; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn-amd-amdhsa -global-isel=0 -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX9,GFX9-SDAG %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -global-isel=1 -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX9,GFX9-GISEL %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -global-isel=0 -mcpu=gfx940 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX940,GFX940-SDAG %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -global-isel=1 -mcpu=gfx940 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX940,GFX940-GISEL %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -global-isel=0 -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10,GFX10-SDAG %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -global-isel=1 -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10,GFX10-GISEL %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -global-isel=0 -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX11,GFX11-SDAG %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -global-isel=1 -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX11,GFX11-GISEL %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -global-isel=0 -mcpu=gfx1200 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX12,GFX12-SDAG %s
; RUN: llc -mtriple=amdgcn-amd-amdhsa -global-isel=1 -mcpu=gfx1200 -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX12,GFX12-GISEL %s

define amdgpu_kernel void @buffer_nontemporal_load_store(ptr addrspace(7) %in, ptr addrspace(7) %out) {
; GFX9-SDAG-LABEL: buffer_nontemporal_load_store:
; GFX9-SDAG:       ; %bb.0: ; %entry
; GFX9-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[8:9], 0x0
; GFX9-SDAG-NEXT:    s_load_dword s11, s[8:9], 0x10
; GFX9-SDAG-NEXT:    s_mov_b32 s10, 0
; GFX9-SDAG-NEXT:    s_add_u32 s0, s0, s15
; GFX9-SDAG-NEXT:    s_mov_b32 s15, s10
; GFX9-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-SDAG-NEXT:    s_mov_b32 s14, s7
; GFX9-SDAG-NEXT:    s_addc_u32 s1, s1, 0
; GFX9-SDAG-NEXT:    s_mov_b32 s12, s5
; GFX9-SDAG-NEXT:    s_or_b64 s[14:15], s[14:15], s[10:11]
; GFX9-SDAG-NEXT:    s_mov_b32 s13, s6
; GFX9-SDAG-NEXT:    v_mov_b32_e32 v0, s4
; GFX9-SDAG-NEXT:    buffer_load_dword v0, v0, s[12:15], 0 offen glc slc
; GFX9-SDAG-NEXT:    s_load_dword s11, s[8:9], 0x30
; GFX9-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[8:9], 0x20
; GFX9-SDAG-NEXT:    s_mov_b32 s9, s10
; GFX9-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-SDAG-NEXT:    s_mov_b32 s8, s7
; GFX9-SDAG-NEXT:    s_or_b64 s[10:11], s[8:9], s[10:11]
; GFX9-SDAG-NEXT:    s_mov_b32 s8, s5
; GFX9-SDAG-NEXT:    s_mov_b32 s9, s6
; GFX9-SDAG-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-SDAG-NEXT:    s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT:    buffer_store_dword v0, v1, s[8:11], 0 offen glc slc
; GFX9-SDAG-NEXT:    s_endpgm
;
; GFX9-GISEL-LABEL: buffer_nontemporal_load_store:
; GFX9-GISEL:       ; %bb.0: ; %entry
; GFX9-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x0
; GFX9-GISEL-NEXT:    s_load_dword s7, s[8:9], 0x10
; GFX9-GISEL-NEXT:    s_mov_b32 s11, 0
; GFX9-GISEL-NEXT:    s_mov_b32 s4, s11
; GFX9-GISEL-NEXT:    s_mov_b32 s6, s11
; GFX9-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-GISEL-NEXT:    s_mov_b32 s10, s1
; GFX9-GISEL-NEXT:    s_mov_b32 s5, s2
; GFX9-GISEL-NEXT:    s_or_b64 s[4:5], s[10:11], s[4:5]
; GFX9-GISEL-NEXT:    s_mov_b32 s10, s3
; GFX9-GISEL-NEXT:    s_or_b64 s[6:7], s[10:11], s[6:7]
; GFX9-GISEL-NEXT:    v_mov_b32_e32 v0, s0
; GFX9-GISEL-NEXT:    buffer_load_dword v0, v0, s[4:7], 0 offen glc slc
; GFX9-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x20
; GFX9-GISEL-NEXT:    s_load_dword s7, s[8:9], 0x30
; GFX9-GISEL-NEXT:    s_mov_b32 s4, s11
; GFX9-GISEL-NEXT:    s_mov_b32 s6, s11
; GFX9-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-GISEL-NEXT:    s_mov_b32 s10, s1
; GFX9-GISEL-NEXT:    s_mov_b32 s5, s2
; GFX9-GISEL-NEXT:    s_or_b64 s[4:5], s[10:11], s[4:5]
; GFX9-GISEL-NEXT:    s_mov_b32 s10, s3
; GFX9-GISEL-NEXT:    s_or_b64 s[6:7], s[10:11], s[6:7]
; GFX9-GISEL-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-GISEL-NEXT:    s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT:    buffer_store_dword v0, v1, s[4:7], 0 offen glc slc
; GFX9-GISEL-NEXT:    s_endpgm
;
; GFX940-SDAG-LABEL: buffer_nontemporal_load_store:
; GFX940-SDAG:       ; %bb.0: ; %entry
; GFX940-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX940-SDAG-NEXT:    s_load_dword s6, s[4:5], 0x10
; GFX940-SDAG-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x20
; GFX940-SDAG-NEXT:    s_load_dword s7, s[4:5], 0x30
; GFX940-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX940-SDAG-NEXT:    v_mov_b64_e32 v[0:1], s[0:1]
; GFX940-SDAG-NEXT:    v_mov_b64_e32 v[2:3], s[2:3]
; GFX940-SDAG-NEXT:    scratch_store_dwordx4 off, v[0:3], off offset:32 sc0 sc1
; GFX940-SDAG-NEXT:    scratch_load_dwordx2 v[10:11], off, off offset:40
; GFX940-SDAG-NEXT:    scratch_load_dword v4, off, off offset:36
; GFX940-SDAG-NEXT:    v_mov_b64_e32 v[0:1], s[8:9]
; GFX940-SDAG-NEXT:    v_mov_b64_e32 v[2:3], s[10:11]
; GFX940-SDAG-NEXT:    scratch_store_dwordx4 off, v[0:3], off sc0 sc1
; GFX940-SDAG-NEXT:    scratch_load_dwordx2 v[12:13], off, off offset:8
; GFX940-SDAG-NEXT:    s_nop 0
; GFX940-SDAG-NEXT:    scratch_load_dword v0, off, off offset:4
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v7, s6
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v3, s7
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v9, s0
; GFX940-SDAG-NEXT:    s_mov_b64 s[2:3], exec
; GFX940-SDAG-NEXT:    s_waitcnt vmcnt(4)
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v6, v11
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v5, v10
; GFX940-SDAG-NEXT:    s_waitcnt vmcnt(1)
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v2, v13
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v1, v12
; GFX940-SDAG-NEXT:  .LBB0_1: ; =>This Inner Loop Header: Depth=1
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s4, v4
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s5, v5
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s6, v6
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s7, v7
; GFX940-SDAG-NEXT:    v_cmp_eq_u64_e32 vcc, s[4:5], v[4:5]
; GFX940-SDAG-NEXT:    s_nop 0
; GFX940-SDAG-NEXT:    v_cmp_eq_u64_e64 s[0:1], s[6:7], v[6:7]
; GFX940-SDAG-NEXT:    s_and_b64 s[0:1], vcc, s[0:1]
; GFX940-SDAG-NEXT:    s_and_saveexec_b64 s[0:1], s[0:1]
; GFX940-SDAG-NEXT:    buffer_load_dword v8, v9, s[4:7], 0 offen nt
; GFX940-SDAG-NEXT:    ; implicit-def: $vgpr4_vgpr5_vgpr6_vgpr7
; GFX940-SDAG-NEXT:    ; implicit-def: $vgpr9
; GFX940-SDAG-NEXT:    s_xor_b64 exec, exec, s[0:1]
; GFX940-SDAG-NEXT:    s_cbranch_execnz .LBB0_1
; GFX940-SDAG-NEXT:  ; %bb.2:
; GFX940-SDAG-NEXT:    s_mov_b64 exec, s[2:3]
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v4, s8
; GFX940-SDAG-NEXT:    s_mov_b64 s[0:1], exec
; GFX940-SDAG-NEXT:    s_waitcnt vmcnt(0)
; GFX940-SDAG-NEXT:  .LBB0_3: ; =>This Inner Loop Header: Depth=1
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s4, v0
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s5, v1
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s6, v2
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s7, v3
; GFX940-SDAG-NEXT:    v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
; GFX940-SDAG-NEXT:    s_nop 0
; GFX940-SDAG-NEXT:    v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
; GFX940-SDAG-NEXT:    s_and_b64 s[0:1], vcc, s[0:1]
; GFX940-SDAG-NEXT:    s_and_saveexec_b64 s[0:1], s[0:1]
; GFX940-SDAG-NEXT:    buffer_store_dword v8, v4, s[4:7], 0 offen sc0 nt sc1
; GFX940-SDAG-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX940-SDAG-NEXT:    ; implicit-def: $vgpr8
; GFX940-SDAG-NEXT:    ; implicit-def: $vgpr4
; GFX940-SDAG-NEXT:    s_xor_b64 exec, exec, s[0:1]
; GFX940-SDAG-NEXT:    s_cbranch_execnz .LBB0_3
; GFX940-SDAG-NEXT:  ; %bb.4:
; GFX940-SDAG-NEXT:    s_endpgm
;
; GFX940-GISEL-LABEL: buffer_nontemporal_load_store:
; GFX940-GISEL:       ; %bb.0: ; %entry
; GFX940-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX940-GISEL-NEXT:    s_load_dword s11, s[4:5], 0x10
; GFX940-GISEL-NEXT:    s_mov_b32 s7, 0
; GFX940-GISEL-NEXT:    s_mov_b32 s8, s7
; GFX940-GISEL-NEXT:    s_mov_b32 s10, s7
; GFX940-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX940-GISEL-NEXT:    s_mov_b32 s6, s1
; GFX940-GISEL-NEXT:    s_mov_b32 s9, s2
; GFX940-GISEL-NEXT:    s_or_b64 s[8:9], s[6:7], s[8:9]
; GFX940-GISEL-NEXT:    s_mov_b32 s6, s3
; GFX940-GISEL-NEXT:    s_or_b64 s[10:11], s[6:7], s[10:11]
; GFX940-GISEL-NEXT:    v_mov_b32_e32 v0, s0
; GFX940-GISEL-NEXT:    buffer_load_dword v0, v0, s[8:11], 0 offen nt
; GFX940-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x20
; GFX940-GISEL-NEXT:    s_load_dword s9, s[4:5], 0x30
; GFX940-GISEL-NEXT:    s_mov_b32 s4, s7
; GFX940-GISEL-NEXT:    s_mov_b32 s8, s7
; GFX940-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX940-GISEL-NEXT:    s_mov_b32 s6, s1
; GFX940-GISEL-NEXT:    s_mov_b32 s5, s2
; GFX940-GISEL-NEXT:    s_or_b64 s[4:5], s[6:7], s[4:5]
; GFX940-GISEL-NEXT:    s_mov_b32 s6, s3
; GFX940-GISEL-NEXT:    s_or_b64 s[6:7], s[6:7], s[8:9]
; GFX940-GISEL-NEXT:    v_mov_b32_e32 v1, s0
; GFX940-GISEL-NEXT:    s_waitcnt vmcnt(0)
; GFX940-GISEL-NEXT:    buffer_store_dword v0, v1, s[4:7], 0 offen sc0 nt sc1
; GFX940-GISEL-NEXT:    s_endpgm
;
; GFX10-SDAG-LABEL: buffer_nontemporal_load_store:
; GFX10-SDAG:       ; %bb.0: ; %entry
; GFX10-SDAG-NEXT:    s_clause 0x1
; GFX10-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[8:9], 0x0
; GFX10-SDAG-NEXT:    s_load_dword s11, s[8:9], 0x10
; GFX10-SDAG-NEXT:    s_mov_b32 s10, 0
; GFX10-SDAG-NEXT:    s_add_u32 s0, s0, s15
; GFX10-SDAG-NEXT:    s_mov_b32 s13, s10
; GFX10-SDAG-NEXT:    s_addc_u32 s1, s1, 0
; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-SDAG-NEXT:    v_mov_b32_e32 v0, s4
; GFX10-SDAG-NEXT:    s_mov_b32 s12, s7
; GFX10-SDAG-NEXT:    s_or_b64 s[14:15], s[12:13], s[10:11]
; GFX10-SDAG-NEXT:    s_mov_b32 s12, s5
; GFX10-SDAG-NEXT:    s_mov_b32 s13, s6
; GFX10-SDAG-NEXT:    buffer_load_dword v0, v0, s[12:15], 0 offen slc
; GFX10-SDAG-NEXT:    s_clause 0x1
; GFX10-SDAG-NEXT:    s_load_dword s11, s[8:9], 0x30
; GFX10-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[8:9], 0x20
; GFX10-SDAG-NEXT:    s_mov_b32 s9, s10
; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-SDAG-NEXT:    v_mov_b32_e32 v1, s4
; GFX10-SDAG-NEXT:    s_mov_b32 s8, s7
; GFX10-SDAG-NEXT:    s_or_b64 s[10:11], s[8:9], s[10:11]
; GFX10-SDAG-NEXT:    s_mov_b32 s8, s5
; GFX10-SDAG-NEXT:    s_mov_b32 s9, s6
; GFX10-SDAG-NEXT:    s_waitcnt vmcnt(0)
; GFX10-SDAG-NEXT:    buffer_store_dword v0, v1, s[8:11], 0 offen glc slc
; GFX10-SDAG-NEXT:    s_endpgm
;
; GFX10-GISEL-LABEL: buffer_nontemporal_load_store:
; GFX10-GISEL:       ; %bb.0: ; %entry
; GFX10-GISEL-NEXT:    s_clause 0x1
; GFX10-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x0
; GFX10-GISEL-NEXT:    s_load_dword s5, s[8:9], 0x10
; GFX10-GISEL-NEXT:    s_mov_b32 s7, 0
; GFX10-GISEL-NEXT:    s_mov_b32 s10, s7
; GFX10-GISEL-NEXT:    s_mov_b32 s4, s7
; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-GISEL-NEXT:    s_mov_b32 s6, s1
; GFX10-GISEL-NEXT:    s_mov_b32 s11, s2
; GFX10-GISEL-NEXT:    v_mov_b32_e32 v0, s0
; GFX10-GISEL-NEXT:    s_or_b64 s[0:1], s[6:7], s[10:11]
; GFX10-GISEL-NEXT:    s_mov_b32 s6, s3
; GFX10-GISEL-NEXT:    s_or_b64 s[2:3], s[6:7], s[4:5]
; GFX10-GISEL-NEXT:    buffer_load_dword v0, v0, s[0:3], 0 offen slc
; GFX10-GISEL-NEXT:    s_clause 0x1
; GFX10-GISEL-NEXT:    s_waitcnt_depctr 0xffe3
; GFX10-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x20
; GFX10-GISEL-NEXT:    s_load_dword s11, s[8:9], 0x30
; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-GISEL-NEXT:    s_mov_b32 s6, s1
; GFX10-GISEL-NEXT:    s_mov_b32 s5, s2
; GFX10-GISEL-NEXT:    v_mov_b32_e32 v1, s0
; GFX10-GISEL-NEXT:    s_or_b64 s[4:5], s[6:7], s[4:5]
; GFX10-GISEL-NEXT:    s_mov_b32 s6, s3
; GFX10-GISEL-NEXT:    s_or_b64 s[6:7], s[6:7], s[10:11]
; GFX10-GISEL-NEXT:    s_waitcnt vmcnt(0)
; GFX10-GISEL-NEXT:    buffer_store_dword v0, v1, s[4:7], 0 offen glc slc
; GFX10-GISEL-NEXT:    s_endpgm
;
; GFX11-SDAG-LABEL: buffer_nontemporal_load_store:
; GFX11-SDAG:       ; %bb.0: ; %entry
; GFX11-SDAG-NEXT:    s_clause 0x2
; GFX11-SDAG-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
; GFX11-SDAG-NEXT:    s_load_b128 s[8:11], s[4:5], 0x20
; GFX11-SDAG-NEXT:    s_load_b32 s6, s[4:5], 0x10
; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-SDAG-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-SDAG-NEXT:    v_dual_mov_b32 v7, s8 :: v_dual_mov_b32 v8, s9
; GFX11-SDAG-NEXT:    v_dual_mov_b32 v9, s10 :: v_dual_mov_b32 v10, s11
; GFX11-SDAG-NEXT:    scratch_store_b128 off, v[0:3], off offset:32
; GFX11-SDAG-NEXT:    s_clause 0x1
; GFX11-SDAG-NEXT:    scratch_load_b64 v[5:6], off, off offset:40
; GFX11-SDAG-NEXT:    scratch_load_b32 v4, off, off offset:36
; GFX11-SDAG-NEXT:    s_load_b32 s1, s[4:5], 0x30
; GFX11-SDAG-NEXT:    scratch_store_b128 off, v[7:10], off
; GFX11-SDAG-NEXT:    s_clause 0x1
; GFX11-SDAG-NEXT:    scratch_load_b64 v[1:2], off, off offset:8
; GFX11-SDAG-NEXT:    scratch_load_b32 v0, off, off offset:4
; GFX11-SDAG-NEXT:    v_mov_b32_e32 v7, s6
; GFX11-SDAG-NEXT:    v_mov_b32_e32 v9, s0
; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, s1
; GFX11-SDAG-NEXT:    s_mov_b32 s1, exec_lo
; GFX11-SDAG-NEXT:    s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT:  .LBB0_1: ; =>This Inner Loop Header: Depth=1
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s4, v4
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s5, v5
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s6, v6
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s7, v7
; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-SDAG-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[4:5]
; GFX11-SDAG-NEXT:    v_cmp_eq_u64_e64 s0, s[6:7], v[6:7]
; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT:    s_and_b32 s0, vcc_lo, s0
; GFX11-SDAG-NEXT:    s_and_saveexec_b32 s0, s0
; GFX11-SDAG-NEXT:    buffer_load_b32 v8, v9, s[4:7], 0 offen slc dlc
; GFX11-SDAG-NEXT:    ; implicit-def: $vgpr4_vgpr5_vgpr6_vgpr7
; GFX11-SDAG-NEXT:    ; implicit-def: $vgpr9
; GFX11-SDAG-NEXT:    s_xor_b32 exec_lo, exec_lo, s0
; GFX11-SDAG-NEXT:    s_cbranch_execnz .LBB0_1
; GFX11-SDAG-NEXT:  ; %bb.2:
; GFX11-SDAG-NEXT:    s_mov_b32 exec_lo, s1
; GFX11-SDAG-NEXT:    v_mov_b32_e32 v4, s8
; GFX11-SDAG-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-SDAG-NEXT:  .LBB0_3: ; =>This Inner Loop Header: Depth=1
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s4, v0
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s5, v1
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s6, v2
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s7, v3
; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-SDAG-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
; GFX11-SDAG-NEXT:    v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT:    s_and_b32 s0, vcc_lo, s0
; GFX11-SDAG-NEXT:    s_and_saveexec_b32 s0, s0
; GFX11-SDAG-NEXT:    s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT:    buffer_store_b32 v8, v4, s[4:7], 0 offen glc slc dlc
; GFX11-SDAG-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-SDAG-NEXT:    ; implicit-def: $vgpr8
; GFX11-SDAG-NEXT:    ; implicit-def: $vgpr4
; GFX11-SDAG-NEXT:    s_xor_b32 exec_lo, exec_lo, s0
; GFX11-SDAG-NEXT:    s_cbranch_execnz .LBB0_3
; GFX11-SDAG-NEXT:  ; %bb.4:
; GFX11-SDAG-NEXT:    s_endpgm
;
; GFX11-GISEL-LABEL: buffer_nontemporal_load_store:
; GFX11-GISEL:       ; %bb.0: ; %entry
; GFX11-GISEL-NEXT:    s_clause 0x1
; GFX11-GISEL-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
; GFX11-GISEL-NEXT:    s_load_b32 s7, s[4:5], 0x10
; GFX11-GISEL-NEXT:    s_mov_b32 s9, 0
; GFX11-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT:    s_mov_b32 s10, s9
; GFX11-GISEL-NEXT:    s_mov_b32 s6, s9
; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT:    s_mov_b32 s8, s1
; GFX11-GISEL-NEXT:    s_mov_b32 s11, s2
; GFX11-GISEL-NEXT:    v_mov_b32_e32 v0, s0
; GFX11-GISEL-NEXT:    s_or_b64 s[0:1], s[8:9], s[10:11]
; GFX11-GISEL-NEXT:    s_mov_b32 s8, s3
; GFX11-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT:    s_or_b64 s[2:3], s[8:9], s[6:7]
; GFX11-GISEL-NEXT:    buffer_load_b32 v0, v0, s[0:3], 0 offen slc dlc
; GFX11-GISEL-NEXT:    s_clause 0x1
; GFX11-GISEL-NEXT:    s_load_b128 s[0:3], s[4:5], 0x20
; GFX11-GISEL-NEXT:    s_load_b32 s7, s[4:5], 0x30
; GFX11-GISEL-NEXT:    s_mov_b32 s4, s9
; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, s0
; GFX11-GISEL-NEXT:    s_mov_b32 s8, s1
; GFX11-GISEL-NEXT:    s_mov_b32 s5, s2
; GFX11-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT:    s_or_b64 s[4:5], s[8:9], s[4:5]
; GFX11-GISEL-NEXT:    s_mov_b32 s8, s3
; GFX11-GISEL-NEXT:    s_or_b64 s[6:7], s[8:9], s[6:7]
; GFX11-GISEL-NEXT:    s_waitcnt vmcnt(0)
; GFX11-GISEL-NEXT:    buffer_store_b32 v0, v1, s[4:7], 0 offen glc slc dlc
; GFX11-GISEL-NEXT:    s_endpgm
;
; GFX12-SDAG-LABEL: buffer_nontemporal_load_store:
; GFX12-SDAG:       ; %bb.0: ; %entry
; GFX12-SDAG-NEXT:    s_clause 0x2
; GFX12-SDAG-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
; GFX12-SDAG-NEXT:    s_load_b128 s[8:11], s[4:5], 0x20
; GFX12-SDAG-NEXT:    s_load_b32 s6, s[4:5], 0x10
; GFX12-SDAG-NEXT:    s_wait_kmcnt 0x0
; GFX12-SDAG-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX12-SDAG-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX12-SDAG-NEXT:    v_dual_mov_b32 v7, s8 :: v_dual_mov_b32 v8, s9
; GFX12-SDAG-NEXT:    v_dual_mov_b32 v9, s10 :: v_dual_mov_b32 v10, s11
; GFX12-SDAG-NEXT:    scratch_store_b128 off, v[0:3], off offset:32
; GFX12-SDAG-NEXT:    s_clause 0x1
; GFX12-SDAG-NEXT:    scratch_load_b64 v[5:6], off, off offset:40
; GFX12-SDAG-NEXT:    scratch_load_b32 v4, off, off offset:36
; GFX12-SDAG-NEXT:    s_load_b32 s1, s[4:5], 0x30
; GFX12-SDAG-NEXT:    scratch_store_b128 off, v[7:10], off
; GFX12-SDAG-NEXT:    s_clause 0x1
; GFX12-SDAG-NEXT:    scratch_load_b64 v[1:2], off, off offset:8
; GFX12-SDAG-NEXT:    scratch_load_b32 v0, off, off offset:4
; GFX12-SDAG-NEXT:    v_mov_b32_e32 v7, s6
; GFX12-SDAG-NEXT:    v_mov_b32_e32 v9, s0
; GFX12-SDAG-NEXT:    s_wait_kmcnt 0x0
; GFX12-SDAG-NEXT:    v_mov_b32_e32 v3, s1
; GFX12-SDAG-NEXT:    s_mov_b32 s1, exec_lo
; GFX12-SDAG-NEXT:  .LBB0_1: ; =>This Inner Loop Header: Depth=1
; GFX12-SDAG-NEXT:    s_wait_loadcnt 0x2
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s4, v4
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s5, v5
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s6, v6
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s7, v7
; GFX12-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-SDAG-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[4:5]
; GFX12-SDAG-NEXT:    v_cmp_eq_u64_e64 s0, s[6:7], v[6:7]
; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
; GFX12-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-SDAG-NEXT:    s_and_b32 s0, vcc_lo, s0
; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
; GFX12-SDAG-NEXT:    s_and_saveexec_b32 s0, s0
; GFX12-SDAG-NEXT:    s_wait_loadcnt 0x0
; GFX12-SDAG-NEXT:    buffer_load_b32 v8, v9, s[4:7], null offen th:TH_LOAD_NT
; GFX12-SDAG-NEXT:    ; implicit-def: $vgpr4_vgpr5_vgpr6_vgpr7
; GFX12-SDAG-NEXT:    ; implicit-def: $vgpr9
; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
; GFX12-SDAG-NEXT:    s_xor_b32 exec_lo, exec_lo, s0
; GFX12-SDAG-NEXT:    s_cbranch_execnz .LBB0_1
; GFX12-SDAG-NEXT:  ; %bb.2:
; GFX12-SDAG-NEXT:    s_mov_b32 exec_lo, s1
; GFX12-SDAG-NEXT:    v_mov_b32_e32 v4, s8
; GFX12-SDAG-NEXT:    s_mov_b32 s0, exec_lo
; GFX12-SDAG-NEXT:  .LBB0_3: ; =>This Inner Loop Header: Depth=1
; GFX12-SDAG-NEXT:    s_wait_loadcnt 0x1
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s4, v0
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s5, v1
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s6, v2
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s7, v3
; GFX12-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-SDAG-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
; GFX12-SDAG-NEXT:    v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
; GFX12-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-SDAG-NEXT:    s_and_b32 s0, vcc_lo, s0
; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
; GFX12-SDAG-NEXT:    s_and_saveexec_b32 s0, s0
; GFX12-SDAG-NEXT:    s_wait_loadcnt 0x0
; GFX12-SDAG-NEXT:    buffer_store_b32 v8, v4, s[4:7], null offen th:TH_STORE_NT
; GFX12-SDAG-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX12-SDAG-NEXT:    ; implicit-def: $vgpr8
; GFX12-SDAG-NEXT:    ; implicit-def: $vgpr4
; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
; GFX12-SDAG-NEXT:    s_xor_b32 exec_lo, exec_lo, s0
; GFX12-SDAG-NEXT:    s_cbranch_execnz .LBB0_3
; GFX12-SDAG-NEXT:  ; %bb.4:
; GFX12-SDAG-NEXT:    s_endpgm
;
; GFX12-GISEL-LABEL: buffer_nontemporal_load_store:
; GFX12-GISEL:       ; %bb.0: ; %entry
; GFX12-GISEL-NEXT:    s_clause 0x1
; GFX12-GISEL-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
; GFX12-GISEL-NEXT:    s_load_b32 s7, s[4:5], 0x10
; GFX12-GISEL-NEXT:    s_mov_b32 s9, 0
; GFX12-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-GISEL-NEXT:    s_mov_b32 s10, s9
; GFX12-GISEL-NEXT:    s_mov_b32 s6, s9
; GFX12-GISEL-NEXT:    s_wait_kmcnt 0x0
; GFX12-GISEL-NEXT:    s_mov_b32 s8, s1
; GFX12-GISEL-NEXT:    s_mov_b32 s11, s2
; GFX12-GISEL-NEXT:    v_mov_b32_e32 v0, s0
; GFX12-GISEL-NEXT:    s_or_b64 s[0:1], s[8:9], s[10:11]
; GFX12-GISEL-NEXT:    s_mov_b32 s8, s3
; GFX12-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-GISEL-NEXT:    s_or_b64 s[2:3], s[8:9], s[6:7]
; GFX12-GISEL-NEXT:    buffer_load_b32 v0, v0, s[0:3], null offen th:TH_LOAD_NT
; GFX12-GISEL-NEXT:    s_clause 0x1
; GFX12-GISEL-NEXT:    s_load_b128 s[0:3], s[4:5], 0x20
; GFX12-GISEL-NEXT:    s_load_b32 s7, s[4:5], 0x30
; GFX12-GISEL-NEXT:    s_mov_b32 s4, s9
; GFX12-GISEL-NEXT:    s_wait_kmcnt 0x0
; GFX12-GISEL-NEXT:    s_wait_alu 0xfffe
; GFX12-GISEL-NEXT:    v_mov_b32_e32 v1, s0
; GFX12-GISEL-NEXT:    s_mov_b32 s8, s1
; GFX12-GISEL-NEXT:    s_mov_b32 s5, s2
; GFX12-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX12-GISEL-NEXT:    s_or_b64 s[4:5], s[8:9], s[4:5]
; GFX12-GISEL-NEXT:    s_mov_b32 s8, s3
; GFX12-GISEL-NEXT:    s_or_b64 s[6:7], s[8:9], s[6:7]
; GFX12-GISEL-NEXT:    s_wait_loadcnt 0x0
; GFX12-GISEL-NEXT:    buffer_store_b32 v0, v1, s[4:7], null offen th:TH_STORE_NT
; GFX12-GISEL-NEXT:    s_endpgm
entry:
  %val = load i32, ptr addrspace(7) %in, !nontemporal !0
  store i32 %val, ptr addrspace(7) %out, !nontemporal !0
  ret void
}

define amdgpu_kernel void @buffer_nontemporal_and_volatile_load_store(ptr addrspace(7) %in, ptr addrspace(7) %out) {
; GFX9-SDAG-LABEL: buffer_nontemporal_and_volatile_load_store:
; GFX9-SDAG:       ; %bb.0: ; %entry
; GFX9-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[8:9], 0x0
; GFX9-SDAG-NEXT:    s_load_dword s11, s[8:9], 0x10
; GFX9-SDAG-NEXT:    s_mov_b32 s10, 0
; GFX9-SDAG-NEXT:    s_add_u32 s0, s0, s15
; GFX9-SDAG-NEXT:    s_mov_b32 s15, s10
; GFX9-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-SDAG-NEXT:    s_mov_b32 s14, s7
; GFX9-SDAG-NEXT:    s_addc_u32 s1, s1, 0
; GFX9-SDAG-NEXT:    s_mov_b32 s12, s5
; GFX9-SDAG-NEXT:    s_or_b64 s[14:15], s[14:15], s[10:11]
; GFX9-SDAG-NEXT:    s_mov_b32 s13, s6
; GFX9-SDAG-NEXT:    v_mov_b32_e32 v0, s4
; GFX9-SDAG-NEXT:    buffer_load_dword v0, v0, s[12:15], 0 offen glc
; GFX9-SDAG-NEXT:    s_load_dword s11, s[8:9], 0x30
; GFX9-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[8:9], 0x20
; GFX9-SDAG-NEXT:    s_mov_b32 s9, s10
; GFX9-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-SDAG-NEXT:    s_mov_b32 s8, s7
; GFX9-SDAG-NEXT:    s_or_b64 s[10:11], s[8:9], s[10:11]
; GFX9-SDAG-NEXT:    s_mov_b32 s8, s5
; GFX9-SDAG-NEXT:    s_mov_b32 s9, s6
; GFX9-SDAG-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-SDAG-NEXT:    s_waitcnt vmcnt(0)
; GFX9-SDAG-NEXT:    buffer_store_dword v0, v1, s[8:11], 0 offen
; GFX9-SDAG-NEXT:    s_endpgm
;
; GFX9-GISEL-LABEL: buffer_nontemporal_and_volatile_load_store:
; GFX9-GISEL:       ; %bb.0: ; %entry
; GFX9-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x0
; GFX9-GISEL-NEXT:    s_load_dword s7, s[8:9], 0x10
; GFX9-GISEL-NEXT:    s_mov_b32 s11, 0
; GFX9-GISEL-NEXT:    s_mov_b32 s4, s11
; GFX9-GISEL-NEXT:    s_mov_b32 s6, s11
; GFX9-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-GISEL-NEXT:    s_mov_b32 s10, s1
; GFX9-GISEL-NEXT:    s_mov_b32 s5, s2
; GFX9-GISEL-NEXT:    s_or_b64 s[4:5], s[10:11], s[4:5]
; GFX9-GISEL-NEXT:    s_mov_b32 s10, s3
; GFX9-GISEL-NEXT:    s_or_b64 s[6:7], s[10:11], s[6:7]
; GFX9-GISEL-NEXT:    v_mov_b32_e32 v0, s0
; GFX9-GISEL-NEXT:    buffer_load_dword v0, v0, s[4:7], 0 offen glc
; GFX9-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x20
; GFX9-GISEL-NEXT:    s_load_dword s7, s[8:9], 0x30
; GFX9-GISEL-NEXT:    s_mov_b32 s4, s11
; GFX9-GISEL-NEXT:    s_mov_b32 s6, s11
; GFX9-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX9-GISEL-NEXT:    s_mov_b32 s10, s1
; GFX9-GISEL-NEXT:    s_mov_b32 s5, s2
; GFX9-GISEL-NEXT:    s_or_b64 s[4:5], s[10:11], s[4:5]
; GFX9-GISEL-NEXT:    s_mov_b32 s10, s3
; GFX9-GISEL-NEXT:    s_or_b64 s[6:7], s[10:11], s[6:7]
; GFX9-GISEL-NEXT:    v_mov_b32_e32 v1, s0
; GFX9-GISEL-NEXT:    s_waitcnt vmcnt(0)
; GFX9-GISEL-NEXT:    buffer_store_dword v0, v1, s[4:7], 0 offen
; GFX9-GISEL-NEXT:    s_endpgm
;
; GFX940-SDAG-LABEL: buffer_nontemporal_and_volatile_load_store:
; GFX940-SDAG:       ; %bb.0: ; %entry
; GFX940-SDAG-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX940-SDAG-NEXT:    s_load_dword s6, s[4:5], 0x10
; GFX940-SDAG-NEXT:    s_load_dwordx4 s[8:11], s[4:5], 0x20
; GFX940-SDAG-NEXT:    s_load_dword s7, s[4:5], 0x30
; GFX940-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX940-SDAG-NEXT:    v_mov_b64_e32 v[0:1], s[0:1]
; GFX940-SDAG-NEXT:    v_mov_b64_e32 v[2:3], s[2:3]
; GFX940-SDAG-NEXT:    scratch_store_dwordx4 off, v[0:3], off offset:32 sc0 sc1
; GFX940-SDAG-NEXT:    scratch_load_dwordx2 v[10:11], off, off offset:40
; GFX940-SDAG-NEXT:    scratch_load_dword v4, off, off offset:36
; GFX940-SDAG-NEXT:    v_mov_b64_e32 v[0:1], s[8:9]
; GFX940-SDAG-NEXT:    v_mov_b64_e32 v[2:3], s[10:11]
; GFX940-SDAG-NEXT:    scratch_store_dwordx4 off, v[0:3], off sc0 sc1
; GFX940-SDAG-NEXT:    scratch_load_dwordx2 v[12:13], off, off offset:8
; GFX940-SDAG-NEXT:    s_nop 0
; GFX940-SDAG-NEXT:    scratch_load_dword v0, off, off offset:4
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v7, s6
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v3, s7
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v9, s0
; GFX940-SDAG-NEXT:    s_mov_b64 s[2:3], exec
; GFX940-SDAG-NEXT:    s_waitcnt vmcnt(4)
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v6, v11
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v5, v10
; GFX940-SDAG-NEXT:    s_waitcnt vmcnt(1)
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v2, v13
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v1, v12
; GFX940-SDAG-NEXT:  .LBB1_1: ; =>This Inner Loop Header: Depth=1
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s4, v4
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s5, v5
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s6, v6
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s7, v7
; GFX940-SDAG-NEXT:    v_cmp_eq_u64_e32 vcc, s[4:5], v[4:5]
; GFX940-SDAG-NEXT:    s_nop 0
; GFX940-SDAG-NEXT:    v_cmp_eq_u64_e64 s[0:1], s[6:7], v[6:7]
; GFX940-SDAG-NEXT:    s_and_b64 s[0:1], vcc, s[0:1]
; GFX940-SDAG-NEXT:    s_and_saveexec_b64 s[0:1], s[0:1]
; GFX940-SDAG-NEXT:    buffer_load_dword v8, v9, s[4:7], 0 offen sc0 sc1
; GFX940-SDAG-NEXT:    ; implicit-def: $vgpr4_vgpr5_vgpr6_vgpr7
; GFX940-SDAG-NEXT:    ; implicit-def: $vgpr9
; GFX940-SDAG-NEXT:    s_xor_b64 exec, exec, s[0:1]
; GFX940-SDAG-NEXT:    s_cbranch_execnz .LBB1_1
; GFX940-SDAG-NEXT:  ; %bb.2:
; GFX940-SDAG-NEXT:    s_mov_b64 exec, s[2:3]
; GFX940-SDAG-NEXT:    v_mov_b32_e32 v4, s8
; GFX940-SDAG-NEXT:    s_mov_b64 s[0:1], exec
; GFX940-SDAG-NEXT:    s_waitcnt vmcnt(0)
; GFX940-SDAG-NEXT:  .LBB1_3: ; =>This Inner Loop Header: Depth=1
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s4, v0
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s5, v1
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s6, v2
; GFX940-SDAG-NEXT:    v_readfirstlane_b32 s7, v3
; GFX940-SDAG-NEXT:    v_cmp_eq_u64_e32 vcc, s[4:5], v[0:1]
; GFX940-SDAG-NEXT:    s_nop 0
; GFX940-SDAG-NEXT:    v_cmp_eq_u64_e64 s[0:1], s[6:7], v[2:3]
; GFX940-SDAG-NEXT:    s_and_b64 s[0:1], vcc, s[0:1]
; GFX940-SDAG-NEXT:    s_and_saveexec_b64 s[0:1], s[0:1]
; GFX940-SDAG-NEXT:    buffer_store_dword v8, v4, s[4:7], 0 offen sc0 sc1
; GFX940-SDAG-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX940-SDAG-NEXT:    ; implicit-def: $vgpr8
; GFX940-SDAG-NEXT:    ; implicit-def: $vgpr4
; GFX940-SDAG-NEXT:    s_xor_b64 exec, exec, s[0:1]
; GFX940-SDAG-NEXT:    s_cbranch_execnz .LBB1_3
; GFX940-SDAG-NEXT:  ; %bb.4:
; GFX940-SDAG-NEXT:    s_endpgm
;
; GFX940-GISEL-LABEL: buffer_nontemporal_and_volatile_load_store:
; GFX940-GISEL:       ; %bb.0: ; %entry
; GFX940-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
; GFX940-GISEL-NEXT:    s_load_dword s11, s[4:5], 0x10
; GFX940-GISEL-NEXT:    s_mov_b32 s7, 0
; GFX940-GISEL-NEXT:    s_mov_b32 s8, s7
; GFX940-GISEL-NEXT:    s_mov_b32 s10, s7
; GFX940-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX940-GISEL-NEXT:    s_mov_b32 s6, s1
; GFX940-GISEL-NEXT:    s_mov_b32 s9, s2
; GFX940-GISEL-NEXT:    s_or_b64 s[8:9], s[6:7], s[8:9]
; GFX940-GISEL-NEXT:    s_mov_b32 s6, s3
; GFX940-GISEL-NEXT:    s_or_b64 s[10:11], s[6:7], s[10:11]
; GFX940-GISEL-NEXT:    v_mov_b32_e32 v0, s0
; GFX940-GISEL-NEXT:    buffer_load_dword v0, v0, s[8:11], 0 offen sc0 sc1
; GFX940-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x20
; GFX940-GISEL-NEXT:    s_load_dword s9, s[4:5], 0x30
; GFX940-GISEL-NEXT:    s_mov_b32 s4, s7
; GFX940-GISEL-NEXT:    s_mov_b32 s8, s7
; GFX940-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX940-GISEL-NEXT:    s_mov_b32 s6, s1
; GFX940-GISEL-NEXT:    s_mov_b32 s5, s2
; GFX940-GISEL-NEXT:    s_or_b64 s[4:5], s[6:7], s[4:5]
; GFX940-GISEL-NEXT:    s_mov_b32 s6, s3
; GFX940-GISEL-NEXT:    s_or_b64 s[6:7], s[6:7], s[8:9]
; GFX940-GISEL-NEXT:    v_mov_b32_e32 v1, s0
; GFX940-GISEL-NEXT:    s_waitcnt vmcnt(0)
; GFX940-GISEL-NEXT:    buffer_store_dword v0, v1, s[4:7], 0 offen sc0 sc1
; GFX940-GISEL-NEXT:    s_endpgm
;
; GFX10-SDAG-LABEL: buffer_nontemporal_and_volatile_load_store:
; GFX10-SDAG:       ; %bb.0: ; %entry
; GFX10-SDAG-NEXT:    s_clause 0x1
; GFX10-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[8:9], 0x0
; GFX10-SDAG-NEXT:    s_load_dword s11, s[8:9], 0x10
; GFX10-SDAG-NEXT:    s_mov_b32 s10, 0
; GFX10-SDAG-NEXT:    s_add_u32 s0, s0, s15
; GFX10-SDAG-NEXT:    s_mov_b32 s13, s10
; GFX10-SDAG-NEXT:    s_addc_u32 s1, s1, 0
; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-SDAG-NEXT:    v_mov_b32_e32 v0, s4
; GFX10-SDAG-NEXT:    s_mov_b32 s12, s7
; GFX10-SDAG-NEXT:    s_or_b64 s[14:15], s[12:13], s[10:11]
; GFX10-SDAG-NEXT:    s_mov_b32 s12, s5
; GFX10-SDAG-NEXT:    s_mov_b32 s13, s6
; GFX10-SDAG-NEXT:    buffer_load_dword v0, v0, s[12:15], 0 offen glc dlc
; GFX10-SDAG-NEXT:    s_clause 0x1
; GFX10-SDAG-NEXT:    s_load_dword s11, s[8:9], 0x30
; GFX10-SDAG-NEXT:    s_load_dwordx4 s[4:7], s[8:9], 0x20
; GFX10-SDAG-NEXT:    s_mov_b32 s9, s10
; GFX10-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-SDAG-NEXT:    v_mov_b32_e32 v1, s4
; GFX10-SDAG-NEXT:    s_mov_b32 s8, s7
; GFX10-SDAG-NEXT:    s_or_b64 s[10:11], s[8:9], s[10:11]
; GFX10-SDAG-NEXT:    s_mov_b32 s8, s5
; GFX10-SDAG-NEXT:    s_mov_b32 s9, s6
; GFX10-SDAG-NEXT:    s_waitcnt vmcnt(0)
; GFX10-SDAG-NEXT:    buffer_store_dword v0, v1, s[8:11], 0 offen
; GFX10-SDAG-NEXT:    s_endpgm
;
; GFX10-GISEL-LABEL: buffer_nontemporal_and_volatile_load_store:
; GFX10-GISEL:       ; %bb.0: ; %entry
; GFX10-GISEL-NEXT:    s_clause 0x1
; GFX10-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x0
; GFX10-GISEL-NEXT:    s_load_dword s5, s[8:9], 0x10
; GFX10-GISEL-NEXT:    s_mov_b32 s7, 0
; GFX10-GISEL-NEXT:    s_mov_b32 s10, s7
; GFX10-GISEL-NEXT:    s_mov_b32 s4, s7
; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-GISEL-NEXT:    s_mov_b32 s6, s1
; GFX10-GISEL-NEXT:    s_mov_b32 s11, s2
; GFX10-GISEL-NEXT:    v_mov_b32_e32 v0, s0
; GFX10-GISEL-NEXT:    s_or_b64 s[0:1], s[6:7], s[10:11]
; GFX10-GISEL-NEXT:    s_mov_b32 s6, s3
; GFX10-GISEL-NEXT:    s_or_b64 s[2:3], s[6:7], s[4:5]
; GFX10-GISEL-NEXT:    buffer_load_dword v0, v0, s[0:3], 0 offen glc dlc
; GFX10-GISEL-NEXT:    s_clause 0x1
; GFX10-GISEL-NEXT:    s_waitcnt_depctr 0xffe3
; GFX10-GISEL-NEXT:    s_load_dwordx4 s[0:3], s[8:9], 0x20
; GFX10-GISEL-NEXT:    s_load_dword s11, s[8:9], 0x30
; GFX10-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX10-GISEL-NEXT:    s_mov_b32 s6, s1
; GFX10-GISEL-NEXT:    s_mov_b32 s5, s2
; GFX10-GISEL-NEXT:    v_mov_b32_e32 v1, s0
; GFX10-GISEL-NEXT:    s_or_b64 s[4:5], s[6:7], s[4:5]
; GFX10-GISEL-NEXT:    s_mov_b32 s6, s3
; GFX10-GISEL-NEXT:    s_or_b64 s[6:7], s[6:7], s[10:11]
; GFX10-GISEL-NEXT:    s_waitcnt vmcnt(0)
; GFX10-GISEL-NEXT:    buffer_store_dword v0, v1, s[4:7], 0 offen
; GFX10-GISEL-NEXT:    s_endpgm
;
; GFX11-SDAG-LABEL: buffer_nontemporal_and_volatile_load_store:
; GFX11-SDAG:       ; %bb.0: ; %entry
; GFX11-SDAG-NEXT:    s_clause 0x2
; GFX11-SDAG-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
; GFX11-SDAG-NEXT:    s_load_b128 s[8:11], s[4:5], 0x20
; GFX11-SDAG-NEXT:    s_load_b32 s6, s[4:5], 0x10
; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-SDAG-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-SDAG-NEXT:    v_dual_mov_b32 v7, s8 :: v_dual_mov_b32 v8, s9
; GFX11-SDAG-NEXT:    v_dual_mov_b32 v9, s10 :: v_dual_mov_b32 v10, s11
; GFX11-SDAG-NEXT:    scratch_store_b128 off, v[0:3], off offset:32
; GFX11-SDAG-NEXT:    s_clause 0x1
; GFX11-SDAG-NEXT:    scratch_load_b64 v[5:6], off, off offset:40
; GFX11-SDAG-NEXT:    scratch_load_b32 v4, off, off offset:36
; GFX11-SDAG-NEXT:    s_load_b32 s1, s[4:5], 0x30
; GFX11-SDAG-NEXT:    scratch_store_b128 off, v[7:10], off
; GFX11-SDAG-NEXT:    s_clause 0x1
; GFX11-SDAG-NEXT:    scratch_load_b64 v[1:2], off, off offset:8
; GFX11-SDAG-NEXT:    scratch_load_b32 v0, off, off offset:4
; GFX11-SDAG-NEXT:    v_mov_b32_e32 v7, s6
; GFX11-SDAG-NEXT:    v_mov_b32_e32 v9, s0
; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, s1
; GFX11-SDAG-NEXT:    s_mov_b32 s1, exec_lo
; GFX11-SDAG-NEXT:    s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT:  .LBB1_1: ; =>This Inner Loop Header: Depth=1
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s4, v4
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s5, v5
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s6, v6
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s7, v7
; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-SDAG-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[4:5]
; GFX11-SDAG-NEXT:    v_cmp_eq_u64_e64 s0, s[6:7], v[6:7]
; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT:    s_and_b32 s0, vcc_lo, s0
; GFX11-SDAG-NEXT:    s_and_saveexec_b32 s0, s0
; GFX11-SDAG-NEXT:    buffer_load_b32 v8, v9, s[4:7], 0 offen glc dlc
; GFX11-SDAG-NEXT:    ; implicit-def: $vgpr4_vgpr5_vgpr6_vgpr7
; GFX11-SDAG-NEXT:    ; implicit-def: $vgpr9
; GFX11-SDAG-NEXT:    s_xor_b32 exec_lo, exec_lo, s0
; GFX11-SDAG-NEXT:    s_cbranch_execnz .LBB1_1
; GFX11-SDAG-NEXT:  ; %bb.2:
; GFX11-SDAG-NEXT:    s_mov_b32 exec_lo, s1
; GFX11-SDAG-NEXT:    v_mov_b32_e32 v4, s8
; GFX11-SDAG-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-SDAG-NEXT:  .LBB1_3: ; =>This Inner Loop Header: Depth=1
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s4, v0
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s5, v1
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s6, v2
; GFX11-SDAG-NEXT:    v_readfirstlane_b32 s7, v3
; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-SDAG-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
; GFX11-SDAG-NEXT:    v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
; GFX11-SDAG-NEXT:    s_and_b32 s0, vcc_lo, s0
; GFX11-SDAG-NEXT:    s_and_saveexec_b32 s0, s0
; GFX11-SDAG-NEXT:    s_waitcnt vmcnt(0)
; GFX11-SDAG-NEXT:    buffer_store_b32 v8, v4, s[4:7], 0 offen dlc
; GFX11-SDAG-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX11-SDAG-NEXT:    ; implicit-def: $vgpr8
; GFX11-SDAG-NEXT:    ; implicit-def: $vgpr4
; GFX11-SDAG-NEXT:    s_xor_b32 exec_lo, exec_lo, s0
; GFX11-SDAG-NEXT:    s_cbranch_execnz .LBB1_3
; GFX11-SDAG-NEXT:  ; %bb.4:
; GFX11-SDAG-NEXT:    s_endpgm
;
; GFX11-GISEL-LABEL: buffer_nontemporal_and_volatile_load_store:
; GFX11-GISEL:       ; %bb.0: ; %entry
; GFX11-GISEL-NEXT:    s_clause 0x1
; GFX11-GISEL-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
; GFX11-GISEL-NEXT:    s_load_b32 s7, s[4:5], 0x10
; GFX11-GISEL-NEXT:    s_mov_b32 s9, 0
; GFX11-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT:    s_mov_b32 s10, s9
; GFX11-GISEL-NEXT:    s_mov_b32 s6, s9
; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT:    s_mov_b32 s8, s1
; GFX11-GISEL-NEXT:    s_mov_b32 s11, s2
; GFX11-GISEL-NEXT:    v_mov_b32_e32 v0, s0
; GFX11-GISEL-NEXT:    s_or_b64 s[0:1], s[8:9], s[10:11]
; GFX11-GISEL-NEXT:    s_mov_b32 s8, s3
; GFX11-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-GISEL-NEXT:    s_or_b64 s[2:3], s[8:9], s[6:7]
; GFX11-GISEL-NEXT:    buffer_load_b32 v0, v0, s[0:3], 0 offen glc dlc
; GFX11-GISEL-NEXT:    s_clause 0x1
; GFX11-GISEL-NEXT:    s_load_b128 s[0:3], s[4:5], 0x20
; GFX11-GISEL-NEXT:    s_load_b32 s7, s[4:5], 0x30
; GFX11-GISEL-NEXT:    s_mov_b32 s4, s9
; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, s0
; GFX11-GISEL-NEXT:    s_mov_b32 s8, s1
; GFX11-GISEL-NEXT:    s_mov_b32 s5, s2
; GFX11-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-GISEL-NEXT:    s_or_b64 s[4:5], s[8:9], s[4:5]
; GFX11-GISEL-NEXT:    s_mov_b32 s8, s3
; GFX11-GISEL-NEXT:    s_or_b64 s[6:7], s[8:9], s[6:7]
; GFX11-GISEL-NEXT:    s_waitcnt vmcnt(0)
; GFX11-GISEL-NEXT:    buffer_store_b32 v0, v1, s[4:7], 0 offen dlc
; GFX11-GISEL-NEXT:    s_endpgm
;
; GFX12-SDAG-LABEL: buffer_nontemporal_and_volatile_load_store:
; GFX12-SDAG:       ; %bb.0: ; %entry
; GFX12-SDAG-NEXT:    s_clause 0x2
; GFX12-SDAG-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
; GFX12-SDAG-NEXT:    s_load_b128 s[8:11], s[4:5], 0x20
; GFX12-SDAG-NEXT:    s_load_b32 s6, s[4:5], 0x10
; GFX12-SDAG-NEXT:    s_wait_kmcnt 0x0
; GFX12-SDAG-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX12-SDAG-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX12-SDAG-NEXT:    v_dual_mov_b32 v7, s8 :: v_dual_mov_b32 v8, s9
; GFX12-SDAG-NEXT:    v_dual_mov_b32 v9, s10 :: v_dual_mov_b32 v10, s11
; GFX12-SDAG-NEXT:    scratch_store_b128 off, v[0:3], off offset:32
; GFX12-SDAG-NEXT:    s_clause 0x1
; GFX12-SDAG-NEXT:    scratch_load_b64 v[5:6], off, off offset:40
; GFX12-SDAG-NEXT:    scratch_load_b32 v4, off, off offset:36
; GFX12-SDAG-NEXT:    s_load_b32 s1, s[4:5], 0x30
; GFX12-SDAG-NEXT:    scratch_store_b128 off, v[7:10], off
; GFX12-SDAG-NEXT:    s_clause 0x1
; GFX12-SDAG-NEXT:    scratch_load_b64 v[1:2], off, off offset:8
; GFX12-SDAG-NEXT:    scratch_load_b32 v0, off, off offset:4
; GFX12-SDAG-NEXT:    v_mov_b32_e32 v7, s6
; GFX12-SDAG-NEXT:    v_mov_b32_e32 v9, s0
; GFX12-SDAG-NEXT:    s_wait_kmcnt 0x0
; GFX12-SDAG-NEXT:    v_mov_b32_e32 v3, s1
; GFX12-SDAG-NEXT:    s_mov_b32 s1, exec_lo
; GFX12-SDAG-NEXT:  .LBB1_1: ; =>This Inner Loop Header: Depth=1
; GFX12-SDAG-NEXT:    s_wait_loadcnt 0x2
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s4, v4
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s5, v5
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s6, v6
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s7, v7
; GFX12-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-SDAG-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[4:5]
; GFX12-SDAG-NEXT:    v_cmp_eq_u64_e64 s0, s[6:7], v[6:7]
; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
; GFX12-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-SDAG-NEXT:    s_and_b32 s0, vcc_lo, s0
; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
; GFX12-SDAG-NEXT:    s_and_saveexec_b32 s0, s0
; GFX12-SDAG-NEXT:    s_wait_loadcnt 0x0
; GFX12-SDAG-NEXT:    buffer_load_b32 v8, v9, s[4:7], null offen th:TH_LOAD_NT scope:SCOPE_SYS
; GFX12-SDAG-NEXT:    ; implicit-def: $vgpr4_vgpr5_vgpr6_vgpr7
; GFX12-SDAG-NEXT:    ; implicit-def: $vgpr9
; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
; GFX12-SDAG-NEXT:    s_xor_b32 exec_lo, exec_lo, s0
; GFX12-SDAG-NEXT:    s_cbranch_execnz .LBB1_1
; GFX12-SDAG-NEXT:  ; %bb.2:
; GFX12-SDAG-NEXT:    s_mov_b32 exec_lo, s1
; GFX12-SDAG-NEXT:    v_mov_b32_e32 v4, s8
; GFX12-SDAG-NEXT:    s_mov_b32 s0, exec_lo
; GFX12-SDAG-NEXT:  .LBB1_3: ; =>This Inner Loop Header: Depth=1
; GFX12-SDAG-NEXT:    s_wait_loadcnt 0x1
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s4, v0
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s5, v1
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s6, v2
; GFX12-SDAG-NEXT:    v_readfirstlane_b32 s7, v3
; GFX12-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX12-SDAG-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[4:5], v[0:1]
; GFX12-SDAG-NEXT:    v_cmp_eq_u64_e64 s0, s[6:7], v[2:3]
; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
; GFX12-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX12-SDAG-NEXT:    s_and_b32 s0, vcc_lo, s0
; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
; GFX12-SDAG-NEXT:    s_and_saveexec_b32 s0, s0
; GFX12-SDAG-NEXT:    s_wait_loadcnt 0x0
; GFX12-SDAG-NEXT:    s_wait_storecnt 0x0
; GFX12-SDAG-NEXT:    buffer_store_b32 v8, v4, s[4:7], null offen th:TH_STORE_NT scope:SCOPE_SYS
; GFX12-SDAG-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3
; GFX12-SDAG-NEXT:    ; implicit-def: $vgpr8
; GFX12-SDAG-NEXT:    ; implicit-def: $vgpr4
; GFX12-SDAG-NEXT:    s_wait_alu 0xfffe
; GFX12-SDAG-NEXT:    s_xor_b32 exec_lo, exec_lo, s0
; GFX12-SDAG-NEXT:    s_cbranch_execnz .LBB1_3
; GFX12-SDAG-NEXT:  ; %bb.4:
; GFX12-SDAG-NEXT:    s_endpgm
;
; GFX12-GISEL-LABEL: buffer_nontemporal_and_volatile_load_store:
; GFX12-GISEL:       ; %bb.0: ; %entry
; GFX12-GISEL-NEXT:    s_clause 0x1
; GFX12-GISEL-NEXT:    s_load_b128 s[0:3], s[4:5], 0x0
; GFX12-GISEL-NEXT:    s_load_b32 s7, s[4:5], 0x10
; GFX12-GISEL-NEXT:    s_mov_b32 s9, 0
; GFX12-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-GISEL-NEXT:    s_mov_b32 s10, s9
; GFX12-GISEL-NEXT:    s_mov_b32 s6, s9
; GFX12-GISEL-NEXT:    s_wait_kmcnt 0x0
; GFX12-GISEL-NEXT:    s_mov_b32 s8, s1
; GFX12-GISEL-NEXT:    s_mov_b32 s11, s2
; GFX12-GISEL-NEXT:    v_mov_b32_e32 v0, s0
; GFX12-GISEL-NEXT:    s_or_b64 s[0:1], s[8:9], s[10:11]
; GFX12-GISEL-NEXT:    s_mov_b32 s8, s3
; GFX12-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX12-GISEL-NEXT:    s_or_b64 s[2:3], s[8:9], s[6:7]
; GFX12-GISEL-NEXT:    buffer_load_b32 v0, v0, s[0:3], null offen th:TH_LOAD_NT scope:SCOPE_SYS
; GFX12-GISEL-NEXT:    s_clause 0x1
; GFX12-GISEL-NEXT:    s_load_b128 s[0:3], s[4:5], 0x20
; GFX12-GISEL-NEXT:    s_load_b32 s7, s[4:5], 0x30
; GFX12-GISEL-NEXT:    s_mov_b32 s4, s9
; GFX12-GISEL-NEXT:    s_wait_kmcnt 0x0
; GFX12-GISEL-NEXT:    s_wait_alu 0xfffe
; GFX12-GISEL-NEXT:    v_mov_b32_e32 v1, s0
; GFX12-GISEL-NEXT:    s_mov_b32 s8, s1
; GFX12-GISEL-NEXT:    s_mov_b32 s5, s2
; GFX12-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX12-GISEL-NEXT:    s_or_b64 s[4:5], s[8:9], s[4:5]
; GFX12-GISEL-NEXT:    s_mov_b32 s8, s3
; GFX12-GISEL-NEXT:    s_or_b64 s[6:7], s[8:9], s[6:7]
; GFX12-GISEL-NEXT:    s_wait_loadcnt 0x0
; GFX12-GISEL-NEXT:    buffer_store_b32 v0, v1, s[4:7], null offen th:TH_STORE_NT scope:SCOPE_SYS
; GFX12-GISEL-NEXT:    s_endpgm
entry:
  %val = load volatile i32, ptr addrspace(7) %in, !nontemporal !0
  store volatile i32 %val, ptr addrspace(7) %out, !nontemporal !0
  ret void
}

!0 = !{i32 1}
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; GFX10: {{.*}}
; GFX11: {{.*}}
; GFX12: {{.*}}
; GFX9: {{.*}}
; GFX940: {{.*}}