmt(E); |
| 1553 | } |
| 1554 | } |
| 1555 | |
| 1556 | void CodeGenFunction::maybeUpdateMCDCCondBitmap(const Expr *E, |
| 1557 | llvm::Value *Val) { |
| 1558 | if (isMCDCCoverageEnabled()) { |
| 1559 | PGO->emitMCDCCondBitmapUpdate(Builder, S: E, MCDCCondBitmapAddr, Val, CGF&: *this); |
| 1560 | PGO->setCurrentStmt(E); |
| 1561 | } |
| 1562 | } |
| 1563 | |
| 1564 | uint64_t CodeGenFunction::getProfileCount(const Stmt *S) { |
| 1565 | return PGO->getStmtCount(S).value_or(u: 0); |
| 1566 | } |
| 1567 | |
| 1568 | /// Set the profiler's current count. |
| 1569 | void CodeGenFunction::setCurrentProfileCount(uint64_t Count) { |
| 1570 | PGO->setCurrentRegionCount(Count); |
| 1571 | } |
| 1572 | |
| 1573 | /// Get the profiler's current count. This is generally the count for the most |
| 1574 | /// recently incremented counter. |
| 1575 | uint64_t CodeGenFunction::getCurrentProfileCount() { |
| 1576 | return PGO->getCurrentRegionCount(); |
| 1577 | } |
| 1578 | |