541 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -mtriple=x86_64-unknown -passes=slp-vectorizer -S -mcpu=x86-64 | FileCheck %s --check-prefixes=CHECK,SCALAR3; RUN: opt < %s -mtriple=x86_64-unknown -passes=slp-vectorizer -S -mcpu=x86-64-v2 -mattr=+prefer-128-bit | FileCheck %s --check-prefixes=CHECK,VEC1284; RUN: opt < %s -mtriple=x86_64-unknown -passes=slp-vectorizer -S -mcpu=x86-64-v2 -mattr=-prefer-128-bit | FileCheck %s --check-prefixes=CHECK,VEC1285; RUN: opt < %s -mtriple=x86_64-unknown -passes=slp-vectorizer -S -mcpu=x86-64-v3 -mattr=+prefer-128-bit | FileCheck %s --check-prefixes=CHECK,VEC1286; RUN: opt < %s -mtriple=x86_64-unknown -passes=slp-vectorizer -S -mcpu=x86-64-v3 -mattr=-prefer-128-bit | FileCheck %s --check-prefixes=CHECK,VEC256,VEC256-AVX27; RUN: opt < %s -mtriple=x86_64-unknown -passes=slp-vectorizer -S -mcpu=x86-64-v4 -mattr=+prefer-256-bit | FileCheck %s --check-prefixes=CHECK,VEC256,VEC256-AVX5128; RUN: opt < %s -mtriple=x86_64-unknown -passes=slp-vectorizer -S -mcpu=x86-64-v4 -mattr=-prefer-256-bit | FileCheck %s --check-prefixes=CHECK,VEC5129 10@f64 = common global [16 x double] zeroinitializer, align 6411@f32 = common global [16 x float] zeroinitializer, align 6412@r64 = common global [16 x i64] zeroinitializer, align 6413@r32 = common global [16 x i32] zeroinitializer, align 6414 15define void @rint_v8f32_v8f32() {16; SCALAR-LABEL: @rint_v8f32_v8f32(17; SCALAR-NEXT: [[A0:%.*]] = load float, ptr @f32, align 818; SCALAR-NEXT: [[A1:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 1), align 819; SCALAR-NEXT: [[A2:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 2), align 820; SCALAR-NEXT: [[A3:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 3), align 821; SCALAR-NEXT: [[A4:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 822; SCALAR-NEXT: [[A5:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 5), align 823; SCALAR-NEXT: [[A6:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 6), align 824; SCALAR-NEXT: [[A7:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 7), align 825; SCALAR-NEXT: [[R0:%.*]] = call float @llvm.rint.f32(float [[A0]])26; SCALAR-NEXT: [[R1:%.*]] = call float @llvm.rint.f32(float [[A1]])27; SCALAR-NEXT: [[R2:%.*]] = call float @llvm.rint.f32(float [[A2]])28; SCALAR-NEXT: [[R3:%.*]] = call float @llvm.rint.f32(float [[A3]])29; SCALAR-NEXT: [[R4:%.*]] = call float @llvm.rint.f32(float [[A4]])30; SCALAR-NEXT: [[R5:%.*]] = call float @llvm.rint.f32(float [[A5]])31; SCALAR-NEXT: [[R6:%.*]] = call float @llvm.rint.f32(float [[A6]])32; SCALAR-NEXT: [[R7:%.*]] = call float @llvm.rint.f32(float [[A7]])33; SCALAR-NEXT: store float [[R0]], ptr @f32, align 834; SCALAR-NEXT: store float [[R1]], ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 1), align 835; SCALAR-NEXT: store float [[R2]], ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 2), align 836; SCALAR-NEXT: store float [[R3]], ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 3), align 837; SCALAR-NEXT: store float [[R4]], ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 838; SCALAR-NEXT: store float [[R5]], ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 5), align 839; SCALAR-NEXT: store float [[R6]], ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 6), align 840; SCALAR-NEXT: store float [[R7]], ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 7), align 841; SCALAR-NEXT: ret void42;43; VEC128-LABEL: @rint_v8f32_v8f32(44; VEC128-NEXT: [[TMP1:%.*]] = load <4 x float>, ptr @f32, align 845; VEC128-NEXT: [[TMP2:%.*]] = call <4 x float> @llvm.rint.v4f32(<4 x float> [[TMP1]])46; VEC128-NEXT: store <4 x float> [[TMP2]], ptr @f32, align 847; VEC128-NEXT: [[TMP3:%.*]] = load <4 x float>, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 848; VEC128-NEXT: [[TMP4:%.*]] = call <4 x float> @llvm.rint.v4f32(<4 x float> [[TMP3]])49; VEC128-NEXT: store <4 x float> [[TMP4]], ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 850; VEC128-NEXT: ret void51;52; VEC256-LABEL: @rint_v8f32_v8f32(53; VEC256-NEXT: [[TMP1:%.*]] = load <8 x float>, ptr @f32, align 854; VEC256-NEXT: [[TMP2:%.*]] = call <8 x float> @llvm.rint.v8f32(<8 x float> [[TMP1]])55; VEC256-NEXT: store <8 x float> [[TMP2]], ptr @f32, align 856; VEC256-NEXT: ret void57;58; VEC512-LABEL: @rint_v8f32_v8f32(59; VEC512-NEXT: [[TMP1:%.*]] = load <8 x float>, ptr @f32, align 860; VEC512-NEXT: [[TMP2:%.*]] = call <8 x float> @llvm.rint.v8f32(<8 x float> [[TMP1]])61; VEC512-NEXT: store <8 x float> [[TMP2]], ptr @f32, align 862; VEC512-NEXT: ret void63;64 %a0 = load float, ptr @f32, align 865 %a1 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 1), align 866 %a2 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 2), align 867 %a3 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 3), align 868 %a4 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 869 %a5 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 5), align 870 %a6 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 6), align 871 %a7 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 7), align 872 %r0 = call float @llvm.rint.i32.f32(float %a0)73 %r1 = call float @llvm.rint.i32.f32(float %a1)74 %r2 = call float @llvm.rint.i32.f32(float %a2)75 %r3 = call float @llvm.rint.i32.f32(float %a3)76 %r4 = call float @llvm.rint.i32.f32(float %a4)77 %r5 = call float @llvm.rint.i32.f32(float %a5)78 %r6 = call float @llvm.rint.i32.f32(float %a6)79 %r7 = call float @llvm.rint.i32.f32(float %a7)80 store float %r0, ptr @f32, align 881 store float %r1, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 1), align 882 store float %r2, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 2), align 883 store float %r3, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 3), align 884 store float %r4, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 885 store float %r5, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 5), align 886 store float %r6, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 6), align 887 store float %r7, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 7), align 888 ret void89}90 91define void @rint_v8f64_v8f64() {92; SCALAR-LABEL: @rint_v8f64_v8f64(93; SCALAR-NEXT: [[A0:%.*]] = load double, ptr @f64, align 894; SCALAR-NEXT: [[A1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 1), align 895; SCALAR-NEXT: [[A2:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 2), align 896; SCALAR-NEXT: [[A3:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 3), align 897; SCALAR-NEXT: [[A4:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 898; SCALAR-NEXT: [[A5:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 5), align 899; SCALAR-NEXT: [[A6:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 6), align 8100; SCALAR-NEXT: [[A7:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 7), align 8101; SCALAR-NEXT: [[R0:%.*]] = call double @llvm.rint.f64(double [[A0]])102; SCALAR-NEXT: [[R1:%.*]] = call double @llvm.rint.f64(double [[A1]])103; SCALAR-NEXT: [[R2:%.*]] = call double @llvm.rint.f64(double [[A2]])104; SCALAR-NEXT: [[R3:%.*]] = call double @llvm.rint.f64(double [[A3]])105; SCALAR-NEXT: [[R4:%.*]] = call double @llvm.rint.f64(double [[A4]])106; SCALAR-NEXT: [[R5:%.*]] = call double @llvm.rint.f64(double [[A5]])107; SCALAR-NEXT: [[R6:%.*]] = call double @llvm.rint.f64(double [[A6]])108; SCALAR-NEXT: [[R7:%.*]] = call double @llvm.rint.f64(double [[A7]])109; SCALAR-NEXT: store double [[R0]], ptr @f64, align 8110; SCALAR-NEXT: store double [[R1]], ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 1), align 8111; SCALAR-NEXT: store double [[R2]], ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 2), align 8112; SCALAR-NEXT: store double [[R3]], ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 3), align 8113; SCALAR-NEXT: store double [[R4]], ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8114; SCALAR-NEXT: store double [[R5]], ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 5), align 8115; SCALAR-NEXT: store double [[R6]], ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 6), align 8116; SCALAR-NEXT: store double [[R7]], ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 7), align 8117; SCALAR-NEXT: ret void118;119; VEC128-LABEL: @rint_v8f64_v8f64(120; VEC128-NEXT: [[TMP1:%.*]] = load <2 x double>, ptr @f64, align 8121; VEC128-NEXT: [[TMP2:%.*]] = call <2 x double> @llvm.rint.v2f64(<2 x double> [[TMP1]])122; VEC128-NEXT: store <2 x double> [[TMP2]], ptr @f64, align 8123; VEC128-NEXT: [[TMP3:%.*]] = load <2 x double>, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 2), align 8124; VEC128-NEXT: [[TMP4:%.*]] = call <2 x double> @llvm.rint.v2f64(<2 x double> [[TMP3]])125; VEC128-NEXT: store <2 x double> [[TMP4]], ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 2), align 8126; VEC128-NEXT: [[TMP5:%.*]] = load <2 x double>, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8127; VEC128-NEXT: [[TMP6:%.*]] = call <2 x double> @llvm.rint.v2f64(<2 x double> [[TMP5]])128; VEC128-NEXT: store <2 x double> [[TMP6]], ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8129; VEC128-NEXT: [[TMP7:%.*]] = load <2 x double>, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 6), align 8130; VEC128-NEXT: [[TMP8:%.*]] = call <2 x double> @llvm.rint.v2f64(<2 x double> [[TMP7]])131; VEC128-NEXT: store <2 x double> [[TMP8]], ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 6), align 8132; VEC128-NEXT: ret void133;134; VEC256-LABEL: @rint_v8f64_v8f64(135; VEC256-NEXT: [[TMP1:%.*]] = load <4 x double>, ptr @f64, align 8136; VEC256-NEXT: [[TMP2:%.*]] = call <4 x double> @llvm.rint.v4f64(<4 x double> [[TMP1]])137; VEC256-NEXT: store <4 x double> [[TMP2]], ptr @f64, align 8138; VEC256-NEXT: [[TMP3:%.*]] = load <4 x double>, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8139; VEC256-NEXT: [[TMP4:%.*]] = call <4 x double> @llvm.rint.v4f64(<4 x double> [[TMP3]])140; VEC256-NEXT: store <4 x double> [[TMP4]], ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8141; VEC256-NEXT: ret void142;143; VEC512-LABEL: @rint_v8f64_v8f64(144; VEC512-NEXT: [[TMP1:%.*]] = load <8 x double>, ptr @f64, align 8145; VEC512-NEXT: [[TMP2:%.*]] = call <8 x double> @llvm.rint.v8f64(<8 x double> [[TMP1]])146; VEC512-NEXT: store <8 x double> [[TMP2]], ptr @f64, align 8147; VEC512-NEXT: ret void148;149 %a0 = load double, ptr @f64, align 8150 %a1 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 1), align 8151 %a2 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 2), align 8152 %a3 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 3), align 8153 %a4 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8154 %a5 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 5), align 8155 %a6 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 6), align 8156 %a7 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 7), align 8157 %r0 = call double @llvm.rint.f64.f64(double %a0)158 %r1 = call double @llvm.rint.f64.f64(double %a1)159 %r2 = call double @llvm.rint.f64.f64(double %a2)160 %r3 = call double @llvm.rint.f64.f64(double %a3)161 %r4 = call double @llvm.rint.f64.f64(double %a4)162 %r5 = call double @llvm.rint.f64.f64(double %a5)163 %r6 = call double @llvm.rint.f64.f64(double %a6)164 %r7 = call double @llvm.rint.f64.f64(double %a7)165 store double %r0, ptr @f64, align 8166 store double %r1, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 1), align 8167 store double %r2, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 2), align 8168 store double %r3, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 3), align 8169 store double %r4, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8170 store double %r5, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 5), align 8171 store double %r6, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 6), align 8172 store double %r7, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 7), align 8173 ret void174}175 176define void @lrint_v8f32_v8i32() {177; SCALAR-LABEL: @lrint_v8f32_v8i32(178; SCALAR-NEXT: [[TMP1:%.*]] = load <4 x float>, ptr @f32, align 8179; SCALAR-NEXT: [[TMP2:%.*]] = call <4 x i32> @llvm.lrint.v4i32.v4f32(<4 x float> [[TMP1]])180; SCALAR-NEXT: store <4 x i32> [[TMP2]], ptr @r32, align 8181; SCALAR-NEXT: [[TMP3:%.*]] = load <4 x float>, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 8182; SCALAR-NEXT: [[TMP4:%.*]] = call <4 x i32> @llvm.lrint.v4i32.v4f32(<4 x float> [[TMP3]])183; SCALAR-NEXT: store <4 x i32> [[TMP4]], ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 4), align 8184; SCALAR-NEXT: ret void185;186; VEC128-LABEL: @lrint_v8f32_v8i32(187; VEC128-NEXT: [[TMP1:%.*]] = load <4 x float>, ptr @f32, align 8188; VEC128-NEXT: [[TMP2:%.*]] = call <4 x i32> @llvm.lrint.v4i32.v4f32(<4 x float> [[TMP1]])189; VEC128-NEXT: store <4 x i32> [[TMP2]], ptr @r32, align 8190; VEC128-NEXT: [[TMP3:%.*]] = load <4 x float>, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 8191; VEC128-NEXT: [[TMP4:%.*]] = call <4 x i32> @llvm.lrint.v4i32.v4f32(<4 x float> [[TMP3]])192; VEC128-NEXT: store <4 x i32> [[TMP4]], ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 4), align 8193; VEC128-NEXT: ret void194;195; VEC256-LABEL: @lrint_v8f32_v8i32(196; VEC256-NEXT: [[TMP1:%.*]] = load <8 x float>, ptr @f32, align 8197; VEC256-NEXT: [[TMP2:%.*]] = call <8 x i32> @llvm.lrint.v8i32.v8f32(<8 x float> [[TMP1]])198; VEC256-NEXT: store <8 x i32> [[TMP2]], ptr @r32, align 8199; VEC256-NEXT: ret void200;201; VEC512-LABEL: @lrint_v8f32_v8i32(202; VEC512-NEXT: [[TMP1:%.*]] = load <8 x float>, ptr @f32, align 8203; VEC512-NEXT: [[TMP2:%.*]] = call <8 x i32> @llvm.lrint.v8i32.v8f32(<8 x float> [[TMP1]])204; VEC512-NEXT: store <8 x i32> [[TMP2]], ptr @r32, align 8205; VEC512-NEXT: ret void206;207 %a0 = load float, ptr @f32, align 8208 %a1 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 1), align 8209 %a2 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 2), align 8210 %a3 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 3), align 8211 %a4 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 8212 %a5 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 5), align 8213 %a6 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 6), align 8214 %a7 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 7), align 8215 %r0 = call i32 @llvm.lrint.i32.f32(float %a0)216 %r1 = call i32 @llvm.lrint.i32.f32(float %a1)217 %r2 = call i32 @llvm.lrint.i32.f32(float %a2)218 %r3 = call i32 @llvm.lrint.i32.f32(float %a3)219 %r4 = call i32 @llvm.lrint.i32.f32(float %a4)220 %r5 = call i32 @llvm.lrint.i32.f32(float %a5)221 %r6 = call i32 @llvm.lrint.i32.f32(float %a6)222 %r7 = call i32 @llvm.lrint.i32.f32(float %a7)223 store i32 %r0, ptr @r32, align 8224 store i32 %r1, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 1), align 8225 store i32 %r2, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 2), align 8226 store i32 %r3, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 3), align 8227 store i32 %r4, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 4), align 8228 store i32 %r5, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 5), align 8229 store i32 %r6, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 6), align 8230 store i32 %r7, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 7), align 8231 ret void232}233 234define void @lrint_v8f64_v8i32() {235; SCALAR-LABEL: @lrint_v8f64_v8i32(236; SCALAR-NEXT: [[TMP1:%.*]] = load <4 x double>, ptr @f64, align 8237; SCALAR-NEXT: [[TMP2:%.*]] = call <4 x i32> @llvm.lrint.v4i32.v4f64(<4 x double> [[TMP1]])238; SCALAR-NEXT: store <4 x i32> [[TMP2]], ptr @r32, align 8239; SCALAR-NEXT: [[TMP3:%.*]] = load <4 x double>, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8240; SCALAR-NEXT: [[TMP4:%.*]] = call <4 x i32> @llvm.lrint.v4i32.v4f64(<4 x double> [[TMP3]])241; SCALAR-NEXT: store <4 x i32> [[TMP4]], ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 4), align 8242; SCALAR-NEXT: ret void243;244; VEC128-LABEL: @lrint_v8f64_v8i32(245; VEC128-NEXT: [[TMP1:%.*]] = load <4 x double>, ptr @f64, align 8246; VEC128-NEXT: [[TMP2:%.*]] = call <4 x i32> @llvm.lrint.v4i32.v4f64(<4 x double> [[TMP1]])247; VEC128-NEXT: store <4 x i32> [[TMP2]], ptr @r32, align 8248; VEC128-NEXT: [[TMP3:%.*]] = load <4 x double>, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8249; VEC128-NEXT: [[TMP4:%.*]] = call <4 x i32> @llvm.lrint.v4i32.v4f64(<4 x double> [[TMP3]])250; VEC128-NEXT: store <4 x i32> [[TMP4]], ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 4), align 8251; VEC128-NEXT: ret void252;253; VEC256-LABEL: @lrint_v8f64_v8i32(254; VEC256-NEXT: [[TMP1:%.*]] = load <8 x double>, ptr @f64, align 8255; VEC256-NEXT: [[TMP2:%.*]] = call <8 x i32> @llvm.lrint.v8i32.v8f64(<8 x double> [[TMP1]])256; VEC256-NEXT: store <8 x i32> [[TMP2]], ptr @r32, align 8257; VEC256-NEXT: ret void258;259; VEC512-LABEL: @lrint_v8f64_v8i32(260; VEC512-NEXT: [[TMP1:%.*]] = load <8 x double>, ptr @f64, align 8261; VEC512-NEXT: [[TMP2:%.*]] = call <8 x i32> @llvm.lrint.v8i32.v8f64(<8 x double> [[TMP1]])262; VEC512-NEXT: store <8 x i32> [[TMP2]], ptr @r32, align 8263; VEC512-NEXT: ret void264;265 %a0 = load double, ptr @f64, align 8266 %a1 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 1), align 8267 %a2 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 2), align 8268 %a3 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 3), align 8269 %a4 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8270 %a5 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 5), align 8271 %a6 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 6), align 8272 %a7 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 7), align 8273 %r0 = call i32 @llvm.lrint.i32.f64(double %a0)274 %r1 = call i32 @llvm.lrint.i32.f64(double %a1)275 %r2 = call i32 @llvm.lrint.i32.f64(double %a2)276 %r3 = call i32 @llvm.lrint.i32.f64(double %a3)277 %r4 = call i32 @llvm.lrint.i32.f64(double %a4)278 %r5 = call i32 @llvm.lrint.i32.f64(double %a5)279 %r6 = call i32 @llvm.lrint.i32.f64(double %a6)280 %r7 = call i32 @llvm.lrint.i32.f64(double %a7)281 store i32 %r0, ptr @r32, align 8282 store i32 %r1, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 1), align 8283 store i32 %r2, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 2), align 8284 store i32 %r3, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 3), align 8285 store i32 %r4, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 4), align 8286 store i32 %r5, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 5), align 8287 store i32 %r6, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 6), align 8288 store i32 %r7, ptr getelementptr inbounds ([8 x i32], ptr @r32, i32 0, i32 7), align 8289 ret void290}291 292define void @llrint_v8f32_v8i64() {293; SCALAR-LABEL: @llrint_v8f32_v8i64(294; SCALAR-NEXT: [[A0:%.*]] = load float, ptr @f32, align 8295; SCALAR-NEXT: [[A1:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 1), align 8296; SCALAR-NEXT: [[A2:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 2), align 8297; SCALAR-NEXT: [[A3:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 3), align 8298; SCALAR-NEXT: [[A4:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 8299; SCALAR-NEXT: [[A5:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 5), align 8300; SCALAR-NEXT: [[A6:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 6), align 8301; SCALAR-NEXT: [[A7:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 7), align 8302; SCALAR-NEXT: [[R0:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A0]])303; SCALAR-NEXT: [[R1:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A1]])304; SCALAR-NEXT: [[R2:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A2]])305; SCALAR-NEXT: [[R3:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A3]])306; SCALAR-NEXT: [[R4:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A4]])307; SCALAR-NEXT: [[R5:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A5]])308; SCALAR-NEXT: [[R6:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A6]])309; SCALAR-NEXT: [[R7:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A7]])310; SCALAR-NEXT: store i64 [[R0]], ptr @r64, align 8311; SCALAR-NEXT: store i64 [[R1]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 1), align 8312; SCALAR-NEXT: store i64 [[R2]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 2), align 8313; SCALAR-NEXT: store i64 [[R3]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 3), align 8314; SCALAR-NEXT: store i64 [[R4]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 4), align 8315; SCALAR-NEXT: store i64 [[R5]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 5), align 8316; SCALAR-NEXT: store i64 [[R6]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 6), align 8317; SCALAR-NEXT: store i64 [[R7]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 7), align 8318; SCALAR-NEXT: ret void319;320; VEC128-LABEL: @llrint_v8f32_v8i64(321; VEC128-NEXT: [[A0:%.*]] = load float, ptr @f32, align 8322; VEC128-NEXT: [[A1:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 1), align 8323; VEC128-NEXT: [[A2:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 2), align 8324; VEC128-NEXT: [[A3:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 3), align 8325; VEC128-NEXT: [[A4:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 8326; VEC128-NEXT: [[A5:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 5), align 8327; VEC128-NEXT: [[A6:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 6), align 8328; VEC128-NEXT: [[A7:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 7), align 8329; VEC128-NEXT: [[R0:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A0]])330; VEC128-NEXT: [[R1:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A1]])331; VEC128-NEXT: [[R2:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A2]])332; VEC128-NEXT: [[R3:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A3]])333; VEC128-NEXT: [[R4:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A4]])334; VEC128-NEXT: [[R5:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A5]])335; VEC128-NEXT: [[R6:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A6]])336; VEC128-NEXT: [[R7:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A7]])337; VEC128-NEXT: store i64 [[R0]], ptr @r64, align 8338; VEC128-NEXT: store i64 [[R1]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 1), align 8339; VEC128-NEXT: store i64 [[R2]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 2), align 8340; VEC128-NEXT: store i64 [[R3]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 3), align 8341; VEC128-NEXT: store i64 [[R4]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 4), align 8342; VEC128-NEXT: store i64 [[R5]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 5), align 8343; VEC128-NEXT: store i64 [[R6]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 6), align 8344; VEC128-NEXT: store i64 [[R7]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 7), align 8345; VEC128-NEXT: ret void346;347; VEC256-AVX2-LABEL: @llrint_v8f32_v8i64(348; VEC256-AVX2-NEXT: [[A0:%.*]] = load float, ptr @f32, align 8349; VEC256-AVX2-NEXT: [[A1:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 1), align 8350; VEC256-AVX2-NEXT: [[A2:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 2), align 8351; VEC256-AVX2-NEXT: [[A3:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 3), align 8352; VEC256-AVX2-NEXT: [[A4:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 8353; VEC256-AVX2-NEXT: [[A5:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 5), align 8354; VEC256-AVX2-NEXT: [[A6:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 6), align 8355; VEC256-AVX2-NEXT: [[A7:%.*]] = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 7), align 8356; VEC256-AVX2-NEXT: [[R0:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A0]])357; VEC256-AVX2-NEXT: [[R1:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A1]])358; VEC256-AVX2-NEXT: [[R2:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A2]])359; VEC256-AVX2-NEXT: [[R3:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A3]])360; VEC256-AVX2-NEXT: [[R4:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A4]])361; VEC256-AVX2-NEXT: [[R5:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A5]])362; VEC256-AVX2-NEXT: [[R6:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A6]])363; VEC256-AVX2-NEXT: [[R7:%.*]] = call i64 @llvm.llrint.i64.f32(float [[A7]])364; VEC256-AVX2-NEXT: store i64 [[R0]], ptr @r64, align 8365; VEC256-AVX2-NEXT: store i64 [[R1]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 1), align 8366; VEC256-AVX2-NEXT: store i64 [[R2]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 2), align 8367; VEC256-AVX2-NEXT: store i64 [[R3]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 3), align 8368; VEC256-AVX2-NEXT: store i64 [[R4]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 4), align 8369; VEC256-AVX2-NEXT: store i64 [[R5]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 5), align 8370; VEC256-AVX2-NEXT: store i64 [[R6]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 6), align 8371; VEC256-AVX2-NEXT: store i64 [[R7]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 7), align 8372; VEC256-AVX2-NEXT: ret void373;374; VEC256-AVX512-LABEL: @llrint_v8f32_v8i64(375; VEC256-AVX512-NEXT: [[TMP1:%.*]] = load <4 x float>, ptr @f32, align 8376; VEC256-AVX512-NEXT: [[TMP2:%.*]] = call <4 x i64> @llvm.llrint.v4i64.v4f32(<4 x float> [[TMP1]])377; VEC256-AVX512-NEXT: store <4 x i64> [[TMP2]], ptr @r64, align 8378; VEC256-AVX512-NEXT: [[TMP3:%.*]] = load <4 x float>, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 8379; VEC256-AVX512-NEXT: [[TMP4:%.*]] = call <4 x i64> @llvm.llrint.v4i64.v4f32(<4 x float> [[TMP3]])380; VEC256-AVX512-NEXT: store <4 x i64> [[TMP4]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 4), align 8381; VEC256-AVX512-NEXT: ret void382;383; VEC512-LABEL: @llrint_v8f32_v8i64(384; VEC512-NEXT: [[TMP1:%.*]] = load <8 x float>, ptr @f32, align 8385; VEC512-NEXT: [[TMP2:%.*]] = call <8 x i64> @llvm.llrint.v8i64.v8f32(<8 x float> [[TMP1]])386; VEC512-NEXT: store <8 x i64> [[TMP2]], ptr @r64, align 8387; VEC512-NEXT: ret void388;389 %a0 = load float, ptr @f32, align 8390 %a1 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 1), align 8391 %a2 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 2), align 8392 %a3 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 3), align 8393 %a4 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 4), align 8394 %a5 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 5), align 8395 %a6 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 6), align 8396 %a7 = load float, ptr getelementptr inbounds ([8 x float], ptr @f32, i32 0, i32 7), align 8397 %r0 = call i64 @llvm.llrint.i64.f32(float %a0)398 %r1 = call i64 @llvm.llrint.i64.f32(float %a1)399 %r2 = call i64 @llvm.llrint.i64.f32(float %a2)400 %r3 = call i64 @llvm.llrint.i64.f32(float %a3)401 %r4 = call i64 @llvm.llrint.i64.f32(float %a4)402 %r5 = call i64 @llvm.llrint.i64.f32(float %a5)403 %r6 = call i64 @llvm.llrint.i64.f32(float %a6)404 %r7 = call i64 @llvm.llrint.i64.f32(float %a7)405 store i64 %r0, ptr @r64, align 8406 store i64 %r1, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 1), align 8407 store i64 %r2, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 2), align 8408 store i64 %r3, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 3), align 8409 store i64 %r4, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 4), align 8410 store i64 %r5, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 5), align 8411 store i64 %r6, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 6), align 8412 store i64 %r7, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 7), align 8413 ret void414}415 416define void @llrint_v8f64_v8i64() {417; SCALAR-LABEL: @llrint_v8f64_v8i64(418; SCALAR-NEXT: [[A0:%.*]] = load double, ptr @f64, align 8419; SCALAR-NEXT: [[A1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 1), align 8420; SCALAR-NEXT: [[A2:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 2), align 8421; SCALAR-NEXT: [[A3:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 3), align 8422; SCALAR-NEXT: [[A4:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8423; SCALAR-NEXT: [[A5:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 5), align 8424; SCALAR-NEXT: [[A6:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 6), align 8425; SCALAR-NEXT: [[A7:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 7), align 8426; SCALAR-NEXT: [[R0:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A0]])427; SCALAR-NEXT: [[R1:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A1]])428; SCALAR-NEXT: [[R2:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A2]])429; SCALAR-NEXT: [[R3:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A3]])430; SCALAR-NEXT: [[R4:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A4]])431; SCALAR-NEXT: [[R5:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A5]])432; SCALAR-NEXT: [[R6:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A6]])433; SCALAR-NEXT: [[R7:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A7]])434; SCALAR-NEXT: store i64 [[R0]], ptr @r64, align 8435; SCALAR-NEXT: store i64 [[R1]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 1), align 8436; SCALAR-NEXT: store i64 [[R2]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 2), align 8437; SCALAR-NEXT: store i64 [[R3]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 3), align 8438; SCALAR-NEXT: store i64 [[R4]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 4), align 8439; SCALAR-NEXT: store i64 [[R5]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 5), align 8440; SCALAR-NEXT: store i64 [[R6]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 6), align 8441; SCALAR-NEXT: store i64 [[R7]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 7), align 8442; SCALAR-NEXT: ret void443;444; VEC128-LABEL: @llrint_v8f64_v8i64(445; VEC128-NEXT: [[A0:%.*]] = load double, ptr @f64, align 8446; VEC128-NEXT: [[A1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 1), align 8447; VEC128-NEXT: [[A2:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 2), align 8448; VEC128-NEXT: [[A3:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 3), align 8449; VEC128-NEXT: [[A4:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8450; VEC128-NEXT: [[A5:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 5), align 8451; VEC128-NEXT: [[A6:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 6), align 8452; VEC128-NEXT: [[A7:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 7), align 8453; VEC128-NEXT: [[R0:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A0]])454; VEC128-NEXT: [[R1:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A1]])455; VEC128-NEXT: [[R2:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A2]])456; VEC128-NEXT: [[R3:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A3]])457; VEC128-NEXT: [[R4:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A4]])458; VEC128-NEXT: [[R5:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A5]])459; VEC128-NEXT: [[R6:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A6]])460; VEC128-NEXT: [[R7:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A7]])461; VEC128-NEXT: store i64 [[R0]], ptr @r64, align 8462; VEC128-NEXT: store i64 [[R1]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 1), align 8463; VEC128-NEXT: store i64 [[R2]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 2), align 8464; VEC128-NEXT: store i64 [[R3]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 3), align 8465; VEC128-NEXT: store i64 [[R4]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 4), align 8466; VEC128-NEXT: store i64 [[R5]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 5), align 8467; VEC128-NEXT: store i64 [[R6]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 6), align 8468; VEC128-NEXT: store i64 [[R7]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 7), align 8469; VEC128-NEXT: ret void470;471; VEC256-AVX2-LABEL: @llrint_v8f64_v8i64(472; VEC256-AVX2-NEXT: [[A0:%.*]] = load double, ptr @f64, align 8473; VEC256-AVX2-NEXT: [[A1:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 1), align 8474; VEC256-AVX2-NEXT: [[A2:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 2), align 8475; VEC256-AVX2-NEXT: [[A3:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 3), align 8476; VEC256-AVX2-NEXT: [[A4:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8477; VEC256-AVX2-NEXT: [[A5:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 5), align 8478; VEC256-AVX2-NEXT: [[A6:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 6), align 8479; VEC256-AVX2-NEXT: [[A7:%.*]] = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 7), align 8480; VEC256-AVX2-NEXT: [[R0:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A0]])481; VEC256-AVX2-NEXT: [[R1:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A1]])482; VEC256-AVX2-NEXT: [[R2:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A2]])483; VEC256-AVX2-NEXT: [[R3:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A3]])484; VEC256-AVX2-NEXT: [[R4:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A4]])485; VEC256-AVX2-NEXT: [[R5:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A5]])486; VEC256-AVX2-NEXT: [[R6:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A6]])487; VEC256-AVX2-NEXT: [[R7:%.*]] = call i64 @llvm.llrint.i64.f64(double [[A7]])488; VEC256-AVX2-NEXT: store i64 [[R0]], ptr @r64, align 8489; VEC256-AVX2-NEXT: store i64 [[R1]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 1), align 8490; VEC256-AVX2-NEXT: store i64 [[R2]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 2), align 8491; VEC256-AVX2-NEXT: store i64 [[R3]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 3), align 8492; VEC256-AVX2-NEXT: store i64 [[R4]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 4), align 8493; VEC256-AVX2-NEXT: store i64 [[R5]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 5), align 8494; VEC256-AVX2-NEXT: store i64 [[R6]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 6), align 8495; VEC256-AVX2-NEXT: store i64 [[R7]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 7), align 8496; VEC256-AVX2-NEXT: ret void497;498; VEC256-AVX512-LABEL: @llrint_v8f64_v8i64(499; VEC256-AVX512-NEXT: [[TMP1:%.*]] = load <4 x double>, ptr @f64, align 8500; VEC256-AVX512-NEXT: [[TMP2:%.*]] = call <4 x i64> @llvm.llrint.v4i64.v4f64(<4 x double> [[TMP1]])501; VEC256-AVX512-NEXT: store <4 x i64> [[TMP2]], ptr @r64, align 8502; VEC256-AVX512-NEXT: [[TMP3:%.*]] = load <4 x double>, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8503; VEC256-AVX512-NEXT: [[TMP4:%.*]] = call <4 x i64> @llvm.llrint.v4i64.v4f64(<4 x double> [[TMP3]])504; VEC256-AVX512-NEXT: store <4 x i64> [[TMP4]], ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 4), align 8505; VEC256-AVX512-NEXT: ret void506;507; VEC512-LABEL: @llrint_v8f64_v8i64(508; VEC512-NEXT: [[TMP1:%.*]] = load <8 x double>, ptr @f64, align 8509; VEC512-NEXT: [[TMP2:%.*]] = call <8 x i64> @llvm.llrint.v8i64.v8f64(<8 x double> [[TMP1]])510; VEC512-NEXT: store <8 x i64> [[TMP2]], ptr @r64, align 8511; VEC512-NEXT: ret void512;513 %a0 = load double, ptr @f64, align 8514 %a1 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 1), align 8515 %a2 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 2), align 8516 %a3 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 3), align 8517 %a4 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 4), align 8518 %a5 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 5), align 8519 %a6 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 6), align 8520 %a7 = load double, ptr getelementptr inbounds ([8 x double], ptr @f64, i32 0, i32 7), align 8521 %r0 = call i64 @llvm.llrint.i64.f64(double %a0)522 %r1 = call i64 @llvm.llrint.i64.f64(double %a1)523 %r2 = call i64 @llvm.llrint.i64.f64(double %a2)524 %r3 = call i64 @llvm.llrint.i64.f64(double %a3)525 %r4 = call i64 @llvm.llrint.i64.f64(double %a4)526 %r5 = call i64 @llvm.llrint.i64.f64(double %a5)527 %r6 = call i64 @llvm.llrint.i64.f64(double %a6)528 %r7 = call i64 @llvm.llrint.i64.f64(double %a7)529 store i64 %r0, ptr @r64, align 8530 store i64 %r1, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 1), align 8531 store i64 %r2, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 2), align 8532 store i64 %r3, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 3), align 8533 store i64 %r4, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 4), align 8534 store i64 %r5, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 5), align 8535 store i64 %r6, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 6), align 8536 store i64 %r7, ptr getelementptr inbounds ([8 x i64], ptr @r64, i32 0, i32 7), align 8537 ret void538}539;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:540; CHECK: {{.*}}541