294 lines · plain
1# RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -run-pass=si-peephole-sdwa -verify-machineinstrs -o - %s | FileCheck -check-prefix=GFX1010 -check-prefix=GCN %s2 3# GCN-LABEL: {{^}}name: vop1_instructions4 5# GFX1010: %{{[0-9]+}}:vgpr_32 = V_MOV_B32_sdwa 0, %{{[0-9]+}}, 0, 5, 0, 5, implicit $exec6# GFX1010: %{{[0-9]+}}:vgpr_32 = V_FRACT_F32_sdwa 0, %{{[0-9]+}}, 0, 0, 5, 0, 6, implicit $mode, implicit $exec7# GFX1010: %{{[0-9]+}}:vgpr_32 = V_SIN_F32_sdwa 0, %{{[0-9]+}}, 0, 0, 5, 0, 5, implicit $mode, implicit $exec8# GFX1010: %{{[0-9]+}}:vgpr_32 = V_CVT_U32_F32_sdwa 0, %{{[0-9]+}}, 0, 5, 0, 5, implicit $mode, implicit $exec9# GFX1010: %{{[0-9]+}}:vgpr_32 = V_CVT_F32_I32_sdwa 0, %{{[0-9]+}}, 0, 0, 5, 0, 6, implicit $mode, implicit $exec10 11# GFX1010: %{{[0-9]+}}:vgpr_32 = V_MOV_B32_sdwa 0, %{{[0-9]+}}, 0, 6, 0, 5, implicit $exec12# GFX1010: %{{[0-9]+}}:vgpr_32 = V_FRACT_F32_sdwa 0, %{{[0-9]+}}, 0, 0, 5, 0, 6, implicit $mode, implicit $exec13# GFX1010: %{{[0-9]+}}:vgpr_32 = V_SIN_F32_sdwa 0, %{{[0-9]+}}, 0, 0, 5, 0, 5, implicit $mode, implicit $exec14# GFX1010: %{{[0-9]+}}:vgpr_32 = V_CVT_U32_F32_sdwa 0, %{{[0-9]+}}, 0, 5, 0, 5, implicit $mode, implicit $exec15# GFX1010: %{{[0-9]+}}:vgpr_32 = V_CVT_F32_I32_sdwa 0, %{{[0-9]+}}, 0, 0, 5, 0, 6, implicit $mode, implicit $exec16 17# GFX1010: %{{[0-9]+}}:vgpr_32 = V_FRACT_F32_sdwa 1, %{{[0-9]+}}, 0, 0, 5, 0, 5, implicit $mode, implicit $exec18# GFX1010: %{{[0-9]+}}:vgpr_32 = V_SIN_F32_sdwa 0, %{{[0-9]+}}, 1, 0, 5, 0, 5, implicit $mode, implicit $exec19# GFX1010: %{{[0-9]+}}:vgpr_32 = V_CVT_U32_F32_sdwa 1, %{{[0-9]+}}, 0, 5, 0, 5, implicit $mode, implicit $exec20# GFX1010: %{{[0-9]+}}:vgpr_32 = V_CVT_F32_I32_sdwa 0, %{{[0-9]+}}, 0, 1, 5, 0, 5, implicit $mode, implicit $exec21 22---23name: vop1_instructions24tracksRegLiveness: true25registers:26 - { id: 0, class: vreg_64 }27 - { id: 1, class: vreg_64 }28 - { id: 2, class: sreg_64 }29 - { id: 3, class: vgpr_32 }30 - { id: 4, class: sreg_32_xm0 }31 - { id: 5, class: sreg_32_xm0 }32 - { id: 6, class: sreg_32_xm0 }33 - { id: 7, class: sreg_32_xm0 }34 - { id: 8, class: sreg_32 }35 - { id: 9, class: vgpr_32 }36 - { id: 10, class: vgpr_32 }37 - { id: 11, class: vgpr_32 }38 - { id: 12, class: vgpr_32 }39 - { id: 13, class: vgpr_32 }40 - { id: 14, class: vgpr_32 }41 - { id: 15, class: vgpr_32 }42 - { id: 16, class: vgpr_32 }43 - { id: 17, class: vgpr_32 }44 - { id: 18, class: vgpr_32 }45 - { id: 19, class: vgpr_32 }46 - { id: 20, class: vgpr_32 }47 - { id: 21, class: vgpr_32 }48 - { id: 22, class: vgpr_32 }49 - { id: 23, class: vgpr_32 }50 - { id: 24, class: vgpr_32 }51 - { id: 25, class: vgpr_32 }52 - { id: 26, class: vgpr_32 }53 - { id: 27, class: vgpr_32 }54 - { id: 28, class: vgpr_32 }55 - { id: 29, class: vgpr_32 }56 - { id: 30, class: vgpr_32 }57 - { id: 31, class: vgpr_32 }58 - { id: 32, class: vgpr_32 }59 - { id: 33, class: vgpr_32 }60 - { id: 34, class: vgpr_32 }61 - { id: 35, class: vgpr_32 }62 - { id: 36, class: vgpr_32 }63 - { id: 37, class: vgpr_32 }64 - { id: 38, class: vgpr_32 }65 - { id: 39, class: vgpr_32 }66 - { id: 40, class: vgpr_32 }67 - { id: 41, class: vgpr_32 }68 - { id: 42, class: vgpr_32 }69 - { id: 43, class: vgpr_32 }70 - { id: 44, class: vgpr_32 }71 - { id: 45, class: vgpr_32 }72 - { id: 46, class: vgpr_32 }73 - { id: 47, class: vgpr_32 }74 - { id: 48, class: vgpr_32 }75 - { id: 100, class: vgpr_32 }76body: |77 bb.0:78 liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $sgpr30_sgpr3179 80 %2 = COPY $sgpr30_sgpr3181 %1 = COPY $vgpr2_vgpr382 %0 = COPY $vgpr0_vgpr183 %3 = FLAT_LOAD_DWORD %1, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32))84 85 %5 = S_MOV_B32 6553586 %6 = S_MOV_B32 6553587 88 %10 = V_LSHRREV_B32_e64 16, %3, implicit $exec89 %11 = V_MOV_B32_e32 %10, implicit $exec90 %12 = V_LSHLREV_B32_e64 16, %11, implicit $exec91 %14 = V_FRACT_F32_e32 123, implicit $mode, implicit $exec92 %15 = V_LSHLREV_B32_e64 16, %14, implicit $exec93 %16 = V_LSHRREV_B32_e64 16, %15, implicit $exec94 %17 = V_SIN_F32_e32 %16, implicit $mode, implicit $exec95 %18 = V_LSHLREV_B32_e64 16, %17, implicit $exec96 %19 = V_LSHRREV_B32_e64 16, %18, implicit $exec97 %20 = V_CVT_U32_F32_e32 %19, implicit $mode, implicit $exec98 %21 = V_LSHLREV_B32_e64 16, %20, implicit $exec99 %23 = V_CVT_F32_I32_e32 123, implicit $mode, implicit $exec100 %24 = V_LSHLREV_B32_e64 16, %23, implicit $exec101 102 %25 = V_LSHRREV_B32_e64 16, %3, implicit $exec103 %26 = V_MOV_B32_e64 %25, implicit $exec104 %26 = V_LSHLREV_B32_e64 16, %26, implicit $exec105 %27 = V_FRACT_F32_e64 0, %6, 0, 0, implicit $mode, implicit $exec106 %28 = V_LSHLREV_B32_e64 16, %27, implicit $exec107 %29 = V_LSHRREV_B32_e64 16, %28, implicit $exec108 %30 = V_SIN_F32_e64 0, %29, 0, 0, implicit $mode, implicit $exec109 %31 = V_LSHLREV_B32_e64 16, %30, implicit $exec110 %32 = V_LSHRREV_B32_e64 16, %31, implicit $exec111 %33 = V_CVT_U32_F32_e64 0, %32, 0, 0, implicit $mode, implicit $exec112 %34 = V_LSHLREV_B32_e64 16, %33, implicit $exec113 %35 = V_CVT_F32_I32_e64 %6, 0, 0, implicit $mode, implicit $exec114 %36 = V_LSHLREV_B32_e64 16, %35, implicit $exec115 116 117 %37 = V_LSHRREV_B32_e64 16, %36, implicit $exec118 %38 = V_FRACT_F32_e64 1, %37, 0, 0, implicit $mode, implicit $exec119 %39 = V_LSHLREV_B32_e64 16, %38, implicit $exec120 %40 = V_LSHRREV_B32_e64 16, %39, implicit $exec121 %41 = V_SIN_F32_e64 0, %40, 1, 0, implicit $mode, implicit $exec122 %42 = V_LSHLREV_B32_e64 16, %41, implicit $exec123 %43 = V_LSHRREV_B32_e64 16, %42, implicit $exec124 %44 = V_CVT_U32_F32_e64 1, %43, 0, 0, implicit $mode, implicit $exec125 %45 = V_LSHLREV_B32_e64 16, %44, implicit $exec126 %46 = V_LSHRREV_B32_e64 16, %45, implicit $exec127 %47 = V_CVT_F32_I32_e64 %46, 0, 1, implicit $mode, implicit $exec128 %48 = V_LSHLREV_B32_e64 16, %47, implicit $exec129 130 131 %100 = V_MOV_B32_e32 %48, implicit $exec132 133 FLAT_STORE_DWORD %0, %100, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32))134 $sgpr30_sgpr31 = COPY %2135 S_SETPC_B64_return $sgpr30_sgpr31136 137...138---139# GCN-LABEL: {{^}}name: vop2_instructions140 141# GFX1010: %{{[0-9]+}}:vgpr_32 = V_AND_B32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 5, 0, 6, 5, implicit $exec142# GFX1010: %{{[0-9]+}}:vgpr_32 = V_ADD_F32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 0, 5, 0, 5, 1, implicit $mode, implicit $exec143# GFX1010: %{{[0-9]+}}:vgpr_32 = V_SUB_F16_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 0, 6, 0, 5, 1, implicit $mode, implicit $exec144# GFX1010: %{{[0-9]+}}:vgpr_32 = V_FMAC_F32_e32 %{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}}, implicit $mode, implicit $exec145# GFX1010: %{{[0-9]+}}:vgpr_32 = V_FMAC_F16_e32 %{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}}, implicit $mode, implicit $exec146 147# GFX1010: %{{[0-9]+}}:vgpr_32 = V_AND_B32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 5, 0, 6, 5, implicit $exec148# GFX1010: %{{[0-9]+}}:vgpr_32 = V_ADD_F32_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 0, 5, 0, 5, 1, implicit $mode, implicit $exec149# GFX1010: %{{[0-9]+}}:vgpr_32 = V_SUB_F16_sdwa 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 0, 5, 0, 6, 1, implicit $mode, implicit $exec150# GFX1010: %{{[0-9]+}}:vgpr_32 = V_FMAC_F32_e64 0, 23, 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 0, implicit $mode, implicit $exec151# GFX1010: %{{[0-9]+}}:vgpr_32 = V_FMAC_F16_e64 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, %{{[0-9]+}}, 0, 0, implicit $mode, implicit $exec152 153# GFX1010: %{{[0-9]+}}:vgpr_32 = V_ADD_F32_sdwa 0, %{{[0-9]+}}, 1, %{{[0-9]+}}, 0, 0, 5, 0, 5, 1, implicit $mode, implicit $exec154# GFX1010: %{{[0-9]+}}:vgpr_32 = V_SUB_F16_sdwa 1, %{{[0-9]+}}, 1, %{{[0-9]+}}, 0, 0, 5, 0, 6, 1, implicit $mode, implicit $exec155# GFX1010: %{{[0-9]+}}:vgpr_32 = V_FMAC_F32_e64 1, 23, 1, %{{[0-9]+}}, 1, %{{[0-9]+}}, 1, 0, implicit $mode, implicit $exec156# GFX1010: %{{[0-9]+}}:vgpr_32 = V_FMAC_F16_e64 1, %{{[0-9]+}}, 1, %{{[0-9]+}}, 1, %{{[0-9]+}}, 0, 2, implicit $mode, implicit $exec157 158name: vop2_instructions159tracksRegLiveness: true160registers:161 - { id: 0, class: vreg_64 }162 - { id: 1, class: vreg_64 }163 - { id: 2, class: sreg_64 }164 - { id: 3, class: vgpr_32 }165 - { id: 4, class: sreg_32_xm0 }166 - { id: 5, class: sreg_32_xm0 }167 - { id: 6, class: sreg_32_xm0 }168 - { id: 7, class: sreg_32_xm0 }169 - { id: 8, class: sreg_32 }170 - { id: 9, class: vgpr_32 }171 - { id: 10, class: vgpr_32 }172 - { id: 11, class: vgpr_32 }173 - { id: 12, class: vgpr_32 }174 - { id: 13, class: vgpr_32 }175 - { id: 14, class: vgpr_32 }176 - { id: 15, class: vgpr_32 }177 - { id: 16, class: vgpr_32 }178 - { id: 17, class: vgpr_32 }179 - { id: 18, class: vgpr_32 }180 - { id: 19, class: vgpr_32 }181 - { id: 20, class: vgpr_32 }182 - { id: 21, class: vgpr_32 }183 - { id: 22, class: vgpr_32 }184 - { id: 23, class: vgpr_32 }185 - { id: 24, class: vgpr_32 }186 - { id: 25, class: vgpr_32 }187 - { id: 26, class: vgpr_32 }188 - { id: 27, class: vgpr_32 }189 - { id: 28, class: vgpr_32 }190 - { id: 29, class: vgpr_32 }191 - { id: 30, class: vgpr_32 }192 - { id: 31, class: vgpr_32 }193 - { id: 32, class: vgpr_32 }194 - { id: 33, class: vgpr_32 }195 - { id: 34, class: vgpr_32 }196 - { id: 35, class: vgpr_32 }197 - { id: 36, class: vgpr_32 }198 - { id: 37, class: vgpr_32 }199 - { id: 38, class: vgpr_32 }200 - { id: 39, class: vgpr_32 }201 - { id: 40, class: vgpr_32 }202 - { id: 41, class: vgpr_32 }203 - { id: 42, class: vgpr_32 }204 - { id: 43, class: vgpr_32 }205 - { id: 44, class: vgpr_32 }206 - { id: 45, class: vgpr_32 }207 - { id: 46, class: vgpr_32 }208 - { id: 47, class: vgpr_32 }209 - { id: 48, class: vgpr_32 }210 - { id: 49, class: vgpr_32 }211 - { id: 50, class: vgpr_32 }212 - { id: 51, class: vgpr_32 }213 - { id: 52, class: vgpr_32 }214 - { id: 53, class: vgpr_32 }215 - { id: 54, class: vgpr_32 }216 - { id: 55, class: vgpr_32 }217 - { id: 56, class: vgpr_32 }218 - { id: 57, class: vgpr_32 }219 - { id: 58, class: vgpr_32 }220 - { id: 59, class: vgpr_32 }221 - { id: 60, class: vgpr_32 }222 - { id: 100, class: vgpr_32 }223body: |224 bb.0:225 liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $sgpr30_sgpr31226 227 %2 = COPY $sgpr30_sgpr31228 %1 = COPY $vgpr2_vgpr3229 %0 = COPY $vgpr0_vgpr1230 %3 = FLAT_LOAD_DWORD %1, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32))231 232 %5 = S_MOV_B32 65535233 %6 = S_MOV_B32 65535234 235 %11 = V_LSHRREV_B32_e64 16, %3, implicit $exec236 %12 = V_AND_B32_e32 %6, %11, implicit $exec237 %13 = V_LSHLREV_B32_e64 16, %12, implicit $exec238 %14 = V_LSHRREV_B32_e64 16, %13, implicit $exec239 %15 = V_BFE_U32_e64 %13, 8, 8, implicit $exec240 %16 = V_ADD_F32_e32 %14, %15, implicit $mode, implicit $exec241 %17 = V_LSHLREV_B32_e64 16, %16, implicit $exec242 %18 = V_LSHRREV_B32_e64 16, %17, implicit $exec243 %19 = V_BFE_U32_e64 %17, 8, 8, implicit $exec244 %20 = V_SUB_F16_e32 %18, %19, implicit $mode, implicit $exec245 %21 = V_LSHLREV_B32_e64 16, %20, implicit $exec246 %22 = V_BFE_U32_e64 %20, 8, 8, implicit $exec247 %23 = V_FMAC_F32_e32 %21, %22, %22, implicit $mode, implicit $exec248 %24 = V_LSHLREV_B32_e64 16, %23, implicit $exec249 %25 = V_LSHRREV_B32_e64 16, %24, implicit $exec250 %26 = V_BFE_U32_e64 %24, 8, 8, implicit $exec251 %27 = V_FMAC_F16_e32 %25, %26, %26, implicit $mode, implicit $exec252 %28 = V_LSHLREV_B32_e64 16, %27, implicit $exec253 254 %29 = V_LSHRREV_B32_e64 16, %28, implicit $exec255 %30 = V_AND_B32_e64 23, %29, implicit $exec256 %31 = V_LSHLREV_B32_e64 16, %30, implicit $exec257 %32 = V_LSHRREV_B32_e64 16, %31, implicit $exec258 %33 = V_BFE_U32_e64 %31, 8, 8, implicit $exec259 %34 = V_ADD_F32_e64 0, %32, 0, %33, 0, 0, implicit $mode, implicit $exec260 %35 = V_LSHLREV_B32_e64 16, %34, implicit $exec261 %37 = V_BFE_U32_e64 %35, 8, 8, implicit $exec262 %38 = V_SUB_F16_e64 0, 23, 0, %37, 0, 0, implicit $mode, implicit $exec263 %39 = V_LSHLREV_B32_e64 16, %38, implicit $exec264 %40 = V_BFE_U32_e64 %39, 8, 8, implicit $exec265 %41 = V_FMAC_F32_e64 0, 23, 0, %40, 0, %40, 0, 0, implicit $mode, implicit $exec266 %42 = V_LSHLREV_B32_e64 16, %41, implicit $exec267 %43 = V_LSHRREV_B32_e64 16, %42, implicit $exec268 %44 = V_BFE_U32_e64 %42, 8, 8, implicit $exec269 %45 = V_FMAC_F16_e64 0, %43, 0, %44, 0, %44, 0, 0, implicit $mode, implicit $exec270 %46 = V_LSHLREV_B32_e64 16, %45, implicit $exec271 272 %47 = V_LSHRREV_B32_e64 16, %46, implicit $exec273 %48 = V_BFE_U32_e64 %46, 8, 8, implicit $exec274 %49 = V_ADD_F32_e64 0, %47, 1, %48, 0, 0, implicit $mode, implicit $exec275 %50 = V_LSHLREV_B32_e64 16, %49, implicit $exec276 %51 = V_BFE_U32_e64 %50, 8, 8, implicit $exec277 %52 = V_SUB_F16_e64 1, 23, 1, %51, 0, 0, implicit $mode, implicit $exec278 %53 = V_LSHLREV_B32_e64 16, %52, implicit $exec279 %54 = V_BFE_U32_e64 %53, 8, 8, implicit $exec280 %55 = V_FMAC_F32_e64 1, 23, 1, %54, 1, %54, 1, 0, implicit $mode, implicit $exec281 %56 = V_LSHLREV_B32_e64 16, %55, implicit $exec282 %57 = V_LSHRREV_B32_e64 16, %56, implicit $exec283 %58 = V_BFE_U32_e64 %56, 8, 8, implicit $exec284 %59 = V_FMAC_F16_e64 1, %57, 1, %58, 1, %58, 0, 2, implicit $mode, implicit $exec285 %60 = V_LSHLREV_B32_e64 16, %59, implicit $exec286 287 %100 = V_MOV_B32_e32 %60, implicit $exec288 289 FLAT_STORE_DWORD %0, %100, 0, 0, implicit $exec, implicit $flat_scr :: (store (s32))290 $sgpr30_sgpr31 = COPY %2291 S_SETPC_B64_return $sgpr30_sgpr31292 293...294