# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -march=amdgcn -mcpu=gfx908 -verify-machineinstrs -run-pass si-insert-waitcnts -o - %s | FileCheck -check-prefixes=GFX9 %s

---
name:            test_waitcnt_preexisting_lgkmcnt_unmodified
body:             |
  bb.0:
    liveins: $vgpr0

    ; GFX9-LABEL: name: test_waitcnt_preexisting_lgkmcnt_unmodified
    ; GFX9: S_WAITCNT 0
    ; GFX9: $vgpr0_vgpr1 = DS_READ2_B32 $vgpr0, 0, 1, 0, implicit $m0, implicit $exec
    ; GFX9: S_WAITCNT 49279
    ; GFX9: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_WAITCNT 112
    ; GFX9: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_ENDPGM 0
    $vgpr0_vgpr1 = DS_READ2_B32 $vgpr0, 0, 1, 0, implicit $m0, implicit $exec
    S_WAITCNT 49279
    $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    S_ENDPGM 0
...

---
name:            test_waitcnt_preexisting_vmcnt_unmodified
body:             |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; GFX9-LABEL: name: test_waitcnt_preexisting_vmcnt_unmodified
    ; GFX9: S_WAITCNT 0
    ; GFX9: $vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
    ; GFX9: S_WAITCNT 3952
    ; GFX9: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_WAITCNT 112
    ; GFX9: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_ENDPGM 0
    $vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
    S_WAITCNT 3952
    $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    S_ENDPGM 0
...

# Respect preexisting waitcnt and add required wait.

---
name:            test_waitcnt_preexisting_vmcnt_needs_lgkmcnt
body:             |
  bb.0:
    liveins: $vgpr0

    ; GFX9-LABEL: name: test_waitcnt_preexisting_vmcnt_needs_lgkmcnt
    ; GFX9: S_WAITCNT 0
    ; GFX9: $vgpr0_vgpr1 = DS_READ2_B32 $vgpr0, 0, 1, 0, implicit $m0, implicit $exec
    ; GFX9: S_WAITCNT 112
    ; GFX9: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_WAITCNT 112
    ; GFX9: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_ENDPGM 0
    $vgpr0_vgpr1 = DS_READ2_B32 $vgpr0, 0, 1, 0, implicit $m0, implicit $exec
    S_WAITCNT 3952
    $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    S_ENDPGM 0
...

---
name:            test_waitcnt_preexisting_lgkmcnt_needs_vmcnt
body:             |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; GFX9-LABEL: name: test_waitcnt_preexisting_lgkmcnt_needs_vmcnt
    ; GFX9: S_WAITCNT 0
    ; GFX9: $vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
    ; GFX9: S_WAITCNT 112
    ; GFX9: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_WAITCNT 112
    ; GFX9: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_ENDPGM 0
    $vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
    S_WAITCNT 49279
    $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    S_ENDPGM 0
...

# Apply wait for all counters from preexisting waitcnt regardless of the wait
# required by the next instruction.

---
name:            test_waitcnt_preexisting_apply_all_counters
body:             |
  bb.0:
    liveins: $vgpr0_vgpr1, $vgpr2

    ; GFX9-LABEL: name: test_waitcnt_preexisting_apply_all_counters
    ; GFX9: S_WAITCNT 0
    ; GFX9: $vgpr4_vgpr5 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
    ; GFX9: $vgpr6_vgpr7 = DS_READ2_B32 $vgpr2, 0, 1, 0, implicit $m0, implicit $exec
    ; GFX9: S_WAITCNT 0
    ; GFX9: $vgpr6 = V_OR_B32_e32 1, killed $vgpr6, implicit $exec
    ; GFX9: $vgpr0 = FLAT_LOAD_DWORD $vgpr4_vgpr5, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_WAITCNT 112
    ; GFX9: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    $vgpr4_vgpr5 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
    $vgpr6_vgpr7 = DS_READ2_B32 $vgpr2, 0, 1, 0, implicit $m0, implicit $exec
    S_WAITCNT 0
    $vgpr6 = V_OR_B32_e32 1, killed $vgpr6, implicit $exec
    $vgpr0 = FLAT_LOAD_DWORD $vgpr4_vgpr5, 0, 0, implicit $exec, implicit $flat_scr
    FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
...

---
name:            test_waitcnt_preexisting_combine_waitcnt
body:             |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; GFX9-LABEL: name: test_waitcnt_preexisting_combine_waitcnt
    ; GFX9: S_WAITCNT 0
    ; GFX9: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_WAITCNT 0
    ; GFX9: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    S_WAITCNT 0
    S_WAITCNT 0
    S_WAITCNT 0
    S_WAITCNT 0
    S_WAITCNT 0
    S_WAITCNT 0
    FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
...

---
name:            test_waitcnt_preexisting_combine_waitcnt_diff_counters
body:             |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; GFX9-LABEL: name: test_waitcnt_preexisting_combine_waitcnt_diff_counters
    ; GFX9: S_WAITCNT 0
    ; GFX9: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_WAITCNT 112
    ; GFX9: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    S_WAITCNT 49279
    S_WAITCNT 3952
    FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
...

# Apply preexisting waitcnt when no wait is immediately needed.
# FIXME: Move waitcnt as late as possible.

---
name:            test_waitcnt_preexisting_early_wait
body:             |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; GFX9-LABEL: name: test_waitcnt_preexisting_early_wait
    ; GFX9: S_WAITCNT 0
    ; GFX9: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_WAITCNT 0
    ; GFX9: S_NOP 0
    ; GFX9: S_NOP 0
    ; GFX9: S_NOP 0
    ; GFX9: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_ENDPGM 0
    $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    S_WAITCNT 0
    S_NOP 0
    S_NOP 0
    S_NOP 0
    FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    S_ENDPGM 0
...

---
name:            test_waitcnt_preexisting_ignore_kill
body:             |
  bb.0:
    liveins: $vgpr0_vgpr1

    ; GFX9-LABEL: name: test_waitcnt_preexisting_ignore_kill
    ; GFX9: S_WAITCNT 0
    ; GFX9: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_WAITCNT 3952
    ; GFX9: KILL $vgpr0
    $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    S_WAITCNT 3952
    KILL $vgpr0
...

# Combine preexisting waitcnt with wait added to the start of a non-entry function.

---
name:            test_waitcnt_preexisting_func_start
body:             |
  bb.0:
    ; GFX9-LABEL: name: test_waitcnt_preexisting_func_start
    ; GFX9: S_WAITCNT 0
    ; GFX9-NOT: S_WAITCNT 0
    ; GFX9: S_ENDPGM 0
    S_WAITCNT 0
    S_ENDPGM 0
...

# Verify that extra waitcnt are not added after buffer invalidate instructions.

---
name:            test_waitcnt_preexisting_buffer_inv
body:             |
  bb.0:
    ; GFX9-LABEL: name: test_waitcnt_preexisting_buffer_inv
    ; GFX9: S_WAITCNT 0
    ; GFX9: $vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
    ; GFX9: S_WAITCNT 3952
    ; GFX9: BUFFER_INVL2 implicit $exec
    ; GFX9: BUFFER_WBINVL1_VOL implicit $exec
    ; GFX9: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_WAITCNT 112
    ; GFX9: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    ; GFX9: S_ENDPGM 0
    $vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
    S_WAITCNT 3952
    BUFFER_INVL2 implicit $exec
    BUFFER_WBINVL1_VOL implicit $exec
    $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
    FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
    S_ENDPGM 0
...