clang/test/CodeGen/AArch64/neon-luti.c

   1 // NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 4
   2 // REQUIRES: aarch64-registered-target
   3 #include <arm_neon.h>
   4 // RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +neon -target-feature +lut -target-feature +bf16 -O3 -emit-llvm -o - %s | FileCheck %s
   5 // RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +neon -target-feature +lut -target-feature +bf16 -S -disable-O0-optnone -Werror -Wall -o /dev/null %s
   6
   7 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti2_lane_u8(
   8 // CHECK-SAME: <8 x i8> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
   9 // CHECK-NEXT:  entry:
  10 // CHECK-NEXT:    [[VLUTI2_LANE:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti2.lane.v16i8.v8i8(<8 x i8> [[VN]], <8 x i8> [[VM]], i32 0)
  11 // CHECK-NEXT:    ret <16 x i8> [[VLUTI2_LANE]]
  12 //
  13 uint8x16_t test_vluti2_lane_u8(uint8x8_t vn, uint8x8_t vm) {
  14   return vluti2_lane_u8(vn, vm, 0);
  15 }
  16
  17 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti2_laneq_u8(
  18 // CHECK-SAME: <8 x i8> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
  19 // CHECK-NEXT:  entry:
  20 // CHECK-NEXT:    [[VLUTI2_LANEQ:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti2.laneq.v16i8.v8i8(<8 x i8> [[VN]], <16 x i8> [[VM]], i32 0)
  21 // CHECK-NEXT:    ret <16 x i8> [[VLUTI2_LANEQ]]
  22 //
  23 uint8x16_t test_vluti2_laneq_u8(uint8x8_t vn, uint8x16_t vm) {
  24   return vluti2_laneq_u8(vn, vm, 0);
  25 }
  26
  27 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti2q_lane_u8(
  28 // CHECK-SAME: <16 x i8> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
  29 // CHECK-NEXT:  entry:
  30 // CHECK-NEXT:    [[VLUTI2_LANE:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti2.lane.v16i8.v16i8(<16 x i8> [[VN]], <8 x i8> [[VM]], i32 1)
  31 // CHECK-NEXT:    ret <16 x i8> [[VLUTI2_LANE]]
  32 //
  33 uint8x16_t test_vluti2q_lane_u8(uint8x16_t vn, uint8x8_t vm) {
  34   return vluti2q_lane_u8(vn, vm, 1);
  35 }
  36
  37 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti2q_laneq_u8(
  38 // CHECK-SAME: <16 x i8> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
  39 // CHECK-NEXT:  entry:
  40 // CHECK-NEXT:    [[VLUTI2_LANEQ:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti2.laneq.v16i8.v16i8(<16 x i8> [[VN]], <16 x i8> [[VM]], i32 3)
  41 // CHECK-NEXT:    ret <16 x i8> [[VLUTI2_LANEQ]]
  42 //
  43 uint8x16_t test_vluti2q_laneq_u8(uint8x16_t vn, uint8x16_t vm) {
  44   return vluti2q_laneq_u8(vn, vm, 3);
  45 }
  46
  47 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti2_lane_s8(
  48 // CHECK-SAME: <8 x i8> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
  49 // CHECK-NEXT:  entry:
  50 // CHECK-NEXT:    [[VLUTI2_LANE:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti2.lane.v16i8.v8i8(<8 x i8> [[VN]], <8 x i8> [[VM]], i32 0)
  51 // CHECK-NEXT:    ret <16 x i8> [[VLUTI2_LANE]]
  52 //
  53 int8x16_t test_vluti2_lane_s8(int8x8_t vn, uint8x8_t vm) {
  54   return vluti2_lane_s8(vn, vm, 0);
  55 }
  56
  57 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti2_laneq_s8(
  58 // CHECK-SAME: <8 x i8> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
  59 // CHECK-NEXT:  entry:
  60 // CHECK-NEXT:    [[VLUTI2_LANEQ:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti2.laneq.v16i8.v8i8(<8 x i8> [[VN]], <16 x i8> [[VM]], i32 0)
  61 // CHECK-NEXT:    ret <16 x i8> [[VLUTI2_LANEQ]]
  62 //
  63 int8x16_t test_vluti2_laneq_s8(int8x8_t vn, uint8x16_t vm) {
  64   return vluti2_laneq_s8(vn, vm, 0);
  65 }
  66
  67 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti2q_lane_s8(
  68 // CHECK-SAME: <16 x i8> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
  69 // CHECK-NEXT:  entry:
  70 // CHECK-NEXT:    [[VLUTI2_LANE:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti2.lane.v16i8.v16i8(<16 x i8> [[VN]], <8 x i8> [[VM]], i32 1)
  71 // CHECK-NEXT:    ret <16 x i8> [[VLUTI2_LANE]]
  72 //
  73 int8x16_t test_vluti2q_lane_s8(int8x16_t vn, uint8x8_t vm) {
  74   return vluti2q_lane_s8(vn, vm, 1);
  75 }
  76
  77 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti2q_laneq_s8(
  78 // CHECK-SAME: <16 x i8> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
  79 // CHECK-NEXT:  entry:
  80 // CHECK-NEXT:    [[VLUTI2_LANEQ:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti2.laneq.v16i8.v16i8(<16 x i8> [[VN]], <16 x i8> [[VM]], i32 3)
  81 // CHECK-NEXT:    ret <16 x i8> [[VLUTI2_LANEQ]]
  82 //
  83 int8x16_t test_vluti2q_laneq_s8(int8x16_t vn, uint8x16_t vm) {
  84   return vluti2q_laneq_s8(vn, vm, 3);
  85 }
  86
  87 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti2_lane_p8(
  88 // CHECK-SAME: <8 x i8> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
  89 // CHECK-NEXT:  entry:
  90 // CHECK-NEXT:    [[VLUTI2_LANE:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti2.lane.v16i8.v8i8(<8 x i8> [[VN]], <8 x i8> [[VM]], i32 0)
  91 // CHECK-NEXT:    ret <16 x i8> [[VLUTI2_LANE]]
  92 //
  93 poly8x16_t test_vluti2_lane_p8(poly8x8_t vn, uint8x8_t vm) {
  94   return vluti2_lane_p8(vn, vm, 0);
  95 }
  96
  97 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti2_laneq_p8(
  98 // CHECK-SAME: <8 x i8> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
  99 // CHECK-NEXT:  entry:
 100 // CHECK-NEXT:    [[VLUTI2_LANEQ:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti2.laneq.v16i8.v8i8(<8 x i8> [[VN]], <16 x i8> [[VM]], i32 0)
 101 // CHECK-NEXT:    ret <16 x i8> [[VLUTI2_LANEQ]]
 102 //
 103 poly8x16_t test_vluti2_laneq_p8(poly8x8_t vn, uint8x16_t vm) {
 104   return vluti2_laneq_p8(vn, vm, 0);
 105 }
 106
 107 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti2q_lane_p8(
 108 // CHECK-SAME: <16 x i8> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 109 // CHECK-NEXT:  entry:
 110 // CHECK-NEXT:    [[VLUTI2_LANE:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti2.lane.v16i8.v16i8(<16 x i8> [[VN]], <8 x i8> [[VM]], i32 1)
 111 // CHECK-NEXT:    ret <16 x i8> [[VLUTI2_LANE]]
 112 //
 113 poly8x16_t test_vluti2q_lane_p8(poly8x16_t vn, uint8x8_t vm) {
 114   return vluti2q_lane_p8(vn, vm, 1);
 115 }
 116
 117 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti2q_laneq_p8(
 118 // CHECK-SAME: <16 x i8> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 119 // CHECK-NEXT:  entry:
 120 // CHECK-NEXT:    [[VLUTI2_LANEQ:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti2.laneq.v16i8.v16i8(<16 x i8> [[VN]], <16 x i8> [[VM]], i32 3)
 121 // CHECK-NEXT:    ret <16 x i8> [[VLUTI2_LANEQ]]
 122 //
 123 poly8x16_t test_vluti2q_laneq_p8(poly8x16_t vn, uint8x16_t vm) {
 124   return vluti2q_laneq_p8(vn, vm, 3);
 125 }
 126
 127 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti2_lane_u16(
 128 // CHECK-SAME: <4 x i16> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 129 // CHECK-NEXT:  entry:
 130 // CHECK-NEXT:    [[VLUTI2_LANE1:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti2.lane.v8i16.v4i16(<4 x i16> [[VN]], <8 x i8> [[VM]], i32 0)
 131 // CHECK-NEXT:    ret <8 x i16> [[VLUTI2_LANE1]]
 132 //
 133 uint16x8_t test_vluti2_lane_u16(uint16x4_t vn, uint8x8_t vm) {
 134   return vluti2_lane_u16(vn, vm, 0);
 135 }
 136
 137 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti2_laneq_u16(
 138 // CHECK-SAME: <4 x i16> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 139 // CHECK-NEXT:  entry:
 140 // CHECK-NEXT:    [[VLUTI2_LANEQ1:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti2.laneq.v8i16.v4i16(<4 x i16> [[VN]], <16 x i8> [[VM]], i32 0)
 141 // CHECK-NEXT:    ret <8 x i16> [[VLUTI2_LANEQ1]]
 142 //
 143 uint16x8_t test_vluti2_laneq_u16(uint16x4_t vn, uint8x16_t vm) {
 144   return vluti2_laneq_u16(vn, vm, 0);
 145 }
 146
 147 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti2q_lane_u16(
 148 // CHECK-SAME: <8 x i16> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 149 // CHECK-NEXT:  entry:
 150 // CHECK-NEXT:    [[VLUTI2_LANE1:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti2.lane.v8i16.v8i16(<8 x i16> [[VN]], <8 x i8> [[VM]], i32 3)
 151 // CHECK-NEXT:    ret <8 x i16> [[VLUTI2_LANE1]]
 152 //
 153 uint16x8_t test_vluti2q_lane_u16(uint16x8_t vn, uint8x8_t vm) {
 154   return vluti2q_lane_u16(vn, vm, 3);
 155 }
 156
 157 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti2q_laneq_u16(
 158 // CHECK-SAME: <8 x i16> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 159 // CHECK-NEXT:  entry:
 160 // CHECK-NEXT:    [[VLUTI2_LANEQ1:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti2.laneq.v8i16.v8i16(<8 x i16> [[VN]], <16 x i8> [[VM]], i32 7)
 161 // CHECK-NEXT:    ret <8 x i16> [[VLUTI2_LANEQ1]]
 162 //
 163 uint16x8_t test_vluti2q_laneq_u16(uint16x8_t vn, uint8x16_t vm) {
 164   return vluti2q_laneq_u16(vn, vm, 7);
 165 }
 166
 167 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti2_lane_s16(
 168 // CHECK-SAME: <4 x i16> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 169 // CHECK-NEXT:  entry:
 170 // CHECK-NEXT:    [[VLUTI2_LANE1:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti2.lane.v8i16.v4i16(<4 x i16> [[VN]], <8 x i8> [[VM]], i32 0)
 171 // CHECK-NEXT:    ret <8 x i16> [[VLUTI2_LANE1]]
 172 //
 173 int16x8_t test_vluti2_lane_s16(int16x4_t vn, uint8x8_t vm) {
 174   return vluti2_lane_s16(vn, vm, 0);
 175 }
 176
 177 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti2_laneq_s16(
 178 // CHECK-SAME: <4 x i16> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 179 // CHECK-NEXT:  entry:
 180 // CHECK-NEXT:    [[VLUTI2_LANEQ1:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti2.laneq.v8i16.v4i16(<4 x i16> [[VN]], <16 x i8> [[VM]], i32 0)
 181 // CHECK-NEXT:    ret <8 x i16> [[VLUTI2_LANEQ1]]
 182 //
 183 int16x8_t test_vluti2_laneq_s16(int16x4_t vn, uint8x16_t vm) {
 184   return vluti2_laneq_s16(vn, vm, 0);
 185 }
 186
 187 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti2q_lane_s16(
 188 // CHECK-SAME: <8 x i16> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 189 // CHECK-NEXT:  entry:
 190 // CHECK-NEXT:    [[VLUTI2_LANE1:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti2.lane.v8i16.v8i16(<8 x i16> [[VN]], <8 x i8> [[VM]], i32 3)
 191 // CHECK-NEXT:    ret <8 x i16> [[VLUTI2_LANE1]]
 192 //
 193 int16x8_t test_vluti2q_lane_s16(int16x8_t vn, uint8x8_t vm) {
 194   return vluti2q_lane_s16(vn, vm, 3);
 195 }
 196
 197 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti2q_laneq_s16(
 198 // CHECK-SAME: <8 x i16> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 199 // CHECK-NEXT:  entry:
 200 // CHECK-NEXT:    [[VLUTI2_LANEQ1:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti2.laneq.v8i16.v8i16(<8 x i16> [[VN]], <16 x i8> [[VM]], i32 7)
 201 // CHECK-NEXT:    ret <8 x i16> [[VLUTI2_LANEQ1]]
 202 //
 203 int16x8_t test_vluti2q_laneq_s16(int16x8_t vn, uint8x16_t vm) {
 204   return vluti2q_laneq_s16(vn, vm, 7);
 205 }
 206
 207 // CHECK-LABEL: define dso_local <8 x half> @test_vluti2_lane_f16(
 208 // CHECK-SAME: <4 x half> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 209 // CHECK-NEXT:  entry:
 210 // CHECK-NEXT:    [[VLUTI2_LANE1:%.*]] = tail call <8 x half> @llvm.aarch64.neon.vluti2.lane.v8f16.v4f16(<4 x half> [[VN]], <8 x i8> [[VM]], i32 0)
 211 // CHECK-NEXT:    ret <8 x half> [[VLUTI2_LANE1]]
 212 //
 213 float16x8_t test_vluti2_lane_f16(float16x4_t vn, uint8x8_t vm) {
 214   return vluti2_lane_f16(vn, vm, 0);
 215 }
 216
 217 // CHECK-LABEL: define dso_local <8 x half> @test_vluti2_laneq_f16(
 218 // CHECK-SAME: <4 x half> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 219 // CHECK-NEXT:  entry:
 220 // CHECK-NEXT:    [[VLUTI2_LANEQ1:%.*]] = tail call <8 x half> @llvm.aarch64.neon.vluti2.laneq.v8f16.v4f16(<4 x half> [[VN]], <16 x i8> [[VM]], i32 0)
 221 // CHECK-NEXT:    ret <8 x half> [[VLUTI2_LANEQ1]]
 222 //
 223 float16x8_t test_vluti2_laneq_f16(float16x4_t vn, uint8x16_t vm) {
 224   return vluti2_laneq_f16(vn, vm, 0);
 225 }
 226
 227 // CHECK-LABEL: define dso_local <8 x half> @test_vluti2q_lane_f16(
 228 // CHECK-SAME: <8 x half> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 229 // CHECK-NEXT:  entry:
 230 // CHECK-NEXT:    [[VLUTI2_LANE1:%.*]] = tail call <8 x half> @llvm.aarch64.neon.vluti2.lane.v8f16.v8f16(<8 x half> [[VN]], <8 x i8> [[VM]], i32 3)
 231 // CHECK-NEXT:    ret <8 x half> [[VLUTI2_LANE1]]
 232 //
 233 float16x8_t test_vluti2q_lane_f16(float16x8_t vn, uint8x8_t vm) {
 234   return vluti2q_lane_f16(vn, vm, 3);
 235 }
 236
 237 // CHECK-LABEL: define dso_local <8 x half> @test_vluti2q_laneq_f16(
 238 // CHECK-SAME: <8 x half> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 239 // CHECK-NEXT:  entry:
 240 // CHECK-NEXT:    [[VLUTI2_LANEQ1:%.*]] = tail call <8 x half> @llvm.aarch64.neon.vluti2.laneq.v8f16.v8f16(<8 x half> [[VN]], <16 x i8> [[VM]], i32 7)
 241 // CHECK-NEXT:    ret <8 x half> [[VLUTI2_LANEQ1]]
 242 //
 243 float16x8_t test_vluti2q_laneq_f16(float16x8_t vn, uint8x16_t vm) {
 244   return vluti2q_laneq_f16(vn, vm, 7);
 245 }
 246
 247 // CHECK-LABEL: define dso_local <8 x bfloat> @test_vluti2_lane_bf16(
 248 // CHECK-SAME: <4 x bfloat> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 249 // CHECK-NEXT:  entry:
 250 // CHECK-NEXT:    [[VLUTI2_LANE1:%.*]] = tail call <8 x bfloat> @llvm.aarch64.neon.vluti2.lane.v8bf16.v4bf16(<4 x bfloat> [[VN]], <8 x i8> [[VM]], i32 0)
 251 // CHECK-NEXT:    ret <8 x bfloat> [[VLUTI2_LANE1]]
 252 //
 253 bfloat16x8_t test_vluti2_lane_bf16(bfloat16x4_t vn, uint8x8_t vm) {
 254   return vluti2_lane_bf16(vn, vm, 0);
 255 }
 256
 257 // CHECK-LABEL: define dso_local <8 x bfloat> @test_vluti2_laneq_bf16(
 258 // CHECK-SAME: <4 x bfloat> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 259 // CHECK-NEXT:  entry:
 260 // CHECK-NEXT:    [[VLUTI2_LANEQ1:%.*]] = tail call <8 x bfloat> @llvm.aarch64.neon.vluti2.laneq.v8bf16.v4bf16(<4 x bfloat> [[VN]], <16 x i8> [[VM]], i32 0)
 261 // CHECK-NEXT:    ret <8 x bfloat> [[VLUTI2_LANEQ1]]
 262 //
 263 bfloat16x8_t test_vluti2_laneq_bf16(bfloat16x4_t vn, uint8x16_t vm) {
 264   return vluti2_laneq_bf16(vn, vm, 0);
 265 }
 266
 267 // CHECK-LABEL: define dso_local <8 x bfloat> @test_vluti2q_lane_bf16(
 268 // CHECK-SAME: <8 x bfloat> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 269 // CHECK-NEXT:  entry:
 270 // CHECK-NEXT:    [[VLUTI2_LANE1:%.*]] = tail call <8 x bfloat> @llvm.aarch64.neon.vluti2.lane.v8bf16.v8bf16(<8 x bfloat> [[VN]], <8 x i8> [[VM]], i32 3)
 271 // CHECK-NEXT:    ret <8 x bfloat> [[VLUTI2_LANE1]]
 272 //
 273 bfloat16x8_t test_vluti2q_lane_bf16(bfloat16x8_t vn, uint8x8_t vm) {
 274   return vluti2q_lane_bf16(vn, vm, 3);
 275 }
 276
 277 // CHECK-LABEL: define dso_local <8 x bfloat> @test_vluti2q_laneq_bf16(
 278 // CHECK-SAME: <8 x bfloat> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 279 // CHECK-NEXT:  entry:
 280 // CHECK-NEXT:    [[VLUTI2_LANEQ1:%.*]] = tail call <8 x bfloat> @llvm.aarch64.neon.vluti2.laneq.v8bf16.v8bf16(<8 x bfloat> [[VN]], <16 x i8> [[VM]], i32 7)
 281 // CHECK-NEXT:    ret <8 x bfloat> [[VLUTI2_LANEQ1]]
 282 //
 283 bfloat16x8_t test_vluti2q_laneq_bf16(bfloat16x8_t vn, uint8x16_t vm) {
 284   return vluti2q_laneq_bf16(vn, vm, 7);
 285 }
 286
 287 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti2_lane_p16(
 288 // CHECK-SAME: <4 x i16> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 289 // CHECK-NEXT:  entry:
 290 // CHECK-NEXT:    [[VLUTI2_LANE1:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti2.lane.v8i16.v4i16(<4 x i16> [[VN]], <8 x i8> [[VM]], i32 0)
 291 // CHECK-NEXT:    ret <8 x i16> [[VLUTI2_LANE1]]
 292 //
 293 poly16x8_t test_vluti2_lane_p16(poly16x4_t vn, uint8x8_t vm) {
 294   return vluti2_lane_p16(vn, vm, 0);
 295 }
 296
 297 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti2_laneq_p16(
 298 // CHECK-SAME: <4 x i16> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 299 // CHECK-NEXT:  entry:
 300 // CHECK-NEXT:    [[VLUTI2_LANEQ1:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti2.laneq.v8i16.v4i16(<4 x i16> [[VN]], <16 x i8> [[VM]], i32 0)
 301 // CHECK-NEXT:    ret <8 x i16> [[VLUTI2_LANEQ1]]
 302 //
 303 poly16x8_t test_vluti2_laneq_p16(poly16x4_t vn, uint8x16_t vm) {
 304   return vluti2_laneq_p16(vn, vm, 0);
 305 }
 306
 307 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti2q_lane_p16(
 308 // CHECK-SAME: <8 x i16> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 309 // CHECK-NEXT:  entry:
 310 // CHECK-NEXT:    [[VLUTI2_LANE1:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti2.lane.v8i16.v8i16(<8 x i16> [[VN]], <8 x i8> [[VM]], i32 3)
 311 // CHECK-NEXT:    ret <8 x i16> [[VLUTI2_LANE1]]
 312 //
 313 poly16x8_t test_vluti2q_lane_p16(poly16x8_t vn, uint8x8_t vm) {
 314   return vluti2q_lane_p16(vn, vm, 3);
 315 }
 316
 317 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti2q_laneq_p16(
 318 // CHECK-SAME: <8 x i16> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 319 // CHECK-NEXT:  entry:
 320 // CHECK-NEXT:    [[VLUTI2_LANEQ1:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti2.laneq.v8i16.v8i16(<8 x i16> [[VN]], <16 x i8> [[VM]], i32 7)
 321 // CHECK-NEXT:    ret <8 x i16> [[VLUTI2_LANEQ1]]
 322 //
 323 poly16x8_t test_vluti2q_laneq_p16(poly16x8_t vn, uint8x16_t vm) {
 324   return vluti2q_laneq_p16(vn, vm, 7);
 325 }
 326
 327 //
 328
 329 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti4q_lane_u8(
 330 // CHECK-SAME: <16 x i8> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 331 // CHECK-NEXT:  entry:
 332 // CHECK-NEXT:    [[VLUTI4Q_LANE:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti4q.lane.v16i8(<16 x i8> [[VN]], <8 x i8> [[VM]], i32 0)
 333 // CHECK-NEXT:    ret <16 x i8> [[VLUTI4Q_LANE]]
 334 //
 335 uint8x16_t test_vluti4q_lane_u8(uint8x16_t vn, uint8x8_t vm) {
 336   return vluti4q_lane_u8(vn, vm, 0);
 337 }
 338
 339 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti4q_laneq_u8(
 340 // CHECK-SAME: <16 x i8> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 341 // CHECK-NEXT:  entry:
 342 // CHECK-NEXT:    [[VLUTI4Q_LANEQ:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti4q.laneq.v16i8(<16 x i8> [[VN]], <16 x i8> [[VM]], i32 0)
 343 // CHECK-NEXT:    ret <16 x i8> [[VLUTI4Q_LANEQ]]
 344 //
 345 uint8x16_t test_vluti4q_laneq_u8(uint8x16_t vn, uint8x16_t vm) {
 346   return vluti4q_laneq_u8(vn, vm, 0);
 347 }
 348
 349 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti4q_lane_s8(
 350 // CHECK-SAME: <16 x i8> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 351 // CHECK-NEXT:  entry:
 352 // CHECK-NEXT:    [[VLUTI4Q_LANE:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti4q.lane.v16i8(<16 x i8> [[VN]], <8 x i8> [[VM]], i32 0)
 353 // CHECK-NEXT:    ret <16 x i8> [[VLUTI4Q_LANE]]
 354 //
 355 int8x16_t test_vluti4q_lane_s8(int8x16_t vn, uint8x8_t vm) {
 356   return vluti4q_lane_s8(vn, vm, 0);
 357 }
 358
 359 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti4q_laneq_s8(
 360 // CHECK-SAME: <16 x i8> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 361 // CHECK-NEXT:  entry:
 362 // CHECK-NEXT:    [[VLUTI4Q_LANEQ:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti4q.laneq.v16i8(<16 x i8> [[VN]], <16 x i8> [[VM]], i32 1)
 363 // CHECK-NEXT:    ret <16 x i8> [[VLUTI4Q_LANEQ]]
 364 //
 365 int8x16_t test_vluti4q_laneq_s8(int8x16_t vn, uint8x16_t vm) {
 366   return vluti4q_laneq_s8(vn, vm, 1);
 367 }
 368
 369 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti4q_lane_p8(
 370 // CHECK-SAME: <16 x i8> noundef [[VN:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 371 // CHECK-NEXT:  entry:
 372 // CHECK-NEXT:    [[VLUTI4Q_LANE:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti4q.lane.v16i8(<16 x i8> [[VN]], <8 x i8> [[VM]], i32 0)
 373 // CHECK-NEXT:    ret <16 x i8> [[VLUTI4Q_LANE]]
 374 //
 375 poly8x16_t test_vluti4q_lane_p8(poly8x16_t vn, uint8x8_t vm) {
 376   return vluti4q_lane_p8(vn, vm, 0);
 377 }
 378
 379 // CHECK-LABEL: define dso_local <16 x i8> @test_vluti4q_laneq_p8(
 380 // CHECK-SAME: <16 x i8> noundef [[VN:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 381 // CHECK-NEXT:  entry:
 382 // CHECK-NEXT:    [[VLUTI4Q_LANEQ:%.*]] = tail call <16 x i8> @llvm.aarch64.neon.vluti4q.laneq.v16i8(<16 x i8> [[VN]], <16 x i8> [[VM]], i32 1)
 383 // CHECK-NEXT:    ret <16 x i8> [[VLUTI4Q_LANEQ]]
 384 //
 385 poly8x16_t test_vluti4q_laneq_p8(poly8x16_t vn, uint8x16_t vm) {
 386   return vluti4q_laneq_p8(vn, vm, 1);
 387 }
 388
 389 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti4q_lane_u16_x2(
 390 // CHECK-SAME: [2 x <8 x i16>] alignstack(16) [[VN_COERCE:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 391 // CHECK-NEXT:  entry:
 392 // CHECK-NEXT:    [[VN_COERCE_FCA_0_EXTRACT:%.*]] = extractvalue [2 x <8 x i16>] [[VN_COERCE]], 0
 393 // CHECK-NEXT:    [[VN_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x i16>] [[VN_COERCE]], 1
 394 // CHECK-NEXT:    [[VLUTI4Q_LANE_X24:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti4q.lane.x2.v8i16(<8 x i16> [[VN_COERCE_FCA_0_EXTRACT]], <8 x i16> [[VN_COERCE_FCA_1_EXTRACT]], <8 x i8> [[VM]], i32 0)
 395 // CHECK-NEXT:    ret <8 x i16> [[VLUTI4Q_LANE_X24]]
 396 //
 397 uint16x8_t test_vluti4q_lane_u16_x2(uint16x8x2_t vn, uint8x8_t vm) {
 398   return vluti4q_lane_u16_x2(vn, vm, 0);
 399 }
 400
 401 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti4q_laneq_u16_x2(
 402 // CHECK-SAME: [2 x <8 x i16>] alignstack(16) [[VN_COERCE:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 403 // CHECK-NEXT:  entry:
 404 // CHECK-NEXT:    [[VN_COERCE_FCA_0_EXTRACT:%.*]] = extractvalue [2 x <8 x i16>] [[VN_COERCE]], 0
 405 // CHECK-NEXT:    [[VN_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x i16>] [[VN_COERCE]], 1
 406 // CHECK-NEXT:    [[VLUTI4Q_LANEQ_X24:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti4q.laneq.x2.v8i16(<8 x i16> [[VN_COERCE_FCA_0_EXTRACT]], <8 x i16> [[VN_COERCE_FCA_1_EXTRACT]], <16 x i8> [[VM]], i32 0)
 407 // CHECK-NEXT:    ret <8 x i16> [[VLUTI4Q_LANEQ_X24]]
 408 //
 409 uint16x8_t test_vluti4q_laneq_u16_x2(uint16x8x2_t vn, uint8x16_t vm) {
 410   return vluti4q_laneq_u16_x2(vn, vm, 0);
 411 }
 412
 413 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti4q_lane_s16_x2(
 414 // CHECK-SAME: [2 x <8 x i16>] alignstack(16) [[VN_COERCE:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 415 // CHECK-NEXT:  entry:
 416 // CHECK-NEXT:    [[VN_COERCE_FCA_0_EXTRACT:%.*]] = extractvalue [2 x <8 x i16>] [[VN_COERCE]], 0
 417 // CHECK-NEXT:    [[VN_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x i16>] [[VN_COERCE]], 1
 418 // CHECK-NEXT:    [[VLUTI4Q_LANE_X24:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti4q.lane.x2.v8i16(<8 x i16> [[VN_COERCE_FCA_0_EXTRACT]], <8 x i16> [[VN_COERCE_FCA_1_EXTRACT]], <8 x i8> [[VM]], i32 1)
 419 // CHECK-NEXT:    ret <8 x i16> [[VLUTI4Q_LANE_X24]]
 420 //
 421 int16x8_t test_vluti4q_lane_s16_x2(int16x8x2_t vn, uint8x8_t vm) {
 422   return vluti4q_lane_s16_x2(vn, vm, 1);
 423 }
 424
 425 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti4q_laneq_s16_x2(
 426 // CHECK-SAME: [2 x <8 x i16>] alignstack(16) [[VN_COERCE:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 427 // CHECK-NEXT:  entry:
 428 // CHECK-NEXT:    [[VN_COERCE_FCA_0_EXTRACT:%.*]] = extractvalue [2 x <8 x i16>] [[VN_COERCE]], 0
 429 // CHECK-NEXT:    [[VN_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x i16>] [[VN_COERCE]], 1
 430 // CHECK-NEXT:    [[VLUTI4Q_LANEQ_X24:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti4q.laneq.x2.v8i16(<8 x i16> [[VN_COERCE_FCA_0_EXTRACT]], <8 x i16> [[VN_COERCE_FCA_1_EXTRACT]], <16 x i8> [[VM]], i32 3)
 431 // CHECK-NEXT:    ret <8 x i16> [[VLUTI4Q_LANEQ_X24]]
 432 //
 433 int16x8_t test_vluti4q_laneq_s16_x2(int16x8x2_t vn, uint8x16_t vm) {
 434   return vluti4q_laneq_s16_x2(vn, vm, 3);
 435 }
 436
 437 // CHECK-LABEL: define dso_local <8 x half> @test_vluti4q_lane_f16_x2(
 438 // CHECK-SAME: [2 x <8 x half>] alignstack(16) [[VN_COERCE:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 439 // CHECK-NEXT:  entry:
 440 // CHECK-NEXT:    [[VN_COERCE_FCA_0_EXTRACT:%.*]] = extractvalue [2 x <8 x half>] [[VN_COERCE]], 0
 441 // CHECK-NEXT:    [[VN_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x half>] [[VN_COERCE]], 1
 442 // CHECK-NEXT:    [[VLUTI4Q_LANE_X24:%.*]] = tail call <8 x half> @llvm.aarch64.neon.vluti4q.lane.x2.v8f16(<8 x half> [[VN_COERCE_FCA_0_EXTRACT]], <8 x half> [[VN_COERCE_FCA_1_EXTRACT]], <8 x i8> [[VM]], i32 1)
 443 // CHECK-NEXT:    ret <8 x half> [[VLUTI4Q_LANE_X24]]
 444 //
 445 float16x8_t test_vluti4q_lane_f16_x2(float16x8x2_t vn, uint8x8_t vm) {
 446   return vluti4q_lane_f16_x2(vn, vm, 1);
 447 }
 448
 449 // CHECK-LABEL: define dso_local <8 x half> @test_vluti4q_laneq_f16_x2(
 450 // CHECK-SAME: [2 x <8 x half>] alignstack(16) [[VN_COERCE:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 451 // CHECK-NEXT:  entry:
 452 // CHECK-NEXT:    [[VN_COERCE_FCA_0_EXTRACT:%.*]] = extractvalue [2 x <8 x half>] [[VN_COERCE]], 0
 453 // CHECK-NEXT:    [[VN_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x half>] [[VN_COERCE]], 1
 454 // CHECK-NEXT:    [[VLUTI4Q_LANEQ_X24:%.*]] = tail call <8 x half> @llvm.aarch64.neon.vluti4q.laneq.x2.v8f16(<8 x half> [[VN_COERCE_FCA_0_EXTRACT]], <8 x half> [[VN_COERCE_FCA_1_EXTRACT]], <16 x i8> [[VM]], i32 1)
 455 // CHECK-NEXT:    ret <8 x half> [[VLUTI4Q_LANEQ_X24]]
 456 //
 457 float16x8_t test_vluti4q_laneq_f16_x2(float16x8x2_t vn, uint8x16_t vm) {
 458   return vluti4q_laneq_f16_x2(vn, vm, 1);
 459 }
 460
 461 // CHECK-LABEL: define dso_local <8 x bfloat> @test_vluti4q_lane_bf16_x2(
 462 // CHECK-SAME: [2 x <8 x bfloat>] alignstack(16) [[VN_COERCE:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 463 // CHECK-NEXT:  entry:
 464 // CHECK-NEXT:    [[VN_COERCE_FCA_0_EXTRACT:%.*]] = extractvalue [2 x <8 x bfloat>] [[VN_COERCE]], 0
 465 // CHECK-NEXT:    [[VN_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x bfloat>] [[VN_COERCE]], 1
 466 // CHECK-NEXT:    [[VLUTI4Q_LANE_X24:%.*]] = tail call <8 x bfloat> @llvm.aarch64.neon.vluti4q.lane.x2.v8bf16(<8 x bfloat> [[VN_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[VN_COERCE_FCA_1_EXTRACT]], <8 x i8> [[VM]], i32 1)
 467 // CHECK-NEXT:    ret <8 x bfloat> [[VLUTI4Q_LANE_X24]]
 468 //
 469 bfloat16x8_t test_vluti4q_lane_bf16_x2(bfloat16x8x2_t vn, uint8x8_t vm) {
 470   return vluti4q_lane_bf16_x2(vn, vm, 1);
 471 }
 472
 473 // CHECK-LABEL: define dso_local <8 x bfloat> @test_vluti4q_laneq_bf16_x2(
 474 // CHECK-SAME: [2 x <8 x bfloat>] alignstack(16) [[VN_COERCE:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 475 // CHECK-NEXT:  entry:
 476 // CHECK-NEXT:    [[VN_COERCE_FCA_0_EXTRACT:%.*]] = extractvalue [2 x <8 x bfloat>] [[VN_COERCE]], 0
 477 // CHECK-NEXT:    [[VN_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x bfloat>] [[VN_COERCE]], 1
 478 // CHECK-NEXT:    [[VLUTI4Q_LANEQ_X24:%.*]] = tail call <8 x bfloat> @llvm.aarch64.neon.vluti4q.laneq.x2.v8bf16(<8 x bfloat> [[VN_COERCE_FCA_0_EXTRACT]], <8 x bfloat> [[VN_COERCE_FCA_1_EXTRACT]], <16 x i8> [[VM]], i32 2)
 479 // CHECK-NEXT:    ret <8 x bfloat> [[VLUTI4Q_LANEQ_X24]]
 480 //
 481 bfloat16x8_t test_vluti4q_laneq_bf16_x2(bfloat16x8x2_t vn, uint8x16_t vm) {
 482   return vluti4q_laneq_bf16_x2(vn, vm, 2);
 483 }
 484
 485 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti4q_lane_p16_x2(
 486 // CHECK-SAME: [2 x <8 x i16>] alignstack(16) [[VN_COERCE:%.*]], <8 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 487 // CHECK-NEXT:  entry:
 488 // CHECK-NEXT:    [[VN_COERCE_FCA_0_EXTRACT:%.*]] = extractvalue [2 x <8 x i16>] [[VN_COERCE]], 0
 489 // CHECK-NEXT:    [[VN_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x i16>] [[VN_COERCE]], 1
 490 // CHECK-NEXT:    [[VLUTI4Q_LANE_X24:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti4q.lane.x2.v8i16(<8 x i16> [[VN_COERCE_FCA_0_EXTRACT]], <8 x i16> [[VN_COERCE_FCA_1_EXTRACT]], <8 x i8> [[VM]], i32 0)
 491 // CHECK-NEXT:    ret <8 x i16> [[VLUTI4Q_LANE_X24]]
 492 //
 493 poly16x8_t test_vluti4q_lane_p16_x2(poly16x8x2_t vn, uint8x8_t vm) {
 494   return vluti4q_lane_p16_x2(vn, vm, 0);
 495 }
 496
 497 // CHECK-LABEL: define dso_local <8 x i16> @test_vluti4q_laneq_p16_x2(
 498 // CHECK-SAME: [2 x <8 x i16>] alignstack(16) [[VN_COERCE:%.*]], <16 x i8> noundef [[VM:%.*]]) local_unnamed_addr #[[ATTR0]] {
 499 // CHECK-NEXT:  entry:
 500 // CHECK-NEXT:    [[VN_COERCE_FCA_0_EXTRACT:%.*]] = extractvalue [2 x <8 x i16>] [[VN_COERCE]], 0
 501 // CHECK-NEXT:    [[VN_COERCE_FCA_1_EXTRACT:%.*]] = extractvalue [2 x <8 x i16>] [[VN_COERCE]], 1
 502 // CHECK-NEXT:    [[VLUTI4Q_LANEQ_X24:%.*]] = tail call <8 x i16> @llvm.aarch64.neon.vluti4q.laneq.x2.v8i16(<8 x i16> [[VN_COERCE_FCA_0_EXTRACT]], <8 x i16> [[VN_COERCE_FCA_1_EXTRACT]], <16 x i8> [[VM]], i32 0)
 503 // CHECK-NEXT:    ret <8 x i16> [[VLUTI4Q_LANEQ_X24]]
 504 //
 505 poly16x8_t test_vluti4q_laneq_p16_x2(poly16x8x2_t vn, uint8x16_t vm) {
 506   return vluti4q_laneq_p16_x2(vn, vm, 0);
 507 }