Matt Arsenault | 4b7938c | 2017-11-13 23:24:26 +0000 | [diff] [blame^] | 1 | ; RUN: llc -march=amdgcn -mcpu=gfx900 -amdgpu-sroa=0 -mattr=-promote-alloca -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX9 %s |
| 2 | ; RUN: llc -march=amdgcn -mcpu=fiji -amdgpu-sroa=0 -mattr=-promote-alloca -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,VI %s |
Matt Arsenault | e1cd482 | 2017-11-13 00:22:09 +0000 | [diff] [blame] | 3 | |
| 4 | ; GCN-LABEL: {{^}}load_local_lo_v2i16_undeflo: |
| 5 | ; GCN: s_waitcnt |
| 6 | ; GFX9-NEXT: ds_read_u16_d16 v0, v0 |
| 7 | ; GFX9-NEXT: s_waitcnt |
| 8 | ; GFX9-NEXT: s_setpc_b64 |
| 9 | |
| 10 | ; VI: ds_read_u16 |
| 11 | define <2 x i16> @load_local_lo_v2i16_undeflo(i16 addrspace(3)* %in) #0 { |
| 12 | entry: |
| 13 | %load = load i16, i16 addrspace(3)* %in |
| 14 | %build = insertelement <2 x i16> undef, i16 %load, i32 0 |
| 15 | ret <2 x i16> %build |
| 16 | } |
| 17 | |
| 18 | ; GCN-LABEL: {{^}}load_local_lo_v2i16_reglo: |
| 19 | ; GCN: s_waitcnt |
| 20 | ; GFX9-NEXT: ds_read_u16_d16 v0, v0 |
| 21 | ; GFX9-NEXT: s_waitcnt |
| 22 | ; GFX9-NEXT: s_setpc_b64 |
| 23 | |
| 24 | ; VI: ds_read_u16 |
| 25 | define <2 x i16> @load_local_lo_v2i16_reglo(i16 addrspace(3)* %in, i16 %reg) #0 { |
| 26 | entry: |
| 27 | %load = load i16, i16 addrspace(3)* %in |
| 28 | %build0 = insertelement <2 x i16> undef, i16 %reg, i32 0 |
| 29 | %build1 = insertelement <2 x i16> %build0, i16 %load, i32 0 |
| 30 | ret <2 x i16> %build1 |
| 31 | } |
| 32 | |
| 33 | ; Show that we get reasonable regalloc without physreg constraints. |
| 34 | ; GCN-LABEL: {{^}}load_local_lo_v2i16_reglo_vreg: |
| 35 | ; GCN: s_waitcnt |
| 36 | ; GFX9-NEXT: ds_read_u16_d16 v0, v0 |
| 37 | ; GFX9-NEXT: s_waitcnt |
| 38 | ; GFX9-NEXT: global_store_dword v[0:1], v0, off{{$}} |
| 39 | ; GFX9-NEXT: s_waitcnt |
| 40 | ; GFX9-NEXT: s_setpc_b64 |
| 41 | |
| 42 | ; VI: ds_read_u16 |
| 43 | define void @load_local_lo_v2i16_reglo_vreg(i16 addrspace(3)* %in, i16 %reg) #0 { |
| 44 | entry: |
| 45 | %load = load i16, i16 addrspace(3)* %in |
| 46 | %build0 = insertelement <2 x i16> undef, i16 %reg, i32 0 |
| 47 | %build1 = insertelement <2 x i16> %build0, i16 %load, i32 0 |
| 48 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 49 | ret void |
| 50 | } |
| 51 | |
| 52 | ; GCN-LABEL: {{^}}load_local_lo_v2i16_zerolo: |
| 53 | ; GCN: s_waitcnt |
| 54 | ; GFX9-NEXT: v_mov_b32_e32 v1, 0 |
| 55 | ; GFX9-NEXT: ds_read_u16_d16 v1, v0 |
| 56 | ; GFX9-NEXT: s_waitcnt |
| 57 | ; GFX9-NEXT: v_mov_b32_e32 v0, v1 |
| 58 | ; GFX9-NEXT: s_setpc_b64 |
| 59 | |
| 60 | ; VI: ds_read_u16 v |
| 61 | define <2 x i16> @load_local_lo_v2i16_zerolo(i16 addrspace(3)* %in) #0 { |
| 62 | entry: |
| 63 | %load = load i16, i16 addrspace(3)* %in |
| 64 | %build = insertelement <2 x i16> zeroinitializer, i16 %load, i32 0 |
| 65 | ret <2 x i16> %build |
| 66 | } |
| 67 | |
| 68 | ; GCN-LABEL: {{^}}load_local_lo_v2f16_fpimm: |
| 69 | ; GCN: s_waitcnt |
| 70 | ; GFX9-NEXT: v_mov_b32_e32 v1, 2.0 |
| 71 | ; GFX9-NEXT: ds_read_u16_d16 v1, v0 |
| 72 | ; GFX9-NEXT: s_waitcnt |
| 73 | ; GFX9-NEXT: v_mov_b32_e32 v0, v1 |
| 74 | ; GFX9-NEXT: s_setpc_b64 |
| 75 | |
| 76 | ; VI: ds_read_u16 v |
| 77 | define <2 x half> @load_local_lo_v2f16_fpimm(half addrspace(3)* %in) #0 { |
| 78 | entry: |
| 79 | %load = load half, half addrspace(3)* %in |
| 80 | %build = insertelement <2 x half> <half 0.0, half 2.0>, half %load, i32 0 |
| 81 | ret <2 x half> %build |
| 82 | } |
| 83 | |
| 84 | ; GCN-LABEL: {{^}}load_local_lo_v2f16_reghi_vreg: |
| 85 | ; GCN: s_waitcnt |
| 86 | ; GFX9-NEXT: ds_read_u16_d16 v1, v0 |
| 87 | ; GFX9-NEXT: s_waitcnt |
| 88 | ; GFX9-NEXT: global_store_dword v[0:1], v1, off{{$}} |
| 89 | ; GFX9-NEXT: s_waitcnt |
| 90 | ; GFX9-NEXT: s_setpc_b64 |
| 91 | |
| 92 | ; VI: ds_read_u16 v |
| 93 | define void @load_local_lo_v2f16_reghi_vreg(half addrspace(3)* %in, i32 %reg) #0 { |
| 94 | entry: |
| 95 | %reg.bc = bitcast i32 %reg to <2 x half> |
| 96 | %load = load half, half addrspace(3)* %in |
| 97 | %build1 = insertelement <2 x half> %reg.bc, half %load, i32 0 |
| 98 | store <2 x half> %build1, <2 x half> addrspace(1)* undef |
| 99 | ret void |
| 100 | } |
| 101 | |
| 102 | ; GCN-LABEL: {{^}}load_local_lo_v2f16_reglo_vreg: |
| 103 | |
| 104 | ; GFX9: ds_read_u16 v |
| 105 | ; GFX9: v_and_b32_e32 v{{[0-9]+}}, 0xffff, v{{[0-9]+}} |
| 106 | ; GFX9: v_lshl_or_b32 v{{[0-9]+}}, v{{[0-9]+}}, 16, v{{[0-9]+}} |
| 107 | ; GFX9: global_store_dword |
| 108 | |
| 109 | ; VI: ds_read_u16 v |
| 110 | define void @load_local_lo_v2f16_reglo_vreg(half addrspace(3)* %in, half %reg) #0 { |
| 111 | entry: |
| 112 | %load = load half, half addrspace(3)* %in |
| 113 | %build0 = insertelement <2 x half> undef, half %reg, i32 1 |
| 114 | %build1 = insertelement <2 x half> %build0, half %load, i32 0 |
| 115 | store <2 x half> %build1, <2 x half> addrspace(1)* undef |
| 116 | ret void |
| 117 | } |
| 118 | |
| 119 | ; GCN-LABEL: {{^}}load_local_lo_v2i16_reghi_vreg_zexti8: |
| 120 | ; GCN: s_waitcnt |
| 121 | ; GFX9-NEXT: ds_read_u8_d16 v1, v0 |
| 122 | ; GFX9-NEXT: s_waitcnt |
| 123 | ; GFX9-NEXT: global_store_dword v[0:1], v1, off{{$}} |
| 124 | ; GFX9-NEXT: s_waitcnt |
| 125 | ; GFX9-NEXT: s_setpc_b64 |
| 126 | |
| 127 | ; VI: ds_read_u8 v |
| 128 | define void @load_local_lo_v2i16_reghi_vreg_zexti8(i8 addrspace(3)* %in, i32 %reg) #0 { |
| 129 | entry: |
| 130 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 131 | %load = load i8, i8 addrspace(3)* %in |
| 132 | %ext = zext i8 %load to i16 |
| 133 | %build1 = insertelement <2 x i16> %reg.bc, i16 %ext, i32 0 |
| 134 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 135 | ret void |
| 136 | } |
| 137 | |
| 138 | ; GCN-LABEL: {{^}}load_local_lo_v2i16_reglo_vreg_zexti8: |
| 139 | ; GCN: s_waitcnt |
| 140 | ; GFX9: ds_read_u8 v |
| 141 | ; GFX9: global_store_dword |
| 142 | ; GFX9-NEXT: s_waitcnt |
| 143 | ; GFX9-NEXT: s_setpc_b64 |
| 144 | |
| 145 | ; VI: ds_read_u8 v |
| 146 | define void @load_local_lo_v2i16_reglo_vreg_zexti8(i8 addrspace(3)* %in, i16 %reg) #0 { |
| 147 | entry: |
| 148 | %load = load i8, i8 addrspace(3)* %in |
| 149 | %ext = zext i8 %load to i16 |
| 150 | %build0 = insertelement <2 x i16> undef, i16 %reg, i32 1 |
| 151 | %build1 = insertelement <2 x i16> %build0, i16 %ext, i32 0 |
| 152 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 153 | ret void |
| 154 | } |
| 155 | |
| 156 | ; GCN-LABEL: {{^}}load_local_lo_v2i16_reghi_vreg_sexti8: |
| 157 | ; GCN: s_waitcnt |
| 158 | ; GFX9-NEXT: ds_read_i8_d16 v1, v0 |
| 159 | ; GFX9-NEXT: s_waitcnt |
| 160 | ; GFX9-NEXT: global_store_dword v[0:1], v1, off{{$}} |
| 161 | ; GFX9-NEXT: s_waitcnt |
| 162 | ; GFX9-NEXT: s_setpc_b64 |
| 163 | |
| 164 | ; VI: ds_read_i8 v |
| 165 | define void @load_local_lo_v2i16_reghi_vreg_sexti8(i8 addrspace(3)* %in, i32 %reg) #0 { |
| 166 | entry: |
| 167 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 168 | %load = load i8, i8 addrspace(3)* %in |
| 169 | %ext = sext i8 %load to i16 |
| 170 | %build1 = insertelement <2 x i16> %reg.bc, i16 %ext, i32 0 |
| 171 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 172 | ret void |
| 173 | } |
| 174 | |
| 175 | ; GCN-LABEL: {{^}}load_local_lo_v2i16_reglo_vreg_sexti8: |
| 176 | ; GCN: s_waitcnt |
| 177 | ; GFX9: ds_read_i8 v |
| 178 | ; GFX9: v_and_b32_e32 v{{[0-9]+}}, 0xffff, v{{[0-9]+}} |
| 179 | ; GFX9: v_lshl_or_b32 v{{[0-9]+}}, v{{[0-9]+}}, 16, v{{[0-9]+}} |
| 180 | |
| 181 | ; VI: ds_read_i8 v |
| 182 | define void @load_local_lo_v2i16_reglo_vreg_sexti8(i8 addrspace(3)* %in, i16 %reg) #0 { |
| 183 | entry: |
| 184 | %load = load i8, i8 addrspace(3)* %in |
| 185 | %ext = sext i8 %load to i16 |
| 186 | %build0 = insertelement <2 x i16> undef, i16 %reg, i32 1 |
| 187 | %build1 = insertelement <2 x i16> %build0, i16 %ext, i32 0 |
| 188 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 189 | ret void |
| 190 | } |
| 191 | |
| 192 | ; GCN-LABEL: {{^}}load_global_lo_v2i16_reglo_vreg: |
| 193 | ; GCN: s_waitcnt |
| 194 | ; GFX9-NEXT: global_load_short_d16 v2, v[0:1], off offset:-4094 |
| 195 | ; GFX9-NEXT: s_waitcnt |
| 196 | ; GFX9-NEXT: global_store_dword |
| 197 | ; GFX9-NEXT: s_waitcnt |
| 198 | ; GFX9-NEXT: s_setpc_b64 |
| 199 | define void @load_global_lo_v2i16_reglo_vreg(i16 addrspace(1)* %in, i32 %reg) #0 { |
| 200 | entry: |
| 201 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 202 | %gep = getelementptr inbounds i16, i16 addrspace(1)* %in, i64 -2047 |
| 203 | %load = load i16, i16 addrspace(1)* %gep |
| 204 | %build1 = insertelement <2 x i16> %reg.bc, i16 %load, i32 0 |
| 205 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 206 | ret void |
| 207 | } |
| 208 | |
| 209 | ; GCN-LABEL: {{^}}load_global_lo_v2f16_reglo_vreg: |
| 210 | ; GCN: s_waitcnt |
| 211 | ; GFX9-NEXT: global_load_short_d16 v2, v[0:1], off offset:-4094 |
| 212 | ; GFX9-NEXT: s_waitcnt |
| 213 | ; GFX9-NEXT: global_store_dword |
| 214 | ; GFX9-NEXT: s_waitcnt |
| 215 | ; GFX9-NEXT: s_setpc_b64 |
| 216 | define void @load_global_lo_v2f16_reglo_vreg(half addrspace(1)* %in, i32 %reg) #0 { |
| 217 | entry: |
| 218 | %reg.bc = bitcast i32 %reg to <2 x half> |
| 219 | %gep = getelementptr inbounds half, half addrspace(1)* %in, i64 -2047 |
| 220 | %load = load half, half addrspace(1)* %gep |
| 221 | %build1 = insertelement <2 x half> %reg.bc, half %load, i32 0 |
| 222 | store <2 x half> %build1, <2 x half> addrspace(1)* undef |
| 223 | ret void |
| 224 | } |
| 225 | |
| 226 | ; GCN-LABEL: {{^}}load_global_lo_v2i16_reglo_vreg_zexti8: |
| 227 | ; GCN: s_waitcnt |
| 228 | ; GFX9-NEXT: global_load_ubyte_d16 v2, v[0:1], off offset:-4095 |
| 229 | ; GFX9-NEXT: s_waitcnt |
| 230 | ; GFX9-NEXT: global_store_dword |
| 231 | ; GFX9-NEXT: s_waitcnt |
| 232 | ; GFX9-NEXT: s_setpc_b64 |
| 233 | define void @load_global_lo_v2i16_reglo_vreg_zexti8(i8 addrspace(1)* %in, i32 %reg) #0 { |
| 234 | entry: |
| 235 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 236 | %gep = getelementptr inbounds i8, i8 addrspace(1)* %in, i64 -4095 |
| 237 | %load = load i8, i8 addrspace(1)* %gep |
| 238 | %ext = zext i8 %load to i16 |
| 239 | %build1 = insertelement <2 x i16> %reg.bc, i16 %ext, i32 0 |
| 240 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 241 | ret void |
| 242 | } |
| 243 | |
| 244 | ; GCN-LABEL: {{^}}load_global_lo_v2i16_reglo_vreg_sexti8: |
| 245 | ; GCN: s_waitcnt |
| 246 | ; GFX9-NEXT: global_load_sbyte_d16 v2, v[0:1], off offset:-4095 |
| 247 | ; GFX9-NEXT: s_waitcnt |
| 248 | ; GFX9-NEXT: global_store_dword |
| 249 | ; GFX9-NEXT: s_waitcnt |
| 250 | ; GFX9-NEXT: s_setpc_b64 |
| 251 | define void @load_global_lo_v2i16_reglo_vreg_sexti8(i8 addrspace(1)* %in, i32 %reg) #0 { |
| 252 | entry: |
| 253 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 254 | %gep = getelementptr inbounds i8, i8 addrspace(1)* %in, i64 -4095 |
| 255 | %load = load i8, i8 addrspace(1)* %gep |
| 256 | %ext = sext i8 %load to i16 |
| 257 | %build1 = insertelement <2 x i16> %reg.bc, i16 %ext, i32 0 |
| 258 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 259 | ret void |
| 260 | } |
| 261 | |
| 262 | ; GCN-LABEL: {{^}}load_flat_lo_v2i16_reghi_vreg: |
| 263 | ; GCN: s_waitcnt |
| 264 | ; GFX9-NEXT: flat_load_short_d16 v2, v[0:1] |
| 265 | ; GFX9-NEXT: s_waitcnt |
| 266 | ; GFX9-NEXT: global_store_dword v[0:1], v2 |
| 267 | ; GFX9-NEXT: s_waitcnt |
| 268 | ; GFX9-NEXT: s_setpc_b64 |
| 269 | |
| 270 | ; VI: flat_load_ushort v{{[0-9]+}} |
| 271 | ; VI: v_or_b32_e32 |
| 272 | define void @load_flat_lo_v2i16_reghi_vreg(i16 addrspace(4)* %in, i32 %reg) #0 { |
| 273 | entry: |
| 274 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 275 | %load = load i16, i16 addrspace(4)* %in |
| 276 | %build1 = insertelement <2 x i16> %reg.bc, i16 %load, i32 0 |
| 277 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 278 | ret void |
| 279 | } |
| 280 | |
| 281 | ; GCN-LABEL: {{^}}load_flat_lo_v2f16_reghi_vreg: |
| 282 | ; GCN: s_waitcnt |
| 283 | ; GFX9-NEXT: flat_load_short_d16 v2, v[0:1] |
| 284 | ; GFX9-NEXT: s_waitcnt |
| 285 | ; GFX9-NEXT: global_store_dword v[0:1], v2 |
| 286 | ; GFX9-NEXT: s_waitcnt |
| 287 | ; GFX9-NEXT: s_setpc_b64 |
| 288 | |
| 289 | ; VI: flat_load_ushort v{{[0-9]+}} |
| 290 | ; VI: v_or_b32_e32 |
| 291 | define void @load_flat_lo_v2f16_reghi_vreg(half addrspace(4)* %in, i32 %reg) #0 { |
| 292 | entry: |
| 293 | %reg.bc = bitcast i32 %reg to <2 x half> |
| 294 | %load = load half, half addrspace(4)* %in |
| 295 | %build1 = insertelement <2 x half> %reg.bc, half %load, i32 0 |
| 296 | store <2 x half> %build1, <2 x half> addrspace(1)* undef |
| 297 | ret void |
| 298 | } |
| 299 | |
| 300 | ; GCN-LABEL: {{^}}load_flat_lo_v2i16_reglo_vreg_zexti8: |
| 301 | ; GCN: s_waitcnt |
| 302 | ; GFX9-NEXT: flat_load_ubyte_d16 v2, v[0:1] |
| 303 | ; GFX9-NEXT: s_waitcnt |
| 304 | ; GFX9-NEXT: global_store_dword v[0:1], v2 |
| 305 | ; GFX9-NEXT: s_waitcnt |
| 306 | ; GFX9-NEXT: s_setpc_b64 |
| 307 | |
| 308 | ; VI: flat_load_ubyte v{{[0-9]+}} |
| 309 | ; VI: v_or_b32_e32 |
| 310 | define void @load_flat_lo_v2i16_reglo_vreg_zexti8(i8 addrspace(4)* %in, i32 %reg) #0 { |
| 311 | entry: |
| 312 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 313 | %load = load i8, i8 addrspace(4)* %in |
| 314 | %ext = zext i8 %load to i16 |
| 315 | %build1 = insertelement <2 x i16> %reg.bc, i16 %ext, i32 0 |
| 316 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 317 | ret void |
| 318 | } |
| 319 | |
| 320 | ; GCN-LABEL: {{^}}load_flat_lo_v2i16_reglo_vreg_sexti8: |
| 321 | ; GCN: s_waitcnt |
| 322 | ; GFX9-NEXT: flat_load_sbyte_d16 v2, v[0:1] |
| 323 | ; GFX9-NEXT: s_waitcnt |
| 324 | ; GFX9-NEXT: global_store_dword v[0:1], v2 |
| 325 | ; GFX9-NEXT: s_waitcnt |
| 326 | ; GFX9-NEXT: s_setpc_b64 |
| 327 | |
| 328 | ; VI: flat_load_sbyte v{{[0-9]+}} |
| 329 | ; VI: v_or_b32_sdwa v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD |
| 330 | |
| 331 | define void @load_flat_lo_v2i16_reglo_vreg_sexti8(i8 addrspace(4)* %in, i32 %reg) #0 { |
| 332 | entry: |
| 333 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 334 | %load = load i8, i8 addrspace(4)* %in |
| 335 | %ext = sext i8 %load to i16 |
| 336 | %build1 = insertelement <2 x i16> %reg.bc, i16 %ext, i32 0 |
| 337 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 338 | ret void |
| 339 | } |
| 340 | |
| 341 | ; GCN-LABEL: {{^}}load_private_lo_v2i16_reglo_vreg: |
| 342 | ; GCN: s_waitcnt |
| 343 | ; GFX9-NEXT: buffer_load_short_d16 v1, v0, s[0:3], s4 offen offset:4094{{$}} |
| 344 | ; GFX9-NEXT: s_waitcnt |
| 345 | ; GFX9-NEXT: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, v1 |
| 346 | ; GFX9-NEXT: s_waitcnt |
| 347 | ; GFX9-NEXT: s_setpc_b64 |
| 348 | |
| 349 | ; VI: buffer_load_ushort v{{[0-9]+}}, v0, s[0:3], s4 offen offset:4094{{$}} |
| 350 | define void @load_private_lo_v2i16_reglo_vreg(i16* %in, i32 %reg) #0 { |
| 351 | entry: |
| 352 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 353 | %gep = getelementptr inbounds i16, i16* %in, i64 2047 |
| 354 | %load = load i16, i16* %gep |
| 355 | %build1 = insertelement <2 x i16> %reg.bc, i16 %load, i32 0 |
| 356 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 357 | ret void |
| 358 | } |
| 359 | |
| 360 | ; GCN-LABEL: {{^}}load_private_lo_v2i16_reghi_vreg: |
| 361 | ; GCN: s_waitcnt |
| 362 | ; GFX9-NEXT: buffer_load_ushort v{{[0-9]+}}, v{{[0-9]+}}, s[0:3], s4 offen offset:4094{{$}} |
| 363 | ; GFX9-NEXT: s_waitcnt |
| 364 | ; GFX9: v_and_b32 |
| 365 | ; GFX9: v_lshl_or_b32 |
| 366 | |
| 367 | ; GFX9-NEXT: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}} |
| 368 | ; GFX9-NEXT: s_waitcnt |
| 369 | ; GFX9-NEXT: s_setpc_b64 |
| 370 | |
| 371 | ; VI: buffer_load_ushort v{{[0-9]+}}, v0, s[0:3], s4 offen offset:4094{{$}} |
| 372 | define void @load_private_lo_v2i16_reghi_vreg(i16* %in, i16 %reg) #0 { |
| 373 | entry: |
| 374 | %gep = getelementptr inbounds i16, i16* %in, i64 2047 |
| 375 | %load = load i16, i16* %gep |
| 376 | %build0 = insertelement <2 x i16> undef, i16 %reg, i32 1 |
| 377 | %build1 = insertelement <2 x i16> %build0, i16 %load, i32 0 |
| 378 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 379 | ret void |
| 380 | } |
| 381 | |
| 382 | ; GCN-LABEL: {{^}}load_private_lo_v2f16_reglo_vreg: |
| 383 | ; GCN: s_waitcnt |
| 384 | ; GFX9-NEXT: buffer_load_short_d16 v1, v0, s[0:3], s4 offen offset:4094{{$}} |
| 385 | ; GFX9-NEXT: s_waitcnt |
| 386 | ; GFX9-NEXT: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, v1 |
| 387 | ; GFX9-NEXT: s_waitcnt |
| 388 | ; GFX9-NEXT: s_setpc_b64 |
| 389 | |
| 390 | ; VI: buffer_load_ushort v{{[0-9]+}}, v0, s[0:3], s4 offen offset:4094{{$}} |
| 391 | define void @load_private_lo_v2f16_reglo_vreg(half* %in, i32 %reg) #0 { |
| 392 | entry: |
| 393 | %reg.bc = bitcast i32 %reg to <2 x half> |
| 394 | %gep = getelementptr inbounds half, half* %in, i64 2047 |
| 395 | %load = load half, half* %gep |
| 396 | %build1 = insertelement <2 x half> %reg.bc, half %load, i32 0 |
| 397 | store <2 x half> %build1, <2 x half> addrspace(1)* undef |
| 398 | ret void |
| 399 | } |
| 400 | |
| 401 | ; GCN-LABEL: {{^}}load_private_lo_v2i16_reglo_vreg_nooff: |
| 402 | ; GCN: s_waitcnt |
| 403 | ; GFX9-NEXT: buffer_load_short_d16 v1, off, s[0:3], s4 offset:4094{{$}} |
| 404 | ; GFX9-NEXT: s_waitcnt |
| 405 | ; GFX9-NEXT: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, v1 |
| 406 | ; GFX9-NEXT: s_waitcnt |
| 407 | ; GFX9-NEXT: s_setpc_b64 |
| 408 | |
| 409 | ; VI: buffer_load_ushort v{{[0-9]+}}, off, s[0:3], s4 offset:4094{{$}} |
| 410 | define void @load_private_lo_v2i16_reglo_vreg_nooff(i16* %in, i32 %reg) #0 { |
| 411 | entry: |
| 412 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 413 | %load = load volatile i16, i16* inttoptr (i32 4094 to i16*) |
| 414 | %build1 = insertelement <2 x i16> %reg.bc, i16 %load, i32 0 |
| 415 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 416 | ret void |
| 417 | } |
| 418 | |
| 419 | ; GCN-LABEL: {{^}}load_private_lo_v2i16_reghi_vreg_nooff: |
| 420 | ; GCN: s_waitcnt |
| 421 | ; GFX9-NEXT: buffer_load_short_d16 v1, off, s[0:3], s4 offset:4094{{$}} |
| 422 | ; GFX9-NEXT: s_waitcnt |
| 423 | ; GFX9-NEXT: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, v1 |
| 424 | ; GFX9-NEXT: s_waitcnt |
| 425 | ; GFX9-NEXT: s_setpc_b64 |
| 426 | |
| 427 | ; VI: buffer_load_ushort v{{[0-9]+}}, off, s[0:3], s4 offset:4094{{$}} |
| 428 | define void @load_private_lo_v2i16_reghi_vreg_nooff(i16* %in, i32 %reg) #0 { |
| 429 | entry: |
| 430 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 431 | %load = load volatile i16, i16* inttoptr (i32 4094 to i16*) |
| 432 | %build1 = insertelement <2 x i16> %reg.bc, i16 %load, i32 0 |
| 433 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 434 | ret void |
| 435 | } |
| 436 | |
| 437 | ; GCN-LABEL: {{^}}load_private_lo_v2f16_reglo_vreg_nooff: |
| 438 | ; GCN: s_waitcnt |
| 439 | ; GFX9-NEXT: buffer_load_short_d16 v1, off, s[0:3], s4 offset:4094{{$}} |
| 440 | ; GFX9-NEXT: s_waitcnt |
| 441 | ; GFX9-NEXT: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, v1 |
| 442 | ; GFX9-NEXT: s_waitcnt |
| 443 | ; GFX9-NEXT: s_setpc_b64 |
| 444 | |
| 445 | ; VI: buffer_load_ushort v{{[0-9]+}}, off, s[0:3], s4 offset:4094{{$}} |
| 446 | define void @load_private_lo_v2f16_reglo_vreg_nooff(half* %in, i32 %reg) #0 { |
| 447 | entry: |
| 448 | %reg.bc = bitcast i32 %reg to <2 x half> |
| 449 | %load = load volatile half, half* inttoptr (i32 4094 to half*) |
| 450 | %build1 = insertelement <2 x half> %reg.bc, half %load, i32 0 |
| 451 | store <2 x half> %build1, <2 x half> addrspace(1)* undef |
| 452 | ret void |
| 453 | } |
| 454 | |
| 455 | ; GCN-LABEL: {{^}}load_private_lo_v2i16_reglo_vreg_zexti8: |
| 456 | ; GCN: s_waitcnt |
| 457 | ; GFX9-NEXT: buffer_load_ubyte_d16 v1, v0, s[0:3], s4 offen offset:2047{{$}} |
| 458 | ; GFX9-NEXT: s_waitcnt |
| 459 | ; GFX9-NEXT: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, v1 |
| 460 | ; GFX9-NEXT: s_waitcnt |
| 461 | ; GFX9-NEXT: s_setpc_b64 |
| 462 | |
| 463 | ; VI: buffer_load_ubyte v{{[0-9]+}}, v0, s[0:3], s4 offen offset:2047{{$}} |
| 464 | define void @load_private_lo_v2i16_reglo_vreg_zexti8(i8* %in, i32 %reg) #0 { |
| 465 | entry: |
| 466 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 467 | %gep = getelementptr inbounds i8, i8* %in, i64 2047 |
| 468 | %load = load i8, i8* %gep |
| 469 | %ext = zext i8 %load to i16 |
| 470 | %build1 = insertelement <2 x i16> %reg.bc, i16 %ext, i32 0 |
| 471 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 472 | ret void |
| 473 | } |
| 474 | |
| 475 | ; GCN-LABEL: {{^}}load_private_lo_v2i16_reglo_vreg_sexti8: |
| 476 | ; GCN: s_waitcnt |
| 477 | ; GFX9-NEXT: buffer_load_sbyte_d16 v1, v0, s[0:3], s4 offen offset:2047{{$}} |
| 478 | ; GFX9-NEXT: s_waitcnt |
| 479 | ; GFX9-NEXT: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, v1 |
| 480 | ; GFX9-NEXT: s_waitcnt |
| 481 | ; GFX9-NEXT: s_setpc_b64 |
| 482 | |
| 483 | ; VI: buffer_load_sbyte v{{[0-9]+}}, v0, s[0:3], s4 offen offset:2047{{$}} |
| 484 | define void @load_private_lo_v2i16_reglo_vreg_sexti8(i8* %in, i32 %reg) #0 { |
| 485 | entry: |
| 486 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 487 | %gep = getelementptr inbounds i8, i8* %in, i64 2047 |
| 488 | %load = load i8, i8* %gep |
| 489 | %ext = sext i8 %load to i16 |
| 490 | %build1 = insertelement <2 x i16> %reg.bc, i16 %ext, i32 0 |
| 491 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 492 | ret void |
| 493 | } |
| 494 | |
| 495 | ; GCN-LABEL: {{^}}load_private_lo_v2i16_reglo_vreg_nooff_zexti8: |
| 496 | ; GCN: s_waitcnt |
| 497 | ; GFX9-NEXT: buffer_load_ubyte_d16 v1, off, s[0:3], s4 offset:4094{{$}} |
| 498 | ; GFX9-NEXT: s_waitcnt |
| 499 | ; GFX9-NEXT: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, v1 |
| 500 | ; GFX9-NEXT: s_waitcnt |
| 501 | ; GFX9-NEXT: s_setpc_b64 |
| 502 | |
| 503 | ; VI: buffer_load_ubyte v0, off, s[0:3], s4 offset:4094{{$}} |
| 504 | define void @load_private_lo_v2i16_reglo_vreg_nooff_zexti8(i8* %in, i32 %reg) #0 { |
| 505 | entry: |
| 506 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 507 | %load = load volatile i8, i8* inttoptr (i32 4094 to i8*) |
| 508 | %ext = zext i8 %load to i16 |
| 509 | %build1 = insertelement <2 x i16> %reg.bc, i16 %ext, i32 0 |
| 510 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 511 | ret void |
| 512 | } |
| 513 | |
| 514 | ; GCN-LABEL: {{^}}load_private_lo_v2i16_reglo_vreg_nooff_sexti8: |
| 515 | ; GCN: s_waitcnt |
| 516 | ; GFX9-NEXT: buffer_load_sbyte_d16 v1, off, s[0:3], s4 offset:4094{{$}} |
| 517 | ; GFX9-NEXT: s_waitcnt |
| 518 | ; GFX9-NEXT: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, v1 |
| 519 | ; GFX9-NEXT: s_waitcnt |
| 520 | ; GFX9-NEXT: s_setpc_b64 |
| 521 | |
| 522 | ; VI: buffer_load_sbyte v0, off, s[0:3], s4 offset:4094{{$}} |
| 523 | define void @load_private_lo_v2i16_reglo_vreg_nooff_sexti8(i8* %in, i32 %reg) #0 { |
| 524 | entry: |
| 525 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 526 | %load = load volatile i8, i8* inttoptr (i32 4094 to i8*) |
| 527 | %ext = sext i8 %load to i16 |
| 528 | %build1 = insertelement <2 x i16> %reg.bc, i16 %ext, i32 0 |
| 529 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 530 | ret void |
| 531 | } |
| 532 | |
| 533 | ; GCN-LABEL: {{^}}load_private_lo_v2f16_reglo_vreg_nooff_zexti8: |
| 534 | ; GCN: s_waitcnt |
| 535 | ; GFX9-NEXT: buffer_load_ubyte_d16 v1, off, s[0:3], s4 offset:4094{{$}} |
| 536 | ; GFX9-NEXT: s_waitcnt |
| 537 | ; GFX9-NEXT: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, v1 |
| 538 | ; GFX9-NEXT: s_waitcnt |
| 539 | ; GFX9-NEXT: s_setpc_b64 |
| 540 | |
| 541 | ; VI: buffer_load_ubyte v0, off, s[0:3], s4 offset:4094{{$}} |
| 542 | define void @load_private_lo_v2f16_reglo_vreg_nooff_zexti8(i8* %in, i32 %reg) #0 { |
| 543 | entry: |
| 544 | %reg.bc = bitcast i32 %reg to <2 x half> |
| 545 | %load = load volatile i8, i8* inttoptr (i32 4094 to i8*) |
| 546 | %ext = zext i8 %load to i16 |
| 547 | %bc.ext = bitcast i16 %ext to half |
| 548 | %build1 = insertelement <2 x half> %reg.bc, half %bc.ext, i32 0 |
| 549 | store <2 x half> %build1, <2 x half> addrspace(1)* undef |
| 550 | ret void |
| 551 | } |
| 552 | |
| 553 | ; GCN-LABEL: {{^}}load_constant_lo_v2i16_reglo_vreg: |
| 554 | ; GCN: s_waitcnt |
| 555 | ; GFX9-NEXT: global_load_short_d16 v2, v[0:1], off offset:-4094 |
| 556 | ; GFX9-NEXT: s_waitcnt |
| 557 | ; GFX9-NEXT: global_store_dword |
| 558 | ; GFX9-NEXT: s_waitcnt |
| 559 | ; GFX9-NEXT: s_setpc_b64 |
| 560 | |
| 561 | ; VI: flat_load_ushort |
| 562 | define void @load_constant_lo_v2i16_reglo_vreg(i16 addrspace(2)* %in, i32 %reg) #0 { |
| 563 | entry: |
| 564 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 565 | %gep = getelementptr inbounds i16, i16 addrspace(2)* %in, i64 -2047 |
| 566 | %load = load i16, i16 addrspace(2)* %gep |
| 567 | %build1 = insertelement <2 x i16> %reg.bc, i16 %load, i32 0 |
| 568 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 569 | ret void |
| 570 | } |
| 571 | |
| 572 | ; GCN-LABEL: load_constant_lo_v2f16_reglo_vreg |
| 573 | ; GCN: s_waitcnt |
| 574 | ; GFX9-NEXT: global_load_short_d16 v2, v[0:1], off offset:-4094 |
| 575 | ; GFX9-NEXT: s_waitcnt |
| 576 | ; GFX9-NEXT: global_store_dword |
| 577 | ; GFX9-NEXT: s_waitcnt |
| 578 | ; GFX9-NEXT: s_setpc_b64 |
| 579 | |
| 580 | ; VI: flat_load_ushort |
| 581 | define void @load_constant_lo_v2f16_reglo_vreg(half addrspace(2)* %in, i32 %reg) #0 { |
| 582 | entry: |
| 583 | %reg.bc = bitcast i32 %reg to <2 x half> |
| 584 | %gep = getelementptr inbounds half, half addrspace(2)* %in, i64 -2047 |
| 585 | %load = load half, half addrspace(2)* %gep |
| 586 | %build1 = insertelement <2 x half> %reg.bc, half %load, i32 0 |
| 587 | store <2 x half> %build1, <2 x half> addrspace(1)* undef |
| 588 | ret void |
| 589 | } |
| 590 | |
Matt Arsenault | 4b7938c | 2017-11-13 23:24:26 +0000 | [diff] [blame^] | 591 | ; GCN-LABEL: {{^}}load_private_lo_v2i16_reglo_vreg_to_offset: |
| 592 | ; GFX9: buffer_store_dword |
| 593 | ; GFX9-NEXT: buffer_load_short_d16 v0, off, s[0:3], s5 offset:4094 |
| 594 | |
| 595 | ; VI: buffer_load_ushort v |
| 596 | define void @load_private_lo_v2i16_reglo_vreg_to_offset(i32 %reg) #0 { |
| 597 | entry: |
| 598 | %obj0 = alloca [10 x i32], align 4 |
| 599 | %obj1 = alloca [4096 x i16], align 2 |
| 600 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 601 | %bc = bitcast [10 x i32]* %obj0 to i32* |
| 602 | store volatile i32 123, i32* %bc |
| 603 | %gep = getelementptr inbounds [4096 x i16], [4096 x i16]* %obj1, i32 0, i32 2025 |
| 604 | %load = load volatile i16, i16* %gep |
| 605 | %build1 = insertelement <2 x i16> %reg.bc, i16 %load, i32 0 |
| 606 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 607 | ret void |
| 608 | } |
| 609 | |
| 610 | ; GCN-LABEL: {{^}}load_private_lo_v2i16_reglo_vreg_sexti8_to_offset: |
| 611 | ; GFX9: buffer_store_dword |
| 612 | ; GFX9-NEXT: buffer_load_sbyte_d16 v0, off, s[0:3], s5 offset:4095 |
| 613 | |
| 614 | ; VI: buffer_load_sbyte v |
| 615 | define void @load_private_lo_v2i16_reglo_vreg_sexti8_to_offset(i32 %reg) #0 { |
| 616 | entry: |
| 617 | %obj0 = alloca [10 x i32], align 4 |
| 618 | %obj1 = alloca [4096 x i8], align 2 |
| 619 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 620 | %bc = bitcast [10 x i32]* %obj0 to i32* |
| 621 | store volatile i32 123, i32* %bc |
| 622 | %gep = getelementptr inbounds [4096 x i8], [4096 x i8]* %obj1, i32 0, i32 4051 |
| 623 | %load = load volatile i8, i8* %gep |
| 624 | %load.ext = sext i8 %load to i16 |
| 625 | %build1 = insertelement <2 x i16> %reg.bc, i16 %load.ext, i32 0 |
| 626 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 627 | ret void |
| 628 | } |
| 629 | |
| 630 | ; GCN-LABEL: {{^}}load_private_lo_v2i16_reglo_vreg_zexti8_to_offset: |
| 631 | ; GFX9: buffer_store_dword |
| 632 | ; GFX9-NEXT: buffer_load_ubyte_d16 v0, off, s[0:3], s5 offset:4095 |
| 633 | |
| 634 | ; VI: buffer_load_ubyte v |
| 635 | define void @load_private_lo_v2i16_reglo_vreg_zexti8_to_offset(i32 %reg) #0 { |
| 636 | entry: |
| 637 | %obj0 = alloca [10 x i32], align 4 |
| 638 | %obj1 = alloca [4096 x i8], align 2 |
| 639 | %reg.bc = bitcast i32 %reg to <2 x i16> |
| 640 | %bc = bitcast [10 x i32]* %obj0 to i32* |
| 641 | store volatile i32 123, i32* %bc |
| 642 | %gep = getelementptr inbounds [4096 x i8], [4096 x i8]* %obj1, i32 0, i32 4051 |
| 643 | %load = load volatile i8, i8* %gep |
| 644 | %load.ext = zext i8 %load to i16 |
| 645 | %build1 = insertelement <2 x i16> %reg.bc, i16 %load.ext, i32 0 |
| 646 | store <2 x i16> %build1, <2 x i16> addrspace(1)* undef |
| 647 | ret void |
| 648 | } |
| 649 | |
Matt Arsenault | e1cd482 | 2017-11-13 00:22:09 +0000 | [diff] [blame] | 650 | attributes #0 = { nounwind } |