# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -march=amdgcn -mcpu=gfx90a -run-pass si-insert-waitcnts -verify-machineinstrs %s -o - | FileCheck -check-prefix=GCN %s

--- |
  define amdgpu_kernel void @flat_zero_waitcnt(i32 addrspace(1)* %global4,
                                 <4 x i32> addrspace(1)* %global16,
                                 i32* %flat4,
                                 <4 x i32>* %flat16) {
    ret void
  }

  define amdgpu_kernel void @single_fallthrough_successor_no_end_block_wait() {
    ret void
  }

  define amdgpu_kernel void @single_branch_successor_not_next_block() {
    ret void
  }

  define amdgpu_kernel void @preexisting_waitcnt() {
    ret void
  }

  define amdgpu_kernel void @bundle_no_waitcnt() {
    ret void
  }

  define amdgpu_kernel void @preexisting_waitcnt_in_bundle() {
    ret void
  }

  define amdgpu_kernel void @insert_in_bundle() {
    ret void
  }

  define amdgpu_kernel void @exit_bundle() {
    ret void
  }

  define amdgpu_kernel void @cross_bundle() {
    ret void
  }

...
---


# Global loads will return in order so we should:
# s_waitcnt vmcnt(1)

# s_waitcnt vmcnt(0)

# s_waitcnt vmcnt(0)

name: flat_zero_waitcnt

body: |
  ; GCN-LABEL: name: flat_zero_waitcnt
  ; GCN: bb.0:
  ; GCN:   successors: %bb.1(0x80000000)
  ; GCN:   S_WAITCNT 0
  ; GCN:   $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %ir.global4, addrspace 1)
  ; GCN:   $agpr4_agpr5_agpr6_agpr7 = FLAT_LOAD_DWORDX4 $vgpr8_vgpr9, 0, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %ir.global16, addrspace 1)
  ; GCN:   S_WAITCNT 3953
  ; GCN:   $agpr0 = V_ACCVGPR_MOV_B32 $agpr1, implicit $exec
  ; GCN:   S_BRANCH %bb.1
  ; GCN: bb.1:
  ; GCN:   successors: %bb.2(0x80000000)
  ; GCN:   $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
  ; GCN:   S_WAITCNT 3952
  ; GCN:   $agpr4_agpr5_agpr6_agpr7 = FLAT_LOAD_DWORDX4 $vgpr8_vgpr9, 0, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %ir.global16, addrspace 1)
  ; GCN:   $vgpr0 = V_ACCVGPR_READ_B32_e64 $agpr1, implicit $exec
  ; GCN:   S_BRANCH %bb.2
  ; GCN: bb.2:
  ; GCN:   S_WAITCNT 49279
  ; GCN:   $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %ir.flat4)
  ; GCN:   S_WAITCNT 3952
  ; GCN:   $agpr4_agpr5_agpr6_agpr7 = FLAT_LOAD_DWORDX4 $vgpr8_vgpr9, 0, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %ir.flat16)
  ; GCN:   $vgpr0 = V_ACCVGPR_READ_B32_e64 $agpr1, implicit $exec
  ; GCN:   S_ENDPGM 0
  bb.0:
    successors: %bb.1
    $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %ir.global4)
    $agpr4_agpr5_agpr6_agpr7 = FLAT_LOAD_DWORDX4 $vgpr8_vgpr9, 0, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %ir.global16)
    $agpr0 = V_ACCVGPR_MOV_B32 $agpr1, implicit $exec
    S_BRANCH %bb.1

  bb.1:
    successors: %bb.2
    $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
    $agpr4_agpr5_agpr6_agpr7 = FLAT_LOAD_DWORDX4 $vgpr8_vgpr9, 0, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %ir.global16)
    $vgpr0 = V_ACCVGPR_READ_B32_e64 $agpr1, implicit $exec
    S_BRANCH %bb.2

  bb.2:
    $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr :: (load (s32) from %ir.flat4)
    $agpr4_agpr5_agpr6_agpr7 = FLAT_LOAD_DWORDX4 $vgpr8_vgpr9, 0, 0, implicit $exec, implicit $flat_scr :: (load (s128) from %ir.flat16)
    $vgpr0 = V_ACCVGPR_READ_B32_e64 $agpr1, implicit $exec
    S_ENDPGM 0
...
---
# There is only a single fallthrough successor block, so there's no
# need to wait immediately.


name: single_fallthrough_successor_no_end_block_wait

body: |
  ; GCN-LABEL: name: single_fallthrough_successor_no_end_block_wait
  ; GCN: bb.0:
  ; GCN:   successors: %bb.1(0x80000000)
  ; GCN:   S_WAITCNT 0
  ; GCN:   $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
  ; GCN: bb.1:
  ; GCN:   $vgpr4_vgpr5 = V_LSHLREV_B64_e64 4, $vgpr8_vgpr9, implicit $exec
  ; GCN:   S_WAITCNT 112
  ; GCN:   FLAT_STORE_DWORD $vgpr4_vgpr5, $agpr0, 0, 0, implicit $exec, implicit $flat_scr
  ; GCN:   S_ENDPGM 0
  bb.0:
    successors: %bb.1
    $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr

  bb.1:
    $vgpr4_vgpr5 = V_LSHLREV_B64_e64 4, $vgpr8_vgpr9, implicit $exec
    FLAT_STORE_DWORD $vgpr4_vgpr5, $agpr0, 0, 0, implicit $exec, implicit $flat_scr
    S_ENDPGM 0
...
---
# The block has a single predecessor with a single successor, but it
# is not the next block so it's non-obvious that the wait is not needed.


name: single_branch_successor_not_next_block

body: |
  ; GCN-LABEL: name: single_branch_successor_not_next_block
  ; GCN: bb.0:
  ; GCN:   successors: %bb.2(0x80000000)
  ; GCN:   S_WAITCNT 0
  ; GCN:   $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
  ; GCN:   S_BRANCH %bb.2
  ; GCN: bb.1:
  ; GCN:   FLAT_STORE_DWORD $vgpr8_vgpr9, $agpr10, 0, 0, implicit $exec, implicit $flat_scr
  ; GCN:   S_ENDPGM 0
  ; GCN: bb.2:
  ; GCN:   $vgpr4_vgpr5 = V_LSHLREV_B64_e64 4, $vgpr8_vgpr9, implicit $exec
  ; GCN:   S_WAITCNT 112
  ; GCN:   FLAT_STORE_DWORD $vgpr4_vgpr5, $agpr0, 0, 0, implicit $exec, implicit $flat_scr
  ; GCN:   S_ENDPGM 0
  bb.0:
    successors: %bb.2
    $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
   S_BRANCH %bb.2

  bb.1:
    FLAT_STORE_DWORD $vgpr8_vgpr9, $agpr10, 0, 0, implicit $exec, implicit $flat_scr
    S_ENDPGM 0

  bb.2:
    $vgpr4_vgpr5 = V_LSHLREV_B64_e64 4, $vgpr8_vgpr9, implicit $exec
    FLAT_STORE_DWORD $vgpr4_vgpr5, $agpr0, 0, 0, implicit $exec, implicit $flat_scr
    S_ENDPGM 0
...

# GCN-LABEL: name: preexisting_waitcnt{{$}}
# GCN: FLAT_LOAD_DWORD
# GCN-NEXT: S_WAITCNT 0
# GCN-NOT: S_WAITCNT
name: preexisting_waitcnt
tracksRegLiveness: true
machineFunctionInfo:
  isEntryFunction: true
body: |
  bb.0:
    liveins: $vgpr1_vgpr2
    $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
    S_WAITCNT 0
    FLAT_STORE_DWORD $vgpr2_vgpr3, $agpr0, 0, 0, implicit $exec, implicit $flat_scr

...

---

name: bundle_no_waitcnt
tracksRegLiveness: true
machineFunctionInfo:
  isEntryFunction: true
body: |
  bb.0:
    liveins: $vgpr1_vgpr2
    ; GCN-LABEL: name: bundle_no_waitcnt
    ; GCN: liveins: $vgpr1_vgpr2
    ; GCN: $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
    ; GCN: BUNDLE {
    ; GCN:   S_NOP 0
    ; GCN:   S_NOP 0
    ; GCN: }
    ; GCN: S_WAITCNT 112
    ; GCN: FLAT_STORE_DWORD $vgpr2_vgpr3, $agpr0, 0, 0, implicit $exec, implicit $flat_scr
    $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
    BUNDLE {
      S_NOP 0
      S_NOP 0
    }
    FLAT_STORE_DWORD $vgpr2_vgpr3, $agpr0, 0, 0, implicit $exec, implicit $flat_scr

...

---

# See the waitcnt inside the bundle and don't insert an extra
name: preexisting_waitcnt_in_bundle
tracksRegLiveness: true
machineFunctionInfo:
  isEntryFunction: true
body: |
  bb.0:
    liveins: $vgpr1_vgpr2
    ; GCN-LABEL: name: preexisting_waitcnt_in_bundle
    ; GCN: liveins: $vgpr1_vgpr2
    ; GCN: $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
    ; GCN: BUNDLE {
    ; GCN:   S_NOP 0
    ; GCN:   S_WAITCNT 0
    ; GCN: }
    ; GCN: FLAT_STORE_DWORD $vgpr2_vgpr3, $agpr0, 0, 0, implicit $exec, implicit $flat_scr
    $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
    BUNDLE {
      S_NOP 0
      S_WAITCNT 0
    }
    FLAT_STORE_DWORD $vgpr2_vgpr3, $agpr0, 0, 0, implicit $exec, implicit $flat_scr

...

---

# Def and use inside bundle

name: insert_in_bundle
tracksRegLiveness: true
machineFunctionInfo:
  isEntryFunction: true
body: |
  bb.0:
    liveins: $vgpr1_vgpr2
    ; GCN-LABEL: name: insert_in_bundle
    ; GCN: liveins: $vgpr1_vgpr2
    ; GCN: BUNDLE implicit-def $agpr0, implicit $vgpr2_vgpr3 {
    ; GCN:   $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
    ; GCN:   S_WAITCNT 112
    ; GCN:   FLAT_STORE_DWORD $vgpr2_vgpr3, internal $agpr0, 0, 0, implicit $exec, implicit $flat_scr
    ; GCN: }
    BUNDLE implicit-def $agpr0, implicit $vgpr2_vgpr3 {
    $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
    FLAT_STORE_DWORD $vgpr2_vgpr3, internal $agpr0, 0, 0, implicit $exec, implicit $flat_scr
    }
...

---

# Def is last instruction in bundle, use is outside bundle


name: exit_bundle
tracksRegLiveness: true
machineFunctionInfo:
  isEntryFunction: true
body: |
  bb.0:
    liveins: $vgpr1_vgpr2
    ; GCN-LABEL: name: exit_bundle
    ; GCN: liveins: $vgpr1_vgpr2
    ; GCN: BUNDLE implicit-def $agpr0, implicit $vgpr2_vgpr3 {
    ; GCN:   $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
    ; GCN: }
    ; GCN: S_WAITCNT 112
    ; GCN: FLAT_STORE_DWORD $vgpr2_vgpr3, $agpr0, 0, 0, implicit $exec, implicit $flat_scr
    BUNDLE implicit-def $agpr0, implicit $vgpr2_vgpr3 {
    $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
    }

    FLAT_STORE_DWORD $vgpr2_vgpr3, $agpr0, 0, 0, implicit $exec, implicit $flat_scr

...

---

# Def is in bundle, use is in another bundle


name: cross_bundle
tracksRegLiveness: true
machineFunctionInfo:
  isEntryFunction: true
body: |
  bb.0:
    liveins: $vgpr1_vgpr2
    ; GCN-LABEL: name: cross_bundle
    ; GCN: liveins: $vgpr1_vgpr2
    ; GCN: BUNDLE implicit-def $agpr0, implicit $vgpr2_vgpr3 {
    ; GCN:   $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
    ; GCN: }
    ; GCN: S_WAITCNT 112
    ; GCN: BUNDLE implicit $agpr0, implicit $vgpr2_vgpr3 {
    ; GCN:   FLAT_STORE_DWORD $vgpr2_vgpr3, $agpr0, 0, 0, implicit $exec, implicit $flat_scr
    ; GCN: }
    BUNDLE implicit-def $agpr0, implicit $vgpr2_vgpr3 {
    $agpr0 = FLAT_LOAD_DWORD $vgpr2_vgpr3, 0, 0, implicit $exec, implicit $flat_scr
    }
    BUNDLE implicit $agpr0, implicit $vgpr2_vgpr3 {
      FLAT_STORE_DWORD $vgpr2_vgpr3, $agpr0, 0, 0, implicit $exec, implicit $flat_scr
    }
...