clang/test/CodeGen/AArch64/sme2p1-intrinsics/acle_sme2p1_movaz.c

   1 // NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 4
   2 // REQUIRES: aarch64-registered-target
   3 //RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sme2p1 -target-feature +sme -target-feature +bf16 -disable-O0-optnone -Werror -Wall -emit-llvm -o - %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s
   4 // RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sme2p1 -target-feature +sme -target-feature +bf16 -disable-O0-optnone -Werror -Wall -emit-llvm -o - -x c++ %s | opt -S -p mem2reg,instcombine,tailcallelim | FileCheck %s -check-prefix=CPP-CHECK
   5 // RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sme2p1 -target-feature +sme -target-feature +bf16 -S -disable-O0-optnone -Werror -Wall -o /dev/null %s
   6
   7 #include <arm_sme.h>
   8
   9 //
  10 // X2- hor
  11 // CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8> } @test_svreadz_hor_za8_s8_x2(
  12 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0:[0-9]+]] {
  13 // CHECK-NEXT:  entry:
  14 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.horiz.x2.nxv16i8(i32 0, i32 [[SLICE]])
  15 // CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
  16 //
  17 // CPP-CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8> } @_Z26test_svreadz_hor_za8_s8_x2j(
  18 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0:[0-9]+]] {
  19 // CPP-CHECK-NEXT:  entry:
  20 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.horiz.x2.nxv16i8(i32 0, i32 [[SLICE]])
  21 // CPP-CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
  22 //
  23 svint8x2_t test_svreadz_hor_za8_s8_x2(uint32_t slice) __arm_streaming __arm_inout("za")
  24 {
  25    return svreadz_hor_za8_s8_vg2(0, slice);
  26 }
  27
  28 // CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8> } @test_svreadz_hor_za8_u8_x2(
  29 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
  30 // CHECK-NEXT:  entry:
  31 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.horiz.x2.nxv16i8(i32 0, i32 [[SLICE]])
  32 // CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
  33 //
  34 // CPP-CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8> } @_Z26test_svreadz_hor_za8_u8_x2j(
  35 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
  36 // CPP-CHECK-NEXT:  entry:
  37 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.horiz.x2.nxv16i8(i32 0, i32 [[SLICE]])
  38 // CPP-CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
  39 //
  40 svuint8x2_t test_svreadz_hor_za8_u8_x2(uint32_t slice) __arm_streaming __arm_inout("za")
  41 {
  42    return svreadz_hor_za8_u8_vg2(0, slice);
  43 }
  44
  45 // CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16> } @test_svreadz_hor_za16_s16_x2(
  46 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
  47 // CHECK-NEXT:  entry:
  48 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.horiz.x2.nxv8i16(i32 0, i32 [[SLICE]])
  49 // CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
  50 //
  51 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16> } @_Z28test_svreadz_hor_za16_s16_x2j(
  52 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
  53 // CPP-CHECK-NEXT:  entry:
  54 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.horiz.x2.nxv8i16(i32 0, i32 [[SLICE]])
  55 // CPP-CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
  56 //
  57 svint16x2_t test_svreadz_hor_za16_s16_x2(uint32_t slice) __arm_streaming __arm_inout("za")
  58 {
  59    return svreadz_hor_za16_s16_vg2(0, slice);
  60 }
  61
  62 // CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16> } @test_svreadz_hor_za16_u16_x2(
  63 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
  64 // CHECK-NEXT:  entry:
  65 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.horiz.x2.nxv8i16(i32 1, i32 [[SLICE]])
  66 // CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
  67 //
  68 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16> } @_Z28test_svreadz_hor_za16_u16_x2j(
  69 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
  70 // CPP-CHECK-NEXT:  entry:
  71 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.horiz.x2.nxv8i16(i32 1, i32 [[SLICE]])
  72 // CPP-CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
  73 //
  74 svuint16x2_t test_svreadz_hor_za16_u16_x2(uint32_t slice) __arm_streaming __arm_inout("za")
  75 {
  76    return svreadz_hor_za16_u16_vg2(1, slice);
  77 }
  78
  79 // CHECK-LABEL: define dso_local { <vscale x 8 x half>, <vscale x 8 x half> } @test_svreadz_hor_za16_f16_x2(
  80 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
  81 // CHECK-NEXT:  entry:
  82 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.readz.horiz.x2.nxv8f16(i32 0, i32 [[SLICE]])
  83 // CHECK-NEXT:    ret { <vscale x 8 x half>, <vscale x 8 x half> } [[TMP0]]
  84 //
  85 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x half>, <vscale x 8 x half> } @_Z28test_svreadz_hor_za16_f16_x2j(
  86 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
  87 // CPP-CHECK-NEXT:  entry:
  88 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.readz.horiz.x2.nxv8f16(i32 0, i32 [[SLICE]])
  89 // CPP-CHECK-NEXT:    ret { <vscale x 8 x half>, <vscale x 8 x half> } [[TMP0]]
  90 //
  91 svfloat16x2_t test_svreadz_hor_za16_f16_x2(uint32_t slice) __arm_streaming __arm_inout("za")
  92 {
  93    return svreadz_hor_za16_f16_vg2(0, slice);
  94 }
  95
  96 // CHECK-LABEL: define dso_local { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @test_svreadz_hor_za16_bf16_x2(
  97 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
  98 // CHECK-NEXT:  entry:
  99 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @llvm.aarch64.sme.readz.horiz.x2.nxv8bf16(i32 1, i32 [[SLICE]])
 100 // CHECK-NEXT:    ret { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } [[TMP0]]
 101 //
 102 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @_Z29test_svreadz_hor_za16_bf16_x2j(
 103 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 104 // CPP-CHECK-NEXT:  entry:
 105 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @llvm.aarch64.sme.readz.horiz.x2.nxv8bf16(i32 1, i32 [[SLICE]])
 106 // CPP-CHECK-NEXT:    ret { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } [[TMP0]]
 107 //
 108 svbfloat16x2_t test_svreadz_hor_za16_bf16_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 109 {
 110    return svreadz_hor_za16_bf16_vg2(1, slice);
 111 }
 112
 113
 114 // CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32> } @test_svreadz_hor_za32_s32_x2(
 115 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 116 // CHECK-NEXT:  entry:
 117 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.horiz.x2.nxv4i32(i32 0, i32 [[SLICE]])
 118 // CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 119 //
 120 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32> } @_Z28test_svreadz_hor_za32_s32_x2j(
 121 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 122 // CPP-CHECK-NEXT:  entry:
 123 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.horiz.x2.nxv4i32(i32 0, i32 [[SLICE]])
 124 // CPP-CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 125 //
 126 svint32x2_t test_svreadz_hor_za32_s32_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 127 {
 128    return svreadz_hor_za32_s32_vg2(0, slice);
 129 }
 130
 131 // CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32> } @test_svreadz_hor_za32_u32_x2(
 132 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 133 // CHECK-NEXT:  entry:
 134 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.horiz.x2.nxv4i32(i32 2, i32 [[SLICE]])
 135 // CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 136 //
 137 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32> } @_Z28test_svreadz_hor_za32_u32_x2j(
 138 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 139 // CPP-CHECK-NEXT:  entry:
 140 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.horiz.x2.nxv4i32(i32 2, i32 [[SLICE]])
 141 // CPP-CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 142 //
 143 svuint32x2_t test_svreadz_hor_za32_u32_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 144 {
 145    return svreadz_hor_za32_u32_vg2(2, slice);
 146 }
 147
 148 // CHECK-LABEL: define dso_local { <vscale x 4 x float>, <vscale x 4 x float> } @test_svreadz_hor_za32_f32_x2(
 149 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 150 // CHECK-NEXT:  entry:
 151 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.readz.horiz.x2.nxv4f32(i32 3, i32 [[SLICE]])
 152 // CHECK-NEXT:    ret { <vscale x 4 x float>, <vscale x 4 x float> } [[TMP0]]
 153 //
 154 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x float>, <vscale x 4 x float> } @_Z28test_svreadz_hor_za32_f32_x2j(
 155 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 156 // CPP-CHECK-NEXT:  entry:
 157 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.readz.horiz.x2.nxv4f32(i32 3, i32 [[SLICE]])
 158 // CPP-CHECK-NEXT:    ret { <vscale x 4 x float>, <vscale x 4 x float> } [[TMP0]]
 159 //
 160 svfloat32x2_t test_svreadz_hor_za32_f32_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 161 {
 162    return svreadz_hor_za32_f32_vg2(3, slice);
 163 }
 164
 165 // CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64> } @test_svreadz_hor_za64_s64_x2(
 166 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 167 // CHECK-NEXT:  entry:
 168 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.horiz.x2.nxv2i64(i32 0, i32 [[SLICE]])
 169 // CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 170 //
 171 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64> } @_Z28test_svreadz_hor_za64_s64_x2j(
 172 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 173 // CPP-CHECK-NEXT:  entry:
 174 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.horiz.x2.nxv2i64(i32 0, i32 [[SLICE]])
 175 // CPP-CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 176 //
 177 svint64x2_t test_svreadz_hor_za64_s64_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 178 {
 179    return svreadz_hor_za64_s64_vg2(0, slice);
 180 }
 181
 182 // CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64> } @test_svreadz_hor_za64_u64_x2(
 183 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 184 // CHECK-NEXT:  entry:
 185 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.horiz.x2.nxv2i64(i32 4, i32 [[SLICE]])
 186 // CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 187 //
 188 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64> } @_Z28test_svreadz_hor_za64_u64_x2j(
 189 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 190 // CPP-CHECK-NEXT:  entry:
 191 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.horiz.x2.nxv2i64(i32 4, i32 [[SLICE]])
 192 // CPP-CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 193 //
 194 svuint64x2_t test_svreadz_hor_za64_u64_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 195 {
 196    return svreadz_hor_za64_u64_vg2(4, slice);
 197 }
 198
 199 // CHECK-LABEL: define dso_local { <vscale x 2 x double>, <vscale x 2 x double> } @test_svreadz_hor_za64_f64_x2(
 200 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 201 // CHECK-NEXT:  entry:
 202 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.readz.horiz.x2.nxv2f64(i32 7, i32 [[SLICE]])
 203 // CHECK-NEXT:    ret { <vscale x 2 x double>, <vscale x 2 x double> } [[TMP0]]
 204 //
 205 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x double>, <vscale x 2 x double> } @_Z28test_svreadz_hor_za64_f64_x2j(
 206 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 207 // CPP-CHECK-NEXT:  entry:
 208 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.readz.horiz.x2.nxv2f64(i32 7, i32 [[SLICE]])
 209 // CPP-CHECK-NEXT:    ret { <vscale x 2 x double>, <vscale x 2 x double> } [[TMP0]]
 210 //
 211 svfloat64x2_t test_svreadz_hor_za64_f64_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 212 {
 213    return svreadz_hor_za64_f64_vg2(7, slice);
 214 }
 215
 216
 217 //
 218 // X2- ver
 219 //
 220
 221 // CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8> } @test_svreadz_ver_za8_s8_x2(
 222 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 223 // CHECK-NEXT:  entry:
 224 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.vert.x2.nxv16i8(i32 0, i32 [[SLICE]])
 225 // CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
 226 //
 227 // CPP-CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8> } @_Z26test_svreadz_ver_za8_s8_x2j(
 228 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 229 // CPP-CHECK-NEXT:  entry:
 230 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.vert.x2.nxv16i8(i32 0, i32 [[SLICE]])
 231 // CPP-CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
 232 //
 233 svint8x2_t test_svreadz_ver_za8_s8_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 234 {
 235    return svreadz_ver_za8_s8_vg2(0, slice);
 236 }
 237
 238 // CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8> } @test_svreadz_ver_za8_u8_x2(
 239 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 240 // CHECK-NEXT:  entry:
 241 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.vert.x2.nxv16i8(i32 0, i32 [[SLICE]])
 242 // CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
 243 //
 244 // CPP-CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8> } @_Z26test_svreadz_ver_za8_u8_x2j(
 245 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 246 // CPP-CHECK-NEXT:  entry:
 247 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.vert.x2.nxv16i8(i32 0, i32 [[SLICE]])
 248 // CPP-CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
 249 //
 250 svuint8x2_t test_svreadz_ver_za8_u8_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 251 {
 252    return svreadz_ver_za8_u8_vg2(0, slice);
 253 }
 254
 255 // CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16> } @test_svreadz_ver_za16_s16_x2(
 256 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 257 // CHECK-NEXT:  entry:
 258 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.vert.x2.nxv8i16(i32 0, i32 [[SLICE]])
 259 // CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
 260 //
 261 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16> } @_Z28test_svreadz_ver_za16_s16_x2j(
 262 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 263 // CPP-CHECK-NEXT:  entry:
 264 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.vert.x2.nxv8i16(i32 0, i32 [[SLICE]])
 265 // CPP-CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
 266 //
 267 svint16x2_t test_svreadz_ver_za16_s16_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 268 {
 269    return svreadz_ver_za16_s16_vg2(0, slice);
 270 }
 271
 272 // CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16> } @test_svreadz_ver_za16_u16_x2(
 273 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 274 // CHECK-NEXT:  entry:
 275 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.vert.x2.nxv8i16(i32 1, i32 [[SLICE]])
 276 // CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
 277 //
 278 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16> } @_Z28test_svreadz_ver_za16_u16_x2j(
 279 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 280 // CPP-CHECK-NEXT:  entry:
 281 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.vert.x2.nxv8i16(i32 1, i32 [[SLICE]])
 282 // CPP-CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
 283 //
 284 svuint16x2_t test_svreadz_ver_za16_u16_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 285 {
 286    return svreadz_ver_za16_u16_vg2(1, slice);
 287 }
 288
 289 // CHECK-LABEL: define dso_local { <vscale x 8 x half>, <vscale x 8 x half> } @test_svreadz_ver_za16_f16_x2(
 290 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 291 // CHECK-NEXT:  entry:
 292 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.readz.vert.x2.nxv8f16(i32 0, i32 [[SLICE]])
 293 // CHECK-NEXT:    ret { <vscale x 8 x half>, <vscale x 8 x half> } [[TMP0]]
 294 //
 295 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x half>, <vscale x 8 x half> } @_Z28test_svreadz_ver_za16_f16_x2j(
 296 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 297 // CPP-CHECK-NEXT:  entry:
 298 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.readz.vert.x2.nxv8f16(i32 0, i32 [[SLICE]])
 299 // CPP-CHECK-NEXT:    ret { <vscale x 8 x half>, <vscale x 8 x half> } [[TMP0]]
 300 //
 301 svfloat16x2_t test_svreadz_ver_za16_f16_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 302 {
 303    return svreadz_ver_za16_f16_vg2(0, slice);
 304 }
 305
 306 // CHECK-LABEL: define dso_local { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @test_svreadz_ver_za16_bf16_x2(
 307 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 308 // CHECK-NEXT:  entry:
 309 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @llvm.aarch64.sme.readz.vert.x2.nxv8bf16(i32 1, i32 [[SLICE]])
 310 // CHECK-NEXT:    ret { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } [[TMP0]]
 311 //
 312 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @_Z29test_svreadz_ver_za16_bf16_x2j(
 313 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 314 // CPP-CHECK-NEXT:  entry:
 315 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @llvm.aarch64.sme.readz.vert.x2.nxv8bf16(i32 1, i32 [[SLICE]])
 316 // CPP-CHECK-NEXT:    ret { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } [[TMP0]]
 317 //
 318 svbfloat16x2_t test_svreadz_ver_za16_bf16_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 319 {
 320    return svreadz_ver_za16_bf16_vg2(1, slice);
 321 }
 322
 323
 324 // CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32> } @test_svreadz_ver_za32_s32_x2(
 325 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 326 // CHECK-NEXT:  entry:
 327 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.vert.x2.nxv4i32(i32 0, i32 [[SLICE]])
 328 // CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 329 //
 330 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32> } @_Z28test_svreadz_ver_za32_s32_x2j(
 331 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 332 // CPP-CHECK-NEXT:  entry:
 333 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.vert.x2.nxv4i32(i32 0, i32 [[SLICE]])
 334 // CPP-CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 335 //
 336 svint32x2_t test_svreadz_ver_za32_s32_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 337 {
 338    return svreadz_ver_za32_s32_vg2(0, slice);
 339 }
 340
 341 // CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32> } @test_svreadz_ver_za32_u32_x2(
 342 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 343 // CHECK-NEXT:  entry:
 344 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.vert.x2.nxv4i32(i32 2, i32 [[SLICE]])
 345 // CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 346 //
 347 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32> } @_Z28test_svreadz_ver_za32_u32_x2j(
 348 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 349 // CPP-CHECK-NEXT:  entry:
 350 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.vert.x2.nxv4i32(i32 2, i32 [[SLICE]])
 351 // CPP-CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 352 //
 353 svuint32x2_t test_svreadz_ver_za32_u32_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 354 {
 355    return svreadz_ver_za32_u32_vg2(2, slice);
 356 }
 357
 358 // CHECK-LABEL: define dso_local { <vscale x 4 x float>, <vscale x 4 x float> } @test_svreadz_ver_za32_f32_x2(
 359 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 360 // CHECK-NEXT:  entry:
 361 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.readz.vert.x2.nxv4f32(i32 3, i32 [[SLICE]])
 362 // CHECK-NEXT:    ret { <vscale x 4 x float>, <vscale x 4 x float> } [[TMP0]]
 363 //
 364 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x float>, <vscale x 4 x float> } @_Z28test_svreadz_ver_za32_f32_x2j(
 365 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 366 // CPP-CHECK-NEXT:  entry:
 367 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.readz.vert.x2.nxv4f32(i32 3, i32 [[SLICE]])
 368 // CPP-CHECK-NEXT:    ret { <vscale x 4 x float>, <vscale x 4 x float> } [[TMP0]]
 369 //
 370 svfloat32x2_t test_svreadz_ver_za32_f32_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 371 {
 372    return svreadz_ver_za32_f32_vg2(3, slice);
 373 }
 374
 375 // CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64> } @test_svreadz_ver_za64_s64_x2(
 376 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 377 // CHECK-NEXT:  entry:
 378 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.vert.x2.nxv2i64(i32 0, i32 [[SLICE]])
 379 // CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 380 //
 381 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64> } @_Z28test_svreadz_ver_za64_s64_x2j(
 382 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 383 // CPP-CHECK-NEXT:  entry:
 384 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.vert.x2.nxv2i64(i32 0, i32 [[SLICE]])
 385 // CPP-CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 386 //
 387 svint64x2_t test_svreadz_ver_za64_s64_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 388 {
 389    return svreadz_ver_za64_s64_vg2(0, slice);
 390 }
 391
 392 // CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64> } @test_svreadz_ver_za64_u64_x2(
 393 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 394 // CHECK-NEXT:  entry:
 395 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.vert.x2.nxv2i64(i32 4, i32 [[SLICE]])
 396 // CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 397 //
 398 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64> } @_Z28test_svreadz_ver_za64_u64_x2j(
 399 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 400 // CPP-CHECK-NEXT:  entry:
 401 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.vert.x2.nxv2i64(i32 4, i32 [[SLICE]])
 402 // CPP-CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 403 //
 404 svuint64x2_t test_svreadz_ver_za64_u64_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 405 {
 406    return svreadz_ver_za64_u64_vg2(4, slice);
 407 }
 408
 409 // CHECK-LABEL: define dso_local { <vscale x 2 x double>, <vscale x 2 x double> } @test_svreadz_ver_za64_f64_x2(
 410 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 411 // CHECK-NEXT:  entry:
 412 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.readz.vert.x2.nxv2f64(i32 7, i32 [[SLICE]])
 413 // CHECK-NEXT:    ret { <vscale x 2 x double>, <vscale x 2 x double> } [[TMP0]]
 414 //
 415 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x double>, <vscale x 2 x double> } @_Z28test_svreadz_ver_za64_f64_x2j(
 416 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 417 // CPP-CHECK-NEXT:  entry:
 418 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.readz.vert.x2.nxv2f64(i32 7, i32 [[SLICE]])
 419 // CPP-CHECK-NEXT:    ret { <vscale x 2 x double>, <vscale x 2 x double> } [[TMP0]]
 420 //
 421 svfloat64x2_t test_svreadz_ver_za64_f64_x2(uint32_t slice) __arm_streaming __arm_inout("za")
 422 {
 423    return svreadz_ver_za64_f64_vg2(7, slice);
 424 }
 425
 426
 427 //
 428 // X4 - hor
 429 // CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @test_svreadz_hor_za8_s8_x4(
 430 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 431 // CHECK-NEXT:  entry:
 432 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.horiz.x4.nxv16i8(i32 0, i32 [[SLICE]])
 433 // CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
 434 //
 435 // CPP-CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @_Z26test_svreadz_hor_za8_s8_x4j(
 436 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 437 // CPP-CHECK-NEXT:  entry:
 438 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.horiz.x4.nxv16i8(i32 0, i32 [[SLICE]])
 439 // CPP-CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
 440 //
 441 svint8x4_t test_svreadz_hor_za8_s8_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 442 {
 443    return svreadz_hor_za8_s8_vg4(0, slice);
 444 }
 445
 446 // CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @test_svreadz_hor_za8_u8_x4(
 447 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 448 // CHECK-NEXT:  entry:
 449 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.horiz.x4.nxv16i8(i32 0, i32 [[SLICE]])
 450 // CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
 451 //
 452 // CPP-CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @_Z26test_svreadz_hor_za8_u8_x4j(
 453 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 454 // CPP-CHECK-NEXT:  entry:
 455 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.horiz.x4.nxv16i8(i32 0, i32 [[SLICE]])
 456 // CPP-CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
 457 //
 458 svuint8x4_t test_svreadz_hor_za8_u8_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 459 {
 460    return svreadz_hor_za8_u8_vg4(0, slice);
 461 }
 462
 463 // CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @test_svreadz_hor_za16_s16_x4(
 464 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 465 // CHECK-NEXT:  entry:
 466 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.horiz.x4.nxv8i16(i32 0, i32 [[SLICE]])
 467 // CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
 468 //
 469 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @_Z28test_svreadz_hor_za16_s16_x4j(
 470 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 471 // CPP-CHECK-NEXT:  entry:
 472 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.horiz.x4.nxv8i16(i32 0, i32 [[SLICE]])
 473 // CPP-CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
 474 //
 475 svint16x4_t test_svreadz_hor_za16_s16_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 476 {
 477    return svreadz_hor_za16_s16_vg4(0, slice);
 478 }
 479
 480 // CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @test_svreadz_hor_za16_u16_x4(
 481 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 482 // CHECK-NEXT:  entry:
 483 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.horiz.x4.nxv8i16(i32 1, i32 [[SLICE]])
 484 // CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
 485 //
 486 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @_Z28test_svreadz_hor_za16_u16_x4j(
 487 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 488 // CPP-CHECK-NEXT:  entry:
 489 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.horiz.x4.nxv8i16(i32 1, i32 [[SLICE]])
 490 // CPP-CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
 491 //
 492 svuint16x4_t test_svreadz_hor_za16_u16_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 493 {
 494    return svreadz_hor_za16_u16_vg4(1, slice);
 495 }
 496
 497 // CHECK-LABEL: define dso_local { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @test_svreadz_hor_za16_f16_x4(
 498 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 499 // CHECK-NEXT:  entry:
 500 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.readz.horiz.x4.nxv8f16(i32 0, i32 [[SLICE]])
 501 // CHECK-NEXT:    ret { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } [[TMP0]]
 502 //
 503 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @_Z28test_svreadz_hor_za16_f16_x4j(
 504 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 505 // CPP-CHECK-NEXT:  entry:
 506 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.readz.horiz.x4.nxv8f16(i32 0, i32 [[SLICE]])
 507 // CPP-CHECK-NEXT:    ret { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } [[TMP0]]
 508 //
 509 svfloat16x4_t test_svreadz_hor_za16_f16_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 510 {
 511    return svreadz_hor_za16_f16_vg4(0, slice);
 512 }
 513
 514 // CHECK-LABEL: define dso_local { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @test_svreadz_hor_za16_bf16_x4(
 515 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 516 // CHECK-NEXT:  entry:
 517 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @llvm.aarch64.sme.readz.horiz.x4.nxv8bf16(i32 1, i32 [[SLICE]])
 518 // CHECK-NEXT:    ret { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } [[TMP0]]
 519 //
 520 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @_Z29test_svreadz_hor_za16_bf16_x4j(
 521 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 522 // CPP-CHECK-NEXT:  entry:
 523 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @llvm.aarch64.sme.readz.horiz.x4.nxv8bf16(i32 1, i32 [[SLICE]])
 524 // CPP-CHECK-NEXT:    ret { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } [[TMP0]]
 525 //
 526 svbfloat16x4_t test_svreadz_hor_za16_bf16_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 527 {
 528    return svreadz_hor_za16_bf16_vg4(1, slice);
 529 }
 530
 531
 532 // CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @test_svreadz_hor_za32_s32_x4(
 533 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 534 // CHECK-NEXT:  entry:
 535 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.horiz.x4.nxv4i32(i32 0, i32 [[SLICE]])
 536 // CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 537 //
 538 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @_Z28test_svreadz_hor_za32_s32_x4j(
 539 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 540 // CPP-CHECK-NEXT:  entry:
 541 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.horiz.x4.nxv4i32(i32 0, i32 [[SLICE]])
 542 // CPP-CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 543 //
 544 svint32x4_t test_svreadz_hor_za32_s32_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 545 {
 546    return svreadz_hor_za32_s32_vg4(0, slice);
 547 }
 548
 549 // CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @test_svreadz_hor_za32_u32_x4(
 550 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 551 // CHECK-NEXT:  entry:
 552 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.horiz.x4.nxv4i32(i32 2, i32 [[SLICE]])
 553 // CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 554 //
 555 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @_Z28test_svreadz_hor_za32_u32_x4j(
 556 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 557 // CPP-CHECK-NEXT:  entry:
 558 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.horiz.x4.nxv4i32(i32 2, i32 [[SLICE]])
 559 // CPP-CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 560 //
 561 svuint32x4_t test_svreadz_hor_za32_u32_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 562 {
 563    return svreadz_hor_za32_u32_vg4(2, slice);
 564 }
 565
 566 // CHECK-LABEL: define dso_local { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @test_svreadz_hor_za32_f32_x4(
 567 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 568 // CHECK-NEXT:  entry:
 569 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.readz.horiz.x4.nxv4f32(i32 3, i32 [[SLICE]])
 570 // CHECK-NEXT:    ret { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } [[TMP0]]
 571 //
 572 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @_Z28test_svreadz_hor_za32_f32_x4j(
 573 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 574 // CPP-CHECK-NEXT:  entry:
 575 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.readz.horiz.x4.nxv4f32(i32 3, i32 [[SLICE]])
 576 // CPP-CHECK-NEXT:    ret { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } [[TMP0]]
 577 //
 578 svfloat32x4_t test_svreadz_hor_za32_f32_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 579 {
 580    return svreadz_hor_za32_f32_vg4(3, slice);
 581 }
 582
 583 // CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @test_svreadz_hor_za64_s64_x4(
 584 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 585 // CHECK-NEXT:  entry:
 586 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.horiz.x4.nxv2i64(i32 0, i32 [[SLICE]])
 587 // CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 588 //
 589 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @_Z28test_svreadz_hor_za64_s64_x4j(
 590 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 591 // CPP-CHECK-NEXT:  entry:
 592 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.horiz.x4.nxv2i64(i32 0, i32 [[SLICE]])
 593 // CPP-CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 594 //
 595 svint64x4_t test_svreadz_hor_za64_s64_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 596 {
 597    return svreadz_hor_za64_s64_vg4(0, slice);
 598 }
 599
 600 // CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @test_svreadz_hor_za64_u64_x4(
 601 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 602 // CHECK-NEXT:  entry:
 603 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.horiz.x4.nxv2i64(i32 4, i32 [[SLICE]])
 604 // CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 605 //
 606 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @_Z28test_svreadz_hor_za64_u64_x4j(
 607 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 608 // CPP-CHECK-NEXT:  entry:
 609 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.horiz.x4.nxv2i64(i32 4, i32 [[SLICE]])
 610 // CPP-CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 611 //
 612 svuint64x4_t test_svreadz_hor_za64_u64_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 613 {
 614    return svreadz_hor_za64_u64_vg4(4, slice);
 615 }
 616
 617 // CHECK-LABEL: define dso_local { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @test_svreadz_hor_za64_f64_x4(
 618 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 619 // CHECK-NEXT:  entry:
 620 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.readz.horiz.x4.nxv2f64(i32 7, i32 [[SLICE]])
 621 // CHECK-NEXT:    ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } [[TMP0]]
 622 //
 623 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @_Z28test_svreadz_hor_za64_f64_x4j(
 624 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 625 // CPP-CHECK-NEXT:  entry:
 626 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.readz.horiz.x4.nxv2f64(i32 7, i32 [[SLICE]])
 627 // CPP-CHECK-NEXT:    ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } [[TMP0]]
 628 //
 629 svfloat64x4_t test_svreadz_hor_za64_f64_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 630 {
 631    return svreadz_hor_za64_f64_vg4(7, slice);
 632 }
 633
 634 //
 635 // X4 - ver
 636 // CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @test_svreadz_ver_za8_s8_x4(
 637 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 638 // CHECK-NEXT:  entry:
 639 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.vert.x4.nxv16i8(i32 0, i32 [[SLICE]])
 640 // CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
 641 //
 642 // CPP-CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @_Z26test_svreadz_ver_za8_s8_x4j(
 643 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 644 // CPP-CHECK-NEXT:  entry:
 645 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.vert.x4.nxv16i8(i32 0, i32 [[SLICE]])
 646 // CPP-CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
 647 //
 648 svint8x4_t test_svreadz_ver_za8_s8_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 649 {
 650    return svreadz_ver_za8_s8_vg4(0, slice);
 651 }
 652
 653 // CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @test_svreadz_ver_za8_u8_x4(
 654 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 655 // CHECK-NEXT:  entry:
 656 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.vert.x4.nxv16i8(i32 0, i32 [[SLICE]])
 657 // CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
 658 //
 659 // CPP-CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @_Z26test_svreadz_ver_za8_u8_x4j(
 660 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 661 // CPP-CHECK-NEXT:  entry:
 662 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.vert.x4.nxv16i8(i32 0, i32 [[SLICE]])
 663 // CPP-CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
 664 //
 665 svuint8x4_t test_svreadz_ver_za8_u8_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 666 {
 667    return svreadz_ver_za8_u8_vg4(0, slice);
 668 }
 669
 670 // CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @test_svreadz_ver_za16_s16_x4(
 671 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 672 // CHECK-NEXT:  entry:
 673 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.vert.x4.nxv8i16(i32 0, i32 [[SLICE]])
 674 // CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
 675 //
 676 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @_Z28test_svreadz_ver_za16_s16_x4j(
 677 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 678 // CPP-CHECK-NEXT:  entry:
 679 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.vert.x4.nxv8i16(i32 0, i32 [[SLICE]])
 680 // CPP-CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
 681 //
 682 svint16x4_t test_svreadz_ver_za16_s16_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 683 {
 684    return svreadz_ver_za16_s16_vg4(0, slice);
 685 }
 686
 687 // CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @test_svreadz_ver_za16_u16_x4(
 688 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 689 // CHECK-NEXT:  entry:
 690 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.vert.x4.nxv8i16(i32 1, i32 [[SLICE]])
 691 // CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
 692 //
 693 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @_Z28test_svreadz_ver_za16_u16_x4j(
 694 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 695 // CPP-CHECK-NEXT:  entry:
 696 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.vert.x4.nxv8i16(i32 1, i32 [[SLICE]])
 697 // CPP-CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
 698 //
 699 svuint16x4_t test_svreadz_ver_za16_u16_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 700 {
 701    return svreadz_ver_za16_u16_vg4(1, slice);
 702 }
 703
 704 // CHECK-LABEL: define dso_local { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @test_svreadz_ver_za16_f16_x4(
 705 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 706 // CHECK-NEXT:  entry:
 707 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.readz.vert.x4.nxv8f16(i32 0, i32 [[SLICE]])
 708 // CHECK-NEXT:    ret { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } [[TMP0]]
 709 //
 710 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @_Z28test_svreadz_ver_za16_f16_x4j(
 711 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 712 // CPP-CHECK-NEXT:  entry:
 713 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.readz.vert.x4.nxv8f16(i32 0, i32 [[SLICE]])
 714 // CPP-CHECK-NEXT:    ret { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } [[TMP0]]
 715 //
 716 svfloat16x4_t test_svreadz_ver_za16_f16_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 717 {
 718    return svreadz_ver_za16_f16_vg4(0, slice);
 719 }
 720
 721 // CHECK-LABEL: define dso_local { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @test_svreadz_ver_za16_bf16_x4(
 722 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 723 // CHECK-NEXT:  entry:
 724 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @llvm.aarch64.sme.readz.vert.x4.nxv8bf16(i32 1, i32 [[SLICE]])
 725 // CHECK-NEXT:    ret { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } [[TMP0]]
 726 //
 727 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @_Z29test_svreadz_ver_za16_bf16_x4j(
 728 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 729 // CPP-CHECK-NEXT:  entry:
 730 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @llvm.aarch64.sme.readz.vert.x4.nxv8bf16(i32 1, i32 [[SLICE]])
 731 // CPP-CHECK-NEXT:    ret { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } [[TMP0]]
 732 //
 733 svbfloat16x4_t test_svreadz_ver_za16_bf16_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 734 {
 735    return svreadz_ver_za16_bf16_vg4(1, slice);
 736 }
 737
 738
 739 // CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @test_svreadz_ver_za32_s32_x4(
 740 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 741 // CHECK-NEXT:  entry:
 742 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.vert.x4.nxv4i32(i32 0, i32 [[SLICE]])
 743 // CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 744 //
 745 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @_Z28test_svreadz_ver_za32_s32_x4j(
 746 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 747 // CPP-CHECK-NEXT:  entry:
 748 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.vert.x4.nxv4i32(i32 0, i32 [[SLICE]])
 749 // CPP-CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 750 //
 751 svint32x4_t test_svreadz_ver_za32_s32_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 752 {
 753    return svreadz_ver_za32_s32_vg4(0, slice);
 754 }
 755
 756 // CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @test_svreadz_ver_za32_u32_x4(
 757 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 758 // CHECK-NEXT:  entry:
 759 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.vert.x4.nxv4i32(i32 2, i32 [[SLICE]])
 760 // CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 761 //
 762 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @_Z28test_svreadz_ver_za32_u32_x4j(
 763 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 764 // CPP-CHECK-NEXT:  entry:
 765 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.vert.x4.nxv4i32(i32 2, i32 [[SLICE]])
 766 // CPP-CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
 767 //
 768 svuint32x4_t test_svreadz_ver_za32_u32_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 769 {
 770    return svreadz_ver_za32_u32_vg4(2, slice);
 771 }
 772
 773 // CHECK-LABEL: define dso_local { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @test_svreadz_ver_za32_f32_x4(
 774 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 775 // CHECK-NEXT:  entry:
 776 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.readz.vert.x4.nxv4f32(i32 3, i32 [[SLICE]])
 777 // CHECK-NEXT:    ret { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } [[TMP0]]
 778 //
 779 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @_Z28test_svreadz_ver_za32_f32_x4j(
 780 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 781 // CPP-CHECK-NEXT:  entry:
 782 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.readz.vert.x4.nxv4f32(i32 3, i32 [[SLICE]])
 783 // CPP-CHECK-NEXT:    ret { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } [[TMP0]]
 784 //
 785 svfloat32x4_t test_svreadz_ver_za32_f32_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 786 {
 787    return svreadz_ver_za32_f32_vg4(3, slice);
 788 }
 789
 790 // CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @test_svreadz_ver_za64_s64_x4(
 791 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 792 // CHECK-NEXT:  entry:
 793 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.vert.x4.nxv2i64(i32 0, i32 [[SLICE]])
 794 // CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 795 //
 796 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @_Z28test_svreadz_ver_za64_s64_x4j(
 797 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 798 // CPP-CHECK-NEXT:  entry:
 799 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.vert.x4.nxv2i64(i32 0, i32 [[SLICE]])
 800 // CPP-CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 801 //
 802 svint64x4_t test_svreadz_ver_za64_s64_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 803 {
 804    return svreadz_ver_za64_s64_vg4(0, slice);
 805 }
 806
 807 // CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @test_svreadz_ver_za64_u64_x4(
 808 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 809 // CHECK-NEXT:  entry:
 810 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.vert.x4.nxv2i64(i32 4, i32 [[SLICE]])
 811 // CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 812 //
 813 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @_Z28test_svreadz_ver_za64_u64_x4j(
 814 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 815 // CPP-CHECK-NEXT:  entry:
 816 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.vert.x4.nxv2i64(i32 4, i32 [[SLICE]])
 817 // CPP-CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
 818 //
 819 svuint64x4_t test_svreadz_ver_za64_u64_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 820 {
 821    return svreadz_ver_za64_u64_vg4(4, slice);
 822 }
 823
 824 // CHECK-LABEL: define dso_local { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @test_svreadz_ver_za64_f64_x4(
 825 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 826 // CHECK-NEXT:  entry:
 827 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.readz.vert.x4.nxv2f64(i32 7, i32 [[SLICE]])
 828 // CHECK-NEXT:    ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } [[TMP0]]
 829 //
 830 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @_Z28test_svreadz_ver_za64_f64_x4j(
 831 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 832 // CPP-CHECK-NEXT:  entry:
 833 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.readz.vert.x4.nxv2f64(i32 7, i32 [[SLICE]])
 834 // CPP-CHECK-NEXT:    ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } [[TMP0]]
 835 //
 836 svfloat64x4_t test_svreadz_ver_za64_f64_x4(uint32_t slice) __arm_streaming __arm_inout("za")
 837 {
 838    return svreadz_ver_za64_f64_vg4(7, slice);
 839 }
 840
 841 // CHECK-LABEL: define dso_local <vscale x 16 x i8> @test_svreadz_hor_za8_s8(
 842 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 843 // CHECK-NEXT:  entry:
 844 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sme.readz.horiz.nxv16i8(i32 0, i32 [[SLICE]])
 845 // CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
 846 //
 847 // CPP-CHECK-LABEL: define dso_local <vscale x 16 x i8> @_Z23test_svreadz_hor_za8_s8j(
 848 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 849 // CPP-CHECK-NEXT:  entry:
 850 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sme.readz.horiz.nxv16i8(i32 0, i32 [[SLICE]])
 851 // CPP-CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
 852 //
 853 svint8_t test_svreadz_hor_za8_s8(uint32_t slice) __arm_streaming __arm_inout("za")
 854 {
 855    return svreadz_hor_za8_s8(0, slice);
 856 }
 857
 858 // CHECK-LABEL: define dso_local <vscale x 16 x i8> @test_svreadz_hor_za8_u8(
 859 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 860 // CHECK-NEXT:  entry:
 861 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sme.readz.horiz.nxv16i8(i32 0, i32 [[SLICE]])
 862 // CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
 863 //
 864 // CPP-CHECK-LABEL: define dso_local <vscale x 16 x i8> @_Z23test_svreadz_hor_za8_u8j(
 865 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 866 // CPP-CHECK-NEXT:  entry:
 867 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sme.readz.horiz.nxv16i8(i32 0, i32 [[SLICE]])
 868 // CPP-CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
 869 //
 870 svuint8_t test_svreadz_hor_za8_u8(uint32_t slice) __arm_streaming __arm_inout("za")
 871 {
 872    return svreadz_hor_za8_u8(0, slice);
 873 }
 874
 875 // CHECK-LABEL: define dso_local <vscale x 8 x i16> @test_svreadz_hor_za16_s16(
 876 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 877 // CHECK-NEXT:  entry:
 878 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x i16> @llvm.aarch64.sme.readz.horiz.nxv8i16(i32 0, i32 [[SLICE]])
 879 // CHECK-NEXT:    ret <vscale x 8 x i16> [[TMP0]]
 880 //
 881 // CPP-CHECK-LABEL: define dso_local <vscale x 8 x i16> @_Z25test_svreadz_hor_za16_s16j(
 882 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 883 // CPP-CHECK-NEXT:  entry:
 884 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x i16> @llvm.aarch64.sme.readz.horiz.nxv8i16(i32 0, i32 [[SLICE]])
 885 // CPP-CHECK-NEXT:    ret <vscale x 8 x i16> [[TMP0]]
 886 //
 887 svint16_t test_svreadz_hor_za16_s16(uint32_t slice) __arm_streaming __arm_inout("za")
 888 {
 889    return svreadz_hor_za16_s16(0, slice);
 890 }
 891
 892 // CHECK-LABEL: define dso_local <vscale x 8 x i16> @test_svreadz_hor_za16_u16(
 893 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 894 // CHECK-NEXT:  entry:
 895 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x i16> @llvm.aarch64.sme.readz.horiz.nxv8i16(i32 1, i32 [[SLICE]])
 896 // CHECK-NEXT:    ret <vscale x 8 x i16> [[TMP0]]
 897 //
 898 // CPP-CHECK-LABEL: define dso_local <vscale x 8 x i16> @_Z25test_svreadz_hor_za16_u16j(
 899 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 900 // CPP-CHECK-NEXT:  entry:
 901 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x i16> @llvm.aarch64.sme.readz.horiz.nxv8i16(i32 1, i32 [[SLICE]])
 902 // CPP-CHECK-NEXT:    ret <vscale x 8 x i16> [[TMP0]]
 903 //
 904 svuint16_t test_svreadz_hor_za16_u16(uint32_t slice) __arm_streaming __arm_inout("za")
 905 {
 906    return svreadz_hor_za16_u16(1, slice);
 907 }
 908
 909 // CHECK-LABEL: define dso_local <vscale x 8 x half> @test_svreadz_hor_za16_f16(
 910 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 911 // CHECK-NEXT:  entry:
 912 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x half> @llvm.aarch64.sme.readz.horiz.nxv8f16(i32 0, i32 [[SLICE]])
 913 // CHECK-NEXT:    ret <vscale x 8 x half> [[TMP0]]
 914 //
 915 // CPP-CHECK-LABEL: define dso_local <vscale x 8 x half> @_Z25test_svreadz_hor_za16_f16j(
 916 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 917 // CPP-CHECK-NEXT:  entry:
 918 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x half> @llvm.aarch64.sme.readz.horiz.nxv8f16(i32 0, i32 [[SLICE]])
 919 // CPP-CHECK-NEXT:    ret <vscale x 8 x half> [[TMP0]]
 920 //
 921 svfloat16_t test_svreadz_hor_za16_f16(uint32_t slice) __arm_streaming __arm_inout("za")
 922 {
 923    return svreadz_hor_za16_f16(0, slice);
 924 }
 925
 926 // CHECK-LABEL: define dso_local <vscale x 8 x bfloat> @test_svreadz_hor_za16_bf16(
 927 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 928 // CHECK-NEXT:  entry:
 929 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.aarch64.sme.readz.horiz.nxv8bf16(i32 1, i32 [[SLICE]])
 930 // CHECK-NEXT:    ret <vscale x 8 x bfloat> [[TMP0]]
 931 //
 932 // CPP-CHECK-LABEL: define dso_local <vscale x 8 x bfloat> @_Z26test_svreadz_hor_za16_bf16j(
 933 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 934 // CPP-CHECK-NEXT:  entry:
 935 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.aarch64.sme.readz.horiz.nxv8bf16(i32 1, i32 [[SLICE]])
 936 // CPP-CHECK-NEXT:    ret <vscale x 8 x bfloat> [[TMP0]]
 937 //
 938 svbfloat16_t test_svreadz_hor_za16_bf16(uint32_t slice) __arm_streaming __arm_inout("za")
 939 {
 940    return svreadz_hor_za16_bf16(1, slice);
 941 }
 942
 943
 944 // CHECK-LABEL: define dso_local <vscale x 4 x i32> @test_svreadz_hor_za32_s32(
 945 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 946 // CHECK-NEXT:  entry:
 947 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x i32> @llvm.aarch64.sme.readz.horiz.nxv4i32(i32 0, i32 [[SLICE]])
 948 // CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP0]]
 949 //
 950 // CPP-CHECK-LABEL: define dso_local <vscale x 4 x i32> @_Z25test_svreadz_hor_za32_s32j(
 951 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 952 // CPP-CHECK-NEXT:  entry:
 953 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x i32> @llvm.aarch64.sme.readz.horiz.nxv4i32(i32 0, i32 [[SLICE]])
 954 // CPP-CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP0]]
 955 //
 956 svint32_t test_svreadz_hor_za32_s32(uint32_t slice) __arm_streaming __arm_inout("za")
 957 {
 958    return svreadz_hor_za32_s32(0, slice);
 959 }
 960
 961 // CHECK-LABEL: define dso_local <vscale x 4 x i32> @test_svreadz_hor_za32_u32(
 962 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 963 // CHECK-NEXT:  entry:
 964 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x i32> @llvm.aarch64.sme.readz.horiz.nxv4i32(i32 2, i32 [[SLICE]])
 965 // CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP0]]
 966 //
 967 // CPP-CHECK-LABEL: define dso_local <vscale x 4 x i32> @_Z25test_svreadz_hor_za32_u32j(
 968 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 969 // CPP-CHECK-NEXT:  entry:
 970 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x i32> @llvm.aarch64.sme.readz.horiz.nxv4i32(i32 2, i32 [[SLICE]])
 971 // CPP-CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP0]]
 972 //
 973 svuint32_t test_svreadz_hor_za32_u32(uint32_t slice) __arm_streaming __arm_inout("za")
 974 {
 975    return svreadz_hor_za32_u32(2, slice);
 976 }
 977
 978 // CHECK-LABEL: define dso_local <vscale x 4 x float> @test_svreadz_hor_za32_f32(
 979 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 980 // CHECK-NEXT:  entry:
 981 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x float> @llvm.aarch64.sme.readz.horiz.nxv4f32(i32 3, i32 [[SLICE]])
 982 // CHECK-NEXT:    ret <vscale x 4 x float> [[TMP0]]
 983 //
 984 // CPP-CHECK-LABEL: define dso_local <vscale x 4 x float> @_Z25test_svreadz_hor_za32_f32j(
 985 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 986 // CPP-CHECK-NEXT:  entry:
 987 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x float> @llvm.aarch64.sme.readz.horiz.nxv4f32(i32 3, i32 [[SLICE]])
 988 // CPP-CHECK-NEXT:    ret <vscale x 4 x float> [[TMP0]]
 989 //
 990 svfloat32_t test_svreadz_hor_za32_f32(uint32_t slice) __arm_streaming __arm_inout("za")
 991 {
 992    return svreadz_hor_za32_f32(3, slice);
 993 }
 994
 995 // CHECK-LABEL: define dso_local <vscale x 2 x i64> @test_svreadz_hor_za64_s64(
 996 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
 997 // CHECK-NEXT:  entry:
 998 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x i64> @llvm.aarch64.sme.readz.horiz.nxv2i64(i32 0, i32 [[SLICE]])
 999 // CHECK-NEXT:    ret <vscale x 2 x i64> [[TMP0]]
1000 //
1001 // CPP-CHECK-LABEL: define dso_local <vscale x 2 x i64> @_Z25test_svreadz_hor_za64_s64j(
1002 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1003 // CPP-CHECK-NEXT:  entry:
1004 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x i64> @llvm.aarch64.sme.readz.horiz.nxv2i64(i32 0, i32 [[SLICE]])
1005 // CPP-CHECK-NEXT:    ret <vscale x 2 x i64> [[TMP0]]
1006 //
1007 svint64_t test_svreadz_hor_za64_s64(uint32_t slice) __arm_streaming __arm_inout("za")
1008 {
1009    return svreadz_hor_za64_s64(0, slice);
1010 }
1011
1012 // CHECK-LABEL: define dso_local <vscale x 2 x i64> @test_svreadz_hor_za64_u64(
1013 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1014 // CHECK-NEXT:  entry:
1015 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x i64> @llvm.aarch64.sme.readz.horiz.nxv2i64(i32 4, i32 [[SLICE]])
1016 // CHECK-NEXT:    ret <vscale x 2 x i64> [[TMP0]]
1017 //
1018 // CPP-CHECK-LABEL: define dso_local <vscale x 2 x i64> @_Z25test_svreadz_hor_za64_u64j(
1019 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1020 // CPP-CHECK-NEXT:  entry:
1021 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x i64> @llvm.aarch64.sme.readz.horiz.nxv2i64(i32 4, i32 [[SLICE]])
1022 // CPP-CHECK-NEXT:    ret <vscale x 2 x i64> [[TMP0]]
1023 //
1024 svuint64_t test_svreadz_hor_za64_u64(uint32_t slice) __arm_streaming __arm_inout("za")
1025 {
1026    return svreadz_hor_za64_u64(4, slice);
1027 }
1028
1029 // CHECK-LABEL: define dso_local <vscale x 2 x double> @test_svreadz_hor_za64_f64(
1030 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1031 // CHECK-NEXT:  entry:
1032 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x double> @llvm.aarch64.sme.readz.horiz.nxv2f64(i32 7, i32 [[SLICE]])
1033 // CHECK-NEXT:    ret <vscale x 2 x double> [[TMP0]]
1034 //
1035 // CPP-CHECK-LABEL: define dso_local <vscale x 2 x double> @_Z25test_svreadz_hor_za64_f64j(
1036 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1037 // CPP-CHECK-NEXT:  entry:
1038 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x double> @llvm.aarch64.sme.readz.horiz.nxv2f64(i32 7, i32 [[SLICE]])
1039 // CPP-CHECK-NEXT:    ret <vscale x 2 x double> [[TMP0]]
1040 //
1041 svfloat64_t test_svreadz_hor_za64_f64(uint32_t slice) __arm_streaming __arm_inout("za")
1042 {
1043    return svreadz_hor_za64_f64(7, slice);
1044 }
1045
1046 // ZA128
1047 // CHECK-LABEL: define dso_local <vscale x 16 x i8> @test_svreadz_hor_za128_s8(
1048 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1049 // CHECK-NEXT:  entry:
1050 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sme.readz.q.horiz.nxv16i8(i32 0, i32 [[SLICE]])
1051 // CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
1052 //
1053 // CPP-CHECK-LABEL: define dso_local <vscale x 16 x i8> @_Z25test_svreadz_hor_za128_s8j(
1054 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1055 // CPP-CHECK-NEXT:  entry:
1056 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sme.readz.q.horiz.nxv16i8(i32 0, i32 [[SLICE]])
1057 // CPP-CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
1058 //
1059 svint8_t test_svreadz_hor_za128_s8(uint32_t slice) __arm_streaming __arm_inout("za")
1060 {
1061    return svreadz_hor_za128_s8(0, slice);
1062 }
1063
1064 // CHECK-LABEL: define dso_local <vscale x 16 x i8> @test_svreadz_hor_za128_u8(
1065 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1066 // CHECK-NEXT:  entry:
1067 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sme.readz.q.horiz.nxv16i8(i32 1, i32 [[SLICE]])
1068 // CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
1069 //
1070 // CPP-CHECK-LABEL: define dso_local <vscale x 16 x i8> @_Z25test_svreadz_hor_za128_u8j(
1071 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1072 // CPP-CHECK-NEXT:  entry:
1073 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sme.readz.q.horiz.nxv16i8(i32 1, i32 [[SLICE]])
1074 // CPP-CHECK-NEXT:    ret <vscale x 16 x i8> [[TMP0]]
1075 //
1076 svuint8_t test_svreadz_hor_za128_u8(uint32_t slice) __arm_streaming __arm_inout("za")
1077 {
1078    return svreadz_hor_za128_u8(1, slice);
1079 }
1080
1081 // CHECK-LABEL: define dso_local <vscale x 8 x i16> @test_svreadz_hor_za128_s16(
1082 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1083 // CHECK-NEXT:  entry:
1084 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x i16> @llvm.aarch64.sme.readz.q.horiz.nxv8i16(i32 2, i32 [[SLICE]])
1085 // CHECK-NEXT:    ret <vscale x 8 x i16> [[TMP0]]
1086 //
1087 // CPP-CHECK-LABEL: define dso_local <vscale x 8 x i16> @_Z26test_svreadz_hor_za128_s16j(
1088 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1089 // CPP-CHECK-NEXT:  entry:
1090 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x i16> @llvm.aarch64.sme.readz.q.horiz.nxv8i16(i32 2, i32 [[SLICE]])
1091 // CPP-CHECK-NEXT:    ret <vscale x 8 x i16> [[TMP0]]
1092 //
1093 svint16_t test_svreadz_hor_za128_s16(uint32_t slice) __arm_streaming __arm_inout("za")
1094 {
1095    return svreadz_hor_za128_s16(2, slice);
1096 }
1097
1098 // CHECK-LABEL: define dso_local <vscale x 8 x i16> @test_svreadz_hor_za128_u16(
1099 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1100 // CHECK-NEXT:  entry:
1101 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x i16> @llvm.aarch64.sme.readz.q.horiz.nxv8i16(i32 3, i32 [[SLICE]])
1102 // CHECK-NEXT:    ret <vscale x 8 x i16> [[TMP0]]
1103 //
1104 // CPP-CHECK-LABEL: define dso_local <vscale x 8 x i16> @_Z26test_svreadz_hor_za128_u16j(
1105 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1106 // CPP-CHECK-NEXT:  entry:
1107 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x i16> @llvm.aarch64.sme.readz.q.horiz.nxv8i16(i32 3, i32 [[SLICE]])
1108 // CPP-CHECK-NEXT:    ret <vscale x 8 x i16> [[TMP0]]
1109 //
1110 svuint16_t test_svreadz_hor_za128_u16(uint32_t slice) __arm_streaming __arm_inout("za")
1111 {
1112    return svreadz_hor_za128_u16(3, slice);
1113 }
1114
1115 // CHECK-LABEL: define dso_local <vscale x 8 x half> @test_svreadz_hor_za128_f16(
1116 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1117 // CHECK-NEXT:  entry:
1118 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x half> @llvm.aarch64.sme.readz.q.horiz.nxv8f16(i32 4, i32 [[SLICE]])
1119 // CHECK-NEXT:    ret <vscale x 8 x half> [[TMP0]]
1120 //
1121 // CPP-CHECK-LABEL: define dso_local <vscale x 8 x half> @_Z26test_svreadz_hor_za128_f16j(
1122 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1123 // CPP-CHECK-NEXT:  entry:
1124 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x half> @llvm.aarch64.sme.readz.q.horiz.nxv8f16(i32 4, i32 [[SLICE]])
1125 // CPP-CHECK-NEXT:    ret <vscale x 8 x half> [[TMP0]]
1126 //
1127 svfloat16_t test_svreadz_hor_za128_f16(uint32_t slice) __arm_streaming __arm_inout("za")
1128 {
1129    return svreadz_hor_za128_f16(4, slice);
1130 }
1131
1132 // CHECK-LABEL: define dso_local <vscale x 8 x bfloat> @test_svreadz_hor_za128_bf16(
1133 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1134 // CHECK-NEXT:  entry:
1135 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.aarch64.sme.readz.q.horiz.nxv8bf16(i32 5, i32 [[SLICE]])
1136 // CHECK-NEXT:    ret <vscale x 8 x bfloat> [[TMP0]]
1137 //
1138 // CPP-CHECK-LABEL: define dso_local <vscale x 8 x bfloat> @_Z27test_svreadz_hor_za128_bf16j(
1139 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1140 // CPP-CHECK-NEXT:  entry:
1141 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 8 x bfloat> @llvm.aarch64.sme.readz.q.horiz.nxv8bf16(i32 5, i32 [[SLICE]])
1142 // CPP-CHECK-NEXT:    ret <vscale x 8 x bfloat> [[TMP0]]
1143 //
1144 svbfloat16_t test_svreadz_hor_za128_bf16(uint32_t slice) __arm_streaming __arm_inout("za")
1145 {
1146    return svreadz_hor_za128_bf16(5, slice);
1147 }
1148
1149
1150 // CHECK-LABEL: define dso_local <vscale x 4 x i32> @test_svreadz_hor_za128_s32(
1151 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1152 // CHECK-NEXT:  entry:
1153 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x i32> @llvm.aarch64.sme.readz.q.horiz.nxv4i32(i32 6, i32 [[SLICE]])
1154 // CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP0]]
1155 //
1156 // CPP-CHECK-LABEL: define dso_local <vscale x 4 x i32> @_Z26test_svreadz_hor_za128_s32j(
1157 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1158 // CPP-CHECK-NEXT:  entry:
1159 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x i32> @llvm.aarch64.sme.readz.q.horiz.nxv4i32(i32 6, i32 [[SLICE]])
1160 // CPP-CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP0]]
1161 //
1162 svint32_t test_svreadz_hor_za128_s32(uint32_t slice) __arm_streaming __arm_inout("za")
1163 {
1164    return svreadz_hor_za128_s32(6, slice);
1165 }
1166
1167 // CHECK-LABEL: define dso_local <vscale x 4 x i32> @test_svreadz_hor_za128_u32(
1168 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1169 // CHECK-NEXT:  entry:
1170 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x i32> @llvm.aarch64.sme.readz.q.horiz.nxv4i32(i32 7, i32 [[SLICE]])
1171 // CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP0]]
1172 //
1173 // CPP-CHECK-LABEL: define dso_local <vscale x 4 x i32> @_Z26test_svreadz_hor_za128_u32j(
1174 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1175 // CPP-CHECK-NEXT:  entry:
1176 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x i32> @llvm.aarch64.sme.readz.q.horiz.nxv4i32(i32 7, i32 [[SLICE]])
1177 // CPP-CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP0]]
1178 //
1179 svuint32_t test_svreadz_hor_za128_u32(uint32_t slice) __arm_streaming __arm_inout("za")
1180 {
1181    return svreadz_hor_za128_u32(7, slice);
1182 }
1183
1184 // CHECK-LABEL: define dso_local <vscale x 4 x float> @test_svreadz_hor_za128_f32(
1185 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1186 // CHECK-NEXT:  entry:
1187 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x float> @llvm.aarch64.sme.readz.q.horiz.nxv4f32(i32 8, i32 [[SLICE]])
1188 // CHECK-NEXT:    ret <vscale x 4 x float> [[TMP0]]
1189 //
1190 // CPP-CHECK-LABEL: define dso_local <vscale x 4 x float> @_Z26test_svreadz_hor_za128_f32j(
1191 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1192 // CPP-CHECK-NEXT:  entry:
1193 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 4 x float> @llvm.aarch64.sme.readz.q.horiz.nxv4f32(i32 8, i32 [[SLICE]])
1194 // CPP-CHECK-NEXT:    ret <vscale x 4 x float> [[TMP0]]
1195 //
1196 svfloat32_t test_svreadz_hor_za128_f32(uint32_t slice) __arm_streaming __arm_inout("za")
1197 {
1198    return svreadz_hor_za128_f32(8, slice);
1199 }
1200
1201 // CHECK-LABEL: define dso_local <vscale x 2 x i64> @test_svreadz_hor_za128_s64(
1202 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1203 // CHECK-NEXT:  entry:
1204 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x i64> @llvm.aarch64.sme.readz.q.horiz.nxv2i64(i32 13, i32 [[SLICE]])
1205 // CHECK-NEXT:    ret <vscale x 2 x i64> [[TMP0]]
1206 //
1207 // CPP-CHECK-LABEL: define dso_local <vscale x 2 x i64> @_Z26test_svreadz_hor_za128_s64j(
1208 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1209 // CPP-CHECK-NEXT:  entry:
1210 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x i64> @llvm.aarch64.sme.readz.q.horiz.nxv2i64(i32 13, i32 [[SLICE]])
1211 // CPP-CHECK-NEXT:    ret <vscale x 2 x i64> [[TMP0]]
1212 //
1213 svint64_t test_svreadz_hor_za128_s64(uint32_t slice) __arm_streaming __arm_inout("za")
1214 {
1215    return svreadz_hor_za128_s64(13, slice);
1216 }
1217 // CHECK-LABEL: define dso_local <vscale x 2 x i64> @test_svreadz_hor_za128_u64(
1218 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1219 // CHECK-NEXT:  entry:
1220 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x i64> @llvm.aarch64.sme.readz.q.horiz.nxv2i64(i32 14, i32 [[SLICE]])
1221 // CHECK-NEXT:    ret <vscale x 2 x i64> [[TMP0]]
1222 //
1223 // CPP-CHECK-LABEL: define dso_local <vscale x 2 x i64> @_Z26test_svreadz_hor_za128_u64j(
1224 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1225 // CPP-CHECK-NEXT:  entry:
1226 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x i64> @llvm.aarch64.sme.readz.q.horiz.nxv2i64(i32 14, i32 [[SLICE]])
1227 // CPP-CHECK-NEXT:    ret <vscale x 2 x i64> [[TMP0]]
1228 //
1229 svuint64_t test_svreadz_hor_za128_u64(uint32_t slice) __arm_streaming __arm_inout("za")
1230 {
1231    return svreadz_hor_za128_u64(14, slice);
1232 }
1233
1234 // CHECK-LABEL: define dso_local <vscale x 2 x double> @test_svreadz_hor_za128_f64(
1235 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1236 // CHECK-NEXT:  entry:
1237 // CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x double> @llvm.aarch64.sme.readz.q.horiz.nxv2f64(i32 15, i32 [[SLICE]])
1238 // CHECK-NEXT:    ret <vscale x 2 x double> [[TMP0]]
1239 //
1240 // CPP-CHECK-LABEL: define dso_local <vscale x 2 x double> @_Z26test_svreadz_hor_za128_f64j(
1241 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1242 // CPP-CHECK-NEXT:  entry:
1243 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call <vscale x 2 x double> @llvm.aarch64.sme.readz.q.horiz.nxv2f64(i32 15, i32 [[SLICE]])
1244 // CPP-CHECK-NEXT:    ret <vscale x 2 x double> [[TMP0]]
1245 //
1246 svfloat64_t test_svreadz_hor_za128_f64(uint32_t slice) __arm_streaming __arm_inout("za")
1247 {
1248    return svreadz_hor_za128_f64(15, slice);
1249 }
1250
1251 // MOVAZ Array to Tile
1252 //
1253 // X2
1254 //
1255
1256 // CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8> } @test_svreadz_za8_s8_x2(
1257 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1258 // CHECK-NEXT:  entry:
1259 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.x2.nxv16i8(i32 [[SLICE]])
1260 // CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
1261 //
1262 // CPP-CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8> } @_Z22test_svreadz_za8_s8_x2j(
1263 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1264 // CPP-CHECK-NEXT:  entry:
1265 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.x2.nxv16i8(i32 [[SLICE]])
1266 // CPP-CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
1267 //
1268 svint8x2_t test_svreadz_za8_s8_x2(uint32_t slice) __arm_streaming __arm_inout("za")
1269 {
1270    return svreadz_za8_s8_vg1x2(slice);
1271 }
1272
1273 // CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8> } @test_svreadz_za8_u8_x2(
1274 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1275 // CHECK-NEXT:  entry:
1276 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.x2.nxv16i8(i32 [[SLICE]])
1277 // CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
1278 //
1279 // CPP-CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8> } @_Z22test_svreadz_za8_u8_x2j(
1280 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1281 // CPP-CHECK-NEXT:  entry:
1282 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.x2.nxv16i8(i32 [[SLICE]])
1283 // CPP-CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
1284 //
1285 svuint8x2_t test_svreadz_za8_u8_x2(uint32_t slice) __arm_streaming __arm_inout("za")
1286 {
1287    return svreadz_za8_u8_vg1x2(slice);
1288 }
1289
1290 // CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16> } @test_svreadz_za16_s16_x2(
1291 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1292 // CHECK-NEXT:  entry:
1293 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.x2.nxv8i16(i32 [[SLICE]])
1294 // CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
1295 //
1296 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16> } @_Z24test_svreadz_za16_s16_x2j(
1297 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1298 // CPP-CHECK-NEXT:  entry:
1299 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.x2.nxv8i16(i32 [[SLICE]])
1300 // CPP-CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
1301 //
1302 svint16x2_t test_svreadz_za16_s16_x2(uint32_t slice) __arm_streaming __arm_inout("za")
1303 {
1304    return svreadz_za16_s16_vg1x2(slice);
1305 }
1306
1307 // CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16> } @test_svreadz_za16_u16_x2(
1308 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1309 // CHECK-NEXT:  entry:
1310 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.x2.nxv8i16(i32 [[SLICE]])
1311 // CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
1312 //
1313 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16> } @_Z24test_svreadz_za16_u16_x2j(
1314 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1315 // CPP-CHECK-NEXT:  entry:
1316 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.x2.nxv8i16(i32 [[SLICE]])
1317 // CPP-CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
1318 //
1319 svuint16x2_t test_svreadz_za16_u16_x2(uint32_t slice) __arm_streaming __arm_inout("za")
1320 {
1321    return svreadz_za16_u16_vg1x2(slice);
1322 }
1323
1324 // CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32> } @test_svreadz_za32_s32_x2(
1325 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1326 // CHECK-NEXT:  entry:
1327 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.x2.nxv4i32(i32 [[SLICE]])
1328 // CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
1329 //
1330 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32> } @_Z24test_svreadz_za32_s32_x2j(
1331 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1332 // CPP-CHECK-NEXT:  entry:
1333 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.x2.nxv4i32(i32 [[SLICE]])
1334 // CPP-CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
1335 //
1336 svint32x2_t test_svreadz_za32_s32_x2(uint32_t slice) __arm_streaming __arm_inout("za")
1337 {
1338    return svreadz_za32_s32_vg1x2(slice);
1339 }
1340
1341 // CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32> } @test_svreadz_za32_u32_x2(
1342 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1343 // CHECK-NEXT:  entry:
1344 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.x2.nxv4i32(i32 [[SLICE]])
1345 // CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
1346 //
1347 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32> } @_Z24test_svreadz_za32_u32_x2j(
1348 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1349 // CPP-CHECK-NEXT:  entry:
1350 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.x2.nxv4i32(i32 [[SLICE]])
1351 // CPP-CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
1352 //
1353 svuint32x2_t test_svreadz_za32_u32_x2(uint32_t slice) __arm_streaming __arm_inout("za")
1354 {
1355    return svreadz_za32_u32_vg1x2(slice);
1356 }
1357
1358 // CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64> } @test_svreadz_za64_s64_x2(
1359 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1360 // CHECK-NEXT:  entry:
1361 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.x2.nxv2i64(i32 [[SLICE]])
1362 // CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
1363 //
1364 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64> } @_Z24test_svreadz_za64_s64_x2j(
1365 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1366 // CPP-CHECK-NEXT:  entry:
1367 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.x2.nxv2i64(i32 [[SLICE]])
1368 // CPP-CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
1369 //
1370 svint64x2_t test_svreadz_za64_s64_x2(uint32_t slice) __arm_streaming __arm_inout("za")
1371 {
1372    return svreadz_za64_s64_vg1x2(slice);
1373 }
1374
1375 // CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64> } @test_svreadz_za64_u64_x2(
1376 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1377 // CHECK-NEXT:  entry:
1378 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.x2.nxv2i64(i32 [[SLICE]])
1379 // CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
1380 //
1381 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64> } @_Z24test_svreadz_za64_u64_x2j(
1382 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1383 // CPP-CHECK-NEXT:  entry:
1384 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.x2.nxv2i64(i32 [[SLICE]])
1385 // CPP-CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
1386 //
1387 svuint64x2_t test_svreadz_za64_u64_x2(uint32_t slice) __arm_streaming __arm_inout("za")
1388 {
1389    return svreadz_za64_u64_vg1x2(slice);
1390 }
1391
1392 // CHECK-LABEL: define dso_local { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @test_svreadz_za16_bf16_x2(
1393 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1394 // CHECK-NEXT:  entry:
1395 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @llvm.aarch64.sme.readz.x2.nxv8bf16(i32 [[SLICE]])
1396 // CHECK-NEXT:    ret { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } [[TMP0]]
1397 //
1398 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @_Z25test_svreadz_za16_bf16_x2j(
1399 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1400 // CPP-CHECK-NEXT:  entry:
1401 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @llvm.aarch64.sme.readz.x2.nxv8bf16(i32 [[SLICE]])
1402 // CPP-CHECK-NEXT:    ret { <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } [[TMP0]]
1403 //
1404 svbfloat16x2_t test_svreadz_za16_bf16_x2(uint32_t slice) __arm_streaming __arm_inout("za")
1405 {
1406    return svreadz_za16_bf16_vg1x2(slice);
1407 }
1408
1409 // CHECK-LABEL: define dso_local { <vscale x 8 x half>, <vscale x 8 x half> } @test_svreadz_za16_f16_x2(
1410 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1411 // CHECK-NEXT:  entry:
1412 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.readz.x2.nxv8f16(i32 [[SLICE]])
1413 // CHECK-NEXT:    ret { <vscale x 8 x half>, <vscale x 8 x half> } [[TMP0]]
1414 //
1415 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x half>, <vscale x 8 x half> } @_Z24test_svreadz_za16_f16_x2j(
1416 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1417 // CPP-CHECK-NEXT:  entry:
1418 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.readz.x2.nxv8f16(i32 [[SLICE]])
1419 // CPP-CHECK-NEXT:    ret { <vscale x 8 x half>, <vscale x 8 x half> } [[TMP0]]
1420 //
1421 svfloat16x2_t test_svreadz_za16_f16_x2(uint32_t slice) __arm_streaming __arm_inout("za")
1422 {
1423    return svreadz_za16_f16_vg1x2(slice);
1424 }
1425
1426 // CHECK-LABEL: define dso_local { <vscale x 4 x float>, <vscale x 4 x float> } @test_svreadz_za32_f32_x2(
1427 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1428 // CHECK-NEXT:  entry:
1429 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.readz.x2.nxv4f32(i32 [[SLICE]])
1430 // CHECK-NEXT:    ret { <vscale x 4 x float>, <vscale x 4 x float> } [[TMP0]]
1431 //
1432 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x float>, <vscale x 4 x float> } @_Z24test_svreadz_za32_f32_x2j(
1433 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1434 // CPP-CHECK-NEXT:  entry:
1435 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.readz.x2.nxv4f32(i32 [[SLICE]])
1436 // CPP-CHECK-NEXT:    ret { <vscale x 4 x float>, <vscale x 4 x float> } [[TMP0]]
1437 //
1438 svfloat32x2_t test_svreadz_za32_f32_x2(uint32_t slice) __arm_streaming __arm_inout("za")
1439 {
1440    return svreadz_za32_f32_vg1x2(slice);
1441 }
1442
1443 // CHECK-LABEL: define dso_local { <vscale x 2 x double>, <vscale x 2 x double> } @test_svreadz_za64_f64_x2(
1444 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1445 // CHECK-NEXT:  entry:
1446 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.readz.x2.nxv2f64(i32 [[SLICE]])
1447 // CHECK-NEXT:    ret { <vscale x 2 x double>, <vscale x 2 x double> } [[TMP0]]
1448 //
1449 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x double>, <vscale x 2 x double> } @_Z24test_svreadz_za64_f64_x2j(
1450 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1451 // CPP-CHECK-NEXT:  entry:
1452 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.readz.x2.nxv2f64(i32 [[SLICE]])
1453 // CPP-CHECK-NEXT:    ret { <vscale x 2 x double>, <vscale x 2 x double> } [[TMP0]]
1454 //
1455 svfloat64x2_t test_svreadz_za64_f64_x2(uint32_t slice) __arm_streaming __arm_inout("za")
1456 {
1457    return svreadz_za64_f64_vg1x2(slice);
1458 }
1459
1460 //
1461 // X4
1462 //
1463
1464 // CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @test_svreadz_za8_s8_x4(
1465 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1466 // CHECK-NEXT:  entry:
1467 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.x4.nxv16i8(i32 [[SLICE]])
1468 // CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
1469 //
1470 // CPP-CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @_Z22test_svreadz_za8_s8_x4j(
1471 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1472 // CPP-CHECK-NEXT:  entry:
1473 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.x4.nxv16i8(i32 [[SLICE]])
1474 // CPP-CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
1475 //
1476 svint8x4_t test_svreadz_za8_s8_x4(uint32_t slice) __arm_streaming __arm_inout("za")
1477 {
1478    return svreadz_za8_s8_vg1x4(slice);
1479 }
1480
1481 // CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @test_svreadz_za8_u8_x4(
1482 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1483 // CHECK-NEXT:  entry:
1484 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.x4.nxv16i8(i32 [[SLICE]])
1485 // CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
1486 //
1487 // CPP-CHECK-LABEL: define dso_local { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @_Z22test_svreadz_za8_u8_x4j(
1488 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1489 // CPP-CHECK-NEXT:  entry:
1490 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } @llvm.aarch64.sme.readz.x4.nxv16i8(i32 [[SLICE]])
1491 // CPP-CHECK-NEXT:    ret { <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8> } [[TMP0]]
1492 //
1493 svuint8x4_t test_svreadz_za8_u8_x4(uint32_t slice) __arm_streaming __arm_inout("za")
1494 {
1495    return svreadz_za8_u8_vg1x4(slice);
1496 }
1497
1498 // CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @test_svreadz_za16_s16_x4(
1499 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1500 // CHECK-NEXT:  entry:
1501 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.x4.nxv8i16(i32 [[SLICE]])
1502 // CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
1503 //
1504 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @_Z24test_svreadz_za16_s16_x4j(
1505 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1506 // CPP-CHECK-NEXT:  entry:
1507 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.x4.nxv8i16(i32 [[SLICE]])
1508 // CPP-CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
1509 //
1510 svint16x4_t test_svreadz_za16_s16_x4(uint32_t slice) __arm_streaming __arm_inout("za")
1511 {
1512    return svreadz_za16_s16_vg1x4(slice);
1513 }
1514
1515 // CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @test_svreadz_za16_u16_x4(
1516 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1517 // CHECK-NEXT:  entry:
1518 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.x4.nxv8i16(i32 [[SLICE]])
1519 // CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
1520 //
1521 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @_Z24test_svreadz_za16_u16_x4j(
1522 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1523 // CPP-CHECK-NEXT:  entry:
1524 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } @llvm.aarch64.sme.readz.x4.nxv8i16(i32 [[SLICE]])
1525 // CPP-CHECK-NEXT:    ret { <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16> } [[TMP0]]
1526 //
1527 svuint16x4_t test_svreadz_za16_u16_x4(uint32_t slice) __arm_streaming __arm_inout("za")
1528 {
1529    return svreadz_za16_u16_vg1x4(slice);
1530 }
1531
1532 // CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @test_svreadz_za32_s32_x4(
1533 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1534 // CHECK-NEXT:  entry:
1535 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.x4.nxv4i32(i32 [[SLICE]])
1536 // CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
1537 //
1538 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @_Z24test_svreadz_za32_s32_x4j(
1539 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1540 // CPP-CHECK-NEXT:  entry:
1541 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.x4.nxv4i32(i32 [[SLICE]])
1542 // CPP-CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
1543 //
1544 svint32x4_t test_svreadz_za32_s32_x4(uint32_t slice) __arm_streaming __arm_inout("za")
1545 {
1546    return svreadz_za32_s32_vg1x4(slice);
1547 }
1548
1549 // CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @test_svreadz_za32_u32_x4(
1550 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1551 // CHECK-NEXT:  entry:
1552 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.x4.nxv4i32(i32 [[SLICE]])
1553 // CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
1554 //
1555 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @_Z24test_svreadz_za32_u32_x4j(
1556 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1557 // CPP-CHECK-NEXT:  entry:
1558 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } @llvm.aarch64.sme.readz.x4.nxv4i32(i32 [[SLICE]])
1559 // CPP-CHECK-NEXT:    ret { <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32> } [[TMP0]]
1560 //
1561 svuint32x4_t test_svreadz_za32_u32_x4(uint32_t slice) __arm_streaming __arm_inout("za")
1562 {
1563    return svreadz_za32_u32_vg1x4(slice);
1564 }
1565
1566 // CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @test_svreadz_za64_s64_x4(
1567 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1568 // CHECK-NEXT:  entry:
1569 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.x4.nxv2i64(i32 [[SLICE]])
1570 // CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
1571 //
1572 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @_Z24test_svreadz_za64_s64_x4j(
1573 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1574 // CPP-CHECK-NEXT:  entry:
1575 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.x4.nxv2i64(i32 [[SLICE]])
1576 // CPP-CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
1577 //
1578 svint64x4_t test_svreadz_za64_s64_x4(uint32_t slice) __arm_streaming __arm_inout("za")
1579 {
1580    return svreadz_za64_s64_vg1x4(slice);
1581 }
1582
1583 // CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @test_svreadz_za64_u64_x4(
1584 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1585 // CHECK-NEXT:  entry:
1586 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.x4.nxv2i64(i32 [[SLICE]])
1587 // CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
1588 //
1589 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @_Z24test_svreadz_za64_u64_x4j(
1590 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1591 // CPP-CHECK-NEXT:  entry:
1592 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } @llvm.aarch64.sme.readz.x4.nxv2i64(i32 [[SLICE]])
1593 // CPP-CHECK-NEXT:    ret { <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64> } [[TMP0]]
1594 //
1595 svuint64x4_t test_svreadz_za64_u64_x4(uint32_t slice) __arm_streaming __arm_inout("za")
1596 {
1597    return svreadz_za64_u64_vg1x4(slice);
1598 }
1599
1600 // CHECK-LABEL: define dso_local { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @test_svreadz_za16_bf16_x4(
1601 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1602 // CHECK-NEXT:  entry:
1603 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @llvm.aarch64.sme.readz.x4.nxv8bf16(i32 [[SLICE]])
1604 // CHECK-NEXT:    ret { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } [[TMP0]]
1605 //
1606 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @_Z25test_svreadz_za16_bf16_x4j(
1607 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1608 // CPP-CHECK-NEXT:  entry:
1609 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } @llvm.aarch64.sme.readz.x4.nxv8bf16(i32 [[SLICE]])
1610 // CPP-CHECK-NEXT:    ret { <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat>, <vscale x 8 x bfloat> } [[TMP0]]
1611 //
1612 svbfloat16x4_t test_svreadz_za16_bf16_x4(uint32_t slice) __arm_streaming __arm_inout("za")
1613 {
1614    return svreadz_za16_bf16_vg1x4(slice);
1615 }
1616
1617 // CHECK-LABEL: define dso_local { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @test_svreadz_za16_f16_x4(
1618 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1619 // CHECK-NEXT:  entry:
1620 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.readz.x4.nxv8f16(i32 [[SLICE]])
1621 // CHECK-NEXT:    ret { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } [[TMP0]]
1622 //
1623 // CPP-CHECK-LABEL: define dso_local { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @_Z24test_svreadz_za16_f16_x4j(
1624 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1625 // CPP-CHECK-NEXT:  entry:
1626 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } @llvm.aarch64.sme.readz.x4.nxv8f16(i32 [[SLICE]])
1627 // CPP-CHECK-NEXT:    ret { <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half> } [[TMP0]]
1628 //
1629 svfloat16x4_t test_svreadz_za16_f16_x4(uint32_t slice) __arm_streaming __arm_inout("za")
1630 {
1631    return svreadz_za16_f16_vg1x4(slice);
1632 }
1633
1634 // CHECK-LABEL: define dso_local { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @test_svreadz_za32_f32_x4(
1635 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1636 // CHECK-NEXT:  entry:
1637 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.readz.x4.nxv4f32(i32 [[SLICE]])
1638 // CHECK-NEXT:    ret { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } [[TMP0]]
1639 //
1640 // CPP-CHECK-LABEL: define dso_local { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @_Z24test_svreadz_za32_f32_x4j(
1641 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1642 // CPP-CHECK-NEXT:  entry:
1643 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } @llvm.aarch64.sme.readz.x4.nxv4f32(i32 [[SLICE]])
1644 // CPP-CHECK-NEXT:    ret { <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float> } [[TMP0]]
1645 //
1646 svfloat32x4_t test_svreadz_za32_f32_x4(uint32_t slice) __arm_streaming __arm_inout("za")
1647 {
1648    return svreadz_za32_f32_vg1x4(slice);
1649 }
1650
1651 // CHECK-LABEL: define dso_local { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @test_svreadz_za64_f64_x4(
1652 // CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1653 // CHECK-NEXT:  entry:
1654 // CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.readz.x4.nxv2f64(i32 [[SLICE]])
1655 // CHECK-NEXT:    ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } [[TMP0]]
1656 //
1657 // CPP-CHECK-LABEL: define dso_local { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @_Z24test_svreadz_za64_f64_x4j(
1658 // CPP-CHECK-SAME: i32 noundef [[SLICE:%.*]]) #[[ATTR0]] {
1659 // CPP-CHECK-NEXT:  entry:
1660 // CPP-CHECK-NEXT:    [[TMP0:%.*]] = tail call { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } @llvm.aarch64.sme.readz.x4.nxv2f64(i32 [[SLICE]])
1661 // CPP-CHECK-NEXT:    ret { <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double> } [[TMP0]]
1662 //
1663 svfloat64x4_t test_svreadz_za64_f64_x4(uint32_t slice) __arm_streaming __arm_inout("za")
1664 {
1665    return svreadz_za64_f64_vg1x4(slice);
1666 }