Skip to content

Commit 731a537

Browse files
dcandlerdcandler
committed
fixup! [AArch64] Add support for C1 CPUs
1 parent 46e20c2 commit 731a537

File tree

2 files changed

+113
-109
lines changed

2 files changed

+113
-109
lines changed

‎llvm/lib/Target/AArch64/AArch64Processors.td‎

Lines changed: 111 additions & 107 deletions
Original file line numberDiff line numberDiff line change
@@ -65,6 +65,14 @@ def TuneA520AE : SubtargetFeature<"a520ae", "ARMProcFamily", "CortexA520",
6565
FeaturePostRAScheduler,
6666
FeatureUseWzrToVecMove]>;
6767

68+
def TuneC1Nano : SubtargetFeature<"c1-nano", "ARMProcFamily",
69+
"C1Nano", "C1-Nano ARM Processors",[
70+
FeatureFuseAES,
71+
FeatureFuseAdrpAdd,
72+
FeaturePostRAScheduler,
73+
FeatureUseWzrToVecMove,
74+
FeatureUseFixedOverScalableIfEqualCost]>;
75+
6876
def TuneA57 : SubtargetFeature<"a57", "ARMProcFamily", "CortexA57",
6977
"Cortex-A57 ARM processors", [
7078
FeatureFuseAES,
@@ -231,6 +239,18 @@ def TuneA725 : SubtargetFeature<"cortex-a725", "ARMProcFamily",
231239
FeatureEnableSelectOptimize,
232240
FeaturePredictableSelectIsExpensive]>;
233241

242+
def TuneC1Pro : SubtargetFeature<"c1-pro", "ARMProcFamily",
243+
"C1Pro", "C1-Pro ARM Processors",[
244+
FeatureFuseAES,
245+
FeaturePostRAScheduler,
246+
FeatureCmpBccFusion,
247+
FeatureALULSLFast,
248+
FeatureFuseAdrpAdd,
249+
FeatureFuseCmpCSel,
250+
FeatureFuseCmpCSet,
251+
FeatureEnableSelectOptimize,
252+
FeaturePredictableSelectIsExpensive]>;
253+
234254
def TuneR82 : SubtargetFeature<"cortex-r82", "ARMProcFamily",
235255
"CortexR82",
236256
"Cortex-R82 ARM processors", [
@@ -300,6 +320,32 @@ def TuneX925 : SubtargetFeature<"cortex-x925", "ARMProcFamily",
300320
FeatureAvoidLDAPUR,
301321
FeaturePredictableSelectIsExpensive]>;
302322

323+
def TuneC1Ultra : SubtargetFeature<"c1-ultra", "ARMProcFamily",
324+
"C1Ultra", "C1-Ultra ARM Processors",[
325+
FeatureALULSLFast,
326+
FeatureFuseAdrpAdd,
327+
FeatureFuseCmpCSel,
328+
FeatureFuseCmpCSet,
329+
FeatureFuseAES,
330+
FeaturePostRAScheduler,
331+
FeatureEnableSelectOptimize,
332+
FeatureUseFixedOverScalableIfEqualCost,
333+
FeatureAvoidLDAPUR,
334+
FeaturePredictableSelectIsExpensive]>;
335+
336+
def TuneC1Premium : SubtargetFeature<"c1-premium", "ARMProcFamily",
337+
"C1Premium", "C1-Premium ARM Processors",[
338+
FeatureALULSLFast,
339+
FeatureFuseAdrpAdd,
340+
FeatureFuseCmpCSel,
341+
FeatureFuseCmpCSet,
342+
FeatureFuseAES,
343+
FeaturePostRAScheduler,
344+
FeatureEnableSelectOptimize,
345+
FeatureUseFixedOverScalableIfEqualCost,
346+
FeatureAvoidLDAPUR,
347+
FeaturePredictableSelectIsExpensive]>;
348+
303349
def TuneA64FX : SubtargetFeature<"a64fx", "ARMProcFamily", "A64FX",
304350
"Fujitsu A64FX processors", [
305351
FeaturePostRAScheduler,
@@ -690,48 +736,6 @@ def TuneNeoverseV3AE : SubtargetFeature<"neoversev3AE", "ARMProcFamily", "Neover
690736
FeatureAvoidLDAPUR,
691737
FeaturePredictableSelectIsExpensive]>;
692738

693-
def TuneC1Nano : SubtargetFeature<"c1-nano", "ARMProcFamily",
694-
"C1Nano", "C1-Nano ARM Processors",[
695-
FeatureFuseAES,
696-
FeatureFuseAdrpAdd,
697-
FeaturePostRAScheduler,
698-
FeatureUseWzrToVecMove,
699-
FeatureUseFixedOverScalableIfEqualCost]>;
700-
701-
def TuneC1Pro : SubtargetFeature<"c1-pro", "ARMProcFamily",
702-
"C1Pro", "C1-Pro ARM Processors",[
703-
FeatureFuseAES,
704-
FeaturePostRAScheduler,
705-
FeatureCmpBccFusion,
706-
FeatureALULSLFast,
707-
FeatureFuseAdrpAdd,
708-
FeatureFuseCmpCSel,
709-
FeatureFuseCmpCSet,
710-
FeatureEnableSelectOptimize,
711-
FeaturePredictableSelectIsExpensive]>;
712-
713-
def TuneC1Premium : SubtargetFeature<"c1-premium", "ARMProcFamily",
714-
"C1Premium", "C1-Premium ARM Processors",[
715-
FeatureALULSLFast,
716-
FeatureFuseAdrpAdd,
717-
FeatureFuseAES,
718-
FeaturePostRAScheduler,
719-
FeatureEnableSelectOptimize,
720-
FeaturePredictableSelectIsExpensive]>;
721-
722-
def TuneC1Ultra : SubtargetFeature<"c1-ultra", "ARMProcFamily",
723-
"C1Ultra", "C1-Ultra ARM Processors",[
724-
FeatureALULSLFast,
725-
FeatureFuseAdrpAdd,
726-
FeatureFuseCmpCSel,
727-
FeatureFuseCmpCSet,
728-
FeatureFuseAES,
729-
FeaturePostRAScheduler,
730-
FeatureEnableSelectOptimize,
731-
FeatureUseFixedOverScalableIfEqualCost,
732-
FeatureAvoidLDAPUR,
733-
FeaturePredictableSelectIsExpensive]>;
734-
735739
def TuneSaphira : SubtargetFeature<"saphira", "ARMProcFamily", "Saphira",
736740
"Qualcomm Saphira processors", [
737741
FeaturePostRAScheduler,
@@ -908,6 +912,19 @@ def ProcessorFeatures {
908912
FeatureFPARMv8, FeatureFullFP16, FeatureMatMulInt8, FeatureJS,
909913
FeatureNEON, FeatureLSE, FeatureRAS, FeatureRCPC, FeatureRDM,
910914
FeatureDotProd, FeatureFPAC];
915+
list<SubtargetFeature> C1Nano = [HasV9_3aOps, FeatureNEON, FeatureCLRBHB,
916+
FeatureCHK, FeatureFPARMv8, FeaturePerfMon,
917+
FeatureSPECRES2, FeatureSSBS, FeatureRDM,
918+
FeatureVH, FeatureBF16, FeatureDotProd,
919+
FeatureFP16FML, FeatureFullFP16, FeatureMPAM,
920+
FeatureSVE, FeatureCCIDX, FeatureComplxNum,
921+
FeatureFPAC, FeatureJS, FeatureAM,
922+
FeatureRAS, FeatureSEL2, FeatureTRACEV8_4,
923+
FeatureAltFPCmp, FeatureFRInt3264,
924+
FeatureMTE, FeatureFineGrainedTraps,
925+
FeatureHCX, FeatureRCPC3, FeatureETE,
926+
FeatureSVEBitPerm, FeatureSVE2, FeatureTRBE,
927+
FeatureSME, FeatureSME2];
911928
list<SubtargetFeature> A65 = [HasV8_2aOps, FeatureSHA2, FeatureAES, FeatureFPARMv8,
912929
FeatureNEON, FeatureFullFP16, FeatureDotProd,
913930
FeatureRCPC, FeatureSSBS, FeatureRAS,
@@ -977,6 +994,20 @@ def ProcessorFeatures {
977994
FeatureDotProd, FeatureFPARMv8, FeatureFullFP16, FeatureMatMulInt8,
978995
FeatureJS, FeatureLSE, FeatureNEON, FeaturePAuth, FeatureRAS,
979996
FeatureRCPC, FeatureRDM, FeatureFPAC];
997+
list<SubtargetFeature> C1Pro = [HasV9_3aOps, FeatureNEON, FeatureCLRBHB,
998+
FeatureCHK, FeatureFPARMv8, FeaturePerfMon,
999+
FeatureSPECRES2, FeatureSSBS, FeatureRDM,
1000+
FeatureVH, FeatureBF16, FeatureDotProd,
1001+
FeatureFP16FML, FeatureFullFP16, FeatureMPAM,
1002+
FeatureSPE, FeatureSVE, FeatureCCIDX,
1003+
FeatureComplxNum, FeatureFPAC, FeatureJS,
1004+
FeatureAM, FeatureRAS, FeatureSEL2,
1005+
FeatureTRACEV8_4, FeatureAltFPCmp,
1006+
FeatureFRInt3264, FeatureMTE,
1007+
FeatureFineGrainedTraps, FeatureHCX,
1008+
FeatureSPE_EEF, FeatureRCPC3, FeatureETE,
1009+
FeatureSVEBitPerm, FeatureSVE2, FeatureTRBE,
1010+
FeatureSME, FeatureSME2];
9801011
list<SubtargetFeature> R82 = [HasV8_0rOps, FeaturePerfMon, FeatureFullFP16,
9811012
FeatureFP16FML, FeatureSSBS, FeaturePredRes,
9821013
FeatureSB, FeatureRDM, FeatureDotProd,
@@ -1036,6 +1067,36 @@ def ProcessorFeatures {
10361067
FeatureDotProd, FeatureFPARMv8, FeatureFullFP16, FeatureMatMulInt8,
10371068
FeatureJS, FeatureLSE, FeatureNEON, FeaturePAuth, FeatureRAS,
10381069
FeatureRCPC, FeatureRDM, FeatureFPAC];
1070+
list<SubtargetFeature> C1Ultra = [HasV9_3aOps, FeatureNEON, FeatureCLRBHB,
1071+
FeatureCHK, FeatureFPARMv8, FeaturePerfMon,
1072+
FeatureSPECRES2, FeatureSSBS, FeatureRDM,
1073+
FeatureVH, FeatureBF16, FeatureDotProd,
1074+
FeatureFP16FML, FeatureFullFP16,
1075+
FeatureMPAM, FeatureSPE, FeatureSVE,
1076+
FeatureCCIDX, FeatureComplxNum, FeatureFPAC,
1077+
FeatureJS, FeatureAM, FeatureRAS,
1078+
FeatureSEL2, FeatureTRACEV8_4,
1079+
FeatureAltFPCmp, FeatureFRInt3264,
1080+
FeatureMTE, FeatureFineGrainedTraps,
1081+
FeatureHCX, FeatureSPE_EEF, FeatureRCPC3,
1082+
FeatureETE, FeatureSVEBitPerm, FeatureSVE2,
1083+
FeatureTRBE, FeatureSME, FeatureSME2];
1084+
list<SubtargetFeature> C1Premium = [HasV9_3aOps, FeatureNEON, FeatureCLRBHB,
1085+
FeatureCHK, FeatureFPARMv8,
1086+
FeaturePerfMon, FeatureSPECRES2,
1087+
FeatureSSBS, FeatureRDM, FeatureVH,
1088+
FeatureBF16, FeatureDotProd,
1089+
FeatureFP16FML, FeatureFullFP16,
1090+
FeatureMPAM, FeatureSPE, FeatureSVE,
1091+
FeatureCCIDX, FeatureComplxNum,
1092+
FeatureFPAC, FeatureJS, FeatureAM,
1093+
FeatureRAS, FeatureSEL2, FeatureTRACEV8_4,
1094+
FeatureAltFPCmp, FeatureFRInt3264,
1095+
FeatureMTE, FeatureFineGrainedTraps,
1096+
FeatureHCX, FeatureSPE_EEF, FeatureRCPC3,
1097+
FeatureETE, FeatureSVEBitPerm,
1098+
FeatureSVE2, FeatureTRBE, FeatureSME,
1099+
FeatureSME2];
10391100
list<SubtargetFeature> A64FX = [HasV8_2aOps, FeatureFPARMv8, FeatureNEON,
10401101
FeatureSHA2, FeaturePerfMon, FeatureFullFP16,
10411102
FeatureSVE, FeatureComplxNum,
@@ -1218,63 +1279,6 @@ def ProcessorFeatures {
12181279
FeatureDotProd, FeatureFPARMv8, FeatureMatMulInt8, FeatureJS,
12191280
FeatureLSE, FeatureNEON, FeatureRAS, FeatureRCPC, FeatureRDM,
12201281
FeatureRME, FeatureFPAC];
1221-
list<SubtargetFeature> C1Nano = [HasV9_3aOps, FeatureNEON, FeatureCLRBHB,
1222-
FeatureCHK, FeatureFPARMv8, FeaturePerfMon,
1223-
FeatureSPECRES2, FeatureSSBS, FeatureRDM,
1224-
FeatureVH, FeatureBF16, FeatureDotProd,
1225-
FeatureFP16FML, FeatureFullFP16, FeatureMPAM,
1226-
FeatureSVE, FeatureCCIDX, FeatureComplxNum,
1227-
FeatureFPAC, FeatureJS, FeatureAM,
1228-
FeatureRAS, FeatureSEL2, FeatureTRACEV8_4,
1229-
FeatureAltFPCmp, FeatureFRInt3264,
1230-
FeatureMTE, FeatureFineGrainedTraps,
1231-
FeatureHCX, FeatureRCPC3, FeatureETE,
1232-
FeatureSVEBitPerm, FeatureSVE2, FeatureTRBE,
1233-
FeatureSME, FeatureSME2];
1234-
list<SubtargetFeature> C1Pro = [HasV9_3aOps, FeatureNEON, FeatureCLRBHB,
1235-
FeatureCHK, FeatureFPARMv8, FeaturePerfMon,
1236-
FeatureSPECRES2, FeatureSSBS, FeatureRDM,
1237-
FeatureVH, FeatureBF16, FeatureDotProd,
1238-
FeatureFP16FML, FeatureFullFP16, FeatureMPAM,
1239-
FeatureSPE, FeatureSVE, FeatureCCIDX,
1240-
FeatureComplxNum, FeatureFPAC, FeatureJS,
1241-
FeatureAM, FeatureRAS, FeatureSEL2,
1242-
FeatureTRACEV8_4, FeatureAltFPCmp,
1243-
FeatureFRInt3264, FeatureMTE,
1244-
FeatureFineGrainedTraps, FeatureHCX,
1245-
FeatureSPE_EEF, FeatureRCPC3, FeatureETE,
1246-
FeatureSVEBitPerm, FeatureSVE2, FeatureTRBE,
1247-
FeatureSME, FeatureSME2];
1248-
list<SubtargetFeature> C1Premium = [HasV9_3aOps, FeatureNEON, FeatureCLRBHB,
1249-
FeatureCHK, FeatureFPARMv8,
1250-
FeaturePerfMon, FeatureSPECRES2,
1251-
FeatureSSBS, FeatureRDM, FeatureVH,
1252-
FeatureBF16, FeatureDotProd,
1253-
FeatureFP16FML, FeatureFullFP16,
1254-
FeatureMPAM, FeatureSPE, FeatureSVE,
1255-
FeatureCCIDX, FeatureComplxNum,
1256-
FeatureFPAC, FeatureJS, FeatureAM,
1257-
FeatureRAS, FeatureSEL2, FeatureTRACEV8_4,
1258-
FeatureAltFPCmp, FeatureFRInt3264,
1259-
FeatureMTE, FeatureFineGrainedTraps,
1260-
FeatureHCX, FeatureSPE_EEF, FeatureRCPC3,
1261-
FeatureETE, FeatureSVEBitPerm,
1262-
FeatureSVE2, FeatureTRBE, FeatureSME,
1263-
FeatureSME2];
1264-
list<SubtargetFeature> C1Ultra = [HasV9_3aOps, FeatureNEON, FeatureCLRBHB,
1265-
FeatureCHK, FeatureFPARMv8, FeaturePerfMon,
1266-
FeatureSPECRES2, FeatureSSBS, FeatureRDM,
1267-
FeatureVH, FeatureBF16, FeatureDotProd,
1268-
FeatureFP16FML, FeatureFullFP16,
1269-
FeatureMPAM, FeatureSPE, FeatureSVE,
1270-
FeatureCCIDX, FeatureComplxNum, FeatureFPAC,
1271-
FeatureJS, FeatureAM, FeatureRAS,
1272-
FeatureSEL2, FeatureTRACEV8_4,
1273-
FeatureAltFPCmp, FeatureFRInt3264,
1274-
FeatureMTE, FeatureFineGrainedTraps,
1275-
FeatureHCX, FeatureSPE_EEF, FeatureRCPC3,
1276-
FeatureETE, FeatureSVEBitPerm, FeatureSVE2,
1277-
FeatureTRBE, FeatureSME, FeatureSME2];
12781282
list<SubtargetFeature> Saphira = [HasV8_4aOps, FeatureSHA2, FeatureAES, FeatureFPARMv8,
12791283
FeatureNEON, FeatureSPE, FeaturePerfMon, FeatureCRC,
12801284
FeatureCCIDX,
@@ -1364,6 +1368,8 @@ def : ProcessorModel<"cortex-a520", CortexA510Model, ProcessorFeatures.A520,
13641368
[TuneA520]>;
13651369
def : ProcessorModel<"cortex-a520ae", CortexA510Model, ProcessorFeatures.A520AE,
13661370
[TuneA520AE]>;
1371+
def : ProcessorModel<"c1-nano", NeoverseN2Model,
1372+
ProcessorFeatures.C1Nano, [TuneC1Nano]>;
13671373
def : ProcessorModel<"cortex-a57", CortexA57Model, ProcessorFeatures.A53,
13681374
[TuneA57]>;
13691375
def : ProcessorModel<"cortex-a65", CortexA53Model, ProcessorFeatures.A65,
@@ -1398,6 +1404,8 @@ def : ProcessorModel<"cortex-a720ae", NeoverseN2Model, ProcessorFeatures.A720AE,
13981404
[TuneA720AE]>;
13991405
def : ProcessorModel<"cortex-a725", NeoverseN3Model, ProcessorFeatures.A725,
14001406
[TuneA725]>;
1407+
def : ProcessorModel<"c1-pro", NeoverseN2Model,
1408+
ProcessorFeatures.C1Pro, [TuneC1Pro]>;
14011409
def : ProcessorModel<"cortex-r82", CortexA55Model, ProcessorFeatures.R82,
14021410
[TuneR82]>;
14031411
def : ProcessorModel<"cortex-r82ae", CortexA55Model, ProcessorFeatures.R82AE,
@@ -1414,6 +1422,10 @@ def : ProcessorModel<"cortex-x4", NeoverseV3Model, ProcessorFeatures.X4,
14141422
[TuneX4]>;
14151423
def : ProcessorModel<"cortex-x925", NeoverseV3Model, ProcessorFeatures.X925,
14161424
[TuneX925]>;
1425+
def : ProcessorModel<"c1-ultra", NeoverseV3Model,
1426+
ProcessorFeatures.C1Ultra, [TuneC1Ultra]>;
1427+
def : ProcessorModel<"c1-premium", NeoverseV3Model,
1428+
ProcessorFeatures.C1Premium, [TuneC1Premium]>;
14171429
def : ProcessorModel<"gb10", NeoverseV3Model, ProcessorFeatures.GB10,
14181430
[TuneX925]>;
14191431
def : ProcessorModel<"grace", NeoverseV2Model, ProcessorFeatures.Grace,
@@ -1437,14 +1449,6 @@ def : ProcessorModel<"neoverse-v3", NeoverseV3Model,
14371449
ProcessorFeatures.NeoverseV3, [TuneNeoverseV3]>;
14381450
def : ProcessorModel<"neoverse-v3ae", NeoverseV3AEModel,
14391451
ProcessorFeatures.NeoverseV3AE, [TuneNeoverseV3AE]>;
1440-
def : ProcessorModel<"c1-nano", CortexA510Model,
1441-
ProcessorFeatures.C1Nano, [TuneC1Nano]>;
1442-
def : ProcessorModel<"c1-pro", NeoverseN2Model,
1443-
ProcessorFeatures.C1Pro, [TuneC1Pro]>;
1444-
def : ProcessorModel<"c1-premium", NeoverseV3Model,
1445-
ProcessorFeatures.C1Premium, [TuneC1Premium]>;
1446-
def : ProcessorModel<"c1-ultra", NeoverseV3Model,
1447-
ProcessorFeatures.C1Ultra, [TuneC1Ultra]>;
14481452
def : ProcessorModel<"exynos-m3", ExynosM3Model, ProcessorFeatures.ExynosM3,
14491453
[TuneExynosM3]>;
14501454
def : ProcessorModel<"exynos-m4", ExynosM4Model, ProcessorFeatures.ExynosM4,

‎llvm/lib/Target/AArch64/AArch64Subtarget.cpp‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -187,13 +187,13 @@ void AArch64Subtarget::initializeProperties(bool HasMinSize) {
187187
case CortexA715:
188188
case CortexA720:
189189
case CortexA725:
190+
case C1Pro:
190191
case CortexX2:
191192
case CortexX3:
192193
case CortexX4:
193194
case CortexX925:
194-
case C1Premium:
195-
case C1Pro:
196195
case C1Ultra:
196+
case C1Premium:
197197
PrefFunctionAlignment = Align(16);
198198
VScaleForTuning = 1;
199199
PrefLoopAlignment = Align(32);

0 commit comments

Comments
 (0)