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