@@ -1113,47 +1113,25 @@ struct AAAMDWavesPerEU : public AAAMDSizeRangeAttribute {
11131113 Function *F = getAssociatedFunction ();
11141114 auto &InfoCache =
static_cast <AMDGPUInformationCache &>(A.
getInfoCache ());
11151115
1116- auto TakeRange = [&](std::pair<unsigned , unsigned > R) {
1117- auto [Min, Max] = R;
1118- ConstantRange Range (APInt (32 , Min), APInt (32 , Max + 1 ));
1119- IntegerRangeState RangeState (Range);
1120- clampStateAndIndicateChange (this ->getState (), RangeState);
1121- indicateOptimisticFixpoint ();
1122- };
1123-
1124- std::pair<unsigned , unsigned > MaxWavesPerEURange{
1125- 1U , InfoCache.getMaxWavesPerEU (*F)};
1126-
11271116 // If the attribute exists, we will honor it if it is not the default.
11281117 if (auto Attr = InfoCache.getWavesPerEUAttr (*F)) {
1118+ std::pair<unsigned , unsigned > MaxWavesPerEURange{
1119+ 1U , InfoCache.getMaxWavesPerEU (*F)};
11291120 if (*Attr != MaxWavesPerEURange) {
1130- TakeRange (*Attr);
1121+ auto [Min, Max] = *Attr;
1122+ ConstantRange Range (APInt (32 , Min), APInt (32 , Max + 1 ));
1123+ IntegerRangeState RangeState (Range);
1124+ this ->getState () = RangeState;
1125+ indicateOptimisticFixpoint ();
11311126 return ;
11321127 }
11331128 }
11341129
1135- // Unlike AAAMDFlatWorkGroupSize, it's getting trickier here. Since the
1136- // calculation of waves per EU involves flat work group size, we can't
1137- // simply use an assumed flat work group size as a start point, because the
1138- // update of flat work group size is in an inverse direction of waves per
1139- // EU. However, we can still do something if it is an entry function. Since
1140- // an entry function is a terminal node, and flat work group size either
1141- // from attribute or default will be used anyway, we can take that value and
1142- // calculate the waves per EU based on it. This result can't be updated by
1143- // no means, but that could still allow us to propagate it.
1144- if (AMDGPU::isEntryFunctionCC (F->getCallingConv ())) {
1145- std::pair<unsigned , unsigned > FlatWorkGroupSize;
1146- if (auto Attr = InfoCache.getFlatWorkGroupSizeAttr (*F))
1147- FlatWorkGroupSize = *Attr;
1148- else
1149- FlatWorkGroupSize = InfoCache.getDefaultFlatWorkGroupSize (*F);
1150- TakeRange (InfoCache.getEffectiveWavesPerEU (*F, MaxWavesPerEURange,
1151- FlatWorkGroupSize));
1152- }
1130+ if (AMDGPU::isEntryFunctionCC (F->getCallingConv ()))
1131+ indicatePessimisticFixpoint ();
11531132 }
11541133
11551134 ChangeStatus updateImpl (Attributor &A) override {
1156- auto &InfoCache =
static_cast <AMDGPUInformationCache &>(A.
getInfoCache ());
11571135 ChangeStatus Change = ChangeStatus::UNCHANGED;
11581136
11591137 auto CheckCallSite = [&](AbstractCallSite CS) {
@@ -1162,24 +1140,21 @@ struct AAAMDWavesPerEU : public AAAMDSizeRangeAttribute {
11621140 LLVM_DEBUG (dbgs () << ' [' getName () << " ] Call " getName ()
11631141 << " ->" getName () << ' \n '
11641142
1165- const auto *
CallerInfo = A.
getAAFor <AAAMDWavesPerEU>(
1143+ const auto *
CallerAA = A.
getAAFor <AAAMDWavesPerEU>(
11661144 *this , IRPosition::function (*Caller), DepClassTy::REQUIRED);
1167- const auto *AssumedGroupSize = A.
getAAFor <AAAMDFlatWorkGroupSize>(
1168- *this , IRPosition::function (*Func), DepClassTy::REQUIRED);
1169- if (!CallerInfo || !AssumedGroupSize || !CallerInfo->isValidState () ||
1170- !AssumedGroupSize->isValidState ())
1145+ if (!CallerAA || !CallerAA->isValidState ())
11711146 return false ;
11721147
1173- unsigned Min, Max ;
1174- std::tie (Min, Max) = InfoCache. getEffectiveWavesPerEU (
1175- *Caller,
1176- {CallerInfo-> getAssumed (). getLower ().getZExtValue (),
1177- CallerInfo ->getAssumed ().getUpper ().getZExtValue () - 1 },
1178- {AssumedGroupSize-> getAssumed (). getLower (). getZExtValue (),
1179- AssumedGroupSize-> getAssumed (). getUpper (). getZExtValue () - 1 } );
1180- ConstantRange CallerRange ( APInt ( 32 , Min), APInt ( 32 , Max + 1 )) ;
1181- IntegerRangeState CallerRangeState (CallerRange);
1182- Change |= clampStateAndIndicateChange ( this -> getState (), CallerRangeState) ;
1148+ ConstantRange Assumed = this -> getAssumed () ;
1149+ unsigned Min = std::max (Assumed. getLower (). getZExtValue (),
1150+ CallerAA-> getAssumed (). getLower (). getZExtValue ());
1151+ unsigned Max = std::max (Assumed. getUpper ().getZExtValue (),
1152+ CallerAA ->getAssumed ().getUpper ().getZExtValue ());
1153+ ConstantRange Range ( APInt ( 32 , Min), APInt ( 32 , Max));
1154+ IntegerRangeState RangeState (Range );
1155+ this -> getState () = RangeState ;
1156+ Change |= this -> getState () == Assumed ? ChangeStatus::UNCHANGED
1157+ : ChangeStatus::CHANGED ;
11831158
11841159 return true ;
11851160 };
@@ -1323,6 +1298,60 @@ struct AAAMDGPUNoAGPR
13231298
13241299const char AAAMDGPUNoAGPR::ID = 0 ;
13251300
1301+ // / The final check and update of the attribute 'amdgpu-waves-per-eu' based on
1302+ // / the determined 'amdgpu-flat-work-group-size' attribute. We can't do this
1303+ // / during attributor run because the two attributes grow in opposite direction,
1304+ // / we should not use any intermediate value to calculate waves per eu until we
1305+ // / have a determined flat workgroup size.
1306+ static void updateWavesPerEU (Module &M, TargetMachine &TM) {
1307+ for (Function &F : M) {
1308+ if (F.isDeclaration ())
1309+ continue ;
1310+
1311+ const GCNSubtarget &ST = TM.getSubtarget <GCNSubtarget>(F);
1312+
1313+ std::optional<std::pair<unsigned , std::optional<unsigned >>>
1314+ FlatWgrpSizeAttr =
1315+ AMDGPU::getIntegerPairAttribute (F, " amdgpu-flat-work-group-size"
1316+
1317+ unsigned MinWavesPerEU = ST.getMinWavesPerEU ();
1318+ unsigned MaxWavesPerEU = ST.getMaxWavesPerEU ();
1319+
1320+ unsigned MinFlatWgrpSize = ST.getMinFlatWorkGroupSize ();
1321+ unsigned MaxFlatWgrpSize = ST.getMaxFlatWorkGroupSize ();
1322+ if (FlatWgrpSizeAttr.has_value ()) {
1323+ MinFlatWgrpSize = FlatWgrpSizeAttr->first ;
1324+ MaxFlatWgrpSize = *(FlatWgrpSizeAttr->second );
1325+ }
1326+
1327+ // Start with the max range.
1328+ unsigned Min = MinWavesPerEU;
1329+ unsigned Max = MinWavesPerEU;
1330+
1331+ // Compute the range from flat workgroup size. `getWavesPerEU` will also
1332+ // account for the 'amdgpu-waves-er-eu' attribute.
1333+ auto [MinFromFlatWgrpSize, MaxFromFlatWgrpSize] =
1334+ ST.getWavesPerEU (F, {MinFlatWgrpSize, MaxFlatWgrpSize});
1335+
1336+ // For the lower bound, we have to "tighten" it.
1337+ Min = std::max (Min, MinFromFlatWgrpSize);
1338+ // For the upper bound, we have to "extend" it.
1339+ Max = std::max (Max, MaxFromFlatWgrpSize);
1340+
1341+ // Clamp the range to the max range.
1342+ Min = std::max (Min, MinWavesPerEU);
1343+ Max = std::min (Max, MaxWavesPerEU);
1344+
1345+ // Update the attribute if it is not the max.
1346+ if (Min != MinWavesPerEU || Max != MaxWavesPerEU) {
1347+ SmallString<10 > Buffer;
1348+ raw_svector_ostream OS (Buffer);
1349+ OS << Min << ' ,'
1350+ F.addFnAttr (" amdgpu-waves-per-eu" str ());
1351+ }
1352+ }
1353+ }
1354+
13261355static bool runImpl (Module &M, AnalysisGetter &AG, TargetMachine &TM,
13271356 AMDGPUAttributorOptions Options,
13281357 ThinOrFullLTOPhase LTOPhase) {
@@ -1396,8 +1425,16 @@ static bool runImpl(Module &M, AnalysisGetter &AG, TargetMachine &TM,
13961425 }
13971426 }
13981427
1399- ChangeStatus Change = A.run ();
1400- return Change == ChangeStatus::CHANGED;
1428+ bool Changed = A.run () == ChangeStatus::CHANGED;
1429+
1430+ // We only update the waves-per-eu attribute at the final stage to avoid
1431+ // setting it with intermediate values.
1432+ if (Changed && (LTOPhase == ThinOrFullLTOPhase::None ||
1433+ LTOPhase == ThinOrFullLTOPhase::FullLTOPostLink ||
1434+ LTOPhase == ThinOrFullLTOPhase::ThinLTOPostLink))
1435+ updateWavesPerEU (M, TM);
1436+
1437+ return Changed;
14011438}
14021439
14031440class AMDGPUAttributorLegacy : public ModulePass {
0 commit comments