llvm/test/CodeGen/AMDGPU/llvm.memcpy.ll

   1 ; RUN: llc -mtriple=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
   2 ; RUN: llc -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
   3
   4 declare void @llvm.memcpy.p3.p3.i32(ptr addrspace(3) nocapture, ptr addrspace(3) nocapture, i32, i1) nounwind
   5 declare void @llvm.memcpy.p1.p1.i64(ptr addrspace(1) nocapture, ptr addrspace(1) nocapture, i64, i1) nounwind
   6 declare void @llvm.memcpy.p1.p2.i64(ptr addrspace(1) nocapture, ptr addrspace(4) nocapture, i64, i1) nounwind
   7
   8
   9 ; FUNC-LABEL: {{^}}test_small_memcpy_i64_lds_to_lds_align1:
  10 ; SI-DAG: ds_read_u8
  11 ; SI-DAG: ds_read_u8
  12 ; SI-DAG: ds_read_u8
  13 ; SI-DAG: ds_read_u8
  14 ; SI-DAG: ds_read_u8
  15 ; SI-DAG: ds_read_u8
  16 ; SI-DAG: ds_read_u8
  17 ; SI-DAG: ds_read_u8
  18
  19 ; SI-DAG: ds_read_u8
  20 ; SI-DAG: ds_read_u8
  21 ; SI-DAG: ds_read_u8
  22 ; SI-DAG: ds_read_u8
  23 ; SI-DAG: ds_read_u8
  24 ; SI-DAG: ds_read_u8
  25 ; SI-DAG: ds_read_u8
  26 ; SI-DAG: ds_read_u8
  27
  28 ; SI-DAG: ds_read_u8
  29 ; SI-DAG: ds_read_u8
  30 ; SI-DAG: ds_read_u8
  31 ; SI-DAG: ds_read_u8
  32 ; SI-DAG: ds_read_u8
  33 ; SI-DAG: ds_read_u8
  34 ; SI-DAG: ds_read_u8
  35 ; SI-DAG: ds_read_u8
  36
  37 ; SI-DAG: ds_read_u8
  38 ; SI-DAG: ds_read_u8
  39 ; SI-DAG: ds_read_u8
  40 ; SI-DAG: ds_read_u8
  41 ; SI-DAG: ds_read_u8
  42 ; SI-DAG: ds_read_u8
  43 ; SI-DAG: ds_read_u8
  44 ; SI-DAG: ds_read_u8
  45
  46 ; SI-DAG: ds_write_b8
  47 ; SI-DAG: ds_write_b8
  48 ; SI-DAG: ds_write_b8
  49 ; SI-DAG: ds_write_b8
  50 ; SI-DAG: ds_write_b8
  51 ; SI-DAG: ds_write_b8
  52 ; SI-DAG: ds_write_b8
  53 ; SI-DAG: ds_write_b8
  54
  55 ; SI-DAG: ds_write_b8
  56 ; SI-DAG: ds_write_b8
  57 ; SI-DAG: ds_write_b8
  58 ; SI-DAG: ds_write_b8
  59 ; SI-DAG: ds_write_b8
  60 ; SI-DAG: ds_write_b8
  61 ; SI-DAG: ds_write_b8
  62 ; SI-DAG: ds_write_b8
  63
  64 ; SI-DAG: ds_write_b8
  65 ; SI-DAG: ds_write_b8
  66 ; SI-DAG: ds_write_b8
  67 ; SI-DAG: ds_write_b8
  68 ; SI-DAG: ds_write_b8
  69 ; SI-DAG: ds_write_b8
  70 ; SI-DAG: ds_write_b8
  71 ; SI-DAG: ds_write_b8
  72
  73 ; SI-DAG: ds_write_b8
  74 ; SI-DAG: ds_write_b8
  75 ; SI-DAG: ds_write_b8
  76 ; SI-DAG: ds_write_b8
  77 ; SI-DAG: ds_write_b8
  78 ; SI-DAG: ds_write_b8
  79 ; SI-DAG: ds_write_b8
  80 ; SI-DAG: ds_write_b8
  81
  82 ; SI: s_endpgm
  83 define amdgpu_kernel void @test_small_memcpy_i64_lds_to_lds_align1(ptr addrspace(3) noalias %out, ptr addrspace(3) noalias %in) nounwind {
  84   call void @llvm.memcpy.p3.p3.i32(ptr addrspace(3) %out, ptr addrspace(3) %in, i32 32, i1 false) nounwind
  85   ret void
  86 }
  87
  88 ; FUNC-LABEL: {{^}}test_small_memcpy_i64_lds_to_lds_align2:
  89 ; SI-DAG: ds_read_u16
  90 ; SI-DAG: ds_read_u16
  91 ; SI-DAG: ds_read_u16
  92 ; SI-DAG: ds_read_u16
  93 ; SI-DAG: ds_read_u16
  94 ; SI-DAG: ds_read_u16
  95 ; SI-DAG: ds_read_u16
  96 ; SI-DAG: ds_read_u16
  97
  98 ; SI-DAG: ds_read_u16
  99 ; SI-DAG: ds_read_u16
 100 ; SI-DAG: ds_read_u16
 101 ; SI-DAG: ds_read_u16
 102 ; SI-DAG: ds_read_u16
 103 ; SI-DAG: ds_read_u16
 104 ; SI-DAG: ds_read_u16
 105 ; SI-DAG: ds_read_u16
 106
 107 ; SI-DAG: ds_write_b16
 108 ; SI-DAG: ds_write_b16
 109 ; SI-DAG: ds_write_b16
 110 ; SI-DAG: ds_write_b16
 111 ; SI-DAG: ds_write_b16
 112 ; SI-DAG: ds_write_b16
 113 ; SI-DAG: ds_write_b16
 114 ; SI-DAG: ds_write_b16
 115
 116 ; SI-DAG: ds_write_b16
 117 ; SI-DAG: ds_write_b16
 118 ; SI-DAG: ds_write_b16
 119 ; SI-DAG: ds_write_b16
 120 ; SI-DAG: ds_write_b16
 121 ; SI-DAG: ds_write_b16
 122 ; SI-DAG: ds_write_b16
 123 ; SI-DAG: ds_write_b16
 124
 125 ; SI: s_endpgm
 126 define amdgpu_kernel void @test_small_memcpy_i64_lds_to_lds_align2(ptr addrspace(3) noalias %out, ptr addrspace(3) noalias %in) nounwind {
 127   call void @llvm.memcpy.p3.p3.i32(ptr addrspace(3) align 2 %out, ptr addrspace(3) align 2 %in, i32 32, i1 false) nounwind
 128   ret void
 129 }
 130
 131 ; FUNC-LABEL: {{^}}test_small_memcpy_i64_lds_to_lds_align4:
 132 ; SI: ds_read2_b32
 133 ; SI: ds_read2_b32
 134 ; SI: ds_read2_b32
 135 ; SI: ds_read2_b32
 136
 137 ; SI: ds_write2_b32
 138 ; SI: ds_write2_b32
 139 ; SI: ds_write2_b32
 140 ; SI: ds_write2_b32
 141
 142 ; SI: s_endpgm
 143 define amdgpu_kernel void @test_small_memcpy_i64_lds_to_lds_align4(ptr addrspace(3) noalias %out, ptr addrspace(3) noalias %in) nounwind {
 144   call void @llvm.memcpy.p3.p3.i32(ptr addrspace(3) align 4 %out, ptr addrspace(3) align 4 %in, i32 32, i1 false) nounwind
 145   ret void
 146 }
 147
 148 ; FIXME: Use 64-bit ops
 149 ; FUNC-LABEL: {{^}}test_small_memcpy_i64_lds_to_lds_align8:
 150
 151 ; SI: ds_read2_b64
 152 ; SI: ds_read2_b64
 153
 154 ; SI: ds_write2_b64
 155 ; SI: ds_write2_b64
 156
 157 ; SI-DAG: s_endpgm
 158 define amdgpu_kernel void @test_small_memcpy_i64_lds_to_lds_align8(ptr addrspace(3) noalias %out, ptr addrspace(3) noalias %in) nounwind {
 159   call void @llvm.memcpy.p3.p3.i32(ptr addrspace(3) align 8 %out, ptr addrspace(3) align 8 %in, i32 32, i1 false) nounwind
 160   ret void
 161 }
 162
 163 ; FUNC-LABEL: {{^}}test_small_memcpy_i64_global_to_global_align1:
 164 ; SI-DAG: buffer_load_ubyte
 165 ; SI-DAG: buffer_store_byte
 166 ; SI-DAG: buffer_load_ubyte
 167 ; SI-DAG: buffer_store_byte
 168 ; SI-DAG: buffer_load_ubyte
 169 ; SI-DAG: buffer_store_byte
 170 ; SI-DAG: buffer_load_ubyte
 171 ; SI-DAG: buffer_store_byte
 172 ; SI-DAG: buffer_load_ubyte
 173 ; SI-DAG: buffer_store_byte
 174 ; SI-DAG: buffer_load_ubyte
 175 ; SI-DAG: buffer_store_byte
 176 ; SI-DAG: buffer_load_ubyte
 177 ; SI-DAG: buffer_store_byte
 178 ; SI-DAG: buffer_load_ubyte
 179 ; SI-DAG: buffer_store_byte
 180
 181 ; SI-DAG: buffer_load_ubyte
 182 ; SI-DAG: buffer_store_byte
 183 ; SI-DAG: buffer_load_ubyte
 184 ; SI-DAG: buffer_store_byte
 185 ; SI-DAG: buffer_load_ubyte
 186 ; SI-DAG: buffer_store_byte
 187 ; SI-DAG: buffer_load_ubyte
 188 ; SI-DAG: buffer_store_byte
 189 ; SI-DAG: buffer_load_ubyte
 190 ; SI-DAG: buffer_store_byte
 191 ; SI-DAG: buffer_load_ubyte
 192 ; SI-DAG: buffer_store_byte
 193 ; SI-DAG: buffer_load_ubyte
 194 ; SI-DAG: buffer_store_byte
 195 ; SI-DAG: buffer_load_ubyte
 196 ; SI-DAG: buffer_store_byte
 197
 198 ; SI-DAG: buffer_load_ubyte
 199 ; SI-DAG: buffer_store_byte
 200 ; SI-DAG: buffer_load_ubyte
 201 ; SI-DAG: buffer_store_byte
 202 ; SI-DAG: buffer_load_ubyte
 203 ; SI-DAG: buffer_store_byte
 204 ; SI-DAG: buffer_load_ubyte
 205 ; SI-DAG: buffer_store_byte
 206 ; SI-DAG: buffer_load_ubyte
 207 ; SI-DAG: buffer_store_byte
 208 ; SI-DAG: buffer_load_ubyte
 209 ; SI-DAG: buffer_store_byte
 210 ; SI-DAG: buffer_load_ubyte
 211 ; SI-DAG: buffer_store_byte
 212 ; SI-DAG: buffer_load_ubyte
 213 ; SI-DAG: buffer_store_byte
 214
 215 ; SI-DAG: buffer_load_ubyte
 216 ; SI-DAG: buffer_store_byte
 217 ; SI-DAG: buffer_load_ubyte
 218 ; SI-DAG: buffer_store_byte
 219 ; SI-DAG: buffer_load_ubyte
 220 ; SI-DAG: buffer_store_byte
 221 ; SI-DAG: buffer_load_ubyte
 222 ; SI-DAG: buffer_store_byte
 223 ; SI-DAG: buffer_load_ubyte
 224 ; SI-DAG: buffer_store_byte
 225 ; SI-DAG: buffer_load_ubyte
 226 ; SI-DAG: buffer_store_byte
 227 ; SI-DAG: buffer_load_ubyte
 228 ; SI-DAG: buffer_store_byte
 229 ; SI-DAG: buffer_load_ubyte
 230 ; SI-DAG: buffer_store_byte
 231
 232 ; SI: s_endpgm
 233 define amdgpu_kernel void @test_small_memcpy_i64_global_to_global_align1(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) nounwind {
 234   call void @llvm.memcpy.p1.p1.i64(ptr addrspace(1) %out, ptr addrspace(1) %in, i64 32, i1 false) nounwind
 235   ret void
 236 }
 237
 238 ; FUNC-LABEL: {{^}}test_small_memcpy_i64_global_to_global_align2:
 239 ; SI-DAG: buffer_load_ushort
 240 ; SI-DAG: buffer_load_ushort
 241 ; SI-DAG: buffer_load_ushort
 242 ; SI-DAG: buffer_load_ushort
 243 ; SI-DAG: buffer_load_ushort
 244 ; SI-DAG: buffer_load_ushort
 245 ; SI-DAG: buffer_load_ushort
 246 ; SI-DAG: buffer_load_ushort
 247 ; SI-DAG: buffer_load_ushort
 248 ; SI-DAG: buffer_load_ushort
 249 ; SI-DAG: buffer_load_ushort
 250 ; SI-DAG: buffer_load_ushort
 251 ; SI-DAG: buffer_load_ushort
 252 ; SI-DAG: buffer_load_ushort
 253 ; SI-DAG: buffer_load_ushort
 254 ; SI-DAG: buffer_load_ushort
 255
 256 ; SI-DAG: buffer_store_short
 257 ; SI-DAG: buffer_store_short
 258 ; SI-DAG: buffer_store_short
 259 ; SI-DAG: buffer_store_short
 260 ; SI-DAG: buffer_store_short
 261 ; SI-DAG: buffer_store_short
 262 ; SI-DAG: buffer_store_short
 263 ; SI-DAG: buffer_store_short
 264 ; SI-DAG: buffer_store_short
 265 ; SI-DAG: buffer_store_short
 266 ; SI-DAG: buffer_store_short
 267 ; SI-DAG: buffer_store_short
 268 ; SI-DAG: buffer_store_short
 269 ; SI-DAG: buffer_store_short
 270 ; SI-DAG: buffer_store_short
 271 ; SI-DAG: buffer_store_short
 272
 273 ; SI: s_endpgm
 274 define amdgpu_kernel void @test_small_memcpy_i64_global_to_global_align2(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) nounwind {
 275   call void @llvm.memcpy.p1.p1.i64(ptr addrspace(1) align 2 %out, ptr addrspace(1) align 2 %in, i64 32, i1 false) nounwind
 276   ret void
 277 }
 278
 279 ; FUNC-LABEL: {{^}}test_small_memcpy_i64_global_to_global_align4:
 280 ; SI: buffer_load_dwordx4
 281 ; SI: buffer_load_dwordx4
 282 ; SI: buffer_store_dwordx4
 283 ; SI: buffer_store_dwordx4
 284 ; SI: s_endpgm
 285 define amdgpu_kernel void @test_small_memcpy_i64_global_to_global_align4(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) nounwind {
 286   call void @llvm.memcpy.p1.p1.i64(ptr addrspace(1) align 4 %out, ptr addrspace(1) align 4 %in, i64 32, i1 false) nounwind
 287   ret void
 288 }
 289
 290 ; FUNC-LABEL: {{^}}test_small_memcpy_i64_global_to_global_align8:
 291 ; SI: buffer_load_dwordx4
 292 ; SI: buffer_load_dwordx4
 293 ; SI: buffer_store_dwordx4
 294 ; SI: buffer_store_dwordx4
 295 ; SI: s_endpgm
 296 define amdgpu_kernel void @test_small_memcpy_i64_global_to_global_align8(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) nounwind {
 297   call void @llvm.memcpy.p1.p1.i64(ptr addrspace(1) align 8 %out, ptr addrspace(1) align 8 %in, i64 32, i1 false) nounwind
 298   ret void
 299 }
 300
 301 ; FUNC-LABEL: {{^}}test_small_memcpy_i64_global_to_global_align16:
 302 ; SI: buffer_load_dwordx4
 303 ; SI: buffer_load_dwordx4
 304 ; SI: buffer_store_dwordx4
 305 ; SI: buffer_store_dwordx4
 306 ; SI: s_endpgm
 307 define amdgpu_kernel void @test_small_memcpy_i64_global_to_global_align16(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) nounwind {
 308   call void @llvm.memcpy.p1.p1.i64(ptr addrspace(1) align 16 %out, ptr addrspace(1) align 16 %in, i64 32, i1 false) nounwind
 309   ret void
 310 }
 311
 312 ; Test shouldConvertConstantLoadToIntImm
 313 @hello.align4 = private unnamed_addr addrspace(4) constant [16 x i8] c"constant string\00", align 4
 314 @hello.align1 = private unnamed_addr addrspace(4) constant [16 x i8] c"constant string\00", align 1
 315
 316 ; FUNC-LABEL: {{^}}test_memcpy_const_string_align4:
 317 ; SI: s_getpc_b64
 318 ; SI: s_add_u32 s{{[0-9]+}}, s{{[0-9]+}}, hello.align4@rel32@lo+4
 319 ; SI: s_addc_u32
 320 ; SI-DAG: s_load_dwordx8
 321 ; SI-DAG: s_load_dwordx2
 322 ; SI-DAG: buffer_store_dwordx4
 323 ; SI-DAG: buffer_store_dwordx4
 324 define amdgpu_kernel void @test_memcpy_const_string_align4(ptr addrspace(1) noalias %out) nounwind {
 325   call void @llvm.memcpy.p1.p2.i64(ptr addrspace(1) align 4 %out, ptr addrspace(4) align 4 @hello.align4, i64 32, i1 false)
 326   ret void
 327 }
 328
 329 ; FUNC-LABEL: {{^}}test_memcpy_const_string_align1:
 330 ; SI-NOT: buffer_load
 331 ; SI: v_mov_b32_e32 v{{[0-9]+}}, 0x
 332 ; SI: buffer_store_byte
 333 ; SI: buffer_store_byte
 334 ; SI: buffer_store_byte
 335 ; SI: buffer_store_byte
 336 ; SI: buffer_store_byte
 337 ; SI: buffer_store_byte
 338 ; SI: buffer_store_byte
 339 ; SI: buffer_store_byte
 340 ; SI: buffer_store_byte
 341 ; SI: buffer_store_byte
 342 ; SI: buffer_store_byte
 343 ; SI: buffer_store_byte
 344 ; SI: buffer_store_byte
 345 ; SI: buffer_store_byte
 346 ; SI: buffer_store_byte
 347 ; SI: buffer_store_byte
 348 define amdgpu_kernel void @test_memcpy_const_string_align1(ptr addrspace(1) noalias %out) nounwind {
 349   call void @llvm.memcpy.p1.p2.i64(ptr addrspace(1) %out, ptr addrspace(4) @hello.align1, i64 32, i1 false)
 350   ret void
 351 }