266 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mattr=+simd128 | FileCheck %s3 4;; Test that SIMD extending operations can be successfully selected5 6target triple = "wasm32-unknown-unknown"7 8define <8 x i16> @extend_low_i8x16_s(<16 x i8> %v) {9; CHECK-LABEL: extend_low_i8x16_s:10; CHECK: .functype extend_low_i8x16_s (v128) -> (v128)11; CHECK-NEXT: # %bb.0:12; CHECK-NEXT: local.get 013; CHECK-NEXT: i16x8.extend_low_i8x16_s14; CHECK-NEXT: # fallthrough-return15 %low = shufflevector <16 x i8> %v, <16 x i8> undef,16 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>17 %extended = sext <8 x i8> %low to <8 x i16>18 ret <8 x i16> %extended19}20 21define <8 x i16> @extend_low_i8x16_u(<16 x i8> %v) {22; CHECK-LABEL: extend_low_i8x16_u:23; CHECK: .functype extend_low_i8x16_u (v128) -> (v128)24; CHECK-NEXT: # %bb.0:25; CHECK-NEXT: local.get 026; CHECK-NEXT: i16x8.extend_low_i8x16_u27; CHECK-NEXT: # fallthrough-return28 %low = shufflevector <16 x i8> %v, <16 x i8> undef,29 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>30 %extended = zext <8 x i8> %low to <8 x i16>31 ret <8 x i16> %extended32}33 34define <8 x i16> @extend_high_i8x16_s(<16 x i8> %v) {35; CHECK-LABEL: extend_high_i8x16_s:36; CHECK: .functype extend_high_i8x16_s (v128) -> (v128)37; CHECK-NEXT: # %bb.0:38; CHECK-NEXT: local.get 039; CHECK-NEXT: i16x8.extend_high_i8x16_s40; CHECK-NEXT: # fallthrough-return41 %low = shufflevector <16 x i8> %v, <16 x i8> undef,42 <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>43 %extended = sext <8 x i8> %low to <8 x i16>44 ret <8 x i16> %extended45}46 47define <8 x i16> @extend_high_i8x16_u(<16 x i8> %v) {48; CHECK-LABEL: extend_high_i8x16_u:49; CHECK: .functype extend_high_i8x16_u (v128) -> (v128)50; CHECK-NEXT: # %bb.0:51; CHECK-NEXT: local.get 052; CHECK-NEXT: i16x8.extend_high_i8x16_u53; CHECK-NEXT: # fallthrough-return54 %low = shufflevector <16 x i8> %v, <16 x i8> undef,55 <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>56 %extended = zext <8 x i8> %low to <8 x i16>57 ret <8 x i16> %extended58}59 60define <4 x i32> @extend_low_i16x8_s(<8 x i16> %v) {61; CHECK-LABEL: extend_low_i16x8_s:62; CHECK: .functype extend_low_i16x8_s (v128) -> (v128)63; CHECK-NEXT: # %bb.0:64; CHECK-NEXT: local.get 065; CHECK-NEXT: i32x4.extend_low_i16x8_s66; CHECK-NEXT: # fallthrough-return67 %low = shufflevector <8 x i16> %v, <8 x i16> undef,68 <4 x i32> <i32 0, i32 1, i32 2, i32 3>69 %extended = sext <4 x i16> %low to <4 x i32>70 ret <4 x i32> %extended71}72 73define <4 x i32> @extend_low_i16x8_u(<8 x i16> %v) {74; CHECK-LABEL: extend_low_i16x8_u:75; CHECK: .functype extend_low_i16x8_u (v128) -> (v128)76; CHECK-NEXT: # %bb.0:77; CHECK-NEXT: local.get 078; CHECK-NEXT: i32x4.extend_low_i16x8_u79; CHECK-NEXT: # fallthrough-return80 %low = shufflevector <8 x i16> %v, <8 x i16> undef,81 <4 x i32> <i32 0, i32 1, i32 2, i32 3>82 %extended = zext <4 x i16> %low to <4 x i32>83 ret <4 x i32> %extended84}85 86define <4 x i32> @extend_high_i16x8_s(<8 x i16> %v) {87; CHECK-LABEL: extend_high_i16x8_s:88; CHECK: .functype extend_high_i16x8_s (v128) -> (v128)89; CHECK-NEXT: # %bb.0:90; CHECK-NEXT: local.get 091; CHECK-NEXT: i32x4.extend_high_i16x8_s92; CHECK-NEXT: # fallthrough-return93 %low = shufflevector <8 x i16> %v, <8 x i16> undef,94 <4 x i32> <i32 4, i32 5, i32 6, i32 7>95 %extended = sext <4 x i16> %low to <4 x i32>96 ret <4 x i32> %extended97}98 99define <4 x i32> @extend_high_i16x8_u(<8 x i16> %v) {100; CHECK-LABEL: extend_high_i16x8_u:101; CHECK: .functype extend_high_i16x8_u (v128) -> (v128)102; CHECK-NEXT: # %bb.0:103; CHECK-NEXT: local.get 0104; CHECK-NEXT: i32x4.extend_high_i16x8_u105; CHECK-NEXT: # fallthrough-return106 %low = shufflevector <8 x i16> %v, <8 x i16> undef,107 <4 x i32> <i32 4, i32 5, i32 6, i32 7>108 %extended = zext <4 x i16> %low to <4 x i32>109 ret <4 x i32> %extended110}111 112define <2 x i64> @extend_low_i32x4_s(<4 x i32> %v) {113; CHECK-LABEL: extend_low_i32x4_s:114; CHECK: .functype extend_low_i32x4_s (v128) -> (v128)115; CHECK-NEXT: # %bb.0:116; CHECK-NEXT: local.get 0117; CHECK-NEXT: i64x2.extend_low_i32x4_s118; CHECK-NEXT: # fallthrough-return119 %low = shufflevector <4 x i32> %v, <4 x i32> undef,120 <2 x i32> <i32 0, i32 1>121 %extended = sext <2 x i32> %low to <2 x i64>122 ret <2 x i64> %extended123}124 125define <2 x i64> @extend_low_i32x4_u(<4 x i32> %v) {126; CHECK-LABEL: extend_low_i32x4_u:127; CHECK: .functype extend_low_i32x4_u (v128) -> (v128)128; CHECK-NEXT: # %bb.0:129; CHECK-NEXT: local.get 0130; CHECK-NEXT: i64x2.extend_low_i32x4_u131; CHECK-NEXT: # fallthrough-return132 %low = shufflevector <4 x i32> %v, <4 x i32> undef,133 <2 x i32> <i32 0, i32 1>134 %extended = zext <2 x i32> %low to <2 x i64>135 ret <2 x i64> %extended136}137 138define <2 x i64> @extend_high_i32x4_s(<4 x i32> %v) {139; CHECK-LABEL: extend_high_i32x4_s:140; CHECK: .functype extend_high_i32x4_s (v128) -> (v128)141; CHECK-NEXT: # %bb.0:142; CHECK-NEXT: local.get 0143; CHECK-NEXT: i64x2.extend_high_i32x4_s144; CHECK-NEXT: # fallthrough-return145 %low = shufflevector <4 x i32> %v, <4 x i32> undef,146 <2 x i32> <i32 2, i32 3>147 %extended = sext <2 x i32> %low to <2 x i64>148 ret <2 x i64> %extended149}150 151define <2 x i64> @extend_high_i32x4_u(<4 x i32> %v) {152; CHECK-LABEL: extend_high_i32x4_u:153; CHECK: .functype extend_high_i32x4_u (v128) -> (v128)154; CHECK-NEXT: # %bb.0:155; CHECK-NEXT: local.get 0156; CHECK-NEXT: i64x2.extend_high_i32x4_u157; CHECK-NEXT: # fallthrough-return158 %low = shufflevector <4 x i32> %v, <4 x i32> undef,159 <2 x i32> <i32 2, i32 3>160 %extended = zext <2 x i32> %low to <2 x i64>161 ret <2 x i64> %extended162}163 164;; Also test that similar patterns with offsets not corresponding to165;; the low or high half are correctly expanded.166 167define <8 x i16> @extend_lowish_i8x16_s(<16 x i8> %v) {168; CHECK-LABEL: extend_lowish_i8x16_s:169; CHECK: .functype extend_lowish_i8x16_s (v128) -> (v128)170; CHECK-NEXT: # %bb.0:171; CHECK-NEXT: local.get 0172; CHECK-NEXT: local.get 0173; CHECK-NEXT: i8x16.shuffle 1, 2, 3, 4, 5, 6, 7, 8, 0, 0, 0, 0, 0, 0, 0, 0174; CHECK-NEXT: i16x8.extend_low_i8x16_s175; CHECK-NEXT: # fallthrough-return176 %lowish = shufflevector <16 x i8> %v, <16 x i8> undef,177 <8 x i32> <i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>178 %extended = sext <8 x i8> %lowish to <8 x i16>179 ret <8 x i16> %extended180}181 182define <4 x i32> @extend_lowish_i16x8_s(<8 x i16> %v) {183; CHECK-LABEL: extend_lowish_i16x8_s:184; CHECK: .functype extend_lowish_i16x8_s (v128) -> (v128)185; CHECK-NEXT: # %bb.0:186; CHECK-NEXT: local.get 0187; CHECK-NEXT: local.get 0188; CHECK-NEXT: i8x16.shuffle 2, 3, 4, 5, 6, 7, 8, 9, 0, 1, 0, 1, 0, 1, 0, 1189; CHECK-NEXT: i32x4.extend_low_i16x8_s190; CHECK-NEXT: # fallthrough-return191 %lowish = shufflevector <8 x i16> %v, <8 x i16> undef,192 <4 x i32> <i32 1, i32 2, i32 3, i32 4>193 %extended = sext <4 x i16> %lowish to <4 x i32>194 ret <4 x i32> %extended195}196 197;; Also test vectors that aren't full 128 bits, or might require198;; multiple extensions199 200define <16 x i8> @extend_i1x16_i8(<16 x i1> %v) {201; CHECK-LABEL: extend_i1x16_i8:202; CHECK: .functype extend_i1x16_i8 (v128) -> (v128)203; CHECK-NEXT: # %bb.0:204; CHECK-NEXT: local.get 0205; CHECK-NEXT: i32.const 7206; CHECK-NEXT: i8x16.shl207; CHECK-NEXT: i32.const 7208; CHECK-NEXT: i8x16.shr_s209; CHECK-NEXT: # fallthrough-return210 %extended = sext <16 x i1> %v to <16 x i8>211 ret <16 x i8> %extended212}213 214define <8 x i8> @extend_i1x8_i8(<8 x i1> %v) {215; CHECK-LABEL: extend_i1x8_i8:216; CHECK: .functype extend_i1x8_i8 (v128) -> (v128)217; CHECK-NEXT: # %bb.0:218; CHECK-NEXT: local.get 0219; CHECK-NEXT: local.get 0220; CHECK-NEXT: i8x16.shuffle 0, 2, 4, 6, 8, 10, 12, 14, 0, 0, 0, 0, 0, 0, 0, 0221; CHECK-NEXT: i32.const 7222; CHECK-NEXT: i8x16.shl223; CHECK-NEXT: i32.const 7224; CHECK-NEXT: i8x16.shr_s225; CHECK-NEXT: # fallthrough-return226 %extended = sext <8 x i1> %v to <8 x i8>227 ret <8 x i8> %extended228}229 230define <8 x i16> @extend_i1x8_i16(<8 x i1> %v) {231; CHECK-LABEL: extend_i1x8_i16:232; CHECK: .functype extend_i1x8_i16 (v128) -> (v128)233; CHECK-NEXT: # %bb.0:234; CHECK-NEXT: local.get 0235; CHECK-NEXT: v128.const 1, 1, 1, 1, 1, 1, 1, 1236; CHECK-NEXT: v128.and237; CHECK-NEXT: # fallthrough-return238 %extended = zext <8 x i1> %v to <8 x i16>239 ret <8 x i16> %extended240}241 242define <4 x i32> @extend_i8x4_i32(<4 x i8> %v) {243; CHECK-LABEL: extend_i8x4_i32:244; CHECK: .functype extend_i8x4_i32 (v128) -> (v128)245; CHECK-NEXT: # %bb.0:246; CHECK-NEXT: local.get 0247; CHECK-NEXT: i16x8.extend_low_i8x16_u248; CHECK-NEXT: i32x4.extend_low_i16x8_u249; CHECK-NEXT: # fallthrough-return250 %extended = zext <4 x i8> %v to <4 x i32>251 ret <4 x i32> %extended252}253 254define <2 x i64> @extend_i8x2_i64(<2 x i8> %v) {255; CHECK-LABEL: extend_i8x2_i64:256; CHECK: .functype extend_i8x2_i64 (v128) -> (v128)257; CHECK-NEXT: # %bb.0:258; CHECK-NEXT: local.get 0259; CHECK-NEXT: i16x8.extend_low_i8x16_s260; CHECK-NEXT: i32x4.extend_low_i16x8_s261; CHECK-NEXT: i64x2.extend_low_i32x4_s262; CHECK-NEXT: # fallthrough-return263 %extended = sext <2 x i8> %v to <2 x i64>264 ret <2 x i64> %extended265}266