148 lines · plain
1; RUN: llc -mtriple=arm-eabi -mattr=+neon %s -o - | FileCheck %s2 3define <8 x i8> @vpmins8(ptr %A, ptr %B) nounwind {4;CHECK-LABEL: vpmins8:5;CHECK: vpmin.s86 %tmp1 = load <8 x i8>, ptr %A7 %tmp2 = load <8 x i8>, ptr %B8 %tmp3 = call <8 x i8> @llvm.arm.neon.vpmins.v8i8(<8 x i8> %tmp1, <8 x i8> %tmp2)9 ret <8 x i8> %tmp310}11 12define <4 x i16> @vpmins16(ptr %A, ptr %B) nounwind {13;CHECK-LABEL: vpmins16:14;CHECK: vpmin.s1615 %tmp1 = load <4 x i16>, ptr %A16 %tmp2 = load <4 x i16>, ptr %B17 %tmp3 = call <4 x i16> @llvm.arm.neon.vpmins.v4i16(<4 x i16> %tmp1, <4 x i16> %tmp2)18 ret <4 x i16> %tmp319}20 21define <2 x i32> @vpmins32(ptr %A, ptr %B) nounwind {22;CHECK-LABEL: vpmins32:23;CHECK: vpmin.s3224 %tmp1 = load <2 x i32>, ptr %A25 %tmp2 = load <2 x i32>, ptr %B26 %tmp3 = call <2 x i32> @llvm.arm.neon.vpmins.v2i32(<2 x i32> %tmp1, <2 x i32> %tmp2)27 ret <2 x i32> %tmp328}29 30define <8 x i8> @vpminu8(ptr %A, ptr %B) nounwind {31;CHECK-LABEL: vpminu8:32;CHECK: vpmin.u833 %tmp1 = load <8 x i8>, ptr %A34 %tmp2 = load <8 x i8>, ptr %B35 %tmp3 = call <8 x i8> @llvm.arm.neon.vpminu.v8i8(<8 x i8> %tmp1, <8 x i8> %tmp2)36 ret <8 x i8> %tmp337}38 39define <4 x i16> @vpminu16(ptr %A, ptr %B) nounwind {40;CHECK-LABEL: vpminu16:41;CHECK: vpmin.u1642 %tmp1 = load <4 x i16>, ptr %A43 %tmp2 = load <4 x i16>, ptr %B44 %tmp3 = call <4 x i16> @llvm.arm.neon.vpminu.v4i16(<4 x i16> %tmp1, <4 x i16> %tmp2)45 ret <4 x i16> %tmp346}47 48define <2 x i32> @vpminu32(ptr %A, ptr %B) nounwind {49;CHECK-LABEL: vpminu32:50;CHECK: vpmin.u3251 %tmp1 = load <2 x i32>, ptr %A52 %tmp2 = load <2 x i32>, ptr %B53 %tmp3 = call <2 x i32> @llvm.arm.neon.vpminu.v2i32(<2 x i32> %tmp1, <2 x i32> %tmp2)54 ret <2 x i32> %tmp355}56 57define <2 x float> @vpminf32(ptr %A, ptr %B) nounwind {58;CHECK-LABEL: vpminf32:59;CHECK: vpmin.f3260 %tmp1 = load <2 x float>, ptr %A61 %tmp2 = load <2 x float>, ptr %B62 %tmp3 = call <2 x float> @llvm.arm.neon.vpmins.v2f32(<2 x float> %tmp1, <2 x float> %tmp2)63 ret <2 x float> %tmp364}65 66declare <8 x i8> @llvm.arm.neon.vpmins.v8i8(<8 x i8>, <8 x i8>) nounwind readnone67declare <4 x i16> @llvm.arm.neon.vpmins.v4i16(<4 x i16>, <4 x i16>) nounwind readnone68declare <2 x i32> @llvm.arm.neon.vpmins.v2i32(<2 x i32>, <2 x i32>) nounwind readnone69 70declare <8 x i8> @llvm.arm.neon.vpminu.v8i8(<8 x i8>, <8 x i8>) nounwind readnone71declare <4 x i16> @llvm.arm.neon.vpminu.v4i16(<4 x i16>, <4 x i16>) nounwind readnone72declare <2 x i32> @llvm.arm.neon.vpminu.v2i32(<2 x i32>, <2 x i32>) nounwind readnone73 74declare <2 x float> @llvm.arm.neon.vpmins.v2f32(<2 x float>, <2 x float>) nounwind readnone75 76define <8 x i8> @vpmaxs8(ptr %A, ptr %B) nounwind {77;CHECK-LABEL: vpmaxs8:78;CHECK: vpmax.s879 %tmp1 = load <8 x i8>, ptr %A80 %tmp2 = load <8 x i8>, ptr %B81 %tmp3 = call <8 x i8> @llvm.arm.neon.vpmaxs.v8i8(<8 x i8> %tmp1, <8 x i8> %tmp2)82 ret <8 x i8> %tmp383}84 85define <4 x i16> @vpmaxs16(ptr %A, ptr %B) nounwind {86;CHECK-LABEL: vpmaxs16:87;CHECK: vpmax.s1688 %tmp1 = load <4 x i16>, ptr %A89 %tmp2 = load <4 x i16>, ptr %B90 %tmp3 = call <4 x i16> @llvm.arm.neon.vpmaxs.v4i16(<4 x i16> %tmp1, <4 x i16> %tmp2)91 ret <4 x i16> %tmp392}93 94define <2 x i32> @vpmaxs32(ptr %A, ptr %B) nounwind {95;CHECK-LABEL: vpmaxs32:96;CHECK: vpmax.s3297 %tmp1 = load <2 x i32>, ptr %A98 %tmp2 = load <2 x i32>, ptr %B99 %tmp3 = call <2 x i32> @llvm.arm.neon.vpmaxs.v2i32(<2 x i32> %tmp1, <2 x i32> %tmp2)100 ret <2 x i32> %tmp3101}102 103define <8 x i8> @vpmaxu8(ptr %A, ptr %B) nounwind {104;CHECK-LABEL: vpmaxu8:105;CHECK: vpmax.u8106 %tmp1 = load <8 x i8>, ptr %A107 %tmp2 = load <8 x i8>, ptr %B108 %tmp3 = call <8 x i8> @llvm.arm.neon.vpmaxu.v8i8(<8 x i8> %tmp1, <8 x i8> %tmp2)109 ret <8 x i8> %tmp3110}111 112define <4 x i16> @vpmaxu16(ptr %A, ptr %B) nounwind {113;CHECK-LABEL: vpmaxu16:114;CHECK: vpmax.u16115 %tmp1 = load <4 x i16>, ptr %A116 %tmp2 = load <4 x i16>, ptr %B117 %tmp3 = call <4 x i16> @llvm.arm.neon.vpmaxu.v4i16(<4 x i16> %tmp1, <4 x i16> %tmp2)118 ret <4 x i16> %tmp3119}120 121define <2 x i32> @vpmaxu32(ptr %A, ptr %B) nounwind {122;CHECK-LABEL: vpmaxu32:123;CHECK: vpmax.u32124 %tmp1 = load <2 x i32>, ptr %A125 %tmp2 = load <2 x i32>, ptr %B126 %tmp3 = call <2 x i32> @llvm.arm.neon.vpmaxu.v2i32(<2 x i32> %tmp1, <2 x i32> %tmp2)127 ret <2 x i32> %tmp3128}129 130define <2 x float> @vpmaxf32(ptr %A, ptr %B) nounwind {131;CHECK-LABEL: vpmaxf32:132;CHECK: vpmax.f32133 %tmp1 = load <2 x float>, ptr %A134 %tmp2 = load <2 x float>, ptr %B135 %tmp3 = call <2 x float> @llvm.arm.neon.vpmaxs.v2f32(<2 x float> %tmp1, <2 x float> %tmp2)136 ret <2 x float> %tmp3137}138 139declare <8 x i8> @llvm.arm.neon.vpmaxs.v8i8(<8 x i8>, <8 x i8>) nounwind readnone140declare <4 x i16> @llvm.arm.neon.vpmaxs.v4i16(<4 x i16>, <4 x i16>) nounwind readnone141declare <2 x i32> @llvm.arm.neon.vpmaxs.v2i32(<2 x i32>, <2 x i32>) nounwind readnone142 143declare <8 x i8> @llvm.arm.neon.vpmaxu.v8i8(<8 x i8>, <8 x i8>) nounwind readnone144declare <4 x i16> @llvm.arm.neon.vpmaxu.v4i16(<4 x i16>, <4 x i16>) nounwind readnone145declare <2 x i32> @llvm.arm.neon.vpmaxu.v2i32(<2 x i32>, <2 x i32>) nounwind readnone146 147declare <2 x float> @llvm.arm.neon.vpmaxs.v2f32(<2 x float>, <2 x float>) nounwind readnone148