277 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 22; Test vector extraction of byte-swapped value to memory.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 extraction from the first element.14define void @f1(<8 x i16> %val, ptr %ptr) {15; CHECK-LABEL: f1:16; CHECK: # %bb.0:17; CHECK-NEXT: vstebrh %v24, 0(%r2), 018; CHECK-NEXT: br %r1419 %element = extractelement <8 x i16> %val, i32 020 %swap = call i16 @llvm.bswap.i16(i16 %element)21 store i16 %swap, ptr %ptr22 ret void23}24 25; Test v8i16 extraction from the last element.26define void @f2(<8 x i16> %val, ptr %ptr) {27; CHECK-LABEL: f2:28; CHECK: # %bb.0:29; CHECK-NEXT: vstebrh %v24, 0(%r2), 730; CHECK-NEXT: br %r1431 %element = extractelement <8 x i16> %val, i32 732 %swap = call i16 @llvm.bswap.i16(i16 %element)33 store i16 %swap, ptr %ptr34 ret void35}36 37; Test v8i16 extraction of an invalid element. This must compile,38; but we don't care what it does.39define void @f3(<8 x i16> %val, ptr %ptr) {40; CHECK-LABEL: f3:41; CHECK: # %bb.0:42; CHECK-NEXT: br %r1443 %element = extractelement <8 x i16> %val, i32 844 %swap = call i16 @llvm.bswap.i16(i16 %element)45 store i16 %swap, ptr %ptr46 ret void47}48 49; Test v8i16 extraction with the highest in-range offset.50define void @f4(<8 x i16> %val, ptr %base) {51; CHECK-LABEL: f4:52; CHECK: # %bb.0:53; CHECK-NEXT: vstebrh %v24, 4094(%r2), 554; CHECK-NEXT: br %r1455 %ptr = getelementptr i16, ptr %base, i32 204756 %element = extractelement <8 x i16> %val, i32 557 %swap = call i16 @llvm.bswap.i16(i16 %element)58 store i16 %swap, ptr %ptr59 ret void60}61 62; Test v8i16 extraction with the first ouf-of-range offset.63define void @f5(<8 x i16> %val, ptr %base) {64; CHECK-LABEL: f5:65; CHECK: # %bb.0:66; CHECK-NEXT: aghi %r2, 409667; CHECK-NEXT: vstebrh %v24, 0(%r2), 168; CHECK-NEXT: br %r1469 %ptr = getelementptr i16, ptr %base, i32 204870 %element = extractelement <8 x i16> %val, i32 171 %swap = call i16 @llvm.bswap.i16(i16 %element)72 store i16 %swap, ptr %ptr73 ret void74}75 76; Test v8i16 extraction from a variable element.77define void @f6(<8 x i16> %val, ptr %ptr, i32 %index) {78; CHECK-LABEL: f6:79; CHECK: # %bb.0:80; CHECK-NEXT: vlgvh %r0, %v24, 0(%r3)81; CHECK-NEXT: strvh %r0, 0(%r2)82; CHECK-NEXT: br %r1483 %element = extractelement <8 x i16> %val, i32 %index84 %swap = call i16 @llvm.bswap.i16(i16 %element)85 store i16 %swap, ptr %ptr86 ret void87}88 89; Test v8i16 extraction using a vector bswap.90define void @f7(<8 x i16> %val, ptr %ptr) {91; CHECK-LABEL: f7:92; CHECK: # %bb.0:93; CHECK-NEXT: vstebrh %v24, 0(%r2), 094; CHECK-NEXT: br %r1495 %swap = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> %val)96 %element = extractelement <8 x i16> %swap, i32 097 store i16 %element, ptr %ptr98 ret void99}100 101; Test v4i32 extraction from the first element.102define void @f8(<4 x i32> %val, ptr %ptr) {103; CHECK-LABEL: f8:104; CHECK: # %bb.0:105; CHECK-NEXT: vstebrf %v24, 0(%r2), 0106; CHECK-NEXT: br %r14107 %element = extractelement <4 x i32> %val, i32 0108 %swap = call i32 @llvm.bswap.i32(i32 %element)109 store i32 %swap, ptr %ptr110 ret void111}112 113; Test v4i32 extraction from the last element.114define void @f9(<4 x i32> %val, ptr %ptr) {115; CHECK-LABEL: f9:116; CHECK: # %bb.0:117; CHECK-NEXT: vstebrf %v24, 0(%r2), 3118; CHECK-NEXT: br %r14119 %element = extractelement <4 x i32> %val, i32 3120 %swap = call i32 @llvm.bswap.i32(i32 %element)121 store i32 %swap, ptr %ptr122 ret void123}124 125; Test v4i32 extraction of an invalid element. This must compile,126; but we don't care what it does.127define void @f10(<4 x i32> %val, ptr %ptr) {128; CHECK-LABEL: f10:129; CHECK: # %bb.0:130; CHECK-NEXT: br %r14131 %element = extractelement <4 x i32> %val, i32 4132 %swap = call i32 @llvm.bswap.i32(i32 %element)133 store i32 %swap, ptr %ptr134 ret void135}136 137; Test v4i32 extraction with the highest in-range offset.138define void @f11(<4 x i32> %val, ptr %base) {139; CHECK-LABEL: f11:140; CHECK: # %bb.0:141; CHECK-NEXT: vstebrf %v24, 4092(%r2), 2142; CHECK-NEXT: br %r14143 %ptr = getelementptr i32, ptr %base, i32 1023144 %element = extractelement <4 x i32> %val, i32 2145 %swap = call i32 @llvm.bswap.i32(i32 %element)146 store i32 %swap, ptr %ptr147 ret void148}149 150; Test v4i32 extraction with the first ouf-of-range offset.151define void @f12(<4 x i32> %val, ptr %base) {152; CHECK-LABEL: f12:153; CHECK: # %bb.0:154; CHECK-NEXT: aghi %r2, 4096155; CHECK-NEXT: vstebrf %v24, 0(%r2), 1156; CHECK-NEXT: br %r14157 %ptr = getelementptr i32, ptr %base, i32 1024158 %element = extractelement <4 x i32> %val, i32 1159 %swap = call i32 @llvm.bswap.i32(i32 %element)160 store i32 %swap, ptr %ptr161 ret void162}163 164; Test v4i32 extraction from a variable element.165define void @f13(<4 x i32> %val, ptr %ptr, i32 %index) {166; CHECK-LABEL: f13:167; CHECK: # %bb.0:168; CHECK-NEXT: vlgvf %r0, %v24, 0(%r3)169; CHECK-NEXT: strv %r0, 0(%r2)170; CHECK-NEXT: br %r14171 %element = extractelement <4 x i32> %val, i32 %index172 %swap = call i32 @llvm.bswap.i32(i32 %element)173 store i32 %swap, ptr %ptr174 ret void175}176 177; Test v4i32 extraction using a vector bswap.178define void @f14(<4 x i32> %val, ptr %ptr) {179; CHECK-LABEL: f14:180; CHECK: # %bb.0:181; CHECK-NEXT: vstebrf %v24, 0(%r2), 0182; CHECK-NEXT: br %r14183 %swap = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> %val)184 %element = extractelement <4 x i32> %swap, i32 0185 store i32 %element, ptr %ptr186 ret void187}188 189; Test v2i64 extraction from the first element.190define void @f15(<2 x i64> %val, ptr %ptr) {191; CHECK-LABEL: f15:192; CHECK: # %bb.0:193; CHECK-NEXT: vstebrg %v24, 0(%r2), 0194; CHECK-NEXT: br %r14195 %element = extractelement <2 x i64> %val, i32 0196 %swap = call i64 @llvm.bswap.i64(i64 %element)197 store i64 %swap, ptr %ptr198 ret void199}200 201; Test v2i64 extraction from the last element.202define void @f16(<2 x i64> %val, ptr %ptr) {203; CHECK-LABEL: f16:204; CHECK: # %bb.0:205; CHECK-NEXT: vstebrg %v24, 0(%r2), 1206; CHECK-NEXT: br %r14207 %element = extractelement <2 x i64> %val, i32 1208 %swap = call i64 @llvm.bswap.i64(i64 %element)209 store i64 %swap, ptr %ptr210 ret void211}212 213; Test v2i64 extraction of an invalid element. This must compile,214; but we don't care what it does.215define void @f17(<2 x i64> %val, ptr %ptr) {216; CHECK-LABEL: f17:217; CHECK: # %bb.0:218; CHECK-NEXT: br %r14219 %element = extractelement <2 x i64> %val, i32 2220 %swap = call i64 @llvm.bswap.i64(i64 %element)221 store i64 %swap, ptr %ptr222 ret void223}224 225; Test v2i64 extraction with the highest in-range offset.226define void @f18(<2 x i64> %val, ptr %base) {227; CHECK-LABEL: f18:228; CHECK: # %bb.0:229; CHECK-NEXT: vstebrg %v24, 4088(%r2), 1230; CHECK-NEXT: br %r14231 %ptr = getelementptr i64, ptr %base, i32 511232 %element = extractelement <2 x i64> %val, i32 1233 %swap = call i64 @llvm.bswap.i64(i64 %element)234 store i64 %swap, ptr %ptr235 ret void236}237 238; Test v2i64 extraction with the first ouf-of-range offset.239define void @f19(<2 x i64> %val, ptr %base) {240; CHECK-LABEL: f19:241; CHECK: # %bb.0:242; CHECK-NEXT: aghi %r2, 4096243; CHECK-NEXT: vstebrg %v24, 0(%r2), 0244; CHECK-NEXT: br %r14245 %ptr = getelementptr i64, ptr %base, i32 512246 %element = extractelement <2 x i64> %val, i32 0247 %swap = call i64 @llvm.bswap.i64(i64 %element)248 store i64 %swap, ptr %ptr249 ret void250}251 252; Test v2i64 extraction from a variable element.253define void @f20(<2 x i64> %val, ptr %ptr, i32 %index) {254; CHECK-LABEL: f20:255; CHECK: # %bb.0:256; CHECK-NEXT: vlgvg %r0, %v24, 0(%r3)257; CHECK-NEXT: strvg %r0, 0(%r2)258; CHECK-NEXT: br %r14259 %element = extractelement <2 x i64> %val, i32 %index260 %swap = call i64 @llvm.bswap.i64(i64 %element)261 store i64 %swap, ptr %ptr262 ret void263}264 265; Test v2i64 extraction using a vector bswap.266define void @f21(<2 x i64> %val, ptr %ptr) {267; CHECK-LABEL: f21:268; CHECK: # %bb.0:269; CHECK-NEXT: vstebrg %v24, 0(%r2), 0270; CHECK-NEXT: br %r14271 %swap = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> %val)272 %element = extractelement <2 x i64> %swap, i32 0273 store i64 %element, ptr %ptr274 ret void275}276 277