|
| 1 | +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4 |
| 2 | +; RUN: llc -mtriple=aarch64 %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-SD |
| 3 | +; RUN: llc -mtriple=aarch64 -global-isel -global-isel-abort=2 %s -o - 2>&1 | FileCheck %s --check-prefixes=CHECK,CHECK-GI |
| 4 | + |
| 5 | +; CHECK-GI: warning: Instruction selection used fallback path for store_v4i8 |
| 6 | + |
| 7 | +; ===== Legal Scalars ===== |
| 8 | +define void @store_i8(i8 %a, ptr %ptr){ |
| 9 | +; CHECK-LABEL: store_i8: |
| 10 | +; CHECK: // %bb.0: |
| 11 | +; CHECK-NEXT: strb w0, [x1] |
| 12 | +; CHECK-NEXT: ret |
| 13 | + store i8 %a, ptr %ptr |
| 14 | + ret void |
| 15 | +} |
| 16 | + |
| 17 | +define void @store_i16(i16 %a, ptr %ptr){ |
| 18 | +; CHECK-LABEL: store_i16: |
| 19 | +; CHECK: // %bb.0: |
| 20 | +; CHECK-NEXT: strh w0, [x1] |
| 21 | +; CHECK-NEXT: ret |
| 22 | + store i16 %a, ptr %ptr |
| 23 | + ret void |
| 24 | +} |
| 25 | + |
| 26 | +define void @store_i32(i32 %a, ptr %ptr){ |
| 27 | +; CHECK-LABEL: store_i32: |
| 28 | +; CHECK: // %bb.0: |
| 29 | +; CHECK-NEXT: str w0, [x1] |
| 30 | +; CHECK-NEXT: ret |
| 31 | + store i32 %a, ptr %ptr |
| 32 | + ret void |
| 33 | +} |
| 34 | + |
| 35 | +define void @store_i64(i64 %a, ptr %ptr){ |
| 36 | +; CHECK-LABEL: store_i64: |
| 37 | +; CHECK: // %bb.0: |
| 38 | +; CHECK-NEXT: str x0, [x1] |
| 39 | +; CHECK-NEXT: ret |
| 40 | + store i64 %a, ptr %ptr |
| 41 | + ret void |
| 42 | +} |
| 43 | + |
| 44 | +; ===== Legal Vector Types ===== |
| 45 | + |
| 46 | +define void @store_v8i8(<8 x i8> %a, ptr %ptr){ |
| 47 | +; CHECK-LABEL: store_v8i8: |
| 48 | +; CHECK: // %bb.0: |
| 49 | +; CHECK-NEXT: str d0, [x0] |
| 50 | +; CHECK-NEXT: ret |
| 51 | + store <8 x i8> %a, ptr %ptr |
| 52 | + ret void |
| 53 | +} |
| 54 | + |
| 55 | +define void @store_v16i8(<16 x i8> %a, ptr %ptr){ |
| 56 | +; CHECK-LABEL: store_v16i8: |
| 57 | +; CHECK: // %bb.0: |
| 58 | +; CHECK-NEXT: str q0, [x0] |
| 59 | +; CHECK-NEXT: ret |
| 60 | + store <16 x i8> %a, ptr %ptr |
| 61 | + ret void |
| 62 | +} |
| 63 | + |
| 64 | +define void @store_v4i16(<4 x i16> %a, ptr %ptr){ |
| 65 | +; CHECK-LABEL: store_v4i16: |
| 66 | +; CHECK: // %bb.0: |
| 67 | +; CHECK-NEXT: str d0, [x0] |
| 68 | +; CHECK-NEXT: ret |
| 69 | + store <4 x i16> %a, ptr %ptr |
| 70 | + ret void |
| 71 | +} |
| 72 | + |
| 73 | +define void @store_v8i16(<8 x i16> %a, ptr %ptr){ |
| 74 | +; CHECK-LABEL: store_v8i16: |
| 75 | +; CHECK: // %bb.0: |
| 76 | +; CHECK-NEXT: str q0, [x0] |
| 77 | +; CHECK-NEXT: ret |
| 78 | + store <8 x i16> %a, ptr %ptr |
| 79 | + ret void |
| 80 | +} |
| 81 | + |
| 82 | +define void @store_v2i32(<2 x i32> %a, ptr %ptr){ |
| 83 | +; CHECK-LABEL: store_v2i32: |
| 84 | +; CHECK: // %bb.0: |
| 85 | +; CHECK-NEXT: str d0, [x0] |
| 86 | +; CHECK-NEXT: ret |
| 87 | + store <2 x i32> %a, ptr %ptr |
| 88 | + ret void |
| 89 | +} |
| 90 | + |
| 91 | +define void @store_v4i32(<4 x i32> %a, ptr %ptr){ |
| 92 | +; CHECK-LABEL: store_v4i32: |
| 93 | +; CHECK: // %bb.0: |
| 94 | +; CHECK-NEXT: str q0, [x0] |
| 95 | +; CHECK-NEXT: ret |
| 96 | + store <4 x i32> %a, ptr %ptr |
| 97 | + ret void |
| 98 | +} |
| 99 | + |
| 100 | +define void @store_v2i64(<2 x i64> %a, ptr %ptr){ |
| 101 | +; CHECK-LABEL: store_v2i64: |
| 102 | +; CHECK: // %bb.0: |
| 103 | +; CHECK-NEXT: str q0, [x0] |
| 104 | +; CHECK-NEXT: ret |
| 105 | + store <2 x i64> %a, ptr %ptr |
| 106 | + ret void |
| 107 | +} |
| 108 | + |
| 109 | +; ===== Smaller/Larger Width Vectors with Legal Element Sizes ===== |
| 110 | + |
| 111 | +define void @store_v4i8(i32 %a, ptr %ptr) { |
| 112 | +; CHECK-LABEL: store_v4i8: |
| 113 | +; CHECK: // %bb.0: |
| 114 | +; CHECK-NEXT: str w0, [x1] |
| 115 | +; CHECK-NEXT: ret |
| 116 | + %c = bitcast i32 %a to <4 x i8> |
| 117 | + store <4 x i8> %c, ptr %ptr |
| 118 | + ret void |
| 119 | +} |
| 120 | + |
| 121 | +define void @store_v32i8(<32 x i8> %a, ptr %ptr){ |
| 122 | +; CHECK-LABEL: store_v32i8: |
| 123 | +; CHECK: // %bb.0: |
| 124 | +; CHECK-NEXT: stp q0, q1, [x0] |
| 125 | +; CHECK-NEXT: ret |
| 126 | + store <32 x i8> %a, ptr %ptr |
| 127 | + ret void |
| 128 | +} |
| 129 | + |
| 130 | +define void @store_v2i16(<2 x i16> %a, ptr %ptr){ |
| 131 | +; CHECK-SD-LABEL: store_v2i16: |
| 132 | +; CHECK-SD: // %bb.0: |
| 133 | +; CHECK-SD-NEXT: // kill: def $d0 killed $d0 def $q0 |
| 134 | +; CHECK-SD-NEXT: mov w8, v0.s[1] |
| 135 | +; CHECK-SD-NEXT: fmov w9, s0 |
| 136 | +; CHECK-SD-NEXT: strh w9, [x0] |
| 137 | +; CHECK-SD-NEXT: strh w8, [x0, #2] |
| 138 | +; CHECK-SD-NEXT: ret |
| 139 | +; |
| 140 | +; CHECK-GI-LABEL: store_v2i16: |
| 141 | +; CHECK-GI: // %bb.0: |
| 142 | +; CHECK-GI-NEXT: // kill: def $d0 killed $d0 def $q0 |
| 143 | +; CHECK-GI-NEXT: mov s1, v0.s[1] |
| 144 | +; CHECK-GI-NEXT: str h0, [x0] |
| 145 | +; CHECK-GI-NEXT: str h1, [x0, #2] |
| 146 | +; CHECK-GI-NEXT: ret |
| 147 | + store <2 x i16> %a, ptr %ptr |
| 148 | + ret void |
| 149 | +} |
| 150 | + |
| 151 | +define void @store_v16i16(<16 x i16> %a, ptr %ptr){ |
| 152 | +; CHECK-LABEL: store_v16i16: |
| 153 | +; CHECK: // %bb.0: |
| 154 | +; CHECK-NEXT: stp q0, q1, [x0] |
| 155 | +; CHECK-NEXT: ret |
| 156 | + store <16 x i16> %a, ptr %ptr |
| 157 | + ret void |
| 158 | +} |
| 159 | + |
| 160 | +define void @store_v1i32(<1 x i32> %a, ptr %ptr){ |
| 161 | +; CHECK-SD-LABEL: store_v1i32: |
| 162 | +; CHECK-SD: // %bb.0: |
| 163 | +; CHECK-SD-NEXT: // kill: def $d0 killed $d0 def $q0 |
| 164 | +; CHECK-SD-NEXT: str s0, [x0] |
| 165 | +; CHECK-SD-NEXT: ret |
| 166 | +; |
| 167 | +; CHECK-GI-LABEL: store_v1i32: |
| 168 | +; CHECK-GI: // %bb.0: |
| 169 | +; CHECK-GI-NEXT: str s0, [x0] |
| 170 | +; CHECK-GI-NEXT: ret |
| 171 | + store <1 x i32> %a, ptr %ptr |
| 172 | + ret void |
| 173 | +} |
| 174 | + |
| 175 | +define void @store_v8i32(<8 x i32> %a, ptr %ptr){ |
| 176 | +; CHECK-LABEL: store_v8i32: |
| 177 | +; CHECK: // %bb.0: |
| 178 | +; CHECK-NEXT: stp q0, q1, [x0] |
| 179 | +; CHECK-NEXT: ret |
| 180 | + store <8 x i32> %a, ptr %ptr |
| 181 | + ret void |
| 182 | +} |
| 183 | + |
| 184 | +define void @store_v4i64(<4 x i64> %a, ptr %ptr){ |
| 185 | +; CHECK-LABEL: store_v4i64: |
| 186 | +; CHECK: // %bb.0: |
| 187 | +; CHECK-NEXT: stp q0, q1, [x0] |
| 188 | +; CHECK-NEXT: ret |
| 189 | + store <4 x i64> %a, ptr %ptr |
| 190 | + ret void |
| 191 | +} |
| 192 | + |
| 193 | +; ===== Vectors with Non-Pow 2 Widths ===== |
| 194 | + |
| 195 | +define void @store_v3i8(<3 x i8> %a, ptr %ptr){ |
| 196 | +; CHECK-SD-LABEL: store_v3i8: |
| 197 | +; CHECK-SD: // %bb.0: |
| 198 | +; CHECK-SD-NEXT: sub sp, sp, #16 |
| 199 | +; CHECK-SD-NEXT: .cfi_def_cfa_offset 16 |
| 200 | +; CHECK-SD-NEXT: fmov s0, w0 |
| 201 | +; CHECK-SD-NEXT: mov v0.h[1], w1 |
| 202 | +; CHECK-SD-NEXT: mov v0.h[2], w2 |
| 203 | +; CHECK-SD-NEXT: xtn v0.8b, v0.8h |
| 204 | +; CHECK-SD-NEXT: str s0, [sp, #12] |
| 205 | +; CHECK-SD-NEXT: ldrh w8, [sp, #12] |
| 206 | +; CHECK-SD-NEXT: strb w2, [x3, #2] |
| 207 | +; CHECK-SD-NEXT: strh w8, [x3] |
| 208 | +; CHECK-SD-NEXT: add sp, sp, #16 |
| 209 | +; CHECK-SD-NEXT: ret |
| 210 | +; |
| 211 | +; CHECK-GI-LABEL: store_v3i8: |
| 212 | +; CHECK-GI: // %bb.0: |
| 213 | +; CHECK-GI-NEXT: strb w0, [x3] |
| 214 | +; CHECK-GI-NEXT: strb w1, [x3, #1] |
| 215 | +; CHECK-GI-NEXT: strb w2, [x3, #2] |
| 216 | +; CHECK-GI-NEXT: ret |
| 217 | + store <3 x i8> %a, ptr %ptr |
| 218 | + ret void |
| 219 | +} |
| 220 | + |
| 221 | +define void @store_v7i8(<7 x i8> %a, ptr %ptr){ |
| 222 | +; CHECK-SD-LABEL: store_v7i8: |
| 223 | +; CHECK-SD: // %bb.0: |
| 224 | +; CHECK-SD-NEXT: add x8, x0, #6 |
| 225 | +; CHECK-SD-NEXT: add x9, x0, #4 |
| 226 | +; CHECK-SD-NEXT: // kill: def $d0 killed $d0 def $q0 |
| 227 | +; CHECK-SD-NEXT: str s0, [x0] |
| 228 | +; CHECK-SD-NEXT: st1 { v0.b }[6], [x8] |
| 229 | +; CHECK-SD-NEXT: st1 { v0.h }[2], [x9] |
| 230 | +; CHECK-SD-NEXT: ret |
| 231 | +; |
| 232 | +; CHECK-GI-LABEL: store_v7i8: |
| 233 | +; CHECK-GI: // %bb.0: |
| 234 | +; CHECK-GI-NEXT: add x8, x0, #1 |
| 235 | +; CHECK-GI-NEXT: // kill: def $d0 killed $d0 def $q0 |
| 236 | +; CHECK-GI-NEXT: add x9, x0, #2 |
| 237 | +; CHECK-GI-NEXT: st1 { v0.b }[0], [x0] |
| 238 | +; CHECK-GI-NEXT: st1 { v0.b }[1], [x8] |
| 239 | +; CHECK-GI-NEXT: add x8, x0, #3 |
| 240 | +; CHECK-GI-NEXT: st1 { v0.b }[3], [x8] |
| 241 | +; CHECK-GI-NEXT: add x8, x0, #4 |
| 242 | +; CHECK-GI-NEXT: st1 { v0.b }[4], [x8] |
| 243 | +; CHECK-GI-NEXT: add x8, x0, #5 |
| 244 | +; CHECK-GI-NEXT: st1 { v0.b }[5], [x8] |
| 245 | +; CHECK-GI-NEXT: add x8, x0, #6 |
| 246 | +; CHECK-GI-NEXT: st1 { v0.b }[2], [x9] |
| 247 | +; CHECK-GI-NEXT: st1 { v0.b }[6], [x8] |
| 248 | +; CHECK-GI-NEXT: ret |
| 249 | + store <7 x i8> %a, ptr %ptr |
| 250 | + ret void |
| 251 | +} |
| 252 | + |
| 253 | +define void @store_v3i16(<3 x i16> %a, ptr %ptr){ |
| 254 | +; CHECK-SD-LABEL: store_v3i16: |
| 255 | +; CHECK-SD: // %bb.0: |
| 256 | +; CHECK-SD-NEXT: add x8, x0, #4 |
| 257 | +; CHECK-SD-NEXT: // kill: def $d0 killed $d0 def $q0 |
| 258 | +; CHECK-SD-NEXT: str s0, [x0] |
| 259 | +; CHECK-SD-NEXT: st1 { v0.h }[2], [x8] |
| 260 | +; CHECK-SD-NEXT: ret |
| 261 | +; |
| 262 | +; CHECK-GI-LABEL: store_v3i16: |
| 263 | +; CHECK-GI: // %bb.0: |
| 264 | +; CHECK-GI-NEXT: add x8, x0, #2 |
| 265 | +; CHECK-GI-NEXT: add x9, x0, #4 |
| 266 | +; CHECK-GI-NEXT: // kill: def $d0 killed $d0 def $q0 |
| 267 | +; CHECK-GI-NEXT: str h0, [x0] |
| 268 | +; CHECK-GI-NEXT: st1 { v0.h }[1], [x8] |
| 269 | +; CHECK-GI-NEXT: st1 { v0.h }[2], [x9] |
| 270 | +; CHECK-GI-NEXT: ret |
| 271 | + store <3 x i16> %a, ptr %ptr |
| 272 | + ret void |
| 273 | +} |
| 274 | + |
| 275 | +define void @store_v7i16(<7 x i16> %a, ptr %ptr){ |
| 276 | +; CHECK-SD-LABEL: store_v7i16: |
| 277 | +; CHECK-SD: // %bb.0: |
| 278 | +; CHECK-SD-NEXT: add x8, x0, #12 |
| 279 | +; CHECK-SD-NEXT: add x9, x0, #8 |
| 280 | +; CHECK-SD-NEXT: str d0, [x0] |
| 281 | +; CHECK-SD-NEXT: st1 { v0.h }[6], [x8] |
| 282 | +; CHECK-SD-NEXT: st1 { v0.s }[2], [x9] |
| 283 | +; CHECK-SD-NEXT: ret |
| 284 | +; |
| 285 | +; CHECK-GI-LABEL: store_v7i16: |
| 286 | +; CHECK-GI: // %bb.0: |
| 287 | +; CHECK-GI-NEXT: add x8, x0, #2 |
| 288 | +; CHECK-GI-NEXT: add x9, x0, #4 |
| 289 | +; CHECK-GI-NEXT: str h0, [x0] |
| 290 | +; CHECK-GI-NEXT: st1 { v0.h }[1], [x8] |
| 291 | +; CHECK-GI-NEXT: add x8, x0, #6 |
| 292 | +; CHECK-GI-NEXT: st1 { v0.h }[3], [x8] |
| 293 | +; CHECK-GI-NEXT: add x8, x0, #8 |
| 294 | +; CHECK-GI-NEXT: st1 { v0.h }[4], [x8] |
| 295 | +; CHECK-GI-NEXT: add x8, x0, #10 |
| 296 | +; CHECK-GI-NEXT: st1 { v0.h }[5], [x8] |
| 297 | +; CHECK-GI-NEXT: add x8, x0, #12 |
| 298 | +; CHECK-GI-NEXT: st1 { v0.h }[2], [x9] |
| 299 | +; CHECK-GI-NEXT: st1 { v0.h }[6], [x8] |
| 300 | +; CHECK-GI-NEXT: ret |
| 301 | + store <7 x i16> %a, ptr %ptr |
| 302 | + ret void |
| 303 | +} |
| 304 | + |
| 305 | +define void @store_v3i32(<3 x i32> %a, ptr %ptr){ |
| 306 | +; CHECK-SD-LABEL: store_v3i32: |
| 307 | +; CHECK-SD: // %bb.0: |
| 308 | +; CHECK-SD-NEXT: add x8, x0, #8 |
| 309 | +; CHECK-SD-NEXT: str d0, [x0] |
| 310 | +; CHECK-SD-NEXT: st1 { v0.s }[2], [x8] |
| 311 | +; CHECK-SD-NEXT: ret |
| 312 | +; |
| 313 | +; CHECK-GI-LABEL: store_v3i32: |
| 314 | +; CHECK-GI: // %bb.0: |
| 315 | +; CHECK-GI-NEXT: add x8, x0, #4 |
| 316 | +; CHECK-GI-NEXT: add x9, x0, #8 |
| 317 | +; CHECK-GI-NEXT: str s0, [x0] |
| 318 | +; CHECK-GI-NEXT: st1 { v0.s }[1], [x8] |
| 319 | +; CHECK-GI-NEXT: st1 { v0.s }[2], [x9] |
| 320 | +; CHECK-GI-NEXT: ret |
| 321 | + store <3 x i32> %a, ptr %ptr |
| 322 | + ret void |
| 323 | +} |
0 commit comments