235 lines · plain
1# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 52# RUN: llc -mtriple=amdgcn -mcpu=gfx90a -run-pass si-insert-waitcnts %s -o - | FileCheck -check-prefix=GFX9 %s3# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -run-pass si-insert-waitcnts %s -o - | FileCheck -check-prefix=GFX12 %s4 5# There shall be no S_WAITCNT between two stores.6 7---8name: spill_vgpr_tuple9 10body: |11 bb.0:12 liveins: $vgpr0_vgpr1, $sgpr76_sgpr77_sgpr78_sgpr7913 14 ; GFX9-LABEL: name: spill_vgpr_tuple15 ; GFX9: liveins: $vgpr0_vgpr1, $sgpr76_sgpr77_sgpr78_sgpr7916 ; GFX9-NEXT: {{ $}}17 ; GFX9-NEXT: S_WAITCNT 018 ; GFX9-NEXT: $vgpr64_vgpr65 = V_MOV_B64_e32 $vgpr0_vgpr1, implicit $exec19 ; GFX9-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr64, $sgpr76_sgpr77_sgpr78_sgpr79, 0, 672, 0, 0, implicit $exec, implicit-def $vgpr64_vgpr65, implicit $vgpr64_vgpr6520 ; GFX9-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr65, $sgpr76_sgpr77_sgpr78_sgpr79, 0, 676, 0, 0, implicit $exec, implicit $vgpr64_vgpr6521 ; GFX9-NEXT: S_ENDPGM 022 ;23 ; GFX12-LABEL: name: spill_vgpr_tuple24 ; GFX12: liveins: $vgpr0_vgpr1, $sgpr76_sgpr77_sgpr78_sgpr7925 ; GFX12-NEXT: {{ $}}26 ; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 027 ; GFX12-NEXT: S_WAIT_EXPCNT 028 ; GFX12-NEXT: S_WAIT_SAMPLECNT 029 ; GFX12-NEXT: S_WAIT_BVHCNT 030 ; GFX12-NEXT: S_WAIT_KMCNT 031 ; GFX12-NEXT: $vgpr64_vgpr65 = V_MOV_B64_e32 $vgpr0_vgpr1, implicit $exec32 ; GFX12-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr64, $sgpr76_sgpr77_sgpr78_sgpr79, 0, 672, 0, 0, implicit $exec, implicit-def $vgpr64_vgpr65, implicit $vgpr64_vgpr6533 ; GFX12-NEXT: BUFFER_STORE_DWORD_OFFSET $vgpr65, $sgpr76_sgpr77_sgpr78_sgpr79, 0, 676, 0, 0, implicit $exec, implicit $vgpr64_vgpr6534 ; GFX12-NEXT: S_ENDPGM 035 $vgpr64_vgpr65 = V_MOV_B64_e32 $vgpr0_vgpr1, implicit $exec36 BUFFER_STORE_DWORD_OFFSET killed $vgpr64, $sgpr76_sgpr77_sgpr78_sgpr79, 0, 672, 0, 0, implicit $exec, implicit-def $vgpr64_vgpr65, implicit $vgpr64_vgpr6537 BUFFER_STORE_DWORD_OFFSET $vgpr65, $sgpr76_sgpr77_sgpr78_sgpr79, 0, 676, 0, 0, implicit $exec, implicit $vgpr64_vgpr6538 S_ENDPGM 039...40 41# Make sure that while ignoring implicit operands we will not ignore implicit $vcc on VALU42 43---44name: load_vcc_wait45 46body: |47 bb.0:48 liveins: $vgpr0, $sgpr10_sgpr1149 50 ; GFX9-LABEL: name: load_vcc_wait51 ; GFX9: liveins: $vgpr0, $sgpr10_sgpr1152 ; GFX9-NEXT: {{ $}}53 ; GFX9-NEXT: S_WAITCNT 054 ; GFX9-NEXT: $vcc_lo = S_LOAD_DWORD_IMM $sgpr10_sgpr11, 0, 055 ; GFX9-NEXT: S_WAITCNT 4927956 ; GFX9-NEXT: $vgpr1 = V_ADDC_U32_e32 0, $vgpr0, implicit-def $vcc, implicit $vcc, implicit $exec57 ; GFX9-NEXT: S_ENDPGM 058 ;59 ; GFX12-LABEL: name: load_vcc_wait60 ; GFX12: liveins: $vgpr0, $sgpr10_sgpr1161 ; GFX12-NEXT: {{ $}}62 ; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 063 ; GFX12-NEXT: S_WAIT_EXPCNT 064 ; GFX12-NEXT: S_WAIT_SAMPLECNT 065 ; GFX12-NEXT: S_WAIT_BVHCNT 066 ; GFX12-NEXT: S_WAIT_KMCNT 067 ; GFX12-NEXT: $vcc_lo = S_LOAD_DWORD_IMM $sgpr10_sgpr11, 0, 068 ; GFX12-NEXT: S_WAIT_KMCNT 069 ; GFX12-NEXT: $vgpr1 = V_ADDC_U32_e32 0, $vgpr0, implicit-def $vcc_lo, implicit $vcc_lo, implicit $exec70 ; GFX12-NEXT: S_ENDPGM 071 $vcc_lo = S_LOAD_DWORD_IMM $sgpr10_sgpr11, 0, 072 $vgpr1 = V_ADDC_U32_e32 0, $vgpr0, implicit-def $vcc, implicit $vcc, implicit $exec73 S_ENDPGM 074...75 76# Make sure that while ignoring implicit operands we will not ignore implicit $flat_src on FLAT77 78---79name: load_flat_scr_lo_flat_load_wait80 81body: |82 bb.0:83 liveins: $sgpr10_sgpr11, $vgpr0_vgpr184 85 ; GFX9-LABEL: name: load_flat_scr_lo_flat_load_wait86 ; GFX9: liveins: $sgpr10_sgpr11, $vgpr0_vgpr187 ; GFX9-NEXT: {{ $}}88 ; GFX9-NEXT: S_WAITCNT 089 ; GFX9-NEXT: $flat_scr_lo = S_LOAD_DWORD_IMM $sgpr10_sgpr11, 0, 090 ; GFX9-NEXT: S_WAITCNT 4927991 ; GFX9-NEXT: $vgpr2 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr92 ; GFX9-NEXT: S_ENDPGM 093 ;94 ; GFX12-LABEL: name: load_flat_scr_lo_flat_load_wait95 ; GFX12: liveins: $sgpr10_sgpr11, $vgpr0_vgpr196 ; GFX12-NEXT: {{ $}}97 ; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 098 ; GFX12-NEXT: S_WAIT_EXPCNT 099 ; GFX12-NEXT: S_WAIT_SAMPLECNT 0100 ; GFX12-NEXT: S_WAIT_BVHCNT 0101 ; GFX12-NEXT: S_WAIT_KMCNT 0102 ; GFX12-NEXT: $flat_scr_lo = S_LOAD_DWORD_IMM $sgpr10_sgpr11, 0, 0103 ; GFX12-NEXT: S_WAIT_KMCNT 0104 ; GFX12-NEXT: $vgpr2 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr105 ; GFX12-NEXT: S_ENDPGM 0106 $flat_scr_lo = S_LOAD_DWORD_IMM $sgpr10_sgpr11, 0, 0107 $vgpr2 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr108 S_ENDPGM 0109...110 111---112name: load_flat_scr_lo_scratch_store_wait113 114body: |115 bb.0:116 liveins: $sgpr10_sgpr11, $vgpr0, $sgpr32117 118 ; GFX9-LABEL: name: load_flat_scr_lo_scratch_store_wait119 ; GFX9: liveins: $sgpr10_sgpr11, $vgpr0, $sgpr32120 ; GFX9-NEXT: {{ $}}121 ; GFX9-NEXT: S_WAITCNT 0122 ; GFX9-NEXT: $flat_scr_hi = S_LOAD_DWORD_IMM $sgpr10_sgpr11, 0, 0123 ; GFX9-NEXT: S_WAITCNT 49279124 ; GFX9-NEXT: SCRATCH_STORE_DWORD_SADDR $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr125 ; GFX9-NEXT: S_ENDPGM 0126 ;127 ; GFX12-LABEL: name: load_flat_scr_lo_scratch_store_wait128 ; GFX12: liveins: $sgpr10_sgpr11, $vgpr0, $sgpr32129 ; GFX12-NEXT: {{ $}}130 ; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0131 ; GFX12-NEXT: S_WAIT_EXPCNT 0132 ; GFX12-NEXT: S_WAIT_SAMPLECNT 0133 ; GFX12-NEXT: S_WAIT_BVHCNT 0134 ; GFX12-NEXT: S_WAIT_KMCNT 0135 ; GFX12-NEXT: $flat_scr_hi = S_LOAD_DWORD_IMM $sgpr10_sgpr11, 0, 0136 ; GFX12-NEXT: S_WAIT_KMCNT 0137 ; GFX12-NEXT: SCRATCH_STORE_DWORD_SADDR $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr138 ; GFX12-NEXT: S_ENDPGM 0139 $flat_scr_hi = S_LOAD_DWORD_IMM $sgpr10_sgpr11, 0, 0140 SCRATCH_STORE_DWORD_SADDR $vgpr0, $sgpr32, 0, 0, implicit $exec, implicit $flat_scr141 S_ENDPGM 0142...143 144# Check that implicit spill defs do not force wait to zero on the first store145 146---147name: spill_load_store148 149body: |150 bb.0:151 liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32152 153 ; GFX9-LABEL: name: spill_load_store154 ; GFX9: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32155 ; GFX9-NEXT: {{ $}}156 ; GFX9-NEXT: S_WAITCNT 0157 ; GFX9-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3158 ; GFX9-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec159 ; GFX9-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec160 ; GFX9-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3161 ; GFX9-NEXT: S_WAITCNT 3955162 ; GFX9-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3163 ; GFX9-NEXT: S_WAITCNT 3955164 ; GFX9-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec165 ; GFX9-NEXT: S_WAITCNT 3955166 ; GFX9-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec167 ; GFX9-NEXT: S_WAITCNT 3955168 ; GFX9-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3169 ; GFX9-NEXT: S_ENDPGM 0170 ;171 ; GFX12-LABEL: name: spill_load_store172 ; GFX12: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32173 ; GFX12-NEXT: {{ $}}174 ; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0175 ; GFX12-NEXT: S_WAIT_EXPCNT 0176 ; GFX12-NEXT: S_WAIT_SAMPLECNT 0177 ; GFX12-NEXT: S_WAIT_BVHCNT 0178 ; GFX12-NEXT: S_WAIT_KMCNT 0179 ; GFX12-NEXT: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3180 ; GFX12-NEXT: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec181 ; GFX12-NEXT: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec182 ; GFX12-NEXT: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3183 ; GFX12-NEXT: S_WAIT_LOADCNT 3184 ; GFX12-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3185 ; GFX12-NEXT: S_WAIT_LOADCNT 2186 ; GFX12-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec187 ; GFX12-NEXT: S_WAIT_LOADCNT 1188 ; GFX12-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec189 ; GFX12-NEXT: S_WAIT_LOADCNT 0190 ; GFX12-NEXT: BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3191 ; GFX12-NEXT: S_ENDPGM 0192 $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 8, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3193 $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 12, 0, 0, implicit $exec194 $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 16, 0, 0, implicit $exec195 $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 20, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3196 BUFFER_STORE_DWORD_OFFSET killed $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 40, 0, 0, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $vgpr0_vgpr1_vgpr2_vgpr3197 BUFFER_STORE_DWORD_OFFSET killed $vgpr1, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 44, 0, 0, implicit $exec198 BUFFER_STORE_DWORD_OFFSET killed $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 48, 0, 0, implicit $exec199 BUFFER_STORE_DWORD_OFFSET killed $vgpr3, $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr32, 52, 0, 0, implicit $exec, implicit killed $vgpr0_vgpr1_vgpr2_vgpr3200 S_ENDPGM 0201...202 203# Make sure we have wait to mitigate WAW on gfx12204 205---206name: scratch_load_waw207body: |208 bb.0.entry:209 liveins: $vgpr0, $sgpr0210 211 ; GFX9-LABEL: name: scratch_load_waw212 ; GFX9: liveins: $vgpr0, $sgpr0213 ; GFX9-NEXT: {{ $}}214 ; GFX9-NEXT: S_WAITCNT 0215 ; GFX9-NEXT: $vgpr2 = SCRATCH_LOAD_DWORD $vgpr0, 0, 0, implicit $exec, implicit $flat_scr216 ; GFX9-NEXT: $vgpr2 = SCRATCH_LOAD_SHORT_D16_HI_SADDR $sgpr0, 0, 0, $vgpr2, implicit $exec, implicit $flat_scr217 ; GFX9-NEXT: S_ENDPGM 0218 ;219 ; GFX12-LABEL: name: scratch_load_waw220 ; GFX12: liveins: $vgpr0, $sgpr0221 ; GFX12-NEXT: {{ $}}222 ; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0223 ; GFX12-NEXT: S_WAIT_EXPCNT 0224 ; GFX12-NEXT: S_WAIT_SAMPLECNT 0225 ; GFX12-NEXT: S_WAIT_BVHCNT 0226 ; GFX12-NEXT: S_WAIT_KMCNT 0227 ; GFX12-NEXT: $vgpr2 = SCRATCH_LOAD_DWORD $vgpr0, 0, 0, implicit $exec, implicit $flat_scr228 ; GFX12-NEXT: S_WAIT_LOADCNT 0229 ; GFX12-NEXT: $vgpr2 = SCRATCH_LOAD_SHORT_D16_HI_SADDR $sgpr0, 0, 0, $vgpr2, implicit $exec, implicit $flat_scr230 ; GFX12-NEXT: S_ENDPGM 0231 $vgpr2 = SCRATCH_LOAD_DWORD $vgpr0, 0, 0, implicit $exec, implicit $flat_scr232 $vgpr2 = SCRATCH_LOAD_SHORT_D16_HI_SADDR $sgpr0, 0, 0, $vgpr2, implicit $exec, implicit $flat_scr233 S_ENDPGM 0234...235