; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX12 %s
; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx942 < %s | FileCheck -check-prefix=GFX942 %s
; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s
; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s
; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx90a < %s | FileCheck -check-prefix=GFX90A %s
; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx908 < %s | FileCheck -check-prefix=GFX908 %s
; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tonga < %s | FileCheck -check-prefix=GFX8 %s
; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=hawaii < %s | FileCheck -check-prefix=GFX7 %s
; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s | FileCheck -check-prefix=GFX6 %s

; --------------------------------------------------------------------
; float
; --------------------------------------------------------------------

define float @local_atomic_fsub_ret_f32(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_ret_f32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v1, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB0_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v2, v1
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB0_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    v_mov_b32_e32 v0, v1
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_f32:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v1, v0
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB0_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v2, v1
; GFX942-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX942-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB0_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v0, v1
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_f32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v1, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB0_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v2, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB0_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    v_mov_b32_e32 v0, v1
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_f32:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v1, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB0_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v2, v1
; GFX10-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB0_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    v_mov_b32_e32 v0, v1
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_f32:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v1, v0
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB0_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v2, v1
; GFX90A-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB0_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v0, v1
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_f32:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v1, v0
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB0_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v2, v1
; GFX908-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX908-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB0_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v0, v1
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_f32:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v1, v0
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB0_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v2, v1
; GFX8-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX8-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB0_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v0, v1
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_f32:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v1, v0
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB0_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_mov_b32_e32 v2, v1
; GFX7-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB0_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v0, v1
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_f32:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v1, v0
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB0_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_mov_b32_e32 v2, v1
; GFX6-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB0_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v0, v1
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, float 4.0 seq_cst
  ret float %result
}

define float @local_atomic_fsub_ret_f32__offset(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_ret_f32__offset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v1, v0 offset:65532
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB1_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v2, v1
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65532
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB1_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    v_mov_b32_e32 v0, v1
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_f32__offset:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v1, v0 offset:65532
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB1_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v2, v1
; GFX942-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX942-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB1_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v0, v1
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_f32__offset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v1, v0 offset:65532
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB1_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v2, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65532
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB1_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    v_mov_b32_e32 v0, v1
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_f32__offset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v1, v0 offset:65532
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB1_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v2, v1
; GFX10-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB1_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    v_mov_b32_e32 v0, v1
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_f32__offset:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v1, v0 offset:65532
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB1_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v2, v1
; GFX90A-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB1_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v0, v1
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_f32__offset:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v1, v0 offset:65532
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB1_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v2, v1
; GFX908-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX908-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB1_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v0, v1
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_f32__offset:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v1, v0 offset:65532
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB1_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v2, v1
; GFX8-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX8-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB1_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v0, v1
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_f32__offset:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v1, v0 offset:65532
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB1_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_mov_b32_e32 v2, v1
; GFX7-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB1_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v0, v1
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_f32__offset:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v1, vcc, 0xfffc, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v0, v1
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB1_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_mov_b32_e32 v2, v0
; GFX6-NEXT:    v_add_f32_e32 v0, -4.0, v2
; GFX6-NEXT:    ds_cmpst_rtn_b32 v0, v1, v2, v0
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v0, v2
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB1_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr float, ptr addrspace(3) %ptr, i32 16383
  %result = atomicrmw fsub ptr addrspace(3) %gep, float 4.0 seq_cst
  ret float %result
}

define void @local_atomic_fsub_noret_f32(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_noret_f32:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v1, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB2_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX12-NEXT:    v_mov_b32_e32 v1, v2
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB2_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_f32:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v1, v0
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB2_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX942-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v1, v2
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB2_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_f32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v1, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB2_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX11-NEXT:    v_mov_b32_e32 v1, v2
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB2_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_f32:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v1, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB2_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX10-NEXT:    v_mov_b32_e32 v1, v2
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB2_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_f32:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v1, v0
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB2_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v1, v2
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB2_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_f32:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v1, v0
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB2_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX908-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v1, v2
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB2_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_f32:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v1, v0
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB2_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX8-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v1, v2
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB2_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_f32:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v1, v0
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB2_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX7-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v1, v2
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB2_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_f32:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v1, v0
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB2_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX6-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v1, v2
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB2_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, float 4.0 seq_cst
  ret void
}

define void @local_atomic_fsub_noret_f32__offset(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_noret_f32__offset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v1, v0 offset:65532
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB3_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65532
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX12-NEXT:    v_mov_b32_e32 v1, v2
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB3_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_f32__offset:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v1, v0 offset:65532
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB3_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX942-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v1, v2
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB3_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_f32__offset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v1, v0 offset:65532
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB3_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65532
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX11-NEXT:    v_mov_b32_e32 v1, v2
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB3_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_f32__offset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v1, v0 offset:65532
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB3_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX10-NEXT:    v_mov_b32_e32 v1, v2
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB3_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_f32__offset:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v1, v0 offset:65532
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB3_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v1, v2
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB3_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_f32__offset:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v1, v0 offset:65532
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB3_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX908-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v1, v2
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB3_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_f32__offset:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v1, v0 offset:65532
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB3_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX8-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v1, v2
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB3_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_f32__offset:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v1, v0 offset:65532
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB3_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX7-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v1, v2
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB3_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_f32__offset:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v0, vcc, 0xfffc, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v1, v0
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB3_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX6-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v1, v2
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB3_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr float, ptr addrspace(3) %ptr, i32 16383
  %unused = atomicrmw fsub ptr addrspace(3) %gep, float 4.0 seq_cst
  ret void
}

; --------------------------------------------------------------------
; double
; --------------------------------------------------------------------

define double @local_atomic_fsub_ret_f64(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_ret_f64:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v2, v0
; GFX12-NEXT:    ds_load_b64 v[0:1], v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB4_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_add_f64_e32 v[0:1], -4.0, v[3:4]
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4]
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB4_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_f64:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v2, v0
; GFX942-NEXT:    ds_read_b64 v[0:1], v0
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB4_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b64_e32 v[4:5], v[0:1]
; GFX942-NEXT:    v_add_f64 v[0:1], v[4:5], -4.0
; GFX942-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1]
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB4_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_f64:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v2, v0
; GFX11-NEXT:    ds_load_b64 v[0:1], v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB4_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_f64 v[0:1], v[3:4], -4.0
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4]
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB4_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_f64:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v2, v0
; GFX10-NEXT:    ds_read_b64 v[0:1], v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB4_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v4, v1
; GFX10-NEXT:    v_mov_b32_e32 v3, v0
; GFX10-NEXT:    v_add_f64 v[0:1], v[3:4], -4.0
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB4_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_f64:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v2, v0
; GFX90A-NEXT:    ds_read_b64 v[0:1], v0
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB4_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
; GFX90A-NEXT:    v_add_f64 v[0:1], v[4:5], -4.0
; GFX90A-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1]
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB4_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_f64:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v2, v0
; GFX908-NEXT:    ds_read_b64 v[0:1], v0
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB4_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v4, v1
; GFX908-NEXT:    v_mov_b32_e32 v3, v0
; GFX908-NEXT:    v_add_f64 v[0:1], v[3:4], -4.0
; GFX908-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB4_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_f64:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    v_mov_b32_e32 v2, v0
; GFX8-NEXT:    ds_read_b64 v[0:1], v0
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB4_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v4, v1
; GFX8-NEXT:    v_mov_b32_e32 v3, v0
; GFX8-NEXT:    v_add_f64 v[0:1], v[3:4], -4.0
; GFX8-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB4_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_f64:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    v_mov_b32_e32 v2, v0
; GFX7-NEXT:    ds_read_b64 v[0:1], v0
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB4_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_mov_b32_e32 v4, v1
; GFX7-NEXT:    v_mov_b32_e32 v3, v0
; GFX7-NEXT:    v_add_f64 v[0:1], v[3:4], -4.0
; GFX7-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB4_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_f64:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    v_mov_b32_e32 v2, v0
; GFX6-NEXT:    ds_read_b64 v[0:1], v0
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB4_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_mov_b32_e32 v4, v1
; GFX6-NEXT:    v_mov_b32_e32 v3, v0
; GFX6-NEXT:    v_add_f64 v[0:1], v[3:4], -4.0
; GFX6-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB4_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, double 4.0 seq_cst
  ret double %result
}

define double @local_atomic_fsub_ret_f64__offset(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_ret_f64__offset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v2, v0
; GFX12-NEXT:    ds_load_b64 v[0:1], v0 offset:65528
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB5_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_add_f64_e32 v[0:1], -4.0, v[3:4]
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4] offset:65528
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB5_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_f64__offset:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v2, v0
; GFX942-NEXT:    ds_read_b64 v[0:1], v0 offset:65528
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB5_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b64_e32 v[4:5], v[0:1]
; GFX942-NEXT:    v_add_f64 v[0:1], v[4:5], -4.0
; GFX942-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1] offset:65528
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB5_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_f64__offset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v2, v0
; GFX11-NEXT:    ds_load_b64 v[0:1], v0 offset:65528
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB5_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v3, v0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_f64 v[0:1], v[3:4], -4.0
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b64 v[0:1], v2, v[0:1], v[3:4] offset:65528
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB5_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_f64__offset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v2, v0
; GFX10-NEXT:    ds_read_b64 v[0:1], v0 offset:65528
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB5_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v4, v1
; GFX10-NEXT:    v_mov_b32_e32 v3, v0
; GFX10-NEXT:    v_add_f64 v[0:1], v[3:4], -4.0
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[0:1], v[3:4]
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB5_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_f64__offset:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v2, v0
; GFX90A-NEXT:    ds_read_b64 v[0:1], v0 offset:65528
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB5_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1]
; GFX90A-NEXT:    v_add_f64 v[0:1], v[4:5], -4.0
; GFX90A-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1] offset:65528
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5]
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB5_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_f64__offset:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v2, v0
; GFX908-NEXT:    ds_read_b64 v[0:1], v0 offset:65528
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB5_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v4, v1
; GFX908-NEXT:    v_mov_b32_e32 v3, v0
; GFX908-NEXT:    v_add_f64 v[0:1], v[3:4], -4.0
; GFX908-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB5_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_f64__offset:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    v_mov_b32_e32 v2, v0
; GFX8-NEXT:    ds_read_b64 v[0:1], v0 offset:65528
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB5_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v4, v1
; GFX8-NEXT:    v_mov_b32_e32 v3, v0
; GFX8-NEXT:    v_add_f64 v[0:1], v[3:4], -4.0
; GFX8-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB5_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_f64__offset:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    v_mov_b32_e32 v2, v0
; GFX7-NEXT:    ds_read_b64 v[0:1], v0 offset:65528
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB5_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_mov_b32_e32 v4, v1
; GFX7-NEXT:    v_mov_b32_e32 v3, v0
; GFX7-NEXT:    v_add_f64 v[0:1], v[3:4], -4.0
; GFX7-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] offset:65528
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB5_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_f64__offset:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v2, vcc, 0xfff8, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b64 v[0:1], v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB5_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_mov_b32_e32 v4, v1
; GFX6-NEXT:    v_mov_b32_e32 v3, v0
; GFX6-NEXT:    v_add_f64 v[0:1], v[3:4], -4.0
; GFX6-NEXT:    ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1]
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4]
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB5_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr double, ptr addrspace(3) %ptr, i32 8191
  %result = atomicrmw fsub ptr addrspace(3) %gep, double 4.0 seq_cst
  ret double %result
}

define void @local_atomic_fsub_noret_f64(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_noret_f64:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b64 v[1:2], v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB6_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_add_f64_e32 v[3:4], -4.0, v[1:2]
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2]
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
; GFX12-NEXT:    v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB6_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_f64:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b64 v[2:3], v0
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB6_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_add_f64 v[4:5], v[2:3], -4.0
; GFX942-NEXT:    ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5]
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b64_e32 v[2:3], v[4:5]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB6_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_f64:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b64 v[1:2], v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB6_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_add_f64 v[3:4], v[1:2], -4.0
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2]
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
; GFX11-NEXT:    v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB6_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_f64:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b64 v[1:2], v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB6_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_add_f64 v[3:4], v[1:2], -4.0
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
; GFX10-NEXT:    v_mov_b32_e32 v1, v3
; GFX10-NEXT:    v_mov_b32_e32 v2, v4
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB6_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_f64:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b64 v[2:3], v0
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB6_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_add_f64 v[4:5], v[2:3], -4.0
; GFX90A-NEXT:    ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5]
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[0,1]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB6_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_f64:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b64 v[1:2], v0
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB6_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_add_f64 v[3:4], v[1:2], -4.0
; GFX908-NEXT:    ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
; GFX908-NEXT:    v_mov_b32_e32 v1, v3
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v2, v4
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB6_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_f64:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b64 v[1:2], v0
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB6_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_add_f64 v[3:4], v[1:2], -4.0
; GFX8-NEXT:    ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
; GFX8-NEXT:    v_mov_b32_e32 v1, v3
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v2, v4
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB6_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_f64:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b64 v[1:2], v0
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB6_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_add_f64 v[3:4], v[1:2], -4.0
; GFX7-NEXT:    ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
; GFX7-NEXT:    v_mov_b32_e32 v1, v3
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v2, v4
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB6_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_f64:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b64 v[1:2], v0
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB6_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_add_f64 v[3:4], v[1:2], -4.0
; GFX6-NEXT:    ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4]
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
; GFX6-NEXT:    v_mov_b32_e32 v1, v3
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v2, v4
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB6_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, double 4.0 seq_cst
  ret void
}

define void @local_atomic_fsub_noret_f64__offset(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_noret_f64__offset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b64 v[1:2], v0 offset:65528
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB7_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_add_f64_e32 v[3:4], -4.0, v[1:2]
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] offset:65528
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
; GFX12-NEXT:    v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB7_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_f64__offset:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b64 v[2:3], v0 offset:65528
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB7_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_add_f64 v[4:5], v[2:3], -4.0
; GFX942-NEXT:    ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5] offset:65528
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b64_e32 v[2:3], v[4:5]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB7_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_f64__offset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b64 v[1:2], v0 offset:65528
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB7_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_add_f64 v[3:4], v[1:2], -4.0
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] offset:65528
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
; GFX11-NEXT:    v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB7_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_f64__offset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b64 v[1:2], v0 offset:65528
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB7_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_add_f64 v[3:4], v[1:2], -4.0
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2]
; GFX10-NEXT:    v_mov_b32_e32 v1, v3
; GFX10-NEXT:    v_mov_b32_e32 v2, v4
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB7_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_f64__offset:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b64 v[2:3], v0 offset:65528
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB7_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_add_f64 v[4:5], v[2:3], -4.0
; GFX90A-NEXT:    ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5] offset:65528
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3]
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[0,1]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB7_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_f64__offset:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b64 v[1:2], v0 offset:65528
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB7_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_add_f64 v[3:4], v[1:2], -4.0
; GFX908-NEXT:    ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
; GFX908-NEXT:    v_mov_b32_e32 v1, v3
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v2, v4
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB7_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_f64__offset:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b64 v[1:2], v0 offset:65528
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB7_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_add_f64 v[3:4], v[1:2], -4.0
; GFX8-NEXT:    ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
; GFX8-NEXT:    v_mov_b32_e32 v1, v3
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v2, v4
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB7_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_f64__offset:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b64 v[1:2], v0 offset:65528
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB7_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_add_f64 v[3:4], v[1:2], -4.0
; GFX7-NEXT:    ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2]
; GFX7-NEXT:    v_mov_b32_e32 v1, v3
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v2, v4
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB7_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_f64__offset:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v2, vcc, 0xfff8, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b64 v[0:1], v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB7_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_add_f64 v[3:4], v[0:1], -4.0
; GFX6-NEXT:    ds_cmpst_rtn_b64 v[3:4], v2, v[0:1], v[3:4]
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u64_e32 vcc, v[3:4], v[0:1]
; GFX6-NEXT:    v_mov_b32_e32 v0, v3
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v1, v4
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB7_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr double, ptr addrspace(3) %ptr, i32 8191
  %unused = atomicrmw fsub ptr addrspace(3) %gep, double 4.0 seq_cst
  ret void
}

; --------------------------------------------------------------------
; half
; --------------------------------------------------------------------

define half @local_atomic_fsub_ret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_ret_f16:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX12-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:    ds_load_b32 v2, v1
; GFX12-NEXT:    v_lshlrev_b32_e64 v3, v0, 0xffff
; GFX12-NEXT:    v_and_b32_e32 v0, 24, v0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-NEXT:    v_not_b32_e32 v3, v3
; GFX12-NEXT:  .LBB8_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v4, v2
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshrrev_b32_e32 v2, v0, v4
; GFX12-NEXT:    v_add_f16_e32 v2, -4.0, v2
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX12-NEXT:    v_lshlrev_b32_e32 v2, v0, v2
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_and_or_b32 v2, v4, v3, v2
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v2, v1, v2, v4
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB8_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    v_lshrrev_b32_e32 v0, v0, v2
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_f16:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX942-NEXT:    ds_read_b32 v2, v1
; GFX942-NEXT:    v_lshlrev_b32_e32 v3, 3, v0
; GFX942-NEXT:    s_mov_b32 s0, 0xffff
; GFX942-NEXT:    v_and_b32_e32 v0, 24, v3
; GFX942-NEXT:    v_lshlrev_b32_e64 v3, v3, s0
; GFX942-NEXT:    v_not_b32_e32 v3, v3
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB8_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v4, v2
; GFX942-NEXT:    v_lshrrev_b32_e32 v2, v0, v4
; GFX942-NEXT:    v_add_f16_e32 v2, -4.0, v2
; GFX942-NEXT:    v_lshlrev_b32_e32 v2, v0, v2
; GFX942-NEXT:    v_and_or_b32 v2, v4, v3, v2
; GFX942-NEXT:    ds_cmpst_rtn_b32 v2, v1, v4, v2
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB8_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    v_lshrrev_b32_e32 v0, v0, v2
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_f16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    ds_load_b32 v2, v1
; GFX11-NEXT:    v_lshlrev_b32_e64 v3, v0, 0xffff
; GFX11-NEXT:    v_and_b32_e32 v0, 24, v0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-NEXT:    v_not_b32_e32 v3, v3
; GFX11-NEXT:  .LBB8_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v4, v2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshrrev_b32_e32 v2, v0, v4
; GFX11-NEXT:    v_add_f16_e32 v2, -4.0, v2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-NEXT:    v_lshlrev_b32_e32 v2, v0, v2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_and_or_b32 v2, v4, v3, v2
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v2, v1, v2, v4
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v4
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB8_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    v_lshrrev_b32_e32 v0, v0, v2
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_f16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:    ds_read_b32 v2, v1
; GFX10-NEXT:    v_lshlrev_b32_e64 v3, v0, 0xffff
; GFX10-NEXT:    v_and_b32_e32 v0, 24, v0
; GFX10-NEXT:    v_not_b32_e32 v3, v3
; GFX10-NEXT:  .LBB8_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v4, v2
; GFX10-NEXT:    v_lshrrev_b32_e32 v2, v0, v4
; GFX10-NEXT:    v_add_f16_e32 v2, -4.0, v2
; GFX10-NEXT:    v_lshlrev_b32_sdwa v2, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; GFX10-NEXT:    v_and_or_b32 v2, v4, v3, v2
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v2, v1, v4, v2
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v4
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB8_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    v_lshrrev_b32_e32 v0, v0, v2
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_f16:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX90A-NEXT:    ds_read_b32 v2, v1
; GFX90A-NEXT:    v_lshlrev_b32_e32 v3, 3, v0
; GFX90A-NEXT:    s_mov_b32 s4, 0xffff
; GFX90A-NEXT:    v_and_b32_e32 v0, 24, v3
; GFX90A-NEXT:    v_lshlrev_b32_e64 v3, v3, s4
; GFX90A-NEXT:    v_not_b32_e32 v3, v3
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB8_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v4, v2
; GFX90A-NEXT:    v_lshrrev_b32_e32 v2, v0, v4
; GFX90A-NEXT:    v_add_f16_e32 v2, -4.0, v2
; GFX90A-NEXT:    v_lshlrev_b32_e32 v2, v0, v2
; GFX90A-NEXT:    v_and_or_b32 v2, v4, v3, v2
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v2, v1, v4, v2
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB8_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    v_lshrrev_b32_e32 v0, v0, v2
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_f16:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX908-NEXT:    ds_read_b32 v2, v1
; GFX908-NEXT:    v_lshlrev_b32_e32 v3, 3, v0
; GFX908-NEXT:    s_mov_b32 s4, 0xffff
; GFX908-NEXT:    v_and_b32_e32 v0, 24, v3
; GFX908-NEXT:    v_lshlrev_b32_e64 v3, v3, s4
; GFX908-NEXT:    v_not_b32_e32 v3, v3
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB8_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v4, v2
; GFX908-NEXT:    v_lshrrev_b32_e32 v2, v0, v4
; GFX908-NEXT:    v_add_f16_e32 v2, -4.0, v2
; GFX908-NEXT:    v_lshlrev_b32_e32 v2, v0, v2
; GFX908-NEXT:    v_and_or_b32 v2, v4, v3, v2
; GFX908-NEXT:    ds_cmpst_rtn_b32 v2, v1, v4, v2
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB8_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    v_lshrrev_b32_e32 v0, v0, v2
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_f16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v3, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX8-NEXT:    s_mov_b32 s4, 0xffff
; GFX8-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX8-NEXT:    v_lshlrev_b32_e64 v2, v2, s4
; GFX8-NEXT:    v_not_b32_e32 v2, v2
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB8_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v4, v3
; GFX8-NEXT:    v_lshrrev_b32_e32 v3, v0, v4
; GFX8-NEXT:    v_add_f16_e32 v3, -4.0, v3
; GFX8-NEXT:    v_and_b32_e32 v5, v4, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, v0, v3
; GFX8-NEXT:    v_or_b32_e32 v3, v5, v3
; GFX8-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB8_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_f16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v1
; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX7-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX7-NEXT:    v_lshl_b32_e32 v2, 0xffff, v2
; GFX7-NEXT:    v_not_b32_e32 v2, v2
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB8_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_mov_b32_e32 v4, v3
; GFX7-NEXT:    v_lshrrev_b32_e32 v3, v0, v4
; GFX7-NEXT:    v_cvt_f32_f16_e32 v3, v3
; GFX7-NEXT:    v_and_b32_e32 v5, v4, v2
; GFX7-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX7-NEXT:    v_cvt_f16_f32_e32 v3, v3
; GFX7-NEXT:    v_lshlrev_b32_e32 v3, v0, v3
; GFX7-NEXT:    v_or_b32_e32 v3, v5, v3
; GFX7-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB8_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v0, v0
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_f16:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v3, v1
; GFX6-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX6-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX6-NEXT:    v_lshl_b32_e32 v2, 0xffff, v2
; GFX6-NEXT:    v_not_b32_e32 v2, v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB8_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_mov_b32_e32 v4, v3
; GFX6-NEXT:    v_lshrrev_b32_e32 v3, v0, v4
; GFX6-NEXT:    v_cvt_f32_f16_e32 v3, v3
; GFX6-NEXT:    v_and_b32_e32 v5, v4, v2
; GFX6-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX6-NEXT:    v_cvt_f16_f32_e32 v3, v3
; GFX6-NEXT:    v_lshlrev_b32_e32 v3, v0, v3
; GFX6-NEXT:    v_or_b32_e32 v3, v5, v3
; GFX6-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB8_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX6-NEXT:    v_cvt_f32_f16_e32 v0, v0
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, half 4.0 seq_cst
  ret half %result
}

define half @local_atomic_fsub_ret_f16__offset(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_ret_f16__offset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    v_add_nc_u32_e32 v1, 0xfffe, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX12-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX12-NEXT:    ds_load_b32 v3, v0
; GFX12-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX12-NEXT:    v_lshlrev_b32_e64 v2, v1, 0xffff
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_not_b32_e32 v2, v2
; GFX12-NEXT:  .LBB9_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v4, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshrrev_b32_e32 v3, v1, v4
; GFX12-NEXT:    v_add_f16_e32 v3, -4.0, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX12-NEXT:    v_lshlrev_b32_e32 v3, v1, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v3, v0, v3, v4
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB9_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    v_lshrrev_b32_e32 v0, v1, v3
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_f16__offset:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    v_add_u32_e32 v0, 0xfffe, v0
; GFX942-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX942-NEXT:    ds_read_b32 v2, v1
; GFX942-NEXT:    v_and_b32_e32 v0, 3, v0
; GFX942-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX942-NEXT:    s_mov_b32 s0, 0xffff
; GFX942-NEXT:    v_lshlrev_b32_e64 v3, v0, s0
; GFX942-NEXT:    v_not_b32_e32 v3, v3
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB9_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v4, v2
; GFX942-NEXT:    v_lshrrev_b32_e32 v2, v0, v4
; GFX942-NEXT:    v_add_f16_e32 v2, -4.0, v2
; GFX942-NEXT:    v_lshlrev_b32_e32 v2, v0, v2
; GFX942-NEXT:    v_and_or_b32 v2, v4, v3, v2
; GFX942-NEXT:    ds_cmpst_rtn_b32 v2, v1, v4, v2
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB9_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    v_lshrrev_b32_e32 v0, v0, v2
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_f16__offset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0xfffe, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX11-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX11-NEXT:    ds_load_b32 v3, v0
; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX11-NEXT:    v_lshlrev_b32_e64 v2, v1, 0xffff
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_not_b32_e32 v2, v2
; GFX11-NEXT:  .LBB9_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v4, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshrrev_b32_e32 v3, v1, v4
; GFX11-NEXT:    v_add_f16_e32 v3, -4.0, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, v1, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v3, v0, v3, v4
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB9_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    v_lshrrev_b32_e32 v0, v1, v3
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_f16__offset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_add_nc_u32_e32 v1, 0xfffe, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX10-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX10-NEXT:    ds_read_b32 v2, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX10-NEXT:    v_lshlrev_b32_e64 v3, v1, 0xffff
; GFX10-NEXT:    v_not_b32_e32 v3, v3
; GFX10-NEXT:  .LBB9_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v4, v2
; GFX10-NEXT:    v_lshrrev_b32_e32 v2, v1, v4
; GFX10-NEXT:    v_add_f16_e32 v2, -4.0, v2
; GFX10-NEXT:    v_lshlrev_b32_sdwa v2, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; GFX10-NEXT:    v_and_or_b32 v2, v4, v3, v2
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v2, v0, v4, v2
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v4
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB9_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    v_lshrrev_b32_e32 v0, v1, v2
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_f16__offset:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    v_add_u32_e32 v0, 0xfffe, v0
; GFX90A-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX90A-NEXT:    ds_read_b32 v2, v1
; GFX90A-NEXT:    v_and_b32_e32 v0, 3, v0
; GFX90A-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX90A-NEXT:    s_mov_b32 s4, 0xffff
; GFX90A-NEXT:    v_lshlrev_b32_e64 v3, v0, s4
; GFX90A-NEXT:    v_not_b32_e32 v3, v3
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB9_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v4, v2
; GFX90A-NEXT:    v_lshrrev_b32_e32 v2, v0, v4
; GFX90A-NEXT:    v_add_f16_e32 v2, -4.0, v2
; GFX90A-NEXT:    v_lshlrev_b32_e32 v2, v0, v2
; GFX90A-NEXT:    v_and_or_b32 v2, v4, v3, v2
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v2, v1, v4, v2
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB9_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    v_lshrrev_b32_e32 v0, v0, v2
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_f16__offset:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    v_add_u32_e32 v0, 0xfffe, v0
; GFX908-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX908-NEXT:    ds_read_b32 v2, v1
; GFX908-NEXT:    v_and_b32_e32 v0, 3, v0
; GFX908-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX908-NEXT:    s_mov_b32 s4, 0xffff
; GFX908-NEXT:    v_lshlrev_b32_e64 v3, v0, s4
; GFX908-NEXT:    v_not_b32_e32 v3, v3
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB9_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v4, v2
; GFX908-NEXT:    v_lshrrev_b32_e32 v2, v0, v4
; GFX908-NEXT:    v_add_f16_e32 v2, -4.0, v2
; GFX908-NEXT:    v_lshlrev_b32_e32 v2, v0, v2
; GFX908-NEXT:    v_and_or_b32 v2, v4, v3, v2
; GFX908-NEXT:    ds_cmpst_rtn_b32 v2, v1, v4, v2
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB9_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    v_lshrrev_b32_e32 v0, v0, v2
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_f16__offset:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_add_u32_e32 v0, vcc, 0xfffe, v0
; GFX8-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v3, v1
; GFX8-NEXT:    v_and_b32_e32 v0, 3, v0
; GFX8-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX8-NEXT:    s_mov_b32 s4, 0xffff
; GFX8-NEXT:    v_lshlrev_b32_e64 v2, v0, s4
; GFX8-NEXT:    v_not_b32_e32 v2, v2
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB9_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v4, v3
; GFX8-NEXT:    v_lshrrev_b32_e32 v3, v0, v4
; GFX8-NEXT:    v_add_f16_e32 v3, -4.0, v3
; GFX8-NEXT:    v_and_b32_e32 v5, v4, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, v0, v3
; GFX8-NEXT:    v_or_b32_e32 v3, v5, v3
; GFX8-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB9_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_f16__offset:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_add_i32_e32 v1, vcc, 0xfffe, v0
; GFX7-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX7-NEXT:    v_lshl_b32_e32 v2, 0xffff, v1
; GFX7-NEXT:    v_not_b32_e32 v2, v2
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB9_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_mov_b32_e32 v4, v3
; GFX7-NEXT:    v_lshrrev_b32_e32 v3, v1, v4
; GFX7-NEXT:    v_cvt_f32_f16_e32 v3, v3
; GFX7-NEXT:    v_and_b32_e32 v5, v4, v2
; GFX7-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX7-NEXT:    v_cvt_f16_f32_e32 v3, v3
; GFX7-NEXT:    v_lshlrev_b32_e32 v3, v1, v3
; GFX7-NEXT:    v_or_b32_e32 v3, v5, v3
; GFX7-NEXT:    ds_cmpst_rtn_b32 v3, v0, v4, v3
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB9_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_lshrrev_b32_e32 v0, v1, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v0, v0
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_f16__offset:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v1, vcc, 0xfffe, v0
; GFX6-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v3, v0
; GFX6-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX6-NEXT:    v_lshl_b32_e32 v2, 0xffff, v1
; GFX6-NEXT:    v_not_b32_e32 v2, v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB9_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_mov_b32_e32 v4, v3
; GFX6-NEXT:    v_lshrrev_b32_e32 v3, v1, v4
; GFX6-NEXT:    v_cvt_f32_f16_e32 v3, v3
; GFX6-NEXT:    v_and_b32_e32 v5, v4, v2
; GFX6-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX6-NEXT:    v_cvt_f16_f32_e32 v3, v3
; GFX6-NEXT:    v_lshlrev_b32_e32 v3, v1, v3
; GFX6-NEXT:    v_or_b32_e32 v3, v5, v3
; GFX6-NEXT:    ds_cmpst_rtn_b32 v3, v0, v4, v3
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB9_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    v_lshrrev_b32_e32 v0, v1, v3
; GFX6-NEXT:    v_cvt_f32_f16_e32 v0, v0
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
  %result = atomicrmw fsub ptr addrspace(3) %gep, half 4.0 seq_cst
  ret half %result
}

define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_noret_f16:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX12-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:    ds_load_b32 v2, v1
; GFX12-NEXT:    v_lshlrev_b32_e64 v3, v0, 0xffff
; GFX12-NEXT:    v_and_b32_e32 v0, 24, v0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-NEXT:    v_not_b32_e32 v3, v3
; GFX12-NEXT:  .LBB10_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshrrev_b32_e32 v4, v0, v2
; GFX12-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX12-NEXT:    v_lshlrev_b32_e32 v4, v0, v4
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_and_or_b32 v4, v2, v3, v4
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v4, v1, v4, v2
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v2
; GFX12-NEXT:    v_mov_b32_e32 v2, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB10_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_f16:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX942-NEXT:    ds_read_b32 v3, v1
; GFX942-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX942-NEXT:    s_mov_b32 s0, 0xffff
; GFX942-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX942-NEXT:    v_lshlrev_b32_e64 v2, v2, s0
; GFX942-NEXT:    v_not_b32_e32 v2, v2
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB10_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_lshrrev_b32_e32 v4, v0, v3
; GFX942-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX942-NEXT:    v_lshlrev_b32_e32 v4, v0, v4
; GFX942-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX942-NEXT:    ds_cmpst_rtn_b32 v4, v1, v3, v4
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v3, v4
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB10_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_f16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    ds_load_b32 v2, v1
; GFX11-NEXT:    v_lshlrev_b32_e64 v3, v0, 0xffff
; GFX11-NEXT:    v_and_b32_e32 v0, 24, v0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-NEXT:    v_not_b32_e32 v3, v3
; GFX11-NEXT:  .LBB10_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshrrev_b32_e32 v4, v0, v2
; GFX11-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-NEXT:    v_lshlrev_b32_e32 v4, v0, v4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_and_or_b32 v4, v2, v3, v4
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v4, v1, v4, v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v2
; GFX11-NEXT:    v_mov_b32_e32 v2, v4
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB10_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_f16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:    ds_read_b32 v2, v1
; GFX10-NEXT:    v_lshlrev_b32_e64 v3, v0, 0xffff
; GFX10-NEXT:    v_and_b32_e32 v0, 24, v0
; GFX10-NEXT:    v_not_b32_e32 v3, v3
; GFX10-NEXT:  .LBB10_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_lshrrev_b32_e32 v4, v0, v2
; GFX10-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX10-NEXT:    v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; GFX10-NEXT:    v_and_or_b32 v4, v2, v3, v4
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v4, v1, v2, v4
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v2
; GFX10-NEXT:    v_mov_b32_e32 v2, v4
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB10_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_f16:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX90A-NEXT:    ds_read_b32 v3, v1
; GFX90A-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX90A-NEXT:    s_mov_b32 s4, 0xffff
; GFX90A-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX90A-NEXT:    v_lshlrev_b32_e64 v2, v2, s4
; GFX90A-NEXT:    v_not_b32_e32 v2, v2
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB10_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_lshrrev_b32_e32 v4, v0, v3
; GFX90A-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX90A-NEXT:    v_lshlrev_b32_e32 v4, v0, v4
; GFX90A-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v4, v1, v3, v4
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v3, v4
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB10_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_f16:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX908-NEXT:    ds_read_b32 v3, v1
; GFX908-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX908-NEXT:    s_mov_b32 s4, 0xffff
; GFX908-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX908-NEXT:    v_lshlrev_b32_e64 v2, v2, s4
; GFX908-NEXT:    v_not_b32_e32 v2, v2
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB10_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_lshrrev_b32_e32 v4, v0, v3
; GFX908-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX908-NEXT:    v_lshlrev_b32_e32 v4, v0, v4
; GFX908-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX908-NEXT:    ds_cmpst_rtn_b32 v4, v1, v3, v4
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v3, v4
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB10_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_f16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v3, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX8-NEXT:    s_mov_b32 s4, 0xffff
; GFX8-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX8-NEXT:    v_lshlrev_b32_e64 v2, v2, s4
; GFX8-NEXT:    v_not_b32_e32 v2, v2
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB10_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_lshrrev_b32_e32 v4, v0, v3
; GFX8-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX8-NEXT:    v_and_b32_e32 v5, v3, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, v0, v4
; GFX8-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX8-NEXT:    ds_cmpst_rtn_b32 v4, v1, v3, v4
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v3, v4
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB10_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_f16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v1
; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX7-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX7-NEXT:    v_lshl_b32_e32 v2, 0xffff, v2
; GFX7-NEXT:    v_not_b32_e32 v2, v2
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB10_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshrrev_b32_e32 v4, v0, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v4, v4
; GFX7-NEXT:    v_and_b32_e32 v5, v3, v2
; GFX7-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX7-NEXT:    v_cvt_f16_f32_e32 v4, v4
; GFX7-NEXT:    v_lshlrev_b32_e32 v4, v0, v4
; GFX7-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX7-NEXT:    ds_cmpst_rtn_b32 v4, v1, v3, v4
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v3, v4
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB10_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_f16:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v3, v1
; GFX6-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX6-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX6-NEXT:    v_lshl_b32_e32 v2, 0xffff, v2
; GFX6-NEXT:    v_not_b32_e32 v2, v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB10_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshrrev_b32_e32 v4, v0, v3
; GFX6-NEXT:    v_cvt_f32_f16_e32 v4, v4
; GFX6-NEXT:    v_and_b32_e32 v5, v3, v2
; GFX6-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX6-NEXT:    v_cvt_f16_f32_e32 v4, v4
; GFX6-NEXT:    v_lshlrev_b32_e32 v4, v0, v4
; GFX6-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX6-NEXT:    ds_cmpst_rtn_b32 v4, v1, v3, v4
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v3, v4
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB10_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, half 4.0 seq_cst
  ret void
}

define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_noret_f16__offset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    v_add_nc_u32_e32 v1, 0xfffe, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX12-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX12-NEXT:    ds_load_b32 v3, v0
; GFX12-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX12-NEXT:    v_lshlrev_b32_e64 v2, v1, 0xffff
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_not_b32_e32 v2, v2
; GFX12-NEXT:  .LBB11_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX12-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshlrev_b32_e32 v4, v1, v4
; GFX12-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v4, v0, v4, v3
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v3
; GFX12-NEXT:    v_mov_b32_e32 v3, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB11_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_f16__offset:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    v_add_u32_e32 v1, 0xfffe, v0
; GFX942-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX942-NEXT:    ds_read_b32 v3, v0
; GFX942-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX942-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX942-NEXT:    s_mov_b32 s0, 0xffff
; GFX942-NEXT:    v_lshlrev_b32_e64 v2, v1, s0
; GFX942-NEXT:    v_not_b32_e32 v2, v2
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB11_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX942-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX942-NEXT:    v_lshlrev_b32_e32 v4, v1, v4
; GFX942-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX942-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v3, v4
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB11_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_f16__offset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0xfffe, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX11-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX11-NEXT:    ds_load_b32 v3, v0
; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX11-NEXT:    v_lshlrev_b32_e64 v2, v1, 0xffff
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_not_b32_e32 v2, v2
; GFX11-NEXT:  .LBB11_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX11-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshlrev_b32_e32 v4, v1, v4
; GFX11-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v4, v0, v4, v3
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v3
; GFX11-NEXT:    v_mov_b32_e32 v3, v4
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB11_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_f16__offset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_add_nc_u32_e32 v1, 0xfffe, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX10-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX10-NEXT:    ds_read_b32 v3, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX10-NEXT:    v_lshlrev_b32_e64 v2, v1, 0xffff
; GFX10-NEXT:    v_not_b32_e32 v2, v2
; GFX10-NEXT:  .LBB11_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX10-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX10-NEXT:    v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; GFX10-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v3
; GFX10-NEXT:    v_mov_b32_e32 v3, v4
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB11_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_f16__offset:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    v_add_u32_e32 v1, 0xfffe, v0
; GFX90A-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX90A-NEXT:    ds_read_b32 v3, v0
; GFX90A-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX90A-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX90A-NEXT:    s_mov_b32 s4, 0xffff
; GFX90A-NEXT:    v_lshlrev_b32_e64 v2, v1, s4
; GFX90A-NEXT:    v_not_b32_e32 v2, v2
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB11_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX90A-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX90A-NEXT:    v_lshlrev_b32_e32 v4, v1, v4
; GFX90A-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v3, v4
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB11_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_f16__offset:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    v_add_u32_e32 v1, 0xfffe, v0
; GFX908-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX908-NEXT:    ds_read_b32 v3, v0
; GFX908-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX908-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX908-NEXT:    s_mov_b32 s4, 0xffff
; GFX908-NEXT:    v_lshlrev_b32_e64 v2, v1, s4
; GFX908-NEXT:    v_not_b32_e32 v2, v2
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB11_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX908-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX908-NEXT:    v_lshlrev_b32_e32 v4, v1, v4
; GFX908-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX908-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v3, v4
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB11_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_f16__offset:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_add_u32_e32 v1, vcc, 0xfffe, v0
; GFX8-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v3, v0
; GFX8-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX8-NEXT:    s_mov_b32 s4, 0xffff
; GFX8-NEXT:    v_lshlrev_b32_e64 v2, v1, s4
; GFX8-NEXT:    v_not_b32_e32 v2, v2
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB11_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX8-NEXT:    v_add_f16_e32 v4, -4.0, v4
; GFX8-NEXT:    v_and_b32_e32 v5, v3, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, v1, v4
; GFX8-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX8-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v3, v4
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB11_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_f16__offset:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_add_i32_e32 v1, vcc, 0xfffe, v0
; GFX7-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX7-NEXT:    v_lshl_b32_e32 v2, 0xffff, v1
; GFX7-NEXT:    v_not_b32_e32 v2, v2
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB11_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v4, v4
; GFX7-NEXT:    v_and_b32_e32 v5, v3, v2
; GFX7-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX7-NEXT:    v_cvt_f16_f32_e32 v4, v4
; GFX7-NEXT:    v_lshlrev_b32_e32 v4, v1, v4
; GFX7-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX7-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v3, v4
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB11_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_f16__offset:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v1, vcc, 0xfffe, v0
; GFX6-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v3, v0
; GFX6-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX6-NEXT:    v_lshl_b32_e32 v2, 0xffff, v1
; GFX6-NEXT:    v_not_b32_e32 v2, v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB11_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX6-NEXT:    v_cvt_f32_f16_e32 v4, v4
; GFX6-NEXT:    v_and_b32_e32 v5, v3, v2
; GFX6-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX6-NEXT:    v_cvt_f16_f32_e32 v4, v4
; GFX6-NEXT:    v_lshlrev_b32_e32 v4, v1, v4
; GFX6-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX6-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v3, v4
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB11_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
  %unused = atomicrmw fsub ptr addrspace(3) %gep, half 4.0 seq_cst
  ret void
}

define half @local_atomic_fsub_ret_f16__offset__align4(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_ret_f16__offset__align4:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v1, v0 offset:65534
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB12_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v2, v1
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_add_f16_e32 v1, -4.0, v2
; GFX12-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB12_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    v_mov_b32_e32 v0, v1
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_f16__offset__align4:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:    s_mov_b32 s2, 0xffff0000
; GFX942-NEXT:  .LBB12_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v2, v1
; GFX942-NEXT:    v_add_f16_e32 v1, -4.0, v2
; GFX942-NEXT:    v_and_or_b32 v1, v2, s2, v1
; GFX942-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB12_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v0, v1
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_f16__offset__align4:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v1, v0 offset:65534
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB12_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v2, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_f16_e32 v1, -4.0, v2
; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB12_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    v_mov_b32_e32 v0, v1
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_f16__offset__align4:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB12_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v2, v1
; GFX10-NEXT:    v_add_f16_e32 v1, -4.0, v2
; GFX10-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; GFX10-NEXT:    v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB12_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    v_mov_b32_e32 v0, v1
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_f16__offset__align4:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:    s_mov_b32 s6, 0xffff0000
; GFX90A-NEXT:  .LBB12_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v2, v1
; GFX90A-NEXT:    v_add_f16_e32 v1, -4.0, v2
; GFX90A-NEXT:    v_and_or_b32 v1, v2, s6, v1
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB12_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v0, v1
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_f16__offset__align4:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:    s_mov_b32 s6, 0xffff0000
; GFX908-NEXT:  .LBB12_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v2, v1
; GFX908-NEXT:    v_add_f16_e32 v1, -4.0, v2
; GFX908-NEXT:    v_and_or_b32 v1, v2, s6, v1
; GFX908-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB12_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v0, v1
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_f16__offset__align4:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB12_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v2, v1
; GFX8-NEXT:    v_add_f16_e32 v1, -4.0, v2
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX8-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX8-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB12_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v0, v1
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_f16__offset__align4:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB12_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_mov_b32_e32 v2, v1
; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v2
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX7-NEXT:    v_add_f32_e32 v1, -4.0, v1
; GFX7-NEXT:    v_cvt_f16_f32_e32 v1, v1
; GFX7-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB12_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_cvt_f32_f16_e32 v0, v1
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_f16__offset__align4:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v0, vcc, 0xfffe, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v1, v0
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB12_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_mov_b32_e32 v2, v1
; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v2
; GFX6-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX6-NEXT:    v_add_f32_e32 v1, -4.0, v1
; GFX6-NEXT:    v_cvt_f16_f32_e32 v1, v1
; GFX6-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB12_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    v_cvt_f32_f16_e32 v0, v1
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
  %result = atomicrmw fsub ptr addrspace(3) %gep, half 4.0 seq_cst, align 4
  ret half %result
}

define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_noret_f16__offset__align4:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v1, v0 offset:65534
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB13_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_add_f16_e32 v2, -4.0, v1
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX12-NEXT:    v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX12-NEXT:    v_mov_b32_e32 v1, v2
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB13_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_f16__offset__align4:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:    s_mov_b32 s2, 0xffff0000
; GFX942-NEXT:  .LBB13_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_add_f16_e32 v2, -4.0, v1
; GFX942-NEXT:    v_and_or_b32 v2, v1, s2, v2
; GFX942-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v1, v2
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB13_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_f16__offset__align4:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v1, v0 offset:65534
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB13_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_add_f16_e32 v2, -4.0, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-NEXT:    v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX11-NEXT:    v_mov_b32_e32 v1, v2
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB13_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_f16__offset__align4:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB13_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_add_f16_e32 v2, -4.0, v1
; GFX10-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX10-NEXT:    v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX10-NEXT:    v_mov_b32_e32 v1, v2
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB13_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_f16__offset__align4:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:    s_mov_b32 s6, 0xffff0000
; GFX90A-NEXT:  .LBB13_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_add_f16_e32 v2, -4.0, v1
; GFX90A-NEXT:    v_and_or_b32 v2, v1, s6, v2
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v1, v2
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB13_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_f16__offset__align4:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:    s_mov_b32 s6, 0xffff0000
; GFX908-NEXT:  .LBB13_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_add_f16_e32 v2, -4.0, v1
; GFX908-NEXT:    v_and_or_b32 v2, v1, s6, v2
; GFX908-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v1, v2
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB13_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_f16__offset__align4:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB13_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_add_f16_e32 v2, -4.0, v1
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX8-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX8-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v1, v2
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB13_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_f16__offset__align4:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB13_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cvt_f32_f16_e32 v2, v1
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX7-NEXT:    v_add_f32_e32 v2, -4.0, v2
; GFX7-NEXT:    v_cvt_f16_f32_e32 v2, v2
; GFX7-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX7-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v1, v2
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB13_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_f16__offset__align4:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v0, vcc, 0xfffe, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v1, v0
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB13_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cvt_f32_f16_e32 v2, v1
; GFX6-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX6-NEXT:    v_add_f32_e32 v2, -4.0, v2
; GFX6-NEXT:    v_cvt_f16_f32_e32 v2, v2
; GFX6-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX6-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v1, v2
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB13_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr half, ptr addrspace(3) %ptr, i32 32767
  %unused = atomicrmw fsub ptr addrspace(3) %gep, half 4.0 seq_cst, align 4
  ret void
}

; --------------------------------------------------------------------
; bfloat
; --------------------------------------------------------------------

define bfloat @local_atomic_fsub_ret_bf16(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_ret_bf16:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX12-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:    ds_load_b32 v3, v1
; GFX12-NEXT:    v_lshlrev_b32_e64 v2, v0, 0xffff
; GFX12-NEXT:    v_and_b32_e32 v0, 24, v0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-NEXT:    v_not_b32_e32 v2, v2
; GFX12-NEXT:  .LBB14_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v4, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshrrev_b32_e32 v3, v0, v4
; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX12-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
; GFX12-NEXT:    s_wait_alu 0xfffd
; GFX12-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX12-NEXT:    v_lshlrev_b32_e32 v3, v0, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v3, v1, v3, v4
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB14_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_bf16:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX942-NEXT:    ds_read_b32 v3, v1
; GFX942-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX942-NEXT:    s_mov_b32 s0, 0xffff
; GFX942-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX942-NEXT:    v_lshlrev_b32_e64 v2, v2, s0
; GFX942-NEXT:    v_not_b32_e32 v2, v2
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:    s_movk_i32 s2, 0x7fff
; GFX942-NEXT:  .LBB14_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v4, v3
; GFX942-NEXT:    v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX942-NEXT:    s_nop 0
; GFX942-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX942-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX942-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX942-NEXT:    v_add3_u32 v5, v5, v3, s2
; GFX942-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX942-NEXT:    s_nop 1
; GFX942-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX942-NEXT:    v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX942-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX942-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB14_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_bf16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    ds_load_b32 v3, v1
; GFX11-NEXT:    v_lshlrev_b32_e64 v2, v0, 0xffff
; GFX11-NEXT:    v_and_b32_e32 v0, 24, v0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-NEXT:    v_not_b32_e32 v2, v2
; GFX11-NEXT:    .p2align 6
; GFX11-NEXT:  .LBB14_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v4, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshrrev_b32_e32 v3, v0, v4
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX11-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
; GFX11-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, v0, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v3, v1, v3, v4
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB14_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_bf16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:    ds_read_b32 v3, v1
; GFX10-NEXT:    v_lshlrev_b32_e64 v2, v0, 0xffff
; GFX10-NEXT:    v_and_b32_e32 v0, 24, v0
; GFX10-NEXT:    v_not_b32_e32 v2, v2
; GFX10-NEXT:  .LBB14_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v4, v3
; GFX10-NEXT:    v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX10-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX10-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX10-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB14_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_bf16:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX90A-NEXT:    ds_read_b32 v3, v1
; GFX90A-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX90A-NEXT:    s_mov_b32 s4, 0xffff
; GFX90A-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX90A-NEXT:    v_lshlrev_b32_e64 v2, v2, s4
; GFX90A-NEXT:    v_not_b32_e32 v2, v2
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:    s_movk_i32 s6, 0x7fff
; GFX90A-NEXT:  .LBB14_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v4, v3
; GFX90A-NEXT:    v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX90A-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX90A-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX90A-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX90A-NEXT:    v_add3_u32 v5, v5, v3, s6
; GFX90A-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX90A-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX90A-NEXT:    v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX90A-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB14_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_bf16:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX908-NEXT:    ds_read_b32 v3, v1
; GFX908-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX908-NEXT:    s_mov_b32 s4, 0xffff
; GFX908-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX908-NEXT:    v_lshlrev_b32_e64 v2, v2, s4
; GFX908-NEXT:    v_not_b32_e32 v2, v2
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:    s_movk_i32 s6, 0x7fff
; GFX908-NEXT:  .LBB14_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v4, v3
; GFX908-NEXT:    v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX908-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX908-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX908-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX908-NEXT:    v_add3_u32 v5, v5, v3, s6
; GFX908-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX908-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX908-NEXT:    v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX908-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX908-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB14_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_bf16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v3, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX8-NEXT:    s_mov_b32 s4, 0xffff
; GFX8-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX8-NEXT:    v_lshlrev_b32_e64 v2, v2, s4
; GFX8-NEXT:    v_not_b32_e32 v2, v2
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB14_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v4, v3
; GFX8-NEXT:    v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX8-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX8-NEXT:    v_bfe_u32 v6, v3, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v3
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, 0x7fff, v6
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v6, v7, vcc
; GFX8-NEXT:    v_and_b32_e32 v5, v4, v2
; GFX8-NEXT:    v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX8-NEXT:    v_or_b32_e32 v3, v5, v3
; GFX8-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB14_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_bf16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v1
; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX7-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX7-NEXT:    v_lshl_b32_e32 v2, 0xffff, v2
; GFX7-NEXT:    v_not_b32_e32 v2, v2
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB14_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_mov_b32_e32 v4, v3
; GFX7-NEXT:    v_lshrrev_b32_e32 v3, v0, v4
; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX7-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX7-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT:    v_and_b32_e32 v5, v4, v2
; GFX7-NEXT:    v_lshlrev_b32_e32 v3, v0, v3
; GFX7-NEXT:    v_or_b32_e32 v3, v5, v3
; GFX7-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB14_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_bf16:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v3, v1
; GFX6-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX6-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX6-NEXT:    v_lshl_b32_e32 v2, 0xffff, v2
; GFX6-NEXT:    v_not_b32_e32 v2, v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB14_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_mov_b32_e32 v4, v3
; GFX6-NEXT:    v_lshrrev_b32_e32 v3, v0, v4
; GFX6-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX6-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX6-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX6-NEXT:    v_and_b32_e32 v5, v4, v2
; GFX6-NEXT:    v_lshlrev_b32_e32 v3, v0, v3
; GFX6-NEXT:    v_or_b32_e32 v3, v5, v3
; GFX6-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB14_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, bfloat 4.0 seq_cst
  ret bfloat %result
}

define bfloat @local_atomic_fsub_ret_bf16__offset(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_ret_bf16__offset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    v_add_nc_u32_e32 v1, 0xfffe, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX12-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX12-NEXT:    ds_load_b32 v3, v0
; GFX12-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX12-NEXT:    v_lshlrev_b32_e64 v2, v1, 0xffff
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_not_b32_e32 v2, v2
; GFX12-NEXT:  .LBB15_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v4, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshrrev_b32_e32 v3, v1, v4
; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX12-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
; GFX12-NEXT:    s_wait_alu 0xfffd
; GFX12-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX12-NEXT:    v_lshlrev_b32_e32 v3, v1, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v3, v0, v3, v4
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB15_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    v_lshrrev_b32_e32 v0, v1, v3
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_bf16__offset:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    v_add_u32_e32 v0, 0xfffe, v0
; GFX942-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX942-NEXT:    ds_read_b32 v3, v1
; GFX942-NEXT:    v_and_b32_e32 v0, 3, v0
; GFX942-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX942-NEXT:    s_mov_b32 s0, 0xffff
; GFX942-NEXT:    v_lshlrev_b32_e64 v2, v0, s0
; GFX942-NEXT:    v_not_b32_e32 v2, v2
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:    s_movk_i32 s2, 0x7fff
; GFX942-NEXT:  .LBB15_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v4, v3
; GFX942-NEXT:    v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX942-NEXT:    s_nop 0
; GFX942-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX942-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX942-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX942-NEXT:    v_add3_u32 v5, v5, v3, s2
; GFX942-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX942-NEXT:    s_nop 1
; GFX942-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX942-NEXT:    v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX942-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX942-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB15_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_bf16__offset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0xfffe, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX11-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX11-NEXT:    ds_load_b32 v3, v0
; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX11-NEXT:    v_lshlrev_b32_e64 v2, v1, 0xffff
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_not_b32_e32 v2, v2
; GFX11-NEXT:    .p2align 6
; GFX11-NEXT:  .LBB15_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v4, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshrrev_b32_e32 v3, v1, v4
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX11-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
; GFX11-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, v1, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v3, v0, v3, v4
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB15_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    v_lshrrev_b32_e32 v0, v1, v3
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_bf16__offset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_add_nc_u32_e32 v1, 0xfffe, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX10-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX10-NEXT:    ds_read_b32 v3, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX10-NEXT:    v_lshlrev_b32_e64 v2, v1, 0xffff
; GFX10-NEXT:    v_not_b32_e32 v2, v2
; GFX10-NEXT:  .LBB15_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v4, v3
; GFX10-NEXT:    v_lshrrev_b32_sdwa v3, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX10-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX10-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX10-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX10-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v3, v0, v4, v3
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v4
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB15_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    v_lshrrev_b32_e32 v0, v1, v3
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_bf16__offset:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    v_add_u32_e32 v0, 0xfffe, v0
; GFX90A-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX90A-NEXT:    ds_read_b32 v3, v1
; GFX90A-NEXT:    v_and_b32_e32 v0, 3, v0
; GFX90A-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX90A-NEXT:    s_mov_b32 s4, 0xffff
; GFX90A-NEXT:    v_lshlrev_b32_e64 v2, v0, s4
; GFX90A-NEXT:    v_not_b32_e32 v2, v2
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:    s_movk_i32 s6, 0x7fff
; GFX90A-NEXT:  .LBB15_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v4, v3
; GFX90A-NEXT:    v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX90A-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX90A-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX90A-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX90A-NEXT:    v_add3_u32 v5, v5, v3, s6
; GFX90A-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX90A-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX90A-NEXT:    v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX90A-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB15_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_bf16__offset:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    v_add_u32_e32 v0, 0xfffe, v0
; GFX908-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX908-NEXT:    ds_read_b32 v3, v1
; GFX908-NEXT:    v_and_b32_e32 v0, 3, v0
; GFX908-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX908-NEXT:    s_mov_b32 s4, 0xffff
; GFX908-NEXT:    v_lshlrev_b32_e64 v2, v0, s4
; GFX908-NEXT:    v_not_b32_e32 v2, v2
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:    s_movk_i32 s6, 0x7fff
; GFX908-NEXT:  .LBB15_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v4, v3
; GFX908-NEXT:    v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX908-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX908-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX908-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX908-NEXT:    v_add3_u32 v5, v5, v3, s6
; GFX908-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX908-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX908-NEXT:    v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX908-NEXT:    v_and_or_b32 v3, v4, v2, v3
; GFX908-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB15_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_bf16__offset:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_add_u32_e32 v0, vcc, 0xfffe, v0
; GFX8-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v3, v1
; GFX8-NEXT:    v_and_b32_e32 v0, 3, v0
; GFX8-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX8-NEXT:    s_mov_b32 s4, 0xffff
; GFX8-NEXT:    v_lshlrev_b32_e64 v2, v0, s4
; GFX8-NEXT:    v_not_b32_e32 v2, v2
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB15_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v4, v3
; GFX8-NEXT:    v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX8-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX8-NEXT:    v_bfe_u32 v6, v3, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v3
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, 0x7fff, v6
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX8-NEXT:    v_cndmask_b32_e32 v3, v6, v7, vcc
; GFX8-NEXT:    v_and_b32_e32 v5, v4, v2
; GFX8-NEXT:    v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX8-NEXT:    v_or_b32_e32 v3, v5, v3
; GFX8-NEXT:    ds_cmpst_rtn_b32 v3, v1, v4, v3
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB15_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    v_lshrrev_b32_e32 v0, v0, v3
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_bf16__offset:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_add_i32_e32 v1, vcc, 0xfffe, v0
; GFX7-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX7-NEXT:    v_lshl_b32_e32 v2, 0xffff, v1
; GFX7-NEXT:    v_not_b32_e32 v2, v2
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB15_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_mov_b32_e32 v4, v3
; GFX7-NEXT:    v_lshrrev_b32_e32 v3, v1, v4
; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX7-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX7-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT:    v_and_b32_e32 v5, v4, v2
; GFX7-NEXT:    v_lshlrev_b32_e32 v3, v1, v3
; GFX7-NEXT:    v_or_b32_e32 v3, v5, v3
; GFX7-NEXT:    ds_cmpst_rtn_b32 v3, v0, v4, v3
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB15_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_lshrrev_b32_e32 v0, v1, v3
; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_bf16__offset:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v1, vcc, 0xfffe, v0
; GFX6-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v3, v0
; GFX6-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX6-NEXT:    v_lshl_b32_e32 v2, 0xffff, v1
; GFX6-NEXT:    v_not_b32_e32 v2, v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB15_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_mov_b32_e32 v4, v3
; GFX6-NEXT:    v_lshrrev_b32_e32 v3, v1, v4
; GFX6-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX6-NEXT:    v_add_f32_e32 v3, -4.0, v3
; GFX6-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX6-NEXT:    v_and_b32_e32 v5, v4, v2
; GFX6-NEXT:    v_lshlrev_b32_e32 v3, v1, v3
; GFX6-NEXT:    v_or_b32_e32 v3, v5, v3
; GFX6-NEXT:    ds_cmpst_rtn_b32 v3, v0, v4, v3
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v4
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB15_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    v_lshrrev_b32_e32 v0, v1, v3
; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
  %result = atomicrmw fsub ptr addrspace(3) %gep, bfloat 4.0 seq_cst
  ret bfloat %result
}

define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_noret_bf16:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX12-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:    ds_load_b32 v2, v1
; GFX12-NEXT:    v_lshlrev_b32_e64 v3, v0, 0xffff
; GFX12-NEXT:    v_and_b32_e32 v0, 24, v0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX12-NEXT:    v_not_b32_e32 v3, v3
; GFX12-NEXT:  .LBB16_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshrrev_b32_e32 v4, v0, v2
; GFX12-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX12-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_add3_u32 v5, v5, v4, 0x7fff
; GFX12-NEXT:    s_wait_alu 0xfffd
; GFX12-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc_lo
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; GFX12-NEXT:    v_lshlrev_b32_e32 v4, v0, v4
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_and_or_b32 v4, v2, v3, v4
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v4, v1, v4, v2
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v2
; GFX12-NEXT:    v_mov_b32_e32 v2, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB16_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_bf16:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX942-NEXT:    ds_read_b32 v3, v1
; GFX942-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX942-NEXT:    s_mov_b32 s0, 0xffff
; GFX942-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX942-NEXT:    v_lshlrev_b32_e64 v2, v2, s0
; GFX942-NEXT:    v_not_b32_e32 v2, v2
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:    s_movk_i32 s2, 0x7fff
; GFX942-NEXT:  .LBB16_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX942-NEXT:    s_nop 0
; GFX942-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX942-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX942-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX942-NEXT:    v_add3_u32 v5, v5, v4, s2
; GFX942-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX942-NEXT:    s_nop 1
; GFX942-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX942-NEXT:    v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX942-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX942-NEXT:    ds_cmpst_rtn_b32 v4, v1, v3, v4
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v3, v4
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB16_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_bf16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    ds_load_b32 v2, v1
; GFX11-NEXT:    v_lshlrev_b32_e64 v3, v0, 0xffff
; GFX11-NEXT:    v_and_b32_e32 v0, 24, v0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-NEXT:    v_not_b32_e32 v3, v3
; GFX11-NEXT:    .p2align 6
; GFX11-NEXT:  .LBB16_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshrrev_b32_e32 v4, v0, v2
; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX11-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add3_u32 v5, v5, v4, 0x7fff
; GFX11-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc_lo
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; GFX11-NEXT:    v_lshlrev_b32_e32 v4, v0, v4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_and_or_b32 v4, v2, v3, v4
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v4, v1, v4, v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v2
; GFX11-NEXT:    v_mov_b32_e32 v2, v4
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB16_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_bf16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:    ds_read_b32 v2, v1
; GFX10-NEXT:    v_lshlrev_b32_e64 v3, v0, 0xffff
; GFX10-NEXT:    v_and_b32_e32 v0, 24, v0
; GFX10-NEXT:    v_not_b32_e32 v3, v3
; GFX10-NEXT:  .LBB16_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_lshrrev_b32_sdwa v4, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX10-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX10-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_add3_u32 v5, v5, v4, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX10-NEXT:    v_and_or_b32 v4, v2, v3, v4
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v4, v1, v2, v4
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v2
; GFX10-NEXT:    v_mov_b32_e32 v2, v4
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB16_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_bf16:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX90A-NEXT:    ds_read_b32 v3, v1
; GFX90A-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX90A-NEXT:    s_mov_b32 s4, 0xffff
; GFX90A-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX90A-NEXT:    v_lshlrev_b32_e64 v2, v2, s4
; GFX90A-NEXT:    v_not_b32_e32 v2, v2
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:    s_movk_i32 s6, 0x7fff
; GFX90A-NEXT:  .LBB16_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX90A-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX90A-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX90A-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX90A-NEXT:    v_add3_u32 v5, v5, v4, s6
; GFX90A-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX90A-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX90A-NEXT:    v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX90A-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v4, v1, v3, v4
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v3, v4
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB16_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_bf16:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX908-NEXT:    ds_read_b32 v3, v1
; GFX908-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX908-NEXT:    s_mov_b32 s4, 0xffff
; GFX908-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX908-NEXT:    v_lshlrev_b32_e64 v2, v2, s4
; GFX908-NEXT:    v_not_b32_e32 v2, v2
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:    s_movk_i32 s6, 0x7fff
; GFX908-NEXT:  .LBB16_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX908-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX908-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX908-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX908-NEXT:    v_add3_u32 v5, v5, v4, s6
; GFX908-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX908-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX908-NEXT:    v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX908-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX908-NEXT:    ds_cmpst_rtn_b32 v4, v1, v3, v4
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v3, v4
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB16_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_bf16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v3, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX8-NEXT:    s_mov_b32 s4, 0xffff
; GFX8-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX8-NEXT:    v_lshlrev_b32_e64 v2, v2, s4
; GFX8-NEXT:    v_not_b32_e32 v2, v2
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB16_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX8-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX8-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v4
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, 0x7fff, v6
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v6, v7, vcc
; GFX8-NEXT:    v_and_b32_e32 v5, v3, v2
; GFX8-NEXT:    v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX8-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX8-NEXT:    ds_cmpst_rtn_b32 v4, v1, v3, v4
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v3, v4
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB16_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_bf16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v1
; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX7-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX7-NEXT:    v_lshl_b32_e32 v2, 0xffff, v2
; GFX7-NEXT:    v_not_b32_e32 v2, v2
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB16_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshrrev_b32_e32 v4, v0, v3
; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX7-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX7-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT:    v_and_b32_e32 v5, v3, v2
; GFX7-NEXT:    v_lshlrev_b32_e32 v4, v0, v4
; GFX7-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX7-NEXT:    ds_cmpst_rtn_b32 v4, v1, v3, v4
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v3, v4
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB16_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_bf16:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_and_b32_e32 v1, -4, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v3, v1
; GFX6-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
; GFX6-NEXT:    v_and_b32_e32 v0, 24, v2
; GFX6-NEXT:    v_lshl_b32_e32 v2, 0xffff, v2
; GFX6-NEXT:    v_not_b32_e32 v2, v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB16_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshrrev_b32_e32 v4, v0, v3
; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX6-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX6-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; GFX6-NEXT:    v_and_b32_e32 v5, v3, v2
; GFX6-NEXT:    v_lshlrev_b32_e32 v4, v0, v4
; GFX6-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX6-NEXT:    ds_cmpst_rtn_b32 v4, v1, v3, v4
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v3, v4
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB16_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, bfloat 4.0 seq_cst
  ret void
}

define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_noret_bf16__offset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    v_add_nc_u32_e32 v1, 0xfffe, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX12-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX12-NEXT:    ds_load_b32 v3, v0
; GFX12-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX12-NEXT:    v_lshlrev_b32_e64 v2, v1, 0xffff
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_not_b32_e32 v2, v2
; GFX12-NEXT:  .LBB17_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX12-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX12-NEXT:    v_add3_u32 v5, v5, v4, 0x7fff
; GFX12-NEXT:    s_wait_alu 0xfffd
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc_lo
; GFX12-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshlrev_b32_e32 v4, v1, v4
; GFX12-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v4, v0, v4, v3
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v3
; GFX12-NEXT:    v_mov_b32_e32 v3, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB17_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_bf16__offset:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    v_add_u32_e32 v1, 0xfffe, v0
; GFX942-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX942-NEXT:    ds_read_b32 v3, v0
; GFX942-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX942-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX942-NEXT:    s_mov_b32 s0, 0xffff
; GFX942-NEXT:    v_lshlrev_b32_e64 v2, v1, s0
; GFX942-NEXT:    v_not_b32_e32 v2, v2
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:    s_movk_i32 s2, 0x7fff
; GFX942-NEXT:  .LBB17_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX942-NEXT:    s_nop 0
; GFX942-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX942-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX942-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX942-NEXT:    v_add3_u32 v5, v5, v4, s2
; GFX942-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX942-NEXT:    s_nop 1
; GFX942-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX942-NEXT:    v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX942-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX942-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v3, v4
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB17_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_bf16__offset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_add_nc_u32_e32 v1, 0xfffe, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX11-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX11-NEXT:    ds_load_b32 v3, v0
; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX11-NEXT:    v_lshlrev_b32_e64 v2, v1, 0xffff
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_not_b32_e32 v2, v2
; GFX11-NEXT:    .p2align 6
; GFX11-NEXT:  .LBB17_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX11-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-NEXT:    v_add3_u32 v5, v5, v4, 0x7fff
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc_lo
; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshlrev_b32_e32 v4, v1, v4
; GFX11-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v4, v0, v4, v3
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v3
; GFX11-NEXT:    v_mov_b32_e32 v3, v4
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB17_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_bf16__offset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    v_add_nc_u32_e32 v1, 0xfffe, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX10-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX10-NEXT:    ds_read_b32 v3, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX10-NEXT:    v_lshlrev_b32_e64 v2, v1, 0xffff
; GFX10-NEXT:    v_not_b32_e32 v2, v2
; GFX10-NEXT:  .LBB17_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX10-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX10-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX10-NEXT:    v_add3_u32 v5, v5, v4, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc_lo
; GFX10-NEXT:    v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX10-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v3
; GFX10-NEXT:    v_mov_b32_e32 v3, v4
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB17_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_bf16__offset:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    v_add_u32_e32 v1, 0xfffe, v0
; GFX90A-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX90A-NEXT:    ds_read_b32 v3, v0
; GFX90A-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX90A-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX90A-NEXT:    s_mov_b32 s4, 0xffff
; GFX90A-NEXT:    v_lshlrev_b32_e64 v2, v1, s4
; GFX90A-NEXT:    v_not_b32_e32 v2, v2
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:    s_movk_i32 s6, 0x7fff
; GFX90A-NEXT:  .LBB17_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX90A-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX90A-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX90A-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX90A-NEXT:    v_add3_u32 v5, v5, v4, s6
; GFX90A-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX90A-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX90A-NEXT:    v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX90A-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v3, v4
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB17_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_bf16__offset:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    v_add_u32_e32 v1, 0xfffe, v0
; GFX908-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX908-NEXT:    ds_read_b32 v3, v0
; GFX908-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX908-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX908-NEXT:    s_mov_b32 s4, 0xffff
; GFX908-NEXT:    v_lshlrev_b32_e64 v2, v1, s4
; GFX908-NEXT:    v_not_b32_e32 v2, v2
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:    s_movk_i32 s6, 0x7fff
; GFX908-NEXT:  .LBB17_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX908-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX908-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX908-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX908-NEXT:    v_add3_u32 v5, v5, v4, s6
; GFX908-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX908-NEXT:    v_cndmask_b32_e32 v4, v5, v6, vcc
; GFX908-NEXT:    v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX908-NEXT:    v_and_or_b32 v4, v3, v2, v4
; GFX908-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v3, v4
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB17_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_bf16__offset:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    v_add_u32_e32 v1, vcc, 0xfffe, v0
; GFX8-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v3, v0
; GFX8-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX8-NEXT:    s_mov_b32 s4, 0xffff
; GFX8-NEXT:    v_lshlrev_b32_e64 v2, v1, s4
; GFX8-NEXT:    v_not_b32_e32 v2, v2
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB17_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX8-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX8-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v4
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, 0x7fff, v6
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v4, v6, v7, vcc
; GFX8-NEXT:    v_and_b32_e32 v5, v3, v2
; GFX8-NEXT:    v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX8-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX8-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v3, v4
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB17_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_bf16__offset:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    v_add_i32_e32 v1, vcc, 0xfffe, v0
; GFX7-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v0
; GFX7-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX7-NEXT:    v_lshl_b32_e32 v2, 0xffff, v1
; GFX7-NEXT:    v_not_b32_e32 v2, v2
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB17_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX7-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX7-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; GFX7-NEXT:    v_and_b32_e32 v5, v3, v2
; GFX7-NEXT:    v_lshlrev_b32_e32 v4, v1, v4
; GFX7-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX7-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v3, v4
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB17_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_bf16__offset:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v1, vcc, 0xfffe, v0
; GFX6-NEXT:    v_and_b32_e32 v0, -4, v1
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v3, v0
; GFX6-NEXT:    v_and_b32_e32 v1, 3, v1
; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 3, v1
; GFX6-NEXT:    v_lshl_b32_e32 v2, 0xffff, v1
; GFX6-NEXT:    v_not_b32_e32 v2, v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB17_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshrrev_b32_e32 v4, v1, v3
; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX6-NEXT:    v_add_f32_e32 v4, -4.0, v4
; GFX6-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; GFX6-NEXT:    v_and_b32_e32 v5, v3, v2
; GFX6-NEXT:    v_lshlrev_b32_e32 v4, v1, v4
; GFX6-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX6-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v3, v4
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB17_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
  %unused = atomicrmw fsub ptr addrspace(3) %gep, bfloat 4.0 seq_cst
  ret void
}

define bfloat @local_atomic_fsub_ret_bf16__offset__align4(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v1, v0 offset:65534
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB18_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v2, v1
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX12-NEXT:    v_add_f32_e32 v1, -4.0, v1
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX12-NEXT:    v_add3_u32 v3, v3, v1, 0x7fff
; GFX12-NEXT:    s_wait_alu 0xfffd
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc_lo
; GFX12-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB18_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    v_mov_b32_e32 v0, v1
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:    s_movk_i32 s2, 0x7fff
; GFX942-NEXT:    s_mov_b32 s3, 0xffff0000
; GFX942-NEXT:  .LBB18_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v2, v1
; GFX942-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX942-NEXT:    v_add_f32_e32 v1, -4.0, v1
; GFX942-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX942-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX942-NEXT:    v_add3_u32 v3, v3, v1, s2
; GFX942-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX942-NEXT:    s_nop 1
; GFX942-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX942-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX942-NEXT:    v_and_or_b32 v1, v2, s3, v1
; GFX942-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB18_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v0, v1
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v1, v0 offset:65534
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    .p2align 6
; GFX11-NEXT:  .LBB18_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v2, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX11-NEXT:    v_add_f32_e32 v1, -4.0, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-NEXT:    v_add3_u32 v3, v3, v1, 0x7fff
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc_lo
; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB18_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    v_mov_b32_e32 v0, v1
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB18_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v2, v1
; GFX10-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX10-NEXT:    v_add_f32_e32 v1, -4.0, v1
; GFX10-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX10-NEXT:    v_add3_u32 v3, v3, v1, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc_lo
; GFX10-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX10-NEXT:    v_and_or_b32 v1, 0xffff0000, v2, v1
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB18_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    v_mov_b32_e32 v0, v1
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:    s_movk_i32 s6, 0x7fff
; GFX90A-NEXT:    s_mov_b32 s7, 0xffff0000
; GFX90A-NEXT:  .LBB18_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v2, v1
; GFX90A-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX90A-NEXT:    v_add_f32_e32 v1, -4.0, v1
; GFX90A-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX90A-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX90A-NEXT:    v_add3_u32 v3, v3, v1, s6
; GFX90A-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX90A-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX90A-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX90A-NEXT:    v_and_or_b32 v1, v2, s7, v1
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB18_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v0, v1
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:    s_movk_i32 s6, 0x7fff
; GFX908-NEXT:    s_mov_b32 s7, 0xffff0000
; GFX908-NEXT:  .LBB18_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v2, v1
; GFX908-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX908-NEXT:    v_add_f32_e32 v1, -4.0, v1
; GFX908-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX908-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX908-NEXT:    v_add3_u32 v3, v3, v1, s6
; GFX908-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX908-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX908-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX908-NEXT:    v_and_or_b32 v1, v2, s7, v1
; GFX908-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB18_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v0, v1
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB18_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v2, v1
; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX8-NEXT:    v_add_f32_e32 v1, -4.0, v1
; GFX8-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; GFX8-NEXT:    v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX8-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB18_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v0, v1
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB18_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_mov_b32_e32 v2, v1
; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX7-NEXT:    v_add_f32_e32 v1, -4.0, v1
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX7-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB18_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 16, v1
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_bf16__offset__align4:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v0, vcc, 0xfffe, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v1, v0
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB18_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_mov_b32_e32 v2, v1
; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX6-NEXT:    v_add_f32_e32 v1, -4.0, v1
; GFX6-NEXT:    v_and_b32_e32 v3, 0xffff0000, v2
; GFX6-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX6-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB18_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 16, v1
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
  %result = atomicrmw fsub ptr addrspace(3) %gep, bfloat 4.0 seq_cst, align 4
  ret bfloat %result
}

define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) nounwind {
; GFX12-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v1, v0 offset:65534
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB19_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_add_f32_e32 v2, -4.0, v2
; GFX12-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX12-NEXT:    s_wait_alu 0xfffd
; GFX12-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX12-NEXT:    v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX12-NEXT:    v_mov_b32_e32 v1, v2
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB19_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:    s_movk_i32 s2, 0x7fff
; GFX942-NEXT:    s_mov_b32 s3, 0xffff0000
; GFX942-NEXT:  .LBB19_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX942-NEXT:    v_add_f32_e32 v2, -4.0, v2
; GFX942-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX942-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX942-NEXT:    v_add3_u32 v3, v3, v2, s2
; GFX942-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX942-NEXT:    s_nop 1
; GFX942-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX942-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX942-NEXT:    v_and_or_b32 v2, v1, s3, v2
; GFX942-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v1, v2
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB19_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v1, v0 offset:65534
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    .p2align 6
; GFX11-NEXT:  .LBB19_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_f32_e32 v2, -4.0, v2
; GFX11-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX11-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-NEXT:    v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX11-NEXT:    v_mov_b32_e32 v1, v2
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB19_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB19_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX10-NEXT:    v_add_f32_e32 v2, -4.0, v2
; GFX10-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX10-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
; GFX10-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
; GFX10-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX10-NEXT:    v_and_or_b32 v2, 0xffff0000, v1, v2
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX10-NEXT:    v_mov_b32_e32 v1, v2
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB19_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:    s_movk_i32 s6, 0x7fff
; GFX90A-NEXT:    s_mov_b32 s7, 0xffff0000
; GFX90A-NEXT:  .LBB19_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX90A-NEXT:    v_add_f32_e32 v2, -4.0, v2
; GFX90A-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX90A-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX90A-NEXT:    v_add3_u32 v3, v3, v2, s6
; GFX90A-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX90A-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX90A-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX90A-NEXT:    v_and_or_b32 v2, v1, s7, v2
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v1, v2
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB19_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:    s_movk_i32 s6, 0x7fff
; GFX908-NEXT:    s_mov_b32 s7, 0xffff0000
; GFX908-NEXT:  .LBB19_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX908-NEXT:    v_add_f32_e32 v2, -4.0, v2
; GFX908-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX908-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX908-NEXT:    v_add3_u32 v3, v3, v2, s6
; GFX908-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX908-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX908-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX908-NEXT:    v_and_or_b32 v2, v1, s7, v2
; GFX908-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v1, v2
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB19_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB19_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT:    v_add_f32_e32 v2, -4.0, v2
; GFX8-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v4, v2
; GFX8-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; GFX8-NEXT:    v_or_b32_e32 v5, 0x400000, v2
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX8-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX8-NEXT:    v_cndmask_b32_e32 v2, v4, v5, vcc
; GFX8-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX8-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v1, v2
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB19_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v1, v0 offset:65534
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB19_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX7-NEXT:    v_add_f32_e32 v2, -4.0, v2
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX7-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX7-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX7-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v1, v2
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB19_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_bf16__offset__align4:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v0, vcc, 0xfffe, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v1, v0
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB19_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX6-NEXT:    v_add_f32_e32 v2, -4.0, v2
; GFX6-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX6-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX6-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX6-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v1, v2
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB19_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr bfloat, ptr addrspace(3) %ptr, i32 32767
  %unused = atomicrmw fsub ptr addrspace(3) %gep, bfloat 4.0 seq_cst, align 4
  ret void
}

; --------------------------------------------------------------------
; <2 x half>
; --------------------------------------------------------------------

define <2 x half> @local_atomic_fsub_ret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val) {
; GFX12-LABEL: local_atomic_fsub_ret_v2f16:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v2, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB20_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v3, v2
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v3
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB20_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    v_mov_b32_e32 v0, v2
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_v2f16:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v2, v0
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB20_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v3, v2
; GFX942-NEXT:    v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX942-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB20_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v0, v2
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_v2f16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v2, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB20_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v3, v2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v3
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB20_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    v_mov_b32_e32 v0, v2
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_v2f16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v2, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB20_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v3, v2
; GFX10-NEXT:    v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB20_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    v_mov_b32_e32 v0, v2
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_v2f16:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v2, v0
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB20_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v3, v2
; GFX90A-NEXT:    v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB20_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v0, v2
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_v2f16:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v2, v0
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB20_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v3, v2
; GFX908-NEXT:    v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX908-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB20_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v0, v2
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_v2f16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v2, v0
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB20_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v3, v2
; GFX8-NEXT:    v_sub_f16_sdwa v2, v3, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
; GFX8-NEXT:    v_sub_f16_e32 v4, v3, v1
; GFX8-NEXT:    v_or_b32_e32 v2, v4, v2
; GFX8-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB20_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v0, v2
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_v2f16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v0
; GFX7-NEXT:    v_cvt_f16_f32_e32 v4, v2
; GFX7-NEXT:    v_cvt_f16_f32_e32 v5, v1
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v2
; GFX7-NEXT:    v_cvt_f32_f16_e32 v2, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v3, v4
; GFX7-NEXT:    v_cvt_f32_f16_e32 v4, v5
; GFX7-NEXT:  .LBB20_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    v_cvt_f16_f32_e32 v1, v1
; GFX7-NEXT:    v_cvt_f16_f32_e32 v2, v2
; GFX7-NEXT:    v_cvt_f32_f16_e32 v5, v1
; GFX7-NEXT:    v_cvt_f32_f16_e32 v6, v2
; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX7-NEXT:    v_or_b32_e32 v7, v2, v1
; GFX7-NEXT:    v_sub_f32_e32 v5, v5, v3
; GFX7-NEXT:    v_sub_f32_e32 v6, v6, v4
; GFX7-NEXT:    v_cvt_f16_f32_e32 v5, v5
; GFX7-NEXT:    v_cvt_f16_f32_e32 v6, v6
; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
; GFX7-NEXT:    v_or_b32_e32 v1, v6, v1
; GFX7-NEXT:    ds_cmpst_rtn_b32 v5, v0, v7, v1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
; GFX7-NEXT:    v_cvt_f32_f16_e32 v2, v5
; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v1
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v7
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB20_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v0, v2
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_v2f16:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v3, v0
; GFX6-NEXT:    v_cvt_f16_f32_e32 v4, v2
; GFX6-NEXT:    v_cvt_f16_f32_e32 v5, v1
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v2
; GFX6-NEXT:    v_cvt_f32_f16_e32 v2, v3
; GFX6-NEXT:    v_cvt_f32_f16_e32 v3, v4
; GFX6-NEXT:    v_cvt_f32_f16_e32 v4, v5
; GFX6-NEXT:  .LBB20_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    v_cvt_f16_f32_e32 v1, v1
; GFX6-NEXT:    v_cvt_f16_f32_e32 v2, v2
; GFX6-NEXT:    v_cvt_f32_f16_e32 v5, v1
; GFX6-NEXT:    v_cvt_f32_f16_e32 v6, v2
; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX6-NEXT:    v_or_b32_e32 v7, v2, v1
; GFX6-NEXT:    v_sub_f32_e32 v5, v5, v3
; GFX6-NEXT:    v_sub_f32_e32 v6, v6, v4
; GFX6-NEXT:    v_cvt_f16_f32_e32 v5, v5
; GFX6-NEXT:    v_cvt_f16_f32_e32 v6, v6
; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
; GFX6-NEXT:    v_or_b32_e32 v1, v6, v1
; GFX6-NEXT:    ds_cmpst_rtn_b32 v5, v0, v7, v1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
; GFX6-NEXT:    v_cvt_f32_f16_e32 v2, v5
; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v1
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v7
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB20_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v0, v2
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
  %result = atomicrmw fsub ptr addrspace(3) %ptr, <2 x half> %val seq_cst
  ret <2 x half> %result
}

define <2 x half> @local_atomic_fsub_ret_v2f16__offset(ptr addrspace(3) %ptr, <2 x half> %val) {
; GFX12-LABEL: local_atomic_fsub_ret_v2f16__offset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v2, v0 offset:65532
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB21_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v3, v2
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v3 offset:65532
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB21_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    v_mov_b32_e32 v0, v2
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_v2f16__offset:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB21_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v3, v2
; GFX942-NEXT:    v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX942-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB21_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v0, v2
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_v2f16__offset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v2, v0 offset:65532
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB21_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v3, v2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v3 offset:65532
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB21_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    v_mov_b32_e32 v0, v2
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_v2f16__offset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB21_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v3, v2
; GFX10-NEXT:    v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v3
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB21_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    v_mov_b32_e32 v0, v2
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_v2f16__offset:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB21_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v3, v2
; GFX90A-NEXT:    v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB21_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v0, v2
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_v2f16__offset:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB21_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v3, v2
; GFX908-NEXT:    v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX908-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB21_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v0, v2
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_v2f16__offset:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB21_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v3, v2
; GFX8-NEXT:    v_sub_f16_sdwa v2, v3, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
; GFX8-NEXT:    v_sub_f16_e32 v4, v3, v1
; GFX8-NEXT:    v_or_b32_e32 v2, v4, v2
; GFX8-NEXT:    ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v3
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB21_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v0, v2
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_v2f16__offset:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v0 offset:65532
; GFX7-NEXT:    v_cvt_f16_f32_e32 v4, v2
; GFX7-NEXT:    v_cvt_f16_f32_e32 v5, v1
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v2
; GFX7-NEXT:    v_cvt_f32_f16_e32 v2, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v3, v4
; GFX7-NEXT:    v_cvt_f32_f16_e32 v4, v5
; GFX7-NEXT:  .LBB21_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    v_cvt_f16_f32_e32 v1, v1
; GFX7-NEXT:    v_cvt_f16_f32_e32 v2, v2
; GFX7-NEXT:    v_cvt_f32_f16_e32 v5, v1
; GFX7-NEXT:    v_cvt_f32_f16_e32 v6, v2
; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX7-NEXT:    v_or_b32_e32 v7, v2, v1
; GFX7-NEXT:    v_sub_f32_e32 v5, v5, v3
; GFX7-NEXT:    v_sub_f32_e32 v6, v6, v4
; GFX7-NEXT:    v_cvt_f16_f32_e32 v5, v5
; GFX7-NEXT:    v_cvt_f16_f32_e32 v6, v6
; GFX7-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
; GFX7-NEXT:    v_or_b32_e32 v1, v6, v1
; GFX7-NEXT:    ds_cmpst_rtn_b32 v5, v0, v7, v1 offset:65532
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
; GFX7-NEXT:    v_cvt_f32_f16_e32 v2, v5
; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v1
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v7
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB21_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v0, v2
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_v2f16__offset:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v3, vcc, 0xfffc, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v4, v3
; GFX6-NEXT:    v_cvt_f16_f32_e32 v2, v2
; GFX6-NEXT:    v_cvt_f16_f32_e32 v5, v1
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshrrev_b32_e32 v1, 16, v4
; GFX6-NEXT:    v_cvt_f32_f16_e32 v0, v4
; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v1
; GFX6-NEXT:    v_cvt_f32_f16_e32 v2, v2
; GFX6-NEXT:    v_cvt_f32_f16_e32 v4, v5
; GFX6-NEXT:  .LBB21_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    v_cvt_f16_f32_e32 v1, v1
; GFX6-NEXT:    v_cvt_f16_f32_e32 v0, v0
; GFX6-NEXT:    v_cvt_f32_f16_e32 v5, v1
; GFX6-NEXT:    v_cvt_f32_f16_e32 v6, v0
; GFX6-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX6-NEXT:    v_or_b32_e32 v7, v0, v1
; GFX6-NEXT:    v_sub_f32_e32 v5, v5, v2
; GFX6-NEXT:    v_sub_f32_e32 v6, v6, v4
; GFX6-NEXT:    v_cvt_f16_f32_e32 v5, v5
; GFX6-NEXT:    v_cvt_f16_f32_e32 v6, v6
; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 16, v5
; GFX6-NEXT:    v_or_b32_e32 v0, v6, v0
; GFX6-NEXT:    ds_cmpst_rtn_b32 v5, v3, v7, v0
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
; GFX6-NEXT:    v_cvt_f32_f16_e32 v0, v5
; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v1
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v7
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB21_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
  %result = atomicrmw fsub ptr addrspace(3) %gep, <2 x half> %val seq_cst
  ret <2 x half> %result
}

define void @local_atomic_fsub_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %val) {
; GFX12-LABEL: local_atomic_fsub_noret_v2f16:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v2, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB22_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v3, v0, v3, v2
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v2
; GFX12-NEXT:    v_mov_b32_e32 v2, v3
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB22_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_v2f16:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v2, v0
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB22_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX942-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v2
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v2, v3
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB22_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_v2f16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v2, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB22_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v3, v0, v3, v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v2
; GFX11-NEXT:    v_mov_b32_e32 v2, v3
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB22_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_v2f16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v2, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB22_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v2
; GFX10-NEXT:    v_mov_b32_e32 v2, v3
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB22_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_v2f16:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v2, v0
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB22_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v2
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v2, v3
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB22_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_v2f16:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v2, v0
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB22_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX908-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v2
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v2, v3
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB22_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_v2f16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v2, v0
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB22_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_sub_f16_sdwa v3, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
; GFX8-NEXT:    v_sub_f16_e32 v4, v2, v1
; GFX8-NEXT:    v_or_b32_e32 v3, v4, v3
; GFX8-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v2
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v2, v3
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB22_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_v2f16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v0
; GFX7-NEXT:    v_cvt_f16_f32_e32 v2, v2
; GFX7-NEXT:    v_cvt_f16_f32_e32 v5, v1
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v2
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v4, v4
; GFX7-NEXT:    v_cvt_f32_f16_e32 v3, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v2, v5
; GFX7-NEXT:  .LBB22_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    v_cvt_f16_f32_e32 v4, v4
; GFX7-NEXT:    v_cvt_f16_f32_e32 v3, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v5, v4
; GFX7-NEXT:    v_cvt_f32_f16_e32 v6, v3
; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX7-NEXT:    v_or_b32_e32 v7, v3, v4
; GFX7-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX7-NEXT:    v_sub_f32_e32 v6, v6, v2
; GFX7-NEXT:    v_cvt_f16_f32_e32 v5, v5
; GFX7-NEXT:    v_cvt_f16_f32_e32 v6, v6
; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
; GFX7-NEXT:    v_or_b32_e32 v3, v6, v3
; GFX7-NEXT:    ds_cmpst_rtn_b32 v5, v0, v7, v3
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
; GFX7-NEXT:    v_cvt_f32_f16_e32 v3, v5
; GFX7-NEXT:    v_cvt_f32_f16_e32 v4, v4
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v7
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB22_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_v2f16:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v3, v0
; GFX6-NEXT:    v_cvt_f16_f32_e32 v2, v2
; GFX6-NEXT:    v_cvt_f16_f32_e32 v5, v1
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v2
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX6-NEXT:    v_cvt_f32_f16_e32 v4, v4
; GFX6-NEXT:    v_cvt_f32_f16_e32 v3, v3
; GFX6-NEXT:    v_cvt_f32_f16_e32 v2, v5
; GFX6-NEXT:  .LBB22_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    v_cvt_f16_f32_e32 v4, v4
; GFX6-NEXT:    v_cvt_f16_f32_e32 v3, v3
; GFX6-NEXT:    v_cvt_f32_f16_e32 v5, v4
; GFX6-NEXT:    v_cvt_f32_f16_e32 v6, v3
; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX6-NEXT:    v_or_b32_e32 v7, v3, v4
; GFX6-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX6-NEXT:    v_sub_f32_e32 v6, v6, v2
; GFX6-NEXT:    v_cvt_f16_f32_e32 v5, v5
; GFX6-NEXT:    v_cvt_f16_f32_e32 v6, v6
; GFX6-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
; GFX6-NEXT:    v_or_b32_e32 v3, v6, v3
; GFX6-NEXT:    ds_cmpst_rtn_b32 v5, v0, v7, v3
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
; GFX6-NEXT:    v_cvt_f32_f16_e32 v3, v5
; GFX6-NEXT:    v_cvt_f32_f16_e32 v4, v4
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v7
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB22_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, <2 x half> %val seq_cst
  ret void
}

define void @local_atomic_fsub_noret_v2f16__offset(ptr addrspace(3) %ptr, <2 x half> %val) {
; GFX12-LABEL: local_atomic_fsub_noret_v2f16__offset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v2, v0 offset:65532
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB23_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v3, v0, v3, v2 offset:65532
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v2
; GFX12-NEXT:    v_mov_b32_e32 v2, v3
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB23_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_v2f16__offset:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB23_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX942-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v2
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v2, v3
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB23_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_v2f16__offset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v2, v0 offset:65532
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB23_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v3, v0, v3, v2 offset:65532
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v2
; GFX11-NEXT:    v_mov_b32_e32 v2, v3
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB23_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_v2f16__offset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB23_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v3, v2
; GFX10-NEXT:    v_mov_b32_e32 v2, v3
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB23_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_v2f16__offset:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB23_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v2
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v2, v3
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB23_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_v2f16__offset:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB23_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1]
; GFX908-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v2
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v2, v3
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB23_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_v2f16__offset:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB23_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_sub_f16_sdwa v3, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
; GFX8-NEXT:    v_sub_f16_e32 v4, v2, v1
; GFX8-NEXT:    v_or_b32_e32 v3, v4, v3
; GFX8-NEXT:    ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v2
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v2, v3
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB23_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_v2f16__offset:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v0 offset:65532
; GFX7-NEXT:    v_cvt_f16_f32_e32 v2, v2
; GFX7-NEXT:    v_cvt_f16_f32_e32 v5, v1
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:    v_cvt_f32_f16_e32 v1, v2
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v4, v4
; GFX7-NEXT:    v_cvt_f32_f16_e32 v3, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v2, v5
; GFX7-NEXT:  .LBB23_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    v_cvt_f16_f32_e32 v4, v4
; GFX7-NEXT:    v_cvt_f16_f32_e32 v3, v3
; GFX7-NEXT:    v_cvt_f32_f16_e32 v5, v4
; GFX7-NEXT:    v_cvt_f32_f16_e32 v6, v3
; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX7-NEXT:    v_or_b32_e32 v7, v3, v4
; GFX7-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX7-NEXT:    v_sub_f32_e32 v6, v6, v2
; GFX7-NEXT:    v_cvt_f16_f32_e32 v5, v5
; GFX7-NEXT:    v_cvt_f16_f32_e32 v6, v6
; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
; GFX7-NEXT:    v_or_b32_e32 v3, v6, v3
; GFX7-NEXT:    ds_cmpst_rtn_b32 v5, v0, v7, v3 offset:65532
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
; GFX7-NEXT:    v_cvt_f32_f16_e32 v3, v5
; GFX7-NEXT:    v_cvt_f32_f16_e32 v4, v4
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v7
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB23_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_v2f16__offset:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v0, vcc, 0xfffc, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v4, v0
; GFX6-NEXT:    v_cvt_f16_f32_e32 v2, v2
; GFX6-NEXT:    v_cvt_f16_f32_e32 v5, v1
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshrrev_b32_e32 v1, 16, v4
; GFX6-NEXT:    v_cvt_f32_f16_e32 v3, v4
; GFX6-NEXT:    v_cvt_f32_f16_e32 v4, v1
; GFX6-NEXT:    v_cvt_f32_f16_e32 v1, v2
; GFX6-NEXT:    v_cvt_f32_f16_e32 v2, v5
; GFX6-NEXT:  .LBB23_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    v_cvt_f16_f32_e32 v4, v4
; GFX6-NEXT:    v_cvt_f16_f32_e32 v3, v3
; GFX6-NEXT:    v_cvt_f32_f16_e32 v5, v4
; GFX6-NEXT:    v_cvt_f32_f16_e32 v6, v3
; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX6-NEXT:    v_or_b32_e32 v7, v3, v4
; GFX6-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX6-NEXT:    v_sub_f32_e32 v6, v6, v2
; GFX6-NEXT:    v_cvt_f16_f32_e32 v5, v5
; GFX6-NEXT:    v_cvt_f16_f32_e32 v6, v6
; GFX6-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
; GFX6-NEXT:    v_or_b32_e32 v3, v6, v3
; GFX6-NEXT:    ds_cmpst_rtn_b32 v5, v0, v7, v3
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
; GFX6-NEXT:    v_cvt_f32_f16_e32 v3, v5
; GFX6-NEXT:    v_cvt_f32_f16_e32 v4, v4
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v7
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB23_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr <2 x half>, ptr addrspace(3) %ptr, i32 16383
  %result = atomicrmw fsub ptr addrspace(3) %gep, <2 x half> %val seq_cst
  ret void
}

; --------------------------------------------------------------------
; <2 x bfloat>
; --------------------------------------------------------------------

define <2 x bfloat> @local_atomic_fsub_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
; GFX12-LABEL: local_atomic_fsub_ret_v2bf16:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v2, v0
; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX12-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX12-NEXT:    s_mov_b32 s1, 0
; GFX12-NEXT:  .LBB24_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v4, v2
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX12-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX12-NEXT:    s_wait_alu 0xfffd
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
; GFX12-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v8, 0x400000, v2
; GFX12-NEXT:    v_cmp_u_f32_e64 s0, v2, v2
; GFX12-NEXT:    v_add3_u32 v6, v6, v2, 0x7fff
; GFX12-NEXT:    s_wait_alu 0xf1ff
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_cndmask_b32_e64 v2, v6, v8, s0
; GFX12-NEXT:    v_perm_b32 v2, v5, v2, 0x7060302
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v4
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s1, vcc_lo, s1
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s1
; GFX12-NEXT:    s_cbranch_execnz .LBB24_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s1
; GFX12-NEXT:    v_mov_b32_e32 v0, v2
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_v2bf16:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v2, v0
; GFX942-NEXT:    s_mov_b64 s[2:3], 0
; GFX942-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX942-NEXT:    s_movk_i32 s4, 0x7fff
; GFX942-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX942-NEXT:    s_mov_b32 s5, 0x7060302
; GFX942-NEXT:  .LBB24_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v4, v2
; GFX942-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX942-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX942-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX942-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX942-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX942-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX942-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX942-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX942-NEXT:    v_add3_u32 v6, v6, v2, s4
; GFX942-NEXT:    v_add3_u32 v8, v8, v5, s4
; GFX942-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX942-NEXT:    v_cmp_u_f32_e64 s[0:1], v2, v2
; GFX942-NEXT:    s_nop 0
; GFX942-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX942-NEXT:    v_cndmask_b32_e64 v2, v6, v7, s[0:1]
; GFX942-NEXT:    v_perm_b32 v2, v5, v2, s5
; GFX942-NEXT:    ds_cmpst_rtn_b32 v2, v0, v4, v2
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX942-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[2:3]
; GFX942-NEXT:    s_cbranch_execnz .LBB24_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[2:3]
; GFX942-NEXT:    v_mov_b32_e32 v0, v2
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_v2bf16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v2, v0
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT:    s_mov_b32 s1, 0
; GFX11-NEXT:    s_set_inst_prefetch_distance 0x1
; GFX11-NEXT:    .p2align 6
; GFX11-NEXT:  .LBB24_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v4, v2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX11-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
; GFX11-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v2
; GFX11-NEXT:    v_cmp_u_f32_e64 s0, v2, v2
; GFX11-NEXT:    v_add3_u32 v6, v6, v2, 0x7fff
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_cndmask_b32_e64 v2, v6, v8, s0
; GFX11-NEXT:    v_perm_b32 v2, v5, v2, 0x7060302
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v4
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v4
; GFX11-NEXT:    s_or_b32 s1, vcc_lo, s1
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s1
; GFX11-NEXT:    s_cbranch_execnz .LBB24_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_set_inst_prefetch_distance 0x2
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s1
; GFX11-NEXT:    v_mov_b32_e32 v0, v2
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_v2bf16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v2, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX10-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT:    s_mov_b32 s5, 0
; GFX10-NEXT:  .LBB24_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v4, v2
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX10-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX10-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX10-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX10-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX10-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v8, 0x400000, v2
; GFX10-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT:    v_add3_u32 v6, v6, v2, 0x7fff
; GFX10-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX10-NEXT:    v_cmp_u_f32_e64 s4, v2, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc_lo
; GFX10-NEXT:    v_cndmask_b32_e64 v2, v6, v8, s4
; GFX10-NEXT:    v_perm_b32 v2, v5, v2, 0x7060302
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v2, v0, v4, v2
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v4
; GFX10-NEXT:    s_or_b32 s5, vcc_lo, s5
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s5
; GFX10-NEXT:    s_cbranch_execnz .LBB24_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s5
; GFX10-NEXT:    v_mov_b32_e32 v0, v2
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_v2bf16:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v2, v0
; GFX90A-NEXT:    s_mov_b64 s[6:7], 0
; GFX90A-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX90A-NEXT:    s_movk_i32 s8, 0x7fff
; GFX90A-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX90A-NEXT:    s_mov_b32 s9, 0x7060302
; GFX90A-NEXT:  .LBB24_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v4, v2
; GFX90A-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX90A-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX90A-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX90A-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX90A-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX90A-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX90A-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX90A-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX90A-NEXT:    v_add3_u32 v6, v6, v2, s8
; GFX90A-NEXT:    v_add3_u32 v8, v8, v5, s8
; GFX90A-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX90A-NEXT:    v_cmp_u_f32_e64 s[4:5], v2, v2
; GFX90A-NEXT:    v_cndmask_b32_e64 v2, v6, v7, s[4:5]
; GFX90A-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX90A-NEXT:    v_perm_b32 v2, v5, v2, s9
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v2, v0, v4, v2
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX90A-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[6:7]
; GFX90A-NEXT:    s_cbranch_execnz .LBB24_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX90A-NEXT:    v_mov_b32_e32 v0, v2
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_v2bf16:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v2, v0
; GFX908-NEXT:    s_mov_b64 s[6:7], 0
; GFX908-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX908-NEXT:    s_movk_i32 s8, 0x7fff
; GFX908-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX908-NEXT:    s_mov_b32 s9, 0x7060302
; GFX908-NEXT:  .LBB24_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v4, v2
; GFX908-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX908-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX908-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX908-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX908-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX908-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX908-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX908-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX908-NEXT:    v_add3_u32 v6, v6, v2, s8
; GFX908-NEXT:    v_add3_u32 v8, v8, v5, s8
; GFX908-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX908-NEXT:    v_cmp_u_f32_e64 s[4:5], v2, v2
; GFX908-NEXT:    v_cndmask_b32_e64 v2, v6, v7, s[4:5]
; GFX908-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX908-NEXT:    v_perm_b32 v2, v5, v2, s9
; GFX908-NEXT:    ds_cmpst_rtn_b32 v2, v0, v4, v2
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX908-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[6:7]
; GFX908-NEXT:    s_cbranch_execnz .LBB24_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX908-NEXT:    v_mov_b32_e32 v0, v2
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_v2bf16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v2, v0
; GFX8-NEXT:    s_mov_b64 s[6:7], 0
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX8-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT:  .LBB24_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v4, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX8-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX8-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX8-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX8-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v2
; GFX8-NEXT:    v_add_u32_e32 v8, vcc, v8, v5
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, 0x7fff, v6
; GFX8-NEXT:    v_add_u32_e32 v8, vcc, 0x7fff, v8
; GFX8-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX8-NEXT:    v_cmp_u_f32_e64 s[4:5], v2, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX8-NEXT:    v_cndmask_b32_e64 v2, v6, v7, s[4:5]
; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX8-NEXT:    v_alignbit_b32 v2, v5, v2, 16
; GFX8-NEXT:    ds_cmpst_rtn_b32 v2, v0, v4, v2
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX8-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[6:7]
; GFX8-NEXT:    s_cbranch_execnz .LBB24_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX8-NEXT:    v_mov_b32_e32 v0, v2
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_v2bf16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v0
; GFX7-NEXT:    v_mov_b32_e32 v4, v1
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:  .LBB24_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX7-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT:    v_sub_f32_e32 v6, v6, v2
; GFX7-NEXT:    v_sub_f32_e32 v5, v5, v4
; GFX7-NEXT:    v_alignbit_b32 v1, v1, v3, 16
; GFX7-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
; GFX7-NEXT:    v_alignbit_b32 v3, v3, v5, 16
; GFX7-NEXT:    ds_cmpst_rtn_b32 v3, v0, v1, v3
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v1
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB24_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v0, v3
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_v2bf16:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v3, v0
; GFX6-NEXT:    v_mov_b32_e32 v4, v1
; GFX6-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX6-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
; GFX6-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX6-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX6-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX6-NEXT:  .LBB24_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX6-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX6-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; GFX6-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX6-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX6-NEXT:    v_sub_f32_e32 v6, v6, v2
; GFX6-NEXT:    v_sub_f32_e32 v5, v5, v4
; GFX6-NEXT:    v_alignbit_b32 v1, v1, v3, 16
; GFX6-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
; GFX6-NEXT:    v_alignbit_b32 v3, v3, v5, 16
; GFX6-NEXT:    ds_cmpst_rtn_b32 v3, v0, v1, v3
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v1
; GFX6-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB24_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v0, v3
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, <2 x bfloat> %val seq_cst
  ret <2 x bfloat> %result
}

define <2 x bfloat> @local_atomic_fsub_ret_v2bf16__offset(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
; GFX12-LABEL: local_atomic_fsub_ret_v2bf16__offset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v2, v0 offset:65532
; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX12-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX12-NEXT:    s_mov_b32 s1, 0
; GFX12-NEXT:  .LBB25_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v4, v2
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX12-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX12-NEXT:    s_wait_alu 0xfffd
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
; GFX12-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX12-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v8, 0x400000, v2
; GFX12-NEXT:    v_cmp_u_f32_e64 s0, v2, v2
; GFX12-NEXT:    v_add3_u32 v6, v6, v2, 0x7fff
; GFX12-NEXT:    s_wait_alu 0xf1ff
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_cndmask_b32_e64 v2, v6, v8, s0
; GFX12-NEXT:    v_perm_b32 v2, v5, v2, 0x7060302
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v4 offset:65532
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s1, vcc_lo, s1
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s1
; GFX12-NEXT:    s_cbranch_execnz .LBB25_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s1
; GFX12-NEXT:    v_mov_b32_e32 v0, v2
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_v2bf16__offset:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX942-NEXT:    s_mov_b64 s[2:3], 0
; GFX942-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX942-NEXT:    s_movk_i32 s4, 0x7fff
; GFX942-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX942-NEXT:    s_mov_b32 s5, 0x7060302
; GFX942-NEXT:  .LBB25_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v4, v2
; GFX942-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX942-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX942-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX942-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX942-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX942-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX942-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX942-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX942-NEXT:    v_add3_u32 v6, v6, v2, s4
; GFX942-NEXT:    v_add3_u32 v8, v8, v5, s4
; GFX942-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX942-NEXT:    v_cmp_u_f32_e64 s[0:1], v2, v2
; GFX942-NEXT:    s_nop 0
; GFX942-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX942-NEXT:    v_cndmask_b32_e64 v2, v6, v7, s[0:1]
; GFX942-NEXT:    v_perm_b32 v2, v5, v2, s5
; GFX942-NEXT:    ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX942-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[2:3]
; GFX942-NEXT:    s_cbranch_execnz .LBB25_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[2:3]
; GFX942-NEXT:    v_mov_b32_e32 v0, v2
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_v2bf16__offset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v2, v0 offset:65532
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT:    s_mov_b32 s1, 0
; GFX11-NEXT:    s_set_inst_prefetch_distance 0x1
; GFX11-NEXT:    .p2align 6
; GFX11-NEXT:  .LBB25_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v4, v2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX11-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_dual_cndmask_b32 v5, v7, v9 :: v_dual_lshlrev_b32 v2, 16, v4
; GFX11-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v2
; GFX11-NEXT:    v_cmp_u_f32_e64 s0, v2, v2
; GFX11-NEXT:    v_add3_u32 v6, v6, v2, 0x7fff
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_cndmask_b32_e64 v2, v6, v8, s0
; GFX11-NEXT:    v_perm_b32 v2, v5, v2, 0x7060302
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v4 offset:65532
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v4
; GFX11-NEXT:    s_or_b32 s1, vcc_lo, s1
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s1
; GFX11-NEXT:    s_cbranch_execnz .LBB25_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_set_inst_prefetch_distance 0x2
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s1
; GFX11-NEXT:    v_mov_b32_e32 v0, v2
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_v2bf16__offset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX10-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT:    s_mov_b32 s5, 0
; GFX10-NEXT:  .LBB25_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v4, v2
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX10-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX10-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX10-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX10-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX10-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v8, 0x400000, v2
; GFX10-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT:    v_add3_u32 v6, v6, v2, 0x7fff
; GFX10-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX10-NEXT:    v_cmp_u_f32_e64 s4, v2, v2
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc_lo
; GFX10-NEXT:    v_cndmask_b32_e64 v2, v6, v8, s4
; GFX10-NEXT:    v_perm_b32 v2, v5, v2, 0x7060302
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v4
; GFX10-NEXT:    s_or_b32 s5, vcc_lo, s5
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s5
; GFX10-NEXT:    s_cbranch_execnz .LBB25_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s5
; GFX10-NEXT:    v_mov_b32_e32 v0, v2
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_v2bf16__offset:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX90A-NEXT:    s_mov_b64 s[6:7], 0
; GFX90A-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX90A-NEXT:    s_movk_i32 s8, 0x7fff
; GFX90A-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX90A-NEXT:    s_mov_b32 s9, 0x7060302
; GFX90A-NEXT:  .LBB25_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v4, v2
; GFX90A-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX90A-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX90A-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX90A-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX90A-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX90A-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX90A-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX90A-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX90A-NEXT:    v_add3_u32 v6, v6, v2, s8
; GFX90A-NEXT:    v_add3_u32 v8, v8, v5, s8
; GFX90A-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX90A-NEXT:    v_cmp_u_f32_e64 s[4:5], v2, v2
; GFX90A-NEXT:    v_cndmask_b32_e64 v2, v6, v7, s[4:5]
; GFX90A-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX90A-NEXT:    v_perm_b32 v2, v5, v2, s9
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX90A-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[6:7]
; GFX90A-NEXT:    s_cbranch_execnz .LBB25_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX90A-NEXT:    v_mov_b32_e32 v0, v2
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_v2bf16__offset:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX908-NEXT:    s_mov_b64 s[6:7], 0
; GFX908-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX908-NEXT:    s_movk_i32 s8, 0x7fff
; GFX908-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX908-NEXT:    s_mov_b32 s9, 0x7060302
; GFX908-NEXT:  .LBB25_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v4, v2
; GFX908-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX908-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX908-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX908-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX908-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX908-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX908-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX908-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX908-NEXT:    v_add3_u32 v6, v6, v2, s8
; GFX908-NEXT:    v_add3_u32 v8, v8, v5, s8
; GFX908-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX908-NEXT:    v_cmp_u_f32_e64 s[4:5], v2, v2
; GFX908-NEXT:    v_cndmask_b32_e64 v2, v6, v7, s[4:5]
; GFX908-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX908-NEXT:    v_perm_b32 v2, v5, v2, s9
; GFX908-NEXT:    ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX908-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[6:7]
; GFX908-NEXT:    s_cbranch_execnz .LBB25_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX908-NEXT:    v_mov_b32_e32 v0, v2
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_v2bf16__offset:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v2, v0 offset:65532
; GFX8-NEXT:    s_mov_b64 s[6:7], 0
; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
; GFX8-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT:  .LBB25_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v4, v2
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX8-NEXT:    v_sub_f32_e32 v2, v2, v3
; GFX8-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX8-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX8-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v2
; GFX8-NEXT:    v_add_u32_e32 v8, vcc, v8, v5
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, 0x7fff, v6
; GFX8-NEXT:    v_add_u32_e32 v8, vcc, 0x7fff, v8
; GFX8-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX8-NEXT:    v_cmp_u_f32_e64 s[4:5], v2, v2
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX8-NEXT:    v_cndmask_b32_e64 v2, v6, v7, s[4:5]
; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX8-NEXT:    v_alignbit_b32 v2, v5, v2, 16
; GFX8-NEXT:    ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v4
; GFX8-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[6:7]
; GFX8-NEXT:    s_cbranch_execnz .LBB25_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX8-NEXT:    v_mov_b32_e32 v0, v2
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_v2bf16__offset:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v3, v0 offset:65532
; GFX7-NEXT:    v_mov_b32_e32 v4, v1
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:  .LBB25_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX7-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX7-NEXT:    v_sub_f32_e32 v6, v6, v2
; GFX7-NEXT:    v_sub_f32_e32 v5, v5, v4
; GFX7-NEXT:    v_alignbit_b32 v1, v1, v3, 16
; GFX7-NEXT:    v_lshrrev_b32_e32 v3, 16, v6
; GFX7-NEXT:    v_alignbit_b32 v3, v3, v5, 16
; GFX7-NEXT:    ds_cmpst_rtn_b32 v3, v0, v1, v3 offset:65532
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v1
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB25_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v0, v3
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_v2bf16__offset:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v4, vcc, 0xfffc, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v0, v4
; GFX6-NEXT:    v_mov_b32_e32 v3, v1
; GFX6-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX6-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_and_b32_e32 v1, 0xffff0000, v0
; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX6-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX6-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX6-NEXT:  .LBB25_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX6-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; GFX6-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; GFX6-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; GFX6-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX6-NEXT:    v_sub_f32_e32 v6, v6, v2
; GFX6-NEXT:    v_sub_f32_e32 v5, v5, v3
; GFX6-NEXT:    v_alignbit_b32 v0, v1, v0, 16
; GFX6-NEXT:    v_lshrrev_b32_e32 v1, 16, v6
; GFX6-NEXT:    v_alignbit_b32 v1, v1, v5, 16
; GFX6-NEXT:    ds_cmpst_rtn_b32 v5, v4, v0, v1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v5, v0
; GFX6-NEXT:    v_and_b32_e32 v1, 0xffff0000, v5
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_lshlrev_b32_e32 v0, 16, v5
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB25_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr <2 x bfloat>, ptr addrspace(3) %ptr, i32 16383
  %result = atomicrmw fsub ptr addrspace(3) %gep, <2 x bfloat> %val seq_cst
  ret <2 x bfloat> %result
}

define void @local_atomic_fsub_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
; GFX12-LABEL: local_atomic_fsub_noret_v2bf16:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v3, v0
; GFX12-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX12-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX12-NEXT:    s_mov_b32 s1, 0
; GFX12-NEXT:  .LBB26_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_dual_sub_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
; GFX12-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX12-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX12-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX12-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX12-NEXT:    v_cmp_u_f32_e64 s0, v4, v4
; GFX12-NEXT:    s_wait_alu 0xfffd
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc_lo
; GFX12-NEXT:    s_wait_alu 0xf1ff
; GFX12-NEXT:    v_cndmask_b32_e64 v4, v6, v8, s0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_perm_b32 v4, v5, v4, 0x7060302
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v4, v0, v4, v3
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v3
; GFX12-NEXT:    v_mov_b32_e32 v3, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s1, vcc_lo, s1
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s1
; GFX12-NEXT:    s_cbranch_execnz .LBB26_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s1
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_v2bf16:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v3, v0
; GFX942-NEXT:    s_mov_b64 s[2:3], 0
; GFX942-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX942-NEXT:    s_movk_i32 s4, 0x7fff
; GFX942-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX942-NEXT:    s_mov_b32 s5, 0x7060302
; GFX942-NEXT:  .LBB26_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX942-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX942-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX942-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX942-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX942-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX942-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX942-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX942-NEXT:    v_add3_u32 v6, v6, v4, s4
; GFX942-NEXT:    v_add3_u32 v8, v8, v5, s4
; GFX942-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX942-NEXT:    v_cmp_u_f32_e64 s[0:1], v4, v4
; GFX942-NEXT:    s_nop 0
; GFX942-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX942-NEXT:    v_cndmask_b32_e64 v4, v6, v7, s[0:1]
; GFX942-NEXT:    v_perm_b32 v4, v5, v4, s5
; GFX942-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX942-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
; GFX942-NEXT:    v_mov_b32_e32 v3, v4
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[2:3]
; GFX942-NEXT:    s_cbranch_execnz .LBB26_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[2:3]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_v2bf16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v3, v0
; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT:    s_mov_b32 s1, 0
; GFX11-NEXT:    s_set_inst_prefetch_distance 0x1
; GFX11-NEXT:    .p2align 6
; GFX11-NEXT:  .LBB26_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_dual_sub_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
; GFX11-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX11-NEXT:    v_cmp_u_f32_e64 s0, v4, v4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc_lo
; GFX11-NEXT:    v_cndmask_b32_e64 v4, v6, v8, s0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_perm_b32 v4, v5, v4, 0x7060302
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v4, v0, v4, v3
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v3
; GFX11-NEXT:    v_mov_b32_e32 v3, v4
; GFX11-NEXT:    s_or_b32 s1, vcc_lo, s1
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s1
; GFX11-NEXT:    s_cbranch_execnz .LBB26_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_set_inst_prefetch_distance 0x2
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s1
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_v2bf16:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v3, v0
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX10-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT:    s_mov_b32 s5, 0
; GFX10-NEXT:  .LBB26_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX10-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX10-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX10-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX10-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX10-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX10-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX10-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX10-NEXT:    v_cmp_u_f32_e64 s4, v4, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc_lo
; GFX10-NEXT:    v_cndmask_b32_e64 v4, v6, v8, s4
; GFX10-NEXT:    v_perm_b32 v4, v5, v4, 0x7060302
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v3
; GFX10-NEXT:    v_mov_b32_e32 v3, v4
; GFX10-NEXT:    s_or_b32 s5, vcc_lo, s5
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s5
; GFX10-NEXT:    s_cbranch_execnz .LBB26_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s5
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_v2bf16:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v3, v0
; GFX90A-NEXT:    s_mov_b64 s[6:7], 0
; GFX90A-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX90A-NEXT:    s_movk_i32 s8, 0x7fff
; GFX90A-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX90A-NEXT:    s_mov_b32 s9, 0x7060302
; GFX90A-NEXT:  .LBB26_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX90A-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX90A-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX90A-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX90A-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX90A-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX90A-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX90A-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX90A-NEXT:    v_add3_u32 v6, v6, v4, s8
; GFX90A-NEXT:    v_add3_u32 v8, v8, v5, s8
; GFX90A-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX90A-NEXT:    v_cmp_u_f32_e64 s[4:5], v4, v4
; GFX90A-NEXT:    v_cndmask_b32_e64 v4, v6, v7, s[4:5]
; GFX90A-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX90A-NEXT:    v_perm_b32 v4, v5, v4, s9
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX90A-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
; GFX90A-NEXT:    v_mov_b32_e32 v3, v4
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[6:7]
; GFX90A-NEXT:    s_cbranch_execnz .LBB26_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_v2bf16:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v3, v0
; GFX908-NEXT:    s_mov_b64 s[6:7], 0
; GFX908-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX908-NEXT:    s_movk_i32 s8, 0x7fff
; GFX908-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX908-NEXT:    s_mov_b32 s9, 0x7060302
; GFX908-NEXT:  .LBB26_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX908-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX908-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX908-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX908-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX908-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX908-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX908-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX908-NEXT:    v_add3_u32 v6, v6, v4, s8
; GFX908-NEXT:    v_add3_u32 v8, v8, v5, s8
; GFX908-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX908-NEXT:    v_cmp_u_f32_e64 s[4:5], v4, v4
; GFX908-NEXT:    v_cndmask_b32_e64 v4, v6, v7, s[4:5]
; GFX908-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX908-NEXT:    v_perm_b32 v4, v5, v4, s9
; GFX908-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX908-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
; GFX908-NEXT:    v_mov_b32_e32 v3, v4
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[6:7]
; GFX908-NEXT:    s_cbranch_execnz .LBB26_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_v2bf16:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v3, v0
; GFX8-NEXT:    s_mov_b64 s[6:7], 0
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT:  .LBB26_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX8-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX8-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX8-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX8-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v4
; GFX8-NEXT:    v_add_u32_e32 v8, vcc, v8, v5
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, 0x7fff, v6
; GFX8-NEXT:    v_add_u32_e32 v8, vcc, 0x7fff, v8
; GFX8-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX8-NEXT:    v_cmp_u_f32_e64 s[4:5], v4, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX8-NEXT:    v_cndmask_b32_e64 v4, v6, v7, s[4:5]
; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX8-NEXT:    v_alignbit_b32 v4, v5, v4, 16
; GFX8-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX8-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
; GFX8-NEXT:    v_mov_b32_e32 v3, v4
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[6:7]
; GFX8-NEXT:    s_cbranch_execnz .LBB26_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_v2bf16:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v4, v0
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:  .LBB26_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX7-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT:    v_sub_f32_e32 v6, v6, v2
; GFX7-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX7-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; GFX7-NEXT:    v_lshrrev_b32_e32 v4, 16, v6
; GFX7-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; GFX7-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB26_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_v2bf16:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v4, v0
; GFX6-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX6-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX6-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX6-NEXT:  .LBB26_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX6-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX6-NEXT:    v_and_b32_e32 v6, 0xffff0000, v3
; GFX6-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX6-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX6-NEXT:    v_sub_f32_e32 v6, v6, v2
; GFX6-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX6-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; GFX6-NEXT:    v_lshrrev_b32_e32 v4, 16, v6
; GFX6-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; GFX6-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX6-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB26_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, <2 x bfloat> %val seq_cst
  ret void
}

define void @local_atomic_fsub_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x bfloat> %val) {
; GFX12-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v3, v0 offset:65532
; GFX12-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX12-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX12-NEXT:    s_mov_b32 s1, 0
; GFX12-NEXT:  .LBB27_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX12-NEXT:    v_dual_sub_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
; GFX12-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX12-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX12-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX12-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX12-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX12-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX12-NEXT:    v_cmp_u_f32_e64 s0, v4, v4
; GFX12-NEXT:    s_wait_alu 0xfffd
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX12-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc_lo
; GFX12-NEXT:    s_wait_alu 0xf1ff
; GFX12-NEXT:    v_cndmask_b32_e64 v4, v6, v8, s0
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_perm_b32 v4, v5, v4, 0x7060302
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v3
; GFX12-NEXT:    v_mov_b32_e32 v3, v4
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s1, vcc_lo, s1
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s1
; GFX12-NEXT:    s_cbranch_execnz .LBB27_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s1
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v3, v0 offset:65532
; GFX942-NEXT:    s_mov_b64 s[2:3], 0
; GFX942-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX942-NEXT:    s_movk_i32 s4, 0x7fff
; GFX942-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX942-NEXT:    s_mov_b32 s5, 0x7060302
; GFX942-NEXT:  .LBB27_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX942-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX942-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX942-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX942-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX942-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX942-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX942-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX942-NEXT:    v_add3_u32 v6, v6, v4, s4
; GFX942-NEXT:    v_add3_u32 v8, v8, v5, s4
; GFX942-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX942-NEXT:    v_cmp_u_f32_e64 s[0:1], v4, v4
; GFX942-NEXT:    s_nop 0
; GFX942-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX942-NEXT:    v_cndmask_b32_e64 v4, v6, v7, s[0:1]
; GFX942-NEXT:    v_perm_b32 v4, v5, v4, s5
; GFX942-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX942-NEXT:    s_or_b64 s[2:3], vcc, s[2:3]
; GFX942-NEXT:    v_mov_b32_e32 v3, v4
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[2:3]
; GFX942-NEXT:    s_cbranch_execnz .LBB27_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[2:3]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v3, v0 offset:65532
; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-NEXT:    s_mov_b32 s1, 0
; GFX11-NEXT:    s_set_inst_prefetch_distance 0x1
; GFX11-NEXT:    .p2align 6
; GFX11-NEXT:  .LBB27_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_dual_sub_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3
; GFX11-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX11-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX11-NEXT:    v_cmp_u_f32_e64 s0, v4, v4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc_lo
; GFX11-NEXT:    v_cndmask_b32_e64 v4, v6, v8, s0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_perm_b32 v4, v5, v4, 0x7060302
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v3
; GFX11-NEXT:    v_mov_b32_e32 v3, v4
; GFX11-NEXT:    s_or_b32 s1, vcc_lo, s1
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s1
; GFX11-NEXT:    s_cbranch_execnz .LBB27_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_set_inst_prefetch_distance 0x2
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s1
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v3, v0 offset:65532
; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX10-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX10-NEXT:    s_mov_b32 s5, 0
; GFX10-NEXT:  .LBB27_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX10-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX10-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX10-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX10-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX10-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX10-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX10-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX10-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX10-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX10-NEXT:    v_cmp_u_f32_e64 s4, v4, v4
; GFX10-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc_lo
; GFX10-NEXT:    v_cndmask_b32_e64 v4, v6, v8, s4
; GFX10-NEXT:    v_perm_b32 v4, v5, v4, 0x7060302
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v4, v3
; GFX10-NEXT:    v_mov_b32_e32 v3, v4
; GFX10-NEXT:    s_or_b32 s5, vcc_lo, s5
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s5
; GFX10-NEXT:    s_cbranch_execnz .LBB27_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s5
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v3, v0 offset:65532
; GFX90A-NEXT:    s_mov_b64 s[6:7], 0
; GFX90A-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX90A-NEXT:    s_movk_i32 s8, 0x7fff
; GFX90A-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX90A-NEXT:    s_mov_b32 s9, 0x7060302
; GFX90A-NEXT:  .LBB27_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX90A-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX90A-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX90A-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX90A-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX90A-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX90A-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX90A-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX90A-NEXT:    v_add3_u32 v6, v6, v4, s8
; GFX90A-NEXT:    v_add3_u32 v8, v8, v5, s8
; GFX90A-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX90A-NEXT:    v_cmp_u_f32_e64 s[4:5], v4, v4
; GFX90A-NEXT:    v_cndmask_b32_e64 v4, v6, v7, s[4:5]
; GFX90A-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX90A-NEXT:    v_perm_b32 v4, v5, v4, s9
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX90A-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
; GFX90A-NEXT:    v_mov_b32_e32 v3, v4
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[6:7]
; GFX90A-NEXT:    s_cbranch_execnz .LBB27_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v3, v0 offset:65532
; GFX908-NEXT:    s_mov_b64 s[6:7], 0
; GFX908-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX908-NEXT:    s_movk_i32 s8, 0x7fff
; GFX908-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX908-NEXT:    s_mov_b32 s9, 0x7060302
; GFX908-NEXT:  .LBB27_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX908-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX908-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX908-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX908-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX908-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX908-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX908-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX908-NEXT:    v_add3_u32 v6, v6, v4, s8
; GFX908-NEXT:    v_add3_u32 v8, v8, v5, s8
; GFX908-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX908-NEXT:    v_cmp_u_f32_e64 s[4:5], v4, v4
; GFX908-NEXT:    v_cndmask_b32_e64 v4, v6, v7, s[4:5]
; GFX908-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX908-NEXT:    v_perm_b32 v4, v5, v4, s9
; GFX908-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX908-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
; GFX908-NEXT:    v_mov_b32_e32 v3, v4
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[6:7]
; GFX908-NEXT:    s_cbranch_execnz .LBB27_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v3, v0 offset:65532
; GFX8-NEXT:    s_mov_b64 s[6:7], 0
; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX8-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX8-NEXT:  .LBB27_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
; GFX8-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX8-NEXT:    v_sub_f32_e32 v4, v4, v2
; GFX8-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX8-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX8-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, v6, v4
; GFX8-NEXT:    v_add_u32_e32 v8, vcc, v8, v5
; GFX8-NEXT:    v_add_u32_e32 v6, vcc, 0x7fff, v6
; GFX8-NEXT:    v_add_u32_e32 v8, vcc, 0x7fff, v8
; GFX8-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX8-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX8-NEXT:    v_cmp_u_f32_e64 s[4:5], v4, v4
; GFX8-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc
; GFX8-NEXT:    v_cndmask_b32_e64 v4, v6, v7, s[4:5]
; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX8-NEXT:    v_alignbit_b32 v4, v5, v4, 16
; GFX8-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX8-NEXT:    s_or_b64 s[6:7], vcc, s[6:7]
; GFX8-NEXT:    v_mov_b32_e32 v3, v4
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[6:7]
; GFX8-NEXT:    s_cbranch_execnz .LBB27_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[6:7]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v4, v0 offset:65532
; GFX7-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX7-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX7-NEXT:  .LBB27_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX7-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX7-NEXT:    v_and_b32_e32 v6, 0xffff0000, v3
; GFX7-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX7-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX7-NEXT:    v_sub_f32_e32 v6, v6, v2
; GFX7-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX7-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; GFX7-NEXT:    v_lshrrev_b32_e32 v4, 16, v6
; GFX7-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; GFX7-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB27_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_v2bf16__ofset:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    v_add_i32_e32 v0, vcc, 0xfffc, v0
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v4, v0
; GFX6-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; GFX6-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX6-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX6-NEXT:  .LBB27_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; GFX6-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; GFX6-NEXT:    v_and_b32_e32 v6, 0xffff0000, v3
; GFX6-NEXT:    v_and_b32_e32 v5, 0xffff0000, v4
; GFX6-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX6-NEXT:    v_sub_f32_e32 v6, v6, v2
; GFX6-NEXT:    v_sub_f32_e32 v5, v5, v1
; GFX6-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; GFX6-NEXT:    v_lshrrev_b32_e32 v4, 16, v6
; GFX6-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; GFX6-NEXT:    ds_cmpst_rtn_b32 v4, v0, v3, v4
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v3
; GFX6-NEXT:    v_and_b32_e32 v3, 0xffff0000, v4
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB27_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %gep = getelementptr <2 x bfloat>, ptr addrspace(3) %ptr, i32 16383
  %result = atomicrmw fsub ptr addrspace(3) %gep, <2 x bfloat> %val seq_cst
  ret void
}

; --------------------------------------------------------------------
; misc
; --------------------------------------------------------------------

define float @local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode(ptr addrspace(3) %ptr) {
; GFX12-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v1, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB28_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_mov_b32_e32 v2, v1
; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB28_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    v_mov_b32_e32 v0, v1
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v1, v0
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB28_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_mov_b32_e32 v2, v1
; GFX942-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX942-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB28_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v0, v1
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v1, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB28_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_mov_b32_e32 v2, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB28_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    v_mov_b32_e32 v0, v1
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v1, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB28_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_mov_b32_e32 v2, v1
; GFX10-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v1, v2
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB28_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    v_mov_b32_e32 v0, v1
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v1, v0
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB28_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_mov_b32_e32 v2, v1
; GFX90A-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB28_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v0, v1
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v1, v0
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB28_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_mov_b32_e32 v2, v1
; GFX908-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX908-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB28_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v0, v1
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v1, v0
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB28_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_mov_b32_e32 v2, v1
; GFX8-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX8-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB28_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v0, v1
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v1, v0
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB28_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_mov_b32_e32 v2, v1
; GFX7-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX7-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB28_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v0, v1
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v1, v0
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB28_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_mov_b32_e32 v2, v1
; GFX6-NEXT:    v_add_f32_e32 v1, -4.0, v2
; GFX6-NEXT:    ds_cmpst_rtn_b32 v1, v0, v2, v1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v1, v2
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB28_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v0, v1
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, float 4.0 seq_cst, !amdgpu.ignore.denormal.mode !0
  ret float %result
}

define void @local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode(ptr addrspace(3) %ptr) {
; GFX12-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
; GFX12:       ; %bb.0:
; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
; GFX12-NEXT:    s_wait_expcnt 0x0
; GFX12-NEXT:    s_wait_samplecnt 0x0
; GFX12-NEXT:    s_wait_bvhcnt 0x0
; GFX12-NEXT:    s_wait_kmcnt 0x0
; GFX12-NEXT:    ds_load_b32 v1, v0
; GFX12-NEXT:    s_mov_b32 s0, 0
; GFX12-NEXT:  .LBB29_1: ; %atomicrmw.start
; GFX12-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX12-NEXT:    s_wait_storecnt 0x0
; GFX12-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v1
; GFX12-NEXT:    s_wait_dscnt 0x0
; GFX12-NEXT:    global_inv scope:SCOPE_SE
; GFX12-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX12-NEXT:    v_mov_b32_e32 v1, v2
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX12-NEXT:    s_wait_alu 0xfffe
; GFX12-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_cbranch_execnz .LBB29_1
; GFX12-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX12-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX12-NEXT:    s_setpc_b64 s[30:31]
;
; GFX942-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
; GFX942:       ; %bb.0:
; GFX942-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX942-NEXT:    ds_read_b32 v1, v0
; GFX942-NEXT:    s_mov_b64 s[0:1], 0
; GFX942-NEXT:  .LBB29_1: ; %atomicrmw.start
; GFX942-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX942-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX942-NEXT:    s_waitcnt lgkmcnt(0)
; GFX942-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX942-NEXT:    s_or_b64 s[0:1], vcc, s[0:1]
; GFX942-NEXT:    v_mov_b32_e32 v1, v2
; GFX942-NEXT:    s_andn2_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_cbranch_execnz .LBB29_1
; GFX942-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX942-NEXT:    s_or_b64 exec, exec, s[0:1]
; GFX942-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    ds_load_b32 v1, v0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:  .LBB29_1: ; %atomicrmw.start
; GFX11-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX11-NEXT:    ds_cmpstore_rtn_b32 v2, v0, v2, v1
; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-NEXT:    buffer_gl0_inv
; GFX11-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX11-NEXT:    v_mov_b32_e32 v1, v2
; GFX11-NEXT:    s_or_b32 s0, vcc_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_execnz .LBB29_1
; GFX11-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_setpc_b64 s[30:31]
;
; GFX10-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
; GFX10:       ; %bb.0:
; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT:    ds_read_b32 v1, v0
; GFX10-NEXT:    s_mov_b32 s4, 0
; GFX10-NEXT:  .LBB29_1: ; %atomicrmw.start
; GFX10-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
; GFX10-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-NEXT:    buffer_gl0_inv
; GFX10-NEXT:    v_cmp_eq_u32_e32 vcc_lo, v2, v1
; GFX10-NEXT:    v_mov_b32_e32 v1, v2
; GFX10-NEXT:    s_or_b32 s4, vcc_lo, s4
; GFX10-NEXT:    s_andn2_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_cbranch_execnz .LBB29_1
; GFX10-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX10-NEXT:    s_or_b32 exec_lo, exec_lo, s4
; GFX10-NEXT:    s_setpc_b64 s[30:31]
;
; GFX90A-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
; GFX90A:       ; %bb.0:
; GFX90A-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX90A-NEXT:    ds_read_b32 v1, v0
; GFX90A-NEXT:    s_mov_b64 s[4:5], 0
; GFX90A-NEXT:  .LBB29_1: ; %atomicrmw.start
; GFX90A-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX90A-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX90A-NEXT:    s_waitcnt lgkmcnt(0)
; GFX90A-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX90A-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX90A-NEXT:    v_mov_b32_e32 v1, v2
; GFX90A-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_cbranch_execnz .LBB29_1
; GFX90A-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX90A-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX90A-NEXT:    s_setpc_b64 s[30:31]
;
; GFX908-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
; GFX908:       ; %bb.0:
; GFX908-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX908-NEXT:    ds_read_b32 v1, v0
; GFX908-NEXT:    s_mov_b64 s[4:5], 0
; GFX908-NEXT:  .LBB29_1: ; %atomicrmw.start
; GFX908-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX908-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX908-NEXT:    s_waitcnt lgkmcnt(0)
; GFX908-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX908-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX908-NEXT:    v_mov_b32_e32 v1, v2
; GFX908-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_cbranch_execnz .LBB29_1
; GFX908-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX908-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX908-NEXT:    s_setpc_b64 s[30:31]
;
; GFX8-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
; GFX8:       ; %bb.0:
; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX8-NEXT:    s_mov_b32 m0, -1
; GFX8-NEXT:    ds_read_b32 v1, v0
; GFX8-NEXT:    s_mov_b64 s[4:5], 0
; GFX8-NEXT:  .LBB29_1: ; %atomicrmw.start
; GFX8-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX8-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
; GFX8-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX8-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX8-NEXT:    v_mov_b32_e32 v1, v2
; GFX8-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_cbranch_execnz .LBB29_1
; GFX8-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX8-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX8-NEXT:    s_setpc_b64 s[30:31]
;
; GFX7-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
; GFX7:       ; %bb.0:
; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX7-NEXT:    s_mov_b32 m0, -1
; GFX7-NEXT:    ds_read_b32 v1, v0
; GFX7-NEXT:    s_mov_b64 s[4:5], 0
; GFX7-NEXT:  .LBB29_1: ; %atomicrmw.start
; GFX7-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX7-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
; GFX7-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX7-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX7-NEXT:    v_mov_b32_e32 v1, v2
; GFX7-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_cbranch_execnz .LBB29_1
; GFX7-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX7-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX7-NEXT:    s_setpc_b64 s[30:31]
;
; GFX6-LABEL: local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode:
; GFX6:       ; %bb.0:
; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX6-NEXT:    s_mov_b32 m0, -1
; GFX6-NEXT:    ds_read_b32 v1, v0
; GFX6-NEXT:    s_mov_b64 s[4:5], 0
; GFX6-NEXT:  .LBB29_1: ; %atomicrmw.start
; GFX6-NEXT:    ; =>This Inner Loop Header: Depth=1
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_add_f32_e32 v2, -4.0, v1
; GFX6-NEXT:    ds_cmpst_rtn_b32 v2, v0, v1, v2
; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
; GFX6-NEXT:    v_cmp_eq_u32_e32 vcc, v2, v1
; GFX6-NEXT:    s_or_b64 s[4:5], vcc, s[4:5]
; GFX6-NEXT:    v_mov_b32_e32 v1, v2
; GFX6-NEXT:    s_andn2_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_cbranch_execnz .LBB29_1
; GFX6-NEXT:  ; %bb.2: ; %atomicrmw.end
; GFX6-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX6-NEXT:    s_setpc_b64 s[30:31]
  %result = atomicrmw fsub ptr addrspace(3) %ptr, float 4.0 seq_cst, !amdgpu.ignore.denormal.mode !0
  ret void
}

!0 = !{}