123#include <unordered_map>
131#define DEBUG_TYPE "pgo-instrumentation"
133STATISTIC(NumOfPGOInstrument,
"Number of edges instrumented.");
134STATISTIC(NumOfPGOSelectInsts,
"Number of select instruction instrumented.");
135STATISTIC(NumOfPGOMemIntrinsics,
"Number of mem intrinsics instrumented.");
138STATISTIC(NumOfPGOSplit,
"Number of critical edge splits.");
139STATISTIC(NumOfPGOFunc,
"Number of functions having valid profile counts.");
140STATISTIC(NumOfPGOMismatch,
"Number of functions having mismatch profile.");
141STATISTIC(NumOfPGOMissing,
"Number of functions without profile.");
142STATISTIC(NumOfPGOICall,
"Number of indirect call value instrumentations.");
143STATISTIC(NumOfCSPGOInstrument,
"Number of edges instrumented in CSPGO.");
145 "Number of select instruction instrumented in CSPGO.");
147 "Number of mem intrinsics instrumented in CSPGO.");
149STATISTIC(NumOfCSPGOBB,
"Number of basic-blocks in CSPGO.");
150STATISTIC(NumOfCSPGOSplit,
"Number of critical edge splits in CSPGO.");
152 "Number of functions having valid profile counts in CSPGO.");
154 "Number of functions having mismatch profile in CSPGO.");
155STATISTIC(NumOfCSPGOMissing,
"Number of functions without profile in CSPGO.");
156STATISTIC(NumCoveredBlocks,
"Number of basic blocks that were executed");
163 cl::desc(
"Specify the path of profile data file. This is"
164 "mainly for test purpose."));
168 cl::desc(
"Specify the path of profile remapping file. This is mainly for "
175 cl::desc(
"Disable Value Profiling"));
181 cl::desc(
"Max number of annotations for a single indirect "
188 cl::desc(
"Max number of preicise value annotations for a single memop"
195 cl::desc(
"Append function hash to the name of COMDAT function to avoid "
196 "function hash mismatch due to the preinliner"));
203 cl::desc(
"Use this option to turn on/off "
204 "warnings about missing profile data for "
211 cl::desc(
"Use this option to turn off/on "
212 "warnings about profile cfg mismatch."));
219 cl::desc(
"The option is used to turn on/off "
220 "warnings about hash mismatch for comdat "
221 "or weak functions."));
227 cl::desc(
"Use this option to turn on/off SELECT "
228 "instruction instrumentation. "));
233 cl::desc(
"A boolean option to show CFG dag or text "
234 "with raw profile counts from "
235 "profile data. See also option "
236 "-pgo-view-counts. To limit graph "
237 "display to only one function, use "
238 "filtering option -view-bfi-func-name."),
246 cl::desc(
"Use this option to turn on/off "
247 "memory intrinsic size profiling."));
252 cl::desc(
"When this option is on, the annotated "
253 "branch probability will be emitted as "
254 "optimization remarks: -{Rpass|"
255 "pass-remarks}=pgo-instrumentation"));
259 cl::desc(
"Force to instrument function entry basicblock."));
264 "Use this option to enable function entry coverage instrumentation."));
267 "pgo-block-coverage",
268 cl::desc(
"Use this option to enable basic block coverage instrumentation"));
272 cl::desc(
"Create a dot file of CFGs with block "
273 "coverage inference information"));
276 "pgo-temporal-instrumentation",
277 cl::desc(
"Use this option to enable temporal instrumentation"));
281 cl::desc(
"Fix function entry count in profile use."));
285 cl::desc(
"Print out the non-match BFI count if a hot raw profile count "
286 "becomes non-hot, or a cold raw profile count becomes hot. "
287 "The print is enabled under -Rpass-analysis=pgo, or "
288 "internal option -pass-remakrs-analysis=pgo."));
292 cl::desc(
"Print out mismatched BFI counts after setting profile metadata "
293 "The print is enabled under -Rpass-analysis=pgo, or "
294 "internal option -pass-remakrs-analysis=pgo."));
298 cl::desc(
"Set the threshold for pgo-verify-bfi: only print out "
299 "mismatched BFI if the difference percentage is greater than "
300 "this value (in percentage)."));
304 cl::desc(
"Set the threshold for pgo-verify-bfi: skip the counts whose "
305 "profile count value is below."));
310 cl::desc(
"Trace the hash of the function with this name."));
314 cl::desc(
"Do not instrument functions smaller than this threshold."));
318 cl::desc(
"Do not instrument functions with the number of critical edges "
319 " greater than this threshold."));
341 return std::string();
346 return std::string();
358 else if (CV->
isOne())
370#define VALUE_PROF_KIND(Enumerator, Value, Descr) Descr,
377 const StringRef VarName(INSTR_PROF_QUOTE(INSTR_PROF_RAW_VERSION_VAR));
379 uint64_t ProfileVersion = (INSTR_PROF_RAW_VERSION | VARIANT_MASK_IR_PROF);
381 ProfileVersion |= VARIANT_MASK_CSIR_PROF;
383 ProfileVersion |= VARIANT_MASK_INSTR_ENTRY;
385 ProfileVersion |= VARIANT_MASK_DBG_CORRELATE;
388 VARIANT_MASK_BYTE_COVERAGE | VARIANT_MASK_FUNCTION_ENTRY_ONLY;
390 ProfileVersion |= VARIANT_MASK_BYTE_COVERAGE;
392 ProfileVersion |= VARIANT_MASK_TEMPORAL_PROF;
397 Triple TT(M.getTargetTriple());
398 if (TT.supportsCOMDAT()) {
400 IRLevelVersionVariable->setComdat(M.getOrInsertComdat(VarName));
402 return IRLevelVersionVariable;
412enum VisitMode { VM_counting, VM_instrument, VM_annotate };
416struct SelectInstVisitor :
public InstVisitor<SelectInstVisitor> {
419 VisitMode
Mode = VM_counting;
420 unsigned *CurCtrIdx =
nullptr;
421 unsigned TotalNumCtrs = 0;
424 PGOUseFunc *UseFunc =
nullptr;
425 bool HasSingleByteCoverage;
427 SelectInstVisitor(
Function &Func,
bool HasSingleByteCoverage)
428 :
F(
Func), HasSingleByteCoverage(HasSingleByteCoverage) {}
430 void countSelects() {
440 void instrumentSelects(
unsigned *Ind,
unsigned TotalNC,
GlobalVariable *FNV,
442 Mode = VM_instrument;
444 TotalNumCtrs = TotalNC;
451 void annotateSelects(PGOUseFunc *UF,
unsigned *Ind) {
466 unsigned getNumOfSelectInsts()
const {
return NSIs; }
478 bool Removed =
false;
479 bool IsCritical =
false;
482 : SrcBB(Src), DestBB(Dest), Weight(
W) {}
485 std::string infoString()
const {
486 return (
Twine(Removed ?
"-" :
" ") + (InMST ?
" " :
"*") +
487 (IsCritical ?
"c" :
" ") +
" W=" +
Twine(Weight))
498 PGOBBInfo(
unsigned IX) : Group(this),
Index(IX) {}
501 std::string infoString()
const {
507template <
class Edge,
class BBInfo>
class FuncPGOInstrumentation {
515 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers;
519 void computeCFGHash();
520 void renameComdatFunction();
524 std::vector<std::vector<VPCandidateInfo>> ValueSites;
525 SelectInstVisitor SIVisitor;
526 std::string FuncName;
527 std::string DeprecatedFuncName;
536 const std::optional<BlockCoverageInference> BCI;
538 static std::optional<BlockCoverageInference>
539 constructBCI(
Function &Func,
bool HasSingleByteCoverage,
540 bool InstrumentFuncEntry) {
541 if (HasSingleByteCoverage)
548 void getInstrumentBBs(std::vector<BasicBlock *> &InstrumentBBs);
561 void dumpInfo(
StringRef Str =
"")
const {
563 " Hash: " +
Twine(FunctionHash) +
"\t" + Str);
566 FuncPGOInstrumentation(
568 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers,
571 bool InstrumentFuncEntry =
true,
bool HasSingleByteCoverage =
false)
572 :
F(
Func), IsCS(IsCS), ComdatMembers(ComdatMembers), VPC(
Func, TLI),
573 TLI(TLI), ValueSites(IPVK_Last + 1),
574 SIVisitor(
Func, HasSingleByteCoverage),
575 MST(
F, InstrumentFuncEntry, BPI,
BFI),
576 BCI(constructBCI(
Func, HasSingleByteCoverage, InstrumentFuncEntry)) {
578 BCI->viewBlockCoverageGraph();
580 SIVisitor.countSelects();
581 ValueSites[IPVK_MemOPSize] = VPC.
get(IPVK_MemOPSize);
583 NumOfPGOSelectInsts += SIVisitor.getNumOfSelectInsts();
584 NumOfPGOMemIntrinsics += ValueSites[IPVK_MemOPSize].size();
586 ValueSites[IPVK_IndirectCallTarget] = VPC.
get(IPVK_IndirectCallTarget);
588 ValueSites[IPVK_VTableTarget] = VPC.
get(IPVK_VTableTarget);
590 NumOfCSPGOSelectInsts += SIVisitor.getNumOfSelectInsts();
591 NumOfCSPGOMemIntrinsics += ValueSites[IPVK_MemOPSize].size();
598 if (!ComdatMembers.empty())
599 renameComdatFunction();
602 for (
const auto &E : MST.
allEdges()) {
605 IsCS ? NumOfCSPGOEdge++ : NumOfPGOEdge++;
607 IsCS ? NumOfCSPGOInstrument++ : NumOfPGOInstrument++;
620template <
class Edge,
class BBInfo>
621void FuncPGOInstrumentation<Edge, BBInfo>::computeCFGHash() {
622 std::vector<uint8_t> Indexes;
626 auto BI = findBBInfo(Succ);
630 for (
int J = 0; J < 4; J++)
631 Indexes.push_back((uint8_t)(
Index >> (J * 8)));
638 auto updateJCH = [&JCH](
uint64_t Num) {
643 updateJCH((
uint64_t)SIVisitor.getNumOfSelectInsts());
644 updateJCH((
uint64_t)ValueSites[IPVK_IndirectCallTarget].
size());
647 updateJCH(BCI->getInstrumentedBlocksHash());
657 FunctionHash &= 0x0FFFFFFFFFFFFFFF;
660 LLVM_DEBUG(
dbgs() <<
"Function Hash Computation for " <<
F.getName() <<
":\n"
661 <<
" CRC = " << JC.
getCRC()
662 <<
", Selects = " << SIVisitor.getNumOfSelectInsts()
663 <<
", Edges = " << MST.
numEdges() <<
", ICSites = "
664 << ValueSites[IPVK_IndirectCallTarget].size()
665 <<
", Memops = " << ValueSites[IPVK_MemOPSize].size()
666 <<
", High32 CRC = " << JCH.
getCRC()
667 <<
", Hash = " << FunctionHash <<
"\n";);
670 dbgs() <<
"Funcname=" <<
F.getName() <<
", Hash=" << FunctionHash
671 <<
" in building " <<
F.getParent()->getSourceFileName() <<
"\n";
677 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) {
689 for (
auto &&CM :
make_range(ComdatMembers.equal_range(
C))) {
690 assert(!isa<GlobalAlias>(CM.second));
691 Function *FM = dyn_cast<Function>(CM.second);
699template <
class Edge,
class BBInfo>
700void FuncPGOInstrumentation<Edge, BBInfo>::renameComdatFunction() {
703 std::string OrigName =
F.getName().str();
704 std::string NewFuncName =
706 F.setName(
Twine(NewFuncName));
708 FuncName =
Twine(FuncName +
"." +
Twine(FunctionHash)).
str();
714 if (!
F.hasComdat()) {
716 NewComdat =
M->getOrInsertComdat(
StringRef(NewFuncName));
718 F.setComdat(NewComdat);
723 Comdat *OrigComdat =
F.getComdat();
724 std::string NewComdatName =
726 NewComdat =
M->getOrInsertComdat(
StringRef(NewComdatName));
729 for (
auto &&CM :
make_range(ComdatMembers.equal_range(OrigComdat))) {
731 cast<Function>(CM.second)->setComdat(NewComdat);
737template <
class Edge,
class BBInfo>
738void FuncPGOInstrumentation<Edge, BBInfo>::getInstrumentBBs(
739 std::vector<BasicBlock *> &InstrumentBBs) {
742 if (BCI->shouldInstrumentBlock(BB))
743 InstrumentBBs.push_back(&BB);
748 std::vector<Edge *> EdgeList;
750 for (
const auto &E : MST.
allEdges())
751 EdgeList.push_back(E.get());
753 for (
auto &E : EdgeList) {
756 InstrumentBBs.push_back(InstrBB);
762template <
class Edge,
class BBInfo>
763BasicBlock *FuncPGOInstrumentation<Edge, BBInfo>::getInstrBB(Edge *E) {
764 if (E->InMST || E->Removed)
770 if (SrcBB ==
nullptr)
772 if (DestBB ==
nullptr)
787 return canInstrument(SrcBB);
789 return canInstrument(DestBB);
798 dbgs() <<
"Fail to split critical edge: not instrument this edge.\n");
803 IsCS ? NumOfCSPGOSplit++ : NumOfPGOSplit++;
805 <<
" --> " << getBBInfo(DestBB).
Index <<
"\n");
807 MST.
addEdge(SrcBB, InstrBB, 0);
809 Edge &NewEdge1 = MST.
addEdge(InstrBB, DestBB, 0);
810 NewEdge1.InMST =
true;
813 return canInstrument(InstrBB);
829 if (!isa<IntrinsicInst>(OrigCall)) {
832 std::optional<OperandBundleUse> ParentFunclet =
840 if (!BlockColors.
empty()) {
841 const ColorVector &CV = BlockColors.
find(OrigCall->getParent())->second;
842 assert(CV.
size() == 1 &&
"non-unique color for block!");
855 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers,
863 FuncPGOInstrumentation<PGOEdge, PGOBBInfo> FuncInfo(
867 auto Name = FuncInfo.FuncNameVar;
869 FuncInfo.FunctionHash);
871 auto &EntryBB =
F.getEntryBlock();
872 IRBuilder<> Builder(&EntryBB, EntryBB.getFirstInsertionPt());
881 std::vector<BasicBlock *> InstrumentBBs;
882 FuncInfo.getInstrumentBBs(InstrumentBBs);
883 unsigned NumCounters =
884 InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts();
889 auto &EntryBB =
F.getEntryBlock();
890 IRBuilder<> Builder(&EntryBB, EntryBB.getFirstInsertionPt());
899 for (
auto *InstrBB : InstrumentBBs) {
902 "Cannot get the Instrumentation point");
907 ? Intrinsic::instrprof_cover
908 : Intrinsic::instrprof_increment),
913 FuncInfo.SIVisitor.instrumentSelects(&
I, NumCounters, FuncInfo.FuncNameVar,
914 FuncInfo.FunctionHash);
920 NumOfPGOICall += FuncInfo.ValueSites[IPVK_IndirectCallTarget].size();
927 if (
F.hasPersonalityFn() &&
932 for (
uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) {
933 unsigned SiteIndex = 0;
939 <<
" site: CallSite Index = " << SiteIndex <<
"\n");
943 "Cannot get the Instrumentation point");
945 Value *ToProfile =
nullptr;
946 if (Cand.V->getType()->isIntegerTy())
948 else if (Cand.V->getType()->isPointerTy())
950 assert(ToProfile &&
"value profiling Value is of unexpected type");
956 {FuncInfo.FuncNameVar, Builder.
getInt64(FuncInfo.FunctionHash),
966struct PGOUseEdge :
public PGOEdge {
967 using PGOEdge::PGOEdge;
969 std::optional<uint64_t> Count;
975 std::string infoString()
const {
977 return PGOEdge::infoString();
978 return (
Twine(PGOEdge::infoString()) +
" Count=" +
Twine(*Count)).str();
985struct PGOUseBBInfo :
public PGOBBInfo {
986 std::optional<uint64_t> Count;
987 int32_t UnknownCountInEdge = 0;
988 int32_t UnknownCountOutEdge = 0;
990 DirectEdges OutEdges;
992 PGOUseBBInfo(
unsigned IX) : PGOBBInfo(IX) {}
998 std::string infoString()
const {
1000 return PGOBBInfo::infoString();
1001 return (
Twine(PGOBBInfo::infoString()) +
" Count=" +
Twine(*Count)).str();
1005 void addOutEdge(PGOUseEdge *E) {
1006 OutEdges.push_back(E);
1007 UnknownCountOutEdge++;
1011 void addInEdge(PGOUseEdge *E) {
1012 InEdges.push_back(E);
1013 UnknownCountInEdge++;
1022 for (
const auto &E : Edges) {
1036 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers,
1039 bool HasSingleByteCoverage)
1040 :
F(
Func),
M(Modu),
BFI(BFIin), PSI(PSI),
1041 FuncInfo(
Func, TLI, ComdatMembers,
false, BPI, BFIin, IsCS,
1042 InstrumentFuncEntry, HasSingleByteCoverage),
1043 FreqAttr(FFA_Normal), IsCS(IsCS) {}
1045 void handleInstrProfError(
Error Err,
uint64_t MismatchedFuncSum);
1052 void populateCounters();
1061 void annotateValueSites();
1064 void annotateValueSites(
uint32_t Kind);
1067 void annotateIrrLoopHeaderWeights();
1070 enum FuncFreqAttr { FFA_Normal, FFA_Cold, FFA_Hot };
1073 FuncFreqAttr getFuncFreqAttr()
const {
return FreqAttr; }
1082 PGOUseBBInfo &getBBInfo(
const BasicBlock *BB)
const {
1083 return FuncInfo.getBBInfo(BB);
1087 PGOUseBBInfo *findBBInfo(
const BasicBlock *BB)
const {
1088 return FuncInfo.findBBInfo(BB);
1093 void dumpInfo(
StringRef Str =
"")
const { FuncInfo.dumpInfo(Str); }
1095 uint64_t getProgramMaxCount()
const {
return ProgramMaxCount; }
1104 FuncPGOInstrumentation<PGOUseEdge, PGOUseBBInfo> FuncInfo;
1120 FuncFreqAttr FreqAttr;
1126 bool setInstrumentedCounts(
const std::vector<uint64_t> &CountFromProfile);
1139 FreqAttr = FFA_Cold;
1147 const FuncPGOInstrumentation<PGOUseEdge, PGOUseBBInfo> &FuncInfo) {
1151 for (
const auto &E : FuncInfo.MST.allEdges()) {
1156 PGOUseBBInfo &SrcInfo = FuncInfo.getBBInfo(SrcBB);
1157 PGOUseBBInfo &DestInfo = FuncInfo.getBBInfo(DestBB);
1158 SrcInfo.addOutEdge(E.get());
1159 DestInfo.addInEdge(E.get());
1165bool PGOUseFunc::setInstrumentedCounts(
1166 const std::vector<uint64_t> &CountFromProfile) {
1168 std::vector<BasicBlock *> InstrumentBBs;
1169 FuncInfo.getInstrumentBBs(InstrumentBBs);
1173 unsigned NumCounters =
1174 InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts();
1177 if (NumCounters != CountFromProfile.size()) {
1180 auto *FuncEntry = &*
F.begin();
1185 uint64_t CountValue = CountFromProfile[
I++];
1186 PGOUseBBInfo &
Info = getBBInfo(InstrBB);
1190 if (InstrBB == FuncEntry && CountValue == 0)
1192 Info.setBBInfoCount(CountValue);
1194 ProfileCountSize = CountFromProfile.size();
1198 auto setEdgeCount = [
this](PGOUseEdge *E,
uint64_t Value) ->
void {
1199 E->setEdgeCount(
Value);
1200 this->getBBInfo(E->SrcBB).UnknownCountOutEdge--;
1201 this->getBBInfo(E->DestBB).UnknownCountInEdge--;
1207 for (
const auto &E : FuncInfo.MST.allEdges()) {
1208 if (E->Removed || E->InMST)
1211 PGOUseBBInfo &SrcInfo = getBBInfo(SrcBB);
1215 if (SrcInfo.Count && SrcInfo.OutEdges.size() == 1)
1216 setEdgeCount(E.get(), *SrcInfo.Count);
1219 PGOUseBBInfo &DestInfo = getBBInfo(DestBB);
1222 if (DestInfo.Count && DestInfo.InEdges.size() == 1)
1223 setEdgeCount(E.get(), *DestInfo.Count);
1229 setEdgeCount(E.get(), 0);
1236void PGOUseFunc::setEdgeCount(DirectEdges &Edges,
uint64_t Value) {
1237 for (
auto &E : Edges) {
1240 E->setEdgeCount(
Value);
1242 getBBInfo(E->SrcBB).UnknownCountOutEdge--;
1243 getBBInfo(E->DestBB).UnknownCountInEdge--;
1251 const char MetadataName[] =
"instr_prof_hash_mismatch";
1254 auto *Existing =
F.getMetadata(LLVMContext::MD_annotation);
1256 MDTuple *Tuple = cast<MDTuple>(Existing);
1257 for (
const auto &
N : Tuple->
operands()) {
1258 if (
N.equalsStr(MetadataName))
1267 F.setMetadata(LLVMContext::MD_annotation, MD);
1270void PGOUseFunc::handleInstrProfError(
Error Err,
uint64_t MismatchedFuncSum) {
1272 auto &Ctx =
M->getContext();
1273 auto Err = IPE.
get();
1274 bool SkipWarning =
false;
1276 << FuncInfo.FuncName <<
": ");
1277 if (Err == instrprof_error::unknown_function) {
1278 IsCS ? NumOfCSPGOMissing++ : NumOfPGOMissing++;
1281 }
else if (Err == instrprof_error::hash_mismatch ||
1282 Err == instrprof_error::malformed) {
1283 IsCS ? NumOfCSPGOMismatch++ : NumOfPGOMismatch++;
1289 LLVM_DEBUG(
dbgs() <<
"hash mismatch (hash= " << FuncInfo.FunctionHash
1290 <<
" skip=" << SkipWarning <<
")");
1300 IPE.
message() + std::string(
" ") +
F.getName().str() +
1301 std::string(
" Hash = ") + std::to_string(FuncInfo.FunctionHash) +
1302 std::string(
" up to ") + std::to_string(MismatchedFuncSum) +
1303 std::string(
" count discarded");
1315 auto &Ctx =
M->getContext();
1318 FuncInfo.FuncName, FuncInfo.FunctionHash, FuncInfo.DeprecatedFuncName,
1319 &MismatchedFuncSum);
1321 handleInstrProfError(std::move(E), MismatchedFuncSum);
1324 ProfileRecord = std::move(
Result.get());
1329 std::vector<uint64_t> &CountFromProfile = ProfileRecord.
Counts;
1331 IsCS ? NumOfCSPGOFunc++ : NumOfPGOFunc++;
1335 for (
unsigned I = 0, S = CountFromProfile.size();
I < S;
I++) {
1337 ValueSum += CountFromProfile[
I];
1339 AllZeros = (ValueSum == 0);
1343 getBBInfo(
nullptr).UnknownCountOutEdge = 2;
1344 getBBInfo(
nullptr).UnknownCountInEdge = 2;
1346 if (!setInstrumentedCounts(CountFromProfile)) {
1348 dbgs() <<
"Inconsistent number of counts, skipping this function");
1350 M->getName().data(),
1351 Twine(
"Inconsistent number of counts in ") +
F.getName().str() +
1352 Twine(
": the profile may be stale or there is a function name "
1364 FuncInfo.FuncName, FuncInfo.FunctionHash, FuncInfo.DeprecatedFuncName,
1365 &MismatchedFuncSum);
1366 if (
auto Err =
Result.takeError()) {
1367 handleInstrProfError(std::move(Err), MismatchedFuncSum);
1371 std::vector<uint64_t> &CountsFromProfile =
Result.get().Counts;
1375 if (FuncInfo.BCI->shouldInstrumentBlock(BB))
1381 InverseDependencies;
1382 for (
auto &BB :
F) {
1383 for (
auto *Dep : FuncInfo.BCI->getDependencies(BB)) {
1385 InverseDependencies[Dep].
insert(&BB);
1390 std::stack<const BasicBlock *> CoveredBlocksToProcess;
1391 for (
auto &[BB, IsCovered] : Coverage)
1393 CoveredBlocksToProcess.push(BB);
1395 while (!CoveredBlocksToProcess.empty()) {
1396 auto *CoveredBlock = CoveredBlocksToProcess.top();
1397 assert(Coverage[CoveredBlock]);
1398 CoveredBlocksToProcess.pop();
1399 for (
auto *BB : InverseDependencies[CoveredBlock]) {
1404 CoveredBlocksToProcess.push(BB);
1412 F.setEntryCount(Coverage[&
F.getEntryBlock()] ? 10000 : 0);
1413 for (
auto &BB :
F) {
1422 Weights.
push_back((Coverage[Succ] || !Coverage[&BB]) ? 1 : 0);
1423 if (Weights.
size() >= 2)
1427 unsigned NumCorruptCoverage = 0;
1432 auto IsBlockDead = [&](
const BasicBlock &BB) -> std::optional<bool> {
1433 if (
auto C =
BFI.getBlockProfileCount(&BB))
1437 LLVM_DEBUG(
dbgs() <<
"Block Coverage: (Instrumented=*, Covered=X)\n");
1438 for (
auto &BB :
F) {
1439 LLVM_DEBUG(
dbgs() << (FuncInfo.BCI->shouldInstrumentBlock(BB) ?
"* " :
" ")
1440 << (Coverage[&BB] ?
"X " :
" ") <<
" " << BB.getName()
1446 if (Coverage[&BB] == IsBlockDead(BB).value_or(
false)) {
1448 dbgs() <<
"Found inconsistent block covearge for " << BB.getName()
1449 <<
": BCI=" << (Coverage[&BB] ?
"Covered" :
"Dead") <<
" BFI="
1450 << (IsBlockDead(BB).
value() ?
"Dead" :
"Covered") <<
"\n");
1451 ++NumCorruptCoverage;
1457 auto &Ctx =
M->getContext();
1459 M->getName().data(),
1460 Twine(
"Found inconsistent block coverage for function ") +
F.getName() +
1461 " in " +
Twine(NumCorruptCoverage) +
" blocks.",
1465 FuncInfo.BCI->viewBlockCoverageGraph(&Coverage);
1470void PGOUseFunc::populateCounters() {
1471 bool Changes =
true;
1472 unsigned NumPasses = 0;
1480 PGOUseBBInfo *UseBBInfo = findBBInfo(&BB);
1481 if (UseBBInfo ==
nullptr)
1483 if (!UseBBInfo->Count) {
1484 if (UseBBInfo->UnknownCountOutEdge == 0) {
1487 }
else if (UseBBInfo->UnknownCountInEdge == 0) {
1492 if (UseBBInfo->Count) {
1493 if (UseBBInfo->UnknownCountOutEdge == 1) {
1499 if (*UseBBInfo->Count > OutSum)
1500 Total = *UseBBInfo->Count - OutSum;
1501 setEdgeCount(UseBBInfo->OutEdges,
Total);
1504 if (UseBBInfo->UnknownCountInEdge == 1) {
1507 if (*UseBBInfo->Count > InSum)
1508 Total = *UseBBInfo->Count - InSum;
1509 setEdgeCount(UseBBInfo->InEdges,
Total);
1516 LLVM_DEBUG(
dbgs() <<
"Populate counts in " << NumPasses <<
" passes.\n");
1520 for (
auto &BB :
F) {
1521 auto BI = findBBInfo(&BB);
1524 assert(BI->Count &&
"BB count is not valid");
1529 for (
auto &BB :
F) {
1530 auto BI = findBBInfo(&BB);
1533 FuncMaxCount = std::max(FuncMaxCount, *BI->Count);
1543 FuncInfo.SIVisitor.annotateSelects(
this, &CountPosition);
1544 assert(CountPosition == ProfileCountSize);
1546 LLVM_DEBUG(FuncInfo.dumpInfo(
"after reading profile."));
1550void PGOUseFunc::setBranchWeights() {
1552 LLVM_DEBUG(
dbgs() <<
"\nSetting branch weights for func " <<
F.getName()
1553 <<
" IsCS=" << IsCS <<
"\n");
1554 for (
auto &BB :
F) {
1558 if (!(isa<BranchInst>(TI) || isa<SwitchInst>(TI) ||
1559 isa<IndirectBrInst>(TI) || isa<InvokeInst>(TI) ||
1560 isa<CallBrInst>(TI)))
1563 const PGOUseBBInfo &BBCountInfo = getBBInfo(&BB);
1564 if (!*BBCountInfo.Count)
1568 unsigned Size = BBCountInfo.OutEdges.size();
1571 for (
unsigned s = 0; s <
Size; s++) {
1572 const PGOUseEdge *E = BBCountInfo.OutEdges[s];
1575 if (DestBB ==
nullptr)
1579 if (EdgeCount > MaxCount)
1580 MaxCount = EdgeCount;
1581 EdgeCounts[SuccNum] = EdgeCount;
1590 auto &Ctx =
M->getContext();
1592 M->getName().data(),
1593 Twine(
"Profile in ") +
F.getName().str() +
1594 Twine(
" partially ignored") +
1595 Twine(
", possibly due to the lack of a return path."),
1603 if (isa<IndirectBrInst>(Pred->getTerminator()))
1609void PGOUseFunc::annotateIrrLoopHeaderWeights() {
1610 LLVM_DEBUG(
dbgs() <<
"\nAnnotating irreducible loop header weights.\n");
1612 for (
auto &BB :
F) {
1618 const PGOUseBBInfo &BBCountInfo = getBBInfo(&BB);
1624void SelectInstVisitor::instrumentOneSelectInst(
SelectInst &SI) {
1628 auto *Step = Builder.CreateZExt(
SI.getCondition(), Int64Ty);
1631 {FuncNameVar, Builder.getInt64(FuncHash), Builder.getInt32(TotalNumCtrs),
1632 Builder.getInt32(*CurCtrIdx), Step});
1636void SelectInstVisitor::annotateOneSelectInst(
SelectInst &SI) {
1637 std::vector<uint64_t> &CountFromProfile = UseFunc->getProfileRecord().Counts;
1638 assert(*CurCtrIdx < CountFromProfile.size() &&
1639 "Out of bound access of counters");
1641 SCounts[0] = CountFromProfile[*CurCtrIdx];
1644 auto BI = UseFunc->findBBInfo(
SI.getParent());
1646 TotalCount = *BI->Count;
1648 SCounts[1] = (TotalCount > SCounts[0] ? TotalCount - SCounts[0] : 0);
1649 uint64_t MaxCount = std::max(SCounts[0], SCounts[1]);
1654void SelectInstVisitor::visitSelectInst(
SelectInst &SI) {
1658 if (
SI.getCondition()->getType()->isVectorTy())
1666 instrumentOneSelectInst(SI);
1669 annotateOneSelectInst(SI);
1677void PGOUseFunc::annotateValueSites() {
1685 annotateValueSites(Kind);
1689void PGOUseFunc::annotateValueSites(
uint32_t Kind) {
1690 assert(Kind <= IPVK_Last);
1691 unsigned ValueSiteIndex = 0;
1692 auto &ValueSites = FuncInfo.ValueSites[
Kind];
1694 if (NumValueSites != ValueSites.size()) {
1695 auto &Ctx =
M->getContext();
1697 M->getName().data(),
1698 Twine(
"Inconsistent number of value sites for ") +
1701 Twine(
"\", possibly due to the use of a stale profile."),
1707 LLVM_DEBUG(
dbgs() <<
"Read one value site profile (kind = " << Kind
1708 <<
"): Index = " << ValueSiteIndex <<
" out of "
1709 << NumValueSites <<
"\n");
1722 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) {
1727 ComdatMembers.insert(std::make_pair(
C, &
F));
1729 if (
Comdat *
C = GV.getComdat())
1730 ComdatMembers.insert(std::make_pair(
C, &GV));
1732 if (
Comdat *
C = GA.getComdat())
1733 ComdatMembers.insert(std::make_pair(
C, &GA));
1738 if (
F.isDeclaration())
1743 unsigned NumCriticalEdges = 0;
1744 for (
auto &BB :
F) {
1753 <<
", NumCriticalEdges=" << NumCriticalEdges
1754 <<
" exceed the threshold. Skip PGO.\n");
1764 if (
F.hasFnAttribute(llvm::Attribute::Naked))
1766 if (
F.hasFnAttribute(llvm::Attribute::NoProfile))
1768 if (
F.hasFnAttribute(llvm::Attribute::SkipProfile))
1784 Triple TT(M.getTargetTriple());
1789 Twine(
"VTable value profiling is presently not "
1790 "supported for non-ELF object formats"),
1792 std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers;
1798 auto &TLI = LookupTLI(
F);
1799 auto *BPI = LookupBPI(
F);
1800 auto *BFI = LookupBFI(
F);
1844 auto BFIEntryCount =
F.getEntryCount();
1845 assert(BFIEntryCount && (BFIEntryCount->getCount() > 0) &&
1846 "Invalid BFI Entrycount");
1850 for (
auto &BBI :
F) {
1853 if (!Func.findBBInfo(&BBI))
1856 CountValue = *Func.getBBInfo(&BBI).Count;
1857 BFICountValue = *BFICount;
1861 if (SumCount.isZero())
1865 "Incorrect sum of BFI counts");
1868 double Scale = (SumCount / SumBFICount).convertToDouble();
1869 if (Scale < 1.001 && Scale > 0.999)
1874 if (NewEntryCount == 0)
1880 << NewEntryCount <<
"\n");
1897 unsigned BBNum = 0, BBMisMatchNum = 0, NonZeroBBNum = 0;
1898 for (
auto &BBI :
F) {
1902 CountValue = Func.getBBInfo(&BBI).Count.value_or(CountValue);
1909 BFICountValue = *BFICount;
1912 bool rawIsHot = CountValue >= HotCountThreshold;
1913 bool BFIIsHot = BFICountValue >= HotCountThreshold;
1915 bool ShowCount =
false;
1916 if (rawIsHot && !BFIIsHot) {
1917 Msg =
"raw-Hot to BFI-nonHot";
1919 }
else if (rawIsCold && BFIIsHot) {
1920 Msg =
"raw-Cold to BFI-Hot";
1929 uint64_t Diff = (BFICountValue >= CountValue)
1930 ? BFICountValue - CountValue
1931 : CountValue - BFICountValue;
1939 F.getSubprogram(), &BBI);
1941 <<
" Count=" <<
ore::NV(
"Count", CountValue)
1942 <<
" BFI_Count=" <<
ore::NV(
"Count", BFICountValue);
1944 Remark <<
" (" << Msg <<
")";
1951 F.getSubprogram(), &
F.getEntryBlock())
1952 <<
"In Func " <<
ore::NV(
"Function",
F.getName())
1953 <<
": Num_of_BB=" <<
ore::NV(
"Count", BBNum)
1954 <<
", Num_of_non_zerovalue_BB=" <<
ore::NV(
"Count", NonZeroBBNum)
1955 <<
", Num_of_mis_matching_BB=" <<
ore::NV(
"Count", BBMisMatchNum);
1967 auto &Ctx = M.getContext();
1970 ProfileRemappingFileName);
1971 if (
Error E = ReaderOrErr.takeError()) {
1979 std::unique_ptr<IndexedInstrProfReader> PGOReader =
1980 std::move(ReaderOrErr.get());
1986 if (!PGOReader->hasCSIRLevelProfile() && IsCS)
1990 if (!PGOReader->isIRLevelProfile()) {
1992 ProfileFileName.
data(),
"Not an IR level instrumentation profile"));
1995 if (PGOReader->functionEntryOnly()) {
1997 ProfileFileName.
data(),
1998 "Function entry profiles are not yet supported for optimization"));
2005 M.setProfileSummary(PGOReader->getSummary(IsCS).getMD(M.getContext()),
2010 std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers;
2012 std::vector<Function *> HotFunctions;
2013 std::vector<Function *> ColdFunctions;
2017 bool InstrumentFuncEntry = PGOReader->instrEntryBBEnabled();
2020 bool HasSingleByteCoverage = PGOReader->hasSingleByteCoverage();
2024 auto &TLI = LookupTLI(
F);
2025 auto *BPI = LookupBPI(
F);
2026 auto *BFI = LookupBFI(
F);
2027 if (!HasSingleByteCoverage) {
2033 PGOUseFunc Func(
F, &M, TLI, ComdatMembers, BPI, BFI, PSI, IsCS,
2034 InstrumentFuncEntry, HasSingleByteCoverage);
2035 if (HasSingleByteCoverage) {
2036 Func.populateCoverage(PGOReader.get());
2044 bool AllZeros =
false;
2045 if (!Func.readCounters(PGOReader.get(), AllZeros, PseudoKind))
2049 if (Func.getProgramMaxCount() != 0)
2050 ColdFunctions.push_back(&
F);
2055 if (
F.hasFnAttribute(Attribute::Cold))
2056 F.removeFnAttr(Attribute::Cold);
2059 F.addFnAttr(Attribute::Hot);
2062 Func.populateCounters();
2063 Func.setBranchWeights();
2064 Func.annotateValueSites();
2065 Func.annotateIrrLoopHeaderWeights();
2066 PGOUseFunc::FuncFreqAttr FreqAttr = Func.getFuncFreqAttr();
2067 if (FreqAttr == PGOUseFunc::FFA_Cold)
2068 ColdFunctions.push_back(&
F);
2069 else if (FreqAttr == PGOUseFunc::FFA_Hot)
2070 HotFunctions.push_back(&
F);
2075 std::unique_ptr<BranchProbabilityInfo> NewBPI =
2076 std::make_unique<BranchProbabilityInfo>(
F, LI);
2077 std::unique_ptr<BlockFrequencyInfo> NewBFI =
2078 std::make_unique<BlockFrequencyInfo>(
F, *NewBPI, LI);
2082 dbgs() <<
"pgo-view-counts: " << Func.getFunc().getName() <<
"\n";
2083 NewBFI->print(
dbgs());
2093 ViewGraph(&Func,
Twine(
"PGORawCounts_") + Func.getFunc().getName());
2095 dbgs() <<
"pgo-view-raw-counts: " << Func.getFunc().getName() <<
"\n";
2122 for (
auto &
F : HotFunctions) {
2123 F->addFnAttr(Attribute::InlineHint);
2124 LLVM_DEBUG(
dbgs() <<
"Set inline attribute to function: " <<
F->getName()
2127 for (
auto &
F : ColdFunctions) {
2130 if (
F->hasFnAttribute(Attribute::Hot)) {
2131 auto &Ctx = M.getContext();
2132 std::string Msg = std::string(
"Function ") +
F->getName().str() +
2133 std::string(
" is annotated as a hot function but"
2134 " the profile is cold");
2139 F->addFnAttr(Attribute::Cold);
2140 LLVM_DEBUG(
dbgs() <<
"Set cold attribute to function: " <<
F->getName()
2147 std::string Filename, std::string RemappingFilename,
bool IsCS,
2149 : ProfileFileName(
std::
move(Filename)),
2150 ProfileRemappingFileName(
std::
move(RemappingFilename)), IsCS(IsCS),
2177 LookupTLI, LookupBPI, LookupBFI, PSI, IsCS))
2184 if (!
Node->getName().empty())
2185 return Node->getName().str();
2187 std::string SimpleNodeName;
2195 assert(MaxCount > 0 &&
"Bad max count");
2198 for (
const auto &ECI : EdgeCounts)
2211 if (BrCondStr.empty())
2223 std::string BranchProbStr;
2226 OS <<
" (total count : " << TotalCount <<
")";
2232 << BrCondStr <<
" is true with probability : " << BranchProbStr;
2251 return &
G->getFunc().front();
2274 return std::string(
G->getFunc().getName());
2282 PGOUseBBInfo *BI = Graph->findBBInfo(Node);
2284 if (BI && BI->Count)
2285 OS << *BI->Count <<
"\\l";
2293 if (!isa<SelectInst>(&
I))
2296 OS <<
"SELECT : { T = ";
2300 OS <<
"Unknown, F = Unknown }\\l";
2302 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)
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)
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."))
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)
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 Function * getParent() const
Return the enclosing method, or null if none.
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
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.
const BasicBlock * getParent() const
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.
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 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.
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.
const CustomOperand< const MCSubtargetInfo & > Msg[]
@ 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...
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)
bool isScopedEHPersonality(EHPersonality Pers)
Returns true if this personality uses scope-style EH IR instructions: catchswitch,...
cl::opt< bool > DebugInfoCorrelate
void setBranchWeights(Instruction &I, ArrayRef< uint32_t > Weights)
Create a new branch_weights metadata node and add or overwrite a prof metadata reference to instructi...
OperandBundleDefT< Value * > OperandBundleDef
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)
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