187 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 22; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sme2 -mattr=+fp8 -force-streaming -verify-machineinstrs < %s | FileCheck %s3 4; FSCALE (Single, x2)5 6define { <vscale x 8 x half>, <vscale x 8 x half> } @multi_vec_scale_single_x2_half( <vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x i16> %zm) {7; CHECK-LABEL: multi_vec_scale_single_x2_half:8; CHECK: // %bb.0:9; CHECK-NEXT: // kill: def $z1 killed $z1 killed $z0_z1 def $z0_z110; CHECK-NEXT: // kill: def $z0 killed $z0 killed $z0_z1 def $z0_z111; CHECK-NEXT: fscale { z0.h, z1.h }, { z0.h, z1.h }, z2.h12; CHECK-NEXT: ret13 %res = call { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.fp8.scale.single.x2.nxv8f16(<vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x i16> %zm)14 ret { <vscale x 8 x half>, <vscale x 8 x half> } %res15}16 17define { <vscale x 4 x float>, <vscale x 4 x float> } @multi_vec_scale_single_x2_float( <vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x i32> %zm) {18; CHECK-LABEL: multi_vec_scale_single_x2_float:19; CHECK: // %bb.0:20; CHECK-NEXT: // kill: def $z1 killed $z1 killed $z0_z1 def $z0_z121; CHECK-NEXT: // kill: def $z0 killed $z0 killed $z0_z1 def $z0_z122; CHECK-NEXT: fscale { z0.s, z1.s }, { z0.s, z1.s }, z2.s23; CHECK-NEXT: ret24 %res = call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.fp8.scale.single.x2.nxv4f32(<vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x i32> %zm)25 ret { <vscale x 4 x float>, <vscale x 4 x float> } %res26}27 28define { <vscale x 2 x double>, <vscale x 2 x double> } @multi_vec_scale_single_x2_double( <vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x i64> %zm) {29; CHECK-LABEL: multi_vec_scale_single_x2_double:30; CHECK: // %bb.0:31; CHECK-NEXT: // kill: def $z1 killed $z1 killed $z0_z1 def $z0_z132; CHECK-NEXT: // kill: def $z0 killed $z0 killed $z0_z1 def $z0_z133; CHECK-NEXT: fscale { z0.d, z1.d }, { z0.d, z1.d }, z2.d34; CHECK-NEXT: ret35 %res = call { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.fp8.scale.single.x2.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x i64> %zm)36 ret { <vscale x 2 x double>, <vscale x 2 x double> } %res37}38 39; FSCALE (Single, x4)40 41define { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @multi_vec_scale_single_x4_half( <vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zdn3, <vscale x 8 x half> %zdn4, <vscale x 8 x i16> %zm) {42; CHECK-LABEL: multi_vec_scale_single_x4_half:43; CHECK: // %bb.0:44; CHECK-NEXT: // kill: def $z3 killed $z3 killed $z0_z1_z2_z3 def $z0_z1_z2_z345; CHECK-NEXT: // kill: def $z2 killed $z2 killed $z0_z1_z2_z3 def $z0_z1_z2_z346; CHECK-NEXT: // kill: def $z1 killed $z1 killed $z0_z1_z2_z3 def $z0_z1_z2_z347; CHECK-NEXT: // kill: def $z0 killed $z0 killed $z0_z1_z2_z3 def $z0_z1_z2_z348; CHECK-NEXT: fscale { z0.h - z3.h }, { z0.h - z3.h }, z4.h49; CHECK-NEXT: ret50 %res = call { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.fp8.scale.single.x4.nxv8f16(<vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zdn3, <vscale x 8 x half> %zdn4, <vscale x 8 x i16> %zm)51 ret { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } %res52}53 54define { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @multi_vec_scale_single_x4_float( <vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zdn3, <vscale x 4 x float> %zdn4, <vscale x 4 x i32> %zm) {55; CHECK-LABEL: multi_vec_scale_single_x4_float:56; CHECK: // %bb.0:57; CHECK-NEXT: // kill: def $z3 killed $z3 killed $z0_z1_z2_z3 def $z0_z1_z2_z358; CHECK-NEXT: // kill: def $z2 killed $z2 killed $z0_z1_z2_z3 def $z0_z1_z2_z359; CHECK-NEXT: // kill: def $z1 killed $z1 killed $z0_z1_z2_z3 def $z0_z1_z2_z360; CHECK-NEXT: // kill: def $z0 killed $z0 killed $z0_z1_z2_z3 def $z0_z1_z2_z361; CHECK-NEXT: fscale { z0.s - z3.s }, { z0.s - z3.s }, z4.s62; CHECK-NEXT: ret63 %res = call { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.fp8.scale.single.x4.nxv4f32(<vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zdn3, <vscale x 4 x float> %zdn4, <vscale x 4 x i32> %zm)64 ret { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } %res65}66 67define { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @multi_vec_scale_single_x4_double( <vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4, <vscale x 2 x i64> %zm) {68; CHECK-LABEL: multi_vec_scale_single_x4_double:69; CHECK: // %bb.0:70; CHECK-NEXT: // kill: def $z3 killed $z3 killed $z0_z1_z2_z3 def $z0_z1_z2_z371; CHECK-NEXT: // kill: def $z2 killed $z2 killed $z0_z1_z2_z3 def $z0_z1_z2_z372; CHECK-NEXT: // kill: def $z1 killed $z1 killed $z0_z1_z2_z3 def $z0_z1_z2_z373; CHECK-NEXT: // kill: def $z0 killed $z0 killed $z0_z1_z2_z3 def $z0_z1_z2_z374; CHECK-NEXT: fscale { z0.d - z3.d }, { z0.d - z3.d }, z4.d75; CHECK-NEXT: ret76 %res = call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.fp8.scale.single.x4.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4, <vscale x 2 x i64> %zm)77 ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } %res78}79 80; FSCALE (Multi, x2)81define { <vscale x 8 x half>, <vscale x 8 x half> } @multi_vec_scale_x2_half( <vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x i16> %zm1, <vscale x 8 x i16> %zm2) {82; CHECK-LABEL: multi_vec_scale_x2_half:83; CHECK: // %bb.0:84; CHECK-NEXT: // kill: def $z1 killed $z1 killed $z0_z1 def $z0_z185; CHECK-NEXT: // kill: def $z3 killed $z3 killed $z2_z3 def $z2_z386; CHECK-NEXT: // kill: def $z0 killed $z0 killed $z0_z1 def $z0_z187; CHECK-NEXT: // kill: def $z2 killed $z2 killed $z2_z3 def $z2_z388; CHECK-NEXT: fscale { z0.h, z1.h }, { z0.h, z1.h }, { z2.h, z3.h }89; CHECK-NEXT: ret90 %res = call { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.fp8.scale.x2.nxv8f16(<vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x i16> %zm1, <vscale x 8 x i16> %zm2)91 ret { <vscale x 8 x half>, <vscale x 8 x half> } %res92}93 94define { <vscale x 4 x float>, <vscale x 4 x float> } @multi_vec_scale_x2_float( <vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x i32> %zm1, <vscale x 4 x i32> %zm2 ) {95; CHECK-LABEL: multi_vec_scale_x2_float:96; CHECK: // %bb.0:97; CHECK-NEXT: // kill: def $z1 killed $z1 killed $z0_z1 def $z0_z198; CHECK-NEXT: // kill: def $z3 killed $z3 killed $z2_z3 def $z2_z399; CHECK-NEXT: // kill: def $z0 killed $z0 killed $z0_z1 def $z0_z1100; CHECK-NEXT: // kill: def $z2 killed $z2 killed $z2_z3 def $z2_z3101; CHECK-NEXT: fscale { z0.s, z1.s }, { z0.s, z1.s }, { z2.s, z3.s }102; CHECK-NEXT: ret103 %res = call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.fp8.scale.x2.nxv4f32(<vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x i32> %zm1, <vscale x 4 x i32> %zm2)104 ret { <vscale x 4 x float>, <vscale x 4 x float> } %res105}106 107define { <vscale x 2 x double>, <vscale x 2 x double> } @multi_vec_scale_x2_double( <vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x i64> %zm1, <vscale x 2 x i64> %zm2) {108; CHECK-LABEL: multi_vec_scale_x2_double:109; CHECK: // %bb.0:110; CHECK-NEXT: // kill: def $z1 killed $z1 killed $z0_z1 def $z0_z1111; CHECK-NEXT: // kill: def $z3 killed $z3 killed $z2_z3 def $z2_z3112; CHECK-NEXT: // kill: def $z0 killed $z0 killed $z0_z1 def $z0_z1113; CHECK-NEXT: // kill: def $z2 killed $z2 killed $z2_z3 def $z2_z3114; CHECK-NEXT: fscale { z0.d, z1.d }, { z0.d, z1.d }, { z2.d, z3.d }115; CHECK-NEXT: ret116 %res = call { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.fp8.scale.x2.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x i64> %zm1, <vscale x 2 x i64> %zm2)117 ret { <vscale x 2 x double>, <vscale x 2 x double> } %res118}119 120; FSCALE (Multi, x4)121define { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @multi_vec_scale_x4_half( <vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zdn3, <vscale x 8 x half> %zdn4, <vscale x 8 x i16> %zm1, <vscale x 8 x i16> %zm2, <vscale x 8 x i16> %zm3, <vscale x 8 x i16> %zm4) {122; CHECK-LABEL: multi_vec_scale_x4_half:123; CHECK: // %bb.0:124; CHECK-NEXT: // kill: def $z3 killed $z3 killed $z0_z1_z2_z3 def $z0_z1_z2_z3125; CHECK-NEXT: // kill: def $z7 killed $z7 killed $z4_z5_z6_z7 def $z4_z5_z6_z7126; CHECK-NEXT: // kill: def $z2 killed $z2 killed $z0_z1_z2_z3 def $z0_z1_z2_z3127; CHECK-NEXT: // kill: def $z6 killed $z6 killed $z4_z5_z6_z7 def $z4_z5_z6_z7128; CHECK-NEXT: // kill: def $z1 killed $z1 killed $z0_z1_z2_z3 def $z0_z1_z2_z3129; CHECK-NEXT: // kill: def $z5 killed $z5 killed $z4_z5_z6_z7 def $z4_z5_z6_z7130; CHECK-NEXT: // kill: def $z0 killed $z0 killed $z0_z1_z2_z3 def $z0_z1_z2_z3131; CHECK-NEXT: // kill: def $z4 killed $z4 killed $z4_z5_z6_z7 def $z4_z5_z6_z7132; CHECK-NEXT: fscale { z0.h - z3.h }, { z0.h - z3.h }, { z4.h - z7.h }133; CHECK-NEXT: ret134 %res = call { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.fp8.scale.x4.nxv8f16(<vscale x 8 x half> %zdn1, <vscale x 8 x half> %zdn2, <vscale x 8 x half> %zdn3, <vscale x 8 x half> %zdn4, <vscale x 8 x i16> %zm1, <vscale x 8 x i16> %zm2, <vscale x 8 x i16> %zm3, <vscale x 8 x i16> %zm4)135 ret { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } %res136}137 138define { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @multi_vec_scale_x4_float( <vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zdn3, <vscale x 4 x float> %zdn4, <vscale x 4 x i32> %zm1, <vscale x 4 x i32> %zm2, <vscale x 4 x i32> %zm3, <vscale x 4 x i32> %zm4) {139; CHECK-LABEL: multi_vec_scale_x4_float:140; CHECK: // %bb.0:141; CHECK-NEXT: // kill: def $z3 killed $z3 killed $z0_z1_z2_z3 def $z0_z1_z2_z3142; CHECK-NEXT: // kill: def $z7 killed $z7 killed $z4_z5_z6_z7 def $z4_z5_z6_z7143; CHECK-NEXT: // kill: def $z2 killed $z2 killed $z0_z1_z2_z3 def $z0_z1_z2_z3144; CHECK-NEXT: // kill: def $z6 killed $z6 killed $z4_z5_z6_z7 def $z4_z5_z6_z7145; CHECK-NEXT: // kill: def $z1 killed $z1 killed $z0_z1_z2_z3 def $z0_z1_z2_z3146; CHECK-NEXT: // kill: def $z5 killed $z5 killed $z4_z5_z6_z7 def $z4_z5_z6_z7147; CHECK-NEXT: // kill: def $z0 killed $z0 killed $z0_z1_z2_z3 def $z0_z1_z2_z3148; CHECK-NEXT: // kill: def $z4 killed $z4 killed $z4_z5_z6_z7 def $z4_z5_z6_z7149; CHECK-NEXT: fscale { z0.s - z3.s }, { z0.s - z3.s }, { z4.s - z7.s }150; CHECK-NEXT: ret151 %res = call { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.fp8.scale.x4.nxv4f32(<vscale x 4 x float> %zdn1, <vscale x 4 x float> %zdn2, <vscale x 4 x float> %zdn3, <vscale x 4 x float> %zdn4, <vscale x 4 x i32> %zm1, <vscale x 4 x i32> %zm2, <vscale x 4 x i32> %zm3, <vscale x 4 x i32> %zm4)152 ret { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } %res153}154 155define { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @multi_vec_scale_x4_double( <vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4, <vscale x 2 x i64> %zm1, <vscale x 2 x i64> %zm2, <vscale x 2 x i64> %zm3, <vscale x 2 x i64> %zm4) {156; CHECK-LABEL: multi_vec_scale_x4_double:157; CHECK: // %bb.0:158; CHECK-NEXT: // kill: def $z3 killed $z3 killed $z0_z1_z2_z3 def $z0_z1_z2_z3159; CHECK-NEXT: // kill: def $z7 killed $z7 killed $z4_z5_z6_z7 def $z4_z5_z6_z7160; CHECK-NEXT: // kill: def $z2 killed $z2 killed $z0_z1_z2_z3 def $z0_z1_z2_z3161; CHECK-NEXT: // kill: def $z6 killed $z6 killed $z4_z5_z6_z7 def $z4_z5_z6_z7162; CHECK-NEXT: // kill: def $z1 killed $z1 killed $z0_z1_z2_z3 def $z0_z1_z2_z3163; CHECK-NEXT: // kill: def $z5 killed $z5 killed $z4_z5_z6_z7 def $z4_z5_z6_z7164; CHECK-NEXT: // kill: def $z0 killed $z0 killed $z0_z1_z2_z3 def $z0_z1_z2_z3165; CHECK-NEXT: // kill: def $z4 killed $z4 killed $z4_z5_z6_z7 def $z4_z5_z6_z7166; CHECK-NEXT: fscale { z0.d - z3.d }, { z0.d - z3.d }, { z4.d - z7.d }167; CHECK-NEXT: ret168 %res = call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.fp8.scale.x4.nxv2f64(<vscale x 2 x double> %zdn1, <vscale x 2 x double> %zdn2, <vscale x 2 x double> %zdn3, <vscale x 2 x double> %zdn4, <vscale x 2 x i64> %zm1, <vscale x 2 x i64> %zm2, <vscale x 2 x i64> %zm3, <vscale x 2 x i64> %zm4)169 ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } %res170}171 172declare { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.fp8.scale.single.x2.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x i16>)173declare { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.fp8.scale.single.x2.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x i32>)174declare { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.fp8.scale.single.x2.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x i64>)175 176declare { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.fp8.scale.single.x4.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>,<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x i16>)177declare { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.fp8.scale.single.x4.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>,<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x i32>)178declare { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.fp8.scale.single.x4.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>,<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x i64>)179 180declare { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.fp8.scale.x2.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x i16>, <vscale x 8 x i16>)181declare { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.fp8.scale.x2.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x i32>, <vscale x 4 x i32>)182declare { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.fp8.scale.x2.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x i64>, <vscale x 2 x i64>)183 184declare { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.fp8.scale.x4.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>,<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)185declare { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.fp8.scale.x4.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>,<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)186declare { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.fp8.scale.x4.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>,<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)187