1//===- PGOInstrumentation.cpp - MST-based PGO Instrumentation -------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements PGO instrumentation using a minimum spanning tree based
10// on the following paper:
11// [1] Donald E. Knuth, Francis R. Stevenson. Optimal measurement of points
12// for program frequency counts. BIT Numerical Mathematics 1973, Volume 13,
13// Issue 3, pp 313-322
14// The idea of the algorithm based on the fact that for each node (except for
15// the entry and exit), the sum of incoming edge counts equals the sum of
16// outgoing edge counts. The count of edge on spanning tree can be derived from
17// those edges not on the spanning tree. Knuth proves this method instruments
18// the minimum number of edges.
19//
20// The minimal spanning tree here is actually a maximum weight tree -- on-tree
21// edges have higher frequencies (more likely to execute). The idea is to
22// instrument those less frequently executed edges to reduce the runtime
23// overhead of instrumented binaries.
24//
25// This file contains two passes:
26// (1) Pass PGOInstrumentationGen which instruments the IR to generate edge
27// count profile, and generates the instrumentation for indirect call
28// profiling.
29// (2) Pass PGOInstrumentationUse which reads the edge count profile and
30// annotates the branch weights. It also reads the indirect call value
31// profiling records and annotate the indirect call instructions.
32//
33// To get the precise counter information, These two passes need to invoke at
34// the same compilation point (so they see the same IR). For pass
35// PGOInstrumentationGen, the real work is done in instrumentOneFunc(). For
36// pass PGOInstrumentationUse, the real work in done in class PGOUseFunc and
37// the profile is opened in module level and passed to each PGOUseFunc instance.
38// The shared code for PGOInstrumentationGen and PGOInstrumentationUse is put
39// in class FuncPGOInstrumentation.
40//
41// Class PGOEdge represents a CFG edge and some auxiliary information. Class
42// BBInfo contains auxiliary information for each BB. These two classes are used
43// in pass PGOInstrumentationGen. Class PGOUseEdge and UseBBInfo are the derived
44// class of PGOEdge and BBInfo, respectively. They contains extra data structure
45// used in populating profile counters.
46// The MST implementation is in Class CFGMST (CFGMST.h).
47//
48//===----------------------------------------------------------------------===//
49
50#include "llvm/Transforms/Instrumentation/PGOInstrumentation.h"
51#include "ValueProfileCollector.h"
52#include "llvm/ADT/APInt.h"
53#include "llvm/ADT/ArrayRef.h"
54#include "llvm/ADT/STLExtras.h"
55#include "llvm/ADT/SmallVector.h"
56#include "llvm/ADT/Statistic.h"
57#include "llvm/ADT/StringRef.h"
58#include "llvm/ADT/StringSet.h"
59#include "llvm/ADT/Twine.h"
60#include "llvm/ADT/iterator.h"
61#include "llvm/ADT/iterator_range.h"
62#include "llvm/Analysis/BlockFrequencyInfo.h"
63#include "llvm/Analysis/BranchProbabilityInfo.h"
64#include "llvm/Analysis/CFG.h"
65#include "llvm/Analysis/LoopInfo.h"
66#include "llvm/Analysis/OptimizationRemarkEmitter.h"
67#include "llvm/Analysis/ProfileSummaryInfo.h"
68#include "llvm/Analysis/TargetLibraryInfo.h"
69#include "llvm/IR/Attributes.h"
70#include "llvm/IR/BasicBlock.h"
71#include "llvm/IR/CFG.h"
72#include "llvm/IR/Comdat.h"
73#include "llvm/IR/Constant.h"
74#include "llvm/IR/Constants.h"
75#include "llvm/IR/DiagnosticInfo.h"
76#include "llvm/IR/Dominators.h"
77#include "llvm/IR/EHPersonalities.h"
78#include "llvm/IR/Function.h"
79#include "llvm/IR/GlobalAlias.h"
80#include "llvm/IR/GlobalValue.h"
81#include "llvm/IR/GlobalVariable.h"
82#include "llvm/IR/IRBuilder.h"
83#include "llvm/IR/InstVisitor.h"
84#include "llvm/IR/InstrTypes.h"
85#include "llvm/IR/Instruction.h"
86#include "llvm/IR/Instructions.h"
87#include "llvm/IR/IntrinsicInst.h"
88#include "llvm/IR/Intrinsics.h"
89#include "llvm/IR/LLVMContext.h"
90#include "llvm/IR/MDBuilder.h"
91#include "llvm/IR/Module.h"
92#include "llvm/IR/PassManager.h"
93#include "llvm/IR/ProfDataUtils.h"
94#include "llvm/IR/ProfileSummary.h"
95#include "llvm/IR/Type.h"
96#include "llvm/IR/Value.h"
97#include "llvm/ProfileData/InstrProf.h"
98#include "llvm/ProfileData/InstrProfReader.h"
99#include "llvm/Support/BranchProbability.h"
100#include "llvm/Support/CRC.h"
101#include "llvm/Support/Casting.h"
102#include "llvm/Support/CommandLine.h"
103#include "llvm/Support/Compiler.h"
104#include "llvm/Support/DOTGraphTraits.h"
105#include "llvm/Support/Debug.h"
106#include "llvm/Support/Error.h"
107#include "llvm/Support/ErrorHandling.h"
108#include "llvm/Support/GraphWriter.h"
109#include "llvm/Support/VirtualFileSystem.h"
110#include "llvm/Support/raw_ostream.h"
111#include "llvm/TargetParser/Triple.h"
112#include "llvm/Transforms/Instrumentation/BlockCoverageInference.h"
113#include "llvm/Transforms/Instrumentation/CFGMST.h"
114#include "llvm/Transforms/Utils/BasicBlockUtils.h"
115#include "llvm/Transforms/Utils/Instrumentation.h"
116#include "llvm/Transforms/Utils/MisExpect.h"
117#include "llvm/Transforms/Utils/ModuleUtils.h"
118#include <algorithm>
119#include <cassert>
120#include <cstdint>
121#include <memory>
122#include <numeric>
123#include <optional>
124#include <stack>
125#include <string>
126#include <unordered_map>
127#include <utility>
128#include <vector>
129
130using namespace llvm;
131using ProfileCount = Function::ProfileCount;
132using VPCandidateInfo = ValueProfileCollector::CandidateInfo;
133
134#define DEBUG_TYPE "pgo-instrumentation"
135
136STATISTIC(NumOfPGOInstrument, "Number of edges instrumented.");
137STATISTIC(NumOfPGOSelectInsts, "Number of select instruction instrumented.");
138STATISTIC(NumOfPGOMemIntrinsics, "Number of mem intrinsics instrumented.");
139STATISTIC(NumOfPGOEdge, "Number of edges.");
140STATISTIC(NumOfPGOBB, "Number of basic-blocks.");
141STATISTIC(NumOfPGOSplit, "Number of critical edge splits.");
142STATISTIC(NumOfPGOFunc, "Number of functions having valid profile counts.");
143STATISTIC(NumOfPGOMismatch, "Number of functions having mismatch profile.");
144STATISTIC(NumOfPGOMissing, "Number of functions without profile.");
145STATISTIC(NumOfPGOICall, "Number of indirect call value instrumentations.");
146STATISTIC(NumOfCSPGOInstrument, "Number of edges instrumented in CSPGO.");
147STATISTIC(NumOfCSPGOSelectInsts,
148 "Number of select instruction instrumented in CSPGO.");
149STATISTIC(NumOfCSPGOMemIntrinsics,
150 "Number of mem intrinsics instrumented in CSPGO.");
151STATISTIC(NumOfCSPGOEdge, "Number of edges in CSPGO.");
152STATISTIC(NumOfCSPGOBB, "Number of basic-blocks in CSPGO.");
153STATISTIC(NumOfCSPGOSplit, "Number of critical edge splits in CSPGO.");
154STATISTIC(NumOfCSPGOFunc,
155 "Number of functions having valid profile counts in CSPGO.");
156STATISTIC(NumOfCSPGOMismatch,
157 "Number of functions having mismatch profile in CSPGO.");
158STATISTIC(NumOfCSPGOMissing, "Number of functions without profile in CSPGO.");
159STATISTIC(NumCoveredBlocks, "Number of basic blocks that were executed");
160
161// Command line option to specify the file to read profile from. This is
162// mainly used for testing.
163static cl::opt<std::string> PGOTestProfileFile(
164 "pgo-test-profile-file", cl::init(Val: ""), cl::Hidden,
165 cl::value_desc("filename"),
166 cl::desc("Specify the path of profile data file. This is "
167 "mainly for test purpose."));
168static cl::opt<std::string> PGOTestProfileRemappingFile(
169 "pgo-test-profile-remapping-file", cl::init(Val: ""), cl::Hidden,
170 cl::value_desc("filename"),
171 cl::desc("Specify the path of profile remapping file. This is mainly for "
172 "test purpose."));
173
174// Command line option to disable value profiling. The default is false:
175// i.e. value profiling is enabled by default. This is for debug purpose.
176static cl::opt<bool> DisableValueProfiling("disable-vp", cl::init(Val: false),
177 cl::Hidden,
178 cl::desc("Disable Value Profiling"));
179
180// Command line option to set the maximum number of VP annotations to write to
181// the metadata for a single indirect call callsite.
182static cl::opt<unsigned> MaxNumAnnotations(
183 "icp-max-annotations", cl::init(Val: 3), cl::Hidden,
184 cl::desc("Max number of annotations for a single indirect "
185 "call callsite"));
186
187// Command line option to set the maximum number of value annotations
188// to write to the metadata for a single memop intrinsic.
189static cl::opt<unsigned> MaxNumMemOPAnnotations(
190 "memop-max-annotations", cl::init(Val: 4), cl::Hidden,
191 cl::desc("Max number of precise value annotations for a single memop"
192 "intrinsic"));
193
194// Command line option to control appending FunctionHash to the name of a COMDAT
195// function. This is to avoid the hash mismatch caused by the preinliner.
196static cl::opt<bool> DoComdatRenaming(
197 "do-comdat-renaming", cl::init(Val: false), cl::Hidden,
198 cl::desc("Append function hash to the name of COMDAT function to avoid "
199 "function hash mismatch due to the preinliner"));
200
201namespace llvm {
202// Command line option to enable/disable the warning about missing profile
203// information.
204cl::opt<bool> PGOWarnMissing("pgo-warn-missing-function", cl::init(Val: false),
205 cl::Hidden,
206 cl::desc("Use this option to turn on/off "
207 "warnings about missing profile data for "
208 "functions."));
209
210// Command line option to enable/disable the warning about a hash mismatch in
211// the profile data.
212cl::opt<bool>
213 NoPGOWarnMismatch("no-pgo-warn-mismatch", cl::init(Val: false), cl::Hidden,
214 cl::desc("Use this option to turn off/on "
215 "warnings about profile cfg mismatch."));
216
217// Command line option to enable/disable the warning about a hash mismatch in
218// the profile data for Comdat functions, which often turns out to be false
219// positive due to the pre-instrumentation inline.
220cl::opt<bool> NoPGOWarnMismatchComdatWeak(
221 "no-pgo-warn-mismatch-comdat-weak", cl::init(Val: true), cl::Hidden,
222 cl::desc("The option is used to turn on/off "
223 "warnings about hash mismatch for comdat "
224 "or weak functions."));
225
226// Command line option to enable/disable select instruction instrumentation.
227static cl::opt<bool>
228 PGOInstrSelect("pgo-instr-select", cl::init(Val: true), cl::Hidden,
229 cl::desc("Use this option to turn on/off SELECT "
230 "instruction instrumentation. "));
231
232// Command line option to turn on CFG dot or text dump of raw profile counts
233static cl::opt<PGOViewCountsType> PGOViewRawCounts(
234 "pgo-view-raw-counts", cl::Hidden,
235 cl::desc("A boolean option to show CFG dag or text "
236 "with raw profile counts from "
237 "profile data. See also option "
238 "-pgo-view-counts. To limit graph "
239 "display to only one function, use "
240 "filtering option -view-bfi-func-name."),
241 cl::values(clEnumValN(PGOVCT_None, "none", "do not show."),
242 clEnumValN(PGOVCT_Graph, "graph", "show a graph."),
243 clEnumValN(PGOVCT_Text, "text", "show in text.")));
244
245// Command line option to enable/disable memop intrinsic call.size profiling.
246static cl::opt<bool>
247 PGOInstrMemOP("pgo-instr-memop", cl::init(Val: true), cl::Hidden,
248 cl::desc("Use this option to turn on/off "
249 "memory intrinsic size profiling."));
250
251// Emit branch probability as optimization remarks.
252static cl::opt<bool>
253 EmitBranchProbability("pgo-emit-branch-prob", cl::init(Val: false), cl::Hidden,
254 cl::desc("When this option is on, the annotated "
255 "branch probability will be emitted as "
256 "optimization remarks: -{Rpass|"
257 "pass-remarks}=pgo-instrumentation"));
258
259static cl::opt<bool> PGOInstrumentEntry(
260 "pgo-instrument-entry", cl::init(Val: false), cl::Hidden,
261 cl::desc("Force to instrument function entry basicblock."));
262
263static cl::opt<bool>
264 PGOInstrumentLoopEntries("pgo-instrument-loop-entries", cl::init(Val: false),
265 cl::Hidden,
266 cl::desc("Force to instrument loop entries."));
267
268static cl::opt<bool> PGOFunctionEntryCoverage(
269 "pgo-function-entry-coverage", cl::Hidden,
270 cl::desc(
271 "Use this option to enable function entry coverage instrumentation."));
272
273static cl::opt<bool> PGOBlockCoverage(
274 "pgo-block-coverage",
275 cl::desc("Use this option to enable basic block coverage instrumentation"));
276
277static cl::opt<bool>
278 PGOViewBlockCoverageGraph("pgo-view-block-coverage-graph",
279 cl::desc("Create a dot file of CFGs with block "
280 "coverage inference information"));
281
282static cl::opt<bool> PGOTemporalInstrumentation(
283 "pgo-temporal-instrumentation",
284 cl::desc("Use this option to enable temporal instrumentation"));
285
286static cl::opt<bool>
287 PGOFixEntryCount("pgo-fix-entry-count", cl::init(Val: true), cl::Hidden,
288 cl::desc("Fix function entry count in profile use."));
289
290static cl::opt<bool> PGOVerifyHotBFI(
291 "pgo-verify-hot-bfi", cl::init(Val: false), cl::Hidden,
292 cl::desc("Print out the non-match BFI count if a hot raw profile count "
293 "becomes non-hot, or a cold raw profile count becomes hot. "
294 "The print is enabled under -Rpass-analysis=pgo, or "
295 "internal option -pass-remarks-analysis=pgo."));
296
297static cl::opt<bool> PGOVerifyBFI(
298 "pgo-verify-bfi", cl::init(Val: false), cl::Hidden,
299 cl::desc("Print out mismatched BFI counts after setting profile metadata "
300 "The print is enabled under -Rpass-analysis=pgo, or "
301 "internal option -pass-remarks-analysis=pgo."));
302
303static cl::opt<unsigned> PGOVerifyBFIRatio(
304 "pgo-verify-bfi-ratio", cl::init(Val: 2), cl::Hidden,
305 cl::desc("Set the threshold for pgo-verify-bfi: only print out "
306 "mismatched BFI if the difference percentage is greater than "
307 "this value (in percentage)."));
308
309static cl::opt<unsigned> PGOVerifyBFICutoff(
310 "pgo-verify-bfi-cutoff", cl::init(Val: 5), cl::Hidden,
311 cl::desc("Set the threshold for pgo-verify-bfi: skip the counts whose "
312 "profile count value is below."));
313
314static cl::opt<std::string> PGOTraceFuncHash(
315 "pgo-trace-func-hash", cl::init(Val: "-"), cl::Hidden,
316 cl::value_desc("function name"),
317 cl::desc("Trace the hash of the function with this name."));
318
319static cl::opt<unsigned> PGOFunctionSizeThreshold(
320 "pgo-function-size-threshold", cl::Hidden,
321 cl::desc("Do not instrument functions smaller than this threshold."));
322
323static cl::opt<unsigned> PGOFunctionCriticalEdgeThreshold(
324 "pgo-critical-edge-threshold", cl::init(Val: 20000), cl::Hidden,
325 cl::desc("Do not instrument functions with the number of critical edges "
326 " greater than this threshold."));
327
328static cl::opt<uint64_t> PGOColdInstrumentEntryThreshold(
329 "pgo-cold-instrument-entry-threshold", cl::init(Val: 0), cl::Hidden,
330 cl::desc("For cold function instrumentation, skip instrumenting functions "
331 "whose entry count is above the given value."));
332
333static cl::opt<bool> PGOTreatUnknownAsCold(
334 "pgo-treat-unknown-as-cold", cl::init(Val: false), cl::Hidden,
335 cl::desc("For cold function instrumentation, treat count unknown(e.g. "
336 "unprofiled) functions as cold."));
337
338cl::opt<bool> PGOInstrumentColdFunctionOnly(
339 "pgo-instrument-cold-function-only", cl::init(Val: false), cl::Hidden,
340 cl::desc("Enable cold function only instrumentation."));
341
342cl::list<std::string> CtxPGOSkipCallsiteInstrument(
343 "ctx-prof-skip-callsite-instr", cl::Hidden,
344 cl::desc("Do not instrument callsites to functions in this list. Intended "
345 "for testing."));
346
347extern cl::opt<unsigned> MaxNumVTableAnnotations;
348
349// Command line option to turn on CFG dot dump after profile annotation.
350// Defined in Analysis/BlockFrequencyInfo.cpp: -pgo-view-counts
351extern cl::opt<PGOViewCountsType> PGOViewCounts;
352
353// Command line option to specify the name of the function for CFG dump
354// Defined in Analysis/BlockFrequencyInfo.cpp: -view-bfi-func-name=
355extern cl::opt<std::string> ViewBlockFreqFuncName;
356
357// Command line option to enable vtable value profiling. Defined in
358// ProfileData/InstrProf.cpp: -enable-vtable-value-profiling=
359extern cl::opt<bool> EnableVTableValueProfiling;
360extern cl::opt<bool> EnableVTableProfileUse;
361LLVM_ABI extern cl::opt<InstrProfCorrelator::ProfCorrelatorKind>
362 ProfileCorrelate;
363} // namespace llvm
364
365namespace {
366class FunctionInstrumenter final {
367 Module &M;
368 Function &F;
369 TargetLibraryInfo &TLI;
370 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers;
371 BranchProbabilityInfo *const BPI;
372 BlockFrequencyInfo *const BFI;
373 LoopInfo *const LI;
374
375 const PGOInstrumentationType InstrumentationType;
376
377 // FIXME(mtrofin): re-enable this for ctx profiling, for non-indirect calls.
378 // Ctx profiling implicitly captures indirect call cases, but not other
379 // values. Supporting other values is relatively straight-forward - just
380 // another counter range within the context.
381 bool isValueProfilingDisabled() const {
382 return DisableValueProfiling ||
383 InstrumentationType == PGOInstrumentationType::CTXPROF;
384 }
385
386 bool shouldInstrumentEntryBB() const {
387 return PGOInstrumentEntry ||
388 InstrumentationType == PGOInstrumentationType::CTXPROF;
389 }
390
391 bool shouldInstrumentLoopEntries() const { return PGOInstrumentLoopEntries; }
392
393public:
394 FunctionInstrumenter(
395 Module &M, Function &F, TargetLibraryInfo &TLI,
396 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers,
397 BranchProbabilityInfo *BPI = nullptr, BlockFrequencyInfo *BFI = nullptr,
398 LoopInfo *LI = nullptr,
399 PGOInstrumentationType InstrumentationType = PGOInstrumentationType::FDO)
400 : M(M), F(F), TLI(TLI), ComdatMembers(ComdatMembers), BPI(BPI), BFI(BFI),
401 LI(LI), InstrumentationType(InstrumentationType) {}
402
403 void instrument();
404};
405} // namespace
406
407// Return a string describing the branch condition that can be
408// used in static branch probability heuristics:
409static std::string getBranchCondString(Instruction *TI) {
410 BranchInst *BI = dyn_cast<BranchInst>(Val: TI);
411 if (!BI || !BI->isConditional())
412 return std::string();
413
414 Value *Cond = BI->getCondition();
415 ICmpInst *CI = dyn_cast<ICmpInst>(Val: Cond);
416 if (!CI)
417 return std::string();
418
419 std::string result;
420 raw_string_ostream OS(result);
421 OS << CI->getPredicate() << "_";
422 CI->getOperand(i_nocapture: 0)->getType()->print(O&: OS, IsForDebug: true);
423
424 Value *RHS = CI->getOperand(i_nocapture: 1);
425 ConstantInt *CV = dyn_cast<ConstantInt>(Val: RHS);
426 if (CV) {
427 if (CV->isZero())
428 OS << "_Zero";
429 else if (CV->isOne())
430 OS << "_One";
431 else if (CV->isMinusOne())
432 OS << "_MinusOne";
433 else
434 OS << "_Const";
435 }
436 return result;
437}
438
439static const char *ValueProfKindDescr[] = {
440#define VALUE_PROF_KIND(Enumerator, Value, Descr) Descr,
441#include "llvm/ProfileData/InstrProfData.inc"
442};
443
444// Create a COMDAT variable INSTR_PROF_RAW_VERSION_VAR to make the runtime
445// aware this is an ir_level profile so it can set the version flag.
446static GlobalVariable *
447createIRLevelProfileFlagVar(Module &M,
448 PGOInstrumentationType InstrumentationType) {
449 const StringRef VarName(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR));
450 Type *IntTy64 = Type::getInt64Ty(C&: M.getContext());
451 uint64_t ProfileVersion = (INSTR_PROF_RAW_VERSION | VARIANT_MASK_IR_PROF);
452 if (InstrumentationType == PGOInstrumentationType::CSFDO)
453 ProfileVersion |= VARIANT_MASK_CSIR_PROF;
454 if (PGOInstrumentEntry ||
455 InstrumentationType == PGOInstrumentationType::CTXPROF)
456 ProfileVersion |= VARIANT_MASK_INSTR_ENTRY;
457 if (PGOInstrumentLoopEntries)
458 ProfileVersion |= VARIANT_MASK_INSTR_LOOP_ENTRIES;
459 if (ProfileCorrelate == InstrProfCorrelator::DEBUG_INFO)
460 ProfileVersion |= VARIANT_MASK_DBG_CORRELATE;
461 if (PGOFunctionEntryCoverage)
462 ProfileVersion |=
463 VARIANT_MASK_BYTE_COVERAGE | VARIANT_MASK_FUNCTION_ENTRY_ONLY;
464 if (PGOBlockCoverage)
465 ProfileVersion |= VARIANT_MASK_BYTE_COVERAGE;
466 if (PGOTemporalInstrumentation)
467 ProfileVersion |= VARIANT_MASK_TEMPORAL_PROF;
468 auto IRLevelVersionVariable = new GlobalVariable(
469 M, IntTy64, true, GlobalValue::WeakAnyLinkage,
470 Constant::getIntegerValue(Ty: IntTy64, V: APInt(64, ProfileVersion)), VarName);
471 IRLevelVersionVariable->setVisibility(GlobalValue::HiddenVisibility);
472 if (isGPUProfTarget(M))
473 IRLevelVersionVariable->setVisibility(
474 llvm::GlobalValue::ProtectedVisibility);
475
476 Triple TT(M.getTargetTriple());
477 if (TT.supportsCOMDAT()) {
478 IRLevelVersionVariable->setLinkage(GlobalValue::ExternalLinkage);
479 IRLevelVersionVariable->setComdat(M.getOrInsertComdat(Name: VarName));
480 }
481 return IRLevelVersionVariable;
482}
483
484namespace {
485
486/// The select instruction visitor plays three roles specified
487/// by the mode. In \c VM_counting mode, it simply counts the number of
488/// select instructions. In \c VM_instrument mode, it inserts code to count
489/// the number times TrueValue of select is taken. In \c VM_annotate mode,
490/// it reads the profile data and annotate the select instruction with metadata.
491enum VisitMode { VM_counting, VM_instrument, VM_annotate };
492class PGOUseFunc;
493
494/// Instruction Visitor class to visit select instructions.
495struct SelectInstVisitor : public InstVisitor<SelectInstVisitor> {
496 Function &F;
497 unsigned NSIs = 0; // Number of select instructions instrumented.
498 VisitMode Mode = VM_counting; // Visiting mode.
499 unsigned *CurCtrIdx = nullptr; // Pointer to current counter index.
500 unsigned TotalNumCtrs = 0; // Total number of counters
501 GlobalValue *FuncNameVar = nullptr;
502 uint64_t FuncHash = 0;
503 PGOUseFunc *UseFunc = nullptr;
504 bool HasSingleByteCoverage;
505
506 SelectInstVisitor(Function &Func, bool HasSingleByteCoverage)
507 : F(Func), HasSingleByteCoverage(HasSingleByteCoverage) {}
508
509 void countSelects() {
510 NSIs = 0;
511 Mode = VM_counting;
512 visit(F);
513 }
514
515 // Visit the IR stream and instrument all select instructions. \p
516 // Ind is a pointer to the counter index variable; \p TotalNC
517 // is the total number of counters; \p FNV is the pointer to the
518 // PGO function name var; \p FHash is the function hash.
519 void instrumentSelects(unsigned *Ind, unsigned TotalNC, GlobalValue *FNV,
520 uint64_t FHash) {
521 Mode = VM_instrument;
522 CurCtrIdx = Ind;
523 TotalNumCtrs = TotalNC;
524 FuncHash = FHash;
525 FuncNameVar = FNV;
526 visit(F);
527 }
528
529 // Visit the IR stream and annotate all select instructions.
530 void annotateSelects(PGOUseFunc *UF, unsigned *Ind) {
531 Mode = VM_annotate;
532 UseFunc = UF;
533 CurCtrIdx = Ind;
534 visit(F);
535 }
536
537 void instrumentOneSelectInst(SelectInst &SI);
538 void annotateOneSelectInst(SelectInst &SI);
539
540 // Visit \p SI instruction and perform tasks according to visit mode.
541 void visitSelectInst(SelectInst &SI);
542
543 // Return the number of select instructions. This needs be called after
544 // countSelects().
545 unsigned getNumOfSelectInsts() const { return NSIs; }
546};
547
548/// This class implements the CFG edges for the Minimum Spanning Tree (MST)
549/// based instrumentation.
550/// Note that the CFG can be a multi-graph. So there might be multiple edges
551/// with the same SrcBB and DestBB.
552struct PGOEdge {
553 BasicBlock *SrcBB;
554 BasicBlock *DestBB;
555 uint64_t Weight;
556 bool InMST = false;
557 bool Removed = false;
558 bool IsCritical = false;
559
560 PGOEdge(BasicBlock *Src, BasicBlock *Dest, uint64_t W = 1)
561 : SrcBB(Src), DestBB(Dest), Weight(W) {}
562
563 /// Return the information string of an edge.
564 std::string infoString() const {
565 return (Twine(Removed ? "-" : " ") + (InMST ? " " : "*") +
566 (IsCritical ? "c" : " ") + " W=" + Twine(Weight))
567 .str();
568 }
569};
570
571/// This class stores the auxiliary information for each BB in the MST.
572struct PGOBBInfo {
573 PGOBBInfo *Group;
574 uint32_t Index;
575 uint32_t Rank = 0;
576
577 PGOBBInfo(unsigned IX) : Group(this), Index(IX) {}
578
579 /// Return the information string of this object.
580 std::string infoString() const {
581 return (Twine("Index=") + Twine(Index)).str();
582 }
583};
584
585// This class implements the CFG edges. Note the CFG can be a multi-graph.
586template <class Edge, class BBInfo> class FuncPGOInstrumentation {
587private:
588 Function &F;
589
590 // Is this is context-sensitive instrumentation.
591 bool IsCS;
592
593 // A map that stores the Comdat group in function F.
594 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers;
595
596 ValueProfileCollector VPC;
597
598 void computeCFGHash();
599 void renameComdatFunction();
600
601public:
602 const TargetLibraryInfo &TLI;
603 std::vector<std::vector<VPCandidateInfo>> ValueSites;
604 SelectInstVisitor SIVisitor;
605 std::string FuncName;
606 std::string DeprecatedFuncName;
607 GlobalVariable *FuncNameVar;
608
609 // CFG hash value for this function.
610 uint64_t FunctionHash = 0;
611
612 // The Minimum Spanning Tree of function CFG.
613 CFGMST<Edge, BBInfo> MST;
614
615 const std::optional<BlockCoverageInference> BCI;
616
617 static std::optional<BlockCoverageInference>
618 constructBCI(Function &Func, bool HasSingleByteCoverage,
619 bool InstrumentFuncEntry) {
620 if (HasSingleByteCoverage)
621 return BlockCoverageInference(Func, InstrumentFuncEntry);
622 return {};
623 }
624
625 // Collect all the BBs that will be instrumented, and store them in
626 // InstrumentBBs.
627 void getInstrumentBBs(std::vector<BasicBlock *> &InstrumentBBs);
628
629 // Give an edge, find the BB that will be instrumented.
630 // Return nullptr if there is no BB to be instrumented.
631 BasicBlock *getInstrBB(Edge *E);
632
633 // Return the auxiliary BB information.
634 BBInfo &getBBInfo(const BasicBlock *BB) const { return MST.getBBInfo(BB); }
635
636 // Return the auxiliary BB information if available.
637 BBInfo *findBBInfo(const BasicBlock *BB) const { return MST.findBBInfo(BB); }
638
639 // Dump edges and BB information.
640 void dumpInfo(StringRef Str = "") const {
641 MST.dumpEdges(dbgs(), Twine("Dump Function ") + FuncName +
642 " Hash: " + Twine(FunctionHash) + "\t" + Str);
643 }
644
645 FuncPGOInstrumentation(
646 Function &Func, TargetLibraryInfo &TLI,
647 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers,
648 bool CreateGlobalVar = false, BranchProbabilityInfo *BPI = nullptr,
649 BlockFrequencyInfo *BFI = nullptr, LoopInfo *LI = nullptr,
650 bool IsCS = false, bool InstrumentFuncEntry = true,
651 bool InstrumentLoopEntries = false, bool HasSingleByteCoverage = false)
652 : F(Func), IsCS(IsCS), ComdatMembers(ComdatMembers), VPC(Func, TLI),
653 TLI(TLI), ValueSites(IPVK_Last + 1),
654 SIVisitor(Func, HasSingleByteCoverage),
655 MST(F, InstrumentFuncEntry, InstrumentLoopEntries, BPI, BFI, LI),
656 BCI(constructBCI(Func, HasSingleByteCoverage, InstrumentFuncEntry)) {
657 if (BCI && PGOViewBlockCoverageGraph)
658 BCI->viewBlockCoverageGraph();
659 // This should be done before CFG hash computation.
660 SIVisitor.countSelects();
661 ValueSites[IPVK_MemOPSize] = VPC.get(Kind: IPVK_MemOPSize);
662 if (!IsCS) {
663 NumOfPGOSelectInsts += SIVisitor.getNumOfSelectInsts();
664 NumOfPGOMemIntrinsics += ValueSites[IPVK_MemOPSize].size();
665 NumOfPGOBB += MST.bbInfoSize();
666 ValueSites[IPVK_IndirectCallTarget] = VPC.get(Kind: IPVK_IndirectCallTarget);
667 if (EnableVTableValueProfiling)
668 ValueSites[IPVK_VTableTarget] = VPC.get(Kind: IPVK_VTableTarget);
669 } else {
670 NumOfCSPGOSelectInsts += SIVisitor.getNumOfSelectInsts();
671 NumOfCSPGOMemIntrinsics += ValueSites[IPVK_MemOPSize].size();
672 NumOfCSPGOBB += MST.bbInfoSize();
673 }
674
675 FuncName = getIRPGOFuncName(F);
676 DeprecatedFuncName = getPGOFuncName(F);
677 computeCFGHash();
678 if (!ComdatMembers.empty())
679 renameComdatFunction();
680 LLVM_DEBUG(dumpInfo("after CFGMST"));
681
682 for (const auto &E : MST.allEdges()) {
683 if (E->Removed)
684 continue;
685 IsCS ? NumOfCSPGOEdge++ : NumOfPGOEdge++;
686 if (!E->InMST)
687 IsCS ? NumOfCSPGOInstrument++ : NumOfPGOInstrument++;
688 }
689
690 if (CreateGlobalVar)
691 FuncNameVar = createPGOFuncNameVar(F, PGOFuncName: FuncName);
692 }
693};
694
695} // end anonymous namespace
696
697// Compute Hash value for the CFG: the lower 32 bits are CRC32 of the index
698// value of each BB in the CFG. The higher 32 bits are the CRC32 of the numbers
699// of selects, indirect calls, mem ops and edges.
700template <class Edge, class BBInfo>
701void FuncPGOInstrumentation<Edge, BBInfo>::computeCFGHash() {
702 std::vector<uint8_t> Indexes;
703 JamCRC JC;
704 for (auto &BB : F) {
705 for (BasicBlock *Succ : successors(BB: &BB)) {
706 auto BI = findBBInfo(BB: Succ);
707 if (BI == nullptr)
708 continue;
709 uint32_t Index = BI->Index;
710 for (int J = 0; J < 4; J++)
711 Indexes.push_back(x: (uint8_t)(Index >> (J * 8)));
712 }
713 }
714 JC.update(Data: Indexes);
715
716 JamCRC JCH;
717 // The higher 32 bits.
718 auto updateJCH = [&JCH](uint64_t Num) {
719 uint8_t Data[8];
720 support::endian::write64le(P: Data, V: Num);
721 JCH.update(Data);
722 };
723 updateJCH((uint64_t)SIVisitor.getNumOfSelectInsts());
724 updateJCH((uint64_t)ValueSites[IPVK_IndirectCallTarget].size());
725 updateJCH((uint64_t)ValueSites[IPVK_MemOPSize].size());
726 if (BCI) {
727 updateJCH(BCI->getInstrumentedBlocksHash());
728 } else {
729 updateJCH((uint64_t)MST.numEdges());
730 }
731
732 // Hash format for context sensitive profile. Reserve 4 bits for other
733 // information.
734 FunctionHash = (((uint64_t)JCH.getCRC()) << 28) + JC.getCRC();
735
736 // Reserve bit 60-63 for other information purpose.
737 FunctionHash &= NamedInstrProfRecord::FUNC_HASH_MASK;
738 if (IsCS)
739 NamedInstrProfRecord::setCSFlagInHash(FunctionHash);
740 LLVM_DEBUG(dbgs() << "Function Hash Computation for " << F.getName() << ":\n"
741 << " CRC = " << JC.getCRC()
742 << ", Selects = " << SIVisitor.getNumOfSelectInsts()
743 << ", Edges = " << MST.numEdges() << ", ICSites = "
744 << ValueSites[IPVK_IndirectCallTarget].size()
745 << ", Memops = " << ValueSites[IPVK_MemOPSize].size()
746 << ", High32 CRC = " << JCH.getCRC()
747 << ", Hash = " << FunctionHash << "\n";);
748
749 if (PGOTraceFuncHash != "-" && F.getName().contains(Other: PGOTraceFuncHash))
750 dbgs() << "Funcname=" << F.getName() << ", Hash=" << FunctionHash
751 << " in building " << F.getParent()->getSourceFileName() << "\n";
752}
753
754// Check if we can safely rename this Comdat function.
755static bool canRenameComdat(
756 Function &F,
757 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) {
758 if (!DoComdatRenaming || !canRenameComdatFunc(F, CheckAddressTaken: true))
759 return false;
760
761 // FIXME: Current only handle those Comdat groups that only containing one
762 // function.
763 // (1) For a Comdat group containing multiple functions, we need to have a
764 // unique postfix based on the hashes for each function. There is a
765 // non-trivial code refactoring to do this efficiently.
766 // (2) Variables can not be renamed, so we can not rename Comdat function in a
767 // group including global vars.
768 Comdat *C = F.getComdat();
769 for (auto &&CM : make_range(p: ComdatMembers.equal_range(x: C))) {
770 assert(!isa<GlobalAlias>(CM.second));
771 Function *FM = dyn_cast<Function>(Val: CM.second);
772 if (FM != &F)
773 return false;
774 }
775 return true;
776}
777
778// Append the CFGHash to the Comdat function name.
779template <class Edge, class BBInfo>
780void FuncPGOInstrumentation<Edge, BBInfo>::renameComdatFunction() {
781 if (!canRenameComdat(F, ComdatMembers))
782 return;
783 std::string OrigName = F.getName().str();
784 std::string NewFuncName =
785 Twine(F.getName() + "." + Twine(FunctionHash)).str();
786 F.setName(Twine(NewFuncName));
787 GlobalAlias::create(Linkage: GlobalValue::WeakAnyLinkage, Name: OrigName, Aliasee: &F);
788 FuncName = Twine(FuncName + "." + Twine(FunctionHash)).str();
789 Comdat *NewComdat;
790 Module *M = F.getParent();
791 // For AvailableExternallyLinkage functions, change the linkage to
792 // LinkOnceODR and put them into comdat. This is because after renaming, there
793 // is no backup external copy available for the function.
794 if (!F.hasComdat()) {
795 assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage);
796 NewComdat = M->getOrInsertComdat(Name: StringRef(NewFuncName));
797 F.setLinkage(GlobalValue::LinkOnceODRLinkage);
798 F.setComdat(NewComdat);
799 return;
800 }
801
802 // This function belongs to a single function Comdat group.
803 Comdat *OrigComdat = F.getComdat();
804 std::string NewComdatName =
805 Twine(OrigComdat->getName() + "." + Twine(FunctionHash)).str();
806 NewComdat = M->getOrInsertComdat(Name: StringRef(NewComdatName));
807 NewComdat->setSelectionKind(OrigComdat->getSelectionKind());
808
809 for (auto &&CM : make_range(p: ComdatMembers.equal_range(x: OrigComdat))) {
810 // Must be a function.
811 cast<Function>(Val: CM.second)->setComdat(NewComdat);
812 }
813}
814
815/// Collect all the BBs that will be instruments and add them to
816/// `InstrumentBBs`.
817template <class Edge, class BBInfo>
818void FuncPGOInstrumentation<Edge, BBInfo>::getInstrumentBBs(
819 std::vector<BasicBlock *> &InstrumentBBs) {
820 if (BCI) {
821 for (auto &BB : F)
822 if (BCI->shouldInstrumentBlock(BB))
823 InstrumentBBs.push_back(x: &BB);
824 return;
825 }
826
827 // Use a worklist as we will update the vector during the iteration.
828 std::vector<Edge *> EdgeList;
829 EdgeList.reserve(MST.numEdges());
830 for (const auto &E : MST.allEdges())
831 EdgeList.push_back(E.get());
832
833 for (auto &E : EdgeList) {
834 BasicBlock *InstrBB = getInstrBB(E);
835 if (InstrBB)
836 InstrumentBBs.push_back(x: InstrBB);
837 }
838}
839
840// Given a CFG E to be instrumented, find which BB to place the instrumented
841// code. The function will split the critical edge if necessary.
842template <class Edge, class BBInfo>
843BasicBlock *FuncPGOInstrumentation<Edge, BBInfo>::getInstrBB(Edge *E) {
844 if (E->InMST || E->Removed)
845 return nullptr;
846
847 BasicBlock *SrcBB = E->SrcBB;
848 BasicBlock *DestBB = E->DestBB;
849 // For a fake edge, instrument the real BB.
850 if (SrcBB == nullptr)
851 return DestBB;
852 if (DestBB == nullptr)
853 return SrcBB;
854
855 auto canInstrument = [](BasicBlock *BB) -> BasicBlock * {
856 // There are basic blocks (such as catchswitch) cannot be instrumented.
857 // If the returned first insertion point is the end of BB, skip this BB.
858 if (BB->getFirstNonPHIOrDbgOrAlloca() == BB->end())
859 return nullptr;
860 return BB;
861 };
862
863 // Instrument the SrcBB if it has a single successor,
864 // otherwise, the DestBB if this is not a critical edge.
865 Instruction *TI = SrcBB->getTerminator();
866 if (TI->getNumSuccessors() <= 1)
867 return canInstrument(SrcBB);
868 if (!E->IsCritical)
869 return canInstrument(DestBB);
870
871 // Some IndirectBr critical edges cannot be split by the previous
872 // SplitIndirectBrCriticalEdges call. Bail out.
873 unsigned SuccNum = GetSuccessorNumber(BB: SrcBB, Succ: DestBB);
874 BasicBlock *InstrBB =
875 isa<IndirectBrInst>(Val: TI) ? nullptr : SplitCriticalEdge(TI, SuccNum);
876 if (!InstrBB) {
877 LLVM_DEBUG(
878 dbgs() << "Fail to split critical edge: not instrument this edge.\n");
879 return nullptr;
880 }
881 // For a critical edge, we have to split. Instrument the newly
882 // created BB.
883 IsCS ? NumOfCSPGOSplit++ : NumOfPGOSplit++;
884 LLVM_DEBUG(dbgs() << "Split critical edge: " << getBBInfo(SrcBB).Index
885 << " --> " << getBBInfo(DestBB).Index << "\n");
886 // Need to add two new edges. First one: Add new edge of SrcBB->InstrBB.
887 MST.addEdge(SrcBB, InstrBB, 0);
888 // Second one: Add new edge of InstrBB->DestBB.
889 Edge &NewEdge1 = MST.addEdge(InstrBB, DestBB, 0);
890 NewEdge1.InMST = true;
891 E->Removed = true;
892
893 return canInstrument(InstrBB);
894}
895
896// When generating value profiling calls on Windows routines that make use of
897// handler funclets for exception processing an operand bundle needs to attached
898// to the called function. This routine will set \p OpBundles to contain the
899// funclet information, if any is needed, that should be placed on the generated
900// value profiling call for the value profile candidate call.
901static void
902populateEHOperandBundle(VPCandidateInfo &Cand,
903 DenseMap<BasicBlock *, ColorVector> &BlockColors,
904 SmallVectorImpl<OperandBundleDef> &OpBundles) {
905 auto *OrigCall = dyn_cast<CallBase>(Val: Cand.AnnotatedInst);
906 if (!OrigCall)
907 return;
908
909 if (!isa<IntrinsicInst>(Val: OrigCall)) {
910 // The instrumentation call should belong to the same funclet as a
911 // non-intrinsic call, so just copy the operand bundle, if any exists.
912 std::optional<OperandBundleUse> ParentFunclet =
913 OrigCall->getOperandBundle(ID: LLVMContext::OB_funclet);
914 if (ParentFunclet)
915 OpBundles.emplace_back(Args: OperandBundleDef(*ParentFunclet));
916 } else {
917 // Intrinsics or other instructions do not get funclet information from the
918 // front-end. Need to use the BlockColors that was computed by the routine
919 // colorEHFunclets to determine whether a funclet is needed.
920 if (!BlockColors.empty()) {
921 const ColorVector &CV = BlockColors.find(Val: OrigCall->getParent())->second;
922 assert(CV.size() == 1 && "non-unique color for block!");
923 BasicBlock::iterator EHPadIt = CV.front()->getFirstNonPHIIt();
924 if (EHPadIt->isEHPad())
925 OpBundles.emplace_back(Args: "funclet", Args: &*EHPadIt);
926 }
927 }
928}
929
930// Visit all edge and instrument the edges not in MST, and do value profiling.
931// Critical edges will be split.
932void FunctionInstrumenter::instrument() {
933 if (!PGOBlockCoverage) {
934 // Split indirectbr critical edges here before computing the MST rather than
935 // later in getInstrBB() to avoid invalidating it.
936 SplitIndirectBrCriticalEdges(F, /*IgnoreBlocksWithoutPHI=*/false, BPI, BFI);
937 }
938
939 const bool IsCtxProf = InstrumentationType == PGOInstrumentationType::CTXPROF;
940 FuncPGOInstrumentation<PGOEdge, PGOBBInfo> FuncInfo(
941 F, TLI, ComdatMembers, /*CreateGlobalVar=*/!IsCtxProf, BPI, BFI, LI,
942 InstrumentationType == PGOInstrumentationType::CSFDO,
943 shouldInstrumentEntryBB(), shouldInstrumentLoopEntries(),
944 PGOBlockCoverage);
945
946 auto *const Name = IsCtxProf ? cast<GlobalValue>(Val: &F) : FuncInfo.FuncNameVar;
947 auto *const CFGHash =
948 ConstantInt::get(Ty: Type::getInt64Ty(C&: M.getContext()), V: FuncInfo.FunctionHash);
949 // Make sure that pointer to global is passed in with zero addrspace
950 // This is relevant during GPU profiling
951 auto *NormalizedNamePtr = ConstantExpr::getPointerBitCastOrAddrSpaceCast(
952 C: Name, Ty: PointerType::get(C&: M.getContext(), AddressSpace: 0));
953 if (PGOFunctionEntryCoverage) {
954 auto &EntryBB = F.getEntryBlock();
955 IRBuilder<> Builder(&EntryBB, EntryBB.getFirstNonPHIOrDbgOrAlloca());
956 // llvm.instrprof.cover(i8* <name>, i64 <hash>, i32 <num-counters>,
957 // i32 <index>)
958 Builder.CreateIntrinsic(
959 ID: Intrinsic::instrprof_cover,
960 Args: {NormalizedNamePtr, CFGHash, Builder.getInt32(C: 1), Builder.getInt32(C: 0)});
961 return;
962 }
963
964 std::vector<BasicBlock *> InstrumentBBs;
965 FuncInfo.getInstrumentBBs(InstrumentBBs);
966 unsigned NumCounters =
967 InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts();
968
969 if (IsCtxProf) {
970 StringSet<> SkipCSInstr(llvm::from_range, CtxPGOSkipCallsiteInstrument);
971
972 auto *CSIntrinsic =
973 Intrinsic::getOrInsertDeclaration(M: &M, id: Intrinsic::instrprof_callsite);
974 // We want to count the instrumentable callsites, then instrument them. This
975 // is because the llvm.instrprof.callsite intrinsic has an argument (like
976 // the other instrprof intrinsics) capturing the total number of
977 // instrumented objects (counters, or callsites, in this case). In this
978 // case, we want that value so we can readily pass it to the compiler-rt
979 // APIs that may have to allocate memory based on the nr of callsites.
980 // The traversal logic is the same for both counting and instrumentation,
981 // just needs to be done in succession.
982 auto Visit = [&](llvm::function_ref<void(CallBase * CB)> Visitor) {
983 for (auto &BB : F)
984 for (auto &Instr : BB)
985 if (auto *CS = dyn_cast<CallBase>(Val: &Instr)) {
986 if (!InstrProfCallsite::canInstrumentCallsite(CB: *CS))
987 continue;
988 if (CS->getCalledFunction() &&
989 SkipCSInstr.contains(key: CS->getCalledFunction()->getName()))
990 continue;
991 Visitor(CS);
992 }
993 };
994 // First, count callsites.
995 uint32_t TotalNumCallsites = 0;
996 Visit([&TotalNumCallsites](auto *) { ++TotalNumCallsites; });
997
998 // Now instrument.
999 uint32_t CallsiteIndex = 0;
1000 Visit([&](auto *CB) {
1001 IRBuilder<> Builder(CB);
1002 Builder.CreateCall(CSIntrinsic,
1003 {Name, CFGHash, Builder.getInt32(C: TotalNumCallsites),
1004 Builder.getInt32(C: CallsiteIndex++),
1005 CB->getCalledOperand()});
1006 });
1007 }
1008
1009 uint32_t I = 0;
1010 if (PGOTemporalInstrumentation) {
1011 NumCounters += PGOBlockCoverage ? 8 : 1;
1012 auto &EntryBB = F.getEntryBlock();
1013 IRBuilder<> Builder(&EntryBB, EntryBB.getFirstNonPHIOrDbgOrAlloca());
1014 // llvm.instrprof.timestamp(i8* <name>, i64 <hash>, i32 <num-counters>,
1015 // i32 <index>)
1016 Builder.CreateIntrinsic(ID: Intrinsic::instrprof_timestamp,
1017 Args: {NormalizedNamePtr, CFGHash,
1018 Builder.getInt32(C: NumCounters),
1019 Builder.getInt32(C: I)});
1020 I += PGOBlockCoverage ? 8 : 1;
1021 }
1022
1023 for (auto *InstrBB : InstrumentBBs) {
1024 IRBuilder<> Builder(InstrBB, InstrBB->getFirstNonPHIOrDbgOrAlloca());
1025 assert(Builder.GetInsertPoint() != InstrBB->end() &&
1026 "Cannot get the Instrumentation point");
1027 // llvm.instrprof.increment(i8* <name>, i64 <hash>, i32 <num-counters>,
1028 // i32 <index>)
1029 Builder.CreateIntrinsic(ID: PGOBlockCoverage ? Intrinsic::instrprof_cover
1030 : Intrinsic::instrprof_increment,
1031 Args: {NormalizedNamePtr, CFGHash,
1032 Builder.getInt32(C: NumCounters),
1033 Builder.getInt32(C: I++)});
1034 }
1035
1036 // Now instrument select instructions:
1037 FuncInfo.SIVisitor.instrumentSelects(Ind: &I, TotalNC: NumCounters, FNV: Name,
1038 FHash: FuncInfo.FunctionHash);
1039 assert(I == NumCounters);
1040
1041 if (isValueProfilingDisabled())
1042 return;
1043
1044 NumOfPGOICall += FuncInfo.ValueSites[IPVK_IndirectCallTarget].size();
1045
1046 // Intrinsic function calls do not have funclet operand bundles needed for
1047 // Windows exception handling attached to them. However, if value profiling is
1048 // inserted for one of these calls, then a funclet value will need to be set
1049 // on the instrumentation call based on the funclet coloring.
1050 DenseMap<BasicBlock *, ColorVector> BlockColors;
1051 if (F.hasPersonalityFn() &&
1052 isScopedEHPersonality(Pers: classifyEHPersonality(Pers: F.getPersonalityFn())))
1053 BlockColors = colorEHFunclets(F);
1054
1055 // For each VP Kind, walk the VP candidates and instrument each one.
1056 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
1057 unsigned SiteIndex = 0;
1058 if (Kind == IPVK_MemOPSize && !PGOInstrMemOP)
1059 continue;
1060
1061 for (VPCandidateInfo Cand : FuncInfo.ValueSites[Kind]) {
1062 LLVM_DEBUG(dbgs() << "Instrument one VP " << ValueProfKindDescr[Kind]
1063 << " site: CallSite Index = " << SiteIndex << "\n");
1064
1065 IRBuilder<> Builder(Cand.InsertPt);
1066 assert(Builder.GetInsertPoint() != Cand.InsertPt->getParent()->end() &&
1067 "Cannot get the Instrumentation point");
1068
1069 Value *ToProfile = nullptr;
1070 if (Cand.V->getType()->isIntegerTy())
1071 ToProfile = Builder.CreateZExtOrTrunc(V: Cand.V, DestTy: Builder.getInt64Ty());
1072 else if (Cand.V->getType()->isPointerTy())
1073 ToProfile = Builder.CreatePtrToInt(V: Cand.V, DestTy: Builder.getInt64Ty());
1074 assert(ToProfile && "value profiling Value is of unexpected type");
1075
1076 auto *NormalizedNamePtr = ConstantExpr::getPointerBitCastOrAddrSpaceCast(
1077 C: Name, Ty: PointerType::get(C&: M.getContext(), AddressSpace: 0));
1078
1079 SmallVector<OperandBundleDef, 1> OpBundles;
1080 populateEHOperandBundle(Cand, BlockColors, OpBundles);
1081 Builder.CreateCall(
1082 Callee: Intrinsic::getOrInsertDeclaration(M: &M,
1083 id: Intrinsic::instrprof_value_profile),
1084 Args: {NormalizedNamePtr, Builder.getInt64(C: FuncInfo.FunctionHash),
1085 ToProfile, Builder.getInt32(C: Kind), Builder.getInt32(C: SiteIndex++)},
1086 OpBundles);
1087 }
1088 } // IPVK_First <= Kind <= IPVK_Last
1089}
1090
1091namespace {
1092
1093// This class represents a CFG edge in profile use compilation.
1094struct PGOUseEdge : public PGOEdge {
1095 using PGOEdge::PGOEdge;
1096
1097 std::optional<uint64_t> Count;
1098
1099 // Set edge count value
1100 void setEdgeCount(uint64_t Value) { Count = Value; }
1101
1102 // Return the information string for this object.
1103 std::string infoString() const {
1104 if (!Count)
1105 return PGOEdge::infoString();
1106 return (Twine(PGOEdge::infoString()) + " Count=" + Twine(*Count)).str();
1107 }
1108};
1109
1110using DirectEdges = SmallVector<PGOUseEdge *, 2>;
1111
1112// This class stores the auxiliary information for each BB.
1113struct PGOUseBBInfo : public PGOBBInfo {
1114 std::optional<uint64_t> Count;
1115 int32_t UnknownCountInEdge = 0;
1116 int32_t UnknownCountOutEdge = 0;
1117 DirectEdges InEdges;
1118 DirectEdges OutEdges;
1119
1120 PGOUseBBInfo(unsigned IX) : PGOBBInfo(IX) {}
1121
1122 // Set the profile count value for this BB.
1123 void setBBInfoCount(uint64_t Value) { Count = Value; }
1124
1125 // Return the information string of this object.
1126 std::string infoString() const {
1127 if (!Count)
1128 return PGOBBInfo::infoString();
1129 return (Twine(PGOBBInfo::infoString()) + " Count=" + Twine(*Count)).str();
1130 }
1131
1132 // Add an OutEdge and update the edge count.
1133 void addOutEdge(PGOUseEdge *E) {
1134 OutEdges.push_back(Elt: E);
1135 UnknownCountOutEdge++;
1136 }
1137
1138 // Add an InEdge and update the edge count.
1139 void addInEdge(PGOUseEdge *E) {
1140 InEdges.push_back(Elt: E);
1141 UnknownCountInEdge++;
1142 }
1143};
1144
1145} // end anonymous namespace
1146
1147// Sum up the count values for all the edges.
1148static uint64_t sumEdgeCount(const ArrayRef<PGOUseEdge *> Edges) {
1149 uint64_t Total = 0;
1150 for (const auto &E : Edges) {
1151 if (E->Removed)
1152 continue;
1153 if (E->Count)
1154 Total += *E->Count;
1155 }
1156 return Total;
1157}
1158
1159namespace {
1160
1161class PGOUseFunc {
1162public:
1163 PGOUseFunc(Function &Func, Module *Modu, TargetLibraryInfo &TLI,
1164 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers,
1165 BranchProbabilityInfo *BPI, BlockFrequencyInfo *BFIin,
1166 LoopInfo *LI, ProfileSummaryInfo *PSI, bool IsCS,
1167 bool InstrumentFuncEntry, bool InstrumentLoopEntries,
1168 bool HasSingleByteCoverage)
1169 : F(Func), M(Modu), BFI(BFIin), PSI(PSI),
1170 FuncInfo(Func, TLI, ComdatMembers, false, BPI, BFIin, LI, IsCS,
1171 InstrumentFuncEntry, InstrumentLoopEntries,
1172 HasSingleByteCoverage),
1173 FreqAttr(FFA_Normal), IsCS(IsCS), VPC(Func, TLI) {}
1174
1175 void handleInstrProfError(Error Err, uint64_t MismatchedFuncSum);
1176
1177 /// Get the profile record, assign it to \p ProfileRecord, handle errors if
1178 /// necessary, and assign \p ProgramMaxCount. \returns true if there are no
1179 /// errors.
1180 bool getRecord(IndexedInstrProfReader *PGOReader);
1181
1182 // Read counts for the instrumented BB from profile.
1183 bool readCounters(bool &AllZeros,
1184 InstrProfRecord::CountPseudoKind &PseudoKind);
1185
1186 // Populate the counts for all BBs.
1187 void populateCounters();
1188
1189 // Set block coverage based on profile coverage values.
1190 void populateCoverage();
1191
1192 // Set the branch weights based on the count values.
1193 void setBranchWeights();
1194
1195 // Annotate the value profile call sites for all value kind.
1196 void annotateValueSites();
1197
1198 // Annotate the value profile call sites for one value kind.
1199 void annotateValueSites(uint32_t Kind);
1200
1201 // Annotate the irreducible loop header weights.
1202 void annotateIrrLoopHeaderWeights();
1203
1204 // The hotness of the function from the profile count.
1205 enum FuncFreqAttr { FFA_Normal, FFA_Cold, FFA_Hot };
1206
1207 // Return the function hotness from the profile.
1208 FuncFreqAttr getFuncFreqAttr() const { return FreqAttr; }
1209
1210 // Return the function hash.
1211 uint64_t getFuncHash() const { return FuncInfo.FunctionHash; }
1212
1213 // Return the profile record for this function;
1214 NamedInstrProfRecord &getProfileRecord() { return ProfileRecord; }
1215
1216 // Return the auxiliary BB information.
1217 PGOUseBBInfo &getBBInfo(const BasicBlock *BB) const {
1218 return FuncInfo.getBBInfo(BB);
1219 }
1220
1221 // Return the auxiliary BB information if available.
1222 PGOUseBBInfo *findBBInfo(const BasicBlock *BB) const {
1223 return FuncInfo.findBBInfo(BB);
1224 }
1225
1226 Function &getFunc() const { return F; }
1227
1228 void dumpInfo(StringRef Str = "") const { FuncInfo.dumpInfo(Str); }
1229
1230 uint64_t getProgramMaxCount() const { return ProgramMaxCount; }
1231
1232private:
1233 Function &F;
1234 Module *M;
1235 BlockFrequencyInfo *BFI;
1236 ProfileSummaryInfo *PSI;
1237
1238 // This member stores the shared information with class PGOGenFunc.
1239 FuncPGOInstrumentation<PGOUseEdge, PGOUseBBInfo> FuncInfo;
1240
1241 // The maximum count value in the profile. This is only used in PGO use
1242 // compilation.
1243 uint64_t ProgramMaxCount;
1244
1245 // Position of counter that remains to be read.
1246 uint32_t CountPosition = 0;
1247
1248 // Total size of the profile count for this function.
1249 uint32_t ProfileCountSize = 0;
1250
1251 // ProfileRecord for this function.
1252 NamedInstrProfRecord ProfileRecord;
1253
1254 // Function hotness info derived from profile.
1255 FuncFreqAttr FreqAttr;
1256
1257 // Is to use the context sensitive profile.
1258 bool IsCS;
1259
1260 ValueProfileCollector VPC;
1261
1262 // Find the Instrumented BB and set the value. Return false on error.
1263 bool setInstrumentedCounts(const std::vector<uint64_t> &CountFromProfile);
1264
1265 // Set the edge counter value for the unknown edge -- there should be only
1266 // one unknown edge.
1267 void setEdgeCount(DirectEdges &Edges, uint64_t Value);
1268
1269 // Set the hot/cold inline hints based on the count values.
1270 // FIXME: This function should be removed once the functionality in
1271 // the inliner is implemented.
1272 void markFunctionAttributes(uint64_t EntryCount, uint64_t MaxCount) {
1273 if (PSI->isHotCount(C: EntryCount))
1274 FreqAttr = FFA_Hot;
1275 else if (PSI->isColdCount(C: MaxCount))
1276 FreqAttr = FFA_Cold;
1277 }
1278};
1279
1280} // end anonymous namespace
1281
1282/// Set up InEdges/OutEdges for all BBs in the MST.
1283static void setupBBInfoEdges(
1284 const FuncPGOInstrumentation<PGOUseEdge, PGOUseBBInfo> &FuncInfo) {
1285 // This is not required when there is block coverage inference.
1286 if (FuncInfo.BCI)
1287 return;
1288 for (const auto &E : FuncInfo.MST.allEdges()) {
1289 if (E->Removed)
1290 continue;
1291 const BasicBlock *SrcBB = E->SrcBB;
1292 const BasicBlock *DestBB = E->DestBB;
1293 PGOUseBBInfo &SrcInfo = FuncInfo.getBBInfo(BB: SrcBB);
1294 PGOUseBBInfo &DestInfo = FuncInfo.getBBInfo(BB: DestBB);
1295 SrcInfo.addOutEdge(E: E.get());
1296 DestInfo.addInEdge(E: E.get());
1297 }
1298}
1299
1300// Visit all the edges and assign the count value for the instrumented
1301// edges and the BB. Return false on error.
1302bool PGOUseFunc::setInstrumentedCounts(
1303 const std::vector<uint64_t> &CountFromProfile) {
1304
1305 std::vector<BasicBlock *> InstrumentBBs;
1306 FuncInfo.getInstrumentBBs(InstrumentBBs);
1307
1308 setupBBInfoEdges(FuncInfo);
1309
1310 unsigned NumCounters =
1311 InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts();
1312 // The number of counters here should match the number of counters
1313 // in profile. Return if they mismatch.
1314 if (NumCounters != CountFromProfile.size()) {
1315 return false;
1316 }
1317 auto *FuncEntry = &*F.begin();
1318
1319 // Set the profile count to the Instrumented BBs.
1320 uint32_t I = 0;
1321 for (BasicBlock *InstrBB : InstrumentBBs) {
1322 uint64_t CountValue = CountFromProfile[I++];
1323 PGOUseBBInfo &Info = getBBInfo(BB: InstrBB);
1324 // If we reach here, we know that we have some nonzero count
1325 // values in this function. The entry count should not be 0.
1326 // Fix it if necessary.
1327 if (InstrBB == FuncEntry && CountValue == 0)
1328 CountValue = 1;
1329 Info.setBBInfoCount(CountValue);
1330 }
1331 ProfileCountSize = CountFromProfile.size();
1332 CountPosition = I;
1333
1334 // Set the edge count and update the count of unknown edges for BBs.
1335 auto setEdgeCount = [this](PGOUseEdge *E, uint64_t Value) -> void {
1336 E->setEdgeCount(Value);
1337 this->getBBInfo(BB: E->SrcBB).UnknownCountOutEdge--;
1338 this->getBBInfo(BB: E->DestBB).UnknownCountInEdge--;
1339 };
1340
1341 // Set the profile count the Instrumented edges. There are BBs that not in
1342 // MST but not instrumented. Need to set the edge count value so that we can
1343 // populate the profile counts later.
1344 for (const auto &E : FuncInfo.MST.allEdges()) {
1345 if (E->Removed || E->InMST)
1346 continue;
1347 const BasicBlock *SrcBB = E->SrcBB;
1348 PGOUseBBInfo &SrcInfo = getBBInfo(BB: SrcBB);
1349
1350 // If only one out-edge, the edge profile count should be the same as BB
1351 // profile count.
1352 if (SrcInfo.Count && SrcInfo.OutEdges.size() == 1)
1353 setEdgeCount(E.get(), *SrcInfo.Count);
1354 else {
1355 const BasicBlock *DestBB = E->DestBB;
1356 PGOUseBBInfo &DestInfo = getBBInfo(BB: DestBB);
1357 // If only one in-edge, the edge profile count should be the same as BB
1358 // profile count.
1359 if (DestInfo.Count && DestInfo.InEdges.size() == 1)
1360 setEdgeCount(E.get(), *DestInfo.Count);
1361 }
1362 if (E->Count)
1363 continue;
1364 // E's count should have been set from profile. If not, this meenas E skips
1365 // the instrumentation. We set the count to 0.
1366 setEdgeCount(E.get(), 0);
1367 }
1368 return true;
1369}
1370
1371// Set the count value for the unknown edge. There should be one and only one
1372// unknown edge in Edges vector.
1373void PGOUseFunc::setEdgeCount(DirectEdges &Edges, uint64_t Value) {
1374 for (auto &E : Edges) {
1375 if (E->Count)
1376 continue;
1377 E->setEdgeCount(Value);
1378
1379 getBBInfo(BB: E->SrcBB).UnknownCountOutEdge--;
1380 getBBInfo(BB: E->DestBB).UnknownCountInEdge--;
1381 return;
1382 }
1383 llvm_unreachable("Cannot find the unknown count edge");
1384}
1385
1386// Emit function metadata indicating PGO profile mismatch.
1387static void annotateFunctionWithHashMismatch(Function &F, LLVMContext &ctx) {
1388 const char MetadataName[] = "instr_prof_hash_mismatch";
1389 SmallVector<Metadata *, 2> Names;
1390 // If this metadata already exists, ignore.
1391 auto *Existing = F.getMetadata(KindID: LLVMContext::MD_annotation);
1392 if (Existing) {
1393 MDTuple *Tuple = cast<MDTuple>(Val: Existing);
1394 for (const auto &N : Tuple->operands()) {
1395 if (N.equalsStr(Str: MetadataName))
1396 return;
1397 Names.push_back(Elt: N.get());
1398 }
1399 }
1400
1401 MDBuilder MDB(ctx);
1402 Names.push_back(Elt: MDB.createString(Str: MetadataName));
1403 MDNode *MD = MDTuple::get(Context&: ctx, MDs: Names);
1404 F.setMetadata(KindID: LLVMContext::MD_annotation, Node: MD);
1405}
1406
1407void PGOUseFunc::handleInstrProfError(Error Err, uint64_t MismatchedFuncSum) {
1408 handleAllErrors(E: std::move(Err), Handlers: [&](const InstrProfError &IPE) {
1409 auto &Ctx = M->getContext();
1410 auto Err = IPE.get();
1411 bool SkipWarning = false;
1412 LLVM_DEBUG(dbgs() << "Error in reading profile for Func "
1413 << FuncInfo.FuncName << ": ");
1414 if (Err == instrprof_error::unknown_function) {
1415 IsCS ? NumOfCSPGOMissing++ : NumOfPGOMissing++;
1416 SkipWarning = !PGOWarnMissing;
1417 LLVM_DEBUG(dbgs() << "unknown function");
1418 } else if (Err == instrprof_error::hash_mismatch ||
1419 Err == instrprof_error::malformed) {
1420 IsCS ? NumOfCSPGOMismatch++ : NumOfPGOMismatch++;
1421 SkipWarning =
1422 NoPGOWarnMismatch ||
1423 (NoPGOWarnMismatchComdatWeak &&
1424 (F.hasComdat() || F.getLinkage() == GlobalValue::WeakAnyLinkage ||
1425 F.getLinkage() == GlobalValue::AvailableExternallyLinkage));
1426 LLVM_DEBUG(dbgs() << "hash mismatch (hash= " << FuncInfo.FunctionHash
1427 << " skip=" << SkipWarning << ")");
1428 // Emit function metadata indicating PGO profile mismatch.
1429 annotateFunctionWithHashMismatch(F, ctx&: M->getContext());
1430 }
1431
1432 LLVM_DEBUG(dbgs() << " IsCS=" << IsCS << "\n");
1433 if (SkipWarning)
1434 return;
1435
1436 std::string Msg =
1437 IPE.message() + std::string(" ") + F.getName().str() +
1438 std::string(" Hash = ") + std::to_string(val: FuncInfo.FunctionHash) +
1439 std::string(" up to ") + std::to_string(val: MismatchedFuncSum) +
1440 std::string(" count discarded");
1441
1442 Ctx.diagnose(
1443 DI: DiagnosticInfoPGOProfile(M->getName().data(), Msg, DS_Warning));
1444 });
1445}
1446
1447bool PGOUseFunc::getRecord(IndexedInstrProfReader *PGOReader) {
1448 uint64_t MismatchedFuncSum = 0;
1449 auto Result = PGOReader->getInstrProfRecord(
1450 FuncName: FuncInfo.FuncName, FuncHash: FuncInfo.FunctionHash, DeprecatedFuncName: FuncInfo.DeprecatedFuncName,
1451 MismatchedFuncSum: &MismatchedFuncSum);
1452 if (Error E = Result.takeError()) {
1453 handleInstrProfError(Err: std::move(E), MismatchedFuncSum);
1454 return false;
1455 }
1456 ProfileRecord = std::move(Result.get());
1457 ProgramMaxCount = PGOReader->getMaximumFunctionCount(UseCS: IsCS);
1458 return true;
1459}
1460
1461// Read the profile from ProfileFileName and assign the value to the
1462// instrumented BB and the edges. Return true if the profile are successfully
1463// read, and false on errors.
1464bool PGOUseFunc::readCounters(bool &AllZeros,
1465 InstrProfRecord::CountPseudoKind &PseudoKind) {
1466 auto &Ctx = M->getContext();
1467 PseudoKind = ProfileRecord.getCountPseudoKind();
1468 if (PseudoKind != InstrProfRecord::NotPseudo) {
1469 return true;
1470 }
1471 std::vector<uint64_t> &CountFromProfile = ProfileRecord.Counts;
1472
1473 IsCS ? NumOfCSPGOFunc++ : NumOfPGOFunc++;
1474 LLVM_DEBUG(dbgs() << CountFromProfile.size() << " counts\n");
1475
1476 uint64_t ValueSum = 0;
1477 for (unsigned I = 0, S = CountFromProfile.size(); I < S; I++) {
1478 LLVM_DEBUG(dbgs() << " " << I << ": " << CountFromProfile[I] << "\n");
1479 ValueSum += CountFromProfile[I];
1480 }
1481 AllZeros = (ValueSum == 0);
1482
1483 LLVM_DEBUG(dbgs() << "SUM = " << ValueSum << "\n");
1484
1485 getBBInfo(BB: nullptr).UnknownCountOutEdge = 2;
1486 getBBInfo(BB: nullptr).UnknownCountInEdge = 2;
1487
1488 if (!setInstrumentedCounts(CountFromProfile)) {
1489 LLVM_DEBUG(
1490 dbgs() << "Inconsistent number of counts, skipping this function");
1491 Ctx.diagnose(DI: DiagnosticInfoPGOProfile(
1492 M->getName().data(),
1493 Twine("Inconsistent number of counts in ") + F.getName().str() +
1494 Twine(": the profile may be stale or there is a function name "
1495 "collision."),
1496 DS_Warning));
1497 return false;
1498 }
1499 return true;
1500}
1501
1502void PGOUseFunc::populateCoverage() {
1503 IsCS ? NumOfCSPGOFunc++ : NumOfPGOFunc++;
1504
1505 ArrayRef<uint64_t> CountsFromProfile = ProfileRecord.Counts;
1506 DenseMap<const BasicBlock *, bool> Coverage;
1507 unsigned Index = 0;
1508 for (auto &BB : F)
1509 if (FuncInfo.BCI->shouldInstrumentBlock(BB))
1510 Coverage[&BB] = (CountsFromProfile[Index++] != 0);
1511 assert(Index == CountsFromProfile.size());
1512
1513 // For each B in InverseDependencies[A], if A is covered then B is covered.
1514 DenseMap<const BasicBlock *, DenseSet<const BasicBlock *>>
1515 InverseDependencies;
1516 for (auto &BB : F) {
1517 for (auto *Dep : FuncInfo.BCI->getDependencies(BB)) {
1518 // If Dep is covered then BB is covered.
1519 InverseDependencies[Dep].insert(V: &BB);
1520 }
1521 }
1522
1523 // Infer coverage of the non-instrumented blocks using a flood-fill algorithm.
1524 std::stack<const BasicBlock *> CoveredBlocksToProcess;
1525 for (auto &[BB, IsCovered] : Coverage)
1526 if (IsCovered)
1527 CoveredBlocksToProcess.push(x: BB);
1528
1529 while (!CoveredBlocksToProcess.empty()) {
1530 auto *CoveredBlock = CoveredBlocksToProcess.top();
1531 assert(Coverage[CoveredBlock]);
1532 CoveredBlocksToProcess.pop();
1533 for (auto *BB : InverseDependencies[CoveredBlock]) {
1534 // If CoveredBlock is covered then BB is covered.
1535 bool &Cov = Coverage[BB];
1536 if (Cov)
1537 continue;
1538 Cov = true;
1539 CoveredBlocksToProcess.push(x: BB);
1540 }
1541 }
1542
1543 // Annotate block coverage.
1544 MDBuilder MDB(F.getContext());
1545 // We set the entry count to 10000 if the entry block is covered so that BFI
1546 // can propagate a fraction of this count to the other covered blocks.
1547 F.setEntryCount(Count: Coverage[&F.getEntryBlock()] ? 10000 : 0);
1548 for (auto &BB : F) {
1549 // For a block A and its successor B, we set the edge weight as follows:
1550 // If A is covered and B is covered, set weight=1.
1551 // If A is covered and B is uncovered, set weight=0.
1552 // If A is uncovered, set weight=1.
1553 // This setup will allow BFI to give nonzero profile counts to only covered
1554 // blocks.
1555 SmallVector<uint32_t, 4> Weights;
1556 for (auto *Succ : successors(BB: &BB))
1557 Weights.push_back(Elt: (Coverage[Succ] || !Coverage[&BB]) ? 1 : 0);
1558 if (Weights.size() >= 2)
1559 llvm::setBranchWeights(I&: *BB.getTerminator(), Weights,
1560 /*IsExpected=*/false);
1561 }
1562
1563 unsigned NumCorruptCoverage = 0;
1564 DominatorTree DT(F);
1565 LoopInfo LI(DT);
1566 BranchProbabilityInfo BPI(F, LI);
1567 BlockFrequencyInfo BFI(F, BPI, LI);
1568 auto IsBlockDead = [&](const BasicBlock &BB) -> std::optional<bool> {
1569 if (auto C = BFI.getBlockProfileCount(BB: &BB))
1570 return C == 0;
1571 return {};
1572 };
1573 LLVM_DEBUG(dbgs() << "Block Coverage: (Instrumented=*, Covered=X)\n");
1574 for (auto &BB : F) {
1575 LLVM_DEBUG(dbgs() << (FuncInfo.BCI->shouldInstrumentBlock(BB) ? "* " : " ")
1576 << (Coverage[&BB] ? "X " : " ") << " " << BB.getName()
1577 << "\n");
1578 // In some cases it is possible to find a covered block that has no covered
1579 // successors, e.g., when a block calls a function that may call exit(). In
1580 // those cases, BFI could find its successor to be covered while BCI could
1581 // find its successor to be dead.
1582 const bool &Cov = Coverage[&BB];
1583 if (Cov == IsBlockDead(BB).value_or(u: false)) {
1584 LLVM_DEBUG(
1585 dbgs() << "Found inconsistent block covearge for " << BB.getName()
1586 << ": BCI=" << (Cov ? "Covered" : "Dead") << " BFI="
1587 << (IsBlockDead(BB).value() ? "Dead" : "Covered") << "\n");
1588 ++NumCorruptCoverage;
1589 }
1590 if (Cov)
1591 ++NumCoveredBlocks;
1592 }
1593 if (PGOVerifyBFI && NumCorruptCoverage) {
1594 auto &Ctx = M->getContext();
1595 Ctx.diagnose(DI: DiagnosticInfoPGOProfile(
1596 M->getName().data(),
1597 Twine("Found inconsistent block coverage for function ") + F.getName() +
1598 " in " + Twine(NumCorruptCoverage) + " blocks.",
1599 DS_Warning));
1600 }
1601 if (PGOViewBlockCoverageGraph)
1602 FuncInfo.BCI->viewBlockCoverageGraph(Coverage: &Coverage);
1603}
1604
1605// Populate the counters from instrumented BBs to all BBs.
1606// In the end of this operation, all BBs should have a valid count value.
1607void PGOUseFunc::populateCounters() {
1608 bool Changes = true;
1609 unsigned NumPasses = 0;
1610 while (Changes) {
1611 NumPasses++;
1612 Changes = false;
1613
1614 // For efficient traversal, it's better to start from the end as most
1615 // of the instrumented edges are at the end.
1616 for (auto &BB : reverse(C&: F)) {
1617 PGOUseBBInfo *UseBBInfo = findBBInfo(BB: &BB);
1618 if (UseBBInfo == nullptr)
1619 continue;
1620 if (!UseBBInfo->Count) {
1621 if (UseBBInfo->UnknownCountOutEdge == 0) {
1622 UseBBInfo->Count = sumEdgeCount(Edges: UseBBInfo->OutEdges);
1623 Changes = true;
1624 } else if (UseBBInfo->UnknownCountInEdge == 0) {
1625 UseBBInfo->Count = sumEdgeCount(Edges: UseBBInfo->InEdges);
1626 Changes = true;
1627 }
1628 }
1629 if (UseBBInfo->Count) {
1630 if (UseBBInfo->UnknownCountOutEdge == 1) {
1631 uint64_t Total = 0;
1632 uint64_t OutSum = sumEdgeCount(Edges: UseBBInfo->OutEdges);
1633 // If the one of the successor block can early terminate (no-return),
1634 // we can end up with situation where out edge sum count is larger as
1635 // the source BB's count is collected by a post-dominated block.
1636 if (*UseBBInfo->Count > OutSum)
1637 Total = *UseBBInfo->Count - OutSum;
1638 setEdgeCount(Edges&: UseBBInfo->OutEdges, Value: Total);
1639 Changes = true;
1640 }
1641 if (UseBBInfo->UnknownCountInEdge == 1) {
1642 uint64_t Total = 0;
1643 uint64_t InSum = sumEdgeCount(Edges: UseBBInfo->InEdges);
1644 if (*UseBBInfo->Count > InSum)
1645 Total = *UseBBInfo->Count - InSum;
1646 setEdgeCount(Edges&: UseBBInfo->InEdges, Value: Total);
1647 Changes = true;
1648 }
1649 }
1650 }
1651 }
1652
1653 LLVM_DEBUG(dbgs() << "Populate counts in " << NumPasses << " passes.\n");
1654 (void)NumPasses;
1655#ifndef NDEBUG
1656 // Assert every BB has a valid counter.
1657 for (auto &BB : F) {
1658 auto BI = findBBInfo(&BB);
1659 if (BI == nullptr)
1660 continue;
1661 assert(BI->Count && "BB count is not valid");
1662 }
1663#endif
1664 // Now annotate select instructions. This may fixup impossible block counts.
1665 FuncInfo.SIVisitor.annotateSelects(UF: this, Ind: &CountPosition);
1666 assert(CountPosition == ProfileCountSize);
1667
1668 uint64_t FuncEntryCount = *getBBInfo(BB: &*F.begin()).Count;
1669 uint64_t FuncMaxCount = FuncEntryCount;
1670 for (auto &BB : F) {
1671 auto BI = findBBInfo(BB: &BB);
1672 if (BI == nullptr)
1673 continue;
1674 FuncMaxCount = std::max(a: FuncMaxCount, b: *BI->Count);
1675 }
1676
1677 // Fix the obviously inconsistent entry count.
1678 if (FuncMaxCount > 0 && FuncEntryCount == 0)
1679 FuncEntryCount = 1;
1680 F.setEntryCount(Count: ProfileCount(FuncEntryCount, Function::PCT_Real));
1681 markFunctionAttributes(EntryCount: FuncEntryCount, MaxCount: FuncMaxCount);
1682
1683 LLVM_DEBUG(FuncInfo.dumpInfo("after reading profile."));
1684}
1685
1686// Assign the scaled count values to the BB with multiple out edges.
1687void PGOUseFunc::setBranchWeights() {
1688 // Generate MD_prof metadata for every branch instruction.
1689 LLVM_DEBUG(dbgs() << "\nSetting branch weights for func " << F.getName()
1690 << " IsCS=" << IsCS << "\n");
1691 for (auto &BB : F) {
1692 Instruction *TI = BB.getTerminator();
1693 if (TI->getNumSuccessors() < 2)
1694 continue;
1695 if (!(isa<BranchInst>(Val: TI) || isa<SwitchInst>(Val: TI) ||
1696 isa<IndirectBrInst>(Val: TI) || isa<InvokeInst>(Val: TI) ||
1697 isa<CallBrInst>(Val: TI)))
1698 continue;
1699
1700 const PGOUseBBInfo &BBCountInfo = getBBInfo(BB: &BB);
1701 if (!*BBCountInfo.Count)
1702 continue;
1703
1704 // We have a non-zero Branch BB.
1705
1706 // SuccessorCount can be greater than OutEdgesCount, because
1707 // removed edges don't appear in OutEdges.
1708 unsigned OutEdgesCount = BBCountInfo.OutEdges.size();
1709 unsigned SuccessorCount = BB.getTerminator()->getNumSuccessors();
1710 assert(OutEdgesCount <= SuccessorCount);
1711
1712 SmallVector<uint64_t, 2> EdgeCounts(SuccessorCount, 0);
1713 uint64_t MaxCount = 0;
1714 for (unsigned It = 0; It < OutEdgesCount; It++) {
1715 const PGOUseEdge *E = BBCountInfo.OutEdges[It];
1716 const BasicBlock *SrcBB = E->SrcBB;
1717 const BasicBlock *DestBB = E->DestBB;
1718 if (DestBB == nullptr)
1719 continue;
1720 unsigned SuccNum = GetSuccessorNumber(BB: SrcBB, Succ: DestBB);
1721 uint64_t EdgeCount = *E->Count;
1722 if (EdgeCount > MaxCount)
1723 MaxCount = EdgeCount;
1724 EdgeCounts[SuccNum] = EdgeCount;
1725 }
1726
1727 if (MaxCount)
1728 setProfMetadata(TI, EdgeCounts, MaxCount);
1729 else {
1730 // A zero MaxCount can come about when we have a BB with a positive
1731 // count, and whose successor blocks all have 0 count. This can happen
1732 // when there is no exit block and the code exits via a noreturn function.
1733 auto &Ctx = M->getContext();
1734 Ctx.diagnose(DI: DiagnosticInfoPGOProfile(
1735 M->getName().data(),
1736 Twine("Profile in ") + F.getName().str() +
1737 Twine(" partially ignored") +
1738 Twine(", possibly due to the lack of a return path."),
1739 DS_Warning));
1740 }
1741 }
1742}
1743
1744static bool isIndirectBrTarget(BasicBlock *BB) {
1745 for (BasicBlock *Pred : predecessors(BB)) {
1746 if (isa<IndirectBrInst>(Val: Pred->getTerminator()))
1747 return true;
1748 }
1749 return false;
1750}
1751
1752void PGOUseFunc::annotateIrrLoopHeaderWeights() {
1753 LLVM_DEBUG(dbgs() << "\nAnnotating irreducible loop header weights.\n");
1754 // Find irr loop headers
1755 for (auto &BB : F) {
1756 // As a heuristic also annotate indrectbr targets as they have a high chance
1757 // to become an irreducible loop header after the indirectbr tail
1758 // duplication.
1759 if (BFI->isIrrLoopHeader(BB: &BB) || isIndirectBrTarget(BB: &BB)) {
1760 Instruction *TI = BB.getTerminator();
1761 const PGOUseBBInfo &BBCountInfo = getBBInfo(BB: &BB);
1762 setIrrLoopHeaderMetadata(M, TI, Count: *BBCountInfo.Count);
1763 }
1764 }
1765}
1766
1767void SelectInstVisitor::instrumentOneSelectInst(SelectInst &SI) {
1768 Module *M = F.getParent();
1769 IRBuilder<> Builder(&SI);
1770 Type *Int64Ty = Builder.getInt64Ty();
1771 auto *Step = Builder.CreateZExt(V: SI.getCondition(), DestTy: Int64Ty);
1772 auto *NormalizedFuncNameVarPtr =
1773 ConstantExpr::getPointerBitCastOrAddrSpaceCast(
1774 C: FuncNameVar, Ty: PointerType::get(C&: M->getContext(), AddressSpace: 0));
1775 Builder.CreateIntrinsic(ID: Intrinsic::instrprof_increment_step,
1776 Args: {NormalizedFuncNameVarPtr, Builder.getInt64(C: FuncHash),
1777 Builder.getInt32(C: TotalNumCtrs),
1778 Builder.getInt32(C: *CurCtrIdx), Step});
1779 ++(*CurCtrIdx);
1780}
1781
1782void SelectInstVisitor::annotateOneSelectInst(SelectInst &SI) {
1783 std::vector<uint64_t> &CountFromProfile = UseFunc->getProfileRecord().Counts;
1784 assert(*CurCtrIdx < CountFromProfile.size() &&
1785 "Out of bound access of counters");
1786 uint64_t SCounts[2];
1787 SCounts[0] = CountFromProfile[*CurCtrIdx]; // True count
1788 ++(*CurCtrIdx);
1789 uint64_t TotalCount = 0;
1790 auto BI = UseFunc->findBBInfo(BB: SI.getParent());
1791 if (BI != nullptr) {
1792 TotalCount = *BI->Count;
1793
1794 // Fix the block count if it is impossible.
1795 if (TotalCount < SCounts[0])
1796 BI->Count = SCounts[0];
1797 }
1798 // False Count
1799 SCounts[1] = (TotalCount > SCounts[0] ? TotalCount - SCounts[0] : 0);
1800 uint64_t MaxCount = std::max(a: SCounts[0], b: SCounts[1]);
1801 if (MaxCount)
1802 setProfMetadata(TI: &SI, EdgeCounts: SCounts, MaxCount);
1803}
1804
1805void SelectInstVisitor::visitSelectInst(SelectInst &SI) {
1806 if (!PGOInstrSelect || PGOFunctionEntryCoverage || HasSingleByteCoverage)
1807 return;
1808 // FIXME: do not handle this yet.
1809 if (SI.getCondition()->getType()->isVectorTy())
1810 return;
1811
1812 switch (Mode) {
1813 case VM_counting:
1814 NSIs++;
1815 return;
1816 case VM_instrument:
1817 instrumentOneSelectInst(SI);
1818 return;
1819 case VM_annotate:
1820 annotateOneSelectInst(SI);
1821 return;
1822 }
1823
1824 llvm_unreachable("Unknown visiting mode");
1825}
1826
1827static uint32_t getMaxNumAnnotations(InstrProfValueKind ValueProfKind) {
1828 if (ValueProfKind == IPVK_MemOPSize)
1829 return MaxNumMemOPAnnotations;
1830 if (ValueProfKind == llvm::IPVK_VTableTarget)
1831 return MaxNumVTableAnnotations;
1832 return MaxNumAnnotations;
1833}
1834
1835// Traverse all valuesites and annotate the instructions for all value kind.
1836void PGOUseFunc::annotateValueSites() {
1837 if (DisableValueProfiling)
1838 return;
1839
1840 // Create the PGOFuncName meta data.
1841 createPGOFuncNameMetadata(F, PGOFuncName: FuncInfo.FuncName);
1842
1843 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind)
1844 annotateValueSites(Kind);
1845}
1846
1847// Annotate the instructions for a specific value kind.
1848void PGOUseFunc::annotateValueSites(uint32_t Kind) {
1849 assert(Kind <= IPVK_Last);
1850 unsigned ValueSiteIndex = 0;
1851
1852 unsigned NumValueSites = ProfileRecord.getNumValueSites(ValueKind: Kind);
1853
1854 // Since there isn't a reliable or fast way for profile reader to tell if a
1855 // profile is generated with `-enable-vtable-value-profiling` on, we run the
1856 // value profile collector over the function IR to find the instrumented sites
1857 // iff function profile records shows the number of instrumented vtable sites
1858 // is not zero. Function cfg already takes the number of instrumented
1859 // indirect call sites into account so it doesn't hash the number of
1860 // instrumented vtables; as a side effect it makes it easier to enable
1861 // profiling and profile use in two steps if needed.
1862 // TODO: Remove this if/when -enable-vtable-value-profiling is on by default.
1863 if (NumValueSites > 0 && Kind == IPVK_VTableTarget &&
1864 NumValueSites != FuncInfo.ValueSites[IPVK_VTableTarget].size() &&
1865 MaxNumVTableAnnotations != 0)
1866 FuncInfo.ValueSites[IPVK_VTableTarget] = VPC.get(Kind: IPVK_VTableTarget);
1867 auto &ValueSites = FuncInfo.ValueSites[Kind];
1868 if (NumValueSites != ValueSites.size()) {
1869 auto &Ctx = M->getContext();
1870 Ctx.diagnose(DI: DiagnosticInfoPGOProfile(
1871 M->getName().data(),
1872 Twine("Inconsistent number of value sites for ") +
1873 Twine(ValueProfKindDescr[Kind]) + Twine(" profiling in \"") +
1874 F.getName().str() +
1875 Twine("\", possibly due to the use of a stale profile."),
1876 DS_Warning));
1877 return;
1878 }
1879
1880 for (VPCandidateInfo &I : ValueSites) {
1881 LLVM_DEBUG(dbgs() << "Read one value site profile (kind = " << Kind
1882 << "): Index = " << ValueSiteIndex << " out of "
1883 << NumValueSites << "\n");
1884 annotateValueSite(
1885 M&: *M, Inst&: *I.AnnotatedInst, InstrProfR: ProfileRecord,
1886 ValueKind: static_cast<InstrProfValueKind>(Kind), SiteIndx: ValueSiteIndex,
1887 MaxMDCount: getMaxNumAnnotations(ValueProfKind: static_cast<InstrProfValueKind>(Kind)));
1888 ValueSiteIndex++;
1889 }
1890}
1891
1892// Collect the set of members for each Comdat in module M and store
1893// in ComdatMembers.
1894static void collectComdatMembers(
1895 Module &M,
1896 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) {
1897 if (!DoComdatRenaming)
1898 return;
1899 for (Function &F : M)
1900 if (Comdat *C = F.getComdat())
1901 ComdatMembers.insert(x: std::make_pair(x&: C, y: &F));
1902 for (GlobalVariable &GV : M.globals())
1903 if (Comdat *C = GV.getComdat())
1904 ComdatMembers.insert(x: std::make_pair(x&: C, y: &GV));
1905 for (GlobalAlias &GA : M.aliases())
1906 if (Comdat *C = GA.getComdat())
1907 ComdatMembers.insert(x: std::make_pair(x&: C, y: &GA));
1908}
1909
1910// Return true if we should not find instrumentation data for this function
1911static bool skipPGOUse(const Function &F) {
1912 if (F.isDeclaration())
1913 return true;
1914 // If there are too many critical edges, PGO might cause
1915 // compiler time problem. Skip PGO if the number of
1916 // critical edges execeed the threshold.
1917 unsigned NumCriticalEdges = 0;
1918 for (auto &BB : F) {
1919 const Instruction *TI = BB.getTerminator();
1920 for (unsigned I = 0, E = TI->getNumSuccessors(); I != E; ++I) {
1921 if (isCriticalEdge(TI, SuccNum: I))
1922 NumCriticalEdges++;
1923 }
1924 }
1925 if (NumCriticalEdges > PGOFunctionCriticalEdgeThreshold) {
1926 LLVM_DEBUG(dbgs() << "In func " << F.getName()
1927 << ", NumCriticalEdges=" << NumCriticalEdges
1928 << " exceed the threshold. Skip PGO.\n");
1929 return true;
1930 }
1931 return false;
1932}
1933
1934// Return true if we should not instrument this function
1935static bool skipPGOGen(const Function &F) {
1936 if (skipPGOUse(F))
1937 return true;
1938 if (F.hasFnAttribute(Kind: llvm::Attribute::Naked))
1939 return true;
1940 if (F.hasFnAttribute(Kind: llvm::Attribute::NoProfile))
1941 return true;
1942 if (F.hasFnAttribute(Kind: llvm::Attribute::SkipProfile))
1943 return true;
1944 if (F.getInstructionCount() < PGOFunctionSizeThreshold)
1945 return true;
1946 if (PGOInstrumentColdFunctionOnly) {
1947 if (auto EntryCount = F.getEntryCount())
1948 return EntryCount->getCount() > PGOColdInstrumentEntryThreshold;
1949 return !PGOTreatUnknownAsCold;
1950 }
1951 return false;
1952}
1953
1954static bool InstrumentAllFunctions(
1955 Module &M, function_ref<TargetLibraryInfo &(Function &)> LookupTLI,
1956 function_ref<BranchProbabilityInfo *(Function &)> LookupBPI,
1957 function_ref<BlockFrequencyInfo *(Function &)> LookupBFI,
1958 function_ref<LoopInfo *(Function &)> LookupLI,
1959 PGOInstrumentationType InstrumentationType) {
1960 // For the context-sensitive instrumentation, we should have a separated pass
1961 // (before LTO/ThinLTO linking) to create these variables.
1962 if (InstrumentationType == PGOInstrumentationType::FDO)
1963 createIRLevelProfileFlagVar(M, InstrumentationType);
1964
1965 Triple TT(M.getTargetTriple());
1966 LLVMContext &Ctx = M.getContext();
1967 if (!TT.isOSBinFormatELF() && EnableVTableValueProfiling)
1968 Ctx.diagnose(DI: DiagnosticInfoPGOProfile(
1969 M.getName().data(),
1970 Twine("VTable value profiling is presently not "
1971 "supported for non-ELF object formats"),
1972 DS_Warning));
1973 std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers;
1974 collectComdatMembers(M, ComdatMembers);
1975
1976 for (auto &F : M) {
1977 if (skipPGOGen(F))
1978 continue;
1979 TargetLibraryInfo &TLI = LookupTLI(F);
1980 BranchProbabilityInfo *BPI = LookupBPI(F);
1981 BlockFrequencyInfo *BFI = LookupBFI(F);
1982 LoopInfo *LI = LookupLI(F);
1983 FunctionInstrumenter FI(M, F, TLI, ComdatMembers, BPI, BFI, LI,
1984 InstrumentationType);
1985 FI.instrument();
1986 }
1987 return true;
1988}
1989
1990PreservedAnalyses
1991PGOInstrumentationGenCreateVar::run(Module &M, ModuleAnalysisManager &MAM) {
1992 createProfileFileNameVar(M, InstrProfileOutput: CSInstrName);
1993 // The variable in a comdat may be discarded by LTO. Ensure the declaration
1994 // will be retained.
1995 appendToCompilerUsed(
1996 M, Values: createIRLevelProfileFlagVar(M, InstrumentationType: PGOInstrumentationType::CSFDO));
1997 if (ProfileSampling)
1998 createProfileSamplingVar(M);
1999 PreservedAnalyses PA;
2000 PA.preserve<FunctionAnalysisManagerModuleProxy>();
2001 PA.preserveSet<AllAnalysesOn<Function>>();
2002 return PA;
2003}
2004
2005PreservedAnalyses PGOInstrumentationGen::run(Module &M,
2006 ModuleAnalysisManager &MAM) {
2007 auto &FAM = MAM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager();
2008 auto LookupTLI = [&FAM](Function &F) -> TargetLibraryInfo & {
2009 return FAM.getResult<TargetLibraryAnalysis>(IR&: F);
2010 };
2011 auto LookupBPI = [&FAM](Function &F) {
2012 return &FAM.getResult<BranchProbabilityAnalysis>(IR&: F);
2013 };
2014 auto LookupBFI = [&FAM](Function &F) {
2015 return &FAM.getResult<BlockFrequencyAnalysis>(IR&: F);
2016 };
2017 auto LookupLI = [&FAM](Function &F) {
2018 return &FAM.getResult<LoopAnalysis>(IR&: F);
2019 };
2020
2021 if (!InstrumentAllFunctions(M, LookupTLI, LookupBPI, LookupBFI, LookupLI,
2022 InstrumentationType))
2023 return PreservedAnalyses::all();
2024
2025 return PreservedAnalyses::none();
2026}
2027
2028// Using the ratio b/w sums of profile count values and BFI count values to
2029// adjust the func entry count.
2030static void fixFuncEntryCount(PGOUseFunc &Func, LoopInfo &LI,
2031 BranchProbabilityInfo &NBPI) {
2032 Function &F = Func.getFunc();
2033 BlockFrequencyInfo NBFI(F, NBPI, LI);
2034#ifndef NDEBUG
2035 auto BFIEntryCount = F.getEntryCount();
2036 assert(BFIEntryCount && (BFIEntryCount->getCount() > 0) &&
2037 "Invalid BFI Entrycount");
2038#endif
2039 auto SumCount = APFloat::getZero(Sem: APFloat::IEEEdouble());
2040 auto SumBFICount = APFloat::getZero(Sem: APFloat::IEEEdouble());
2041 for (auto &BBI : F) {
2042 uint64_t CountValue = 0;
2043 uint64_t BFICountValue = 0;
2044 if (!Func.findBBInfo(BB: &BBI))
2045 continue;
2046 auto BFICount = NBFI.getBlockProfileCount(BB: &BBI);
2047 CountValue = *Func.getBBInfo(BB: &BBI).Count;
2048 BFICountValue = *BFICount;
2049 SumCount.add(RHS: APFloat(CountValue * 1.0), RM: APFloat::rmNearestTiesToEven);
2050 SumBFICount.add(RHS: APFloat(BFICountValue * 1.0), RM: APFloat::rmNearestTiesToEven);
2051 }
2052 if (SumCount.isZero())
2053 return;
2054
2055 assert(SumBFICount.compare(APFloat(0.0)) == APFloat::cmpGreaterThan &&
2056 "Incorrect sum of BFI counts");
2057 if (SumBFICount.compare(RHS: SumCount) == APFloat::cmpEqual)
2058 return;
2059 double Scale = (SumCount / SumBFICount).convertToDouble();
2060 if (Scale < 1.001 && Scale > 0.999)
2061 return;
2062
2063 uint64_t FuncEntryCount = *Func.getBBInfo(BB: &*F.begin()).Count;
2064 uint64_t NewEntryCount = 0.5 + FuncEntryCount * Scale;
2065 if (NewEntryCount == 0)
2066 NewEntryCount = 1;
2067 if (NewEntryCount != FuncEntryCount) {
2068 F.setEntryCount(Count: ProfileCount(NewEntryCount, Function::PCT_Real));
2069 LLVM_DEBUG(dbgs() << "FixFuncEntryCount: in " << F.getName()
2070 << ", entry_count " << FuncEntryCount << " --> "
2071 << NewEntryCount << "\n");
2072 }
2073}
2074
2075// Compare the profile count values with BFI count values, and print out
2076// the non-matching ones.
2077static void verifyFuncBFI(PGOUseFunc &Func, LoopInfo &LI,
2078 BranchProbabilityInfo &NBPI,
2079 uint64_t HotCountThreshold,
2080 uint64_t ColdCountThreshold) {
2081 Function &F = Func.getFunc();
2082 BlockFrequencyInfo NBFI(F, NBPI, LI);
2083 // bool PrintFunc = false;
2084 bool HotBBOnly = PGOVerifyHotBFI;
2085 StringRef Msg;
2086 OptimizationRemarkEmitter ORE(&F);
2087
2088 unsigned BBNum = 0, BBMisMatchNum = 0, NonZeroBBNum = 0;
2089 for (auto &BBI : F) {
2090 PGOUseBBInfo *BBInfo = Func.findBBInfo(BB: &BBI);
2091 if (!BBInfo)
2092 continue;
2093
2094 uint64_t CountValue = BBInfo->Count.value_or(u&: CountValue);
2095 uint64_t BFICountValue = 0;
2096
2097 BBNum++;
2098 if (CountValue)
2099 NonZeroBBNum++;
2100 auto BFICount = NBFI.getBlockProfileCount(BB: &BBI);
2101 if (BFICount)
2102 BFICountValue = *BFICount;
2103
2104 if (HotBBOnly) {
2105 bool rawIsHot = CountValue >= HotCountThreshold;
2106 bool BFIIsHot = BFICountValue >= HotCountThreshold;
2107 bool rawIsCold = CountValue <= ColdCountThreshold;
2108 bool ShowCount = false;
2109 if (rawIsHot && !BFIIsHot) {
2110 Msg = "raw-Hot to BFI-nonHot";
2111 ShowCount = true;
2112 } else if (rawIsCold && BFIIsHot) {
2113 Msg = "raw-Cold to BFI-Hot";
2114 ShowCount = true;
2115 }
2116 if (!ShowCount)
2117 continue;
2118 } else {
2119 if ((CountValue < PGOVerifyBFICutoff) &&
2120 (BFICountValue < PGOVerifyBFICutoff))
2121 continue;
2122 uint64_t Diff = (BFICountValue >= CountValue)
2123 ? BFICountValue - CountValue
2124 : CountValue - BFICountValue;
2125 if (Diff <= CountValue / 100 * PGOVerifyBFIRatio)
2126 continue;
2127 }
2128 BBMisMatchNum++;
2129
2130 ORE.emit(RemarkBuilder: [&]() {
2131 OptimizationRemarkAnalysis Remark(DEBUG_TYPE, "bfi-verify",
2132 F.getSubprogram(), &BBI);
2133 Remark << "BB " << ore::NV("Block", BBI.getName())
2134 << " Count=" << ore::NV("Count", CountValue)
2135 << " BFI_Count=" << ore::NV("Count", BFICountValue);
2136 if (!Msg.empty())
2137 Remark << " (" << Msg << ")";
2138 return Remark;
2139 });
2140 }
2141 if (BBMisMatchNum)
2142 ORE.emit(RemarkBuilder: [&]() {
2143 return OptimizationRemarkAnalysis(DEBUG_TYPE, "bfi-verify",
2144 F.getSubprogram(), &F.getEntryBlock())
2145 << "In Func " << ore::NV("Function", F.getName())
2146 << ": Num_of_BB=" << ore::NV("Count", BBNum)
2147 << ", Num_of_non_zerovalue_BB=" << ore::NV("Count", NonZeroBBNum)
2148 << ", Num_of_mis_matching_BB=" << ore::NV("Count", BBMisMatchNum);
2149 });
2150}
2151
2152static bool annotateAllFunctions(
2153 Module &M, StringRef ProfileFileName, StringRef ProfileRemappingFileName,
2154 vfs::FileSystem &FS,
2155 function_ref<TargetLibraryInfo &(Function &)> LookupTLI,
2156 function_ref<BranchProbabilityInfo *(Function &)> LookupBPI,
2157 function_ref<BlockFrequencyInfo *(Function &)> LookupBFI,
2158 function_ref<LoopInfo *(Function &)> LookupLI, ProfileSummaryInfo *PSI,
2159 bool IsCS) {
2160 LLVM_DEBUG(dbgs() << "Read in profile counters: ");
2161 auto &Ctx = M.getContext();
2162 // Read the counter array from file.
2163 auto ReaderOrErr = IndexedInstrProfReader::create(Path: ProfileFileName, FS,
2164 RemappingPath: ProfileRemappingFileName);
2165 if (Error E = ReaderOrErr.takeError()) {
2166 handleAllErrors(E: std::move(E), Handlers: [&](const ErrorInfoBase &EI) {
2167 Ctx.diagnose(
2168 DI: DiagnosticInfoPGOProfile(ProfileFileName.data(), EI.message()));
2169 });
2170 return false;
2171 }
2172
2173 std::unique_ptr<IndexedInstrProfReader> PGOReader =
2174 std::move(ReaderOrErr.get());
2175 if (!PGOReader) {
2176 Ctx.diagnose(DI: DiagnosticInfoPGOProfile(ProfileFileName.data(),
2177 StringRef("Cannot get PGOReader")));
2178 return false;
2179 }
2180 if (!PGOReader->hasCSIRLevelProfile() && IsCS)
2181 return false;
2182
2183 // TODO: might need to change the warning once the clang option is finalized.
2184 if (!PGOReader->isIRLevelProfile()) {
2185 Ctx.diagnose(DI: DiagnosticInfoPGOProfile(
2186 ProfileFileName.data(), "Not an IR level instrumentation profile"));
2187 return false;
2188 }
2189 if (PGOReader->functionEntryOnly()) {
2190 Ctx.diagnose(DI: DiagnosticInfoPGOProfile(
2191 ProfileFileName.data(),
2192 "Function entry profiles are not yet supported for optimization"));
2193 return false;
2194 }
2195
2196 if (EnableVTableProfileUse) {
2197 for (GlobalVariable &G : M.globals()) {
2198 if (!G.hasName() || !G.hasMetadata(KindID: LLVMContext::MD_type))
2199 continue;
2200
2201 // Create the PGOFuncName meta data.
2202 createPGONameMetadata(GO&: G, PGOName: getPGOName(V: G, InLTO: false /* InLTO*/));
2203 }
2204 }
2205
2206 // Add the profile summary (read from the header of the indexed summary) here
2207 // so that we can use it below when reading counters (which checks if the
2208 // function should be marked with a cold or inlinehint attribute).
2209 M.setProfileSummary(M: PGOReader->getSummary(UseCS: IsCS).getMD(Context&: M.getContext()),
2210 Kind: IsCS ? ProfileSummary::PSK_CSInstr
2211 : ProfileSummary::PSK_Instr);
2212 PSI->refresh();
2213
2214 std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers;
2215 collectComdatMembers(M, ComdatMembers);
2216 std::vector<Function *> HotFunctions;
2217 std::vector<Function *> ColdFunctions;
2218
2219 // If the profile marked as always instrument the entry BB, do the
2220 // same. Note this can be overwritten by the internal option in CFGMST.h
2221 bool InstrumentFuncEntry = PGOReader->instrEntryBBEnabled();
2222 if (PGOInstrumentEntry.getNumOccurrences() > 0)
2223 InstrumentFuncEntry = PGOInstrumentEntry;
2224 bool InstrumentLoopEntries = PGOReader->instrLoopEntriesEnabled();
2225 if (PGOInstrumentLoopEntries.getNumOccurrences() > 0)
2226 InstrumentLoopEntries = PGOInstrumentLoopEntries;
2227
2228 bool HasSingleByteCoverage = PGOReader->hasSingleByteCoverage();
2229 for (auto &F : M) {
2230 if (skipPGOUse(F))
2231 continue;
2232 TargetLibraryInfo &TLI = LookupTLI(F);
2233 BranchProbabilityInfo *BPI = LookupBPI(F);
2234 BlockFrequencyInfo *BFI = LookupBFI(F);
2235 LoopInfo *LI = LookupLI(F);
2236 if (!HasSingleByteCoverage) {
2237 // Split indirectbr critical edges here before computing the MST rather
2238 // than later in getInstrBB() to avoid invalidating it.
2239 SplitIndirectBrCriticalEdges(F, /*IgnoreBlocksWithoutPHI=*/false, BPI,
2240 BFI);
2241 }
2242 PGOUseFunc Func(F, &M, TLI, ComdatMembers, BPI, BFI, LI, PSI, IsCS,
2243 InstrumentFuncEntry, InstrumentLoopEntries,
2244 HasSingleByteCoverage);
2245 if (!Func.getRecord(PGOReader: PGOReader.get()))
2246 continue;
2247 if (HasSingleByteCoverage) {
2248 Func.populateCoverage();
2249 continue;
2250 }
2251 // When PseudoKind is set to a value other than InstrProfRecord::NotPseudo,
2252 // it means the profile for the function is unrepresentative and this
2253 // function is actually hot / warm. We will reset the function hot / cold
2254 // attribute and drop all the profile counters.
2255 InstrProfRecord::CountPseudoKind PseudoKind = InstrProfRecord::NotPseudo;
2256 bool AllZeros = false;
2257 if (!Func.readCounters(AllZeros, PseudoKind))
2258 continue;
2259 if (AllZeros) {
2260 F.setEntryCount(Count: ProfileCount(0, Function::PCT_Real));
2261 if (Func.getProgramMaxCount() != 0)
2262 ColdFunctions.push_back(x: &F);
2263 continue;
2264 }
2265 if (PseudoKind != InstrProfRecord::NotPseudo) {
2266 // Clear function attribute cold.
2267 if (F.hasFnAttribute(Kind: Attribute::Cold))
2268 F.removeFnAttr(Kind: Attribute::Cold);
2269 // Set function attribute as hot.
2270 if (PseudoKind == InstrProfRecord::PseudoHot)
2271 F.addFnAttr(Kind: Attribute::Hot);
2272 continue;
2273 }
2274 Func.populateCounters();
2275 Func.setBranchWeights();
2276 Func.annotateValueSites();
2277 Func.annotateIrrLoopHeaderWeights();
2278 PGOUseFunc::FuncFreqAttr FreqAttr = Func.getFuncFreqAttr();
2279 if (FreqAttr == PGOUseFunc::FFA_Cold)
2280 ColdFunctions.push_back(x: &F);
2281 else if (FreqAttr == PGOUseFunc::FFA_Hot)
2282 HotFunctions.push_back(x: &F);
2283 if (PGOViewCounts != PGOVCT_None &&
2284 (ViewBlockFreqFuncName.empty() ||
2285 F.getName() == ViewBlockFreqFuncName)) {
2286 LoopInfo LI{DominatorTree(F)};
2287 std::unique_ptr<BranchProbabilityInfo> NewBPI =
2288 std::make_unique<BranchProbabilityInfo>(args&: F, args&: LI);
2289 std::unique_ptr<BlockFrequencyInfo> NewBFI =
2290 std::make_unique<BlockFrequencyInfo>(args&: F, args&: *NewBPI, args&: LI);
2291 if (PGOViewCounts == PGOVCT_Graph)
2292 NewBFI->view();
2293 else if (PGOViewCounts == PGOVCT_Text) {
2294 dbgs() << "pgo-view-counts: " << Func.getFunc().getName() << "\n";
2295 NewBFI->print(OS&: dbgs());
2296 }
2297 }
2298 if (PGOViewRawCounts != PGOVCT_None &&
2299 (ViewBlockFreqFuncName.empty() ||
2300 F.getName() == ViewBlockFreqFuncName)) {
2301 if (PGOViewRawCounts == PGOVCT_Graph)
2302 if (ViewBlockFreqFuncName.empty())
2303 WriteGraph(G: &Func, Name: Twine("PGORawCounts_") + Func.getFunc().getName());
2304 else
2305 ViewGraph(G: &Func, Name: Twine("PGORawCounts_") + Func.getFunc().getName());
2306 else if (PGOViewRawCounts == PGOVCT_Text) {
2307 dbgs() << "pgo-view-raw-counts: " << Func.getFunc().getName() << "\n";
2308 Func.dumpInfo();
2309 }
2310 }
2311
2312 if (PGOVerifyBFI || PGOVerifyHotBFI || PGOFixEntryCount) {
2313 LoopInfo LI{DominatorTree(F)};
2314 BranchProbabilityInfo NBPI(F, LI);
2315
2316 // Fix func entry count.
2317 if (PGOFixEntryCount)
2318 fixFuncEntryCount(Func, LI, NBPI);
2319
2320 // Verify BlockFrequency information.
2321 uint64_t HotCountThreshold = 0, ColdCountThreshold = 0;
2322 if (PGOVerifyHotBFI) {
2323 HotCountThreshold = PSI->getOrCompHotCountThreshold();
2324 ColdCountThreshold = PSI->getOrCompColdCountThreshold();
2325 }
2326 verifyFuncBFI(Func, LI, NBPI, HotCountThreshold, ColdCountThreshold);
2327 }
2328 }
2329
2330 // Set function hotness attribute from the profile.
2331 // We have to apply these attributes at the end because their presence
2332 // can affect the BranchProbabilityInfo of any callers, resulting in an
2333 // inconsistent MST between prof-gen and prof-use.
2334 for (auto &F : HotFunctions) {
2335 F->addFnAttr(Kind: Attribute::InlineHint);
2336 LLVM_DEBUG(dbgs() << "Set inline attribute to function: " << F->getName()
2337 << "\n");
2338 }
2339 for (auto &F : ColdFunctions) {
2340 // Only set when there is no Attribute::Hot set by the user. For Hot
2341 // attribute, user's annotation has the precedence over the profile.
2342 if (F->hasFnAttribute(Kind: Attribute::Hot)) {
2343 auto &Ctx = M.getContext();
2344 std::string Msg = std::string("Function ") + F->getName().str() +
2345 std::string(" is annotated as a hot function but"
2346 " the profile is cold");
2347 Ctx.diagnose(
2348 DI: DiagnosticInfoPGOProfile(M.getName().data(), Msg, DS_Warning));
2349 continue;
2350 }
2351 F->addFnAttr(Kind: Attribute::Cold);
2352 LLVM_DEBUG(dbgs() << "Set cold attribute to function: " << F->getName()
2353 << "\n");
2354 }
2355 return true;
2356}
2357
2358PGOInstrumentationUse::PGOInstrumentationUse(
2359 std::string Filename, std::string RemappingFilename, bool IsCS,
2360 IntrusiveRefCntPtr<vfs::FileSystem> VFS)
2361 : ProfileFileName(std::move(Filename)),
2362 ProfileRemappingFileName(std::move(RemappingFilename)), IsCS(IsCS),
2363 FS(std::move(VFS)) {
2364 if (!PGOTestProfileFile.empty())
2365 ProfileFileName = PGOTestProfileFile;
2366 if (!PGOTestProfileRemappingFile.empty())
2367 ProfileRemappingFileName = PGOTestProfileRemappingFile;
2368 if (!FS)
2369 FS = vfs::getRealFileSystem();
2370}
2371
2372PreservedAnalyses PGOInstrumentationUse::run(Module &M,
2373 ModuleAnalysisManager &MAM) {
2374
2375 auto &FAM = MAM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager();
2376 auto LookupTLI = [&FAM](Function &F) -> TargetLibraryInfo & {
2377 return FAM.getResult<TargetLibraryAnalysis>(IR&: F);
2378 };
2379 auto LookupBPI = [&FAM](Function &F) {
2380 return &FAM.getResult<BranchProbabilityAnalysis>(IR&: F);
2381 };
2382 auto LookupBFI = [&FAM](Function &F) {
2383 return &FAM.getResult<BlockFrequencyAnalysis>(IR&: F);
2384 };
2385 auto LookupLI = [&FAM](Function &F) {
2386 return &FAM.getResult<LoopAnalysis>(IR&: F);
2387 };
2388
2389 auto *PSI = &MAM.getResult<ProfileSummaryAnalysis>(IR&: M);
2390 if (!annotateAllFunctions(M, ProfileFileName, ProfileRemappingFileName, FS&: *FS,
2391 LookupTLI, LookupBPI, LookupBFI, LookupLI, PSI,
2392 IsCS))
2393 return PreservedAnalyses::all();
2394
2395 return PreservedAnalyses::none();
2396}
2397
2398static std::string getSimpleNodeName(const BasicBlock *Node) {
2399 if (!Node->getName().empty())
2400 return Node->getName().str();
2401
2402 std::string SimpleNodeName;
2403 raw_string_ostream OS(SimpleNodeName);
2404 Node->printAsOperand(O&: OS, PrintType: false);
2405 return SimpleNodeName;
2406}
2407
2408void llvm::setProfMetadata(Instruction *TI, ArrayRef<uint64_t> EdgeCounts,
2409 uint64_t MaxCount) {
2410 auto Weights = downscaleWeights(Weights: EdgeCounts, KnownMaxCount: MaxCount);
2411
2412 LLVM_DEBUG(dbgs() << "Weight is: "; for (const auto &W
2413 : Weights) {
2414 dbgs() << W << " ";
2415 } dbgs() << "\n";);
2416
2417 misexpect::checkExpectAnnotations(I: *TI, ExistingWeights: Weights, /*IsFrontend=*/false);
2418
2419 setBranchWeights(I&: *TI, Weights, /*IsExpected=*/false);
2420 if (EmitBranchProbability) {
2421 std::string BrCondStr = getBranchCondString(TI);
2422 if (BrCondStr.empty())
2423 return;
2424
2425 uint64_t WSum =
2426 std::accumulate(first: Weights.begin(), last: Weights.end(), init: (uint64_t)0,
2427 binary_op: [](uint64_t w1, uint64_t w2) { return w1 + w2; });
2428 uint64_t TotalCount =
2429 std::accumulate(first: EdgeCounts.begin(), last: EdgeCounts.end(), init: (uint64_t)0,
2430 binary_op: [](uint64_t c1, uint64_t c2) { return c1 + c2; });
2431 uint64_t Scale = calculateCountScale(MaxCount: WSum);
2432 BranchProbability BP(scaleBranchCount(Count: Weights[0], Scale),
2433 scaleBranchCount(Count: WSum, Scale));
2434 std::string BranchProbStr;
2435 raw_string_ostream OS(BranchProbStr);
2436 OS << BP;
2437 OS << " (total count : " << TotalCount << ")";
2438 Function *F = TI->getParent()->getParent();
2439 OptimizationRemarkEmitter ORE(F);
2440 ORE.emit(RemarkBuilder: [&]() {
2441 return OptimizationRemark(DEBUG_TYPE, "pgo-instrumentation", TI)
2442 << BrCondStr << " is true with probability : " << BranchProbStr;
2443 });
2444 }
2445}
2446
2447namespace llvm {
2448
2449void setIrrLoopHeaderMetadata(Module *M, Instruction *TI, uint64_t Count) {
2450 MDBuilder MDB(M->getContext());
2451 TI->setMetadata(KindID: llvm::LLVMContext::MD_irr_loop,
2452 Node: MDB.createIrrLoopHeaderWeight(Weight: Count));
2453}
2454
2455template <> struct GraphTraits<PGOUseFunc *> {
2456 using NodeRef = const BasicBlock *;
2457 using ChildIteratorType = const_succ_iterator;
2458 using nodes_iterator = pointer_iterator<Function::const_iterator>;
2459
2460 static NodeRef getEntryNode(const PGOUseFunc *G) {
2461 return &G->getFunc().front();
2462 }
2463
2464 static ChildIteratorType child_begin(const NodeRef N) {
2465 return succ_begin(BB: N);
2466 }
2467
2468 static ChildIteratorType child_end(const NodeRef N) { return succ_end(BB: N); }
2469
2470 static nodes_iterator nodes_begin(const PGOUseFunc *G) {
2471 return nodes_iterator(G->getFunc().begin());
2472 }
2473
2474 static nodes_iterator nodes_end(const PGOUseFunc *G) {
2475 return nodes_iterator(G->getFunc().end());
2476 }
2477};
2478
2479template <> struct DOTGraphTraits<PGOUseFunc *> : DefaultDOTGraphTraits {
2480 explicit DOTGraphTraits(bool isSimple = false)
2481 : DefaultDOTGraphTraits(isSimple) {}
2482
2483 static std::string getGraphName(const PGOUseFunc *G) {
2484 return std::string(G->getFunc().getName());
2485 }
2486
2487 std::string getNodeLabel(const BasicBlock *Node, const PGOUseFunc *Graph) {
2488 std::string Result;
2489 raw_string_ostream OS(Result);
2490
2491 OS << getSimpleNodeName(Node) << ":\\l";
2492 PGOUseBBInfo *BI = Graph->findBBInfo(BB: Node);
2493 OS << "Count : ";
2494 if (BI && BI->Count)
2495 OS << *BI->Count << "\\l";
2496 else
2497 OS << "Unknown\\l";
2498
2499 if (!PGOInstrSelect)
2500 return Result;
2501
2502 for (const Instruction &I : *Node) {
2503 if (!isa<SelectInst>(Val: &I))
2504 continue;
2505 // Display scaled counts for SELECT instruction:
2506 OS << "SELECT : { T = ";
2507 uint64_t TC, FC;
2508 bool HasProf = extractBranchWeights(I, TrueVal&: TC, FalseVal&: FC);
2509 if (!HasProf)
2510 OS << "Unknown, F = Unknown }\\l";
2511 else
2512 OS << TC << ", F = " << FC << " }\\l";
2513 }
2514 return Result;
2515 }
2516};
2517
2518} // end namespace llvm
2519