243 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 22; Test vector insertion of byte-swapped memory values.3;4; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z15 | FileCheck %s5 6declare i16 @llvm.bswap.i16(i16)7declare i32 @llvm.bswap.i32(i32)8declare i64 @llvm.bswap.i64(i64)9declare <8 x i16> @llvm.bswap.v8i16(<8 x i16>)10declare <4 x i32> @llvm.bswap.v4i32(<4 x i32>)11declare <2 x i64> @llvm.bswap.v2i64(<2 x i64>)12 13; Test v8i16 insertion into the first element.14define <8 x i16> @f1(<8 x i16> %val, ptr %ptr) {15; CHECK-LABEL: f1:16; CHECK: # %bb.0:17; CHECK-NEXT: vlebrh %v24, 0(%r2), 018; CHECK-NEXT: br %r1419 %element = load i16, ptr %ptr20 %swap = call i16 @llvm.bswap.i16(i16 %element)21 %ret = insertelement <8 x i16> %val, i16 %swap, i32 022 ret <8 x i16> %ret23}24 25; Test v8i16 insertion into the last element.26define <8 x i16> @f2(<8 x i16> %val, ptr %ptr) {27; CHECK-LABEL: f2:28; CHECK: # %bb.0:29; CHECK-NEXT: vlebrh %v24, 0(%r2), 730; CHECK-NEXT: br %r1431 %element = load i16, ptr %ptr32 %swap = call i16 @llvm.bswap.i16(i16 %element)33 %ret = insertelement <8 x i16> %val, i16 %swap, i32 734 ret <8 x i16> %ret35}36 37; Test v8i16 insertion with the highest in-range offset.38define <8 x i16> @f3(<8 x i16> %val, ptr %base) {39; CHECK-LABEL: f3:40; CHECK: # %bb.0:41; CHECK-NEXT: vlebrh %v24, 4094(%r2), 542; CHECK-NEXT: br %r1443 %ptr = getelementptr i16, ptr %base, i32 204744 %element = load i16, ptr %ptr45 %swap = call i16 @llvm.bswap.i16(i16 %element)46 %ret = insertelement <8 x i16> %val, i16 %swap, i32 547 ret <8 x i16> %ret48}49 50; Test v8i16 insertion with the first ouf-of-range offset.51define <8 x i16> @f4(<8 x i16> %val, ptr %base) {52; CHECK-LABEL: f4:53; CHECK: # %bb.0:54; CHECK-NEXT: aghi %r2, 409655; CHECK-NEXT: vlebrh %v24, 0(%r2), 156; CHECK-NEXT: br %r1457 %ptr = getelementptr i16, ptr %base, i32 204858 %element = load i16, ptr %ptr59 %swap = call i16 @llvm.bswap.i16(i16 %element)60 %ret = insertelement <8 x i16> %val, i16 %swap, i32 161 ret <8 x i16> %ret62}63 64; Test v8i16 insertion into a variable element.65define <8 x i16> @f5(<8 x i16> %val, ptr %ptr, i32 %index) {66; CHECK-LABEL: f5:67; CHECK: # %bb.0:68; CHECK-NEXT: lrvh %r0, 0(%r2)69; CHECK-NEXT: vlvgh %v24, %r0, 0(%r3)70; CHECK-NEXT: br %r1471 %element = load i16, ptr %ptr72 %swap = call i16 @llvm.bswap.i16(i16 %element)73 %ret = insertelement <8 x i16> %val, i16 %swap, i32 %index74 ret <8 x i16> %ret75}76 77; Test v8i16 insertion using a pair of vector bswaps.78define <8 x i16> @f6(<8 x i16> %val, ptr %ptr) {79; CHECK-LABEL: f6:80; CHECK: # %bb.0:81; CHECK-NEXT: vlebrh %v24, 0(%r2), 082; CHECK-NEXT: br %r1483 %element = load i16, ptr %ptr84 %swapval = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> %val)85 %insert = insertelement <8 x i16> %swapval, i16 %element, i32 086 %ret = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> %insert)87 ret <8 x i16> %ret88}89 90; Test v4i32 insertion into the first element.91define <4 x i32> @f7(<4 x i32> %val, ptr %ptr) {92; CHECK-LABEL: f7:93; CHECK: # %bb.0:94; CHECK-NEXT: vlebrf %v24, 0(%r2), 095; CHECK-NEXT: br %r1496 %element = load i32, ptr %ptr97 %swap = call i32 @llvm.bswap.i32(i32 %element)98 %ret = insertelement <4 x i32> %val, i32 %swap, i32 099 ret <4 x i32> %ret100}101 102; Test v4i32 insertion into the last element.103define <4 x i32> @f8(<4 x i32> %val, ptr %ptr) {104; CHECK-LABEL: f8:105; CHECK: # %bb.0:106; CHECK-NEXT: vlebrf %v24, 0(%r2), 3107; CHECK-NEXT: br %r14108 %element = load i32, ptr %ptr109 %swap = call i32 @llvm.bswap.i32(i32 %element)110 %ret = insertelement <4 x i32> %val, i32 %swap, i32 3111 ret <4 x i32> %ret112}113 114; Test v4i32 insertion with the highest in-range offset.115define <4 x i32> @f9(<4 x i32> %val, ptr %base) {116; CHECK-LABEL: f9:117; CHECK: # %bb.0:118; CHECK-NEXT: vlebrf %v24, 4092(%r2), 2119; CHECK-NEXT: br %r14120 %ptr = getelementptr i32, ptr %base, i32 1023121 %element = load i32, ptr %ptr122 %swap = call i32 @llvm.bswap.i32(i32 %element)123 %ret = insertelement <4 x i32> %val, i32 %swap, i32 2124 ret <4 x i32> %ret125}126 127; Test v4i32 insertion with the first ouf-of-range offset.128define <4 x i32> @f10(<4 x i32> %val, ptr %base) {129; CHECK-LABEL: f10:130; CHECK: # %bb.0:131; CHECK-NEXT: aghi %r2, 4096132; CHECK-NEXT: vlebrf %v24, 0(%r2), 1133; CHECK-NEXT: br %r14134 %ptr = getelementptr i32, ptr %base, i32 1024135 %element = load i32, ptr %ptr136 %swap = call i32 @llvm.bswap.i32(i32 %element)137 %ret = insertelement <4 x i32> %val, i32 %swap, i32 1138 ret <4 x i32> %ret139}140 141; Test v4i32 insertion into a variable element.142define <4 x i32> @f11(<4 x i32> %val, ptr %ptr, i32 %index) {143; CHECK-LABEL: f11:144; CHECK: # %bb.0:145; CHECK-NEXT: lrv %r0, 0(%r2)146; CHECK-NEXT: vlvgf %v24, %r0, 0(%r3)147; CHECK-NEXT: br %r14148 %element = load i32, ptr %ptr149 %swap = call i32 @llvm.bswap.i32(i32 %element)150 %ret = insertelement <4 x i32> %val, i32 %swap, i32 %index151 ret <4 x i32> %ret152}153 154; Test v4i32 insertion using a pair of vector bswaps.155define <4 x i32> @f12(<4 x i32> %val, ptr %ptr) {156; CHECK-LABEL: f12:157; CHECK: # %bb.0:158; CHECK-NEXT: vlebrf %v24, 0(%r2), 0159; CHECK-NEXT: br %r14160 %element = load i32, ptr %ptr161 %swapval = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> %val)162 %insert = insertelement <4 x i32> %swapval, i32 %element, i32 0163 %ret = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> %insert)164 ret <4 x i32> %ret165}166 167; Test v2i64 insertion into the first element.168define <2 x i64> @f13(<2 x i64> %val, ptr %ptr) {169; CHECK-LABEL: f13:170; CHECK: # %bb.0:171; CHECK-NEXT: vlebrg %v24, 0(%r2), 0172; CHECK-NEXT: br %r14173 %element = load i64, ptr %ptr174 %swap = call i64 @llvm.bswap.i64(i64 %element)175 %ret = insertelement <2 x i64> %val, i64 %swap, i32 0176 ret <2 x i64> %ret177}178 179; Test v2i64 insertion into the last element.180define <2 x i64> @f14(<2 x i64> %val, ptr %ptr) {181; CHECK-LABEL: f14:182; CHECK: # %bb.0:183; CHECK-NEXT: vlebrg %v24, 0(%r2), 1184; CHECK-NEXT: br %r14185 %element = load i64, ptr %ptr186 %swap = call i64 @llvm.bswap.i64(i64 %element)187 %ret = insertelement <2 x i64> %val, i64 %swap, i32 1188 ret <2 x i64> %ret189}190 191; Test v2i64 insertion with the highest in-range offset.192define <2 x i64> @f15(<2 x i64> %val, ptr %base) {193; CHECK-LABEL: f15:194; CHECK: # %bb.0:195; CHECK-NEXT: vlebrg %v24, 4088(%r2), 1196; CHECK-NEXT: br %r14197 %ptr = getelementptr i64, ptr %base, i32 511198 %element = load i64, ptr %ptr199 %swap = call i64 @llvm.bswap.i64(i64 %element)200 %ret = insertelement <2 x i64> %val, i64 %swap, i32 1201 ret <2 x i64> %ret202}203 204; Test v2i64 insertion with the first ouf-of-range offset.205define <2 x i64> @f16(<2 x i64> %val, ptr %base) {206; CHECK-LABEL: f16:207; CHECK: # %bb.0:208; CHECK-NEXT: aghi %r2, 4096209; CHECK-NEXT: vlebrg %v24, 0(%r2), 0210; CHECK-NEXT: br %r14211 %ptr = getelementptr i64, ptr %base, i32 512212 %element = load i64, ptr %ptr213 %swap = call i64 @llvm.bswap.i64(i64 %element)214 %ret = insertelement <2 x i64> %val, i64 %swap, i32 0215 ret <2 x i64> %ret216}217 218; Test v2i64 insertion into a variable element.219define <2 x i64> @f17(<2 x i64> %val, ptr %ptr, i32 %index) {220; CHECK-LABEL: f17:221; CHECK: # %bb.0:222; CHECK-NEXT: lrvg %r0, 0(%r2)223; CHECK-NEXT: vlvgg %v24, %r0, 0(%r3)224; CHECK-NEXT: br %r14225 %element = load i64, ptr %ptr226 %swap = call i64 @llvm.bswap.i64(i64 %element)227 %ret = insertelement <2 x i64> %val, i64 %swap, i32 %index228 ret <2 x i64> %ret229}230 231; Test v2i64 insertion using a pair of vector bswaps.232define <2 x i64> @f18(<2 x i64> %val, ptr %ptr) {233; CHECK-LABEL: f18:234; CHECK: # %bb.0:235; CHECK-NEXT: vlebrg %v24, 0(%r2), 0236; CHECK-NEXT: br %r14237 %element = load i64, ptr %ptr238 %swapval = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> %val)239 %insert = insertelement <2 x i64> %swapval, i64 %element, i32 0240 %ret = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> %insert)241 ret <2 x i64> %ret242}243