125#include <unordered_map>
133#define DEBUG_TYPE "pgo-instrumentation"
135STATISTIC(NumOfPGOInstrument,
"Number of edges instrumented.");
136STATISTIC(NumOfPGOSelectInsts,
"Number of select instruction instrumented.");
137STATISTIC(NumOfPGOMemIntrinsics,
"Number of mem intrinsics instrumented.");
140STATISTIC(NumOfPGOSplit,
"Number of critical edge splits.");
141STATISTIC(NumOfPGOFunc,
"Number of functions having valid profile counts.");
142STATISTIC(NumOfPGOMismatch,
"Number of functions having mismatch profile.");
143STATISTIC(NumOfPGOMissing,
"Number of functions without profile.");
144STATISTIC(NumOfPGOICall,
"Number of indirect call value instrumentations.");
145STATISTIC(NumOfCSPGOInstrument,
"Number of edges instrumented in CSPGO.");
147 "Number of select instruction instrumented in CSPGO.");
149 "Number of mem intrinsics instrumented in CSPGO.");
151STATISTIC(NumOfCSPGOBB,
"Number of basic-blocks in CSPGO.");
152STATISTIC(NumOfCSPGOSplit,
"Number of critical edge splits in CSPGO.");
154 "Number of functions having valid profile counts in CSPGO.");
156 "Number of functions having mismatch profile in CSPGO.");
157STATISTIC(NumOfCSPGOMissing,
"Number of functions without profile in CSPGO.");
158STATISTIC(NumCoveredBlocks,
"Number of basic blocks that were executed");
165 cl::desc(
"Specify the path of profile data file. This is"
166 "mainly for test purpose."));
170 cl::desc(
"Specify the path of profile remapping file. This is mainly for "
177 cl::desc(
"Disable Value Profiling"));
183 cl::desc(
"Max number of annotations for a single indirect "
190 cl::desc(
"Max number of preicise value annotations for a single memop"
197 cl::desc(
"Append function hash to the name of COMDAT function to avoid "
198 "function hash mismatch due to the preinliner"));
205 cl::desc(
"Use this option to turn on/off "
206 "warnings about missing profile data for "
213 cl::desc(
"Use this option to turn off/on "
214 "warnings about profile cfg mismatch."));
221 cl::desc(
"The option is used to turn on/off "
222 "warnings about hash mismatch for comdat "
223 "or weak functions."));
229 cl::desc(
"Use this option to turn on/off SELECT "
230 "instruction instrumentation. "));
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."),
248 cl::desc(
"Use this option to turn on/off "
249 "memory intrinsic size profiling."));
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"));
261 cl::desc(
"Force to instrument function entry basicblock."));
266 "Use this option to enable function entry coverage instrumentation."));
269 "pgo-block-coverage",
270 cl::desc(
"Use this option to enable basic block coverage instrumentation"));
274 cl::desc(
"Create a dot file of CFGs with block "
275 "coverage inference information"));
278 "pgo-temporal-instrumentation",
279 cl::desc(
"Use this option to enable temporal instrumentation"));
283 cl::desc(
"Fix function entry count in profile use."));
287 cl::desc(
"Print out the non-match BFI count if a hot raw profile count "
288 "becomes non-hot, or a cold raw profile count becomes hot. "
289 "The print is enabled under -Rpass-analysis=pgo, or "
290 "internal option -pass-remakrs-analysis=pgo."));
294 cl::desc(
"Print out mismatched BFI counts after setting profile metadata "
295 "The print is enabled under -Rpass-analysis=pgo, or "
296 "internal option -pass-remakrs-analysis=pgo."));
300 cl::desc(
"Set the threshold for pgo-verify-bfi: only print out "
301 "mismatched BFI if the difference percentage is greater than "
302 "this value (in percentage)."));
306 cl::desc(
"Set the threshold for pgo-verify-bfi: skip the counts whose "
307 "profile count value is below."));
312 cl::desc(
"Trace the hash of the function with this name."));
316 cl::desc(
"Do not instrument functions smaller than this threshold."));
320 cl::desc(
"Do not instrument functions with the number of critical edges "
321 " greater than this threshold."));
360 return std::string();
365 return std::string();
377 else if (CV->
isOne())
389#define VALUE_PROF_KIND(Enumerator, Value, Descr) Descr,
396 const StringRef VarName(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR));
398 uint64_t ProfileVersion = (INSTR_PROF_RAW_VERSION | VARIANT_MASK_IR_PROF);
400 ProfileVersion |= VARIANT_MASK_CSIR_PROF;
402 ProfileVersion |= VARIANT_MASK_INSTR_ENTRY;
404 ProfileVersion |= VARIANT_MASK_DBG_CORRELATE;
407 VARIANT_MASK_BYTE_COVERAGE | VARIANT_MASK_FUNCTION_ENTRY_ONLY;
409 ProfileVersion |= VARIANT_MASK_BYTE_COVERAGE;
411 ProfileVersion |= VARIANT_MASK_TEMPORAL_PROF;
416 Triple TT(M.getTargetTriple());
417 if (TT.supportsCOMDAT()) {
419 IRLevelVersionVariable->setComdat(M.getOrInsertComdat(VarName));
421 return IRLevelVersionVariable;
431enum VisitMode { VM_counting, VM_instrument, VM_annotate };
435struct SelectInstVisitor :
public InstVisitor<SelectInstVisitor> {
438 VisitMode
Mode = VM_counting;
439 unsigned *CurCtrIdx =
nullptr;
440 unsigned TotalNumCtrs = 0;
443 PGOUseFunc *UseFunc =
nullptr;
444 bool HasSingleByteCoverage;
446 SelectInstVisitor(
Function &Func,
bool HasSingleByteCoverage)
447 :
F(
Func), HasSingleByteCoverage(HasSingleByteCoverage) {}
449 void countSelects() {
459 void instrumentSelects(
unsigned *Ind,
unsigned TotalNC,
GlobalVariable *FNV,
461 Mode = VM_instrument;
463 TotalNumCtrs = TotalNC;
470 void annotateSelects(PGOUseFunc *UF,
unsigned *Ind) {
485 unsigned getNumOfSelectInsts()
const {
return NSIs; }
497 bool Removed =
false;
498 bool IsCritical =
false;
501 : SrcBB(Src), DestBB(Dest), Weight(
W) {}
504 std::string infoString()
const {
505 return (
Twine(Removed ?
"-" :
" ") + (InMST ?
" " :
"*") +
506 (IsCritical ?
"c" :
" ") +
" W=" +
Twine(Weight))
517 PGOBBInfo(
unsigned IX) : Group(this),
Index(IX) {}
520 std::string infoString()
const {
526template <
class Edge,
class BBInfo>
class FuncPGOInstrumentation {
534 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers;
538 void computeCFGHash();
539 void renameComdatFunction();
543 std::vector<std::vector<VPCandidateInfo>> ValueSites;
544 SelectInstVisitor SIVisitor;
545 std::string FuncName;
546 std::string DeprecatedFuncName;
555 const std::optional<BlockCoverageInference> BCI;
557 static std::optional<BlockCoverageInference>
558 constructBCI(
Function &Func,
bool HasSingleByteCoverage,
559 bool InstrumentFuncEntry) {
560 if (HasSingleByteCoverage)
567 void getInstrumentBBs(std::vector<BasicBlock *> &InstrumentBBs);
580 void dumpInfo(
StringRef Str =
"")
const {
582 " Hash: " +
Twine(FunctionHash) +
"\t" + Str);
585 FuncPGOInstrumentation(
587 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers,
590 bool InstrumentFuncEntry =
true,
bool HasSingleByteCoverage =
false)
591 :
F(
Func), IsCS(IsCS), ComdatMembers(ComdatMembers), VPC(
Func, TLI),
592 TLI(TLI), ValueSites(IPVK_Last + 1),
593 SIVisitor(
Func, HasSingleByteCoverage),
594 MST(
F, InstrumentFuncEntry, BPI,
BFI),
595 BCI(constructBCI(
Func, HasSingleByteCoverage, InstrumentFuncEntry)) {
597 BCI->viewBlockCoverageGraph();
599 SIVisitor.countSelects();
600 ValueSites[IPVK_MemOPSize] = VPC.
get(IPVK_MemOPSize);
602 NumOfPGOSelectInsts += SIVisitor.getNumOfSelectInsts();
603 NumOfPGOMemIntrinsics += ValueSites[IPVK_MemOPSize].size();
605 ValueSites[IPVK_IndirectCallTarget] = VPC.
get(IPVK_IndirectCallTarget);
607 ValueSites[IPVK_VTableTarget] = VPC.
get(IPVK_VTableTarget);
609 NumOfCSPGOSelectInsts += SIVisitor.getNumOfSelectInsts();
610 NumOfCSPGOMemIntrinsics += ValueSites[IPVK_MemOPSize].size();
617 if (!ComdatMembers.empty())
618 renameComdatFunction();
621 for (
const auto &E : MST.
allEdges()) {
624 IsCS ? NumOfCSPGOEdge++ : NumOfPGOEdge++;
626 IsCS ? NumOfCSPGOInstrument++ : NumOfPGOInstrument++;
639template <
class Edge,
class BBInfo>
640void FuncPGOInstrumentation<Edge, BBInfo>::computeCFGHash() {
641 std::vector<uint8_t> Indexes;
645 auto BI = findBBInfo(Succ);
649 for (
int J = 0; J < 4; J++)
650 Indexes.push_back((uint8_t)(
Index >> (J * 8)));
657 auto updateJCH = [&JCH](
uint64_t Num) {
662 updateJCH((
uint64_t)SIVisitor.getNumOfSelectInsts());
663 updateJCH((
uint64_t)ValueSites[IPVK_IndirectCallTarget].
size());
666 updateJCH(BCI->getInstrumentedBlocksHash());
676 FunctionHash &= 0x0FFFFFFFFFFFFFFF;
679 LLVM_DEBUG(
dbgs() <<
"Function Hash Computation for " <<
F.getName() <<
":\n"
680 <<
" CRC = " << JC.
getCRC()
681 <<
", Selects = " << SIVisitor.getNumOfSelectInsts()
682 <<
", Edges = " << MST.
numEdges() <<
", ICSites = "
683 << ValueSites[IPVK_IndirectCallTarget].size()
684 <<
", Memops = " << ValueSites[IPVK_MemOPSize].size()
685 <<
", High32 CRC = " << JCH.
getCRC()
686 <<
", Hash = " << FunctionHash <<
"\n";);
689 dbgs() <<
"Funcname=" <<
F.getName() <<
", Hash=" << FunctionHash
690 <<
" in building " <<
F.getParent()->getSourceFileName() <<
"\n";
696 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) {
708 for (
auto &&CM :
make_range(ComdatMembers.equal_range(
C))) {
709 assert(!isa<GlobalAlias>(CM.second));
710 Function *FM = dyn_cast<Function>(CM.second);
718template <
class Edge,
class BBInfo>
719void FuncPGOInstrumentation<Edge, BBInfo>::renameComdatFunction() {
722 std::string OrigName =
F.getName().str();
723 std::string NewFuncName =
725 F.setName(
Twine(NewFuncName));
727 FuncName =
Twine(FuncName +
"." +
Twine(FunctionHash)).
str();
733 if (!
F.hasComdat()) {
735 NewComdat =
M->getOrInsertComdat(
StringRef(NewFuncName));
737 F.setComdat(NewComdat);
742 Comdat *OrigComdat =
F.getComdat();
743 std::string NewComdatName =
745 NewComdat =
M->getOrInsertComdat(
StringRef(NewComdatName));
748 for (
auto &&CM :
make_range(ComdatMembers.equal_range(OrigComdat))) {
750 cast<Function>(CM.second)->setComdat(NewComdat);
756template <
class Edge,
class BBInfo>
757void FuncPGOInstrumentation<Edge, BBInfo>::getInstrumentBBs(
758 std::vector<BasicBlock *> &InstrumentBBs) {
761 if (BCI->shouldInstrumentBlock(BB))
762 InstrumentBBs.push_back(&BB);
767 std::vector<Edge *> EdgeList;
769 for (
const auto &E : MST.
allEdges())
770 EdgeList.push_back(E.get());
772 for (
auto &E : EdgeList) {
775 InstrumentBBs.push_back(InstrBB);
781template <
class Edge,
class BBInfo>
782BasicBlock *FuncPGOInstrumentation<Edge, BBInfo>::getInstrBB(Edge *E) {
783 if (E->InMST || E->Removed)
789 if (SrcBB ==
nullptr)
791 if (DestBB ==
nullptr)
806 return canInstrument(SrcBB);
808 return canInstrument(DestBB);
817 dbgs() <<
"Fail to split critical edge: not instrument this edge.\n");
822 IsCS ? NumOfCSPGOSplit++ : NumOfPGOSplit++;
824 <<
" --> " << getBBInfo(DestBB).
Index <<
"\n");
826 MST.
addEdge(SrcBB, InstrBB, 0);
828 Edge &NewEdge1 = MST.
addEdge(InstrBB, DestBB, 0);
829 NewEdge1.InMST =
true;
832 return canInstrument(InstrBB);
848 if (!isa<IntrinsicInst>(OrigCall)) {
851 std::optional<OperandBundleUse> ParentFunclet =
859 if (!BlockColors.
empty()) {
860 const ColorVector &CV = BlockColors.
find(OrigCall->getParent())->second;
861 assert(CV.
size() == 1 &&
"non-unique color for block!");
874 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers,
882 FuncPGOInstrumentation<PGOEdge, PGOBBInfo> FuncInfo(
886 auto Name = FuncInfo.FuncNameVar;
888 FuncInfo.FunctionHash);
890 auto &EntryBB =
F.getEntryBlock();
891 IRBuilder<> Builder(&EntryBB, EntryBB.getFirstInsertionPt());
900 std::vector<BasicBlock *> InstrumentBBs;
901 FuncInfo.getInstrumentBBs(InstrumentBBs);
902 unsigned NumCounters =
903 InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts();
918 for (
auto &Instr : BB)
919 if (
auto *CS = dyn_cast<CallBase>(&Instr)) {
920 if ((CS->getCalledFunction() &&
921 CS->getCalledFunction()->isIntrinsic()) ||
922 dyn_cast<InlineAsm>(CS->getCalledOperand()))
929 Visit([&TotalNrCallsites](
auto *) { ++TotalNrCallsites; });
933 Visit([&](
auto *CB) {
945 auto &EntryBB =
F.getEntryBlock();
946 IRBuilder<> Builder(&EntryBB, EntryBB.getFirstInsertionPt());
955 for (
auto *InstrBB : InstrumentBBs) {
958 "Cannot get the Instrumentation point");
963 ? Intrinsic::instrprof_cover
964 : Intrinsic::instrprof_increment),
969 FuncInfo.SIVisitor.instrumentSelects(&
I, NumCounters, FuncInfo.FuncNameVar,
970 FuncInfo.FunctionHash);
976 NumOfPGOICall += FuncInfo.ValueSites[IPVK_IndirectCallTarget].size();
983 if (
F.hasPersonalityFn() &&
988 for (
uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
989 unsigned SiteIndex = 0;
995 <<
" site: CallSite Index = " << SiteIndex <<
"\n");
999 "Cannot get the Instrumentation point");
1001 Value *ToProfile =
nullptr;
1002 if (Cand.V->getType()->isIntegerTy())
1004 else if (Cand.V->getType()->isPointerTy())
1006 assert(ToProfile &&
"value profiling Value is of unexpected type");
1012 {FuncInfo.FuncNameVar, Builder.
getInt64(FuncInfo.FunctionHash),
1022struct PGOUseEdge :
public PGOEdge {
1023 using PGOEdge::PGOEdge;
1025 std::optional<uint64_t> Count;
1031 std::string infoString()
const {
1033 return PGOEdge::infoString();
1034 return (
Twine(PGOEdge::infoString()) +
" Count=" +
Twine(*Count)).str();
1041struct PGOUseBBInfo :
public PGOBBInfo {
1042 std::optional<uint64_t> Count;
1043 int32_t UnknownCountInEdge = 0;
1044 int32_t UnknownCountOutEdge = 0;
1045 DirectEdges InEdges;
1046 DirectEdges OutEdges;
1048 PGOUseBBInfo(
unsigned IX) : PGOBBInfo(IX) {}
1054 std::string infoString()
const {
1056 return PGOBBInfo::infoString();
1057 return (
Twine(PGOBBInfo::infoString()) +
" Count=" +
Twine(*Count)).str();
1061 void addOutEdge(PGOUseEdge *E) {
1062 OutEdges.push_back(E);
1063 UnknownCountOutEdge++;
1067 void addInEdge(PGOUseEdge *E) {
1068 InEdges.push_back(E);
1069 UnknownCountInEdge++;
1078 for (
const auto &E : Edges) {
1092 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers,
1095 bool HasSingleByteCoverage)
1096 :
F(
Func),
M(Modu),
BFI(BFIin), PSI(PSI),
1097 FuncInfo(
Func, TLI, ComdatMembers,
false, BPI, BFIin, IsCS,
1098 InstrumentFuncEntry, HasSingleByteCoverage),
1099 FreqAttr(FFA_Normal), IsCS(IsCS) {}
1101 void handleInstrProfError(
Error Err,
uint64_t MismatchedFuncSum);
1108 void populateCounters();
1117 void annotateValueSites();
1120 void annotateValueSites(
uint32_t Kind);
1123 void annotateIrrLoopHeaderWeights();
1126 enum FuncFreqAttr { FFA_Normal, FFA_Cold, FFA_Hot };
1129 FuncFreqAttr getFuncFreqAttr()
const {
return FreqAttr; }
1138 PGOUseBBInfo &getBBInfo(
const BasicBlock *BB)
const {
1139 return FuncInfo.getBBInfo(BB);
1143 PGOUseBBInfo *findBBInfo(
const BasicBlock *BB)
const {
1144 return FuncInfo.findBBInfo(BB);
1149 void dumpInfo(
StringRef Str =
"")
const { FuncInfo.dumpInfo(Str); }
1151 uint64_t getProgramMaxCount()
const {
return ProgramMaxCount; }
1160 FuncPGOInstrumentation<PGOUseEdge, PGOUseBBInfo> FuncInfo;
1176 FuncFreqAttr FreqAttr;
1182 bool setInstrumentedCounts(
const std::vector<uint64_t> &CountFromProfile);
1195 FreqAttr = FFA_Cold;
1203 const FuncPGOInstrumentation<PGOUseEdge, PGOUseBBInfo> &FuncInfo) {
1207 for (
const auto &E : FuncInfo.MST.allEdges()) {
1212 PGOUseBBInfo &SrcInfo = FuncInfo.getBBInfo(SrcBB);
1213 PGOUseBBInfo &DestInfo = FuncInfo.getBBInfo(DestBB);
1214 SrcInfo.addOutEdge(E.get());
1215 DestInfo.addInEdge(E.get());
1221bool PGOUseFunc::setInstrumentedCounts(
1222 const std::vector<uint64_t> &CountFromProfile) {
1224 std::vector<BasicBlock *> InstrumentBBs;
1225 FuncInfo.getInstrumentBBs(InstrumentBBs);
1229 unsigned NumCounters =
1230 InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts();
1233 if (NumCounters != CountFromProfile.size()) {
1236 auto *FuncEntry = &*
F.begin();
1241 uint64_t CountValue = CountFromProfile[
I++];
1242 PGOUseBBInfo &
Info = getBBInfo(InstrBB);
1246 if (InstrBB == FuncEntry && CountValue == 0)
1248 Info.setBBInfoCount(CountValue);
1250 ProfileCountSize = CountFromProfile.size();
1254 auto setEdgeCount = [
this](PGOUseEdge *E,
uint64_t Value) ->
void {
1255 E->setEdgeCount(
Value);
1256 this->getBBInfo(E->SrcBB).UnknownCountOutEdge--;
1257 this->getBBInfo(E->DestBB).UnknownCountInEdge--;
1263 for (
const auto &E : FuncInfo.MST.allEdges()) {
1264 if (E->Removed || E->InMST)
1267 PGOUseBBInfo &SrcInfo = getBBInfo(SrcBB);
1271 if (SrcInfo.Count && SrcInfo.OutEdges.size() == 1)
1272 setEdgeCount(E.get(), *SrcInfo.Count);
1275 PGOUseBBInfo &DestInfo = getBBInfo(DestBB);
1278 if (DestInfo.Count && DestInfo.InEdges.size() == 1)
1279 setEdgeCount(E.get(), *DestInfo.Count);
1285 setEdgeCount(E.get(), 0);
1292void PGOUseFunc::setEdgeCount(DirectEdges &Edges,
uint64_t Value) {
1293 for (
auto &E : Edges) {
1296 E->setEdgeCount(
Value);
1298 getBBInfo(E->SrcBB).UnknownCountOutEdge--;
1299 getBBInfo(E->DestBB).UnknownCountInEdge--;
1307 const char MetadataName[] =
"instr_prof_hash_mismatch";
1310 auto *Existing =
F.getMetadata(LLVMContext::MD_annotation);
1312 MDTuple *Tuple = cast<MDTuple>(Existing);
1313 for (
const auto &
N : Tuple->
operands()) {
1314 if (
N.equalsStr(MetadataName))
1323 F.setMetadata(LLVMContext::MD_annotation, MD);
1326void PGOUseFunc::handleInstrProfError(
Error Err,
uint64_t MismatchedFuncSum) {
1328 auto &Ctx =
M->getContext();
1329 auto Err = IPE.
get();
1330 bool SkipWarning =
false;
1332 << FuncInfo.FuncName <<
": ");
1333 if (Err == instrprof_error::unknown_function) {
1334 IsCS ? NumOfCSPGOMissing++ : NumOfPGOMissing++;
1337 }
else if (Err == instrprof_error::hash_mismatch ||
1338 Err == instrprof_error::malformed) {
1339 IsCS ? NumOfCSPGOMismatch++ : NumOfPGOMismatch++;
1345 LLVM_DEBUG(
dbgs() <<
"hash mismatch (hash= " << FuncInfo.FunctionHash
1346 <<
" skip=" << SkipWarning <<
")");
1356 IPE.
message() + std::string(
" ") +
F.getName().str() +
1357 std::string(
" Hash = ") + std::to_string(FuncInfo.FunctionHash) +
1358 std::string(
" up to ") + std::to_string(MismatchedFuncSum) +
1359 std::string(
" count discarded");
1371 auto &Ctx =
M->getContext();
1374 FuncInfo.FuncName, FuncInfo.FunctionHash, FuncInfo.DeprecatedFuncName,
1375 &MismatchedFuncSum);
1377 handleInstrProfError(std::move(E), MismatchedFuncSum);
1380 ProfileRecord = std::move(
Result.get());
1385 std::vector<uint64_t> &CountFromProfile = ProfileRecord.
Counts;
1387 IsCS ? NumOfCSPGOFunc++ : NumOfPGOFunc++;
1391 for (
unsigned I = 0, S = CountFromProfile.size();
I < S;
I++) {
1393 ValueSum += CountFromProfile[
I];
1395 AllZeros = (ValueSum == 0);
1399 getBBInfo(
nullptr).UnknownCountOutEdge = 2;
1400 getBBInfo(
nullptr).UnknownCountInEdge = 2;
1402 if (!setInstrumentedCounts(CountFromProfile)) {
1404 dbgs() <<
"Inconsistent number of counts, skipping this function");
1406 M->getName().data(),
1407 Twine(
"Inconsistent number of counts in ") +
F.getName().str() +
1408 Twine(
": the profile may be stale or there is a function name "
1420 FuncInfo.FuncName, FuncInfo.FunctionHash, FuncInfo.DeprecatedFuncName,
1421 &MismatchedFuncSum);
1422 if (
auto Err =
Result.takeError()) {
1423 handleInstrProfError(std::move(Err), MismatchedFuncSum);
1426 IsCS ? NumOfCSPGOFunc++ : NumOfPGOFunc++;
1428 std::vector<uint64_t> &CountsFromProfile =
Result.get().Counts;
1432 if (FuncInfo.BCI->shouldInstrumentBlock(BB))
1438 InverseDependencies;
1439 for (
auto &BB :
F) {
1440 for (
auto *Dep : FuncInfo.BCI->getDependencies(BB)) {
1442 InverseDependencies[Dep].
insert(&BB);
1447 std::stack<const BasicBlock *> CoveredBlocksToProcess;
1448 for (
auto &[BB, IsCovered] : Coverage)
1450 CoveredBlocksToProcess.push(BB);
1452 while (!CoveredBlocksToProcess.empty()) {
1453 auto *CoveredBlock = CoveredBlocksToProcess.top();
1454 assert(Coverage[CoveredBlock]);
1455 CoveredBlocksToProcess.pop();
1456 for (
auto *BB : InverseDependencies[CoveredBlock]) {
1461 CoveredBlocksToProcess.push(BB);
1469 F.setEntryCount(Coverage[&
F.getEntryBlock()] ? 10000 : 0);
1470 for (
auto &BB :
F) {
1479 Weights.
push_back((Coverage[Succ] || !Coverage[&BB]) ? 1 : 0);
1480 if (Weights.
size() >= 2)
1485 unsigned NumCorruptCoverage = 0;
1490 auto IsBlockDead = [&](
const BasicBlock &BB) -> std::optional<bool> {
1491 if (
auto C =
BFI.getBlockProfileCount(&BB))
1495 LLVM_DEBUG(
dbgs() <<
"Block Coverage: (Instrumented=*, Covered=X)\n");
1496 for (
auto &BB :
F) {
1497 LLVM_DEBUG(
dbgs() << (FuncInfo.BCI->shouldInstrumentBlock(BB) ?
"* " :
" ")
1498 << (Coverage[&BB] ?
"X " :
" ") <<
" " << BB.getName()
1504 if (Coverage[&BB] == IsBlockDead(BB).value_or(
false)) {
1506 dbgs() <<
"Found inconsistent block covearge for " << BB.getName()
1507 <<
": BCI=" << (Coverage[&BB] ?
"Covered" :
"Dead") <<
" BFI="
1508 << (IsBlockDead(BB).
value() ?
"Dead" :
"Covered") <<
"\n");
1509 ++NumCorruptCoverage;
1515 auto &Ctx =
M->getContext();
1517 M->getName().data(),
1518 Twine(
"Found inconsistent block coverage for function ") +
F.getName() +
1519 " in " +
Twine(NumCorruptCoverage) +
" blocks.",
1523 FuncInfo.BCI->viewBlockCoverageGraph(&Coverage);
1528void PGOUseFunc::populateCounters() {
1529 bool Changes =
true;
1530 unsigned NumPasses = 0;
1538 PGOUseBBInfo *UseBBInfo = findBBInfo(&BB);
1539 if (UseBBInfo ==
nullptr)
1541 if (!UseBBInfo->Count) {
1542 if (UseBBInfo->UnknownCountOutEdge == 0) {
1545 }
else if (UseBBInfo->UnknownCountInEdge == 0) {
1550 if (UseBBInfo->Count) {
1551 if (UseBBInfo->UnknownCountOutEdge == 1) {
1557 if (*UseBBInfo->Count > OutSum)
1558 Total = *UseBBInfo->Count - OutSum;
1559 setEdgeCount(UseBBInfo->OutEdges,
Total);
1562 if (UseBBInfo->UnknownCountInEdge == 1) {
1565 if (*UseBBInfo->Count > InSum)
1566 Total = *UseBBInfo->Count - InSum;
1567 setEdgeCount(UseBBInfo->InEdges,
Total);
1574 LLVM_DEBUG(
dbgs() <<
"Populate counts in " << NumPasses <<
" passes.\n");
1578 for (
auto &BB :
F) {
1579 auto BI = findBBInfo(&BB);
1582 assert(BI->Count &&
"BB count is not valid");
1587 for (
auto &BB :
F) {
1588 auto BI = findBBInfo(&BB);
1591 FuncMaxCount = std::max(FuncMaxCount, *BI->Count);
1601 FuncInfo.SIVisitor.annotateSelects(
this, &CountPosition);
1602 assert(CountPosition == ProfileCountSize);
1604 LLVM_DEBUG(FuncInfo.dumpInfo(
"after reading profile."));
1608void PGOUseFunc::setBranchWeights() {
1610 LLVM_DEBUG(
dbgs() <<
"\nSetting branch weights for func " <<
F.getName()
1611 <<
" IsCS=" << IsCS <<
"\n");
1612 for (
auto &BB :
F) {
1616 if (!(isa<BranchInst>(TI) || isa<SwitchInst>(TI) ||
1617 isa<IndirectBrInst>(TI) || isa<InvokeInst>(TI) ||
1618 isa<CallBrInst>(TI)))
1621 const PGOUseBBInfo &BBCountInfo = getBBInfo(&BB);
1622 if (!*BBCountInfo.Count)
1626 unsigned Size = BBCountInfo.OutEdges.size();
1629 for (
unsigned s = 0; s <
Size; s++) {
1630 const PGOUseEdge *E = BBCountInfo.OutEdges[s];
1633 if (DestBB ==
nullptr)
1637 if (EdgeCount > MaxCount)
1638 MaxCount = EdgeCount;
1639 EdgeCounts[SuccNum] = EdgeCount;
1648 auto &Ctx =
M->getContext();
1650 M->getName().data(),
1651 Twine(
"Profile in ") +
F.getName().str() +
1652 Twine(
" partially ignored") +
1653 Twine(
", possibly due to the lack of a return path."),
1661 if (isa<IndirectBrInst>(Pred->getTerminator()))
1667void PGOUseFunc::annotateIrrLoopHeaderWeights() {
1668 LLVM_DEBUG(
dbgs() <<
"\nAnnotating irreducible loop header weights.\n");
1670 for (
auto &BB :
F) {
1676 const PGOUseBBInfo &BBCountInfo = getBBInfo(&BB);
1682void SelectInstVisitor::instrumentOneSelectInst(
SelectInst &SI) {
1686 auto *Step = Builder.CreateZExt(
SI.getCondition(), Int64Ty);
1689 {FuncNameVar, Builder.getInt64(FuncHash), Builder.getInt32(TotalNumCtrs),
1690 Builder.getInt32(*CurCtrIdx), Step});
1694void SelectInstVisitor::annotateOneSelectInst(
SelectInst &SI) {
1695 std::vector<uint64_t> &CountFromProfile = UseFunc->getProfileRecord().Counts;
1696 assert(*CurCtrIdx < CountFromProfile.size() &&
1697 "Out of bound access of counters");
1699 SCounts[0] = CountFromProfile[*CurCtrIdx];
1702 auto BI = UseFunc->findBBInfo(
SI.getParent());
1704 TotalCount = *BI->Count;
1706 SCounts[1] = (TotalCount > SCounts[0] ? TotalCount - SCounts[0] : 0);
1707 uint64_t MaxCount = std::max(SCounts[0], SCounts[1]);
1712void SelectInstVisitor::visitSelectInst(
SelectInst &SI) {
1716 if (
SI.getCondition()->getType()->isVectorTy())
1724 instrumentOneSelectInst(SI);
1727 annotateOneSelectInst(SI);
1735 if (ValueProfKind == IPVK_MemOPSize)
1737 if (ValueProfKind == llvm::IPVK_VTableTarget)
1743void PGOUseFunc::annotateValueSites() {
1751 annotateValueSites(Kind);
1755void PGOUseFunc::annotateValueSites(
uint32_t Kind) {
1756 assert(Kind <= IPVK_Last);
1757 unsigned ValueSiteIndex = 0;
1758 auto &ValueSites = FuncInfo.ValueSites[
Kind];
1760 if (NumValueSites != ValueSites.size()) {
1761 auto &Ctx =
M->getContext();
1763 M->getName().data(),
1764 Twine(
"Inconsistent number of value sites for ") +
1767 Twine(
"\", possibly due to the use of a stale profile."),
1773 LLVM_DEBUG(
dbgs() <<
"Read one value site profile (kind = " << Kind
1774 <<
"): Index = " << ValueSiteIndex <<
" out of "
1775 << NumValueSites <<
"\n");
1777 *M, *
I.AnnotatedInst, ProfileRecord,
1788 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) {
1793 ComdatMembers.insert(std::make_pair(
C, &
F));
1795 if (
Comdat *
C = GV.getComdat())
1796 ComdatMembers.insert(std::make_pair(
C, &GV));
1798 if (
Comdat *
C = GA.getComdat())
1799 ComdatMembers.insert(std::make_pair(
C, &GA));
1804 if (
F.isDeclaration())
1809 unsigned NumCriticalEdges = 0;
1810 for (
auto &BB :
F) {
1819 <<
", NumCriticalEdges=" << NumCriticalEdges
1820 <<
" exceed the threshold. Skip PGO.\n");
1830 if (
F.hasFnAttribute(llvm::Attribute::Naked))
1832 if (
F.hasFnAttribute(llvm::Attribute::NoProfile))
1834 if (
F.hasFnAttribute(llvm::Attribute::SkipProfile))
1850 Triple TT(M.getTargetTriple());
1855 Twine(
"VTable value profiling is presently not "
1856 "supported for non-ELF object formats"),
1858 std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers;
1864 auto &TLI = LookupTLI(
F);
1865 auto *BPI = LookupBPI(
F);
1866 auto *BFI = LookupBFI(
F);
1910 auto BFIEntryCount =
F.getEntryCount();
1911 assert(BFIEntryCount && (BFIEntryCount->getCount() > 0) &&
1912 "Invalid BFI Entrycount");
1916 for (
auto &BBI :
F) {
1919 if (!Func.findBBInfo(&BBI))
1922 CountValue = *Func.getBBInfo(&BBI).Count;
1923 BFICountValue = *BFICount;
1927 if (SumCount.isZero())
1931 "Incorrect sum of BFI counts");
1934 double Scale = (SumCount / SumBFICount).convertToDouble();
1935 if (Scale < 1.001 && Scale > 0.999)
1940 if (NewEntryCount == 0)
1946 << NewEntryCount <<
"\n");
1963 unsigned BBNum = 0, BBMisMatchNum = 0, NonZeroBBNum = 0;
1964 for (
auto &BBI :
F) {
1968 CountValue = Func.getBBInfo(&BBI).Count.value_or(CountValue);
1975 BFICountValue = *BFICount;
1978 bool rawIsHot = CountValue >= HotCountThreshold;
1979 bool BFIIsHot = BFICountValue >= HotCountThreshold;
1981 bool ShowCount =
false;
1982 if (rawIsHot && !BFIIsHot) {
1983 Msg =
"raw-Hot to BFI-nonHot";
1985 }
else if (rawIsCold && BFIIsHot) {
1986 Msg =
"raw-Cold to BFI-Hot";
1995 uint64_t Diff = (BFICountValue >= CountValue)
1996 ? BFICountValue - CountValue
1997 : CountValue - BFICountValue;
2005 F.getSubprogram(), &BBI);
2007 <<
" Count=" <<
ore::NV(
"Count", CountValue)
2008 <<
" BFI_Count=" <<
ore::NV(
"Count", BFICountValue);
2010 Remark <<
" (" << Msg <<
")";
2017 F.getSubprogram(), &
F.getEntryBlock())
2018 <<
"In Func " <<
ore::NV(
"Function",
F.getName())
2019 <<
": Num_of_BB=" <<
ore::NV(
"Count", BBNum)
2020 <<
", Num_of_non_zerovalue_BB=" <<
ore::NV(
"Count", NonZeroBBNum)
2021 <<
", Num_of_mis_matching_BB=" <<
ore::NV(
"Count", BBMisMatchNum);
2033 auto &Ctx = M.getContext();
2036 ProfileRemappingFileName);
2037 if (
Error E = ReaderOrErr.takeError()) {
2045 std::unique_ptr<IndexedInstrProfReader> PGOReader =
2046 std::move(ReaderOrErr.get());
2052 if (!PGOReader->hasCSIRLevelProfile() && IsCS)
2056 if (!PGOReader->isIRLevelProfile()) {
2058 ProfileFileName.
data(),
"Not an IR level instrumentation profile"));
2061 if (PGOReader->functionEntryOnly()) {
2063 ProfileFileName.
data(),
2064 "Function entry profiles are not yet supported for optimization"));
2070 if (!
G.hasName() || !
G.hasMetadata(LLVMContext::MD_type))
2081 M.setProfileSummary(PGOReader->getSummary(IsCS).getMD(M.getContext()),
2086 std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers;
2088 std::vector<Function *> HotFunctions;
2089 std::vector<Function *> ColdFunctions;
2093 bool InstrumentFuncEntry = PGOReader->instrEntryBBEnabled();
2098 bool HasSingleByteCoverage = PGOReader->hasSingleByteCoverage();
2102 auto &TLI = LookupTLI(
F);
2103 auto *BPI = LookupBPI(
F);
2104 auto *BFI = LookupBFI(
F);
2105 if (!HasSingleByteCoverage) {
2111 PGOUseFunc Func(
F, &M, TLI, ComdatMembers, BPI, BFI, PSI, IsCS,
2112 InstrumentFuncEntry, HasSingleByteCoverage);
2113 if (HasSingleByteCoverage) {
2114 Func.populateCoverage(PGOReader.get());
2122 bool AllZeros =
false;
2123 if (!Func.readCounters(PGOReader.get(), AllZeros, PseudoKind))
2127 if (Func.getProgramMaxCount() != 0)
2128 ColdFunctions.push_back(&
F);
2133 if (
F.hasFnAttribute(Attribute::Cold))
2134 F.removeFnAttr(Attribute::Cold);
2137 F.addFnAttr(Attribute::Hot);
2140 Func.populateCounters();
2141 Func.setBranchWeights();
2142 Func.annotateValueSites();
2143 Func.annotateIrrLoopHeaderWeights();
2144 PGOUseFunc::FuncFreqAttr FreqAttr = Func.getFuncFreqAttr();
2145 if (FreqAttr == PGOUseFunc::FFA_Cold)
2146 ColdFunctions.push_back(&
F);
2147 else if (FreqAttr == PGOUseFunc::FFA_Hot)
2148 HotFunctions.push_back(&
F);
2153 std::unique_ptr<BranchProbabilityInfo> NewBPI =
2154 std::make_unique<BranchProbabilityInfo>(
F, LI);
2155 std::unique_ptr<BlockFrequencyInfo> NewBFI =
2156 std::make_unique<BlockFrequencyInfo>(
F, *NewBPI, LI);
2160 dbgs() <<
"pgo-view-counts: " << Func.getFunc().getName() <<
"\n";
2161 NewBFI->print(
dbgs());
2171 ViewGraph(&Func,
Twine(
"PGORawCounts_") + Func.getFunc().getName());
2173 dbgs() <<
"pgo-view-raw-counts: " << Func.getFunc().getName() <<
"\n";
2200 for (
auto &
F : HotFunctions) {
2201 F->addFnAttr(Attribute::InlineHint);
2202 LLVM_DEBUG(
dbgs() <<
"Set inline attribute to function: " <<
F->getName()
2205 for (
auto &
F : ColdFunctions) {
2208 if (
F->hasFnAttribute(Attribute::Hot)) {
2209 auto &Ctx = M.getContext();
2210 std::string Msg = std::string(
"Function ") +
F->getName().str() +
2211 std::string(
" is annotated as a hot function but"
2212 " the profile is cold");
2217 F->addFnAttr(Attribute::Cold);
2218 LLVM_DEBUG(
dbgs() <<
"Set cold attribute to function: " <<
F->getName()
2225 std::string Filename, std::string RemappingFilename,
bool IsCS,
2227 : ProfileFileName(
std::
move(Filename)),
2228 ProfileRemappingFileName(
std::
move(RemappingFilename)), IsCS(IsCS),
2254 LookupTLI, LookupBPI, LookupBFI, PSI, IsCS))
2261 if (!
Node->getName().empty())
2262 return Node->getName().str();
2264 std::string SimpleNodeName;
2267 return SimpleNodeName;
2272 assert(MaxCount > 0 &&
"Bad max count");
2275 for (
const auto &ECI : EdgeCounts)
2288 if (BrCondStr.empty())
2300 std::string BranchProbStr;
2303 OS <<
" (total count : " << TotalCount <<
")";
2309 << BrCondStr <<
" is true with probability : " << BranchProbStr;
2328 return &
G->getFunc().front();
2351 return std::string(
G->getFunc().getName());
2359 PGOUseBBInfo *BI = Graph->findBBInfo(Node);
2361 if (BI && BI->Count)
2362 OS << *BI->Count <<
"\\l";
2370 if (!isa<SelectInst>(&
I))
2373 OS <<
"SELECT : { T = ";
2377 OS <<
"Unknown, F = Unknown }\\l";
2379 OS << TC <<
", F = " << FC <<
" }\\l";
This file implements a class to represent arbitrary precision integral constant values and operations...
This file contains the simple types necessary to represent the attributes associated with functions a...
This file finds the minimum set of blocks on a CFG that must be instrumented to infer execution cover...
Analysis containing CSE Info
#define clEnumValN(ENUMVAL, FLAGNAME, DESC)
This file contains the declarations for the subclasses of Constant, which represent the different fla...
Given that RA is a live value
static BasicBlock * getInstrBB(CFGMST< Edge, BBInfo > &MST, Edge &E, const DenseSet< const BasicBlock * > &ExecBlocks)
This file provides various utilities for inspecting and working with the control flow graph in LLVM I...
static cl::opt< unsigned > ColdCountThreshold("mfs-count-threshold", cl::desc("Minimum number of times a block must be executed to be retained."), cl::init(1), cl::Hidden)
Module.h This file contains the declarations for the Module class.
static cl::opt< bool > PGOInstrumentEntry("pgo-instrument-entry", cl::init(false), cl::Hidden, cl::desc("Force to instrument function entry basicblock."))
static cl::opt< std::string > PGOTestProfileRemappingFile("pgo-test-profile-remapping-file", cl::init(""), cl::Hidden, cl::value_desc("filename"), cl::desc("Specify the path of profile remapping file. This is mainly for " "test purpose."))
static cl::opt< bool > PGOFixEntryCount("pgo-fix-entry-count", cl::init(true), cl::Hidden, cl::desc("Fix function entry count in profile use."))
static void fixFuncEntryCount(PGOUseFunc &Func, LoopInfo &LI, BranchProbabilityInfo &NBPI)
static cl::opt< bool > PGOVerifyHotBFI("pgo-verify-hot-bfi", cl::init(false), cl::Hidden, cl::desc("Print out the non-match BFI count if a hot raw profile count " "becomes non-hot, or a cold raw profile count becomes hot. " "The print is enabled under -Rpass-analysis=pgo, or " "internal option -pass-remakrs-analysis=pgo."))
static void annotateFunctionWithHashMismatch(Function &F, LLVMContext &ctx)
cl::opt< unsigned > MaxNumVTableAnnotations
static cl::opt< bool > PGOTemporalInstrumentation("pgo-temporal-instrumentation", cl::desc("Use this option to enable temporal instrumentation"))
static cl::opt< unsigned > PGOFunctionSizeThreshold("pgo-function-size-threshold", cl::Hidden, cl::desc("Do not instrument functions smaller than this threshold."))
static cl::opt< unsigned > MaxNumAnnotations("icp-max-annotations", cl::init(3), cl::Hidden, cl::desc("Max number of annotations for a single indirect " "call callsite"))
static bool skipPGOGen(const Function &F)
static void collectComdatMembers(Module &M, std::unordered_multimap< Comdat *, GlobalValue * > &ComdatMembers)
static cl::opt< unsigned > PGOVerifyBFICutoff("pgo-verify-bfi-cutoff", cl::init(5), cl::Hidden, cl::desc("Set the threshold for pgo-verify-bfi: skip the counts whose " "profile count value is below."))
static cl::opt< std::string > PGOTraceFuncHash("pgo-trace-func-hash", cl::init("-"), cl::Hidden, cl::value_desc("function name"), cl::desc("Trace the hash of the function with this name."))
static void instrumentOneFunc(Function &F, Module *M, TargetLibraryInfo &TLI, BranchProbabilityInfo *BPI, BlockFrequencyInfo *BFI, std::unordered_multimap< Comdat *, GlobalValue * > &ComdatMembers, bool IsCS)
bool isValueProfilingDisabled()
static void populateEHOperandBundle(VPCandidateInfo &Cand, DenseMap< BasicBlock *, ColorVector > &BlockColors, SmallVectorImpl< OperandBundleDef > &OpBundles)
static cl::opt< bool > PGOInstrSelect("pgo-instr-select", cl::init(true), cl::Hidden, cl::desc("Use this option to turn on/off SELECT " "instruction instrumentation. "))
static cl::opt< bool > PGOFunctionEntryCoverage("pgo-function-entry-coverage", cl::Hidden, cl::desc("Use this option to enable function entry coverage instrumentation."))
static bool InstrumentAllFunctions(Module &M, function_ref< TargetLibraryInfo &(Function &)> LookupTLI, function_ref< BranchProbabilityInfo *(Function &)> LookupBPI, function_ref< BlockFrequencyInfo *(Function &)> LookupBFI, bool IsCS)
static void verifyFuncBFI(PGOUseFunc &Func, LoopInfo &LI, BranchProbabilityInfo &NBPI, uint64_t HotCountThreshold, uint64_t ColdCountThreshold)
static cl::opt< unsigned > PGOVerifyBFIRatio("pgo-verify-bfi-ratio", cl::init(2), cl::Hidden, cl::desc("Set the threshold for pgo-verify-bfi: only print out " "mismatched BFI if the difference percentage is greater than " "this value (in percentage)."))
static cl::opt< bool > DoComdatRenaming("do-comdat-renaming", cl::init(false), cl::Hidden, cl::desc("Append function hash to the name of COMDAT function to avoid " "function hash mismatch due to the preinliner"))
static cl::opt< unsigned > PGOFunctionCriticalEdgeThreshold("pgo-critical-edge-threshold", cl::init(20000), cl::Hidden, cl::desc("Do not instrument functions with the number of critical edges " " greater than this threshold."))
static void setupBBInfoEdges(const FuncPGOInstrumentation< PGOUseEdge, PGOUseBBInfo > &FuncInfo)
Set up InEdges/OutEdges for all BBs in the MST.
static cl::opt< std::string > PGOTestProfileFile("pgo-test-profile-file", cl::init(""), cl::Hidden, cl::value_desc("filename"), cl::desc("Specify the path of profile data file. This is" "mainly for test purpose."))
static bool skipPGOUse(const Function &F)
static bool canRenameComdat(Function &F, std::unordered_multimap< Comdat *, GlobalValue * > &ComdatMembers)
static cl::opt< bool > PGOVerifyBFI("pgo-verify-bfi", cl::init(false), cl::Hidden, cl::desc("Print out mismatched BFI counts after setting profile metadata " "The print is enabled under -Rpass-analysis=pgo, or " "internal option -pass-remakrs-analysis=pgo."))
static cl::opt< bool > PGOBlockCoverage("pgo-block-coverage", cl::desc("Use this option to enable basic block coverage instrumentation"))
static uint64_t sumEdgeCount(const ArrayRef< PGOUseEdge * > Edges)
static cl::opt< bool > PGOInstrMemOP("pgo-instr-memop", cl::init(true), cl::Hidden, cl::desc("Use this option to turn on/off " "memory intrinsic size profiling."))
static uint32_t getMaxNumAnnotations(InstrProfValueKind ValueProfKind)
Function::ProfileCount ProfileCount
static cl::opt< bool > EmitBranchProbability("pgo-emit-branch-prob", cl::init(false), cl::Hidden, cl::desc("When this option is on, the annotated " "branch probability will be emitted as " "optimization remarks: -{Rpass|" "pass-remarks}=pgo-instrumentation"))
static cl::opt< unsigned > MaxNumMemOPAnnotations("memop-max-annotations", cl::init(4), cl::Hidden, cl::desc("Max number of preicise value annotations for a single memop" "intrinsic"))
static cl::opt< bool > DisableValueProfiling("disable-vp", cl::init(false), cl::Hidden, cl::desc("Disable Value Profiling"))
static std::string getSimpleNodeName(const BasicBlock *Node)
static cl::opt< bool > PGOViewBlockCoverageGraph("pgo-view-block-coverage-graph", cl::desc("Create a dot file of CFGs with block " "coverage inference information"))
static GlobalVariable * createIRLevelProfileFlagVar(Module &M, bool IsCS)
static bool isIndirectBrTarget(BasicBlock *BB)
static std::string getBranchCondString(Instruction *TI)
bool shouldInstrumentEntryBB()
static bool annotateAllFunctions(Module &M, StringRef ProfileFileName, StringRef ProfileRemappingFileName, vfs::FileSystem &FS, function_ref< TargetLibraryInfo &(Function &)> LookupTLI, function_ref< BranchProbabilityInfo *(Function &)> LookupBPI, function_ref< BlockFrequencyInfo *(Function &)> LookupBFI, ProfileSummaryInfo *PSI, bool IsCS)
static cl::opt< PGOViewCountsType > PGOViewRawCounts("pgo-view-raw-counts", cl::Hidden, cl::desc("A boolean option to show CFG dag or text " "with raw profile counts from " "profile data. See also option " "-pgo-view-counts. To limit graph " "display to only one function, use " "filtering option -view-bfi-func-name."), cl::values(clEnumValN(PGOVCT_None, "none", "do not show."), clEnumValN(PGOVCT_Graph, "graph", "show a graph."), clEnumValN(PGOVCT_Text, "text", "show in text.")))
static const char * ValueProfKindDescr[]
This file provides the interface for IR based instrumentation passes ( (profile-gen,...
FunctionAnalysisManager FAM
ModuleAnalysisManager MAM
This header defines various interfaces for pass management in LLVM.
This file contains the declarations for profiling metadata utility functions.
const SmallVectorImpl< MachineOperand > & Cond
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
static bool isSimple(Instruction *I)
This file defines the SmallVector class.
This file defines the 'Statistic' class, which is designed to be an easy way to expose various metric...
#define STATISTIC(VARNAME, DESC)
Defines the virtual file system interface vfs::FileSystem.
void printAsOperand(OutputBuffer &OB, Prec P=Prec::Default, bool StrictlyWorse=false) const
static APFloat getZero(const fltSemantics &Sem, bool Negative=false)
Factory for Positive and Negative Zero.
Class for arbitrary precision integers.
This templated class represents "all analyses that operate over <a particular IR unit>" (e....
A container for analyses that lazily runs them and caches their results.
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...
LLVM Basic Block Representation.
const_iterator getFirstInsertionPt() const
Returns an iterator to the first instruction in this block that is suitable for inserting a non-PHI i...
const Instruction * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Analysis pass which computes BlockFrequencyInfo.
BlockFrequencyInfo pass uses BlockFrequencyInfoImpl implementation to estimate IR basic block frequen...
std::optional< uint64_t > getBlockProfileCount(const BasicBlock *BB, bool AllowSynthetic=false) const
Returns the estimated profile count of BB.
Conditional or Unconditional Branch instruction.
bool isConditional() const
Value * getCondition() const
Analysis pass which computes BranchProbabilityInfo.
Analysis providing branch probability information.
An union-find based Minimum Spanning Tree for CFG.
Edge & addEdge(BasicBlock *Src, BasicBlock *Dest, uint64_t W)
const std::vector< std::unique_ptr< Edge > > & allEdges() const
size_t bbInfoSize() const
BBInfo * findBBInfo(const BasicBlock *BB) const
BBInfo & getBBInfo(const BasicBlock *BB) const
void dumpEdges(raw_ostream &OS, const Twine &Message) const
Base class for all callable instructions (InvokeInst and CallInst) Holds everything related to callin...
Value * getCalledOperand() const
Predicate getPredicate() const
Return the predicate for this instruction.
StringRef getName() const
void setSelectionKind(SelectionKind Val)
SelectionKind getSelectionKind() const
This is the shared class of boolean and integer constants.
bool isMinusOne() const
This function will return true iff every bit in this constant is set to true.
bool isOne() const
This is just a convenience method to make client code smaller for a common case.
bool isZero() const
This is just a convenience method to make client code smaller for a common code.
static Constant * getIntegerValue(Type *Ty, const APInt &V)
Return the value for an integer or pointer constant, or a vector thereof, with the given scalar value...
iterator find(const_arg_type_t< KeyT > Val)
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Diagnostic information for the PGO profiler.
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree.
Base class for error info classes.
virtual std::string message() const
Return the error message as a string.
Lightweight error class with error context and mandatory checking.
Tagged union holding either a T or a Error.
Class to represent profile counts.
static GlobalAlias * create(Type *Ty, unsigned AddressSpace, LinkageTypes Linkage, const Twine &Name, Constant *Aliasee, Module *Parent)
If a parent module is specified, the alias is automatically inserted into the end of the specified mo...
@ HiddenVisibility
The GV is hidden.
@ ExternalLinkage
Externally visible function.
@ WeakAnyLinkage
Keep one copy of named function when linking (weak)
@ AvailableExternallyLinkage
Available for inspection, not emission.
@ LinkOnceODRLinkage
Same, but only replaced by something equivalent.
This instruction compares its operands according to the predicate given to the constructor.
Value * CreateZExtOrTrunc(Value *V, Type *DestTy, const Twine &Name="")
Create a ZExt or Trunc from the integer value V to DestTy.
BasicBlock::iterator GetInsertPoint() const
IntegerType * getInt64Ty()
Fetch the type representing a 64-bit integer.
ConstantInt * getInt64(uint64_t C)
Get a constant 64-bit value.
ConstantInt * getInt32(uint32_t C)
Get a constant 32-bit value.
Value * CreatePtrToInt(Value *V, Type *DestTy, const Twine &Name="")
CallInst * CreateCall(FunctionType *FTy, Value *Callee, ArrayRef< Value * > Args=std::nullopt, const Twine &Name="", MDNode *FPMathTag=nullptr)
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Reader for the indexed binary instrprof format.
static Expected< std::unique_ptr< IndexedInstrProfReader > > create(const Twine &Path, vfs::FileSystem &FS, const Twine &RemappingPath="")
Factory method to create an indexed reader.
Expected< InstrProfRecord > getInstrProfRecord(StringRef FuncName, uint64_t FuncHash, StringRef DeprecatedFuncName="", uint64_t *MismatchedFuncSum=nullptr)
Return the NamedInstrProfRecord associated with FuncName and FuncHash.
uint64_t getMaximumFunctionCount(bool UseCS)
Return the maximum of all known function counts.
An analysis over an "outer" IR unit that provides access to an analysis manager over an "inner" IR un...
Base class for instruction visitors.
void visit(Iterator Start, Iterator End)
RetTy visitSelectInst(SelectInst &I)
instrprof_error get() const
std::string message() const override
Return the error message as a string.
unsigned getNumSuccessors() const LLVM_READONLY
Return the number of successors that this instruction has.
bool isEHPad() const
Return true if the instruction is a variety of EH-block.
void setMetadata(unsigned KindID, MDNode *Node)
Set the metadata of the specified kind to the specified node.
A smart pointer to a reference-counted object that inherits from RefCountedBase or ThreadSafeRefCount...
void update(ArrayRef< uint8_t > Data)
This is an important class for using LLVM in a threaded context.
void diagnose(const DiagnosticInfo &DI)
Report a message to the currently installed diagnostic handler.
MDString * createString(StringRef Str)
Return the given string as metadata.
MDNode * createIrrLoopHeaderWeight(uint64_t Weight)
Return metadata containing an irreducible loop header weight.
ArrayRef< MDOperand > operands() const
static MDTuple * get(LLVMContext &Context, ArrayRef< Metadata * > MDs)
A Module instance is used to store all the information related to an LLVM module.
static bool isContextualIRPGOEnabled()
PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM)
PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM)
PGOInstrumentationUse(std::string Filename="", std::string RemappingFilename="", bool IsCS=false, IntrusiveRefCntPtr< vfs::FileSystem > FS=nullptr)
PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM)
A set of analyses that are preserved following a run of a transformation pass.
static PreservedAnalyses none()
Convenience factory function for the empty preserved set.
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
void preserveSet()
Mark an analysis set as preserved.
void preserve()
Mark an analysis as preserved.
An analysis pass based on the new PM to deliver ProfileSummaryInfo.
Analysis providing profile information.
uint64_t getOrCompColdCountThreshold() const
Returns ColdCountThreshold if set.
void refresh()
If no summary is present, attempt to refresh.
bool isColdCount(uint64_t C) const
Returns true if count C is considered cold.
bool isHotCount(uint64_t C) const
Returns true if count C is considered hot.
uint64_t getOrCompHotCountThreshold() const
Returns HotCountThreshold if set.
This class represents the LLVM 'select' instruction.
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
reference emplace_back(ArgTypes &&... Args)
void push_back(const T &Elt)
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
StringRef - Represent a constant reference to a string, i.e.
constexpr bool empty() const
empty - Check if the string is empty.
constexpr const char * data() const
data - Get a pointer to the start of the string (which may not be null terminated).
Analysis pass providing the TargetLibraryInfo.
Provides information about what library functions are available for the current target.
TinyPtrVector - This class is specialized for cases where there are normally 0 or 1 element in a vect...
Triple - Helper class for working with autoconf configuration names.
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
std::string str() const
Return the twine contents as a std::string.
The instances of the Type class are immutable: once they are created, they are never changed.
void print(raw_ostream &O, bool IsForDebug=false, bool NoDetails=false) const
Print the current type.
static IntegerType * getInt64Ty(LLVMContext &C)
Value * getOperand(unsigned i) const
Utility analysis that determines what values are worth profiling.
std::vector< CandidateInfo > get(InstrProfValueKind Kind) const
returns a list of value profiling candidates of the given kind
LLVM Value Representation.
Type * getType() const
All values are typed, get the type of this value.
An efficient, type-erasing, non-owning reference to a callable.
const ParentTy * getParent() const
A raw_ostream that writes to an std::string.
The virtual file system interface.
This provides a very simple, boring adaptor for a begin and end iterator into a range type.
This file contains the declaration of the Comdat class, which represents a single COMDAT in LLVM.
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
@ C
The default llvm calling convention, compatible with C.
Function * getDeclaration(Module *M, ID id, ArrayRef< Type * > Tys=std::nullopt)
Create or insert an LLVM Function declaration for an intrinsic, and return it.
ValuesClass values(OptsTy... Options)
Helper to build a ValuesClass by forwarding a variable number of arguments as an initializer list to ...
initializer< Ty > init(const Ty &Val)
uint64_t getFuncHash(const FuncRecordTy *Record)
Return the structural hash associated with the function.
void checkExpectAnnotations(Instruction &I, const ArrayRef< uint32_t > ExistingWeights, bool IsFrontend)
checkExpectAnnotations - compares PGO counters to the thresholds used for llvm.expect and warns if th...
DiagnosticInfoOptimizationBase::Argument NV
NodeAddr< FuncNode * > Func
void write64le(void *P, uint64_t V)
IntrusiveRefCntPtr< FileSystem > getRealFileSystem()
Gets an vfs::FileSystem for the 'real' file system, as seen by the operating system.
This is an optimization pass for GlobalISel generic memory operations.
void setIrrLoopHeaderMetadata(Module *M, Instruction *TI, uint64_t Count)
auto size(R &&Range, std::enable_if_t< std::is_base_of< std::random_access_iterator_tag, typename std::iterator_traits< decltype(Range.begin())>::iterator_category >::value, void > *=nullptr)
Get the size of a range.
std::string getPGOFuncName(const Function &F, bool InLTO=false, uint64_t Version=INSTR_PROF_INDEX_VERSION)
Please use getIRPGOFuncName for LLVM IR instrumentation.
void createPGOFuncNameMetadata(Function &F, StringRef PGOFuncName)
Create the PGOFuncName meta data if PGOFuncName is different from function's raw name.
unsigned GetSuccessorNumber(const BasicBlock *BB, const BasicBlock *Succ)
Search for the specified successor of basic block BB and return its position in the terminator instru...
std::string getIRPGOFuncName(const Function &F, bool InLTO=false)
Function::ProfileCount ProfileCount
auto successors(const MachineBasicBlock *BB)
void handleAllErrors(Error E, HandlerTs &&... Handlers)
Behaves the same as handleErrors, except that by contract all errors must be handled by the given han...
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
cl::opt< InstrProfCorrelator::ProfCorrelatorKind > ProfileCorrelate("profile-correlate", cl::desc("Use debug info or binary file to correlate profiles."), cl::init(InstrProfCorrelator::NONE), cl::values(clEnumValN(InstrProfCorrelator::NONE, "", "No profile correlation"), clEnumValN(InstrProfCorrelator::DEBUG_INFO, "debug-info", "Use debug info to correlate"), clEnumValN(InstrProfCorrelator::BINARY, "binary", "Use binary to correlate")))
DenseMap< BasicBlock *, ColorVector > colorEHFunclets(Function &F)
If an EH funclet personality is in use (see isFuncletEHPersonality), this will recompute which blocks...
void createPGONameMetadata(GlobalObject &GO, StringRef PGOName)
Create the PGOName metadata if a global object's PGO name is different from its mangled name.
cl::opt< bool > PGOWarnMissing
raw_ostream & WriteGraph(raw_ostream &O, const GraphType &G, bool ShortNames=false, const Twine &Title="")
bool SplitIndirectBrCriticalEdges(Function &F, bool IgnoreBlocksWithoutPHI, BranchProbabilityInfo *BPI=nullptr, BlockFrequencyInfo *BFI=nullptr)
cl::opt< bool > EnableVTableProfileUse("enable-vtable-profile-use", cl::init(false), cl::desc("If ThinLTO and WPD is enabled and this option is true, vtable " "profiles will be used by ICP pass for more efficient indirect " "call sequence. If false, type profiles won't be used."))
bool isScopedEHPersonality(EHPersonality Pers)
Returns true if this personality uses scope-style EH IR instructions: catchswitch,...
cl::opt< bool > DebugInfoCorrelate
OperandBundleDefT< Value * > OperandBundleDef
std::string getPGOName(const GlobalVariable &V, bool InLTO=false)
cl::opt< std::string > ViewBlockFreqFuncName("view-bfi-func-name", cl::Hidden, cl::desc("The option to specify " "the name of the function " "whose CFG will be displayed."))
GlobalVariable * createPGOFuncNameVar(Function &F, StringRef PGOFuncName)
Create and return the global variable for function name used in PGO instrumentation.
void annotateValueSite(Module &M, Instruction &Inst, const InstrProfRecord &InstrProfR, InstrProfValueKind ValueKind, uint32_t SiteIndx, uint32_t MaxMDCount=3)
Get the value profile data for value site SiteIdx from InstrProfR and annotate the instruction Inst w...
auto reverse(ContainerTy &&C)
void setBranchWeights(Instruction &I, ArrayRef< uint32_t > Weights, bool IsExpected)
Create a new branch_weights metadata node and add or overwrite a prof metadata reference to instructi...
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
EHPersonality classifyEHPersonality(const Value *Pers)
See if the given exception handling personality function is one that we understand.
cl::opt< bool > NoPGOWarnMismatch
RNSuccIterator< NodeRef, BlockT, RegionT > succ_begin(NodeRef Node)
cl::opt< PGOViewCountsType > PGOViewCounts("pgo-view-counts", cl::Hidden, cl::desc("A boolean option to show CFG dag or text with " "block profile counts and branch probabilities " "right after PGO profile annotation step. The " "profile counts are computed using branch " "probabilities from the runtime profile data and " "block frequency propagation algorithm. To view " "the raw counts from the profile, use option " "-pgo-view-raw-counts instead. To limit graph " "display to only one function, use filtering option " "-view-bfi-func-name."), cl::values(clEnumValN(PGOVCT_None, "none", "do not show."), clEnumValN(PGOVCT_Graph, "graph", "show a graph."), clEnumValN(PGOVCT_Text, "text", "show in text.")))
RNSuccIterator< NodeRef, BlockT, RegionT > succ_end(NodeRef Node)
static uint32_t scaleBranchCount(uint64_t Count, uint64_t Scale)
Scale an individual branch count.
void appendToCompilerUsed(Module &M, ArrayRef< GlobalValue * > Values)
Adds global values to the llvm.compiler.used list.
BasicBlock * SplitCriticalEdge(Instruction *TI, unsigned SuccNum, const CriticalEdgeSplittingOptions &Options=CriticalEdgeSplittingOptions(), const Twine &BBName="")
If this edge is a critical edge, insert a new node to split the critical edge.
void ViewGraph(const GraphType &G, const Twine &Name, bool ShortNames=false, const Twine &Title="", GraphProgram::Name Program=GraphProgram::DOT)
ViewGraph - Emit a dot graph, run 'dot', run gv on the postscript file, then cleanup.
bool isCriticalEdge(const Instruction *TI, unsigned SuccNum, bool AllowIdenticalEdges=false)
Return true if the specified edge is a critical edge.
static uint64_t calculateCountScale(uint64_t MaxCount)
Calculate what to divide by to scale counts.
bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken=false)
Check if we can safely rename this Comdat function.
void createProfileFileNameVar(Module &M, StringRef InstrProfileOutput)
OutputIt move(R &&Range, OutputIt Out)
Provide wrappers to std::move which take ranges instead of having to pass begin/end explicitly.
bool extractBranchWeights(const MDNode *ProfileData, SmallVectorImpl< uint32_t > &Weights)
Extract branch weights from MD_prof metadata.
auto predecessors(const MachineBasicBlock *BB)
void setProfMetadata(Module *M, Instruction *TI, ArrayRef< uint64_t > EdgeCounts, uint64_t MaxCount)
cl::opt< bool > EnableVTableValueProfiling("enable-vtable-value-profiling", cl::init(false), cl::desc("If true, the virtual table address will be instrumented to know " "the types of a C++ pointer. The information is used in indirect " "call promotion to do selective vtable-based comparison."))
SuccIterator< const Instruction, const BasicBlock > const_succ_iterator
cl::opt< bool > NoPGOWarnMismatchComdatWeak
Implement std::hash so that hash_code can be used in STL containers.
static constexpr roundingMode rmNearestTiesToEven
static const fltSemantics & IEEEdouble() LLVM_READNONE
DOTGraphTraits(bool isSimple=false)
static std::string getGraphName(const PGOUseFunc *G)
std::string getNodeLabel(const BasicBlock *Node, const PGOUseFunc *Graph)
DOTGraphTraits - Template class that can be specialized to customize how graphs are converted to 'dot...
DefaultDOTGraphTraits - This class provides the default implementations of all of the DOTGraphTraits ...
static ChildIteratorType child_end(const NodeRef N)
static NodeRef getEntryNode(const PGOUseFunc *G)
static ChildIteratorType child_begin(const NodeRef N)
static nodes_iterator nodes_end(const PGOUseFunc *G)
static nodes_iterator nodes_begin(const PGOUseFunc *G)
Profiling information for a single function.
std::vector< uint64_t > Counts
CountPseudoKind getCountPseudoKind() const
uint32_t getNumValueSites(uint32_t ValueKind) const
Return the number of instrumented sites for ValueKind.
static void setCSFlagInHash(uint64_t &FuncHash)
Instruction * AnnotatedInst